From patchwork Tue Mar 7 14:25:26 2023 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 8bit X-Patchwork-Submitter: Asahi Lina X-Patchwork-Id: 13163697 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.4.0 (2014-02-07) on aws-us-west-2-korg-lkml-1.web.codeaurora.org Received: from gabe.freedesktop.org (gabe.freedesktop.org [131.252.210.177]) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (No client certificate requested) by smtp.lore.kernel.org (Postfix) with ESMTPS id 7570EC678D5 for ; Tue, 7 Mar 2023 14:26:53 +0000 (UTC) Received: from gabe.freedesktop.org (localhost [127.0.0.1]) by gabe.freedesktop.org (Postfix) with ESMTP id 10C2510E4A6; Tue, 7 Mar 2023 14:26:52 +0000 (UTC) Received: from mail.marcansoft.com (marcansoft.com [212.63.210.85]) by gabe.freedesktop.org (Postfix) with ESMTPS id 2A89710E4A1 for ; Tue, 7 Mar 2023 14:26:49 +0000 (UTC) Received: from [127.0.0.1] (localhost [127.0.0.1]) (using TLSv1.3 with cipher TLS_AES_256_GCM_SHA384 (256/256 bits) key-exchange X25519 server-signature RSA-PSS (4096 bits) server-digest SHA256) (No client certificate requested) (Authenticated sender: linasend@asahilina.net) by mail.marcansoft.com (Postfix) with ESMTPSA id 81BB142058; Tue, 7 Mar 2023 14:26:40 +0000 (UTC) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/simple; d=asahilina.net; s=default; t=1678199207; bh=odEitMmgQnLv6N8I5beA3x4xEAfKw5Q0o5W+1zb1Uq0=; h=From:Date:Subject:References:In-Reply-To:To:Cc; b=RC0KNZdRAjm8I6dSNwqjpDJWYHk41dpDkvn1rl6ayJTvGPU1ivz3uCnyKA6WQy/n1 R60Zce/vvBHuycoyN3hawXRPjGiwzLplHZbKrY8XWrRpJC1z6uRT0zp4hXKXKEMYz4 ppO7x3pDfCTBzt2pEd3dgZJXKlN3hyWqDa6wkGCypMmXLAtxw+M+YoYG21bI+drXYR bAvp1yvGdGMzlV/5JE0P3bkmKnhKyr5Uyy58J7zNpC6uZNFBgNCQ4YVKa+wH8Ppmza 9iJ1EqjZ+pSZ3vRoyX2/GZMkELY03THlB6TTOJ6uqM6tSgSCWD+O61EUgj56wgwlFH pE1feT/nwHBIQ== From: Asahi Lina Date: Tue, 07 Mar 2023 23:25:26 +0900 Subject: [PATCH RFC 01/18] rust: drm: ioctl: Add DRM ioctl abstraction MIME-Version: 1.0 Message-Id: <20230307-rust-drm-v1-1-917ff5bc80a8@asahilina.net> References: <20230307-rust-drm-v1-0-917ff5bc80a8@asahilina.net> In-Reply-To: <20230307-rust-drm-v1-0-917ff5bc80a8@asahilina.net> To: Maarten Lankhorst , Maxime Ripard , Thomas Zimmermann , David Airlie , Daniel Vetter , Miguel Ojeda , Alex Gaynor , Wedson Almeida Filho , Boqun Feng , Gary Guo , =?utf-8?q?Bj=C3=B6rn_Roy_Baron?= , Sumit Semwal , =?utf-8?q?Christian_K=C3=B6nig?= , Luben Tuikov , Jarkko Sakkinen , Dave Hansen X-Mailer: b4 0.12.0 X-Developer-Signature: v=1; a=ed25519-sha256; t=1678199191; l=9667; i=lina@asahilina.net; s=20230221; h=from:subject:message-id; bh=odEitMmgQnLv6N8I5beA3x4xEAfKw5Q0o5W+1zb1Uq0=; b=pVDYzlEsrw9so/OnmS15cqXJNcC01JKoowKgcvrbpThVxIO7BRruz7od9ntExpWQktPx4YCXO EwvtHQIBKysAyqQfsbYwJ6r30+3rae8H4xWsD5WZClGGRTizDqcEogo X-Developer-Key: i=lina@asahilina.net; a=ed25519; pk=Qn8jZuOtR1m5GaiDfTrAoQ4NE1XoYVZ/wmt5YtXWFC4= X-BeenThere: dri-devel@lists.freedesktop.org X-Mailman-Version: 2.1.29 Precedence: list List-Id: Direct Rendering Infrastructure - Development List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Cc: linaro-mm-sig@lists.linaro.org, rust-for-linux@vger.kernel.org, Karol Herbst , asahi@lists.linux.dev, Asahi Lina , linux-kernel@vger.kernel.org, dri-devel@lists.freedesktop.org, Mary , Alyssa Rosenzweig , linux-sgx@vger.kernel.org, Ella Stanforth , Faith Ekstrand , linux-media@vger.kernel.org Errors-To: dri-devel-bounces@lists.freedesktop.org Sender: "dri-devel" DRM drivers need to be able to declare which driver-specific ioctls they support. This abstraction adds the required types and a helper macro to generate the ioctl definition inside the DRM driver. Note that this macro is not usable until further bits of the abstraction are in place (but it will not fail to compile on its own, if not called). Signed-off-by: Asahi Lina --- drivers/gpu/drm/Kconfig | 7 ++ rust/bindings/bindings_helper.h | 2 + rust/kernel/drm/ioctl.rs | 147 ++++++++++++++++++++++++++++++++++++++++ rust/kernel/drm/mod.rs | 5 ++ rust/kernel/lib.rs | 2 + 5 files changed, 163 insertions(+) diff --git a/drivers/gpu/drm/Kconfig b/drivers/gpu/drm/Kconfig index dc0f94f02a82..dab8f0f9aa96 100644 --- a/drivers/gpu/drm/Kconfig +++ b/drivers/gpu/drm/Kconfig @@ -27,6 +27,13 @@ menuconfig DRM details. You should also select and configure AGP (/dev/agpgart) support if it is available for your platform. +# Rust abstractions cannot be built as modules currently, so force them as +# bool by using these intermediate symbols. In the future these could be +# tristate once abstractions themselves can be built as modules. +config RUST_DRM + bool "Rust support for the DRM subsystem" + depends on DRM=y + config DRM_MIPI_DBI tristate depends on DRM diff --git a/rust/bindings/bindings_helper.h b/rust/bindings/bindings_helper.h index 91bb7906ca5a..2687bef1676f 100644 --- a/rust/bindings/bindings_helper.h +++ b/rust/bindings/bindings_helper.h @@ -6,6 +6,7 @@ * Sorted alphabetically. */ +#include #include #include #include @@ -23,6 +24,7 @@ #include #include #include +#include /* `bindgen` gets confused at certain things. */ const gfp_t BINDINGS_GFP_KERNEL = GFP_KERNEL; diff --git a/rust/kernel/drm/ioctl.rs b/rust/kernel/drm/ioctl.rs new file mode 100644 index 000000000000..10304efbd5f1 --- /dev/null +++ b/rust/kernel/drm/ioctl.rs @@ -0,0 +1,147 @@ +// SPDX-License-Identifier: GPL-2.0 OR MIT +#![allow(non_snake_case)] + +//! DRM IOCTL definitions. +//! +//! C header: [`include/linux/drm/drm_ioctl.h`](../../../../include/linux/drm/drm_ioctl.h) + +use crate::ioctl; + +const BASE: u32 = bindings::DRM_IOCTL_BASE as u32; + +/// Construct a DRM ioctl number with no argument. +pub const fn IO(nr: u32) -> u32 { + ioctl::_IO(BASE, nr) +} + +/// Construct a DRM ioctl number with a read-only argument. +pub const fn IOR(nr: u32) -> u32 { + ioctl::_IOR::(BASE, nr) +} + +/// Construct a DRM ioctl number with a write-only argument. +pub const fn IOW(nr: u32) -> u32 { + ioctl::_IOW::(BASE, nr) +} + +/// Construct a DRM ioctl number with a read-write argument. +pub const fn IOWR(nr: u32) -> u32 { + ioctl::_IOWR::(BASE, nr) +} + +/// Descriptor type for DRM ioctls. Use the `declare_drm_ioctls!{}` macro to construct them. +pub type DrmIoctlDescriptor = bindings::drm_ioctl_desc; + +/// This is for ioctl which are used for rendering, and require that the file descriptor is either +/// for a render node, or if it’s a legacy/primary node, then it must be authenticated. +pub const AUTH: u32 = bindings::drm_ioctl_flags_DRM_AUTH; + +/// This must be set for any ioctl which can change the modeset or display state. Userspace must +/// call the ioctl through a primary node, while it is the active master. +/// +/// Note that read-only modeset ioctl can also be called by unauthenticated clients, or when a +/// master is not the currently active one. +pub const MASTER: u32 = bindings::drm_ioctl_flags_DRM_MASTER; + +/// Anything that could potentially wreak a master file descriptor needs to have this flag set. +/// +/// Current that’s only for the SETMASTER and DROPMASTER ioctl, which e.g. logind can call to force +/// a non-behaving master (display compositor) into compliance. +/// +/// This is equivalent to callers with the SYSADMIN capability. +pub const ROOT_ONLY: u32 = bindings::drm_ioctl_flags_DRM_ROOT_ONLY; + +/// Whether drm_ioctl_desc.func should be called with the DRM BKL held or not. Enforced as the +/// default for all modern drivers, hence there should never be a need to set this flag. +/// +/// Do not use anywhere else than for the VBLANK_WAIT IOCTL, which is the only legacy IOCTL which +/// needs this. +pub const UNLOCKED: u32 = bindings::drm_ioctl_flags_DRM_UNLOCKED; + +/// This is used for all ioctl needed for rendering only, for drivers which support render nodes. +/// This should be all new render drivers, and hence it should be always set for any ioctl with +/// `AUTH` set. Note though that read-only query ioctl might have this set, but have not set +/// DRM_AUTH because they do not require authentication. +pub const RENDER_ALLOW: u32 = bindings::drm_ioctl_flags_DRM_RENDER_ALLOW; + +/// Declare the DRM ioctls for a driver. +/// +/// Each entry in the list should have the form: +/// +/// `(ioctl_number, argument_type, flags, user_callback),` +/// +/// `argument_type` is the type name within the `bindings` crate. +/// `user_callback` should have the following prototype: +/// +/// ``` +/// fn foo(device: &kernel::drm::device::Device, +/// data: &mut bindings::argument_type, +/// file: &kernel::drm::file::File, +/// ) +/// ``` +/// where `Self` is the drm::drv::Driver implementation these ioctls are being declared within. +/// +/// # Examples +/// +/// ``` +/// kernel::declare_drm_ioctls! { +/// (FOO_GET_PARAM, drm_foo_get_param, ioctl::RENDER_ALLOW, my_get_param_handler), +/// } +/// ``` +/// +#[macro_export] +macro_rules! declare_drm_ioctls { + ( $(($cmd:ident, $struct:ident, $flags:expr, $func:expr)),* $(,)? ) => { + const IOCTLS: &'static [$crate::drm::ioctl::DrmIoctlDescriptor] = { + const _:() = { + let i: u32 = $crate::bindings::DRM_COMMAND_BASE; + // Assert that all the IOCTLs are in the right order and there are no gaps, + // and that the sizeof of the specified type is correct. + $( + let cmd: u32 = $crate::macros::concat_idents!($crate::bindings::DRM_IOCTL_, $cmd); + ::core::assert!(i == $crate::ioctl::_IOC_NR(cmd)); + ::core::assert!(core::mem::size_of::<$crate::bindings::$struct>() == $crate::ioctl::_IOC_SIZE(cmd)); + let i: u32 = i + 1; + )* + }; + + let ioctls = &[$( + $crate::bindings::drm_ioctl_desc { + cmd: $crate::macros::concat_idents!($crate::bindings::DRM_IOCTL_, $cmd) as u32, + func: { + #[allow(non_snake_case)] + unsafe extern "C" fn $cmd( + raw_dev: *mut $crate::bindings::drm_device, + raw_data: *mut ::core::ffi::c_void, + raw_file_priv: *mut $crate::bindings::drm_file, + ) -> core::ffi::c_int { + // SAFETY: We never drop this, and the DRM core ensures the device lives + // while callbacks are being called. + // + // FIXME: Currently there is nothing enforcing that the types of the + // dev/file match the current driver these ioctls are being declared + // for, and it's not clear how to enforce this within the type system. + let dev = ::core::mem::ManuallyDrop::new(unsafe { + $crate::drm::device::Device::from_raw(raw_dev) + }); + // SAFETY: This is just the ioctl argument, which hopefully has the right type + // (we've done our best checking the size). + let data = unsafe { &mut *(raw_data as *mut $crate::bindings::$struct) }; + // SAFETY: This is just the DRM file structure + let file = unsafe { $crate::drm::file::File::from_raw(raw_file_priv) }; + + match $func(&*dev, data, &file) { + Err(e) => e.to_kernel_errno(), + Ok(i) => i.try_into().unwrap_or(ERANGE.to_kernel_errno()), + } + } + Some($cmd) + }, + flags: $flags, + name: $crate::c_str!(::core::stringify!($cmd)).as_char_ptr(), + } + ),*]; + ioctls + }; + }; +} diff --git a/rust/kernel/drm/mod.rs b/rust/kernel/drm/mod.rs new file mode 100644 index 000000000000..9ec6d7cbcaf3 --- /dev/null +++ b/rust/kernel/drm/mod.rs @@ -0,0 +1,5 @@ +// SPDX-License-Identifier: GPL-2.0 OR MIT + +//! DRM subsystem abstractions. + +pub mod ioctl; diff --git a/rust/kernel/lib.rs b/rust/kernel/lib.rs index 7903490816bf..cb23d24c6718 100644 --- a/rust/kernel/lib.rs +++ b/rust/kernel/lib.rs @@ -37,6 +37,8 @@ mod build_assert; pub mod delay; pub mod device; pub mod driver; +#[cfg(CONFIG_RUST_DRM)] +pub mod drm; pub mod error; pub mod io_buffer; pub mod io_mem; From patchwork Tue Mar 7 14:25:27 2023 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Asahi Lina X-Patchwork-Id: 13163698 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.4.0 (2014-02-07) on aws-us-west-2-korg-lkml-1.web.codeaurora.org Received: from gabe.freedesktop.org (gabe.freedesktop.org [131.252.210.177]) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (No client certificate requested) by smtp.lore.kernel.org (Postfix) with ESMTPS id B14B5C678D5 for ; Tue, 7 Mar 2023 14:27:00 +0000 (UTC) Received: from gabe.freedesktop.org (localhost [127.0.0.1]) by gabe.freedesktop.org (Postfix) with ESMTP id CF7BA10E4A8; Tue, 7 Mar 2023 14:26:59 +0000 (UTC) Received: from mail.marcansoft.com (marcansoft.com [212.63.210.85]) by gabe.freedesktop.org (Postfix) with ESMTPS id ADE1110E4A8 for ; Tue, 7 Mar 2023 14:26:56 +0000 (UTC) Received: from [127.0.0.1] (localhost [127.0.0.1]) (using TLSv1.3 with cipher TLS_AES_256_GCM_SHA384 (256/256 bits) key-exchange X25519 server-signature RSA-PSS (4096 bits) server-digest SHA256) (No client certificate requested) (Authenticated sender: linasend@asahilina.net) by mail.marcansoft.com (Postfix) with ESMTPSA id EFF01420D8; Tue, 7 Mar 2023 14:26:47 +0000 (UTC) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/simple; d=asahilina.net; s=default; t=1678199215; bh=ybXc676YDutsjlMaSuuenGoZJO/fukuHqULLdgV/v/M=; h=From:Date:Subject:References:In-Reply-To:To:Cc; b=OpjjgA7LC7q4Glsexbn+nW/mRuBv6rhD+pd1NYciXMLw6u+H1RakS8DC5sXggG0cG Yd7SjxsIlxh0AkS1OthspkDO4sz81hphi9x5hXBrb9l/LKUV4XxP5Dr6x0Bv1VN7QJ JPOYSKEkvlAHwMATnLAjSaRiCKmnokEd6OTQ0Mu0ay6GhsN1g+hUgG/simriZqZCsB XPcBmLMCgB8rUy+OQn8gKjvKQg1SAGrorvM6lQgXslEQ3cdS94LXjOjhrOIoaeqrUB ASxFLNkLA2ydu5CGN9DDkJlr3E2Ck/1gD0mzAb3jDjmuc4YM1C3WPKe/AianXlSmwS Cg3r/1An4X4Dg== From: Asahi Lina Date: Tue, 07 Mar 2023 23:25:27 +0900 Subject: [PATCH RFC 02/18] rust: drm: Add Device and Driver abstractions MIME-Version: 1.0 Message-Id: <20230307-rust-drm-v1-2-917ff5bc80a8@asahilina.net> References: <20230307-rust-drm-v1-0-917ff5bc80a8@asahilina.net> In-Reply-To: <20230307-rust-drm-v1-0-917ff5bc80a8@asahilina.net> To: Maarten Lankhorst , Maxime Ripard , Thomas Zimmermann , David Airlie , Daniel Vetter , Miguel Ojeda , Alex Gaynor , Wedson Almeida Filho , Boqun Feng , Gary Guo , =?utf-8?q?Bj=C3=B6rn_Roy_Baron?= , Sumit Semwal , =?utf-8?q?Christian_K=C3=B6nig?= , Luben Tuikov , Jarkko Sakkinen , Dave Hansen X-Mailer: b4 0.12.0 X-Developer-Signature: v=1; a=ed25519-sha256; t=1678199191; l=16844; i=lina@asahilina.net; s=20230221; h=from:subject:message-id; bh=ybXc676YDutsjlMaSuuenGoZJO/fukuHqULLdgV/v/M=; b=ZAJvSXMpknOJW8GmoecAuTMvhH+w8WU/LojoH//ra2xRSLnc2949i7/QPHqsvsnwzxoW56X/r CCeFhXwJVd9AxNJFrQPow+MTQsL1NH1jvQhvrIsCC5yTlyWY2NPGHs8 X-Developer-Key: i=lina@asahilina.net; a=ed25519; pk=Qn8jZuOtR1m5GaiDfTrAoQ4NE1XoYVZ/wmt5YtXWFC4= X-BeenThere: dri-devel@lists.freedesktop.org X-Mailman-Version: 2.1.29 Precedence: list List-Id: Direct Rendering Infrastructure - Development List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Cc: linaro-mm-sig@lists.linaro.org, rust-for-linux@vger.kernel.org, Karol Herbst , asahi@lists.linux.dev, Asahi Lina , linux-kernel@vger.kernel.org, dri-devel@lists.freedesktop.org, Mary , Alyssa Rosenzweig , linux-sgx@vger.kernel.org, Ella Stanforth , Faith Ekstrand , linux-media@vger.kernel.org Errors-To: dri-devel-bounces@lists.freedesktop.org Sender: "dri-devel" Add the initial abstractions for DRM drivers and devices. These go together in one commit since they are fairly tightly coupled types. A few things have been stubbed out, to be implemented as further bits of the DRM subsystem are introduced. Signed-off-by: Asahi Lina --- rust/bindings/bindings_helper.h | 3 + rust/kernel/drm/device.rs | 76 +++++++++ rust/kernel/drm/drv.rs | 339 ++++++++++++++++++++++++++++++++++++++++ rust/kernel/drm/mod.rs | 2 + 4 files changed, 420 insertions(+) diff --git a/rust/bindings/bindings_helper.h b/rust/bindings/bindings_helper.h index 2687bef1676f..2a999138c4ae 100644 --- a/rust/bindings/bindings_helper.h +++ b/rust/bindings/bindings_helper.h @@ -6,10 +6,13 @@ * Sorted alphabetically. */ +#include +#include #include #include #include #include +#include #include #include #include diff --git a/rust/kernel/drm/device.rs b/rust/kernel/drm/device.rs new file mode 100644 index 000000000000..6007f941137a --- /dev/null +++ b/rust/kernel/drm/device.rs @@ -0,0 +1,76 @@ +// SPDX-License-Identifier: GPL-2.0 OR MIT + +//! DRM device. +//! +//! C header: [`include/linux/drm/drm_device.h`](../../../../include/linux/drm/drm_device.h) + +use crate::{bindings, device, drm, types::ForeignOwnable}; +use core::marker::PhantomData; + +/// Represents a reference to a DRM device. The device is reference-counted and is guaranteed to +/// not be dropped while this object is alive. +pub struct Device { + // Type invariant: ptr must be a valid and initialized drm_device, + // and this value must either own a reference to it or the caller + // must ensure that it is never dropped if the reference is borrowed. + pub(super) ptr: *mut bindings::drm_device, + _p: PhantomData, +} + +impl Device { + // Not intended to be called externally, except via declare_drm_ioctls!() + #[doc(hidden)] + pub unsafe fn from_raw(raw: *mut bindings::drm_device) -> Device { + Device { + ptr: raw, + _p: PhantomData, + } + } + + #[allow(dead_code)] + pub(crate) fn raw(&self) -> *const bindings::drm_device { + self.ptr + } + + pub(crate) fn raw_mut(&mut self) -> *mut bindings::drm_device { + self.ptr + } + + /// Returns a borrowed reference to the user data associated with this Device. + pub fn data(&self) -> ::Borrowed<'_> { + unsafe { T::Data::borrow((*self.ptr).dev_private) } + } +} + +impl Drop for Device { + fn drop(&mut self) { + // SAFETY: By the type invariants, we know that `self` owns a reference, so it is safe to + // relinquish it now. + unsafe { bindings::drm_dev_put(self.ptr) }; + } +} + +impl Clone for Device { + fn clone(&self) -> Self { + // SAFETY: We get a new reference and then create a new owning object from the raw pointer + unsafe { + bindings::drm_dev_get(self.ptr); + Device::from_raw(self.ptr) + } + } +} + +// SAFETY: `Device` only holds a pointer to a C device, which is safe to be used from any thread. +unsafe impl Send for Device {} + +// SAFETY: `Device` only holds a pointer to a C device, references to which are safe to be used +// from any thread. +unsafe impl Sync for Device {} + +// Make drm::Device work for dev_info!() and friends +unsafe impl device::RawDevice for Device { + fn raw_device(&self) -> *mut bindings::device { + // SAFETY: ptr must be valid per the type invariant + unsafe { (*self.ptr).dev } + } +} diff --git a/rust/kernel/drm/drv.rs b/rust/kernel/drm/drv.rs new file mode 100644 index 000000000000..29a465515dc9 --- /dev/null +++ b/rust/kernel/drm/drv.rs @@ -0,0 +1,339 @@ +// SPDX-License-Identifier: GPL-2.0 OR MIT + +//! DRM driver core. +//! +//! C header: [`include/linux/drm/drm_drv.h`](../../../../include/linux/drm/drm_drv.h) + +use crate::{ + bindings, device, drm, + error::code::*, + error::from_kernel_err_ptr, + error::{Error, Result}, + prelude::*, + private::Sealed, + str::CStr, + types::ForeignOwnable, + ThisModule, +}; +use core::{ + marker::{PhantomData, PhantomPinned}, + pin::Pin, +}; +use macros::vtable; + +/// Driver use the GEM memory manager. This should be set for all modern drivers. +pub const FEAT_GEM: u32 = bindings::drm_driver_feature_DRIVER_GEM; +/// Driver supports mode setting interfaces (KMS). +pub const FEAT_MODESET: u32 = bindings::drm_driver_feature_DRIVER_MODESET; +/// Driver supports dedicated render nodes. +pub const FEAT_RENDER: u32 = bindings::drm_driver_feature_DRIVER_RENDER; +/// Driver supports the full atomic modesetting userspace API. +/// +/// Drivers which only use atomic internally, but do not support the full userspace API (e.g. not +/// all properties converted to atomic, or multi-plane updates are not guaranteed to be tear-free) +/// should not set this flag. +pub const FEAT_ATOMIC: u32 = bindings::drm_driver_feature_DRIVER_ATOMIC; +/// Driver supports DRM sync objects for explicit synchronization of command submission. +pub const FEAT_SYNCOBJ: u32 = bindings::drm_driver_feature_DRIVER_SYNCOBJ; +/// Driver supports the timeline flavor of DRM sync objects for explicit synchronization of command +/// submission. +pub const FEAT_SYNCOBJ_TIMELINE: u32 = bindings::drm_driver_feature_DRIVER_SYNCOBJ_TIMELINE; + +/// Information data for a DRM Driver. +pub struct DriverInfo { + /// Driver major version. + pub major: i32, + /// Driver minor version. + pub minor: i32, + /// Driver patchlevel version. + pub patchlevel: i32, + /// Driver name. + pub name: &'static CStr, + /// Driver description. + pub desc: &'static CStr, + /// Driver date. + pub date: &'static CStr, +} + +/// Internal memory management operation set, normally created by memory managers (e.g. GEM). +/// +/// See `kernel::drm::gem` and `kernel::drm::gem::shmem`. +pub struct AllocOps { + pub(crate) gem_create_object: Option< + unsafe extern "C" fn( + dev: *mut bindings::drm_device, + size: usize, + ) -> *mut bindings::drm_gem_object, + >, + pub(crate) prime_handle_to_fd: Option< + unsafe extern "C" fn( + dev: *mut bindings::drm_device, + file_priv: *mut bindings::drm_file, + handle: u32, + flags: u32, + prime_fd: *mut core::ffi::c_int, + ) -> core::ffi::c_int, + >, + pub(crate) prime_fd_to_handle: Option< + unsafe extern "C" fn( + dev: *mut bindings::drm_device, + file_priv: *mut bindings::drm_file, + prime_fd: core::ffi::c_int, + handle: *mut u32, + ) -> core::ffi::c_int, + >, + pub(crate) gem_prime_import: Option< + unsafe extern "C" fn( + dev: *mut bindings::drm_device, + dma_buf: *mut bindings::dma_buf, + ) -> *mut bindings::drm_gem_object, + >, + pub(crate) gem_prime_import_sg_table: Option< + unsafe extern "C" fn( + dev: *mut bindings::drm_device, + attach: *mut bindings::dma_buf_attachment, + sgt: *mut bindings::sg_table, + ) -> *mut bindings::drm_gem_object, + >, + pub(crate) gem_prime_mmap: Option< + unsafe extern "C" fn( + obj: *mut bindings::drm_gem_object, + vma: *mut bindings::vm_area_struct, + ) -> core::ffi::c_int, + >, + pub(crate) dumb_create: Option< + unsafe extern "C" fn( + file_priv: *mut bindings::drm_file, + dev: *mut bindings::drm_device, + args: *mut bindings::drm_mode_create_dumb, + ) -> core::ffi::c_int, + >, + pub(crate) dumb_map_offset: Option< + unsafe extern "C" fn( + file_priv: *mut bindings::drm_file, + dev: *mut bindings::drm_device, + handle: u32, + offset: *mut u64, + ) -> core::ffi::c_int, + >, + pub(crate) dumb_destroy: Option< + unsafe extern "C" fn( + file_priv: *mut bindings::drm_file, + dev: *mut bindings::drm_device, + handle: u32, + ) -> core::ffi::c_int, + >, +} + +/// Trait for memory manager implementations. Implemented internally. +pub trait AllocImpl: Sealed { + /// The C callback operations for this memory manager. + const ALLOC_OPS: AllocOps; +} + +/// A DRM driver implementation. +#[vtable] +pub trait Driver { + /// Context data associated with the DRM driver + /// + /// Determines the type of the context data passed to each of the methods of the trait. + type Data: ForeignOwnable + Sync + Send; + + /// The type used to manage memory for this driver. + /// + /// Should be either `drm::gem::Object` or `drm::gem::shmem::Object`. + type Object: AllocImpl; + + /// Driver metadata + const INFO: DriverInfo; + + /// Feature flags + const FEATURES: u32; + + /// IOCTL list. See `kernel::drm::ioctl::declare_drm_ioctls!{}`. + const IOCTLS: &'static [drm::ioctl::DrmIoctlDescriptor]; +} + +/// A registration of a DRM device +/// +/// # Invariants: +/// +/// drm is always a valid pointer to an allocated drm_device +pub struct Registration { + drm: drm::device::Device, + registered: bool, + fops: bindings::file_operations, + vtable: Pin>, + _p: PhantomData, + _pin: PhantomPinned, +} + +#[cfg(CONFIG_DRM_LEGACY)] +macro_rules! drm_legacy_fields { + ( $($field:ident: $val:expr),* $(,)? ) => { + bindings::drm_driver { + $( $field: $val ),*, + firstopen: None, + preclose: None, + dma_ioctl: None, + dma_quiescent: None, + context_dtor: None, + irq_handler: None, + irq_preinstall: None, + irq_postinstall: None, + irq_uninstall: None, + get_vblank_counter: None, + enable_vblank: None, + disable_vblank: None, + dev_priv_size: 0, + } + } +} + +#[cfg(not(CONFIG_DRM_LEGACY))] +macro_rules! drm_legacy_fields { + ( $($field:ident: $val:expr),* $(,)? ) => { + bindings::drm_driver { + $( $field: $val ),* + } + } +} + +/// Registers a DRM device with the rest of the kernel. +/// +/// It automatically picks up THIS_MODULE. +#[allow(clippy::crate_in_macro_def)] +#[macro_export] +macro_rules! drm_device_register { + ($reg:expr, $data:expr, $flags:expr $(,)?) => {{ + $crate::drm::drv::Registration::register($reg, $data, $flags, &crate::THIS_MODULE) + }}; +} + +impl Registration { + const VTABLE: bindings::drm_driver = drm_legacy_fields! { + load: None, + open: None, // TODO: File abstraction + postclose: None, // TODO: File abstraction + lastclose: None, + unload: None, + release: None, + master_set: None, + master_drop: None, + debugfs_init: None, + gem_create_object: T::Object::ALLOC_OPS.gem_create_object, + prime_handle_to_fd: T::Object::ALLOC_OPS.prime_handle_to_fd, + prime_fd_to_handle: T::Object::ALLOC_OPS.prime_fd_to_handle, + gem_prime_import: T::Object::ALLOC_OPS.gem_prime_import, + gem_prime_import_sg_table: T::Object::ALLOC_OPS.gem_prime_import_sg_table, + gem_prime_mmap: T::Object::ALLOC_OPS.gem_prime_mmap, + dumb_create: T::Object::ALLOC_OPS.dumb_create, + dumb_map_offset: T::Object::ALLOC_OPS.dumb_map_offset, + dumb_destroy: T::Object::ALLOC_OPS.dumb_destroy, + + major: T::INFO.major, + minor: T::INFO.minor, + patchlevel: T::INFO.patchlevel, + name: T::INFO.name.as_char_ptr() as *mut _, + desc: T::INFO.desc.as_char_ptr() as *mut _, + date: T::INFO.date.as_char_ptr() as *mut _, + + driver_features: T::FEATURES, + ioctls: T::IOCTLS.as_ptr(), + num_ioctls: T::IOCTLS.len() as i32, + fops: core::ptr::null_mut(), + }; + + /// Creates a new [`Registration`] but does not register it yet. + /// + /// It is allowed to move. + pub fn new(parent: &dyn device::RawDevice) -> Result { + let vtable = Pin::new(Box::try_new(Self::VTABLE)?); + let raw_drm = unsafe { bindings::drm_dev_alloc(&*vtable, parent.raw_device()) }; + let raw_drm = from_kernel_err_ptr(raw_drm)?; + + // The reference count is one, and now we take ownership of that reference as a + // drm::device::Device. + let drm = unsafe { drm::device::Device::from_raw(raw_drm) }; + + Ok(Self { + drm, + registered: false, + vtable, + fops: Default::default(), // TODO: GEM abstraction + _pin: PhantomPinned, + _p: PhantomData, + }) + } + + /// Registers a DRM device with the rest of the kernel. + /// + /// Users are encouraged to use the [`drm_device_register!()`] macro because it automatically + /// picks up the current module. + pub fn register( + self: Pin<&mut Self>, + data: T::Data, + flags: usize, + module: &'static ThisModule, + ) -> Result { + if self.registered { + // Already registered. + return Err(EINVAL); + } + + // SAFETY: We never move out of `this`. + let this = unsafe { self.get_unchecked_mut() }; + let data_pointer = ::into_foreign(data); + // SAFETY: `drm` is valid per the type invariant + unsafe { + (*this.drm.raw_mut()).dev_private = data_pointer as *mut _; + } + + this.fops.owner = module.0; + this.vtable.fops = &this.fops; + + // SAFETY: The device is now initialized and ready to be registered. + let ret = unsafe { bindings::drm_dev_register(this.drm.raw_mut(), flags as u64) }; + if ret < 0 { + // SAFETY: `data_pointer` was returned by `into_foreign` above. + unsafe { T::Data::from_foreign(data_pointer) }; + return Err(Error::from_kernel_errno(ret)); + } + + this.registered = true; + Ok(()) + } + + /// Returns a reference to the `Device` instance for this registration. + pub fn device(&self) -> &drm::device::Device { + &self.drm + } +} + +// SAFETY: `Registration` doesn't offer any methods or access to fields when shared between threads +// or CPUs, so it is safe to share it. +unsafe impl Sync for Registration {} + +// SAFETY: Registration with and unregistration from the drm subsystem can happen from any thread. +// Additionally, `T::Data` (which is dropped during unregistration) is `Send`, so it is ok to move +// `Registration` to different threads. +#[allow(clippy::non_send_fields_in_send_ty)] +unsafe impl Send for Registration {} + +impl Drop for Registration { + /// Removes the registration from the kernel if it has completed successfully before. + fn drop(&mut self) { + if self.registered { + // Get a pointer to the data stored in device before destroying it. + // SAFETY: `drm` is valid per the type invariant + let data_pointer = unsafe { (*self.drm.raw_mut()).dev_private }; + + // SAFETY: Since `registered` is true, `self.drm` is both valid and registered. + unsafe { bindings::drm_dev_unregister(self.drm.raw_mut()) }; + + // Free data as well. + // SAFETY: `data_pointer` was returned by `into_foreign` during registration. + unsafe { ::from_foreign(data_pointer) }; + } + } +} diff --git a/rust/kernel/drm/mod.rs b/rust/kernel/drm/mod.rs index 9ec6d7cbcaf3..69376b3c6db9 100644 --- a/rust/kernel/drm/mod.rs +++ b/rust/kernel/drm/mod.rs @@ -2,4 +2,6 @@ //! DRM subsystem abstractions. +pub mod device; +pub mod drv; pub mod ioctl; From patchwork Tue Mar 7 14:25:28 2023 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Asahi Lina X-Patchwork-Id: 13163699 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.4.0 (2014-02-07) on aws-us-west-2-korg-lkml-1.web.codeaurora.org Received: from gabe.freedesktop.org (gabe.freedesktop.org [131.252.210.177]) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (No client certificate requested) by smtp.lore.kernel.org (Postfix) with ESMTPS id 68762C678D5 for ; Tue, 7 Mar 2023 14:27:06 +0000 (UTC) Received: from gabe.freedesktop.org (localhost [127.0.0.1]) by gabe.freedesktop.org (Postfix) with ESMTP id AE64810E4AA; Tue, 7 Mar 2023 14:27:05 +0000 (UTC) Received: from mail.marcansoft.com (marcansoft.com [212.63.210.85]) by gabe.freedesktop.org (Postfix) with ESMTPS id 18FF810E4AA for ; Tue, 7 Mar 2023 14:27:04 +0000 (UTC) Received: from [127.0.0.1] (localhost [127.0.0.1]) (using TLSv1.3 with cipher TLS_AES_256_GCM_SHA384 (256/256 bits) key-exchange X25519 server-signature RSA-PSS (4096 bits) server-digest SHA256) (No client certificate requested) (Authenticated sender: linasend@asahilina.net) by mail.marcansoft.com (Postfix) with ESMTPSA id A3475424AC; Tue, 7 Mar 2023 14:26:55 +0000 (UTC) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/simple; d=asahilina.net; s=default; t=1678199222; bh=QBiytwbsc0G1We1dJW1nIZFmB+ztwRRDlj9w4vMV0Ro=; h=From:Date:Subject:References:In-Reply-To:To:Cc; b=xAGB4P/oo/JGUzu7/9TffIgLeqScyHn1BdOEXg7j24rlLZwfpKL/pW783VTJLoZDd UUUwvllNfz+n8CEH7/VkCJfz2Kt3HCxhYc/EJZjVmLNa1vGNSB69UYBsBpQovNlcoh 47pv2L5TfESCngho/MbRa1R/Qyn3RvwWrlZmhNgbqrCevLiIoGZkBStabFgTUDOae6 EORSq8TCMuYNpvigWU0w1OENwm1hxxYY3R9b2vemSUiKCBpX1ZjdRzQRuNu1MU7U9X tjCGS3DMxQykbNKsmAIuPbPSkWE6vsMwcLH0hy4JZPjUONAvu86zjL0knadsl9l23V 0nrfhIIc5PVpA== From: Asahi Lina Date: Tue, 07 Mar 2023 23:25:28 +0900 Subject: [PATCH RFC 03/18] rust: drm: file: Add File abstraction MIME-Version: 1.0 Message-Id: <20230307-rust-drm-v1-3-917ff5bc80a8@asahilina.net> References: <20230307-rust-drm-v1-0-917ff5bc80a8@asahilina.net> In-Reply-To: <20230307-rust-drm-v1-0-917ff5bc80a8@asahilina.net> To: Maarten Lankhorst , Maxime Ripard , Thomas Zimmermann , David Airlie , Daniel Vetter , Miguel Ojeda , Alex Gaynor , Wedson Almeida Filho , Boqun Feng , Gary Guo , =?utf-8?q?Bj=C3=B6rn_Roy_Baron?= , Sumit Semwal , =?utf-8?q?Christian_K=C3=B6nig?= , Luben Tuikov , Jarkko Sakkinen , Dave Hansen X-Mailer: b4 0.12.0 X-Developer-Signature: v=1; a=ed25519-sha256; t=1678199191; l=5871; i=lina@asahilina.net; s=20230221; h=from:subject:message-id; bh=QBiytwbsc0G1We1dJW1nIZFmB+ztwRRDlj9w4vMV0Ro=; b=NkHywoaUixWj5shg1jl7LNEBfnrmuiLisYFWU1LZZmq0/+I3LVSV8lK+ExlBi3b/t/O2ZUFyH rzDEcCifeKOBMCzcXcZoiQK3N2xZUcpC/i3ZFKNDrBE1tR6HrJ806z4 X-Developer-Key: i=lina@asahilina.net; a=ed25519; pk=Qn8jZuOtR1m5GaiDfTrAoQ4NE1XoYVZ/wmt5YtXWFC4= X-BeenThere: dri-devel@lists.freedesktop.org X-Mailman-Version: 2.1.29 Precedence: list List-Id: Direct Rendering Infrastructure - Development List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Cc: linaro-mm-sig@lists.linaro.org, rust-for-linux@vger.kernel.org, Karol Herbst , asahi@lists.linux.dev, Asahi Lina , linux-kernel@vger.kernel.org, dri-devel@lists.freedesktop.org, Mary , Alyssa Rosenzweig , linux-sgx@vger.kernel.org, Ella Stanforth , Faith Ekstrand , linux-media@vger.kernel.org Errors-To: dri-devel-bounces@lists.freedesktop.org Sender: "dri-devel" A DRM File is the DRM counterpart to a kernel file structure, representing an open DRM file descriptor. Add a Rust abstraction to allow drivers to implement their own File types that implement the DriverFile trait. Signed-off-by: Asahi Lina --- rust/bindings/bindings_helper.h | 1 + rust/kernel/drm/drv.rs | 7 ++- rust/kernel/drm/file.rs | 113 ++++++++++++++++++++++++++++++++++++++++ rust/kernel/drm/mod.rs | 1 + 4 files changed, 120 insertions(+), 2 deletions(-) diff --git a/rust/bindings/bindings_helper.h b/rust/bindings/bindings_helper.h index 2a999138c4ae..7d7828faf89c 100644 --- a/rust/bindings/bindings_helper.h +++ b/rust/bindings/bindings_helper.h @@ -8,6 +8,7 @@ #include #include +#include #include #include #include diff --git a/rust/kernel/drm/drv.rs b/rust/kernel/drm/drv.rs index 29a465515dc9..1dcb651e1417 100644 --- a/rust/kernel/drm/drv.rs +++ b/rust/kernel/drm/drv.rs @@ -144,6 +144,9 @@ pub trait Driver { /// Should be either `drm::gem::Object` or `drm::gem::shmem::Object`. type Object: AllocImpl; + /// The type used to represent a DRM File (client) + type File: drm::file::DriverFile; + /// Driver metadata const INFO: DriverInfo; @@ -213,8 +216,8 @@ macro_rules! drm_device_register { impl Registration { const VTABLE: bindings::drm_driver = drm_legacy_fields! { load: None, - open: None, // TODO: File abstraction - postclose: None, // TODO: File abstraction + open: Some(drm::file::open_callback::), + postclose: Some(drm::file::postclose_callback::), lastclose: None, unload: None, release: None, diff --git a/rust/kernel/drm/file.rs b/rust/kernel/drm/file.rs new file mode 100644 index 000000000000..48751e93c38a --- /dev/null +++ b/rust/kernel/drm/file.rs @@ -0,0 +1,113 @@ +// SPDX-License-Identifier: GPL-2.0 OR MIT + +//! DRM File objects. +//! +//! C header: [`include/linux/drm/drm_file.h`](../../../../include/linux/drm/drm_file.h) + +use crate::{bindings, drm, error::Result}; +use alloc::boxed::Box; +use core::marker::PhantomData; +use core::ops::Deref; + +/// Trait that must be implemented by DRM drivers to represent a DRM File (a client instance). +pub trait DriverFile { + /// The parent `Driver` implementation for this `DriverFile`. + type Driver: drm::drv::Driver; + + /// Open a new file (called when a client opens the DRM device). + fn open(device: &drm::device::Device) -> Result>; +} + +/// An open DRM File. +/// +/// # Invariants +/// `raw` is a valid pointer to a `drm_file` struct. +#[repr(transparent)] +pub struct File { + raw: *mut bindings::drm_file, + _p: PhantomData, +} + +pub(super) unsafe extern "C" fn open_callback( + raw_dev: *mut bindings::drm_device, + raw_file: *mut bindings::drm_file, +) -> core::ffi::c_int { + let drm = core::mem::ManuallyDrop::new(unsafe { drm::device::Device::from_raw(raw_dev) }); + // SAFETY: This reference won't escape this function + let file = unsafe { &mut *raw_file }; + + let inner = match T::open(&drm) { + Err(e) => { + return e.to_kernel_errno(); + } + Ok(i) => i, + }; + + file.driver_priv = Box::into_raw(inner) as *mut _; + + 0 +} + +pub(super) unsafe extern "C" fn postclose_callback( + _dev: *mut bindings::drm_device, + raw_file: *mut bindings::drm_file, +) { + // SAFETY: This reference won't escape this function + let file = unsafe { &*raw_file }; + + // Drop the DriverFile + unsafe { Box::from_raw(file.driver_priv as *mut T) }; +} + +impl File { + // Not intended to be called externally, except via declare_drm_ioctls!() + #[doc(hidden)] + pub unsafe fn from_raw(raw_file: *mut bindings::drm_file) -> File { + File { + raw: raw_file, + _p: PhantomData, + } + } + + #[allow(dead_code)] + /// Return the raw pointer to the underlying `drm_file`. + pub(super) fn raw(&self) -> *const bindings::drm_file { + self.raw + } + + /// Return an immutable reference to the raw `drm_file` structure. + pub(super) fn file(&self) -> &bindings::drm_file { + unsafe { &*self.raw } + } +} + +impl Deref for File { + type Target = T; + + fn deref(&self) -> &T { + unsafe { &*(self.file().driver_priv as *const T) } + } +} + +impl crate::private::Sealed for File {} + +/// Generic trait to allow users that don't care about driver specifics to accept any File. +/// +/// # Safety +/// Must only be implemented for File and return the pointer, following the normal invariants +/// of that type. +pub unsafe trait GenericFile: crate::private::Sealed { + /// Returns the raw const pointer to the `struct drm_file` + fn raw(&self) -> *const bindings::drm_file; + /// Returns the raw mut pointer to the `struct drm_file` + fn raw_mut(&mut self) -> *mut bindings::drm_file; +} + +unsafe impl GenericFile for File { + fn raw(&self) -> *const bindings::drm_file { + self.raw + } + fn raw_mut(&mut self) -> *mut bindings::drm_file { + self.raw + } +} diff --git a/rust/kernel/drm/mod.rs b/rust/kernel/drm/mod.rs index 69376b3c6db9..a767942d0b52 100644 --- a/rust/kernel/drm/mod.rs +++ b/rust/kernel/drm/mod.rs @@ -4,4 +4,5 @@ pub mod device; pub mod drv; +pub mod file; pub mod ioctl; From patchwork Tue Mar 7 14:25:29 2023 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Asahi Lina X-Patchwork-Id: 13163700 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.4.0 (2014-02-07) on aws-us-west-2-korg-lkml-1.web.codeaurora.org Received: from gabe.freedesktop.org (gabe.freedesktop.org [131.252.210.177]) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (No client certificate requested) by smtp.lore.kernel.org (Postfix) with ESMTPS id 1689CC678D4 for ; Tue, 7 Mar 2023 14:27:14 +0000 (UTC) Received: from gabe.freedesktop.org (localhost [127.0.0.1]) by gabe.freedesktop.org (Postfix) with ESMTP id 51F1310E4AC; Tue, 7 Mar 2023 14:27:13 +0000 (UTC) Received: from mail.marcansoft.com (marcansoft.com [212.63.210.85]) by gabe.freedesktop.org (Postfix) with ESMTPS id C9BC110E4AC for ; Tue, 7 Mar 2023 14:27:11 +0000 (UTC) Received: from [127.0.0.1] (localhost [127.0.0.1]) (using TLSv1.3 with cipher TLS_AES_256_GCM_SHA384 (256/256 bits) key-exchange X25519 server-signature RSA-PSS (4096 bits) server-digest SHA256) (No client certificate requested) (Authenticated sender: linasend@asahilina.net) by mail.marcansoft.com (Postfix) with ESMTPSA id 1D5B642528; Tue, 7 Mar 2023 14:27:02 +0000 (UTC) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/simple; d=asahilina.net; s=default; t=1678199230; bh=F1SRAlKo7BSIaLQ7gj8v5YlQfBwivdnzeJ8qnwB97A4=; h=From:Date:Subject:References:In-Reply-To:To:Cc; b=jWGXNdok7booYlvp21MSywrjaBbrNPUOU0mT/8sAntEMvNSBysxntLc98cv75xdYX RCg5ekT5IfZ64CjaIwqYAHnRPzV0dg4ueuM5PywDJLOQXLV+mXp9ChItffn2Tw//mj IgoNtG0Ir/bLvMRfauIUURFSixaOirYJxf6Q967INfBaKJ/RJYBAMbCluh6QM/w1Yq HAUu4C9jJb4E1HshphKRyDPOn0poCGThNi7WvyL/qXF6/JXRrcXHLUDLfS+bNLh427 hd2mLg7zRccSoXkAo+kYOoFpHgWtQObypyS1+Rz7x8gmNGF2zunralOQU3+uKMn3x+ ch1YhPHtgI3ow== From: Asahi Lina Date: Tue, 07 Mar 2023 23:25:29 +0900 Subject: [PATCH RFC 04/18] rust: drm: gem: Add GEM object abstraction MIME-Version: 1.0 Message-Id: <20230307-rust-drm-v1-4-917ff5bc80a8@asahilina.net> References: <20230307-rust-drm-v1-0-917ff5bc80a8@asahilina.net> In-Reply-To: <20230307-rust-drm-v1-0-917ff5bc80a8@asahilina.net> To: Maarten Lankhorst , Maxime Ripard , Thomas Zimmermann , David Airlie , Daniel Vetter , Miguel Ojeda , Alex Gaynor , Wedson Almeida Filho , Boqun Feng , Gary Guo , =?utf-8?q?Bj=C3=B6rn_Roy_Baron?= , Sumit Semwal , =?utf-8?q?Christian_K=C3=B6nig?= , Luben Tuikov , Jarkko Sakkinen , Dave Hansen X-Mailer: b4 0.12.0 X-Developer-Signature: v=1; a=ed25519-sha256; t=1678199191; l=16269; i=lina@asahilina.net; s=20230221; h=from:subject:message-id; bh=F1SRAlKo7BSIaLQ7gj8v5YlQfBwivdnzeJ8qnwB97A4=; b=nXvhHnuR0S4HW6xMU182DkD+eGloZ8eCqt3T0yyYFehUpkIdq6wh+Tsm/p0kPWP2XWtbmp6Ix ZKYhDt6DYHVBwLvs6pXLABRbM2g0wJ1QRbmnk+fGSuC5ChaB4x0OthD X-Developer-Key: i=lina@asahilina.net; a=ed25519; pk=Qn8jZuOtR1m5GaiDfTrAoQ4NE1XoYVZ/wmt5YtXWFC4= X-BeenThere: dri-devel@lists.freedesktop.org X-Mailman-Version: 2.1.29 Precedence: list List-Id: Direct Rendering Infrastructure - Development List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Cc: linaro-mm-sig@lists.linaro.org, rust-for-linux@vger.kernel.org, Karol Herbst , asahi@lists.linux.dev, Asahi Lina , linux-kernel@vger.kernel.org, dri-devel@lists.freedesktop.org, Mary , Alyssa Rosenzweig , linux-sgx@vger.kernel.org, Ella Stanforth , Faith Ekstrand , linux-media@vger.kernel.org Errors-To: dri-devel-bounces@lists.freedesktop.org Sender: "dri-devel" The DRM GEM subsystem is the DRM memory management subsystem used by most modern drivers. Add a Rust abstraction to allow Rust DRM driver implementations to use it. Signed-off-by: Asahi Lina --- rust/bindings/bindings_helper.h | 1 + rust/helpers.c | 23 +++ rust/kernel/drm/drv.rs | 4 +- rust/kernel/drm/gem/mod.rs | 374 ++++++++++++++++++++++++++++++++++++++++ rust/kernel/drm/mod.rs | 1 + 5 files changed, 401 insertions(+), 2 deletions(-) diff --git a/rust/bindings/bindings_helper.h b/rust/bindings/bindings_helper.h index 7d7828faf89c..7183dfe6473f 100644 --- a/rust/bindings/bindings_helper.h +++ b/rust/bindings/bindings_helper.h @@ -9,6 +9,7 @@ #include #include #include +#include #include #include #include diff --git a/rust/helpers.c b/rust/helpers.c index 73b2ce607f27..78ec4162b03b 100644 --- a/rust/helpers.c +++ b/rust/helpers.c @@ -18,6 +18,7 @@ * accidentally exposed. */ +#include #include #include #include @@ -374,6 +375,28 @@ void rust_helper_init_completion(struct completion *c) } EXPORT_SYMBOL_GPL(rust_helper_init_completion); +#ifdef CONFIG_DRM + +void rust_helper_drm_gem_object_get(struct drm_gem_object *obj) +{ + drm_gem_object_get(obj); +} +EXPORT_SYMBOL_GPL(rust_helper_drm_gem_object_get); + +void rust_helper_drm_gem_object_put(struct drm_gem_object *obj) +{ + drm_gem_object_put(obj); +} +EXPORT_SYMBOL_GPL(rust_helper_drm_gem_object_put); + +__u64 rust_helper_drm_vma_node_offset_addr(struct drm_vma_offset_node *node) +{ + return drm_vma_node_offset_addr(node); +} +EXPORT_SYMBOL_GPL(rust_helper_drm_vma_node_offset_addr); + +#endif + /* * We use `bindgen`'s `--size_t-is-usize` option to bind the C `size_t` type * as the Rust `usize` type, so we can use it in contexts where Rust diff --git a/rust/kernel/drm/drv.rs b/rust/kernel/drm/drv.rs index 1dcb651e1417..c138352cb489 100644 --- a/rust/kernel/drm/drv.rs +++ b/rust/kernel/drm/drv.rs @@ -126,7 +126,7 @@ pub struct AllocOps { } /// Trait for memory manager implementations. Implemented internally. -pub trait AllocImpl: Sealed { +pub trait AllocImpl: Sealed + drm::gem::IntoGEMObject { /// The C callback operations for this memory manager. const ALLOC_OPS: AllocOps; } @@ -263,7 +263,7 @@ impl Registration { drm, registered: false, vtable, - fops: Default::default(), // TODO: GEM abstraction + fops: drm::gem::create_fops(), _pin: PhantomPinned, _p: PhantomData, }) diff --git a/rust/kernel/drm/gem/mod.rs b/rust/kernel/drm/gem/mod.rs new file mode 100644 index 000000000000..8a7d99613718 --- /dev/null +++ b/rust/kernel/drm/gem/mod.rs @@ -0,0 +1,374 @@ +// SPDX-License-Identifier: GPL-2.0 OR MIT + +//! DRM GEM API +//! +//! C header: [`include/linux/drm/drm_gem.h`](../../../../include/linux/drm/drm_gem.h) + +use alloc::boxed::Box; + +use crate::{ + bindings, + drm::{device, drv, file}, + error::{to_result, Result}, + prelude::*, +}; +use core::{mem, mem::ManuallyDrop, ops::Deref, ops::DerefMut}; + +/// GEM object functions, which must be implemented by drivers. +pub trait BaseDriverObject: Sync + Send + Sized { + /// Create a new driver data object for a GEM object of a given size. + fn new(dev: &device::Device, size: usize) -> Result; + + /// Open a new handle to an existing object, associated with a File. + fn open( + _obj: &<::Driver as drv::Driver>::Object, + _file: &file::File<<::Driver as drv::Driver>::File>, + ) -> Result { + Ok(()) + } + + /// Close a handle to an existing object, associated with a File. + fn close( + _obj: &<::Driver as drv::Driver>::Object, + _file: &file::File<<::Driver as drv::Driver>::File>, + ) { + } +} + +/// Trait that represents a GEM object subtype +pub trait IntoGEMObject: Sized + crate::private::Sealed { + /// Owning driver for this type + type Driver: drv::Driver; + + /// Returns a pointer to the raw `drm_gem_object` structure, which must be valid as long as + /// this owning object is valid. + fn gem_obj(&self) -> *mut bindings::drm_gem_object; + + /// Returns a reference to the raw `drm_gem_object` structure, which must be valid as long as + /// this owning object is valid. + fn gem_ref(&self) -> &bindings::drm_gem_object { + // SAFETY: gem_obj() must be valid per the above requirement. + unsafe { &*self.gem_obj() } + } + + /// Converts a pointer to a `drm_gem_object` into a pointer to this type. + fn from_gem_obj(obj: *mut bindings::drm_gem_object) -> *mut Self; +} + +/// Trait which must be implemented by drivers using base GEM objects. +pub trait DriverObject: BaseDriverObject> { + /// Parent `Driver` for this object. + type Driver: drv::Driver; +} + +unsafe extern "C" fn free_callback(obj: *mut bindings::drm_gem_object) { + // SAFETY: All of our objects are Object. + let this = crate::container_of!(obj, Object, obj) as *mut Object; + + // SAFETY: The pointer we got has to be valid + unsafe { bindings::drm_gem_object_release(obj) }; + + // SAFETY: All of our objects are allocated via Box<>, and we're in the + // free callback which guarantees this object has zero remaining references, + // so we can drop it + unsafe { Box::from_raw(this) }; +} + +unsafe extern "C" fn open_callback, U: BaseObject>( + raw_obj: *mut bindings::drm_gem_object, + raw_file: *mut bindings::drm_file, +) -> core::ffi::c_int { + // SAFETY: The pointer we got has to be valid. + let file = unsafe { + file::File::<<::Driver as drv::Driver>::File>::from_raw(raw_file) + }; + let obj = + <<::Driver as drv::Driver>::Object as IntoGEMObject>::from_gem_obj( + raw_obj, + ); + + // SAFETY: from_gem_obj() returns a valid pointer as long as the type is + // correct and the raw_obj we got is valid. + match T::open(unsafe { &*obj }, &file) { + Err(e) => e.to_kernel_errno(), + Ok(()) => 0, + } +} + +unsafe extern "C" fn close_callback, U: BaseObject>( + raw_obj: *mut bindings::drm_gem_object, + raw_file: *mut bindings::drm_file, +) { + // SAFETY: The pointer we got has to be valid. + let file = unsafe { + file::File::<<::Driver as drv::Driver>::File>::from_raw(raw_file) + }; + let obj = + <<::Driver as drv::Driver>::Object as IntoGEMObject>::from_gem_obj( + raw_obj, + ); + + // SAFETY: from_gem_obj() returns a valid pointer as long as the type is + // correct and the raw_obj we got is valid. + T::close(unsafe { &*obj }, &file); +} + +impl IntoGEMObject for Object { + type Driver = T::Driver; + + fn gem_obj(&self) -> *mut bindings::drm_gem_object { + &self.obj as *const _ as *mut _ + } + + fn from_gem_obj(obj: *mut bindings::drm_gem_object) -> *mut Object { + crate::container_of!(obj, Object, obj) as *mut Object + } +} + +/// Base operations shared by all GEM object classes +pub trait BaseObject: IntoGEMObject { + /// Returns the size of the object in bytes. + fn size(&self) -> usize { + self.gem_ref().size + } + + /// Creates a new reference to the object. + fn reference(&self) -> ObjectRef { + // SAFETY: Having a reference to an Object implies holding a GEM reference + unsafe { + bindings::drm_gem_object_get(self.gem_obj()); + } + ObjectRef { + ptr: self as *const _, + } + } + + /// Creates a new handle for the object associated with a given `File` + /// (or returns an existing one). + fn create_handle( + &self, + file: &file::File<<::Driver as drv::Driver>::File>, + ) -> Result { + let mut handle: u32 = 0; + // SAFETY: The arguments are all valid per the type invariants. + to_result(unsafe { + bindings::drm_gem_handle_create(file.raw() as *mut _, self.gem_obj(), &mut handle) + })?; + Ok(handle) + } + + /// Looks up an object by its handle for a given `File`. + fn lookup_handle( + file: &file::File<<::Driver as drv::Driver>::File>, + handle: u32, + ) -> Result> { + // SAFETY: The arguments are all valid per the type invariants. + let ptr = unsafe { bindings::drm_gem_object_lookup(file.raw() as *mut _, handle) }; + + if ptr.is_null() { + Err(ENOENT) + } else { + Ok(ObjectRef { + ptr: ptr as *const _, + }) + } + } + + /// Creates an mmap offset to map the object from userspace. + fn create_mmap_offset(&self) -> Result { + // SAFETY: The arguments are valid per the type invariant. + to_result(unsafe { + // TODO: is this threadsafe? + bindings::drm_gem_create_mmap_offset(self.gem_obj()) + })?; + Ok(unsafe { + bindings::drm_vma_node_offset_addr(&self.gem_ref().vma_node as *const _ as *mut _) + }) + } +} + +impl BaseObject for T {} + +/// A base GEM object. +#[repr(C)] +pub struct Object { + obj: bindings::drm_gem_object, + // The DRM core ensures the Device exists as long as its objects exist, so we don't need to + // manage the reference count here. + dev: ManuallyDrop>, + inner: T, +} + +impl Object { + /// The size of this object's structure. + pub const SIZE: usize = mem::size_of::(); + + const OBJECT_FUNCS: bindings::drm_gem_object_funcs = bindings::drm_gem_object_funcs { + free: Some(free_callback::), + open: Some(open_callback::>), + close: Some(close_callback::>), + print_info: None, + export: None, + pin: None, + unpin: None, + get_sg_table: None, + vmap: None, + vunmap: None, + mmap: None, + vm_ops: core::ptr::null_mut(), + }; + + /// Create a new GEM object. + pub fn new(dev: &device::Device, size: usize) -> Result> { + let mut obj: Box = Box::try_new(Self { + // SAFETY: This struct is expected to be zero-initialized + obj: unsafe { mem::zeroed() }, + // SAFETY: The drm subsystem guarantees that the drm_device will live as long as + // the GEM object lives, so we can conjure a reference out of thin air. + dev: ManuallyDrop::new(unsafe { device::Device::from_raw(dev.ptr) }), + inner: T::new(dev, size)?, + })?; + + obj.obj.funcs = &Self::OBJECT_FUNCS; + to_result(unsafe { + bindings::drm_gem_object_init(dev.raw() as *mut _, &mut obj.obj, size) + })?; + + let obj_ref = UniqueObjectRef { + ptr: Box::leak(obj), + }; + + Ok(obj_ref) + } + + /// Returns the `Device` that owns this GEM object. + pub fn dev(&self) -> &device::Device { + &self.dev + } +} + +impl crate::private::Sealed for Object {} + +impl Deref for Object { + type Target = T; + + fn deref(&self) -> &Self::Target { + &self.inner + } +} + +impl DerefMut for Object { + fn deref_mut(&mut self) -> &mut Self::Target { + &mut self.inner + } +} + +impl drv::AllocImpl for Object { + const ALLOC_OPS: drv::AllocOps = drv::AllocOps { + gem_create_object: None, + prime_handle_to_fd: Some(bindings::drm_gem_prime_handle_to_fd), + prime_fd_to_handle: Some(bindings::drm_gem_prime_fd_to_handle), + gem_prime_import: None, + gem_prime_import_sg_table: None, + gem_prime_mmap: Some(bindings::drm_gem_prime_mmap), + dumb_create: None, + dumb_map_offset: None, + dumb_destroy: None, + }; +} + +/// A reference-counted shared reference to a base GEM object. +pub struct ObjectRef { + // Invariant: the pointer is valid and initialized, and this ObjectRef owns a reference to it. + ptr: *const T, +} + +/// SAFETY: GEM object references are safe to share between threads. +unsafe impl Send for ObjectRef {} +unsafe impl Sync for ObjectRef {} + +impl Clone for ObjectRef { + fn clone(&self) -> Self { + self.reference() + } +} + +impl Drop for ObjectRef { + fn drop(&mut self) { + // SAFETY: Having an ObjectRef implies holding a GEM reference. + // The free callback will take care of deallocation. + unsafe { + bindings::drm_gem_object_put((*self.ptr).gem_obj()); + } + } +} + +impl Deref for ObjectRef { + type Target = T; + + fn deref(&self) -> &Self::Target { + // SAFETY: The pointer is valid per the invariant + unsafe { &*self.ptr } + } +} + +/// A unique reference to a base GEM object. +pub struct UniqueObjectRef { + // Invariant: the pointer is valid and initialized, and this ObjectRef owns the only reference + // to it. + ptr: *mut T, +} + +impl UniqueObjectRef { + /// Downgrade this reference to a shared reference. + pub fn into_ref(self) -> ObjectRef { + let ptr = self.ptr as *const _; + core::mem::forget(self); + + ObjectRef { ptr } + } +} + +impl Drop for UniqueObjectRef { + fn drop(&mut self) { + // SAFETY: Having a UniqueObjectRef implies holding a GEM + // reference. The free callback will take care of deallocation. + unsafe { + bindings::drm_gem_object_put((*self.ptr).gem_obj()); + } + } +} + +impl Deref for UniqueObjectRef { + type Target = T; + + fn deref(&self) -> &Self::Target { + // SAFETY: The pointer is valid per the invariant + unsafe { &*self.ptr } + } +} + +impl DerefMut for UniqueObjectRef { + fn deref_mut(&mut self) -> &mut Self::Target { + // SAFETY: The pointer is valid per the invariant + unsafe { &mut *self.ptr } + } +} + +pub(super) fn create_fops() -> bindings::file_operations { + bindings::file_operations { + owner: core::ptr::null_mut(), + open: Some(bindings::drm_open), + release: Some(bindings::drm_release), + unlocked_ioctl: Some(bindings::drm_ioctl), + #[cfg(CONFIG_COMPAT)] + compat_ioctl: Some(bindings::drm_compat_ioctl), + #[cfg(not(CONFIG_COMPAT))] + compat_ioctl: None, + poll: Some(bindings::drm_poll), + read: Some(bindings::drm_read), + llseek: Some(bindings::noop_llseek), + mmap: Some(bindings::drm_gem_mmap), + ..Default::default() + } +} diff --git a/rust/kernel/drm/mod.rs b/rust/kernel/drm/mod.rs index a767942d0b52..c44760a1332f 100644 --- a/rust/kernel/drm/mod.rs +++ b/rust/kernel/drm/mod.rs @@ -5,4 +5,5 @@ pub mod device; pub mod drv; pub mod file; +pub mod gem; pub mod ioctl; From patchwork Tue Mar 7 14:25:30 2023 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Asahi Lina X-Patchwork-Id: 13163701 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.4.0 (2014-02-07) on aws-us-west-2-korg-lkml-1.web.codeaurora.org Received: from gabe.freedesktop.org (gabe.freedesktop.org [131.252.210.177]) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (No client certificate requested) by smtp.lore.kernel.org (Postfix) with ESMTPS id F3E23C678D4 for ; Tue, 7 Mar 2023 14:27:21 +0000 (UTC) Received: from gabe.freedesktop.org (localhost [127.0.0.1]) by gabe.freedesktop.org (Postfix) with ESMTP id 3233C10E4AE; Tue, 7 Mar 2023 14:27:21 +0000 (UTC) Received: from mail.marcansoft.com (marcansoft.com [212.63.210.85]) by gabe.freedesktop.org (Postfix) with ESMTPS id 2FC2210E4AE for ; Tue, 7 Mar 2023 14:27:19 +0000 (UTC) Received: from [127.0.0.1] (localhost [127.0.0.1]) (using TLSv1.3 with cipher TLS_AES_256_GCM_SHA384 (256/256 bits) key-exchange X25519 server-signature RSA-PSS (4096 bits) server-digest SHA256) (No client certificate requested) (Authenticated sender: linasend@asahilina.net) by mail.marcansoft.com (Postfix) with ESMTPSA id C587C424CD; Tue, 7 Mar 2023 14:27:10 +0000 (UTC) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/simple; d=asahilina.net; s=default; t=1678199237; bh=BAqN8OMTaXAhucXUriExOzQpMpPUhvwTh5OsDzrSkpo=; h=From:Date:Subject:References:In-Reply-To:To:Cc; b=sdJUn/+TKQ7o8T8HGDCq3r2jd7AAtYK04d56jZmlkVNBzcSniG/yHCP6fDIbG9ehZ kSFIfPufTMbHtE461pYVmVq6FGQDfJAAWzZ+90gdyChquINRMvLM7Mg7xqF3pK2ikK qiAnhXlPguV04IZom9jBVBQYf/B2wNZqnpf4IeV4QUImr7ycS2rQu7K+yITkcDLmhI VqIH6ZmHfnCKkRScdcKV56rpQVzl1pvVoqyVKpWimedHzo3r7Xioq4dddbVh9q9+4x X/8VYqv9CrMVzNN3zky2OB830unrsw9d1jxOn5VWOFJIiSIKvp0HQl9KEC13yFGfV1 vfcseqiRi4TEw== From: Asahi Lina Date: Tue, 07 Mar 2023 23:25:30 +0900 Subject: [PATCH RFC 05/18] drm/gem-shmem: Export VM ops functions MIME-Version: 1.0 Message-Id: <20230307-rust-drm-v1-5-917ff5bc80a8@asahilina.net> References: <20230307-rust-drm-v1-0-917ff5bc80a8@asahilina.net> In-Reply-To: <20230307-rust-drm-v1-0-917ff5bc80a8@asahilina.net> To: Maarten Lankhorst , Maxime Ripard , Thomas Zimmermann , David Airlie , Daniel Vetter , Miguel Ojeda , Alex Gaynor , Wedson Almeida Filho , Boqun Feng , Gary Guo , =?utf-8?q?Bj=C3=B6rn_Roy_Baron?= , Sumit Semwal , =?utf-8?q?Christian_K=C3=B6nig?= , Luben Tuikov , Jarkko Sakkinen , Dave Hansen X-Mailer: b4 0.12.0 X-Developer-Signature: v=1; a=ed25519-sha256; t=1678199191; l=2745; i=lina@asahilina.net; s=20230221; h=from:subject:message-id; bh=BAqN8OMTaXAhucXUriExOzQpMpPUhvwTh5OsDzrSkpo=; b=g9pqskVqJS3ljWoohfpCNH4NDQdA/LBauCNqtM+x0fz44GOfEY/N8GW4sqZzjgPhnydgE4j/x Si9bKq4ETQwDZpvyu6TbEWKU1rzJi6sGVUQQEwWgtVhiU5HoHMICe3I X-Developer-Key: i=lina@asahilina.net; a=ed25519; pk=Qn8jZuOtR1m5GaiDfTrAoQ4NE1XoYVZ/wmt5YtXWFC4= X-BeenThere: dri-devel@lists.freedesktop.org X-Mailman-Version: 2.1.29 Precedence: list List-Id: Direct Rendering Infrastructure - Development List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Cc: linaro-mm-sig@lists.linaro.org, rust-for-linux@vger.kernel.org, Karol Herbst , asahi@lists.linux.dev, Asahi Lina , linux-kernel@vger.kernel.org, dri-devel@lists.freedesktop.org, Mary , Alyssa Rosenzweig , linux-sgx@vger.kernel.org, Ella Stanforth , Faith Ekstrand , linux-media@vger.kernel.org Errors-To: dri-devel-bounces@lists.freedesktop.org Sender: "dri-devel" There doesn't seem to be a way for the Rust bindings to get a compile-time constant reference to drm_gem_shmem_vm_ops, so we need to duplicate that structure in Rust... this isn't nice... Signed-off-by: Asahi Lina --- drivers/gpu/drm/drm_gem_shmem_helper.c | 9 ++++++--- include/drm/drm_gem_shmem_helper.h | 3 +++ 2 files changed, 9 insertions(+), 3 deletions(-) diff --git a/drivers/gpu/drm/drm_gem_shmem_helper.c b/drivers/gpu/drm/drm_gem_shmem_helper.c index 75185a960fc4..10c09819410e 100644 --- a/drivers/gpu/drm/drm_gem_shmem_helper.c +++ b/drivers/gpu/drm/drm_gem_shmem_helper.c @@ -534,7 +534,7 @@ int drm_gem_shmem_dumb_create(struct drm_file *file, struct drm_device *dev, } EXPORT_SYMBOL_GPL(drm_gem_shmem_dumb_create); -static vm_fault_t drm_gem_shmem_fault(struct vm_fault *vmf) +vm_fault_t drm_gem_shmem_fault(struct vm_fault *vmf) { struct vm_area_struct *vma = vmf->vma; struct drm_gem_object *obj = vma->vm_private_data; @@ -563,8 +563,9 @@ static vm_fault_t drm_gem_shmem_fault(struct vm_fault *vmf) return ret; } +EXPORT_SYMBOL_GPL(drm_gem_shmem_fault); -static void drm_gem_shmem_vm_open(struct vm_area_struct *vma) +void drm_gem_shmem_vm_open(struct vm_area_struct *vma) { struct drm_gem_object *obj = vma->vm_private_data; struct drm_gem_shmem_object *shmem = to_drm_gem_shmem_obj(obj); @@ -585,8 +586,9 @@ static void drm_gem_shmem_vm_open(struct vm_area_struct *vma) drm_gem_vm_open(vma); } +EXPORT_SYMBOL_GPL(drm_gem_shmem_vm_open); -static void drm_gem_shmem_vm_close(struct vm_area_struct *vma) +void drm_gem_shmem_vm_close(struct vm_area_struct *vma) { struct drm_gem_object *obj = vma->vm_private_data; struct drm_gem_shmem_object *shmem = to_drm_gem_shmem_obj(obj); @@ -594,6 +596,7 @@ static void drm_gem_shmem_vm_close(struct vm_area_struct *vma) drm_gem_shmem_put_pages(shmem); drm_gem_vm_close(vma); } +EXPORT_SYMBOL_GPL(drm_gem_shmem_vm_close); const struct vm_operations_struct drm_gem_shmem_vm_ops = { .fault = drm_gem_shmem_fault, diff --git a/include/drm/drm_gem_shmem_helper.h b/include/drm/drm_gem_shmem_helper.h index a2201b2488c5..b9f349b3ed76 100644 --- a/include/drm/drm_gem_shmem_helper.h +++ b/include/drm/drm_gem_shmem_helper.h @@ -138,6 +138,9 @@ void drm_gem_shmem_print_info(const struct drm_gem_shmem_object *shmem, struct drm_printer *p, unsigned int indent); extern const struct vm_operations_struct drm_gem_shmem_vm_ops; +vm_fault_t drm_gem_shmem_fault(struct vm_fault *vmf); +void drm_gem_shmem_vm_open(struct vm_area_struct *vma); +void drm_gem_shmem_vm_close(struct vm_area_struct *vma); /* * GEM object functions From patchwork Tue Mar 7 14:25:31 2023 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 8bit X-Patchwork-Submitter: Asahi Lina X-Patchwork-Id: 13163702 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.4.0 (2014-02-07) on aws-us-west-2-korg-lkml-1.web.codeaurora.org Received: from gabe.freedesktop.org (gabe.freedesktop.org [131.252.210.177]) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (No client certificate requested) by smtp.lore.kernel.org (Postfix) with ESMTPS id F3705C678D5 for ; Tue, 7 Mar 2023 14:27:28 +0000 (UTC) Received: from gabe.freedesktop.org (localhost [127.0.0.1]) by gabe.freedesktop.org (Postfix) with ESMTP id 2A14710E4AB; Tue, 7 Mar 2023 14:27:28 +0000 (UTC) Received: from mail.marcansoft.com (marcansoft.com [212.63.210.85]) by gabe.freedesktop.org (Postfix) with ESMTPS id D5FF610E4AB for ; Tue, 7 Mar 2023 14:27:26 +0000 (UTC) Received: from [127.0.0.1] (localhost [127.0.0.1]) (using TLSv1.3 with cipher TLS_AES_256_GCM_SHA384 (256/256 bits) key-exchange X25519 server-signature RSA-PSS (4096 bits) server-digest SHA256) (No client certificate requested) (Authenticated sender: linasend@asahilina.net) by mail.marcansoft.com (Postfix) with ESMTPSA id 40213424E8; Tue, 7 Mar 2023 14:27:18 +0000 (UTC) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/simple; d=asahilina.net; s=default; t=1678199245; bh=Myw41AQHjD5rzjM1j2FKHqXtGruOkIC1O9lJa78/vVw=; h=From:Date:Subject:References:In-Reply-To:To:Cc; b=m7rJ9bubpU2eWduCvFK5/ydxBV740I3KYlHlxZSn7BX2f/GmL3Hchi/ZRvs3b51Im ahhGj9ECY0CINGig8I7i8iFaPycNS+M4narw/lGH7Dwg4xDH3ndYU+2EIPNz6FJfXM hoCntsHB68dZSgOtMB4G1HYLkypMNEvJCWVMROd7Db9CmH1+kUBNAGwdmQVirB+CTU O3MWJv4BeL/ioBsxWvVAQxp05cAWDGfZJyQ2GyOBv8mn4fXCv0/taSZbd61CWNUwL3 AxRIvjB2j7NjYWFL7GH3XCUsf4l/nsfza/f6oXJR5cTLJhHwQknv52NIS8FJ1BWmNs tlH2eg5ICU+Rg== From: Asahi Lina Date: Tue, 07 Mar 2023 23:25:31 +0900 Subject: [PATCH RFC 06/18] rust: drm: gem: shmem: Add DRM shmem helper abstraction MIME-Version: 1.0 Message-Id: <20230307-rust-drm-v1-6-917ff5bc80a8@asahilina.net> References: <20230307-rust-drm-v1-0-917ff5bc80a8@asahilina.net> In-Reply-To: <20230307-rust-drm-v1-0-917ff5bc80a8@asahilina.net> To: Maarten Lankhorst , Maxime Ripard , Thomas Zimmermann , David Airlie , Daniel Vetter , Miguel Ojeda , Alex Gaynor , Wedson Almeida Filho , Boqun Feng , Gary Guo , =?utf-8?q?Bj=C3=B6rn_Roy_Baron?= , Sumit Semwal , =?utf-8?q?Christian_K=C3=B6nig?= , Luben Tuikov , Jarkko Sakkinen , Dave Hansen X-Mailer: b4 0.12.0 X-Developer-Signature: v=1; a=ed25519-sha256; t=1678199191; l=17977; i=lina@asahilina.net; s=20230221; h=from:subject:message-id; bh=Myw41AQHjD5rzjM1j2FKHqXtGruOkIC1O9lJa78/vVw=; b=2bzmg2GBDxzMtheU4BnU4u1o1j34J+4xXpa+rnINpwVPG76o6w5BnYjHYSkPaD11xDz8jxumt Pd/z5Mqop8tBilCmR38aV6+KhByVJWsiP0GiYgvMZgiB9wwgcWYKRLZ X-Developer-Key: i=lina@asahilina.net; a=ed25519; pk=Qn8jZuOtR1m5GaiDfTrAoQ4NE1XoYVZ/wmt5YtXWFC4= X-BeenThere: dri-devel@lists.freedesktop.org X-Mailman-Version: 2.1.29 Precedence: list List-Id: Direct Rendering Infrastructure - Development List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Cc: linaro-mm-sig@lists.linaro.org, rust-for-linux@vger.kernel.org, Karol Herbst , asahi@lists.linux.dev, Asahi Lina , linux-kernel@vger.kernel.org, dri-devel@lists.freedesktop.org, Mary , Alyssa Rosenzweig , linux-sgx@vger.kernel.org, Ella Stanforth , Faith Ekstrand , linux-media@vger.kernel.org Errors-To: dri-devel-bounces@lists.freedesktop.org Sender: "dri-devel" The DRM shmem helper includes common code useful for drivers which allocate GEM objects as anonymous shmem. Add a Rust abstraction for this. Drivers can choose the raw GEM implementation or the shmem layer, depending on their needs. Signed-off-by: Asahi Lina Signed-off-by: Maíra Canal --- drivers/gpu/drm/Kconfig | 5 + rust/bindings/bindings_helper.h | 2 + rust/helpers.c | 67 +++++++ rust/kernel/drm/gem/mod.rs | 3 + rust/kernel/drm/gem/shmem.rs | 381 ++++++++++++++++++++++++++++++++++++++++ 5 files changed, 458 insertions(+) diff --git a/drivers/gpu/drm/Kconfig b/drivers/gpu/drm/Kconfig index dab8f0f9aa96..70a983a17ac2 100644 --- a/drivers/gpu/drm/Kconfig +++ b/drivers/gpu/drm/Kconfig @@ -34,6 +34,11 @@ config RUST_DRM bool "Rust support for the DRM subsystem" depends on DRM=y +config RUST_DRM_GEM_SHMEM_HELPER + bool + depends on RUST_DRM + select DRM_GEM_SHMEM_HELPER + config DRM_MIPI_DBI tristate depends on DRM diff --git a/rust/bindings/bindings_helper.h b/rust/bindings/bindings_helper.h index 7183dfe6473f..9f152d373df8 100644 --- a/rust/bindings/bindings_helper.h +++ b/rust/bindings/bindings_helper.h @@ -10,6 +10,7 @@ #include #include #include +#include #include #include #include @@ -17,6 +18,7 @@ #include #include #include +#include #include #include #include diff --git a/rust/helpers.c b/rust/helpers.c index 78ec4162b03b..388ff1100ea5 100644 --- a/rust/helpers.c +++ b/rust/helpers.c @@ -19,6 +19,7 @@ */ #include +#include #include #include #include @@ -375,6 +376,18 @@ void rust_helper_init_completion(struct completion *c) } EXPORT_SYMBOL_GPL(rust_helper_init_completion); +dma_addr_t rust_helper_sg_dma_address(const struct scatterlist *sg) +{ + return sg_dma_address(sg); +} +EXPORT_SYMBOL_GPL(rust_helper_sg_dma_address); + +int rust_helper_sg_dma_len(const struct scatterlist *sg) +{ + return sg_dma_len(sg); +} +EXPORT_SYMBOL_GPL(rust_helper_sg_dma_len); + #ifdef CONFIG_DRM void rust_helper_drm_gem_object_get(struct drm_gem_object *obj) @@ -395,6 +408,60 @@ __u64 rust_helper_drm_vma_node_offset_addr(struct drm_vma_offset_node *node) } EXPORT_SYMBOL_GPL(rust_helper_drm_vma_node_offset_addr); +#ifdef CONFIG_DRM_GEM_SHMEM_HELPER + +void rust_helper_drm_gem_shmem_object_free(struct drm_gem_object *obj) +{ + return drm_gem_shmem_object_free(obj); +} +EXPORT_SYMBOL_GPL(rust_helper_drm_gem_shmem_object_free); + +void rust_helper_drm_gem_shmem_object_print_info(struct drm_printer *p, unsigned int indent, + const struct drm_gem_object *obj) +{ + drm_gem_shmem_object_print_info(p, indent, obj); +} +EXPORT_SYMBOL_GPL(rust_helper_drm_gem_shmem_object_print_info); + +int rust_helper_drm_gem_shmem_object_pin(struct drm_gem_object *obj) +{ + return drm_gem_shmem_object_pin(obj); +} +EXPORT_SYMBOL_GPL(rust_helper_drm_gem_shmem_object_pin); + +void rust_helper_drm_gem_shmem_object_unpin(struct drm_gem_object *obj) +{ + drm_gem_shmem_object_unpin(obj); +} +EXPORT_SYMBOL_GPL(rust_helper_drm_gem_shmem_object_unpin); + +struct sg_table *rust_helper_drm_gem_shmem_object_get_sg_table(struct drm_gem_object *obj) +{ + return drm_gem_shmem_object_get_sg_table(obj); +} +EXPORT_SYMBOL_GPL(rust_helper_drm_gem_shmem_object_get_sg_table); + +int rust_helper_drm_gem_shmem_object_vmap(struct drm_gem_object *obj, + struct iosys_map *map) +{ + return drm_gem_shmem_object_vmap(obj, map); +} +EXPORT_SYMBOL_GPL(rust_helper_drm_gem_shmem_object_vmap); + +void rust_helper_drm_gem_shmem_object_vunmap(struct drm_gem_object *obj, + struct iosys_map *map) +{ + drm_gem_shmem_object_vunmap(obj, map); +} +EXPORT_SYMBOL_GPL(rust_helper_drm_gem_shmem_object_vunmap); + +int rust_helper_drm_gem_shmem_object_mmap(struct drm_gem_object *obj, struct vm_area_struct *vma) +{ + return drm_gem_shmem_object_mmap(obj, vma); +} +EXPORT_SYMBOL_GPL(rust_helper_drm_gem_shmem_object_mmap); + +#endif #endif /* diff --git a/rust/kernel/drm/gem/mod.rs b/rust/kernel/drm/gem/mod.rs index 8a7d99613718..e66bdef35c2e 100644 --- a/rust/kernel/drm/gem/mod.rs +++ b/rust/kernel/drm/gem/mod.rs @@ -4,6 +4,9 @@ //! //! C header: [`include/linux/drm/drm_gem.h`](../../../../include/linux/drm/drm_gem.h) +#[cfg(CONFIG_RUST_DRM_GEM_SHMEM_HELPER)] +pub mod shmem; + use alloc::boxed::Box; use crate::{ diff --git a/rust/kernel/drm/gem/shmem.rs b/rust/kernel/drm/gem/shmem.rs new file mode 100644 index 000000000000..15446ea1113e --- /dev/null +++ b/rust/kernel/drm/gem/shmem.rs @@ -0,0 +1,381 @@ +// SPDX-License-Identifier: GPL-2.0 + +//! DRM GEM shmem helper objects +//! +//! C header: [`include/linux/drm/drm_gem_shmem_helper.h`](../../../../include/linux/drm/drm_gem_shmem_helper.h) + +use crate::drm::{device, drv, gem}; +use crate::{ + error::{from_kernel_err_ptr, to_result}, + prelude::*, +}; +use core::{ + marker::PhantomData, + mem, + mem::{ManuallyDrop, MaybeUninit}, + ops::{Deref, DerefMut}, + ptr::addr_of_mut, + slice, +}; + +use gem::BaseObject; + +/// Trait which must be implemented by drivers using shmem-backed GEM objects. +pub trait DriverObject: gem::BaseDriverObject> { + /// Parent `Driver` for this object. + type Driver: drv::Driver; +} + +// FIXME: This is terrible and I don't know how to avoid it +#[cfg(CONFIG_NUMA)] +macro_rules! vm_numa_fields { + ( $($field:ident: $val:expr),* $(,)? ) => { + bindings::vm_operations_struct { + $( $field: $val ),*, + set_policy: None, + get_policy: None, + } + } +} + +#[cfg(not(CONFIG_NUMA))] +macro_rules! vm_numa_fields { + ( $($field:ident: $val:expr),* $(,)? ) => { + bindings::vm_operations_struct { + $( $field: $val ),* + } + } +} + +const SHMEM_VM_OPS: bindings::vm_operations_struct = vm_numa_fields! { + open: Some(bindings::drm_gem_shmem_vm_open), + close: Some(bindings::drm_gem_shmem_vm_close), + may_split: None, + mremap: None, + mprotect: None, + fault: Some(bindings::drm_gem_shmem_fault), + huge_fault: None, + map_pages: None, + pagesize: None, + page_mkwrite: None, + pfn_mkwrite: None, + access: None, + name: None, + find_special_page: None, +}; + +/// A shmem-backed GEM object. +#[repr(C)] +pub struct Object { + obj: bindings::drm_gem_shmem_object, + // The DRM core ensures the Device exists as long as its objects exist, so we don't need to + // manage the reference count here. + dev: ManuallyDrop>, + inner: T, +} + +unsafe extern "C" fn gem_create_object( + raw_dev: *mut bindings::drm_device, + size: usize, +) -> *mut bindings::drm_gem_object { + // SAFETY: GEM ensures the device lives as long as its objects live, + // so we can conjure up a reference from thin air and never drop it. + let dev = ManuallyDrop::new(unsafe { device::Device::from_raw(raw_dev) }); + + let inner = match T::new(&*dev, size) { + Ok(v) => v, + Err(e) => return e.to_ptr(), + }; + + let p = unsafe { + bindings::krealloc( + core::ptr::null(), + Object::::SIZE, + bindings::GFP_KERNEL | bindings::__GFP_ZERO, + ) as *mut Object + }; + + if p.is_null() { + return ENOMEM.to_ptr(); + } + + // SAFETY: p is valid as long as the alloc succeeded + unsafe { + addr_of_mut!((*p).dev).write(dev); + addr_of_mut!((*p).inner).write(inner); + } + + // SAFETY: drm_gem_shmem_object is safe to zero-init, and + // the rest of Object has been initialized + let new: &mut Object = unsafe { &mut *(p as *mut _) }; + + new.obj.base.funcs = &Object::::VTABLE; + &mut new.obj.base +} + +unsafe extern "C" fn free_callback(obj: *mut bindings::drm_gem_object) { + // SAFETY: All of our objects are Object. + let p = crate::container_of!(obj, Object, obj) as *mut Object; + + // SAFETY: p is never used after this + unsafe { + core::ptr::drop_in_place(&mut (*p).inner); + } + + // SAFETY: This pointer has to be valid, since p is valid + unsafe { + bindings::drm_gem_shmem_free(&mut (*p).obj); + } +} + +impl Object { + /// The size of this object's structure. + const SIZE: usize = mem::size_of::(); + + /// `drm_gem_object_funcs` vtable suitable for GEM shmem objects. + const VTABLE: bindings::drm_gem_object_funcs = bindings::drm_gem_object_funcs { + free: Some(free_callback::), + open: Some(super::open_callback::>), + close: Some(super::close_callback::>), + print_info: Some(bindings::drm_gem_shmem_object_print_info), + export: None, + pin: Some(bindings::drm_gem_shmem_object_pin), + unpin: Some(bindings::drm_gem_shmem_object_unpin), + get_sg_table: Some(bindings::drm_gem_shmem_object_get_sg_table), + vmap: Some(bindings::drm_gem_shmem_object_vmap), + vunmap: Some(bindings::drm_gem_shmem_object_vunmap), + mmap: Some(bindings::drm_gem_shmem_object_mmap), + vm_ops: &SHMEM_VM_OPS, + }; + + // SAFETY: Must only be used with DRM functions that are thread-safe + unsafe fn mut_shmem(&self) -> *mut bindings::drm_gem_shmem_object { + &self.obj as *const _ as *mut _ + } + + /// Create a new shmem-backed DRM object of the given size. + pub fn new(dev: &device::Device, size: usize) -> Result> { + // SAFETY: This function can be called as long as the ALLOC_OPS are set properly + // for this driver, and the gem_create_object is called. + let p = unsafe { bindings::drm_gem_shmem_create(dev.raw() as *mut _, size) }; + let p = crate::container_of!(p, Object, obj) as *mut _; + + // SAFETY: The gem_create_object callback ensures this is a valid Object, + // so we can take a unique reference to it. + let obj_ref = gem::UniqueObjectRef { ptr: p }; + + Ok(obj_ref) + } + + /// Returns the `Device` that owns this GEM object. + pub fn dev(&self) -> &device::Device { + &self.dev + } + + /// Creates (if necessary) and returns a scatter-gather table of DMA pages for this object. + /// + /// This will pin the object in memory. + pub fn sg_table(&self) -> Result> { + // SAFETY: drm_gem_shmem_get_pages_sgt is thread-safe. + let sgt = from_kernel_err_ptr(unsafe { + bindings::drm_gem_shmem_get_pages_sgt(self.mut_shmem()) + })?; + + Ok(SGTable { + sgt, + _owner: self.reference(), + }) + } + + /// Creates and returns a virtual kernel memory mapping for this object. + pub fn vmap(&self) -> Result> { + let mut map: MaybeUninit = MaybeUninit::uninit(); + + // SAFETY: drm_gem_shmem_vmap is thread-safe + to_result(unsafe { bindings::drm_gem_shmem_vmap(self.mut_shmem(), map.as_mut_ptr()) })?; + + // SAFETY: if drm_gem_shmem_vmap did not fail, map is initialized now + let map = unsafe { map.assume_init() }; + + Ok(VMap { + map, + owner: self.reference(), + }) + } + + /// Set the write-combine flag for this object. + /// + /// Should be called before any mappings are made. + pub fn set_wc(&mut self, map_wc: bool) { + unsafe { (*self.mut_shmem()).map_wc = map_wc }; + } +} + +impl Deref for Object { + type Target = T; + + fn deref(&self) -> &Self::Target { + &self.inner + } +} + +impl DerefMut for Object { + fn deref_mut(&mut self) -> &mut Self::Target { + &mut self.inner + } +} + +impl crate::private::Sealed for Object {} + +impl gem::IntoGEMObject for Object { + type Driver = T::Driver; + + fn gem_obj(&self) -> *mut bindings::drm_gem_object { + &self.obj.base as *const _ as *mut _ + } + + fn from_gem_obj(obj: *mut bindings::drm_gem_object) -> *mut Object { + crate::container_of!(obj, Object, obj) as *mut Object + } +} + +impl drv::AllocImpl for Object { + const ALLOC_OPS: drv::AllocOps = drv::AllocOps { + gem_create_object: Some(gem_create_object::), + prime_handle_to_fd: Some(bindings::drm_gem_prime_handle_to_fd), + prime_fd_to_handle: Some(bindings::drm_gem_prime_fd_to_handle), + gem_prime_import: None, + gem_prime_import_sg_table: Some(bindings::drm_gem_shmem_prime_import_sg_table), + gem_prime_mmap: Some(bindings::drm_gem_prime_mmap), + dumb_create: Some(bindings::drm_gem_shmem_dumb_create), + dumb_map_offset: None, + dumb_destroy: None, + }; +} + +/// A virtual mapping for a shmem-backed GEM object in kernel address space. +pub struct VMap { + map: bindings::iosys_map, + owner: gem::ObjectRef>, +} + +impl VMap { + /// Returns a const raw pointer to the start of the mapping. + pub fn as_ptr(&self) -> *const core::ffi::c_void { + // SAFETY: The shmem helpers always return non-iomem maps + unsafe { self.map.__bindgen_anon_1.vaddr } + } + + /// Returns a mutable raw pointer to the start of the mapping. + pub fn as_mut_ptr(&mut self) -> *mut core::ffi::c_void { + // SAFETY: The shmem helpers always return non-iomem maps + unsafe { self.map.__bindgen_anon_1.vaddr } + } + + /// Returns a byte slice view of the mapping. + pub fn as_slice(&self) -> &[u8] { + // SAFETY: The vmap maps valid memory up to the owner size + unsafe { slice::from_raw_parts(self.as_ptr() as *const u8, self.owner.size()) } + } + + /// Returns mutable a byte slice view of the mapping. + pub fn as_mut_slice(&mut self) -> &mut [u8] { + // SAFETY: The vmap maps valid memory up to the owner size + unsafe { slice::from_raw_parts_mut(self.as_mut_ptr() as *mut u8, self.owner.size()) } + } + + /// Borrows a reference to the object that owns this virtual mapping. + pub fn owner(&self) -> &gem::ObjectRef> { + &self.owner + } +} + +impl Drop for VMap { + fn drop(&mut self) { + // SAFETY: This function is thread-safe + unsafe { + bindings::drm_gem_shmem_vunmap(self.owner.mut_shmem(), &mut self.map); + } + } +} + +/// SAFETY: `iosys_map` objects are safe to send across threads. +unsafe impl Send for VMap {} +unsafe impl Sync for VMap {} + +/// A single scatter-gather entry, representing a span of pages in the device's DMA address space. +/// +/// For devices not behind a standalone IOMMU, this corresponds to physical addresses. +#[repr(transparent)] +pub struct SGEntry(bindings::scatterlist); + +impl SGEntry { + /// Returns the starting DMA address of this span + pub fn dma_address(&self) -> usize { + (unsafe { bindings::sg_dma_address(&self.0) }) as usize + } + + /// Returns the length of this span in bytes + pub fn dma_len(&self) -> usize { + (unsafe { bindings::sg_dma_len(&self.0) }) as usize + } +} + +/// A scatter-gather table of DMA address spans for a GEM shmem object. +/// +/// # Invariants +/// `sgt` must be a valid pointer to the `sg_table`, which must correspond to the owned +/// object in `_owner` (which ensures it remains valid). +pub struct SGTable { + sgt: *const bindings::sg_table, + _owner: gem::ObjectRef>, +} + +impl SGTable { + /// Returns an iterator through the SGTable's entries + pub fn iter(&'_ self) -> SGTableIter<'_> { + SGTableIter { + left: unsafe { (*self.sgt).nents } as usize, + sg: unsafe { (*self.sgt).sgl }, + _p: PhantomData, + } + } +} + +impl<'a, T: DriverObject> IntoIterator for &'a SGTable { + type Item = &'a SGEntry; + type IntoIter = SGTableIter<'a>; + + fn into_iter(self) -> Self::IntoIter { + self.iter() + } +} + +/// SAFETY: `sg_table` objects are safe to send across threads. +unsafe impl Send for SGTable {} +unsafe impl Sync for SGTable {} + +/// An iterator through `SGTable` entries. +/// +/// # Invariants +/// `sg` must be a valid pointer to the scatterlist, which must outlive our lifetime. +pub struct SGTableIter<'a> { + sg: *mut bindings::scatterlist, + left: usize, + _p: PhantomData<&'a ()>, +} + +impl<'a> Iterator for SGTableIter<'a> { + type Item = &'a SGEntry; + + fn next(&mut self) -> Option { + if self.left == 0 { + None + } else { + let sg = self.sg; + self.sg = unsafe { bindings::sg_next(self.sg) }; + self.left -= 1; + Some(unsafe { &(*(sg as *const SGEntry)) }) + } + } +} From patchwork Tue Mar 7 14:25:32 2023 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Asahi Lina X-Patchwork-Id: 13163703 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.4.0 (2014-02-07) on aws-us-west-2-korg-lkml-1.web.codeaurora.org Received: from gabe.freedesktop.org (gabe.freedesktop.org [131.252.210.177]) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (No client certificate requested) by smtp.lore.kernel.org (Postfix) with ESMTPS id 1068BC678D5 for ; Tue, 7 Mar 2023 14:27:37 +0000 (UTC) Received: from gabe.freedesktop.org (localhost [127.0.0.1]) by gabe.freedesktop.org (Postfix) with ESMTP id 11EA210E4B2; Tue, 7 Mar 2023 14:27:37 +0000 (UTC) Received: from mail.marcansoft.com (marcansoft.com [212.63.210.85]) by gabe.freedesktop.org (Postfix) with ESMTPS id 5A92C10E4B2 for ; Tue, 7 Mar 2023 14:27:34 +0000 (UTC) Received: from [127.0.0.1] (localhost [127.0.0.1]) (using TLSv1.3 with cipher TLS_AES_256_GCM_SHA384 (256/256 bits) key-exchange X25519 server-signature RSA-PSS (4096 bits) server-digest SHA256) (No client certificate requested) (Authenticated sender: linasend@asahilina.net) by mail.marcansoft.com (Postfix) with ESMTPSA id E7AF6424EB; Tue, 7 Mar 2023 14:27:25 +0000 (UTC) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/simple; d=asahilina.net; s=default; t=1678199252; bh=tiuyBWTl8Qdxvj79LVgsGFbR06YAjmOHoitbb98rzLA=; h=From:Date:Subject:References:In-Reply-To:To:Cc; b=pQIUoDU+zvY4rYeN9Otpg73BvCX394ybfNniG50+pCGTU/dRQzMT0GNTyCUTJcoaF RuCPNmEqUTzp0gTrn+45EXTK7X2QBEnupjZrbnmdsvEZPpih7MsFnL5QFTX8FW7Kc9 9VGU1FQjOdn2u08IylzPfQb120kGPd4BjesZwebSTj1+waUoiNLJUenkhaY6jZQ92s ozWFKoHyKSYwPX//CtaLhNDUABDjXC7dgtbgB83UozvJ0nGDFxMk/yIyutNqihzeDx eWtOIqOkRibkTJxe3W27ePR7Rn1BbGEWj0B3nqb0CUwX3+/XSl9hrfIh1f6CudPzC2 a2Q0ItXbVVq7g== From: Asahi Lina Date: Tue, 07 Mar 2023 23:25:32 +0900 Subject: [PATCH RFC 07/18] rust: drm: mm: Add DRM MM Range Allocator abstraction MIME-Version: 1.0 Message-Id: <20230307-rust-drm-v1-7-917ff5bc80a8@asahilina.net> References: <20230307-rust-drm-v1-0-917ff5bc80a8@asahilina.net> In-Reply-To: <20230307-rust-drm-v1-0-917ff5bc80a8@asahilina.net> To: Maarten Lankhorst , Maxime Ripard , Thomas Zimmermann , David Airlie , Daniel Vetter , Miguel Ojeda , Alex Gaynor , Wedson Almeida Filho , Boqun Feng , Gary Guo , =?utf-8?q?Bj=C3=B6rn_Roy_Baron?= , Sumit Semwal , =?utf-8?q?Christian_K=C3=B6nig?= , Luben Tuikov , Jarkko Sakkinen , Dave Hansen X-Mailer: b4 0.12.0 X-Developer-Signature: v=1; a=ed25519-sha256; t=1678199191; l=11527; i=lina@asahilina.net; s=20230221; h=from:subject:message-id; bh=tiuyBWTl8Qdxvj79LVgsGFbR06YAjmOHoitbb98rzLA=; b=1YWNy7i7Y7pqn4yADDd+wawlXS8A7BCpNq2HvNp13CJ6/glYN+hiWD6duc+ImAlFB7hpC4Nj5 qKVdeJNweZNBkjTZR9M7ZwRdGs9mp0ON/OFHohmu5CGOmuAhGYSEnZ9 X-Developer-Key: i=lina@asahilina.net; a=ed25519; pk=Qn8jZuOtR1m5GaiDfTrAoQ4NE1XoYVZ/wmt5YtXWFC4= X-BeenThere: dri-devel@lists.freedesktop.org X-Mailman-Version: 2.1.29 Precedence: list List-Id: Direct Rendering Infrastructure - Development List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Cc: linaro-mm-sig@lists.linaro.org, rust-for-linux@vger.kernel.org, Karol Herbst , asahi@lists.linux.dev, Asahi Lina , linux-kernel@vger.kernel.org, dri-devel@lists.freedesktop.org, Mary , Alyssa Rosenzweig , linux-sgx@vger.kernel.org, Ella Stanforth , Faith Ekstrand , linux-media@vger.kernel.org Errors-To: dri-devel-bounces@lists.freedesktop.org Sender: "dri-devel" drm_mm provides a simple range allocator, useful for managing virtual address ranges. Add a Rust abstraction to expose this module to Rust drivers. Signed-off-by: Asahi Lina --- rust/kernel/drm/mm.rs | 309 +++++++++++++++++++++++++++++++++++++++++++++++++ rust/kernel/drm/mod.rs | 1 + 2 files changed, 310 insertions(+) diff --git a/rust/kernel/drm/mm.rs b/rust/kernel/drm/mm.rs new file mode 100644 index 000000000000..83e27a7dcc7e --- /dev/null +++ b/rust/kernel/drm/mm.rs @@ -0,0 +1,309 @@ +// SPDX-License-Identifier: GPL-2.0 OR MIT + +//! DRM MM range allocator +//! +//! C header: [`include/linux/drm/drm_mm.h`](../../../../include/linux/drm/drm_mm.h) + +use crate::{ + bindings, + error::{to_result, Result}, + str::CStr, + sync::{Arc, LockClassKey, LockIniter, Mutex, UniqueArc}, + types::Opaque, +}; + +use alloc::boxed::Box; + +use core::{ + marker::{PhantomData, PhantomPinned}, + ops::Deref, + pin::Pin, +}; + +/// Type alias representing a DRM MM node. +pub type Node = Pin>>; + +/// Trait which must be implemented by the inner allocator state type provided by the user. +pub trait AllocInner { + /// Notification that a node was dropped from the allocator. + fn drop_object(&mut self, _start: u64, _size: u64, _color: usize, _object: &mut T) {} +} + +impl AllocInner for () {} + +/// Wrapper type for a `struct drm_mm` plus user AllocInner object. +/// +/// # Invariants +/// The `drm_mm` struct is valid and initialized. +struct MmInner, T>(Opaque, A, PhantomData); + +/// Represents a single allocated node in the MM allocator +pub struct NodeData, T> { + node: bindings::drm_mm_node, + mm: Arc>>, + valid: bool, + /// A drm_mm_node needs to be pinned because nodes reference each other in a linked list. + _pin: PhantomPinned, + inner: T, +} + +// SAFETY: Allocator ops take the mutex, and there are no mutable actions on the node. +unsafe impl, T: Send> Send for NodeData {} +unsafe impl, T: Sync> Sync for NodeData {} + +/// Available MM node insertion modes +#[repr(u32)] +pub enum InsertMode { + /// Search for the smallest hole (within the search range) that fits the desired node. + /// + /// Allocates the node from the bottom of the found hole. + Best = bindings::drm_mm_insert_mode_DRM_MM_INSERT_BEST, + + /// Search for the lowest hole (address closest to 0, within the search range) that fits the + /// desired node. + /// + /// Allocates the node from the bottom of the found hole. + Low = bindings::drm_mm_insert_mode_DRM_MM_INSERT_LOW, + + /// Search for the highest hole (address closest to U64_MAX, within the search range) that fits + /// the desired node. + /// + /// Allocates the node from the top of the found hole. The specified alignment for the node is + /// applied to the base of the node (`Node.start()`). + High = bindings::drm_mm_insert_mode_DRM_MM_INSERT_HIGH, + + /// Search for the most recently evicted hole (within the search range) that fits the desired + /// node. This is appropriate for use immediately after performing an eviction scan and removing + /// the selected nodes to form a hole. + /// + /// Allocates the node from the bottom of the found hole. + Evict = bindings::drm_mm_insert_mode_DRM_MM_INSERT_EVICT, +} + +/// A clonable, interlocked reference to the allocator state. +/// +/// This is useful to perform actions on the user-supplied `AllocInner` type given just a Node, +/// without immediately taking the lock. +#[derive(Clone)] +pub struct InnerRef, T>(Arc>>); + +impl, T> InnerRef { + /// Operate on the user `AllocInner` implementation, taking the lock. + pub fn with(&self, cb: impl FnOnce(&mut A) -> RetVal) -> RetVal { + let mut l = self.0.lock(); + cb(&mut l.1) + } +} + +impl, T> NodeData { + /// Returns the color of the node (an opaque value) + pub fn color(&self) -> usize { + self.node.color as usize + } + + /// Returns the start address of the node + pub fn start(&self) -> u64 { + self.node.start + } + + /// Returns the size of the node in bytes + pub fn size(&self) -> u64 { + self.node.size + } + + /// Operate on the user `AllocInner` implementation associated with this node's allocator. + pub fn with_inner(&self, cb: impl FnOnce(&mut A) -> RetVal) -> RetVal { + let mut l = self.mm.lock(); + cb(&mut l.1) + } + + /// Return a clonable, detached reference to the allocator inner data. + pub fn alloc_ref(&self) -> InnerRef { + InnerRef(self.mm.clone()) + } + + /// Return a mutable reference to the inner data. + pub fn inner_mut(self: Pin<&mut Self>) -> &mut T { + // SAFETY: This is okay because inner is not structural + unsafe { &mut self.get_unchecked_mut().inner } + } +} + +impl, T> Deref for NodeData { + type Target = T; + + fn deref(&self) -> &Self::Target { + &self.inner + } +} + +impl, T> Drop for NodeData { + fn drop(&mut self) { + if self.valid { + let mut guard = self.mm.lock(); + + // Inform the user allocator that a node is being dropped. + guard + .1 + .drop_object(self.start(), self.size(), self.color(), &mut self.inner); + // SAFETY: The MM lock is still taken, so we can safely remove the node. + unsafe { bindings::drm_mm_remove_node(&mut self.node) }; + } + } +} + +/// An instance of a DRM MM range allocator. +pub struct Allocator, T> { + mm: Arc>>, + _p: PhantomData, +} + +impl, T> Allocator { + /// Create a new range allocator for the given start and size range of addresses. + /// + /// The user may optionally provide an inner object representing allocator state, which will + /// be protected by the same lock. If not required, `()` can be used. + pub fn new( + start: u64, + size: u64, + inner: A, + name: &'static CStr, + lock_key: &'static LockClassKey, + ) -> Result> { + // SAFETY: We call `Mutex::init_lock` below. + let mut mm: Pin>>> = UniqueArc::try_new(unsafe { + Mutex::new(MmInner(Opaque::uninit(), inner, PhantomData)) + })? + .into(); + + mm.as_mut().init_lock(name, lock_key); + + unsafe { + // SAFETY: The Opaque instance provides a valid pointer, and it is initialized after + // this call. + bindings::drm_mm_init(mm.lock().0.get(), start, size); + } + + Ok(Allocator { + mm: mm.into(), + _p: PhantomData, + }) + } + + /// Insert a new node into the allocator of a given size. + /// + /// `node` is the user `T` type data to store into the node. + pub fn insert_node(&mut self, node: T, size: u64) -> Result> { + self.insert_node_generic(node, size, 0, 0, InsertMode::Best) + } + + /// Insert a new node into the allocator of a given size, with configurable alignment, + /// color, and insertion mode. + /// + /// `node` is the user `T` type data to store into the node. + pub fn insert_node_generic( + &mut self, + node: T, + size: u64, + alignment: u64, + color: usize, + mode: InsertMode, + ) -> Result> { + self.insert_node_in_range(node, size, alignment, color, 0, u64::MAX, mode) + } + + /// Insert a new node into the allocator of a given size, with configurable alignment, + /// color, insertion mode, and sub-range to allocate from. + /// + /// `node` is the user `T` type data to store into the node. + #[allow(clippy::too_many_arguments)] + pub fn insert_node_in_range( + &mut self, + node: T, + size: u64, + alignment: u64, + color: usize, + start: u64, + end: u64, + mode: InsertMode, + ) -> Result> { + let mut mm_node = Box::try_new(NodeData { + // SAFETY: This C struct should be zero-initialized. + node: unsafe { core::mem::zeroed() }, + valid: false, + inner: node, + mm: self.mm.clone(), + _pin: PhantomPinned, + })?; + + let guard = self.mm.lock(); + // SAFETY: We hold the lock and all pointers are valid. + to_result(unsafe { + bindings::drm_mm_insert_node_in_range( + guard.0.get(), + &mut mm_node.node, + size, + alignment, + color as core::ffi::c_ulong, + start, + end, + mode as u32, + ) + })?; + + mm_node.valid = true; + + Ok(Pin::from(mm_node)) + } + + /// Insert a node into the allocator at a fixed start address. + /// + /// `node` is the user `T` type data to store into the node. + pub fn reserve_node( + &mut self, + node: T, + start: u64, + size: u64, + color: usize, + ) -> Result> { + let mut mm_node = Box::try_new(NodeData { + // SAFETY: This C struct should be zero-initialized. + node: unsafe { core::mem::zeroed() }, + valid: false, + inner: node, + mm: self.mm.clone(), + _pin: PhantomPinned, + })?; + + mm_node.node.start = start; + mm_node.node.size = size; + mm_node.node.color = color as core::ffi::c_ulong; + + let guard = self.mm.lock(); + // SAFETY: We hold the lock and all pointers are valid. + to_result(unsafe { bindings::drm_mm_reserve_node(guard.0.get(), &mut mm_node.node) })?; + + mm_node.valid = true; + + Ok(Pin::from(mm_node)) + } + + /// Operate on the inner user type `A`, taking the allocator lock + pub fn with_inner(&self, cb: impl FnOnce(&mut A) -> RetVal) -> RetVal { + let mut guard = self.mm.lock(); + cb(&mut guard.1) + } +} + +impl, T> Drop for MmInner { + fn drop(&mut self) { + // SAFETY: If the MmInner is dropped then all nodes are gone (since they hold references), + // so it is safe to tear down the allocator. + unsafe { + bindings::drm_mm_takedown(self.0.get()); + } + } +} + +// MmInner is safely Send if the AllocInner user type is Send. +unsafe impl, T> Send for MmInner {} diff --git a/rust/kernel/drm/mod.rs b/rust/kernel/drm/mod.rs index c44760a1332f..73fab2dee3af 100644 --- a/rust/kernel/drm/mod.rs +++ b/rust/kernel/drm/mod.rs @@ -7,3 +7,4 @@ pub mod drv; pub mod file; pub mod gem; pub mod ioctl; +pub mod mm; From patchwork Tue Mar 7 14:25:33 2023 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Asahi Lina X-Patchwork-Id: 13163704 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.4.0 (2014-02-07) on aws-us-west-2-korg-lkml-1.web.codeaurora.org Received: from gabe.freedesktop.org (gabe.freedesktop.org [131.252.210.177]) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (No client certificate requested) by smtp.lore.kernel.org (Postfix) with ESMTPS id 4B674C678D4 for ; Tue, 7 Mar 2023 14:27:45 +0000 (UTC) Received: from gabe.freedesktop.org (localhost [127.0.0.1]) by gabe.freedesktop.org (Postfix) with ESMTP id 630BD10E4B3; Tue, 7 Mar 2023 14:27:44 +0000 (UTC) Received: from mail.marcansoft.com (marcansoft.com [212.63.210.85]) by gabe.freedesktop.org (Postfix) with ESMTPS id 343F510E4B3 for ; Tue, 7 Mar 2023 14:27:42 +0000 (UTC) Received: from [127.0.0.1] (localhost [127.0.0.1]) (using TLSv1.3 with cipher TLS_AES_256_GCM_SHA384 (256/256 bits) key-exchange X25519 server-signature RSA-PSS (4096 bits) server-digest SHA256) (No client certificate requested) (Authenticated sender: linasend@asahilina.net) by mail.marcansoft.com (Postfix) with ESMTPSA id 61DA14261B; Tue, 7 Mar 2023 14:27:33 +0000 (UTC) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/simple; d=asahilina.net; s=default; t=1678199260; bh=qxI6GJmNo0Y4+6c3Pa+ta2DSc4p6o+YuxslYcsGJFM4=; h=From:Date:Subject:References:In-Reply-To:To:Cc; b=ZcXSxP3Nj8rQiVFU5QoKVbKGsKDqHqTuptTxSXb/T++6CSJflmJ6UyjAAT2xKxLCF 5AeJuJyumpsVwn46UZijLtezdC7Z9aEhd734nS5JGxluZXzZQ/HlRiaXond8FeBdxU 45QdqLToECPJ/lWxzcqXZVR0u9dZzKqa5CuI643IONL5Uh0pUiBnRU1bAj8OOnjLdP L0krGvpuwnciNiPCT8lvsbIlRYzrCBerjLMwTQIH486utIZNAtUCgDcq2wsbTjN4h7 KWq1FU+enwWdsvDLytFsYG5znXhtrKP1ziisN/AN1ufxC3thyf1Ohxy5oQLpUP+n6c aT0njZjUwaXFA== From: Asahi Lina Date: Tue, 07 Mar 2023 23:25:33 +0900 Subject: [PATCH RFC 08/18] rust: dma_fence: Add DMA Fence abstraction MIME-Version: 1.0 Message-Id: <20230307-rust-drm-v1-8-917ff5bc80a8@asahilina.net> References: <20230307-rust-drm-v1-0-917ff5bc80a8@asahilina.net> In-Reply-To: <20230307-rust-drm-v1-0-917ff5bc80a8@asahilina.net> To: Maarten Lankhorst , Maxime Ripard , Thomas Zimmermann , David Airlie , Daniel Vetter , Miguel Ojeda , Alex Gaynor , Wedson Almeida Filho , Boqun Feng , Gary Guo , =?utf-8?q?Bj=C3=B6rn_Roy_Baron?= , Sumit Semwal , =?utf-8?q?Christian_K=C3=B6nig?= , Luben Tuikov , Jarkko Sakkinen , Dave Hansen X-Mailer: b4 0.12.0 X-Developer-Signature: v=1; a=ed25519-sha256; t=1678199191; l=21370; i=lina@asahilina.net; s=20230221; h=from:subject:message-id; bh=qxI6GJmNo0Y4+6c3Pa+ta2DSc4p6o+YuxslYcsGJFM4=; b=mX8/ZtYERXNdbWtwabzrHCLIOywZes6UBii6Z13L/frmDi6cg/WtPxMZuURZwWM32yFmjCOEN JT3Dr3DG2L2AMfhB4cY+egFO/xkLF5EEkuU/9sOZLjXXs5VnxyMt2Ji X-Developer-Key: i=lina@asahilina.net; a=ed25519; pk=Qn8jZuOtR1m5GaiDfTrAoQ4NE1XoYVZ/wmt5YtXWFC4= X-BeenThere: dri-devel@lists.freedesktop.org X-Mailman-Version: 2.1.29 Precedence: list List-Id: Direct Rendering Infrastructure - Development List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Cc: linaro-mm-sig@lists.linaro.org, rust-for-linux@vger.kernel.org, Karol Herbst , asahi@lists.linux.dev, Asahi Lina , linux-kernel@vger.kernel.org, dri-devel@lists.freedesktop.org, Mary , Alyssa Rosenzweig , linux-sgx@vger.kernel.org, Ella Stanforth , Faith Ekstrand , linux-media@vger.kernel.org Errors-To: dri-devel-bounces@lists.freedesktop.org Sender: "dri-devel" DMA fences are the internal synchronization primitive used for DMA operations like GPU rendering, video en/decoding, etc. Add an abstraction to allow Rust drivers to interact with this subsystem. Note: This uses a raw spinlock living next to the fence, since we do not interact with it other than for initialization. TODO: Expose this to the user at some point with a safe abstraction. Signed-off-by: Asahi Lina --- rust/bindings/bindings_helper.h | 2 + rust/helpers.c | 53 ++++ rust/kernel/dma_fence.rs | 532 ++++++++++++++++++++++++++++++++++++++++ rust/kernel/lib.rs | 2 + 4 files changed, 589 insertions(+) diff --git a/rust/bindings/bindings_helper.h b/rust/bindings/bindings_helper.h index 9f152d373df8..705af292a5b4 100644 --- a/rust/bindings/bindings_helper.h +++ b/rust/bindings/bindings_helper.h @@ -14,6 +14,8 @@ #include #include #include +#include +#include #include #include #include diff --git a/rust/helpers.c b/rust/helpers.c index 388ff1100ea5..8e906a7a7d8a 100644 --- a/rust/helpers.c +++ b/rust/helpers.c @@ -23,6 +23,8 @@ #include #include #include +#include +#include #include #include #include @@ -30,6 +32,7 @@ #include #include #include +#include #include #include #include @@ -388,6 +391,56 @@ int rust_helper_sg_dma_len(const struct scatterlist *sg) } EXPORT_SYMBOL_GPL(rust_helper_sg_dma_len); +void rust_helper___spin_lock_init(spinlock_t *lock, const char *name, + struct lock_class_key *key) +{ +#ifdef CONFIG_DEBUG_SPINLOCK +# ifndef CONFIG_PREEMPT_RT + __raw_spin_lock_init(spinlock_check(lock), name, key, LD_WAIT_CONFIG); +# else + rt_mutex_base_init(&lock->lock); + __rt_spin_lock_init(lock, name, key, false); +# endif +#else + spin_lock_init(lock); +#endif +} +EXPORT_SYMBOL_GPL(rust_helper___spin_lock_init); + +#ifdef CONFIG_DMA_SHARED_BUFFER + +void rust_helper_dma_fence_get(struct dma_fence *fence) +{ + dma_fence_get(fence); +} +EXPORT_SYMBOL_GPL(rust_helper_dma_fence_get); + +void rust_helper_dma_fence_put(struct dma_fence *fence) +{ + dma_fence_put(fence); +} +EXPORT_SYMBOL_GPL(rust_helper_dma_fence_put); + +struct dma_fence_chain *rust_helper_dma_fence_chain_alloc(void) +{ + return dma_fence_chain_alloc(); +} +EXPORT_SYMBOL_GPL(rust_helper_dma_fence_chain_alloc); + +void rust_helper_dma_fence_chain_free(struct dma_fence_chain *chain) +{ + dma_fence_chain_free(chain); +} +EXPORT_SYMBOL_GPL(rust_helper_dma_fence_chain_free); + +void rust_helper_dma_fence_set_error(struct dma_fence *fence, int error) +{ + dma_fence_set_error(fence, error); +} +EXPORT_SYMBOL_GPL(rust_helper_dma_fence_set_error); + +#endif + #ifdef CONFIG_DRM void rust_helper_drm_gem_object_get(struct drm_gem_object *obj) diff --git a/rust/kernel/dma_fence.rs b/rust/kernel/dma_fence.rs new file mode 100644 index 000000000000..ca93380d9da2 --- /dev/null +++ b/rust/kernel/dma_fence.rs @@ -0,0 +1,532 @@ +// SPDX-License-Identifier: GPL-2.0 + +//! DMA fence abstraction. +//! +//! C header: [`include/linux/dma_fence.h`](../../include/linux/dma_fence.h) + +use crate::{ + bindings, + error::{to_result, Result}, + prelude::*, + sync::LockClassKey, + types::Opaque, +}; +use core::fmt::Write; +use core::ops::{Deref, DerefMut}; +use core::ptr::addr_of_mut; +use core::sync::atomic::{AtomicU64, Ordering}; + +/// Any kind of DMA Fence Object +/// +/// # Invariants +/// raw() returns a valid pointer to a dma_fence and we own a reference to it. +pub trait RawDmaFence: crate::private::Sealed { + /// Returns the raw `struct dma_fence` pointer. + fn raw(&self) -> *mut bindings::dma_fence; + + /// Returns the raw `struct dma_fence` pointer and consumes the object. + /// + /// The caller is responsible for dropping the reference. + fn into_raw(self) -> *mut bindings::dma_fence + where + Self: Sized, + { + let ptr = self.raw(); + core::mem::forget(self); + ptr + } + + /// Advances this fence to the chain node which will signal this sequence number. + /// If no sequence number is provided, this returns `self` again. + fn chain_find_seqno(self, seqno: u64) -> Result + where + Self: Sized, + { + let mut ptr = self.into_raw(); + + // SAFETY: This will safely fail if this DmaFence is not a chain. + // `ptr` is valid per the type invariant. + let ret = unsafe { bindings::dma_fence_chain_find_seqno(&mut ptr, seqno) }; + + if ret != 0 { + // SAFETY: This is either an owned reference or NULL, dma_fence_put can handle both. + unsafe { bindings::dma_fence_put(ptr) }; + Err(Error::from_kernel_errno(ret)) + } else if ptr.is_null() { + Err(EINVAL) // When can this happen? + } else { + // SAFETY: ptr is valid and non-NULL as checked above. + Ok(unsafe { Fence::from_raw(ptr) }) + } + } + + /// Signal completion of this fence + fn signal(&self) -> Result { + to_result(unsafe { bindings::dma_fence_signal(self.raw()) }) + } + + /// Set the error flag on this fence + fn set_error(&self, err: Error) { + unsafe { bindings::dma_fence_set_error(self.raw(), err.to_kernel_errno()) }; + } +} + +/// A generic DMA Fence Object +/// +/// # Invariants +/// ptr is a valid pointer to a dma_fence and we own a reference to it. +pub struct Fence { + ptr: *mut bindings::dma_fence, +} + +impl Fence { + /// Create a new Fence object from a raw pointer to a dma_fence. + /// + /// # Safety + /// The caller must own a reference to the dma_fence, which is transferred to the new object. + pub(crate) unsafe fn from_raw(ptr: *mut bindings::dma_fence) -> Fence { + Fence { ptr } + } + + /// Create a new Fence object from a raw pointer to a dma_fence. + /// + /// # Safety + /// Takes a borrowed reference to the dma_fence, and increments the reference count. + pub(crate) unsafe fn get_raw(ptr: *mut bindings::dma_fence) -> Fence { + // SAFETY: Pointer is valid per the safety contract + unsafe { bindings::dma_fence_get(ptr) }; + Fence { ptr } + } + + /// Create a new Fence object from a RawDmaFence. + pub fn from_fence(fence: &dyn RawDmaFence) -> Fence { + // SAFETY: Pointer is valid per the RawDmaFence contract + unsafe { Self::get_raw(fence.raw()) } + } +} + +impl crate::private::Sealed for Fence {} + +impl RawDmaFence for Fence { + fn raw(&self) -> *mut bindings::dma_fence { + self.ptr + } +} + +impl Drop for Fence { + fn drop(&mut self) { + // SAFETY: We own a reference to this syncobj. + unsafe { bindings::dma_fence_put(self.ptr) }; + } +} + +impl Clone for Fence { + fn clone(&self) -> Self { + // SAFETY: `ptr` is valid per the type invariant and we own a reference to it. + unsafe { + bindings::dma_fence_get(self.ptr); + Self::from_raw(self.ptr) + } + } +} + +unsafe impl Sync for Fence {} +unsafe impl Send for Fence {} + +/// Trait which must be implemented by driver-specific fence objects. +#[vtable] +pub trait FenceOps: Sized + Send + Sync { + /// True if this dma_fence implementation uses 64bit seqno, false otherwise. + const USE_64BIT_SEQNO: bool; + + /// Returns the driver name. This is a callback to allow drivers to compute the name at + /// runtime, without having it to store permanently for each fence, or build a cache of + /// some sort. + fn get_driver_name<'a>(self: &'a FenceObject) -> &'a CStr; + + /// Return the name of the context this fence belongs to. This is a callback to allow drivers + /// to compute the name at runtime, without having it to store permanently for each fence, or + /// build a cache of some sort. + fn get_timeline_name<'a>(self: &'a FenceObject) -> &'a CStr; + + /// Enable software signaling of fence. + fn enable_signaling(self: &FenceObject) -> bool { + false + } + + /// Peek whether the fence is signaled, as a fastpath optimization for e.g. dma_fence_wait() or + /// dma_fence_add_callback(). + fn signaled(self: &FenceObject) -> bool { + false + } + + /// Callback to fill in free-form debug info specific to this fence, like the sequence number. + fn fence_value_str(self: &FenceObject, _output: &mut dyn Write) {} + + /// Fills in the current value of the timeline as a string, like the sequence number. Note that + /// the specific fence passed to this function should not matter, drivers should only use it to + /// look up the corresponding timeline structures. + fn timeline_value_str(self: &FenceObject, _output: &mut dyn Write) {} +} + +unsafe extern "C" fn get_driver_name_cb( + fence: *mut bindings::dma_fence, +) -> *const core::ffi::c_char { + // SAFETY: All of our fences are FenceObject. + let p = crate::container_of!(fence, FenceObject, fence) as *mut FenceObject; + + // SAFETY: The caller is responsible for passing a valid dma_fence subtype + T::get_driver_name(unsafe { &mut *p }).as_char_ptr() +} + +unsafe extern "C" fn get_timeline_name_cb( + fence: *mut bindings::dma_fence, +) -> *const core::ffi::c_char { + // SAFETY: All of our fences are FenceObject. + let p = crate::container_of!(fence, FenceObject, fence) as *mut FenceObject; + + // SAFETY: The caller is responsible for passing a valid dma_fence subtype + T::get_timeline_name(unsafe { &mut *p }).as_char_ptr() +} + +unsafe extern "C" fn enable_signaling_cb(fence: *mut bindings::dma_fence) -> bool { + // SAFETY: All of our fences are FenceObject. + let p = crate::container_of!(fence, FenceObject, fence) as *mut FenceObject; + + // SAFETY: The caller is responsible for passing a valid dma_fence subtype + T::enable_signaling(unsafe { &mut *p }) +} + +unsafe extern "C" fn signaled_cb(fence: *mut bindings::dma_fence) -> bool { + // SAFETY: All of our fences are FenceObject. + let p = crate::container_of!(fence, FenceObject, fence) as *mut FenceObject; + + // SAFETY: The caller is responsible for passing a valid dma_fence subtype + T::signaled(unsafe { &mut *p }) +} + +unsafe extern "C" fn release_cb(fence: *mut bindings::dma_fence) { + // SAFETY: All of our fences are FenceObject. + let p = crate::container_of!(fence, FenceObject, fence) as *mut FenceObject; + + // SAFETY: p is never used after this + unsafe { + core::ptr::drop_in_place(&mut (*p).inner); + } + + // SAFETY: All of our fences are allocated using kmalloc, so this is safe. + unsafe { bindings::dma_fence_free(fence) }; +} + +unsafe extern "C" fn fence_value_str_cb( + fence: *mut bindings::dma_fence, + string: *mut core::ffi::c_char, + size: core::ffi::c_int, +) { + let size: usize = size.try_into().unwrap_or(0); + + if size == 0 { + return; + } + + // SAFETY: All of our fences are FenceObject. + let p = crate::container_of!(fence, FenceObject, fence) as *mut FenceObject; + + // SAFETY: The caller is responsible for the validity of string/size + let mut f = unsafe { crate::str::Formatter::from_buffer(string as *mut _, size) }; + + // SAFETY: The caller is responsible for passing a valid dma_fence subtype + T::fence_value_str(unsafe { &mut *p }, &mut f); + let _ = f.write_str("\0"); + + // SAFETY: `size` is at least 1 per the check above + unsafe { *string.add(size - 1) = 0 }; +} + +unsafe extern "C" fn timeline_value_str_cb( + fence: *mut bindings::dma_fence, + string: *mut core::ffi::c_char, + size: core::ffi::c_int, +) { + let size: usize = size.try_into().unwrap_or(0); + + if size == 0 { + return; + } + + // SAFETY: All of our fences are FenceObject. + let p = crate::container_of!(fence, FenceObject, fence) as *mut FenceObject; + + // SAFETY: The caller is responsible for the validity of string/size + let mut f = unsafe { crate::str::Formatter::from_buffer(string as *mut _, size) }; + + // SAFETY: The caller is responsible for passing a valid dma_fence subtype + T::timeline_value_str(unsafe { &mut *p }, &mut f); + let _ = f.write_str("\0"); + + // SAFETY: `size` is at least 1 per the check above + unsafe { *string.add(size - 1) = 0 }; +} + +// Allow FenceObject to be used as a self argument, for ergonomics +impl core::ops::Receiver for FenceObject {} + +/// A driver-specific DMA Fence Object +/// +/// # Invariants +/// ptr is a valid pointer to a dma_fence and we own a reference to it. +#[repr(C)] +pub struct FenceObject { + fence: bindings::dma_fence, + lock: Opaque, + inner: T, +} + +impl FenceObject { + const SIZE: usize = core::mem::size_of::(); + + const VTABLE: bindings::dma_fence_ops = bindings::dma_fence_ops { + use_64bit_seqno: T::USE_64BIT_SEQNO, + get_driver_name: Some(get_driver_name_cb::), + get_timeline_name: Some(get_timeline_name_cb::), + enable_signaling: if T::HAS_ENABLE_SIGNALING { + Some(enable_signaling_cb::) + } else { + None + }, + signaled: if T::HAS_SIGNALED { + Some(signaled_cb::) + } else { + None + }, + wait: None, // Deprecated + release: Some(release_cb::), + fence_value_str: if T::HAS_FENCE_VALUE_STR { + Some(fence_value_str_cb::) + } else { + None + }, + timeline_value_str: if T::HAS_TIMELINE_VALUE_STR { + Some(timeline_value_str_cb::) + } else { + None + }, + }; +} + +impl Deref for FenceObject { + type Target = T; + + fn deref(&self) -> &T { + &self.inner + } +} + +impl DerefMut for FenceObject { + fn deref_mut(&mut self) -> &mut T { + &mut self.inner + } +} + +impl crate::private::Sealed for FenceObject {} +impl RawDmaFence for FenceObject { + fn raw(&self) -> *mut bindings::dma_fence { + &self.fence as *const _ as *mut _ + } +} + +/// A unique reference to a driver-specific fence object +pub struct UniqueFence(*mut FenceObject); + +impl Deref for UniqueFence { + type Target = FenceObject; + + fn deref(&self) -> &FenceObject { + unsafe { &*self.0 } + } +} + +impl DerefMut for UniqueFence { + fn deref_mut(&mut self) -> &mut FenceObject { + unsafe { &mut *self.0 } + } +} + +impl crate::private::Sealed for UniqueFence {} +impl RawDmaFence for UniqueFence { + fn raw(&self) -> *mut bindings::dma_fence { + unsafe { addr_of_mut!((*self.0).fence) } + } +} + +impl From> for UserFence { + fn from(value: UniqueFence) -> Self { + let ptr = value.0; + core::mem::forget(value); + + UserFence(ptr) + } +} + +impl Drop for UniqueFence { + fn drop(&mut self) { + // SAFETY: We own a reference to this fence. + unsafe { bindings::dma_fence_put(self.raw()) }; + } +} + +unsafe impl Sync for UniqueFence {} +unsafe impl Send for UniqueFence {} + +/// A shared reference to a driver-specific fence object +pub struct UserFence(*mut FenceObject); + +impl Deref for UserFence { + type Target = FenceObject; + + fn deref(&self) -> &FenceObject { + unsafe { &*self.0 } + } +} + +impl Clone for UserFence { + fn clone(&self) -> Self { + // SAFETY: `ptr` is valid per the type invariant and we own a reference to it. + unsafe { + bindings::dma_fence_get(self.raw()); + Self(self.0) + } + } +} + +impl crate::private::Sealed for UserFence {} +impl RawDmaFence for UserFence { + fn raw(&self) -> *mut bindings::dma_fence { + unsafe { addr_of_mut!((*self.0).fence) } + } +} + +impl Drop for UserFence { + fn drop(&mut self) { + // SAFETY: We own a reference to this fence. + unsafe { bindings::dma_fence_put(self.raw()) }; + } +} + +unsafe impl Sync for UserFence {} +unsafe impl Send for UserFence {} + +/// An array of fence contexts, out of which fences can be created. +pub struct FenceContexts { + start: u64, + count: u32, + seqnos: Vec, + lock_name: &'static CStr, + lock_key: &'static LockClassKey, +} + +impl FenceContexts { + /// Create a new set of fence contexts. + pub fn new( + count: u32, + name: &'static CStr, + key: &'static LockClassKey, + ) -> Result { + let mut seqnos: Vec = Vec::new(); + + seqnos.try_reserve(count as usize)?; + + for _ in 0..count { + seqnos.try_push(Default::default())?; + } + + let start = unsafe { bindings::dma_fence_context_alloc(count as core::ffi::c_uint) }; + + Ok(FenceContexts { + start, + count, + seqnos, + lock_name: name, + lock_key: key, + }) + } + + /// Create a new fence in a given context index. + pub fn new_fence(&self, context: u32, inner: T) -> Result> { + if context > self.count { + return Err(EINVAL); + } + + let p = unsafe { + bindings::krealloc( + core::ptr::null_mut(), + FenceObject::::SIZE, + bindings::GFP_KERNEL | bindings::__GFP_ZERO, + ) as *mut FenceObject + }; + + if p.is_null() { + return Err(ENOMEM); + } + + let seqno = self.seqnos[context as usize].fetch_add(1, Ordering::Relaxed); + + // SAFETY: The pointer is valid, so pointers to members are too. + // After this, all fields are initialized. + unsafe { + addr_of_mut!((*p).inner).write(inner); + bindings::__spin_lock_init( + addr_of_mut!((*p).lock) as *mut _, + self.lock_name.as_char_ptr(), + self.lock_key.get(), + ); + bindings::dma_fence_init( + addr_of_mut!((*p).fence), + &FenceObject::::VTABLE, + addr_of_mut!((*p).lock) as *mut _, + self.start + context as u64, + seqno, + ); + }; + + Ok(UniqueFence(p)) + } +} + +/// A DMA Fence Chain Object +/// +/// # Invariants +/// ptr is a valid pointer to a dma_fence_chain which we own. +pub struct FenceChain { + ptr: *mut bindings::dma_fence_chain, +} + +impl FenceChain { + /// Create a new DmaFenceChain object. + pub fn new() -> Result { + // SAFETY: This function is safe to call and takes no arguments. + let ptr = unsafe { bindings::dma_fence_chain_alloc() }; + + if ptr.is_null() { + Err(ENOMEM) + } else { + Ok(FenceChain { ptr }) + } + } + + /// Convert the DmaFenceChain into the underlying raw pointer. + /// + /// This assumes the caller will take ownership of the object. + pub(crate) fn into_raw(self) -> *mut bindings::dma_fence_chain { + let ptr = self.ptr; + core::mem::forget(self); + ptr + } +} + +impl Drop for FenceChain { + fn drop(&mut self) { + // SAFETY: We own this dma_fence_chain. + unsafe { bindings::dma_fence_chain_free(self.ptr) }; + } +} diff --git a/rust/kernel/lib.rs b/rust/kernel/lib.rs index cb23d24c6718..31866069e0bc 100644 --- a/rust/kernel/lib.rs +++ b/rust/kernel/lib.rs @@ -36,6 +36,8 @@ mod allocator; mod build_assert; pub mod delay; pub mod device; +#[cfg(CONFIG_DMA_SHARED_BUFFER)] +pub mod dma_fence; pub mod driver; #[cfg(CONFIG_RUST_DRM)] pub mod drm; From patchwork Tue Mar 7 14:25:34 2023 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Asahi Lina X-Patchwork-Id: 13163705 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.4.0 (2014-02-07) on aws-us-west-2-korg-lkml-1.web.codeaurora.org Received: from gabe.freedesktop.org (gabe.freedesktop.org [131.252.210.177]) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (No client certificate requested) by smtp.lore.kernel.org (Postfix) with ESMTPS id 7EA2BC6FD1A for ; Tue, 7 Mar 2023 14:27:51 +0000 (UTC) Received: from gabe.freedesktop.org (localhost [127.0.0.1]) by gabe.freedesktop.org (Postfix) with ESMTP id 90EB610E4BA; Tue, 7 Mar 2023 14:27:50 +0000 (UTC) Received: from mail.marcansoft.com (marcansoft.com [212.63.210.85]) by gabe.freedesktop.org (Postfix) with ESMTPS id 6F38610E4BA for ; Tue, 7 Mar 2023 14:27:49 +0000 (UTC) Received: from [127.0.0.1] (localhost [127.0.0.1]) (using TLSv1.3 with cipher TLS_AES_256_GCM_SHA384 (256/256 bits) key-exchange X25519 server-signature RSA-PSS (4096 bits) server-digest SHA256) (No client certificate requested) (Authenticated sender: linasend@asahilina.net) by mail.marcansoft.com (Postfix) with ESMTPSA id 125B5426E8; Tue, 7 Mar 2023 14:27:40 +0000 (UTC) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/simple; d=asahilina.net; s=default; t=1678199268; bh=RXHMryixfrrNFrvpRW0yNj/9C2ku8y/xcsih3nh/K2g=; h=From:Date:Subject:References:In-Reply-To:To:Cc; b=NxIVR99EHxOkBhHnp9AFtyvk5YP6jlHLfi/ymoHar6CZ0n7F505Dd6iVAMRTL5QRp +4SIidMOCx4/g+KSGpisxU2/M5dz7+AESAreRIsoj9D6ABquOse1rW23VrVV5DYbdM REyTm1j7bW0+bVAeTDeyoqOU7COJjNXzHkCy4kyohj45fj1MAlXMoBWKCalQdM7zMj MtpJdv4agSVmqjY+7miY7bQxiwl1gmN1DLafPoxC+ALSEJyPXxHL/+LeObnZ25V6Jg j2vb34b1egvDVkBqiB9E+SgBbgXQ3G1g7lZs3x0lH84VkYPBFju/+dFy0LwyhA2PwA 0Rm7RI3NbyHYg== From: Asahi Lina Date: Tue, 07 Mar 2023 23:25:34 +0900 Subject: [PATCH RFC 09/18] rust: drm: syncobj: Add DRM Sync Object abstraction MIME-Version: 1.0 Message-Id: <20230307-rust-drm-v1-9-917ff5bc80a8@asahilina.net> References: <20230307-rust-drm-v1-0-917ff5bc80a8@asahilina.net> In-Reply-To: <20230307-rust-drm-v1-0-917ff5bc80a8@asahilina.net> To: Maarten Lankhorst , Maxime Ripard , Thomas Zimmermann , David Airlie , Daniel Vetter , Miguel Ojeda , Alex Gaynor , Wedson Almeida Filho , Boqun Feng , Gary Guo , =?utf-8?q?Bj=C3=B6rn_Roy_Baron?= , Sumit Semwal , =?utf-8?q?Christian_K=C3=B6nig?= , Luben Tuikov , Jarkko Sakkinen , Dave Hansen X-Mailer: b4 0.12.0 X-Developer-Signature: v=1; a=ed25519-sha256; t=1678199191; l=5195; i=lina@asahilina.net; s=20230221; h=from:subject:message-id; bh=RXHMryixfrrNFrvpRW0yNj/9C2ku8y/xcsih3nh/K2g=; b=8Rqbn6Odm2BgzSdkvy7KVrq1XGagKtlDqBAW0tDTf1KXpEEO3aWA7ck+4e3ZI7TTnura3hbcK h1M+YNWSgnKCLVM3okF8IHi1ZpkpnTFRO7R1YkEfo84kVI4X2h78Sq4 X-Developer-Key: i=lina@asahilina.net; a=ed25519; pk=Qn8jZuOtR1m5GaiDfTrAoQ4NE1XoYVZ/wmt5YtXWFC4= X-BeenThere: dri-devel@lists.freedesktop.org X-Mailman-Version: 2.1.29 Precedence: list List-Id: Direct Rendering Infrastructure - Development List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Cc: linaro-mm-sig@lists.linaro.org, rust-for-linux@vger.kernel.org, Karol Herbst , asahi@lists.linux.dev, Asahi Lina , linux-kernel@vger.kernel.org, dri-devel@lists.freedesktop.org, Mary , Alyssa Rosenzweig , linux-sgx@vger.kernel.org, Ella Stanforth , Faith Ekstrand , linux-media@vger.kernel.org Errors-To: dri-devel-bounces@lists.freedesktop.org Sender: "dri-devel" DRM Sync Objects are a container for a DMA fence, and can be waited on signaled, exported, and imported from userspace. Add a Rust abstraction so Rust DRM drivers can support this functionality. Signed-off-by: Asahi Lina --- rust/bindings/bindings_helper.h | 1 + rust/helpers.c | 19 ++++++++++ rust/kernel/drm/mod.rs | 1 + rust/kernel/drm/syncobj.rs | 77 +++++++++++++++++++++++++++++++++++++++++ 4 files changed, 98 insertions(+) diff --git a/rust/bindings/bindings_helper.h b/rust/bindings/bindings_helper.h index 705af292a5b4..b6696011f3a4 100644 --- a/rust/bindings/bindings_helper.h +++ b/rust/bindings/bindings_helper.h @@ -12,6 +12,7 @@ #include #include #include +#include #include #include #include diff --git a/rust/helpers.c b/rust/helpers.c index 8e906a7a7d8a..11965b1e2f4e 100644 --- a/rust/helpers.c +++ b/rust/helpers.c @@ -20,6 +20,7 @@ #include #include +#include #include #include #include @@ -461,6 +462,24 @@ __u64 rust_helper_drm_vma_node_offset_addr(struct drm_vma_offset_node *node) } EXPORT_SYMBOL_GPL(rust_helper_drm_vma_node_offset_addr); +void rust_helper_drm_syncobj_get(struct drm_syncobj *obj) +{ + drm_syncobj_get(obj); +} +EXPORT_SYMBOL_GPL(rust_helper_drm_syncobj_get); + +void rust_helper_drm_syncobj_put(struct drm_syncobj *obj) +{ + drm_syncobj_put(obj); +} +EXPORT_SYMBOL_GPL(rust_helper_drm_syncobj_put); + +struct dma_fence *rust_helper_drm_syncobj_fence_get(struct drm_syncobj *syncobj) +{ + return drm_syncobj_fence_get(syncobj); +} +EXPORT_SYMBOL_GPL(rust_helper_drm_syncobj_fence_get); + #ifdef CONFIG_DRM_GEM_SHMEM_HELPER void rust_helper_drm_gem_shmem_object_free(struct drm_gem_object *obj) diff --git a/rust/kernel/drm/mod.rs b/rust/kernel/drm/mod.rs index 73fab2dee3af..dae98826edfd 100644 --- a/rust/kernel/drm/mod.rs +++ b/rust/kernel/drm/mod.rs @@ -8,3 +8,4 @@ pub mod file; pub mod gem; pub mod ioctl; pub mod mm; +pub mod syncobj; diff --git a/rust/kernel/drm/syncobj.rs b/rust/kernel/drm/syncobj.rs new file mode 100644 index 000000000000..10eed05eb27a --- /dev/null +++ b/rust/kernel/drm/syncobj.rs @@ -0,0 +1,77 @@ +// SPDX-License-Identifier: GPL-2.0 OR MIT + +//! DRM Sync Objects +//! +//! C header: [`include/linux/drm/drm_syncobj.h`](../../../../include/linux/drm/drm_syncobj.h) + +use crate::{bindings, dma_fence::*, drm, error::Result, prelude::*}; + +/// A DRM Sync Object +/// +/// # Invariants +/// ptr is a valid pointer to a drm_syncobj and we own a reference to it. +pub struct SyncObj { + ptr: *mut bindings::drm_syncobj, +} + +impl SyncObj { + /// Looks up a sync object by its handle for a given `File`. + pub fn lookup_handle(file: &impl drm::file::GenericFile, handle: u32) -> Result { + // SAFETY: The arguments are all valid per the type invariants. + let ptr = unsafe { bindings::drm_syncobj_find(file.raw() as *mut _, handle) }; + + if ptr.is_null() { + Err(ENOENT) + } else { + Ok(SyncObj { ptr }) + } + } + + /// Returns the DMA fence associated with this sync object, if any. + pub fn fence_get(&self) -> Option { + let fence = unsafe { bindings::drm_syncobj_fence_get(self.ptr) }; + if fence.is_null() { + None + } else { + // SAFETY: The pointer is non-NULL and drm_syncobj_fence_get acquired an + // additional reference. + Some(unsafe { Fence::from_raw(fence) }) + } + } + + /// Replaces the DMA fence with a new one, or removes it if fence is None. + pub fn replace_fence(&self, fence: Option<&Fence>) { + unsafe { + bindings::drm_syncobj_replace_fence( + self.ptr, + fence.map_or(core::ptr::null_mut(), |a| a.raw()), + ) + }; + } + + /// Adds a new timeline point to the syncobj. + pub fn add_point(&self, chain: FenceChain, fence: &Fence, point: u64) { + // SAFETY: All arguments should be valid per the respective type invariants. + // This takes over the FenceChain ownership. + unsafe { bindings::drm_syncobj_add_point(self.ptr, chain.into_raw(), fence.raw(), point) }; + } +} + +impl Drop for SyncObj { + fn drop(&mut self) { + // SAFETY: We own a reference to this syncobj. + unsafe { bindings::drm_syncobj_put(self.ptr) }; + } +} + +impl Clone for SyncObj { + fn clone(&self) -> Self { + // SAFETY: `ptr` is valid per the type invariant and we own a reference to it. + unsafe { bindings::drm_syncobj_get(self.ptr) }; + SyncObj { ptr: self.ptr } + } +} + +// SAFETY: drm_syncobj operations are internally locked. +unsafe impl Sync for SyncObj {} +unsafe impl Send for SyncObj {} From patchwork Tue Mar 7 14:25:35 2023 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Asahi Lina X-Patchwork-Id: 13163706 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.4.0 (2014-02-07) on aws-us-west-2-korg-lkml-1.web.codeaurora.org Received: from gabe.freedesktop.org (gabe.freedesktop.org [131.252.210.177]) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (No client certificate requested) by smtp.lore.kernel.org (Postfix) with ESMTPS id B2C31C678D5 for ; Tue, 7 Mar 2023 14:27:59 +0000 (UTC) Received: from gabe.freedesktop.org (localhost [127.0.0.1]) by gabe.freedesktop.org (Postfix) with ESMTP id EED0110E4B6; Tue, 7 Mar 2023 14:27:58 +0000 (UTC) Received: from mail.marcansoft.com (marcansoft.com [212.63.210.85]) by gabe.freedesktop.org (Postfix) with ESMTPS id DF9D710E4BB for ; Tue, 7 Mar 2023 14:27:56 +0000 (UTC) Received: from [127.0.0.1] (localhost [127.0.0.1]) (using TLSv1.3 with cipher TLS_AES_256_GCM_SHA384 (256/256 bits) key-exchange X25519 server-signature RSA-PSS (4096 bits) server-digest SHA256) (No client certificate requested) (Authenticated sender: linasend@asahilina.net) by mail.marcansoft.com (Postfix) with ESMTPSA id 7CD0C4261B; Tue, 7 Mar 2023 14:27:48 +0000 (UTC) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/simple; d=asahilina.net; s=default; t=1678199275; bh=HxG6/03qUzMmm6Reqvh7X3TDLICxkEPU4hm2AoOUuLM=; h=From:Date:Subject:References:In-Reply-To:To:Cc; b=g3xFbOLpqvFL/hWEeMzgFc3LbdSVcVVsMJh/7XdjhA21c4wx1kxrwdRtnbBmxPrrB 2CGtpUGuIbw1gEu5imAbrxziNlTiqNnrk9HQk6htxBgOV6yuRyAeYQAftV86CAMQyw rFeLzqMi0pBOmRpRumEPybDrr5v4ua3EE16+Kga4TDShw9EzMtRDc5h6TWQR4AZqq7 3g7gcLKGYG8RlArpFBJ0vDqyTFx613T1iY05PGhXC9pjVYp6nnc64K8hGhUGFW22jy LTR2hm4pngwLC6oJFdPUBUSRu+WicCyQcDZg4hcZo9sny6I9PNUD4JoUMn1kTrd0Pt ZdyZOrGEziyIA== From: Asahi Lina Date: Tue, 07 Mar 2023 23:25:35 +0900 Subject: [PATCH RFC 10/18] drm/scheduler: Add can_run_job callback MIME-Version: 1.0 Message-Id: <20230307-rust-drm-v1-10-917ff5bc80a8@asahilina.net> References: <20230307-rust-drm-v1-0-917ff5bc80a8@asahilina.net> In-Reply-To: <20230307-rust-drm-v1-0-917ff5bc80a8@asahilina.net> To: Maarten Lankhorst , Maxime Ripard , Thomas Zimmermann , David Airlie , Daniel Vetter , Miguel Ojeda , Alex Gaynor , Wedson Almeida Filho , Boqun Feng , Gary Guo , =?utf-8?q?Bj=C3=B6rn_Roy_Baron?= , Sumit Semwal , =?utf-8?q?Christian_K=C3=B6nig?= , Luben Tuikov , Jarkko Sakkinen , Dave Hansen X-Mailer: b4 0.12.0 X-Developer-Signature: v=1; a=ed25519-sha256; t=1678199191; l=1943; i=lina@asahilina.net; s=20230221; h=from:subject:message-id; bh=HxG6/03qUzMmm6Reqvh7X3TDLICxkEPU4hm2AoOUuLM=; b=gWfErbdeUWtPqx2eFx2Rt2xD+1XYyU5Cx/mGxM2DFb9ZT5eqFV3aK1C6shksKCRoFpXgeFx1G VYNphDGVo+oCMIwMR6FaFIFL6uXkU2jraVwiGVuBeBb8B9h5YjTCH5I X-Developer-Key: i=lina@asahilina.net; a=ed25519; pk=Qn8jZuOtR1m5GaiDfTrAoQ4NE1XoYVZ/wmt5YtXWFC4= X-BeenThere: dri-devel@lists.freedesktop.org X-Mailman-Version: 2.1.29 Precedence: list List-Id: Direct Rendering Infrastructure - Development List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Cc: linaro-mm-sig@lists.linaro.org, rust-for-linux@vger.kernel.org, Karol Herbst , asahi@lists.linux.dev, Asahi Lina , linux-kernel@vger.kernel.org, dri-devel@lists.freedesktop.org, Mary , Alyssa Rosenzweig , linux-sgx@vger.kernel.org, Ella Stanforth , Faith Ekstrand , linux-media@vger.kernel.org Errors-To: dri-devel-bounces@lists.freedesktop.org Sender: "dri-devel" Some hardware may require more complex resource utilization accounting than the simple job count supported by drm_sched internally. Add a can_run_job callback to allow drivers to implement more logic before deciding whether to run a GPU job. Signed-off-by: Asahi Lina --- drivers/gpu/drm/scheduler/sched_main.c | 10 ++++++++++ include/drm/gpu_scheduler.h | 8 ++++++++ 2 files changed, 18 insertions(+) diff --git a/drivers/gpu/drm/scheduler/sched_main.c b/drivers/gpu/drm/scheduler/sched_main.c index 4e6ad6e122bc..5c0add2c7546 100644 --- a/drivers/gpu/drm/scheduler/sched_main.c +++ b/drivers/gpu/drm/scheduler/sched_main.c @@ -1001,6 +1001,16 @@ static int drm_sched_main(void *param) if (!entity) continue; + if (sched->ops->can_run_job) { + sched_job = to_drm_sched_job(spsc_queue_peek(&entity->job_queue)); + if (!sched_job) { + complete_all(&entity->entity_idle); + continue; + } + if (!sched->ops->can_run_job(sched_job)) + continue; + } + sched_job = drm_sched_entity_pop_job(entity); if (!sched_job) { diff --git a/include/drm/gpu_scheduler.h b/include/drm/gpu_scheduler.h index 9db9e5e504ee..bd89ea9507b9 100644 --- a/include/drm/gpu_scheduler.h +++ b/include/drm/gpu_scheduler.h @@ -396,6 +396,14 @@ struct drm_sched_backend_ops { struct dma_fence *(*prepare_job)(struct drm_sched_job *sched_job, struct drm_sched_entity *s_entity); + /** + * @can_run_job: Called before job execution to check whether the + * hardware is free enough to run the job. This can be used to + * implement more complex hardware resource policies than the + * hw_submission limit. + */ + bool (*can_run_job)(struct drm_sched_job *sched_job); + /** * @run_job: Called to execute the job once all of the dependencies * have been resolved. This may be called multiple times, if From patchwork Tue Mar 7 14:25:36 2023 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Asahi Lina X-Patchwork-Id: 13163707 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.4.0 (2014-02-07) on aws-us-west-2-korg-lkml-1.web.codeaurora.org Received: from gabe.freedesktop.org (gabe.freedesktop.org [131.252.210.177]) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (No client certificate requested) by smtp.lore.kernel.org (Postfix) with ESMTPS id 3B9DFC678D5 for ; Tue, 7 Mar 2023 14:28:08 +0000 (UTC) Received: from gabe.freedesktop.org (localhost [127.0.0.1]) by gabe.freedesktop.org (Postfix) with ESMTP id 4033310E4B0; Tue, 7 Mar 2023 14:28:07 +0000 (UTC) Received: from mail.marcansoft.com (marcansoft.com [212.63.210.85]) by gabe.freedesktop.org (Postfix) with ESMTPS id 51E1C10E4B0 for ; Tue, 7 Mar 2023 14:28:04 +0000 (UTC) Received: from [127.0.0.1] (localhost [127.0.0.1]) (using TLSv1.3 with cipher TLS_AES_256_GCM_SHA384 (256/256 bits) key-exchange X25519 server-signature RSA-PSS (4096 bits) server-digest SHA256) (No client certificate requested) (Authenticated sender: linasend@asahilina.net) by mail.marcansoft.com (Postfix) with ESMTPSA id E6EB8426E8; Tue, 7 Mar 2023 14:27:55 +0000 (UTC) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/simple; d=asahilina.net; s=default; t=1678199282; bh=mMV3DK/LniFSMGRFdp6WL8dzfr5kUnLaVi0jQqujgO4=; h=From:Date:Subject:References:In-Reply-To:To:Cc; b=YJhY78Ggk6+W6QBEOqupDX8aw5fWn6A/G56oEd7cyWJWvL0r748JHim/QFX3ga48h YlBahQ3QVQCtfglCOCwzF6GQnjnzKV0/4hYvwC31Y0uSRr4V+b3CUQ18zAkJuPAhzS ag95heC6renT1r1t7WaaCKqFa+8tqdzSsHCfhzZdGi7T9Kwk/2BnkgTAnLfc6MW1Au QQuJCcei56DnsCxfewGchiUhK7mQhwLnKfv3ijhcbx+NBUlpUiX39s0hxPnFjbzUg0 pFeU78xSX4yDPScrSCDZRGPIo+U0IsltRbUa79WeKf7JcpsOlh5dgZBcvyy0t2tEmP 6/bMok2XoIg2Q== From: Asahi Lina Date: Tue, 07 Mar 2023 23:25:36 +0900 Subject: [PATCH RFC 11/18] drm/scheduler: Clean up jobs when the scheduler is torn down MIME-Version: 1.0 Message-Id: <20230307-rust-drm-v1-11-917ff5bc80a8@asahilina.net> References: <20230307-rust-drm-v1-0-917ff5bc80a8@asahilina.net> In-Reply-To: <20230307-rust-drm-v1-0-917ff5bc80a8@asahilina.net> To: Maarten Lankhorst , Maxime Ripard , Thomas Zimmermann , David Airlie , Daniel Vetter , Miguel Ojeda , Alex Gaynor , Wedson Almeida Filho , Boqun Feng , Gary Guo , =?utf-8?q?Bj=C3=B6rn_Roy_Baron?= , Sumit Semwal , =?utf-8?q?Christian_K=C3=B6nig?= , Luben Tuikov , Jarkko Sakkinen , Dave Hansen X-Mailer: b4 0.12.0 X-Developer-Signature: v=1; a=ed25519-sha256; t=1678199191; l=2245; i=lina@asahilina.net; s=20230221; h=from:subject:message-id; bh=mMV3DK/LniFSMGRFdp6WL8dzfr5kUnLaVi0jQqujgO4=; b=my5v1HGy930nc+3OP/Xw50JgUtnm2HGl3Rz4cVd2S5YoFUr5PVzWJZx9x3QmCcStWOHa4pvE4 +sIUSbGMdJ7CD+mn/ehaerq/gWofZP8Suo1RhZdOI9n3Y8dnSIkG167 X-Developer-Key: i=lina@asahilina.net; a=ed25519; pk=Qn8jZuOtR1m5GaiDfTrAoQ4NE1XoYVZ/wmt5YtXWFC4= X-BeenThere: dri-devel@lists.freedesktop.org X-Mailman-Version: 2.1.29 Precedence: list List-Id: Direct Rendering Infrastructure - Development List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Cc: linaro-mm-sig@lists.linaro.org, rust-for-linux@vger.kernel.org, Karol Herbst , asahi@lists.linux.dev, Asahi Lina , linux-kernel@vger.kernel.org, dri-devel@lists.freedesktop.org, Mary , Alyssa Rosenzweig , linux-sgx@vger.kernel.org, Ella Stanforth , Faith Ekstrand , linux-media@vger.kernel.org Errors-To: dri-devel-bounces@lists.freedesktop.org Sender: "dri-devel" drm_sched_fini() currently leaves any pending jobs dangling, which causes segfaults and other badness when job completion fences are signaled after the scheduler is torn down. Explicitly detach all jobs from their completion callbacks and free them. This makes it possible to write a sensible safe abstraction for drm_sched, without having to externally duplicate the tracking of in-flight jobs. This shouldn't regress any existing drivers, since calling drm_sched_fini() with any pending jobs is broken and this change should be a no-op if there are no pending jobs. Signed-off-by: Asahi Lina --- drivers/gpu/drm/scheduler/sched_main.c | 27 +++++++++++++++++++++++++-- 1 file changed, 25 insertions(+), 2 deletions(-) diff --git a/drivers/gpu/drm/scheduler/sched_main.c b/drivers/gpu/drm/scheduler/sched_main.c index 5c0add2c7546..0aab1e0aebdd 100644 --- a/drivers/gpu/drm/scheduler/sched_main.c +++ b/drivers/gpu/drm/scheduler/sched_main.c @@ -1119,10 +1119,33 @@ EXPORT_SYMBOL(drm_sched_init); void drm_sched_fini(struct drm_gpu_scheduler *sched) { struct drm_sched_entity *s_entity; + struct drm_sched_job *s_job, *tmp; int i; - if (sched->thread) - kthread_stop(sched->thread); + if (!sched->thread) + return; + + /* + * Stop the scheduler, detaching all jobs from their hardware callbacks + * and cleaning up complete jobs. + */ + drm_sched_stop(sched, NULL); + + /* + * Iterate through the pending job list and free all jobs. + * This assumes the driver has either guaranteed jobs are already stopped, or that + * otherwise it is responsible for keeping any necessary data structures for + * in-progress jobs alive even when the free_job() callback is called early (e.g. by + * putting them in its own queue or doing its own refcounting). + */ + list_for_each_entry_safe(s_job, tmp, &sched->pending_list, list) { + spin_lock(&sched->job_list_lock); + list_del_init(&s_job->list); + spin_unlock(&sched->job_list_lock); + sched->ops->free_job(s_job); + } + + kthread_stop(sched->thread); for (i = DRM_SCHED_PRIORITY_COUNT - 1; i >= DRM_SCHED_PRIORITY_MIN; i--) { struct drm_sched_rq *rq = &sched->sched_rq[i]; From patchwork Tue Mar 7 14:25:37 2023 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Asahi Lina X-Patchwork-Id: 13163708 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.4.0 (2014-02-07) on aws-us-west-2-korg-lkml-1.web.codeaurora.org Received: from gabe.freedesktop.org (gabe.freedesktop.org [131.252.210.177]) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (No client certificate requested) by smtp.lore.kernel.org (Postfix) with ESMTPS id 98D63C678D4 for ; Tue, 7 Mar 2023 14:28:15 +0000 (UTC) Received: from gabe.freedesktop.org (localhost [127.0.0.1]) by gabe.freedesktop.org (Postfix) with ESMTP id CC78510E4B5; Tue, 7 Mar 2023 14:28:14 +0000 (UTC) Received: from mail.marcansoft.com (marcansoft.com [212.63.210.85]) by gabe.freedesktop.org (Postfix) with ESMTPS id 0660C10E4B5 for ; Tue, 7 Mar 2023 14:28:12 +0000 (UTC) Received: from [127.0.0.1] (localhost [127.0.0.1]) (using TLSv1.3 with cipher TLS_AES_256_GCM_SHA384 (256/256 bits) key-exchange X25519 server-signature RSA-PSS (4096 bits) server-digest SHA256) (No client certificate requested) (Authenticated sender: linasend@asahilina.net) by mail.marcansoft.com (Postfix) with ESMTPSA id 5EAB44261B; Tue, 7 Mar 2023 14:28:03 +0000 (UTC) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/simple; d=asahilina.net; s=default; t=1678199290; bh=doeN3VcfbBkHfCBMFKL0Le01kutm2Kr1LQUm5Ywem7w=; h=From:Date:Subject:References:In-Reply-To:To:Cc; b=GYqT9JNT3L6LU++v+IEfjknM+AZUbhwj4/2tCosIWi63HlcrPjIkHAsYVH4O+HRoh /MK1Fz+pXAA8W6r/40fGfpVTaTKKc7Jo6IXfnzkk3SbGW1DPwKrmU7Bd4s/hQINSYD p/V4RdECpng1NGRC7Gsa8n1UBVwiSfagqcvn87ZDyAlrxPxXWA8m2WZD7c4YJPxvJs pBgJzqOZ63Fjk0pzr9F7uPMoz/At2X9VopIDHRyYgPG1W0B2AnWOSUuPyGjroFd94S SiWiTJPQ7imIUx0glBwHPZD1xbXzx17vvXwhGCsW/2YoBRuSvMgjiYMlhW9ISpQB/L S0pQ4/m2/oTEw== From: Asahi Lina Date: Tue, 07 Mar 2023 23:25:37 +0900 Subject: [PATCH RFC 12/18] rust: drm: sched: Add GPU scheduler abstraction MIME-Version: 1.0 Message-Id: <20230307-rust-drm-v1-12-917ff5bc80a8@asahilina.net> References: <20230307-rust-drm-v1-0-917ff5bc80a8@asahilina.net> In-Reply-To: <20230307-rust-drm-v1-0-917ff5bc80a8@asahilina.net> To: Maarten Lankhorst , Maxime Ripard , Thomas Zimmermann , David Airlie , Daniel Vetter , Miguel Ojeda , Alex Gaynor , Wedson Almeida Filho , Boqun Feng , Gary Guo , =?utf-8?q?Bj=C3=B6rn_Roy_Baron?= , Sumit Semwal , =?utf-8?q?Christian_K=C3=B6nig?= , Luben Tuikov , Jarkko Sakkinen , Dave Hansen X-Mailer: b4 0.12.0 X-Developer-Signature: v=1; a=ed25519-sha256; t=1678199191; l=14900; i=lina@asahilina.net; s=20230221; h=from:subject:message-id; bh=doeN3VcfbBkHfCBMFKL0Le01kutm2Kr1LQUm5Ywem7w=; b=v3jijRtVxCOMOLquRKYMJ9SpdkHB/Pr+MGkxG/X/94rpF3w8oSPPtB6SFTvw8sm6mScbBbVDZ G3PGDIj1/m4BKQgENwxSO1WUOpji5kpOOvLfL2+9gKEtTcQLwlr5OE8 X-Developer-Key: i=lina@asahilina.net; a=ed25519; pk=Qn8jZuOtR1m5GaiDfTrAoQ4NE1XoYVZ/wmt5YtXWFC4= X-BeenThere: dri-devel@lists.freedesktop.org X-Mailman-Version: 2.1.29 Precedence: list List-Id: Direct Rendering Infrastructure - Development List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Cc: linaro-mm-sig@lists.linaro.org, rust-for-linux@vger.kernel.org, Karol Herbst , asahi@lists.linux.dev, Asahi Lina , linux-kernel@vger.kernel.org, dri-devel@lists.freedesktop.org, Mary , Alyssa Rosenzweig , linux-sgx@vger.kernel.org, Ella Stanforth , Faith Ekstrand , linux-media@vger.kernel.org Errors-To: dri-devel-bounces@lists.freedesktop.org Sender: "dri-devel" The GPU scheduler manages scheduling GPU jobs and dependencies between them. This Rust abstraction allows Rust DRM drivers to use this functionality. Signed-off-by: Asahi Lina --- drivers/gpu/drm/Kconfig | 5 + rust/bindings/bindings_helper.h | 1 + rust/helpers.c | 6 + rust/kernel/drm/mod.rs | 2 + rust/kernel/drm/sched.rs | 358 ++++++++++++++++++++++++++++++++++++++++ 5 files changed, 372 insertions(+) diff --git a/drivers/gpu/drm/Kconfig b/drivers/gpu/drm/Kconfig index 70a983a17ac2..8b5ad6aee126 100644 --- a/drivers/gpu/drm/Kconfig +++ b/drivers/gpu/drm/Kconfig @@ -39,6 +39,11 @@ config RUST_DRM_GEM_SHMEM_HELPER depends on RUST_DRM select DRM_GEM_SHMEM_HELPER +config RUST_DRM_SCHED + bool + depends on RUST_DRM + select DRM_SCHED + config DRM_MIPI_DBI tristate depends on DRM diff --git a/rust/bindings/bindings_helper.h b/rust/bindings/bindings_helper.h index b6696011f3a4..dc01be08676e 100644 --- a/rust/bindings/bindings_helper.h +++ b/rust/bindings/bindings_helper.h @@ -13,6 +13,7 @@ #include #include #include +#include #include #include #include diff --git a/rust/helpers.c b/rust/helpers.c index 11965b1e2f4e..1b33ed602090 100644 --- a/rust/helpers.c +++ b/rust/helpers.c @@ -408,6 +408,12 @@ void rust_helper___spin_lock_init(spinlock_t *lock, const char *name, } EXPORT_SYMBOL_GPL(rust_helper___spin_lock_init); +unsigned long rust_helper_msecs_to_jiffies(const unsigned int m) +{ + return msecs_to_jiffies(m); +} +EXPORT_SYMBOL_GPL(rust_helper_msecs_to_jiffies); + #ifdef CONFIG_DMA_SHARED_BUFFER void rust_helper_dma_fence_get(struct dma_fence *fence) diff --git a/rust/kernel/drm/mod.rs b/rust/kernel/drm/mod.rs index dae98826edfd..3ddf7712aab3 100644 --- a/rust/kernel/drm/mod.rs +++ b/rust/kernel/drm/mod.rs @@ -8,4 +8,6 @@ pub mod file; pub mod gem; pub mod ioctl; pub mod mm; +#[cfg(CONFIG_RUST_DRM_SCHED)] +pub mod sched; pub mod syncobj; diff --git a/rust/kernel/drm/sched.rs b/rust/kernel/drm/sched.rs new file mode 100644 index 000000000000..a5275cc16179 --- /dev/null +++ b/rust/kernel/drm/sched.rs @@ -0,0 +1,358 @@ +// SPDX-License-Identifier: GPL-2.0 OR MIT + +//! DRM Scheduler +//! +//! C header: [`include/linux/drm/gpu_scheduler.h`](../../../../include/linux/drm/gpu_scheduler.h) + +use crate::{ + bindings, device, + dma_fence::*, + error::{to_result, Result}, + prelude::*, + sync::{Arc, UniqueArc}, +}; +use alloc::boxed::Box; +use core::marker::PhantomData; +use core::mem::MaybeUninit; +use core::ops::{Deref, DerefMut}; +use core::ptr::addr_of_mut; + +/// Scheduler status after timeout recovery +#[repr(u32)] +pub enum Status { + /// Device recovered from the timeout and can execute jobs again + Nominal = bindings::drm_gpu_sched_stat_DRM_GPU_SCHED_STAT_NOMINAL, + /// Device is no longer available + NoDevice = bindings::drm_gpu_sched_stat_DRM_GPU_SCHED_STAT_ENODEV, +} + +/// Scheduler priorities +#[repr(i32)] +pub enum Priority { + /// Low userspace priority + Min = bindings::drm_sched_priority_DRM_SCHED_PRIORITY_MIN, + /// Normal userspace priority + Normal = bindings::drm_sched_priority_DRM_SCHED_PRIORITY_NORMAL, + /// High userspace priority + High = bindings::drm_sched_priority_DRM_SCHED_PRIORITY_HIGH, + /// Kernel priority (highest) + Kernel = bindings::drm_sched_priority_DRM_SCHED_PRIORITY_KERNEL, +} + +/// Trait to be implemented by driver job objects. +pub trait JobImpl: Sized { + /// Called when the scheduler is considering scheduling this job next, to get another Fence + /// for this job to block on. Once it returns None, run() may be called. + fn prepare(_job: &mut Job) -> Option { + None // Equivalent to NULL function pointer + } + + /// Called before job execution to check whether the hardware is free enough to run the job. + /// This can be used to implement more complex hardware resource policies than the hw_submission + /// limit. + fn can_run(_job: &mut Job) -> bool { + true + } + + /// Called to execute the job once all of the dependencies have been resolved. This may be + /// called multiple times, if timed_out() has happened and drm_sched_job_recovery() decides + /// to try it again. + fn run(job: &mut Job) -> Result>; + + /// Called when a job has taken too long to execute, to trigger GPU recovery. + /// + /// This method is called in a workqueue context. + fn timed_out(job: &mut Job) -> Status; +} + +unsafe extern "C" fn prepare_job_cb( + sched_job: *mut bindings::drm_sched_job, + _s_entity: *mut bindings::drm_sched_entity, +) -> *mut bindings::dma_fence { + // SAFETY: All of our jobs are Job. + let p = crate::container_of!(sched_job, Job, job) as *mut Job; + + match T::prepare(unsafe { &mut *p }) { + None => core::ptr::null_mut(), + Some(fence) => fence.into_raw(), + } +} + +unsafe extern "C" fn run_job_cb( + sched_job: *mut bindings::drm_sched_job, +) -> *mut bindings::dma_fence { + // SAFETY: All of our jobs are Job. + let p = crate::container_of!(sched_job, Job, job) as *mut Job; + + match T::run(unsafe { &mut *p }) { + Err(e) => e.to_ptr(), + Ok(None) => core::ptr::null_mut(), + Ok(Some(fence)) => fence.into_raw(), + } +} + +unsafe extern "C" fn can_run_job_cb(sched_job: *mut bindings::drm_sched_job) -> bool { + // SAFETY: All of our jobs are Job. + let p = crate::container_of!(sched_job, Job, job) as *mut Job; + + T::can_run(unsafe { &mut *p }) +} + +unsafe extern "C" fn timedout_job_cb( + sched_job: *mut bindings::drm_sched_job, +) -> bindings::drm_gpu_sched_stat { + // SAFETY: All of our jobs are Job. + let p = crate::container_of!(sched_job, Job, job) as *mut Job; + + T::timed_out(unsafe { &mut *p }) as bindings::drm_gpu_sched_stat +} + +unsafe extern "C" fn free_job_cb(sched_job: *mut bindings::drm_sched_job) { + // SAFETY: All of our jobs are Job. + let p = crate::container_of!(sched_job, Job, job) as *mut Job; + + // Convert the job back to a Box and drop it + // SAFETY: All of our Jobs are created inside a box. + unsafe { Box::from_raw(p) }; +} + +/// A DRM scheduler job. +pub struct Job { + job: bindings::drm_sched_job, + inner: T, +} + +impl Deref for Job { + type Target = T; + + fn deref(&self) -> &Self::Target { + &self.inner + } +} + +impl DerefMut for Job { + fn deref_mut(&mut self) -> &mut Self::Target { + &mut self.inner + } +} + +impl Drop for Job { + fn drop(&mut self) { + // SAFETY: At this point the job has either been submitted and this is being called from + // `free_job_cb` above, or it hasn't and it is safe to call `drm_sched_job_cleanup`. + unsafe { bindings::drm_sched_job_cleanup(&mut self.job) }; + } +} + +/// A pending DRM scheduler job (not yet armed) +pub struct PendingJob<'a, T: JobImpl>(Box>, PhantomData<&'a T>); + +impl<'a, T: JobImpl> PendingJob<'a, T> { + /// Add a fence as a dependency to the job + pub fn add_dependency(&mut self, fence: Fence) -> Result { + to_result(unsafe { + bindings::drm_sched_job_add_dependency(&mut self.0.job, fence.into_raw()) + }) + } + + /// Arm the job to make it ready for execution + pub fn arm(mut self) -> ArmedJob<'a, T> { + unsafe { bindings::drm_sched_job_arm(&mut self.0.job) }; + ArmedJob(self.0, PhantomData) + } +} + +impl<'a, T: JobImpl> Deref for PendingJob<'a, T> { + type Target = Job; + + fn deref(&self) -> &Self::Target { + &self.0 + } +} + +impl<'a, T: JobImpl> DerefMut for PendingJob<'a, T> { + fn deref_mut(&mut self) -> &mut Self::Target { + &mut self.0 + } +} + +/// An armed DRM scheduler job (not yet submitted) +pub struct ArmedJob<'a, T: JobImpl>(Box>, PhantomData<&'a T>); + +impl<'a, T: JobImpl> ArmedJob<'a, T> { + /// Returns the job fences + pub fn fences(&self) -> JobFences<'_> { + JobFences(unsafe { &mut *self.0.job.s_fence }) + } + + /// Push the job for execution into the scheduler + pub fn push(self) { + // After this point, the job is submitted and owned by the scheduler + let ptr = match self { + ArmedJob(job, _) => Box::>::into_raw(job), + }; + + // SAFETY: We are passing in ownership of a valid Box raw pointer. + unsafe { bindings::drm_sched_entity_push_job(addr_of_mut!((*ptr).job)) }; + } +} +impl<'a, T: JobImpl> Deref for ArmedJob<'a, T> { + type Target = Job; + + fn deref(&self) -> &Self::Target { + &self.0 + } +} + +impl<'a, T: JobImpl> DerefMut for ArmedJob<'a, T> { + fn deref_mut(&mut self) -> &mut Self::Target { + &mut self.0 + } +} + +/// Reference to the bundle of fences attached to a DRM scheduler job +pub struct JobFences<'a>(&'a mut bindings::drm_sched_fence); + +impl<'a> JobFences<'a> { + /// Returns a new reference to the job scheduled fence. + pub fn scheduled(&mut self) -> Fence { + unsafe { Fence::get_raw(&mut self.0.scheduled) } + } + + /// Returns a new reference to the job finished fence. + pub fn finished(&mut self) -> Fence { + unsafe { Fence::get_raw(&mut self.0.finished) } + } +} + +struct EntityInner { + entity: bindings::drm_sched_entity, + // TODO: Allow users to share guilty flag between entities + sched: Arc>, + guilty: bindings::atomic_t, + _p: PhantomData, +} + +impl Drop for EntityInner { + fn drop(&mut self) { + // SAFETY: The EntityInner is initialized. This will cancel/free all jobs. + unsafe { bindings::drm_sched_entity_destroy(&mut self.entity) }; + } +} + +// SAFETY: TODO +unsafe impl Sync for EntityInner {} +unsafe impl Send for EntityInner {} + +/// A DRM scheduler entity. +pub struct Entity(Pin>>); + +impl Entity { + /// Create a new scheduler entity. + pub fn new(sched: &Scheduler, priority: Priority) -> Result { + let mut entity: Box>> = Box::try_new_zeroed()?; + + let mut sched_ptr = &sched.0.sched as *const _ as *mut _; + + // SAFETY: The Box is allocated above and valid. + unsafe { + bindings::drm_sched_entity_init( + addr_of_mut!((*entity.as_mut_ptr()).entity), + priority as _, + &mut sched_ptr, + 1, + addr_of_mut!((*entity.as_mut_ptr()).guilty), + ) + }; + + // SAFETY: The Box is allocated above and valid. + unsafe { addr_of_mut!((*entity.as_mut_ptr()).sched).write(sched.0.clone()) }; + + // SAFETY: entity is now initialized. + Ok(Self(Pin::from(unsafe { entity.assume_init() }))) + } + + /// Create a new job on this entity. + /// + /// The entity must outlive the pending job until it transitions into the submitted state, + /// after which the scheduler owns it. + pub fn new_job(&self, inner: T) -> Result> { + let mut job: Box>> = Box::try_new_zeroed()?; + + // SAFETY: We hold a reference to the entity (which is a valid pointer), + // and the job object was just allocated above. + to_result(unsafe { + bindings::drm_sched_job_init( + addr_of_mut!((*job.as_mut_ptr()).job), + &self.0.as_ref().get_ref().entity as *const _ as *mut _, + core::ptr::null_mut(), + ) + })?; + + // SAFETY: The Box pointer is valid, and this initializes the inner member. + unsafe { addr_of_mut!((*job.as_mut_ptr()).inner).write(inner) }; + + // SAFETY: All fields of the Job are now initialized. + Ok(PendingJob(unsafe { job.assume_init() }, PhantomData)) + } +} + +/// DRM scheduler inner data +pub struct SchedulerInner { + sched: bindings::drm_gpu_scheduler, + _p: PhantomData, +} + +impl Drop for SchedulerInner { + fn drop(&mut self) { + // SAFETY: The scheduler is valid. This assumes drm_sched_fini() will take care of + // freeing all in-progress jobs. + unsafe { bindings::drm_sched_fini(&mut self.sched) }; + } +} + +// SAFETY: TODO +unsafe impl Sync for SchedulerInner {} +unsafe impl Send for SchedulerInner {} + +/// A DRM Scheduler +pub struct Scheduler(Arc>); + +impl Scheduler { + const OPS: bindings::drm_sched_backend_ops = bindings::drm_sched_backend_ops { + prepare_job: Some(prepare_job_cb::), + can_run_job: Some(can_run_job_cb::), + run_job: Some(run_job_cb::), + timedout_job: Some(timedout_job_cb::), + free_job: Some(free_job_cb::), + }; + /// Creates a new DRM Scheduler object + // TODO: Shared timeout workqueues & scores + pub fn new( + device: &impl device::RawDevice, + hw_submission: u32, + hang_limit: u32, + timeout_ms: usize, + name: &'static CStr, + ) -> Result> { + let mut sched: UniqueArc>> = UniqueArc::try_new_uninit()?; + + // SAFETY: The drm_sched pointer is valid and pinned as it was just allocated above. + to_result(unsafe { + bindings::drm_sched_init( + addr_of_mut!((*sched.as_mut_ptr()).sched), + &Self::OPS, + hw_submission, + hang_limit, + bindings::msecs_to_jiffies(timeout_ms.try_into()?).try_into()?, + core::ptr::null_mut(), + core::ptr::null_mut(), + name.as_char_ptr(), + device.raw_device(), + ) + })?; + + // SAFETY: All fields of SchedulerInner are now initialized. + Ok(Scheduler(unsafe { sched.assume_init() }.into())) + } +} From patchwork Tue Mar 7 14:25:38 2023 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Asahi Lina X-Patchwork-Id: 13163709 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.4.0 (2014-02-07) on aws-us-west-2-korg-lkml-1.web.codeaurora.org Received: from gabe.freedesktop.org (gabe.freedesktop.org [131.252.210.177]) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (No client certificate requested) by smtp.lore.kernel.org (Postfix) with ESMTPS id 1BC97C678D4 for ; Tue, 7 Mar 2023 14:28:22 +0000 (UTC) Received: from gabe.freedesktop.org (localhost [127.0.0.1]) by gabe.freedesktop.org (Postfix) with ESMTP id 2233E10E4A1; Tue, 7 Mar 2023 14:28:21 +0000 (UTC) Received: from mail.marcansoft.com (marcansoft.com [212.63.210.85]) by gabe.freedesktop.org (Postfix) with ESMTPS id 6FC3410E4BB for ; Tue, 7 Mar 2023 14:28:19 +0000 (UTC) Received: from [127.0.0.1] (localhost [127.0.0.1]) (using TLSv1.3 with cipher TLS_AES_256_GCM_SHA384 (256/256 bits) key-exchange X25519 server-signature RSA-PSS (4096 bits) server-digest SHA256) (No client certificate requested) (Authenticated sender: linasend@asahilina.net) by mail.marcansoft.com (Postfix) with ESMTPSA id 10542426FB; Tue, 7 Mar 2023 14:28:10 +0000 (UTC) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/simple; d=asahilina.net; s=default; t=1678199298; bh=MCSNjz0XdUk8QMw6DI+sZ1Eor9e+x9M609zyvRsQaWY=; h=From:Date:Subject:References:In-Reply-To:To:Cc; b=UbEW7nheyb0tOOAygWaSPCZ+F55R6/L037rcLQv+Flnc4JXqhYMzSr0ujBWUAvUtY /Yc3Zf8Wp+e5wAkA0caTWNAvHNUYTAHI7RPinFoVJDVSncnHyhMu80eFQcet54/Qup 6KxeRC+wlz8S1quYAcncu1i8tEl6ddeRK6zpRS9ELDktUWB763a/cIoAqoqBLKd0ZY CC54HSd0IMFq7UN3me2x2MhagmLdM4mOFAEQlgdJzf10bDCALIdJEBtAh6wES8KDxP dHJOPH2J/BuKksEJPo8MtoA1aKpeNKOLxpwJGKMShbj5usWTz5hi7SwD2pYAw6OwfN n1AwDNbf6T3NA== From: Asahi Lina Date: Tue, 07 Mar 2023 23:25:38 +0900 Subject: [PATCH RFC 13/18] drm/gem: Add a flag to control whether objects can be exported MIME-Version: 1.0 Message-Id: <20230307-rust-drm-v1-13-917ff5bc80a8@asahilina.net> References: <20230307-rust-drm-v1-0-917ff5bc80a8@asahilina.net> In-Reply-To: <20230307-rust-drm-v1-0-917ff5bc80a8@asahilina.net> To: Maarten Lankhorst , Maxime Ripard , Thomas Zimmermann , David Airlie , Daniel Vetter , Miguel Ojeda , Alex Gaynor , Wedson Almeida Filho , Boqun Feng , Gary Guo , =?utf-8?q?Bj=C3=B6rn_Roy_Baron?= , Sumit Semwal , =?utf-8?q?Christian_K=C3=B6nig?= , Luben Tuikov , Jarkko Sakkinen , Dave Hansen X-Mailer: b4 0.12.0 X-Developer-Signature: v=1; a=ed25519-sha256; t=1678199191; l=1912; i=lina@asahilina.net; s=20230221; h=from:subject:message-id; bh=MCSNjz0XdUk8QMw6DI+sZ1Eor9e+x9M609zyvRsQaWY=; b=xT82aRgu2Cl7t2vQWiQlj1BNR6l2Ur2R/6RgHsZ1B0Ituy+ZDVtv/o0c0RjUcXwEGpBtoJKVX +XfQ4py0kEHAfwGzh14DmvngfQjE0HhULuaio4lqxNvMzYCWNi4IIAA X-Developer-Key: i=lina@asahilina.net; a=ed25519; pk=Qn8jZuOtR1m5GaiDfTrAoQ4NE1XoYVZ/wmt5YtXWFC4= X-BeenThere: dri-devel@lists.freedesktop.org X-Mailman-Version: 2.1.29 Precedence: list List-Id: Direct Rendering Infrastructure - Development List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Cc: linaro-mm-sig@lists.linaro.org, rust-for-linux@vger.kernel.org, Karol Herbst , asahi@lists.linux.dev, Asahi Lina , linux-kernel@vger.kernel.org, dri-devel@lists.freedesktop.org, Mary , Alyssa Rosenzweig , linux-sgx@vger.kernel.org, Ella Stanforth , Faith Ekstrand , linux-media@vger.kernel.org Errors-To: dri-devel-bounces@lists.freedesktop.org Sender: "dri-devel" Drivers may want to support driver-private objects, which cannot be shared. This allows them to share a single lock and enables other optimizations. Add an `exportable` field to drm_gem_object, which blocks PRIME export if set to false. It is initialized to true in drm_gem_private_object_init. Signed-off-by: Asahi Lina --- drivers/gpu/drm/drm_gem.c | 1 + drivers/gpu/drm/drm_prime.c | 5 +++++ include/drm/drm_gem.h | 8 ++++++++ 3 files changed, 14 insertions(+) diff --git a/drivers/gpu/drm/drm_gem.c b/drivers/gpu/drm/drm_gem.c index 7a3cb08dc942..152ad9295a8d 100644 --- a/drivers/gpu/drm/drm_gem.c +++ b/drivers/gpu/drm/drm_gem.c @@ -166,6 +166,7 @@ void drm_gem_private_object_init(struct drm_device *dev, drm_vma_node_reset(&obj->vma_node); INIT_LIST_HEAD(&obj->lru_node); + obj->exportable = true; } EXPORT_SYMBOL(drm_gem_private_object_init); diff --git a/drivers/gpu/drm/drm_prime.c b/drivers/gpu/drm/drm_prime.c index f924b8b4ab6b..9d2dd982580e 100644 --- a/drivers/gpu/drm/drm_prime.c +++ b/drivers/gpu/drm/drm_prime.c @@ -391,6 +391,11 @@ static struct dma_buf *export_and_register_object(struct drm_device *dev, return dmabuf; } + if (!obj->exportable) { + dmabuf = ERR_PTR(-EINVAL); + return dmabuf; + } + if (obj->funcs && obj->funcs->export) dmabuf = obj->funcs->export(obj, flags); else diff --git a/include/drm/drm_gem.h b/include/drm/drm_gem.h index 772a4adf5287..852dec3cf763 100644 --- a/include/drm/drm_gem.h +++ b/include/drm/drm_gem.h @@ -361,6 +361,14 @@ struct drm_gem_object { * The current LRU list that the GEM object is on. */ struct drm_gem_lru *lru; + + /** + * @exportable: + * + * Whether this GEM object can be exported via the drm_gem_object_funcs->export + * callback. Defaults to true. + */ + bool exportable; }; /** From patchwork Tue Mar 7 14:25:39 2023 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Asahi Lina X-Patchwork-Id: 13163710 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.4.0 (2014-02-07) on aws-us-west-2-korg-lkml-1.web.codeaurora.org Received: from gabe.freedesktop.org (gabe.freedesktop.org [131.252.210.177]) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (No client certificate requested) by smtp.lore.kernel.org (Postfix) with ESMTPS id AB0EAC678D5 for ; Tue, 7 Mar 2023 14:28:29 +0000 (UTC) Received: from gabe.freedesktop.org (localhost [127.0.0.1]) by gabe.freedesktop.org (Postfix) with ESMTP id D3B5F10E4BE; Tue, 7 Mar 2023 14:28:28 +0000 (UTC) Received: from mail.marcansoft.com (marcansoft.com [212.63.210.85]) by gabe.freedesktop.org (Postfix) with ESMTPS id D620D10E4BE for ; Tue, 7 Mar 2023 14:28:26 +0000 (UTC) Received: from [127.0.0.1] (localhost [127.0.0.1]) (using TLSv1.3 with cipher TLS_AES_256_GCM_SHA384 (256/256 bits) key-exchange X25519 server-signature RSA-PSS (4096 bits) server-digest SHA256) (No client certificate requested) (Authenticated sender: linasend@asahilina.net) by mail.marcansoft.com (Postfix) with ESMTPSA id 7AEE7426E8; Tue, 7 Mar 2023 14:28:18 +0000 (UTC) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/simple; d=asahilina.net; s=default; t=1678199305; bh=PLNwTx966zByawYI9/x7zPg7nR2YAfsS2pH/0itkzOI=; h=From:Date:Subject:References:In-Reply-To:To:Cc; b=NzazbWrvDmMavYRtui5Bhfs+u+J04jGyXpZVZtoYyF3xQqMtKMYT2GZx5PvDRsm/e w3R2yDcXKlby6CeAko/nDjv33/b+TqqkjCDjjxqGZoryIAT09hbYc5lBBtC20slhoX PM4NsdCzE/6fPFfIZtI5Y8URge/uP2kj2LHqzKFieNg37kJBpwG8nP7nzj51FgWN4C /0LfQa+Re1R9H4/EB2yMqoE8dZhcitUfrwcLo44JkslEVJTwgQ14i+xnO3BdHQL0/N jIi1IOSkF6gG/2F2gFlcWSoWYUqhIkf2J8wOoWwGFMv5Qm1HHamYDJ81VZKLDWzQLY SFVO72HCjH8eA== From: Asahi Lina Date: Tue, 07 Mar 2023 23:25:39 +0900 Subject: [PATCH RFC 14/18] rust: drm: gem: Add set_exportable() method MIME-Version: 1.0 Message-Id: <20230307-rust-drm-v1-14-917ff5bc80a8@asahilina.net> References: <20230307-rust-drm-v1-0-917ff5bc80a8@asahilina.net> In-Reply-To: <20230307-rust-drm-v1-0-917ff5bc80a8@asahilina.net> To: Maarten Lankhorst , Maxime Ripard , Thomas Zimmermann , David Airlie , Daniel Vetter , Miguel Ojeda , Alex Gaynor , Wedson Almeida Filho , Boqun Feng , Gary Guo , =?utf-8?q?Bj=C3=B6rn_Roy_Baron?= , Sumit Semwal , =?utf-8?q?Christian_K=C3=B6nig?= , Luben Tuikov , Jarkko Sakkinen , Dave Hansen X-Mailer: b4 0.12.0 X-Developer-Signature: v=1; a=ed25519-sha256; t=1678199191; l=1043; i=lina@asahilina.net; s=20230221; h=from:subject:message-id; bh=PLNwTx966zByawYI9/x7zPg7nR2YAfsS2pH/0itkzOI=; b=MjFE04ydpFUIDjxSVNQLI84A5BtCxwGQPMyQgPalsSvbYqggHyxs6a9C0P0EIP/PLFM9RVKhD fHOskSTk3+WA8ZFMMr/ZHXxfKAlDqdNPAOYtZsZ2Vxzc7VvQoLpGV/a X-Developer-Key: i=lina@asahilina.net; a=ed25519; pk=Qn8jZuOtR1m5GaiDfTrAoQ4NE1XoYVZ/wmt5YtXWFC4= X-BeenThere: dri-devel@lists.freedesktop.org X-Mailman-Version: 2.1.29 Precedence: list List-Id: Direct Rendering Infrastructure - Development List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Cc: linaro-mm-sig@lists.linaro.org, rust-for-linux@vger.kernel.org, Karol Herbst , asahi@lists.linux.dev, Asahi Lina , linux-kernel@vger.kernel.org, dri-devel@lists.freedesktop.org, Mary , Alyssa Rosenzweig , linux-sgx@vger.kernel.org, Ella Stanforth , Faith Ekstrand , linux-media@vger.kernel.org Errors-To: dri-devel-bounces@lists.freedesktop.org Sender: "dri-devel" This allows drivers to control whether a given GEM object is allowed to be exported via PRIME to other drivers. --- rust/kernel/drm/gem/mod.rs | 7 +++++++ 1 file changed, 7 insertions(+) diff --git a/rust/kernel/drm/gem/mod.rs b/rust/kernel/drm/gem/mod.rs index e66bdef35c2e..196252a25b5a 100644 --- a/rust/kernel/drm/gem/mod.rs +++ b/rust/kernel/drm/gem/mod.rs @@ -135,6 +135,13 @@ pub trait BaseObject: IntoGEMObject { self.gem_ref().size } + /// Sets the exportable flag, which controls whether the object can be exported via PRIME. + fn set_exportable(&mut self, exportable: bool) { + // SAFETY: gem_obj() is valid per the type invariant, and this is safe to write if we + // are the only holder (mutable ref). + unsafe { (*self.gem_obj()).exportable = exportable }; + } + /// Creates a new reference to the object. fn reference(&self) -> ObjectRef { // SAFETY: Having a reference to an Object implies holding a GEM reference From patchwork Tue Mar 7 14:25:40 2023 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Asahi Lina X-Patchwork-Id: 13163711 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.4.0 (2014-02-07) on aws-us-west-2-korg-lkml-1.web.codeaurora.org Received: from gabe.freedesktop.org (gabe.freedesktop.org [131.252.210.177]) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (No client certificate requested) by smtp.lore.kernel.org (Postfix) with ESMTPS id 3C80FC678D4 for ; Tue, 7 Mar 2023 14:28:37 +0000 (UTC) Received: from gabe.freedesktop.org (localhost [127.0.0.1]) by gabe.freedesktop.org (Postfix) with ESMTP id 4FA4110E4BC; Tue, 7 Mar 2023 14:28:36 +0000 (UTC) Received: from mail.marcansoft.com (marcansoft.com [212.63.210.85]) by gabe.freedesktop.org (Postfix) with ESMTPS id 9177F10E4BD for ; Tue, 7 Mar 2023 14:28:34 +0000 (UTC) Received: from [127.0.0.1] (localhost [127.0.0.1]) (using TLSv1.3 with cipher TLS_AES_256_GCM_SHA384 (256/256 bits) key-exchange X25519 server-signature RSA-PSS (4096 bits) server-digest SHA256) (No client certificate requested) (Authenticated sender: linasend@asahilina.net) by mail.marcansoft.com (Postfix) with ESMTPSA id E74D142764; Tue, 7 Mar 2023 14:28:25 +0000 (UTC) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/simple; d=asahilina.net; s=default; t=1678199313; bh=Jq3knni5cItYPY4maGWHJKzv9GcNr7rANmisguhTkqw=; h=From:Date:Subject:References:In-Reply-To:To:Cc; b=VXklyo2qp1wQ5IvlmjFlkqnNs11o9Ehak0nlwhPLI4H/0SQyn++PanVy4VxEDK3aA cfAj0FDdFYUWr+gQIBmjX3UGTd5cZ0HM/Y9OCENlrXrniyWUCvE5wsS4af/9C9sTgO bAZ6fzDKPoTp25Nj6scJL+V2sZ/+gDQUSMtaI0jIwz/nA8lLH/uCE7FRn+PvI6eGMg TK3lebth6AZUGnclQO2obd+JhTjknOiv7eZnDCKuXtVc+a4B7WXIZhg0f0bjPTKN/R nZbPNWDSsYkhbMr0zTTTBjXQ9iko3DMGnltxO9hqJbomSEm43Cty9VCH2Gsu1oOMZf MwL9f/eTmf7uQ== From: Asahi Lina Date: Tue, 07 Mar 2023 23:25:40 +0900 Subject: [PATCH RFC 15/18] drm/asahi: Add the Asahi driver UAPI [DO NOT MERGE] MIME-Version: 1.0 Message-Id: <20230307-rust-drm-v1-15-917ff5bc80a8@asahilina.net> References: <20230307-rust-drm-v1-0-917ff5bc80a8@asahilina.net> In-Reply-To: <20230307-rust-drm-v1-0-917ff5bc80a8@asahilina.net> To: Maarten Lankhorst , Maxime Ripard , Thomas Zimmermann , David Airlie , Daniel Vetter , Miguel Ojeda , Alex Gaynor , Wedson Almeida Filho , Boqun Feng , Gary Guo , =?utf-8?q?Bj=C3=B6rn_Roy_Baron?= , Sumit Semwal , =?utf-8?q?Christian_K=C3=B6nig?= , Luben Tuikov , Jarkko Sakkinen , Dave Hansen X-Mailer: b4 0.12.0 X-Developer-Signature: v=1; a=ed25519-sha256; t=1678199192; l=15026; i=lina@asahilina.net; s=20230221; h=from:subject:message-id; bh=Jq3knni5cItYPY4maGWHJKzv9GcNr7rANmisguhTkqw=; b=+CPI6me7U8gNU6GFx/Kk3FLo6usIjsxOjnCG6EeHvsDNF5fLTpd+SHX7P+SqoQkQli3K+KP4h EIlAgKFpBlAApg7Ngmkw0husHyvZg1h71Ray/X2/wCrPEoIs/+bhZLv X-Developer-Key: i=lina@asahilina.net; a=ed25519; pk=Qn8jZuOtR1m5GaiDfTrAoQ4NE1XoYVZ/wmt5YtXWFC4= X-BeenThere: dri-devel@lists.freedesktop.org X-Mailman-Version: 2.1.29 Precedence: list List-Id: Direct Rendering Infrastructure - Development List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Cc: linaro-mm-sig@lists.linaro.org, rust-for-linux@vger.kernel.org, Karol Herbst , asahi@lists.linux.dev, Asahi Lina , linux-kernel@vger.kernel.org, dri-devel@lists.freedesktop.org, Mary , Alyssa Rosenzweig , linux-sgx@vger.kernel.org, Ella Stanforth , Faith Ekstrand , linux-media@vger.kernel.org Errors-To: dri-devel-bounces@lists.freedesktop.org Sender: "dri-devel" Adds the Asahi GPU driver UAPI. Note: this API is not yet stable and therefore not ready for merging! Signed-off-by: Asahi Lina --- include/uapi/drm/asahi_drm.h | 556 +++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 556 insertions(+) diff --git a/include/uapi/drm/asahi_drm.h b/include/uapi/drm/asahi_drm.h new file mode 100644 index 000000000000..7b15b486d03d --- /dev/null +++ b/include/uapi/drm/asahi_drm.h @@ -0,0 +1,556 @@ +/* SPDX-License-Identifier: MIT */ +/* + * Copyright (C) The Asahi Linux Contributors + * + * Heavily inspired by xe_drm.h. + */ +#ifndef _ASAHI_DRM_H_ +#define _ASAHI_DRM_H_ + +#include "drm.h" + +#if defined(__cplusplus) +extern "C" { +#endif + +#define DRM_ASAHI_UNSTABLE_UABI_VERSION 10006 + +#define DRM_ASAHI_GET_PARAMS 0x00 +#define DRM_ASAHI_VM_CREATE 0x01 +#define DRM_ASAHI_VM_DESTROY 0x02 +#define DRM_ASAHI_GEM_CREATE 0x03 +#define DRM_ASAHI_GEM_MMAP_OFFSET 0x04 +#define DRM_ASAHI_GEM_BIND 0x05 +#define DRM_ASAHI_QUEUE_CREATE 0x06 +#define DRM_ASAHI_QUEUE_DESTROY 0x07 +#define DRM_ASAHI_SUBMIT 0x08 +#define DRM_ASAHI_GET_TIME 0x09 + +#define DRM_ASAHI_MAX_CLUSTERS 32 + +struct drm_asahi_params_global { + __u32 unstable_uabi_version; + __u32 pad0; + + __u64 feat_compat; + __u64 feat_incompat; + + __u32 gpu_generation; + __u32 gpu_variant; + __u32 gpu_revision; + __u32 chip_id; + + __u32 num_dies; + __u32 num_clusters_total; + __u32 num_cores_per_cluster; + __u32 num_frags_per_cluster; + __u32 num_gps_per_cluster; + __u32 num_cores_total_active; + __u64 core_masks[DRM_ASAHI_MAX_CLUSTERS]; + + __u32 vm_page_size; + __u32 pad1; + __u64 vm_user_start; + __u64 vm_user_end; + __u64 vm_shader_start; + __u64 vm_shader_end; + + __u32 max_syncs_per_submission; + __u32 max_commands_per_submission; + __u32 max_commands_in_flight; + __u32 max_attachments; + + __u32 timer_frequency_hz; + __u32 min_frequency_khz; + __u32 max_frequency_khz; + __u32 max_power_mw; + + __u32 result_render_size; + __u32 result_compute_size; +}; + +/* +enum drm_asahi_feat_compat { +}; +*/ + +enum drm_asahi_feat_incompat { + DRM_ASAHI_FEAT_MANDATORY_ZS_COMPRESSION = (1UL) << 0, +}; + +struct drm_asahi_get_params { + /** @extensions: Pointer to the first extension struct, if any */ + __u64 extensions; + + /** @param: Parameter group to fetch (MBZ) */ + __u32 param_group; + + /** @pad: MBZ */ + __u32 pad; + + /** @value: User pointer to write parameter struct */ + __u64 pointer; + + /** @value: Size of user buffer, max size supported on return */ + __u64 size; +}; + +struct drm_asahi_vm_create { + /** @extensions: Pointer to the first extension struct, if any */ + __u64 extensions; + + /** @value: Returned VM ID */ + __u32 vm_id; + + /** @pad: MBZ */ + __u32 pad; +}; + +struct drm_asahi_vm_destroy { + /** @extensions: Pointer to the first extension struct, if any */ + __u64 extensions; + + /** @value: VM ID to be destroyed */ + __u32 vm_id; + + /** @pad: MBZ */ + __u32 pad; +}; + +#define ASAHI_GEM_WRITEBACK (1L << 0) +#define ASAHI_GEM_VM_PRIVATE (1L << 1) + +struct drm_asahi_gem_create { + /** @extensions: Pointer to the first extension struct, if any */ + __u64 extensions; + + /** @size: Size of the BO */ + __u64 size; + + /** @flags: BO creation flags */ + __u32 flags; + + /** @handle: VM ID to assign to the BO, if ASAHI_GEM_VM_PRIVATE is set. */ + __u32 vm_id; + + /** @handle: Returned GEM handle for the BO */ + __u32 handle; +}; + +struct drm_asahi_gem_mmap_offset { + /** @extensions: Pointer to the first extension struct, if any */ + __u64 extensions; + + /** @handle: Handle for the object being mapped. */ + __u32 handle; + + /** @flags: Must be zero */ + __u32 flags; + + /** @offset: The fake offset to use for subsequent mmap call */ + __u64 offset; +}; + +enum drm_asahi_bind_op { + ASAHI_BIND_OP_BIND = 0, + ASAHI_BIND_OP_UNBIND = 1, + ASAHI_BIND_OP_UNBIND_ALL = 2, +}; + +#define ASAHI_BIND_READ (1L << 0) +#define ASAHI_BIND_WRITE (1L << 1) + +struct drm_asahi_gem_bind { + /** @extensions: Pointer to the first extension struct, if any */ + __u64 extensions; + + /** @obj: Bind operation */ + __u32 op; + + /** @flags: One or more of ASAHI_BIND_* */ + __u32 flags; + + /** @obj: GEM object to bind */ + __u32 handle; + + /** @vm_id: The ID of the VM to bind to */ + __u32 vm_id; + + /** @offset: Offset into the object */ + __u64 offset; + + /** @range: Number of bytes from the object to bind to addr */ + __u64 range; + + /** @addr: Address to bind to */ + __u64 addr; +}; + +enum drm_asahi_cmd_type { + DRM_ASAHI_CMD_RENDER = 0, + DRM_ASAHI_CMD_BLIT = 1, + DRM_ASAHI_CMD_COMPUTE = 2, +}; + +/* Note: this is an enum so that it can be resolved by Rust bindgen. */ +enum drm_asahi_queue_cap { + DRM_ASAHI_QUEUE_CAP_RENDER = (1UL << DRM_ASAHI_CMD_RENDER), + DRM_ASAHI_QUEUE_CAP_BLIT = (1UL << DRM_ASAHI_CMD_BLIT), + DRM_ASAHI_QUEUE_CAP_COMPUTE = (1UL << DRM_ASAHI_CMD_COMPUTE), +}; + +struct drm_asahi_queue_create { + /** @extensions: Pointer to the first extension struct, if any */ + __u64 extensions; + + /** @flags: MBZ */ + __u32 flags; + + /** @vm_id: The ID of the VM this queue is bound to */ + __u32 vm_id; + + /** @type: Bitmask of DRM_ASAHI_QUEUE_CAP_* */ + __u32 queue_caps; + + /** @priority: Queue priority, 0-3 */ + __u32 priority; + + /** @queue_id: The returned queue ID */ + __u32 queue_id; +}; + +struct drm_asahi_queue_destroy { + /** @extensions: Pointer to the first extension struct, if any */ + __u64 extensions; + + /** @queue_id: The queue ID to be destroyed */ + __u32 queue_id; +}; + +enum drm_asahi_sync_type { + DRM_ASAHI_SYNC_SYNCOBJ = 0, + DRM_ASAHI_SYNC_TIMELINE_SYNCOBJ = 1, +}; + +struct drm_asahi_sync { + /** @extensions: Pointer to the first extension struct, if any */ + __u64 extensions; + + /** @sync_type: One of drm_asahi_sync_type */ + __u32 sync_type; + + /** @handle: The sync object handle */ + __u32 handle; + + /** @timeline_value: Timeline value for timeline sync objects */ + __u64 timeline_value; +}; + +enum drm_asahi_subqueue { + DRM_ASAHI_SUBQUEUE_RENDER = 0, /* Also blit */ + DRM_ASAHI_SUBQUEUE_COMPUTE = 1, + DRM_ASAHI_SUBQUEUE_COUNT = 2, +}; + +#define DRM_ASAHI_BARRIER_NONE ~(0U) + +struct drm_asahi_command { + /** @extensions: Pointer to the first extension struct, if any */ + __u64 extensions; + + /** @type: One of drm_asahi_cmd_type */ + __u32 cmd_type; + + /** @flags: Flags for command submission */ + __u32 flags; + + /** @cmdbuf: Pointer to the appropriate command buffer structure */ + __u64 cmd_buffer; + + /** @cmdbuf: Size of the command buffer structure */ + __u64 cmd_buffer_size; + + /** @cmdbuf: Offset into the result BO to return information about this command */ + __u64 result_offset; + + /** @cmdbuf: Size of the result data structure */ + __u64 result_size; + + /** @barriers: Array of command indices per subqueue to wait on */ + __u32 barriers[DRM_ASAHI_SUBQUEUE_COUNT]; +}; + +struct drm_asahi_submit { + /** @extensions: Pointer to the first extension struct, if any */ + __u64 extensions; + + /** @in_syncs: An optional array of drm_asahi_sync to wait on before starting this job. */ + __u64 in_syncs; + + /** @in_syncs: An optional array of drm_asahi_sync objects to signal upon completion. */ + __u64 out_syncs; + + /** @commands: Pointer to the drm_asahi_command array of commands to submit. */ + __u64 commands; + + /** @flags: Flags for command submission (MBZ) */ + __u32 flags; + + /** @queue_id: The queue ID to be submitted to */ + __u32 queue_id; + + /** @result_handle: An optional BO handle to place result data in */ + __u32 result_handle; + + /** @in_sync_count: Number of sync objects to wait on before starting this job. */ + __u32 in_sync_count; + + /** @in_sync_count: Number of sync objects to signal upon completion of this job. */ + __u32 out_sync_count; + + /** @pad: Number of commands to be submitted */ + __u32 command_count; +}; + +/* FIXME: This doesn't make any sense, figure out exactly what the attachment flags are */ +#define ASAHI_ATTACHMENT_C 0 +#define ASAHI_ATTACHMENT_Z 1 +#define ASAHI_ATTACHMENT_S 2 + +struct drm_asahi_attachment { + __u32 type; + __u32 size; + __u64 pointer; +}; + +#define ASAHI_RENDER_NO_CLEAR_PIPELINE_TEXTURES (1UL << 0) +#define ASAHI_RENDER_SET_WHEN_RELOADING_Z_OR_S (1UL << 1) +#define ASAHI_RENDER_MEMORYLESS_RTS_USED (1UL << 2) /* Not yet implemented */ +#define ASAHI_RENDER_PROCESS_EMPTY_TILES (1UL << 3) +#define ASAHI_RENDER_NO_VERTEX_CLUSTERING (1UL << 4) + +struct drm_asahi_cmd_render { + /** @extensions: Pointer to the first extension struct, if any */ + __u64 extensions; + + __u64 flags; + + __u64 encoder_ptr; + + __u64 attachments; + __u32 attachment_count; + __u32 pad; + + __u64 depth_buffer_1; + __u64 depth_buffer_2; + __u64 depth_buffer_3; + __u64 depth_meta_buffer_1; + __u64 depth_meta_buffer_2; + __u64 depth_meta_buffer_3; + + __u64 stencil_buffer_1; + __u64 stencil_buffer_2; + __u64 stencil_buffer_3; + __u64 stencil_meta_buffer_1; + __u64 stencil_meta_buffer_2; + __u64 stencil_meta_buffer_3; + + __u64 scissor_array; + __u64 depth_bias_array; + __u64 visibility_result_buffer; + + __u64 zls_ctrl; + __u64 ppp_multisamplectl; + __u32 ppp_ctrl; + + __u32 fb_width; + __u32 fb_height; + + __u32 utile_width; + __u32 utile_height; + + __u32 samples; + __u32 layers; + + __u32 encoder_id; + __u32 cmd_ta_id; + __u32 cmd_3d_id; + + __u32 iogpu_unk_49; + __u32 iogpu_unk_212; + __u32 iogpu_unk_214; + + __u32 merge_upper_x; + __u32 merge_upper_y; + + __u32 load_pipeline; + __u32 load_pipeline_bind; + + __u32 store_pipeline; + __u32 store_pipeline_bind; + + __u32 partial_reload_pipeline; + __u32 partial_reload_pipeline_bind; + + __u32 partial_store_pipeline; + __u32 partial_store_pipeline_bind; + + __u32 depth_dimensions; + __u32 isp_bgobjdepth; + __u32 isp_bgobjvals; +}; + +struct drm_asahi_cmd_compute { + __u64 flags; + + __u64 encoder_ptr; + __u64 encoder_end; + + __u64 attachments; + __u32 attachment_count; + __u32 pad; + + __u64 buffer_descriptor; + + __u32 buffer_descriptor_size; /* ? */ + __u32 ctx_switch_prog; + + __u32 encoder_id; + __u32 cmd_id; + + __u32 iogpu_unk_40; + __u32 iogpu_unk_44; +}; + +enum drm_asahi_status { + DRM_ASAHI_STATUS_PENDING = 0, + DRM_ASAHI_STATUS_COMPLETE, + DRM_ASAHI_STATUS_UNKNOWN_ERROR, + DRM_ASAHI_STATUS_TIMEOUT, + DRM_ASAHI_STATUS_FAULT, + DRM_ASAHI_STATUS_KILLED, + DRM_ASAHI_STATUS_NO_DEVICE, +}; + +enum drm_asahi_fault { + DRM_ASAHI_FAULT_NONE = 0, + DRM_ASAHI_FAULT_UNKNOWN, + DRM_ASAHI_FAULT_UNMAPPED, + DRM_ASAHI_FAULT_AF_FAULT, + DRM_ASAHI_FAULT_WRITE_ONLY, + DRM_ASAHI_FAULT_READ_ONLY, + DRM_ASAHI_FAULT_NO_ACCESS, +}; + +struct drm_asahi_result_info { + /** @status: One of enum drm_asahi_status */ + __u32 status; + + /** @reason: One of drm_asahi_fault_type */ + __u32 fault_type; + + /** @unit: Unit number, hardware dependent */ + __u32 unit; + + /** @sideband: Sideband information, hardware dependent */ + __u32 sideband; + + /** @level: Page table level at which the fault occurred, hardware dependent */ + __u8 level; + + /** @read: Fault was a read */ + __u8 is_read; + + /** @pad: MBZ */ + __u16 pad; + + /** @unk_5: Extra bits, hardware dependent */ + __u32 extra; + + /** @address: Fault address, cache line aligned */ + __u64 address; +}; + +#define DRM_ASAHI_RESULT_RENDER_TVB_GROW_OVF (1UL << 0) +#define DRM_ASAHI_RESULT_RENDER_TVB_GROW_MIN (1UL << 1) +#define DRM_ASAHI_RESULT_RENDER_TVB_OVERFLOWED (1UL << 2) + +struct drm_asahi_result_render { + /** @address: Common result information */ + struct drm_asahi_result_info info; + + /** @flags: Zero or more of of DRM_ASAHI_RESULT_RENDER_* */ + __u64 flags; + + /** @vertex_ts_start: Timestamp of the start of vertex processing */ + __u64 vertex_ts_start; + + /** @vertex_ts_end: Timestamp of the end of vertex processing */ + __u64 vertex_ts_end; + + /** @fragment_ts_start: Timestamp of the start of fragment processing */ + __u64 fragment_ts_start; + + /** @fragment_ts_end: Timestamp of the end of fragment processing */ + __u64 fragment_ts_end; + + /** @tvb_size_bytes: TVB size at the start of this render */ + __u64 tvb_size_bytes; + + /** @tvb_usage_bytes: Total TVB usage in bytes for this render */ + __u64 tvb_usage_bytes; + + /** @num_tvb_overflows: Number of TVB overflows that occurred for this render */ + __u32 num_tvb_overflows; +}; + +struct drm_asahi_result_compute { + /** @address: Common result information */ + struct drm_asahi_result_info info; + + /** @flags: Zero or more of of DRM_ASAHI_RESULT_COMPUTE_* */ + __u64 flags; + + /** @ts_start: Timestamp of the start of this compute command */ + __u64 ts_start; + + /** @vertex_ts_end: Timestamp of the end of this compute command */ + __u64 ts_end; +}; + +struct drm_asahi_get_time { + /** @extensions: Pointer to the first extension struct, if any */ + __u64 extensions; + + /** @flags: MBZ. */ + __u64 flags; + + /** @tv_sec: On return, seconds part of a point in time */ + __s64 tv_sec; + + /** @tv_nsec: On return, nanoseconds part of a point in time */ + __s64 tv_nsec; + + /** @gpu_timestamp: On return, the GPU timestamp at that point in time */ + __u64 gpu_timestamp; +}; + +/* Note: this is an enum so that it can be resolved by Rust bindgen. */ +enum { + DRM_IOCTL_ASAHI_GET_PARAMS = DRM_IOWR(DRM_COMMAND_BASE + DRM_ASAHI_GET_PARAMS, struct drm_asahi_get_params), + DRM_IOCTL_ASAHI_VM_CREATE = DRM_IOWR(DRM_COMMAND_BASE + DRM_ASAHI_VM_CREATE, struct drm_asahi_vm_create), + DRM_IOCTL_ASAHI_VM_DESTROY = DRM_IOW(DRM_COMMAND_BASE + DRM_ASAHI_VM_DESTROY, struct drm_asahi_vm_destroy), + DRM_IOCTL_ASAHI_GEM_CREATE = DRM_IOWR(DRM_COMMAND_BASE + DRM_ASAHI_GEM_CREATE, struct drm_asahi_gem_create), + DRM_IOCTL_ASAHI_GEM_MMAP_OFFSET = DRM_IOWR(DRM_COMMAND_BASE + DRM_ASAHI_GEM_MMAP_OFFSET, struct drm_asahi_gem_mmap_offset), + DRM_IOCTL_ASAHI_GEM_BIND = DRM_IOW(DRM_COMMAND_BASE + DRM_ASAHI_GEM_BIND, struct drm_asahi_gem_bind), + DRM_IOCTL_ASAHI_QUEUE_CREATE = DRM_IOWR(DRM_COMMAND_BASE + DRM_ASAHI_QUEUE_CREATE, struct drm_asahi_queue_create), + DRM_IOCTL_ASAHI_QUEUE_DESTROY = DRM_IOW(DRM_COMMAND_BASE + DRM_ASAHI_QUEUE_DESTROY, struct drm_asahi_queue_destroy), + DRM_IOCTL_ASAHI_SUBMIT = DRM_IOW(DRM_COMMAND_BASE + DRM_ASAHI_SUBMIT, struct drm_asahi_submit), + DRM_IOCTL_ASAHI_GET_TIME = DRM_IOWR(DRM_COMMAND_BASE + DRM_ASAHI_GET_TIME, struct drm_asahi_get_time), +}; + +#if defined(__cplusplus) +} +#endif + +#endif /* _ASAHI_DRM_H_ */ From patchwork Tue Mar 7 14:25:41 2023 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Asahi Lina X-Patchwork-Id: 13163712 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.4.0 (2014-02-07) on aws-us-west-2-korg-lkml-1.web.codeaurora.org Received: from gabe.freedesktop.org (gabe.freedesktop.org [131.252.210.177]) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (No client certificate requested) by smtp.lore.kernel.org (Postfix) with ESMTPS id CB00CC678D4 for ; Tue, 7 Mar 2023 14:28:43 +0000 (UTC) Received: from gabe.freedesktop.org (localhost [127.0.0.1]) by gabe.freedesktop.org (Postfix) with ESMTP id 29C1E10E4BD; Tue, 7 Mar 2023 14:28:43 +0000 (UTC) Received: from mail.marcansoft.com (marcansoft.com [212.63.210.85]) by gabe.freedesktop.org (Postfix) with ESMTPS id 08A6F10E4BD for ; Tue, 7 Mar 2023 14:28:42 +0000 (UTC) Received: from [127.0.0.1] (localhost [127.0.0.1]) (using TLSv1.3 with cipher TLS_AES_256_GCM_SHA384 (256/256 bits) key-exchange X25519 server-signature RSA-PSS (4096 bits) server-digest SHA256) (No client certificate requested) (Authenticated sender: linasend@asahilina.net) by mail.marcansoft.com (Postfix) with ESMTPSA id 9773D4261B; Tue, 7 Mar 2023 14:28:33 +0000 (UTC) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/simple; d=asahilina.net; s=default; t=1678199320; bh=IHBFh1p5NyHgt8gvoelp3FjaqhLt2Kb0Z3qVKYeXOdA=; h=From:Date:Subject:References:In-Reply-To:To:Cc; b=h2srMFberg2ar42BciBp2T7cZs0r5shq2P0JO3QjUsxOSin6WJjNGQt8O9BFC3Mm9 RmOByQ5GjAayiskLq3j6O537VbLpqEpJECfmqh7HVRN3loyCsyMrUheWI9qWm9s1Eo jT61rZQD7uUbslS2NbKDSA8O2Q7jgY0T7o0VSfs7OEqbf/cEUP9KBk2b2aB0UwOX5r TUgy/AjztHbk1UKMJq34coDNE/EPadHFGI118XgAMrbdBK5FIdzyiVtudxfbhPtgUv L7b8YoB/MD/NkNCpHht09obReHlsuwunHAv/PZUBC1XXY/6YkLymjrCX8tmvTyqpE6 wvZlhqdR8YD9g== From: Asahi Lina Date: Tue, 07 Mar 2023 23:25:41 +0900 Subject: [PATCH RFC 16/18] rust: bindings: Bind the Asahi DRM UAPI MIME-Version: 1.0 Message-Id: <20230307-rust-drm-v1-16-917ff5bc80a8@asahilina.net> References: <20230307-rust-drm-v1-0-917ff5bc80a8@asahilina.net> In-Reply-To: <20230307-rust-drm-v1-0-917ff5bc80a8@asahilina.net> To: Maarten Lankhorst , Maxime Ripard , Thomas Zimmermann , David Airlie , Daniel Vetter , Miguel Ojeda , Alex Gaynor , Wedson Almeida Filho , Boqun Feng , Gary Guo , =?utf-8?q?Bj=C3=B6rn_Roy_Baron?= , Sumit Semwal , =?utf-8?q?Christian_K=C3=B6nig?= , Luben Tuikov , Jarkko Sakkinen , Dave Hansen X-Mailer: b4 0.12.0 X-Developer-Signature: v=1; a=ed25519-sha256; t=1678199192; l=634; i=lina@asahilina.net; s=20230221; h=from:subject:message-id; bh=IHBFh1p5NyHgt8gvoelp3FjaqhLt2Kb0Z3qVKYeXOdA=; b=wkeUw1/4v9Ab0PfwNo+jl9En/WC72prGsQJCMySNRiL0kR7MponwPpCU1MUn2TIUBmAy343h5 5r7eQI3fZzOChdQd3fPBlNVcvU1ltpg4wdV0XFkDAdRHwRgurJDACCZ X-Developer-Key: i=lina@asahilina.net; a=ed25519; pk=Qn8jZuOtR1m5GaiDfTrAoQ4NE1XoYVZ/wmt5YtXWFC4= X-BeenThere: dri-devel@lists.freedesktop.org X-Mailman-Version: 2.1.29 Precedence: list List-Id: Direct Rendering Infrastructure - Development List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Cc: linaro-mm-sig@lists.linaro.org, rust-for-linux@vger.kernel.org, Karol Herbst , asahi@lists.linux.dev, Asahi Lina , linux-kernel@vger.kernel.org, dri-devel@lists.freedesktop.org, Mary , Alyssa Rosenzweig , linux-sgx@vger.kernel.org, Ella Stanforth , Faith Ekstrand , linux-media@vger.kernel.org Errors-To: dri-devel-bounces@lists.freedesktop.org Sender: "dri-devel" Add the Asahi UAPI to bindings_helper.h so Rust code can use it. Signed-off-by: Asahi Lina --- rust/bindings/bindings_helper.h | 1 + 1 file changed, 1 insertion(+) diff --git a/rust/bindings/bindings_helper.h b/rust/bindings/bindings_helper.h index dc01be08676e..e21c87e6d317 100644 --- a/rust/bindings/bindings_helper.h +++ b/rust/bindings/bindings_helper.h @@ -35,6 +35,7 @@ #include #include #include +#include #include /* `bindgen` gets confused at certain things. */ From patchwork Tue Mar 7 14:25:42 2023 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Asahi Lina X-Patchwork-Id: 13163726 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.4.0 (2014-02-07) on aws-us-west-2-korg-lkml-1.web.codeaurora.org Received: from gabe.freedesktop.org (gabe.freedesktop.org [131.252.210.177]) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (No client certificate requested) by smtp.lore.kernel.org (Postfix) with ESMTPS id DEF0BC678D4 for ; Tue, 7 Mar 2023 14:28:52 +0000 (UTC) Received: from gabe.freedesktop.org (localhost [127.0.0.1]) by gabe.freedesktop.org (Postfix) with ESMTP id CE0CE10E4C1; Tue, 7 Mar 2023 14:28:51 +0000 (UTC) Received: from mail.marcansoft.com (marcansoft.com [212.63.210.85]) by gabe.freedesktop.org (Postfix) with ESMTPS id 787DE10E4C1 for ; Tue, 7 Mar 2023 14:28:49 +0000 (UTC) Received: from [127.0.0.1] (localhost [127.0.0.1]) (using TLSv1.3 with cipher TLS_AES_256_GCM_SHA384 (256/256 bits) key-exchange X25519 server-signature RSA-PSS (4096 bits) server-digest SHA256) (No client certificate requested) (Authenticated sender: linasend@asahilina.net) by mail.marcansoft.com (Postfix) with ESMTPSA id 108E3426E8; Tue, 7 Mar 2023 14:28:40 +0000 (UTC) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/simple; d=asahilina.net; s=default; t=1678199328; bh=K1giKwscy2LOuMZ2AqcEY9ztp05XIoWRr3vQhLNQ3Zg=; h=From:Date:Subject:References:In-Reply-To:To:Cc; b=C8Q+7fYXmoCQcs6Av2OYMPTMZIyuSzXfTLkRhGh0NT0PjTClFUXRliaBU7pDr54pc CCRT3Ih84MQr+URhQR/da7Vkr9ohffUpQSUH8C0hElH04mbmcbhXz0otXJgZ0ASq9o brXr4tr0KGx6AtvwOvV+POQhw0i2gEXsCjAA/XoJ3omx5A4mBGFlkW5saWhQv6gXQV OfnYzJWZNqgvnXU/NPf0IOW787Qql2IlX9qU5XarYRbGgX6EDohlAfZ82itsIDS8MC UhRyv2YUrx4OoUxhS6v+tskkGEbPNFm1p6NBDmkfVOXFuJ+PuUK09FwYmrGqF/OgN+ dCEpNS3Zn/XGQ== From: Asahi Lina Date: Tue, 07 Mar 2023 23:25:42 +0900 Subject: [PATCH RFC 17/18] rust: macros: Add versions macro MIME-Version: 1.0 Message-Id: <20230307-rust-drm-v1-17-917ff5bc80a8@asahilina.net> References: <20230307-rust-drm-v1-0-917ff5bc80a8@asahilina.net> In-Reply-To: <20230307-rust-drm-v1-0-917ff5bc80a8@asahilina.net> To: Maarten Lankhorst , Maxime Ripard , Thomas Zimmermann , David Airlie , Daniel Vetter , Miguel Ojeda , Alex Gaynor , Wedson Almeida Filho , Boqun Feng , Gary Guo , =?utf-8?q?Bj=C3=B6rn_Roy_Baron?= , Sumit Semwal , =?utf-8?q?Christian_K=C3=B6nig?= , Luben Tuikov , Jarkko Sakkinen , Dave Hansen X-Mailer: b4 0.12.0 X-Developer-Signature: v=1; a=ed25519-sha256; t=1678199192; l=11529; i=lina@asahilina.net; s=20230221; h=from:subject:message-id; bh=K1giKwscy2LOuMZ2AqcEY9ztp05XIoWRr3vQhLNQ3Zg=; b=yTmTdTgnDEPP5bvMtuxBQS8eHVdb5iATCaf0YsmsBIxfZyfyeyRxQS1UEyj404IIncEWIlNcL /1Q1SDH60DgB9Omehl/rG5FxK3HalovAA8A3Q7ZfE+MaYw72aFfcD01 X-Developer-Key: i=lina@asahilina.net; a=ed25519; pk=Qn8jZuOtR1m5GaiDfTrAoQ4NE1XoYVZ/wmt5YtXWFC4= X-BeenThere: dri-devel@lists.freedesktop.org X-Mailman-Version: 2.1.29 Precedence: list List-Id: Direct Rendering Infrastructure - Development List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Cc: linaro-mm-sig@lists.linaro.org, rust-for-linux@vger.kernel.org, Karol Herbst , asahi@lists.linux.dev, Asahi Lina , linux-kernel@vger.kernel.org, dri-devel@lists.freedesktop.org, Mary , Alyssa Rosenzweig , linux-sgx@vger.kernel.org, Ella Stanforth , Faith Ekstrand , linux-media@vger.kernel.org Errors-To: dri-devel-bounces@lists.freedesktop.org Sender: "dri-devel" This macro allows Rust code to build multiple versions of the same code, conditionally including certain fields or code segments. The asahi driver uses this to support multiple GPU types and firmware revisions in the same codebase, without duplicating everything. Signed-off-by: Asahi Lina --- rust/macros/lib.rs | 7 ++ rust/macros/versions.rs | 267 ++++++++++++++++++++++++++++++++++++++++++++++++ 2 files changed, 274 insertions(+) diff --git a/rust/macros/lib.rs b/rust/macros/lib.rs index c1d385e345b9..3ab9bae4ab52 100644 --- a/rust/macros/lib.rs +++ b/rust/macros/lib.rs @@ -5,6 +5,7 @@ mod concat_idents; mod helpers; mod module; +mod versions; mod vtable; use proc_macro::TokenStream; @@ -73,6 +74,12 @@ pub fn module(ts: TokenStream) -> TokenStream { module::module(ts) } +/// Declares multiple variants of a structure or impl code +#[proc_macro_attribute] +pub fn versions(attr: TokenStream, item: TokenStream) -> TokenStream { + versions::versions(attr, item) +} + /// Declares or implements a vtable trait. /// /// Linux's use of pure vtables is very close to Rust traits, but they differ diff --git a/rust/macros/versions.rs b/rust/macros/versions.rs new file mode 100644 index 000000000000..3bcd5f557289 --- /dev/null +++ b/rust/macros/versions.rs @@ -0,0 +1,267 @@ +use proc_macro::{token_stream, Group, Ident, Punct, Spacing, Span, TokenStream, TokenTree}; + +use crate::helpers::{expect_group, expect_punct}; + +fn drop_until_punct(it: &mut impl Iterator, delimiter: &str) { + let mut depth: isize = 0; + for token in it.by_ref() { + if let TokenTree::Punct(punct) = token { + match punct.as_char() { + '<' => { + depth += 1; + } + '>' => { + depth -= 1; + } + _ => { + if depth == 0 && delimiter.contains(&punct.to_string()) { + break; + } + } + } + } + } +} + +struct VersionConfig { + fields: &'static [&'static str], + enums: &'static [&'static [&'static str]], + versions: &'static [&'static [&'static str]], +} + +static AGX_VERSIONS: VersionConfig = VersionConfig { + fields: &["G", "V"], + enums: &[&["G13", "G14"], &["V12_3", "V12_4", "V13_0B4", "V13_2"]], + versions: &[ + &["G13", "V12_3"], + &["G14", "V12_4"], + &["G13", "V13_2"], + &["G14", "V13_2"], + ], +}; + +fn check_version(config: &VersionConfig, ver: &[usize], it: &mut token_stream::IntoIter) -> bool { + let first = it.next().unwrap(); + let val: bool = match &first { + TokenTree::Group(group) => check_version(config, ver, &mut group.stream().into_iter()), + TokenTree::Ident(ident) => { + let key = config + .fields + .iter() + .position(|&r| r == ident.to_string()) + .unwrap_or_else(|| panic!("Unknown field {}", ident)); + let mut operator = expect_punct(it).to_string(); + let mut rhs_token = it.next().unwrap(); + if let TokenTree::Punct(punct) = &rhs_token { + operator.extend(std::iter::once(punct.as_char())); + rhs_token = it.next().unwrap(); + } + let rhs_name = if let TokenTree::Ident(ident) = &rhs_token { + ident.to_string() + } else { + panic!("Unexpected token {}", ident) + }; + + let rhs = config.enums[key] + .iter() + .position(|&r| r == rhs_name) + .unwrap_or_else(|| panic!("Unknown value for {}:{}", ident, rhs_name)); + let lhs = ver[key]; + + match operator.as_str() { + "==" => lhs == rhs, + "!=" => lhs != rhs, + ">" => lhs > rhs, + ">=" => lhs >= rhs, + "<" => lhs < rhs, + "<=" => lhs <= rhs, + _ => panic!("Unknown operator {}", operator), + } + } + _ => { + panic!("Unknown token {}", first) + } + }; + + let boolop = it.next(); + match boolop { + Some(TokenTree::Punct(punct)) => { + let right = expect_punct(it).to_string(); + if right != punct.to_string() { + panic!("Unexpected op {}{}", punct, right); + } + match punct.as_char() { + '&' => val && check_version(config, ver, it), + '|' => val || check_version(config, ver, it), + _ => panic!("Unexpected op {}{}", right, right), + } + } + Some(a) => panic!("Unexpected op {}", a), + None => val, + } +} + +fn filter_versions( + config: &VersionConfig, + tag: &str, + ver: &[usize], + mut it: &mut token_stream::IntoIter, + is_struct: bool, +) -> Vec { + let mut out = Vec::::new(); + + while let Some(token) = it.next() { + let mut tail: Option = None; + match &token { + TokenTree::Punct(punct) if punct.to_string() == "#" => { + let group = expect_group(it); + let mut grp_it = group.stream().into_iter(); + let attr = grp_it.next().unwrap(); + match attr { + TokenTree::Ident(ident) if ident.to_string() == "ver" => { + if check_version(config, ver, &mut grp_it) { + } else if is_struct { + drop_until_punct(&mut it, ","); + } else { + let first = it.next().unwrap(); + match &first { + TokenTree::Group(_) => (), + _ => { + drop_until_punct(&mut it, ",;"); + } + } + } + } + _ => { + out.push(token.clone()); + out.push(TokenTree::Group(group.clone())); + } + } + continue; + } + TokenTree::Punct(punct) if punct.to_string() == ":" => { + let next = it.next(); + match next { + Some(TokenTree::Punct(punct)) if punct.to_string() == ":" => { + let next = it.next(); + match next { + Some(TokenTree::Ident(idtag)) if idtag.to_string() == "ver" => { + let ident = match out.pop() { + Some(TokenTree::Ident(ident)) => ident, + a => panic!("$ver not following ident: {:?}", a), + }; + let name = ident.to_string() + tag; + let new_ident = Ident::new(name.as_str(), ident.span()); + out.push(TokenTree::Ident(new_ident)); + continue; + } + Some(a) => { + out.push(token.clone()); + out.push(token.clone()); + tail = Some(a); + } + None => { + out.push(token.clone()); + out.push(token.clone()); + } + } + } + Some(a) => { + out.push(token.clone()); + tail = Some(a); + } + None => { + out.push(token.clone()); + continue; + } + } + } + _ => { + tail = Some(token); + } + } + match &tail { + Some(TokenTree::Group(group)) => { + let new_body = + filter_versions(config, tag, ver, &mut group.stream().into_iter(), is_struct); + let mut stream = TokenStream::new(); + stream.extend(new_body); + let mut filtered_group = Group::new(group.delimiter(), stream); + filtered_group.set_span(group.span()); + out.push(TokenTree::Group(filtered_group)); + } + Some(token) => { + out.push(token.clone()); + } + None => {} + } + } + + out +} + +pub(crate) fn versions(attr: TokenStream, item: TokenStream) -> TokenStream { + let config = match attr.to_string().as_str() { + "AGX" => &AGX_VERSIONS, + _ => panic!("Unknown version group {}", attr), + }; + + let mut it = item.into_iter(); + let mut out = TokenStream::new(); + let mut body: Vec = Vec::new(); + let mut is_struct = false; + + while let Some(token) = it.next() { + match token { + TokenTree::Punct(punct) if punct.to_string() == "#" => { + body.push(TokenTree::Punct(punct)); + body.push(it.next().unwrap()); + } + TokenTree::Ident(ident) + if ["struct", "enum", "union", "const", "type"] + .contains(&ident.to_string().as_str()) => + { + is_struct = ident.to_string() != "const"; + body.push(TokenTree::Ident(ident)); + body.push(it.next().unwrap()); + // This isn't valid syntax in a struct definition, so add it for the user + body.push(TokenTree::Punct(Punct::new(':', Spacing::Joint))); + body.push(TokenTree::Punct(Punct::new(':', Spacing::Alone))); + body.push(TokenTree::Ident(Ident::new("ver", Span::call_site()))); + break; + } + TokenTree::Ident(ident) if ident.to_string() == "impl" => { + body.push(TokenTree::Ident(ident)); + break; + } + TokenTree::Ident(ident) if ident.to_string() == "fn" => { + body.push(TokenTree::Ident(ident)); + break; + } + _ => { + body.push(token); + } + } + } + + body.extend(it); + + for ver in config.versions { + let tag = ver.join(""); + let mut ver_num = Vec::::new(); + for (i, comp) in ver.iter().enumerate() { + let idx = config.enums[i].iter().position(|&r| r == *comp).unwrap(); + ver_num.push(idx); + } + let tt = TokenStream::from_iter(body.clone().into_iter()); + out.extend(filter_versions( + config, + &tag, + &ver_num, + &mut tt.into_iter(), + is_struct, + )); + } + + out +} From patchwork Tue Mar 7 14:25:43 2023 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Asahi Lina X-Patchwork-Id: 13169460 Return-Path: Received: from gabe.freedesktop.org (gabe.freedesktop.org [131.252.210.177]) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (No client certificate requested) by smtp.lore.kernel.org (Postfix) with ESMTPS id 78AB1C6FD1A for ; Tue, 7 Mar 2023 16:54:01 +0000 (UTC) Received: from gabe.freedesktop.org (localhost [127.0.0.1]) by gabe.freedesktop.org (Postfix) with ESMTP id 81C7610E11D; Tue, 7 Mar 2023 16:54:00 +0000 (UTC) Received: from mail.marcansoft.com (marcansoft.com [212.63.210.85]) by gabe.freedesktop.org (Postfix) with ESMTPS id 9EE5810E4C2 for ; Tue, 7 Mar 2023 14:28:57 +0000 (UTC) Received: from [127.0.0.1] (localhost [127.0.0.1]) (using TLSv1.3 with cipher TLS_AES_256_GCM_SHA384 (256/256 bits) key-exchange X25519 server-signature RSA-PSS (4096 bits) server-digest SHA256) (No client certificate requested) (Authenticated sender: linasend@asahilina.net) by mail.marcansoft.com (Postfix) with ESMTPSA id 834294261B; Tue, 7 Mar 2023 14:28:48 +0000 (UTC) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/simple; d=asahilina.net; s=default; t=1678199336; bh=rbyDSQ8tId0O/EQiEITiKBz/Vmt5DRUBiAX5GBvX45o=; h=From:Date:Subject:References:In-Reply-To:To:Cc; b=RVLmnV0/YMUKqABLwS+qijHF/EtRUtgOFaOwz0YOsI2V/UxoqTTyFj0zPa5Ac8aCN 6Qtye7Dq7M0d10xMnFnkwolr/3MoCOc6hh6Xll8fWOCQyedayIwD1SzexAUxlwB/WR YdGFxbtPKDqTh8DtiHMTW3oMm1zhTu4n5HQWXIvgKH5T2xTxSa87JBQ1kNdpQ+ayYI d+k1IdRzhYS12EMT6ZqU8yWLiTKmpMh555v+lD+OP4IQ9lJWCTYQK157n1BQcgoJEg YSQE4kZaxFSsJPgEq2btHlCHkmik5EgkCNSn4+500MWpJmUCRA3j1j40UJr23niSU6 /wV4bhHxrQdtA== From: Asahi Lina Date: Tue, 07 Mar 2023 23:25:43 +0900 Subject: [PATCH RFC 18/18] drm/asahi: Add the Asahi driver for Apple AGX GPUs MIME-Version: 1.0 Message-Id: <20230307-rust-drm-v1-18-917ff5bc80a8@asahilina.net> References: <20230307-rust-drm-v1-0-917ff5bc80a8@asahilina.net> In-Reply-To: <20230307-rust-drm-v1-0-917ff5bc80a8@asahilina.net> To: Maarten Lankhorst , Maxime Ripard , Thomas Zimmermann , David Airlie , Daniel Vetter , Miguel Ojeda , Alex Gaynor , Wedson Almeida Filho , Boqun Feng , Gary Guo , =?utf-8?q?Bj=C3=B6rn_Roy_Baron?= , Sumit Semwal , =?utf-8?q?Christian_K=C3=B6nig?= , Luben Tuikov , Jarkko Sakkinen , Dave Hansen X-Mailer: b4 0.12.0 X-Developer-Signature: v=1; a=ed25519-sha256; t=1678199192; l=611461; i=lina@asahilina.net; s=20230221; h=from:subject:message-id; bh=rbyDSQ8tId0O/EQiEITiKBz/Vmt5DRUBiAX5GBvX45o=; b=+xTqD/ms7jYv7yA8bjhmT16anAh/cJSMn2I24IUCv0UasgldkyBxEbRT5Jvl3CIn2L1yr2Sft CwLB1J5sbjzAEDusbQ+JeU7qErlOS4FosSX/D8jiEBtfDssy5yNWkKN X-Developer-Key: i=lina@asahilina.net; a=ed25519; pk=Qn8jZuOtR1m5GaiDfTrAoQ4NE1XoYVZ/wmt5YtXWFC4= X-Mailman-Approved-At: Tue, 07 Mar 2023 16:53:59 +0000 X-BeenThere: dri-devel@lists.freedesktop.org X-Mailman-Version: 2.1.29 Precedence: list List-Id: Direct Rendering Infrastructure - Development List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Cc: linaro-mm-sig@lists.linaro.org, rust-for-linux@vger.kernel.org, Karol Herbst , asahi@lists.linux.dev, Asahi Lina , linux-kernel@vger.kernel.org, dri-devel@lists.freedesktop.org, Mary , Alyssa Rosenzweig , linux-sgx@vger.kernel.org, Ella Stanforth , Faith Ekstrand , linux-media@vger.kernel.org Errors-To: dri-devel-bounces@lists.freedesktop.org Sender: "dri-devel" The `asahi` drm driver supports Apple AGX GPUs of the following generations: - G13G (Apple M1) - G13S (Apple M1 Pro) - G13C (Apple M1 Max) - G13D (Apple M1 Ultra) - G14G (Apple M2) Signed-off-by: Asahi Lina --- drivers/gpu/drm/Kconfig | 2 + drivers/gpu/drm/Makefile | 1 + drivers/gpu/drm/asahi/Kconfig | 35 + drivers/gpu/drm/asahi/Makefile | 3 + drivers/gpu/drm/asahi/alloc.rs | 1046 ++++++++++++++++++++++++++ drivers/gpu/drm/asahi/asahi.rs | 53 ++ drivers/gpu/drm/asahi/buffer.rs | 694 ++++++++++++++++++ drivers/gpu/drm/asahi/channel.rs | 542 ++++++++++++++ drivers/gpu/drm/asahi/debug.rs | 129 ++++ drivers/gpu/drm/asahi/driver.rs | 166 +++++ drivers/gpu/drm/asahi/event.rs | 229 ++++++ drivers/gpu/drm/asahi/file.rs | 718 ++++++++++++++++++ drivers/gpu/drm/asahi/float.rs | 381 ++++++++++ drivers/gpu/drm/asahi/fw/buffer.rs | 170 +++++ drivers/gpu/drm/asahi/fw/channels.rs | 385 ++++++++++ drivers/gpu/drm/asahi/fw/compute.rs | 107 +++ drivers/gpu/drm/asahi/fw/event.rs | 100 +++ drivers/gpu/drm/asahi/fw/fragment.rs | 276 +++++++ drivers/gpu/drm/asahi/fw/initdata.rs | 1264 ++++++++++++++++++++++++++++++++ drivers/gpu/drm/asahi/fw/job.rs | 56 ++ drivers/gpu/drm/asahi/fw/microseq.rs | 384 ++++++++++ drivers/gpu/drm/asahi/fw/mod.rs | 15 + drivers/gpu/drm/asahi/fw/types.rs | 233 ++++++ drivers/gpu/drm/asahi/fw/vertex.rs | 177 +++++ drivers/gpu/drm/asahi/fw/workqueue.rs | 168 +++++ drivers/gpu/drm/asahi/gem.rs | 301 ++++++++ drivers/gpu/drm/asahi/gpu.rs | 1088 +++++++++++++++++++++++++++ drivers/gpu/drm/asahi/hw/mod.rs | 522 +++++++++++++ drivers/gpu/drm/asahi/hw/t600x.rs | 140 ++++ drivers/gpu/drm/asahi/hw/t8103.rs | 80 ++ drivers/gpu/drm/asahi/hw/t8112.rs | 82 +++ drivers/gpu/drm/asahi/initdata.rs | 777 ++++++++++++++++++++ drivers/gpu/drm/asahi/mem.rs | 133 ++++ drivers/gpu/drm/asahi/microseq.rs | 61 ++ drivers/gpu/drm/asahi/mmu.rs | 1249 +++++++++++++++++++++++++++++++ drivers/gpu/drm/asahi/object.rs | 704 ++++++++++++++++++ drivers/gpu/drm/asahi/place.rs | 343 +++++++++ drivers/gpu/drm/asahi/queue/common.rs | 52 ++ drivers/gpu/drm/asahi/queue/compute.rs | 371 ++++++++++ drivers/gpu/drm/asahi/queue/mod.rs | 725 ++++++++++++++++++ drivers/gpu/drm/asahi/queue/render.rs | 1173 +++++++++++++++++++++++++++++ drivers/gpu/drm/asahi/regs.rs | 387 ++++++++++ drivers/gpu/drm/asahi/slotalloc.rs | 292 ++++++++ drivers/gpu/drm/asahi/util.rs | 44 ++ drivers/gpu/drm/asahi/workqueue.rs | 880 ++++++++++++++++++++++ 45 files changed, 16738 insertions(+) diff --git a/drivers/gpu/drm/Kconfig b/drivers/gpu/drm/Kconfig index 8b5ad6aee126..acdd42fc543a 100644 --- a/drivers/gpu/drm/Kconfig +++ b/drivers/gpu/drm/Kconfig @@ -384,6 +384,8 @@ source "drivers/gpu/drm/solomon/Kconfig" source "drivers/gpu/drm/sprd/Kconfig" +source "drivers/gpu/drm/asahi/Kconfig" + config DRM_HYPERV tristate "DRM Support for Hyper-V synthetic video device" depends on DRM && PCI && MMU && HYPERV diff --git a/drivers/gpu/drm/Makefile b/drivers/gpu/drm/Makefile index ab4460fcd63f..f144501d9c60 100644 --- a/drivers/gpu/drm/Makefile +++ b/drivers/gpu/drm/Makefile @@ -190,3 +190,4 @@ obj-y += gud/ obj-$(CONFIG_DRM_HYPERV) += hyperv/ obj-y += solomon/ obj-$(CONFIG_DRM_SPRD) += sprd/ +obj-$(CONFIG_DRM_ASAHI) += asahi/ diff --git a/drivers/gpu/drm/asahi/Kconfig b/drivers/gpu/drm/asahi/Kconfig new file mode 100644 index 000000000000..9d7caa49b64c --- /dev/null +++ b/drivers/gpu/drm/asahi/Kconfig @@ -0,0 +1,35 @@ +# SPDX-License-Identifier: GPL-2.0 + +config DRM_ASAHI + tristate "Asahi (DRM support for Apple AGX GPUs)" + depends on RUST + depends on RUST_DRM + depends on RUST_APPLE_RTKIT + depends on (ARM64 && ARCH_APPLE) || (COMPILE_TEST && !GENERIC_ATOMIC64) + depends on MMU + select IOMMU_SUPPORT + select IOMMU_IO_PGTABLE_LPAE + select RUST_DRM_SCHED + select RUST_DRM_GEM_SHMEM_HELPER + help + DRM driver for Apple AGX GPUs (G13x/G14). + + This driver supports the following SoCs: + + - T8103 "M1" + - T8112 "M2" + - T6000 "M1 Pro" + - T6001 "M1 Max" + - T6002 "M1 Ultra" + +config DRM_ASAHI_DEBUG_ALLOCATOR + bool "Use debug allocator" + depends on DRM_ASAHI + help + Use an alternate, simpler allocator which significantly reduces + performance, but can help find firmware- or GPU-side memory safety + issues. However, it can also trigger firmware bugs more easily, + so expect GPU crashes. + + Say N unless you are debugging firmware structures or porting to a + new firmware version. diff --git a/drivers/gpu/drm/asahi/Makefile b/drivers/gpu/drm/asahi/Makefile new file mode 100644 index 000000000000..e67248667987 --- /dev/null +++ b/drivers/gpu/drm/asahi/Makefile @@ -0,0 +1,3 @@ +# SPDX-License-Identifier: GPL-2.0 + +obj-$(CONFIG_DRM_ASAHI) += asahi.o diff --git a/drivers/gpu/drm/asahi/alloc.rs b/drivers/gpu/drm/asahi/alloc.rs new file mode 100644 index 000000000000..d918b19e9721 --- /dev/null +++ b/drivers/gpu/drm/asahi/alloc.rs @@ -0,0 +1,1046 @@ +// SPDX-License-Identifier: GPL-2.0-only OR MIT + +//! GPU kernel object allocator. +//! +//! This kernel driver needs to manage a large number of GPU objects, in both firmware/kernel +//! address space and user address space. This module implements a simple grow-only heap allocator +//! based on the DRM MM range allocator, and a debug allocator that allocates each object as a +//! separate GEM object. +//! +//! Allocations may optionally have debugging enabled, which adds preambles that store metadata +//! about the allocation. This is useful for live debugging using the hypervisor or postmortem +//! debugging with a GPU memory snapshot, since it makes it easier to identify use-after-free and +//! caching issues. + +use kernel::{c_str, drm::mm, error::Result, prelude::*, str::CString, sync::LockClassKey}; + +use crate::debug::*; +use crate::driver::AsahiDevice; +use crate::fw::types::Zeroed; +use crate::mmu; +use crate::object::{GpuArray, GpuObject, GpuOnlyArray, GpuStruct, GpuWeakPointer}; + +use core::cmp::Ordering; +use core::fmt; +use core::fmt::{Debug, Formatter}; +use core::marker::PhantomData; +use core::mem; +use core::mem::MaybeUninit; +use core::ptr::NonNull; + +const DEBUG_CLASS: DebugFlags = DebugFlags::Alloc; + +#[cfg(not(CONFIG_DRM_ASAHI_DEBUG_ALLOCATOR))] +/// The driver-global allocator type +pub(crate) type DefaultAllocator = HeapAllocator; + +#[cfg(not(CONFIG_DRM_ASAHI_DEBUG_ALLOCATOR))] +/// The driver-global allocation type +pub(crate) type DefaultAllocation = HeapAllocation; + +#[cfg(CONFIG_DRM_ASAHI_DEBUG_ALLOCATOR)] +/// The driver-global allocator type +pub(crate) type DefaultAllocator = SimpleAllocator; + +#[cfg(CONFIG_DRM_ASAHI_DEBUG_ALLOCATOR)] +/// The driver-global allocation type +pub(crate) type DefaultAllocation = SimpleAllocation; + +/// Represents a raw allocation (without any type information). +pub(crate) trait RawAllocation { + /// Returns the CPU-side pointer (if CPU mapping is enabled) as a byte non-null pointer. + fn ptr(&self) -> Option>; + /// Returns the GPU VA pointer as a u64. + fn gpu_ptr(&self) -> u64; + /// Returns the size of the allocation in bytes. + fn size(&self) -> usize; + /// Returns the AsahiDevice that owns this allocation. + fn device(&self) -> &AsahiDevice; +} + +/// Represents a typed allocation. +pub(crate) trait Allocation: Debug { + /// Returns the typed CPU-side pointer (if CPU mapping is enabled). + fn ptr(&self) -> Option>; + /// Returns the GPU VA pointer as a u64. + fn gpu_ptr(&self) -> u64; + /// Returns the size of the allocation in bytes. + fn size(&self) -> usize; + /// Returns the AsahiDevice that owns this allocation. + fn device(&self) -> &AsahiDevice; +} + +/// A generic typed allocation wrapping a RawAllocation. +/// +/// This is currently the only Allocation implementation, since it is shared by all allocators. +pub(crate) struct GenericAlloc { + alloc: U, + alloc_size: usize, + debug_offset: usize, + padding: usize, + _p: PhantomData, +} + +impl Allocation for GenericAlloc { + fn ptr(&self) -> Option> { + self.alloc + .ptr() + .map(|p| unsafe { NonNull::new_unchecked(p.as_ptr().add(self.debug_offset) as *mut T) }) + } + fn gpu_ptr(&self) -> u64 { + self.alloc.gpu_ptr() + self.debug_offset as u64 + } + fn size(&self) -> usize { + self.alloc_size + } + fn device(&self) -> &AsahiDevice { + self.alloc.device() + } +} + +impl Debug for GenericAlloc { + fn fmt(&self, f: &mut Formatter<'_>) -> fmt::Result { + f.debug_struct(core::any::type_name::>()) + .field("ptr", &format_args!("{:?}", self.ptr())) + .field("gpu_ptr", &format_args!("{:#X?}", self.gpu_ptr())) + .field("size", &format_args!("{:#X?}", self.size())) + .finish() + } +} + +/// Debugging data associated with an allocation, when debugging is enabled. +#[repr(C)] +struct AllocDebugData { + state: u32, + _pad: u32, + size: u64, + base_gpuva: u64, + obj_gpuva: u64, + name: [u8; 0x20], +} + +/// Magic flag indicating a live allocation. +const STATE_LIVE: u32 = 0x4556494c; +/// Magic flag indicating a freed allocation. +const STATE_DEAD: u32 = 0x44414544; + +/// Marker byte to identify when firmware/GPU write beyond the end of an allocation. +const GUARD_MARKER: u8 = 0x93; + +impl Drop for GenericAlloc { + fn drop(&mut self) { + let debug_len = mem::size_of::(); + if self.debug_offset >= debug_len { + if let Some(p) = self.alloc.ptr() { + unsafe { + let p = p.as_ptr().add(self.debug_offset - debug_len); + (p as *mut u32).write(STATE_DEAD); + } + } + } + if debug_enabled(DebugFlags::FillAllocations) { + if let Some(p) = self.ptr() { + unsafe { (p.as_ptr() as *mut u8).write_bytes(0xde, self.size()) }; + } + } + if self.padding != 0 { + if let Some(p) = self.ptr() { + let guard = unsafe { + core::slice::from_raw_parts( + (p.as_ptr() as *mut u8 as *const u8).add(self.size()), + self.padding, + ) + }; + if let Some(first_err) = guard.iter().position(|&r| r != GUARD_MARKER) { + let last_err = guard + .iter() + .rev() + .position(|&r| r != GUARD_MARKER) + .unwrap_or(0); + dev_warn!( + self.device(), + "Allocator: Corruption after object of type {} at {:#x}:{:#x} + {:#x}..={:#x}\n", + core::any::type_name::(), + self.gpu_ptr(), + self.size(), + first_err, + self.padding - last_err - 1 + ); + } + } + } + } +} + +static_assert!(mem::size_of::() == 0x40); + +/// A trait representing an allocator. +pub(crate) trait Allocator { + /// The raw allocation type used by this allocator. + type Raw: RawAllocation; + // TODO: Needs associated_type_defaults + // type Allocation = GenericAlloc; + + /// Returns the `AsahiDevice` associated with this allocator. + fn device(&self) -> &AsahiDevice; + /// Returns whether CPU-side mapping is enabled. + fn cpu_maps(&self) -> bool; + /// Returns the minimum alignment for allocations. + fn min_align(&self) -> usize; + /// Allocate an object of the given size in bytes with the given alignment. + fn alloc(&mut self, size: usize, align: usize) -> Result; + + /// Returns a tuple of (count, size) of how much garbage (freed but not yet reusable objects) + /// exists in this allocator. Optional. + fn garbage(&self) -> (usize, usize) { + (0, 0) + } + /// Collect garbage for this allocator, up to the given object count. Optional. + fn collect_garbage(&mut self, _count: usize) {} + + /// Allocate a new GpuStruct object. See [`GpuObject::new`]. + #[inline(never)] + fn new_object( + &mut self, + inner: T, + callback: impl for<'a> FnOnce(&'a T) -> T::Raw<'a>, + ) -> Result>> { + GpuObject::>::new(self.alloc_object()?, inner, callback) + } + + /// Allocate a new GpuStruct object. See [`GpuObject::new_boxed`]. + #[inline(never)] + fn new_boxed( + &mut self, + inner: Box, + callback: impl for<'a> FnOnce( + &'a T, + &'a mut MaybeUninit>, + ) -> Result<&'a mut T::Raw<'a>>, + ) -> Result>> { + GpuObject::>::new_boxed(self.alloc_object()?, inner, callback) + } + + /// Allocate a new GpuStruct object. See [`GpuObject::new_inplace`]. + #[inline(never)] + fn new_inplace( + &mut self, + inner: T, + callback: impl for<'a> FnOnce( + &'a T, + &'a mut MaybeUninit>, + ) -> Result<&'a mut T::Raw<'a>>, + ) -> Result>> { + GpuObject::>::new_inplace( + self.alloc_object()?, + inner, + callback, + ) + } + + /// Allocate a new GpuStruct object. See [`GpuObject::new_default`]. + #[inline(never)] + fn new_default( + &mut self, + ) -> Result>> + where + for<'a> ::Raw<'a>: Default + Zeroed, + { + GpuObject::>::new_default(self.alloc_object()?) + } + + /// Allocate a new GpuStruct object. See [`GpuObject::new_prealloc`]. + #[inline(never)] + fn new_prealloc( + &mut self, + inner_cb: impl FnOnce(GpuWeakPointer) -> Result>, + raw_cb: impl for<'a> FnOnce( + &'a T, + &'a mut MaybeUninit>, + ) -> Result<&'a mut T::Raw<'a>>, + ) -> Result>> { + GpuObject::>::new_prealloc( + self.alloc_object()?, + inner_cb, + raw_cb, + ) + } + + /// Allocate a generic buffer of the given size and alignment, applying the debug features if + /// enabled to tag it and detect overflows. + fn alloc_generic( + &mut self, + size: usize, + align: usize, + ) -> Result> { + let padding = if debug_enabled(DebugFlags::DetectOverflows) { + size + } else { + 0 + }; + + let ret: GenericAlloc = + if self.cpu_maps() && debug_enabled(debug::DebugFlags::DebugAllocations) { + let debug_align = self.min_align().max(align); + let debug_len = mem::size_of::(); + let debug_offset = (debug_len * 2 + debug_align - 1) & !(debug_align - 1); + + let alloc = self.alloc(size + debug_offset + padding, align)?; + + let mut debug = AllocDebugData { + state: STATE_LIVE, + _pad: 0, + size: size as u64, + base_gpuva: alloc.gpu_ptr(), + obj_gpuva: alloc.gpu_ptr() + debug_offset as u64, + name: [0; 0x20], + }; + + let name = core::any::type_name::().as_bytes(); + let len = name.len().min(debug.name.len() - 1); + debug.name[..len].copy_from_slice(&name[..len]); + + if let Some(p) = alloc.ptr() { + unsafe { + let p = p.as_ptr(); + p.write_bytes(0x42, debug_offset - 2 * debug_len); + let cur = p.add(debug_offset - debug_len) as *mut AllocDebugData; + let prev = p.add(debug_offset - 2 * debug_len) as *mut AllocDebugData; + prev.copy_from(cur, 1); + cur.copy_from(&debug, 1); + }; + } + + GenericAlloc { + alloc, + alloc_size: size, + debug_offset, + padding, + _p: PhantomData, + } + } else { + GenericAlloc { + alloc: self.alloc(size + padding, align)?, + alloc_size: size, + debug_offset: 0, + padding, + _p: PhantomData, + } + }; + + if debug_enabled(DebugFlags::FillAllocations) { + if let Some(p) = ret.ptr() { + unsafe { (p.as_ptr() as *mut u8).write_bytes(0xaa, ret.size()) }; + } + } + + if padding != 0 { + if let Some(p) = ret.ptr() { + unsafe { + (p.as_ptr() as *mut u8) + .add(ret.size()) + .write_bytes(GUARD_MARKER, padding); + } + } + } + + Ok(ret) + } + + /// Allocate an object of a given type, without actually initializing the allocation. + /// + /// This is useful to directly call [`GpuObject::new_*`], without borrowing a reference to the + /// allocator for the entire duration (e.g. if further allocations need to happen inside the + /// callbacks). + fn alloc_object(&mut self) -> Result> { + let size = mem::size_of::>(); + let align = mem::align_of::>(); + + self.alloc_generic(size, align) + } + + /// Allocate an empty `GpuArray` of a given type and length. + fn array_empty( + &mut self, + count: usize, + ) -> Result>> { + let size = mem::size_of::() * count; + let align = mem::align_of::(); + + let alloc = self.alloc_generic(size, align)?; + GpuArray::>::empty(alloc, count) + } + + /// Allocate an empty `GpuOnlyArray` of a given type and length. + fn array_gpuonly( + &mut self, + count: usize, + ) -> Result>> { + let size = mem::size_of::() * count; + let align = mem::align_of::(); + + let alloc = self.alloc_generic(size, align)?; + GpuOnlyArray::>::new(alloc, count) + } +} + +/// A simple allocation backed by a separate GEM object. +/// +/// # Invariants +/// `ptr` is either None or a valid, non-null pointer to the CPU view of the object. +/// `gpu_ptr` is the GPU-side VA of the object. +pub(crate) struct SimpleAllocation { + dev: AsahiDevice, + ptr: Option>, + gpu_ptr: u64, + size: usize, + vm: mmu::Vm, + obj: crate::gem::ObjectRef, +} + +/// SAFETY: `SimpleAllocation` just points to raw memory and should be safe to send across threads. +unsafe impl Send for SimpleAllocation {} +unsafe impl Sync for SimpleAllocation {} + +impl Drop for SimpleAllocation { + fn drop(&mut self) { + mod_dev_dbg!( + self.device(), + "SimpleAllocator: drop object @ {:#x}\n", + self.gpu_ptr() + ); + if debug_enabled(DebugFlags::FillAllocations) { + if let Ok(vmap) = self.obj.vmap() { + vmap.as_mut_slice().fill(0x42); + } + } + self.obj.drop_vm_mappings(self.vm.id()); + } +} + +impl RawAllocation for SimpleAllocation { + fn ptr(&self) -> Option> { + self.ptr + } + fn gpu_ptr(&self) -> u64 { + self.gpu_ptr + } + fn size(&self) -> usize { + self.size + } + + fn device(&self) -> &AsahiDevice { + &self.dev + } +} + +/// A simple allocator that allocates each object as its own GEM object, aligned to the end of a +/// page. +/// +/// This is very slow, but it has the advantage that over-reads by the firmware or GPU will fault on +/// the guard page after the allocation, which can be useful to validate that the firmware's or +/// GPU's idea of object size what we expect. +pub(crate) struct SimpleAllocator { + dev: AsahiDevice, + start: u64, + end: u64, + prot: u32, + vm: mmu::Vm, + min_align: usize, + cpu_maps: bool, +} + +impl SimpleAllocator { + /// Create a new `SimpleAllocator` for a given address range and `Vm`. + #[allow(dead_code)] + #[allow(clippy::too_many_arguments)] + pub(crate) fn new( + dev: &AsahiDevice, + vm: &mmu::Vm, + start: u64, + end: u64, + min_align: usize, + prot: u32, + _block_size: usize, + mut cpu_maps: bool, + _name: fmt::Arguments<'_>, + _keep_garbage: bool, + ) -> Result { + if debug_enabled(DebugFlags::ForceCPUMaps) { + cpu_maps = true; + } + Ok(SimpleAllocator { + dev: dev.clone(), + vm: vm.clone(), + start, + end, + prot, + min_align, + cpu_maps, + }) + } +} + +impl Allocator for SimpleAllocator { + type Raw = SimpleAllocation; + + fn device(&self) -> &AsahiDevice { + &self.dev + } + + fn cpu_maps(&self) -> bool { + self.cpu_maps + } + + fn min_align(&self) -> usize { + self.min_align + } + + #[inline(never)] + fn alloc(&mut self, size: usize, align: usize) -> Result { + let size_aligned = (size + mmu::UAT_PGSZ - 1) & !mmu::UAT_PGMSK; + let align = self.min_align.max(align); + let offset = (size_aligned - size) & !(align - 1); + + mod_dev_dbg!( + &self.dev, + "SimpleAllocator::new: size={:#x} size_al={:#x} al={:#x} off={:#x}\n", + size, + size_aligned, + align, + offset + ); + + let mut obj = crate::gem::new_kernel_object(&self.dev, size_aligned)?; + let p = obj.vmap()?.as_mut_ptr() as *mut u8; + if debug_enabled(DebugFlags::FillAllocations) { + obj.vmap()?.as_mut_slice().fill(0xde); + } + let iova = obj.map_into_range( + &self.vm, + self.start, + self.end, + self.min_align.max(mmu::UAT_PGSZ) as u64, + self.prot, + true, + )?; + + let ptr = unsafe { p.add(offset) } as *mut u8; + let gpu_ptr = (iova + offset) as u64; + + mod_dev_dbg!( + &self.dev, + "SimpleAllocator::new -> {:#?} / {:#?} | {:#x} / {:#x}\n", + p, + ptr, + iova, + gpu_ptr + ); + + Ok(SimpleAllocation { + dev: self.dev.clone(), + ptr: NonNull::new(ptr), + gpu_ptr, + size, + vm: self.vm.clone(), + obj, + }) + } +} + +/// Inner data for an allocation from the heap allocator. +/// +/// This is wrapped in an `mm::Node`. +pub(crate) struct HeapAllocationInner { + dev: AsahiDevice, + ptr: Option>, + real_size: usize, +} + +/// SAFETY: `SimpleAllocation` just points to raw memory and should be safe to send across threads. +unsafe impl Send for HeapAllocationInner {} +unsafe impl Sync for HeapAllocationInner {} + +/// Outer view of a heap allocation. +/// +/// This uses an Option<> so we can move the internal `Node` into the garbage pool when it gets +/// dropped. +/// +/// # Invariants +/// The `Option` must always be `Some(...)` while this object is alive. +pub(crate) struct HeapAllocation(Option>); + +impl Drop for HeapAllocation { + fn drop(&mut self) { + let node = self.0.take().unwrap(); + let size = node.size(); + let alloc = node.alloc_ref(); + + alloc.with(|a| { + if let Some(garbage) = a.garbage.as_mut() { + if garbage.try_push(node).is_err() { + dev_err!( + &a.dev, + "HeapAllocation[{}]::drop: Failed to keep garbage\n", + &*a.name, + ); + } + a.total_garbage += size as usize; + None + } else { + // We need to ensure node survives this scope, since dropping it + // will try to take the mm lock and deadlock us + Some(node) + } + }); + } +} + +impl mm::AllocInner for HeapAllocatorInner { + fn drop_object( + &mut self, + start: u64, + _size: u64, + _color: usize, + obj: &mut HeapAllocationInner, + ) { + /* real_size == 0 means it's a guard node */ + if obj.real_size > 0 { + mod_dev_dbg!( + obj.dev, + "HeapAllocator[{}]: drop object @ {:#x} ({} bytes)\n", + &*self.name, + start, + obj.real_size, + ); + self.allocated -= obj.real_size; + } + } +} + +impl RawAllocation for HeapAllocation { + // SAFETY: This function must always return a valid pointer. + // Since the HeapAllocation contains a reference to the + // backing_objects array that contains the object backing this pointer, + // and objects are only ever added to it, this pointer is guaranteed to + // remain valid for the lifetime of the HeapAllocation. + fn ptr(&self) -> Option> { + self.0.as_ref().unwrap().ptr + } + // SAFETY: This function must always return a valid GPU pointer. + // See the explanation in ptr(). + fn gpu_ptr(&self) -> u64 { + self.0.as_ref().unwrap().start() + } + fn size(&self) -> usize { + self.0.as_ref().unwrap().size() as usize + } + fn device(&self) -> &AsahiDevice { + &self.0.as_ref().unwrap().dev + } +} + +/// Inner data for a heap allocator which uses the DRM MM range allocator to manage the heap. +/// +/// This is wrapped by an `mm::Allocator`. +struct HeapAllocatorInner { + dev: AsahiDevice, + allocated: usize, + backing_objects: Vec<(crate::gem::ObjectRef, u64)>, + garbage: Option>>, + total_garbage: usize, + name: CString, + vm_id: u64, +} + +/// A heap allocator which uses the DRM MM range allocator to manage its objects. +/// +/// The heap is composed of a series of GEM objects. This implementation only ever grows the heap, +/// never shrinks it. +pub(crate) struct HeapAllocator { + dev: AsahiDevice, + start: u64, + end: u64, + top: u64, + prot: u32, + vm: mmu::Vm, + min_align: usize, + block_size: usize, + cpu_maps: bool, + guard_nodes: Vec>, + mm: mm::Allocator, + name: CString, +} + +static LOCK_KEY: LockClassKey = LockClassKey::new(); + +impl HeapAllocator { + /// Create a new HeapAllocator for a given `Vm` and address range. + #[allow(dead_code)] + #[allow(clippy::too_many_arguments)] + pub(crate) fn new( + dev: &AsahiDevice, + vm: &mmu::Vm, + start: u64, + end: u64, + min_align: usize, + prot: u32, + block_size: usize, + mut cpu_maps: bool, + name: fmt::Arguments<'_>, + keep_garbage: bool, + ) -> Result { + if !min_align.is_power_of_two() { + return Err(EINVAL); + } + if debug_enabled(DebugFlags::ForceCPUMaps) { + cpu_maps = true; + } + + let name = CString::try_from_fmt(name)?; + + let inner = HeapAllocatorInner { + dev: dev.clone(), + allocated: 0, + backing_objects: Vec::new(), + // TODO: This clearly needs a try_clone() or similar + name: CString::try_from_fmt(fmt!("{}", &*name))?, + vm_id: vm.id(), + garbage: if keep_garbage { Some(Vec::new()) } else { None }, + total_garbage: 0, + }; + + let mm = mm::Allocator::new( + start, + end - start + 1, + inner, + c_str!("HeapAllocator"), + &LOCK_KEY, + )?; + + Ok(HeapAllocator { + dev: dev.clone(), + vm: vm.clone(), + start, + end, + top: start, + prot, + min_align, + block_size: block_size.max(min_align), + cpu_maps, + guard_nodes: Vec::new(), + mm, + name, + }) + } + + /// Add a new backing block of the given size to this heap. + /// + /// If CPU mapping is enabled, this also adds a guard node to the range allocator to ensure that + /// objects cannot straddle backing block boundaries, since we cannot easily create a contiguous + /// CPU VA mapping for them. This can create some fragmentation. If CPU mapping is disabled, we + /// skip the guard blocks, since the GPU view of the heap is always contiguous. + fn add_block(&mut self, size: usize) -> Result { + let size_aligned = (size + mmu::UAT_PGSZ - 1) & !mmu::UAT_PGMSK; + + mod_dev_dbg!( + &self.dev, + "HeapAllocator[{}]::add_block: size={:#x} size_al={:#x}\n", + &*self.name, + size, + size_aligned, + ); + + if self.top.saturating_add(size_aligned as u64) >= self.end { + dev_err!( + &self.dev, + "HeapAllocator[{}]::add_block: Exhausted VA space\n", + &*self.name, + ); + } + + let mut obj = crate::gem::new_kernel_object(&self.dev, size_aligned)?; + if self.cpu_maps && debug_enabled(DebugFlags::FillAllocations) { + obj.vmap()?.as_mut_slice().fill(0xde); + } + + let gpu_ptr = self.top; + if let Err(e) = obj.map_at(&self.vm, gpu_ptr, self.prot, self.cpu_maps) { + dev_err!( + &self.dev, + "HeapAllocator[{}]::add_block: Failed to map at {:#x} ({:?})\n", + &*self.name, + gpu_ptr, + e + ); + return Err(e); + } + + self.mm + .with_inner(|inner| inner.backing_objects.try_reserve(1))?; + + let mut new_top = self.top + size_aligned as u64; + if self.cpu_maps { + let guard = self.min_align.max(mmu::UAT_PGSZ); + mod_dev_dbg!( + &self.dev, + "HeapAllocator[{}]::add_block: Adding guard node {:#x}:{:#x}\n", + &*self.name, + new_top, + guard + ); + + let inner = HeapAllocationInner { + dev: self.dev.clone(), + ptr: None, + real_size: 0, + }; + + let node = match self.mm.reserve_node(inner, new_top, guard as u64, 0) { + Ok(a) => a, + Err(a) => { + dev_err!( + &self.dev, + "HeapAllocator[{}]::add_block: Failed to reserve guard node {:#x}:{:#x}: {:?}\n", + &*self.name, + guard, + new_top, + a + ); + return Err(EIO); + } + }; + + self.guard_nodes.try_push(node)?; + + new_top += guard as u64; + } + mod_dev_dbg!( + &self.dev, + "HeapAllocator[{}]::add_block: top={:#x}\n", + &*self.name, + new_top + ); + + self.mm + .with_inner(|inner| inner.backing_objects.try_push((obj, gpu_ptr)))?; + + self.top = new_top; + + cls_dev_dbg!( + MemStats, + &self.dev, + "{} Heap: grow to {} bytes\n", + &*self.name, + self.top - self.start + ); + + Ok(()) + } + + /// Find the backing object index that backs a given GPU address. + fn find_obj(&mut self, addr: u64) -> Result { + self.mm.with_inner(|inner| { + inner + .backing_objects + .binary_search_by(|obj| { + let start = obj.1; + let end = obj.1 + obj.0.size() as u64; + if start > addr { + Ordering::Greater + } else if end <= addr { + Ordering::Less + } else { + Ordering::Equal + } + }) + .or(Err(ENOENT)) + }) + } +} + +impl Allocator for HeapAllocator { + type Raw = HeapAllocation; + + fn device(&self) -> &AsahiDevice { + &self.dev + } + + fn cpu_maps(&self) -> bool { + self.cpu_maps + } + + fn min_align(&self) -> usize { + self.min_align + } + + fn alloc(&mut self, size: usize, align: usize) -> Result { + if align != 0 && !align.is_power_of_two() { + return Err(EINVAL); + } + let align = self.min_align.max(align); + let size_aligned = (size + align - 1) & !(align - 1); + + mod_dev_dbg!( + &self.dev, + "HeapAllocator[{}]::new: size={:#x} size_al={:#x}\n", + &*self.name, + size, + size_aligned, + ); + + let inner = HeapAllocationInner { + dev: self.dev.clone(), + ptr: None, + real_size: size, + }; + + let mut node = match self.mm.insert_node_generic( + inner, + size_aligned as u64, + align as u64, + 0, + mm::InsertMode::Best, + ) { + Ok(a) => a, + Err(a) => { + dev_err!( + &self.dev, + "HeapAllocator[{}]::new: Failed to insert node of size {:#x} / align {:#x}: {:?}\n", + &*self.name, size_aligned, align, a + ); + return Err(a); + } + }; + + self.mm.with_inner(|inner| inner.allocated += size); + + let mut new_object = false; + let start = node.start(); + let end = start + node.size(); + if end > self.top { + if start > self.top { + dev_warn!( + self.dev, + "HeapAllocator[{}]::alloc: top={:#x}, start={:#x}\n", + &*self.name, + self.top, + start + ); + } + let block_size = self.block_size.max((end - self.top) as usize); + self.add_block(block_size)?; + new_object = true; + } + assert!(end <= self.top); + + if self.cpu_maps { + mod_dev_dbg!( + self.dev, + "HeapAllocator[{}]::alloc: mapping to CPU\n", + &*self.name + ); + + let idx = if new_object { + None + } else { + Some(match self.find_obj(start) { + Ok(a) => a, + Err(_) => { + dev_warn!( + self.dev, + "HeapAllocator[{}]::alloc: Failed to find object at {:#x}\n", + &*self.name, + start + ); + return Err(EIO); + } + }) + }; + let (obj_start, obj_size, p) = self.mm.with_inner(|inner| -> Result<_> { + let idx = idx.unwrap_or(inner.backing_objects.len() - 1); + let obj = &mut inner.backing_objects[idx]; + let p = obj.0.vmap()?.as_mut_ptr() as *mut u8; + Ok((obj.1, obj.0.size(), p)) + })?; + assert!(obj_start <= start); + assert!(obj_start + obj_size as u64 >= end); + node.as_mut().inner_mut().ptr = + NonNull::new(unsafe { p.add((start - obj_start) as usize) }); + mod_dev_dbg!( + self.dev, + "HeapAllocator[{}]::alloc: CPU pointer = {:?}\n", + &*self.name, + node.ptr + ); + } + + mod_dev_dbg!( + self.dev, + "HeapAllocator[{}]::alloc: Allocated {:#x} bytes @ {:#x}\n", + &*self.name, + end - start, + start + ); + + Ok(HeapAllocation(Some(node))) + } + + fn garbage(&self) -> (usize, usize) { + self.mm.with_inner(|inner| { + if let Some(g) = inner.garbage.as_ref() { + (g.len(), inner.total_garbage) + } else { + (0, 0) + } + }) + } + + fn collect_garbage(&mut self, count: usize) { + // Take the garbage out of the inner block, so we can safely drop it without deadlocking + let mut garbage = Vec::new(); + + if garbage.try_reserve(count).is_err() { + dev_crit!( + self.dev, + "HeapAllocator[{}]:collect_garbage: failed to reserve space\n", + &*self.name, + ); + return; + } + + self.mm.with_inner(|inner| { + if let Some(g) = inner.garbage.as_mut() { + for node in g.drain(0..count) { + inner.total_garbage -= node.size() as usize; + garbage + .try_push(node) + .expect("try_push() failed after reserve()"); + } + } + }); + } +} + +impl Drop for HeapAllocatorInner { + fn drop(&mut self) { + mod_dev_dbg!( + self.dev, + "HeapAllocator[{}]: dropping allocator\n", + &*self.name + ); + if self.allocated > 0 { + // This should never happen + dev_crit!( + self.dev, + "HeapAllocator[{}]: dropping with {} bytes allocated\n", + &*self.name, + self.allocated + ); + } else { + for mut obj in self.backing_objects.drain(..) { + obj.0.drop_vm_mappings(self.vm_id); + } + } + } +} diff --git a/drivers/gpu/drm/asahi/asahi.rs b/drivers/gpu/drm/asahi/asahi.rs new file mode 100644 index 000000000000..e511d83f4cd1 --- /dev/null +++ b/drivers/gpu/drm/asahi/asahi.rs @@ -0,0 +1,53 @@ +// SPDX-License-Identifier: GPL-2.0-only OR MIT +#![recursion_limit = "1024"] + +//! Driver for the Apple AGX GPUs found in Apple Silicon SoCs. + +mod alloc; +mod buffer; +mod channel; +mod debug; +mod driver; +mod event; +mod file; +mod float; +mod fw; +mod gem; +mod gpu; +mod hw; +mod initdata; +mod mem; +mod microseq; +mod mmu; +mod object; +mod place; +mod queue; +mod regs; +mod slotalloc; +mod util; +mod workqueue; + +use kernel::module_platform_driver; + +module_platform_driver! { + type: driver::AsahiDriver, + name: "asahi", + license: "Dual MIT/GPL", + params: { + debug_flags: u64 { + default: 0, + permissions: 0o644, + description: "Debug flags", + }, + fault_control: u32 { + default: 0, + permissions: 0, + description: "Fault control (0x0: hard faults, 0xb: macOS default)", + }, + initial_tvb_size: usize { + default: 0x8, + permissions: 0o644, + description: "Initial TVB size in blocks", + }, + }, +} diff --git a/drivers/gpu/drm/asahi/buffer.rs b/drivers/gpu/drm/asahi/buffer.rs new file mode 100644 index 000000000000..767ea161176f --- /dev/null +++ b/drivers/gpu/drm/asahi/buffer.rs @@ -0,0 +1,694 @@ +// SPDX-License-Identifier: GPL-2.0-only OR MIT + +//! Tiled Vertex Buffer management +//! +//! This module manages the Tiled Vertex Buffer, also known as the Parameter Buffer (in imgtec +//! parlance) or the tiler heap (on other architectures). This buffer holds transformed primitive +//! data between the vertex/tiling stage and the fragment stage. +//! +//! On AGX, the buffer is a heap of 128K blocks split into 32K pages (which must be aligned to a +//! multiple of 32K in VA space). The buffer can be shared between multiple render jobs, and each +//! will allocate pages from it during vertex processing and return them during fragment processing. +//! +//! If the buffer runs out of free pages, the vertex pass stops and a partial fragment pass occurs, +//! spilling the intermediate render target state to RAM (a partial render). This is all managed +//! transparently by the firmware. Since partial renders are less efficient, the kernel must grow +//! the heap in response to feedback from the firmware to avoid partial renders in the future. +//! Currently, we only ever grow the heap, and never shrink it. +//! +//! AGX also supports memoryless render targets, which can be used for intermediate results within +//! a render pass. To support partial renders, it seems the GPU/firmware has the ability to borrow +//! pages from the TVB buffer as a temporary render target buffer. Since this happens during a +//! partial render itself, if the buffer runs out of space, it requires synchronous growth in +//! response to a firmware interrupt. This is not currently supported, but may be in the future, +//! though it is unclear whether it is worth the effort. +//! +//! This module is also in charge of managing the temporary objects associated with a single render +//! pass, which includes the top-level tile array, the tail pointer cache, preemption buffers, and +//! other miscellaneous structures collectively managed as a "scene". +//! +//! To avoid runaway memory usage, there is a maximum size for buffers (at that point it's unlikely +//! that partial renders will incur much overhead over the buffer data access itself). This is +//! different depending on whether memoryless render targets are in use, and is currently hardcoded. +//! to the most common value used by macOS. + +use crate::debug::*; +use crate::fw::buffer; +use crate::fw::types::*; +use crate::util::*; +use crate::{alloc, fw, gpu, mmu, slotalloc}; +use crate::{box_in_place, place}; +use core::sync::atomic::Ordering; +use kernel::prelude::*; +use kernel::sync::{smutex::Mutex, Arc}; + +const DEBUG_CLASS: DebugFlags = DebugFlags::Buffer; + +/// There are 127 GPU/firmware-side buffer manager slots (yes, 127, not 128). +const NUM_BUFFERS: u32 = 127; + +/// Page size bits for buffer pages (32K). VAs must be aligned to this size. +pub(crate) const PAGE_SHIFT: usize = 15; +/// Page size for buffer pages. +pub(crate) const PAGE_SIZE: usize = 1 << PAGE_SHIFT; +/// Number of pages in a buffer block, which should be contiguous in VA space. +pub(crate) const PAGES_PER_BLOCK: usize = 4; +/// Size of a buffer block. +pub(crate) const BLOCK_SIZE: usize = PAGE_SIZE * PAGES_PER_BLOCK; + +/// Metadata about the tiling configuration for a scene. This is computed in the `render` module. +/// based on dimensions, tile size, and other info. +pub(crate) struct TileInfo { + /// Tile count in the X dimension. Tiles are always 32x32. + pub(crate) tiles_x: u32, + /// Tile count in the Y dimension. Tiles are always 32x32. + pub(crate) tiles_y: u32, + /// Total tile count. + pub(crate) tiles: u32, + /// Micro-tile width (16 or 32). + pub(crate) utile_width: u32, + /// Micro-tile height (16 or 32). + pub(crate) utile_height: u32, + // Macro-tiles in the X dimension. Always 4. + //pub(crate) mtiles_x: u32, + // Macro-tiles in the Y dimension. Always 4. + //pub(crate) mtiles_y: u32, + /// Tiles per macro-tile in the X dimension. + pub(crate) tiles_per_mtile_x: u32, + /// Tiles per macro-tile in the Y dimension. + pub(crate) tiles_per_mtile_y: u32, + // Total tiles per macro-tile. + //pub(crate) tiles_per_mtile: u32, + /// Micro-tiles per macro-tile in the X dimension. + pub(crate) utiles_per_mtile_x: u32, + /// Micro-tiles per macro-tile in the Y dimension. + pub(crate) utiles_per_mtile_y: u32, + // Total micro-tiles per macro-tile. + //pub(crate) utiles_per_mtile: u32, + /// Size of the top-level tilemap, in bytes (for all layers, one cluster). + pub(crate) tilemap_size: usize, + /// Size of the Tail Pointer Cache, in bytes (for all layers * clusters). + pub(crate) tpc_size: usize, + /// Number of blocks in the clustering meta buffer (for clustering). + pub(crate) meta1_blocks: u32, + /// Minimum number of TVB blocks for this render. + pub(crate) min_tvb_blocks: usize, + /// XXX: Allocation factor for cluster tilemaps and meta4. Always 2? + pub(crate) cluster_factor: usize, + /// Tiling parameter structure passed to firmware. + pub(crate) params: fw::vertex::raw::TilingParameters, +} + +/// A single scene, representing a render pass and its required buffers. +#[versions(AGX)] +#[derive(Debug)] +pub(crate) struct Scene { + object: GpuObject, + slot: u32, + rebind: bool, + preempt2_off: usize, + preempt3_off: usize, + // Note: these are dead code only on some version variants. + // It's easier to do this than to propagate the version conditionals everywhere. + #[allow(dead_code)] + meta2_off: usize, + #[allow(dead_code)] + meta3_off: usize, + #[allow(dead_code)] + meta4_off: usize, +} + +#[versions(AGX)] +impl Scene::ver { + /// Returns true if the buffer was bound to a fresh manager slot, and therefore needs an init + /// command before a render. + pub(crate) fn rebind(&self) -> bool { + self.rebind + } + + /// Returns the buffer manager slot this scene's buffer was bound to. + pub(crate) fn slot(&self) -> u32 { + self.slot + } + + /// Returns the GPU pointer to the [`buffer::Scene::ver`]. + pub(crate) fn gpu_pointer(&self) -> GpuPointer<'_, buffer::Scene::ver> { + self.object.gpu_pointer() + } + + /// Returns the GPU weak pointer to the [`buffer::Scene::ver`]. + pub(crate) fn weak_pointer(&self) -> GpuWeakPointer { + self.object.weak_pointer() + } + + /// Returns the GPU weak pointer to the kernel-side temp buffer. + /// (purpose unknown...) + pub(crate) fn kernel_buffer_pointer(&self) -> GpuWeakPointer<[u8]> { + self.object.buffer.inner.lock().kernel_buffer.weak_pointer() + } + + /// Returns the GPU pointer to the `buffer::Info::ver` object associated with this Scene. + pub(crate) fn buffer_pointer(&self) -> GpuPointer<'_, buffer::Info::ver> { + // We can't return the strong pointer directly since its lifetime crosses a lock, but we know + // its lifetime will be valid as long as &self since we hold a reference to the buffer, + // so just construct the strong pointer with the right lifetime here. + unsafe { self.weak_buffer_pointer().upgrade() } + } + + /// Returns the GPU weak pointer to the `buffer::Info::ver` object associated with this Scene. + pub(crate) fn weak_buffer_pointer(&self) -> GpuWeakPointer { + self.object.buffer.inner.lock().info.weak_pointer() + } + + /// Returns the GPU pointer to the TVB heap metadata buffer. + pub(crate) fn tvb_heapmeta_pointer(&self) -> GpuPointer<'_, &'_ [u8]> { + self.object.tvb_heapmeta.gpu_pointer() + } + + /// Returns the GPU pointer to the top-level TVB tilemap buffer. + pub(crate) fn tvb_tilemap_pointer(&self) -> GpuPointer<'_, &'_ [u8]> { + self.object.tvb_tilemap.gpu_pointer() + } + + /// Returns the GPU pointer to the Tail Pointer Cache buffer. + pub(crate) fn tpc_pointer(&self) -> GpuPointer<'_, &'_ [u8]> { + self.object.tpc.gpu_pointer() + } + + /// Returns the GPU pointer to the first preemption scratch buffer. + pub(crate) fn preempt_buf_1_pointer(&self) -> GpuPointer<'_, &'_ [u8]> { + self.object.preempt_buf.gpu_pointer() + } + + /// Returns the GPU pointer to the second preemption scratch buffer. + pub(crate) fn preempt_buf_2_pointer(&self) -> GpuPointer<'_, &'_ [u8]> { + self.object + .preempt_buf + .gpu_offset_pointer(self.preempt2_off) + } + + /// Returns the GPU pointer to the third preemption scratch buffer. + pub(crate) fn preempt_buf_3_pointer(&self) -> GpuPointer<'_, &'_ [u8]> { + self.object + .preempt_buf + .gpu_offset_pointer(self.preempt3_off) + } + + /// Returns the GPU pointer to the per-cluster tilemap buffer, if clustering is enabled. + #[allow(dead_code)] + pub(crate) fn cluster_tilemaps_pointer(&self) -> Option> { + self.object + .clustering + .as_ref() + .map(|c| c.tilemaps.gpu_pointer()) + } + + /// Returns the GPU pointer to the clustering metadata 1 buffer, if clustering is enabled. + #[allow(dead_code)] + pub(crate) fn meta_1_pointer(&self) -> Option> { + self.object + .clustering + .as_ref() + .map(|c| c.meta.gpu_pointer()) + } + + /// Returns the GPU pointer to the clustering metadata 2 buffer, if clustering is enabled. + #[allow(dead_code)] + pub(crate) fn meta_2_pointer(&self) -> Option> { + self.object + .clustering + .as_ref() + .map(|c| c.meta.gpu_offset_pointer(self.meta2_off)) + } + + /// Returns the GPU pointer to the clustering metadata 3 buffer, if clustering is enabled. + #[allow(dead_code)] + pub(crate) fn meta_3_pointer(&self) -> Option> { + self.object + .clustering + .as_ref() + .map(|c| c.meta.gpu_offset_pointer(self.meta3_off)) + } + + /// Returns the GPU pointer to the clustering metadata 4 buffer, if clustering is enabled. + #[allow(dead_code)] + pub(crate) fn meta_4_pointer(&self) -> Option> { + self.object + .clustering + .as_ref() + .map(|c| c.meta.gpu_offset_pointer(self.meta4_off)) + } + + /// Returns the GPU pointer to an unknown buffer with incrementing numbers. + pub(crate) fn seq_buf_pointer(&self) -> GpuPointer<'_, &'_ [u64]> { + self.object.seq_buf.gpu_pointer() + } + + /// Returns the number of TVB bytes used for this scene. + pub(crate) fn used_bytes(&self) -> usize { + self.object + .with(|raw, _inner| raw.total_page_count.load(Ordering::Relaxed) as usize * PAGE_SIZE) + } + + /// Returns whether the TVB overflowed while rendering this scene. + pub(crate) fn overflowed(&self) -> bool { + self.object.with(|raw, _inner| { + raw.total_page_count.load(Ordering::Relaxed) + > raw.pass_page_count.load(Ordering::Relaxed) + }) + } +} + +#[versions(AGX)] +impl Drop for Scene::ver { + fn drop(&mut self) { + let mut inner = self.object.buffer.inner.lock(); + assert_ne!(inner.active_scenes, 0); + inner.active_scenes -= 1; + + if inner.active_scenes == 0 { + mod_pr_debug!( + "Buffer: no scenes left, dropping slot {}", + inner.active_slot.take().unwrap().slot() + ); + inner.active_slot = None; + } + } +} + +/// Inner data for a single TVB buffer object. +#[versions(AGX)] +struct BufferInner { + info: GpuObject, + ualloc: Arc>, + ualloc_priv: Arc>, + blocks: Vec>, + max_blocks: usize, + max_blocks_nomemless: usize, + mgr: BufferManager, + active_scenes: usize, + active_slot: Option>, + last_token: Option, + tpc: Option>>, + kernel_buffer: GpuArray, + stats: GpuObject, + preempt1_size: usize, + preempt2_size: usize, + preempt3_size: usize, + num_clusters: usize, +} + +/// Locked and reference counted TVB buffer. +#[versions(AGX)] +pub(crate) struct Buffer { + inner: Arc>, +} + +#[versions(AGX)] +impl Buffer::ver { + /// Create a new Buffer for a given VM, given the per-VM allocators. + pub(crate) fn new( + gpu: &dyn gpu::GpuManager, + alloc: &mut gpu::KernelAllocators, + ualloc: Arc>, + ualloc_priv: Arc>, + mgr: &BufferManager, + ) -> Result { + // These are the typical max numbers on macOS. + // 8GB machines have this halved. + let max_size: usize = 862_322_688; // bytes + let max_size_nomemless = max_size / 3; + + let max_blocks = max_size / BLOCK_SIZE; + let max_blocks_nomemless = max_size_nomemless / BLOCK_SIZE; + let max_pages = max_blocks * PAGES_PER_BLOCK; + let max_pages_nomemless = max_blocks_nomemless * PAGES_PER_BLOCK; + + let num_clusters = gpu.get_dyncfg().id.num_clusters as usize; + let num_clusters_adj = if num_clusters > 1 { + num_clusters + 1 + } else { + 1 + }; + + let preempt1_size = num_clusters_adj * gpu.get_cfg().preempt1_size; + let preempt2_size = num_clusters_adj * gpu.get_cfg().preempt2_size; + let preempt3_size = num_clusters_adj * gpu.get_cfg().preempt3_size; + + let inner = box_in_place!(buffer::Info::ver { + block_ctl: alloc.shared.new_default::()?, + counter: alloc.shared.new_default::()?, + page_list: ualloc_priv.lock().array_empty(max_pages)?, + block_list: ualloc_priv.lock().array_empty(max_blocks * 2)?, + })?; + + let info = alloc.private.new_boxed(inner, |inner, ptr| { + Ok(place!( + ptr, + buffer::raw::Info::ver { + gpu_counter: 0x0, + unk_4: 0, + last_id: 0x0, + cur_id: -1, + unk_10: 0x0, + gpu_counter2: 0x0, + unk_18: 0x0, + #[ver(V < V13_0B4)] + unk_1c: 0x0, + page_list: inner.page_list.gpu_pointer(), + page_list_size: (4 * max_pages).try_into()?, + page_count: AtomicU32::new(0), + max_blocks: max_blocks.try_into()?, + block_count: AtomicU32::new(0), + unk_38: 0x0, + block_list: inner.block_list.gpu_pointer(), + block_ctl: inner.block_ctl.gpu_pointer(), + last_page: AtomicU32::new(0), + gpu_page_ptr1: 0x0, + gpu_page_ptr2: 0x0, + unk_58: 0x0, + block_size: BLOCK_SIZE as u32, + unk_60: U64(0x0), + counter: inner.counter.gpu_pointer(), + unk_70: 0x0, + unk_74: 0x0, + unk_78: 0x0, + unk_7c: 0x0, + unk_80: 0x1, + max_pages: max_pages.try_into()?, + max_pages_nomemless: max_pages_nomemless.try_into()?, + unk_8c: 0x0, + unk_90: Default::default(), + } + )) + })?; + + // Technically similar to Scene below, let's play it safe. + let kernel_buffer = alloc.shared.array_empty(0x40)?; + let stats = alloc + .shared + .new_object(Default::default(), |_inner| buffer::raw::Stats { + reset: AtomicU32::from(1), + ..Default::default() + })?; + + Ok(Buffer::ver { + inner: Arc::try_new(Mutex::new(BufferInner::ver { + info, + ualloc, + ualloc_priv, + blocks: Vec::new(), + max_blocks, + max_blocks_nomemless, + mgr: mgr.clone(), + active_scenes: 0, + active_slot: None, + last_token: None, + tpc: None, + kernel_buffer, + stats, + preempt1_size, + preempt2_size, + preempt3_size, + num_clusters, + }))?, + }) + } + + /// Returns the total block count allocated to this Buffer. + pub(crate) fn block_count(&self) -> u32 { + self.inner.lock().blocks.len() as u32 + } + + /// Returns the total size in bytes allocated to this Buffer. + pub(crate) fn size(&self) -> usize { + self.block_count() as usize * BLOCK_SIZE + } + + /// Automatically grow the Buffer based on feedback from the statistics. + pub(crate) fn auto_grow(&self) -> Result { + let inner = self.inner.lock(); + + let used_pages = inner.stats.with(|raw, _inner| { + let used = raw.max_pages.load(Ordering::Relaxed); + raw.reset.store(1, Ordering::Release); + used as usize + }); + + let need_blocks = div_ceil(used_pages * 2, PAGES_PER_BLOCK).min(inner.max_blocks_nomemless); + let want_blocks = div_ceil(used_pages * 3, PAGES_PER_BLOCK).min(inner.max_blocks_nomemless); + + let cur_count = inner.blocks.len(); + + if need_blocks <= cur_count { + Ok(false) + } else { + // Grow to 3x requested size (same logic as macOS) + core::mem::drop(inner); + self.ensure_blocks(want_blocks)?; + Ok(true) + } + } + + /// Ensure that the buffer has at least a certain minimum size in blocks. + pub(crate) fn ensure_blocks(&self, min_blocks: usize) -> Result { + let mut inner = self.inner.lock(); + + let cur_count = inner.blocks.len(); + if cur_count >= min_blocks { + return Ok(false); + } + if min_blocks > inner.max_blocks { + return Err(ENOMEM); + } + + let add_blocks = min_blocks - cur_count; + let new_count = min_blocks; + + let mut new_blocks: Vec> = Vec::new(); + + // Allocate the new blocks first, so if it fails they will be dropped + let mut ualloc = inner.ualloc.lock(); + for _i in 0..add_blocks { + new_blocks.try_push(ualloc.array_gpuonly(BLOCK_SIZE)?)?; + } + core::mem::drop(ualloc); + + // Then actually commit them + inner.blocks.try_reserve(add_blocks)?; + + for (i, block) in new_blocks.into_iter().enumerate() { + let page_num = (block.gpu_va().get() >> PAGE_SHIFT) as u32; + + inner + .blocks + .try_push(block) + .expect("try_push() failed after try_reserve()"); + inner.info.block_list[2 * (cur_count + i)] = page_num; + for j in 0..PAGES_PER_BLOCK { + inner.info.page_list[(cur_count + i) * PAGES_PER_BLOCK + j] = page_num + j as u32; + } + } + + inner.info.block_ctl.with(|raw, _inner| { + raw.total.store(new_count as u32, Ordering::SeqCst); + raw.wptr.store(new_count as u32, Ordering::SeqCst); + }); + + let page_count = (new_count * PAGES_PER_BLOCK) as u32; + inner.info.with(|raw, _inner| { + raw.page_count.store(page_count, Ordering::Relaxed); + raw.block_count.store(new_count as u32, Ordering::Relaxed); + raw.last_page.store(page_count - 1, Ordering::Relaxed); + }); + + Ok(true) + } + + /// Create a new [`Scene::ver`] (render pass) using this buffer. + pub(crate) fn new_scene( + &self, + alloc: &mut gpu::KernelAllocators, + tile_info: &TileInfo, + ) -> Result { + let mut inner = self.inner.lock(); + + let tilemap_size = tile_info.tilemap_size; + let tpc_size = tile_info.tpc_size; + + // TODO: what is this exactly? + mod_pr_debug!("Buffer: Allocating TVB buffers\n"); + + // This seems to be a list, with 4x2 bytes of headers and 8 bytes per entry. + // On single-cluster devices, the used length always seems to be 1. + // On M1 Ultra, it can grow and usually doesn't exceed 8 * cluster_factor + // entries. macOS allocates a whole 64K * 0x80 for this, so let's go with + // that to be safe... + let user_buffer = inner.ualloc.lock().array_empty(if inner.num_clusters > 1 { + 0x10080 + } else { + 0x80 + })?; + + let tvb_heapmeta = inner.ualloc.lock().array_empty(0x200)?; + let tvb_tilemap = inner.ualloc.lock().array_empty(tilemap_size)?; + + mod_pr_debug!("Buffer: Allocating misc buffers\n"); + let preempt_buf = inner + .ualloc + .lock() + .array_empty(inner.preempt1_size + inner.preempt2_size + inner.preempt3_size)?; + + let mut seq_buf = inner.ualloc.lock().array_empty(0x800)?; + for i in 1..0x400 { + seq_buf[i] = (i + 1) as u64; + } + + let tpc = match inner.tpc.as_ref() { + Some(buf) if buf.len() >= tpc_size => buf.clone(), + _ => { + // MacOS allocates this as shared GPU+FW, but + // priv seems to work and might be faster? + // Needs to be FW-writable anyway, so ualloc + // won't work. + let buf = Arc::try_new( + inner + .ualloc_priv + .lock() + .array_empty((tpc_size + mmu::UAT_PGMSK) & !mmu::UAT_PGMSK)?, + )?; + inner.tpc = Some(buf.clone()); + buf + } + }; + + // Maybe: (4x4 macro tiles + 1 global page)*n, 32bit each (17*4*n) + let meta1_size = align(tile_info.meta1_blocks as usize * 0x44, 0x80); + // check + let meta2_size = align(0x190 * inner.num_clusters, 0x80); + let meta3_size = align(0x280 * inner.num_clusters, 0x80); + // Like user_buffer for single-cluster modes, 0x30 per cluster * the cluster + // factor. + let meta4_size = align(0x30 * inner.num_clusters * tile_info.cluster_factor, 0x80); + let meta_size = meta1_size + meta2_size + meta3_size + meta4_size; + + let clustering = if inner.num_clusters > 1 { + mod_pr_debug!("Buffer: Allocating clustering buffers\n"); + let tilemaps = inner + .ualloc + .lock() + .array_empty(inner.num_clusters * tilemap_size * tile_info.cluster_factor)?; + let meta = inner.ualloc.lock().array_empty(meta_size)?; + Some(buffer::ClusterBuffers { tilemaps, meta }) + } else { + None + }; + + let scene_inner = box_in_place!(buffer::Scene::ver { + user_buffer: user_buffer, + buffer: self.clone(), + tvb_heapmeta: tvb_heapmeta, + tvb_tilemap: tvb_tilemap, + tpc: tpc, + clustering: clustering, + preempt_buf: preempt_buf, + seq_buf: seq_buf, + })?; + + // Could be made strong, but we wind up with a deadlock if we try to grab the + // pointer through the inner.buffer path inside the closure. + let stats_pointer = inner.stats.weak_pointer(); + + // macOS allocates this as private. However, the firmware does not + // DC CIVAC this before reading it (like it does most other things), + // which causes odd cache incoherency bugs when combined with + // speculation on the firmware side (maybe). This doesn't happen + // on macOS because these structs are a circular pool that is mapped + // already initialized. Just mark this shared for now. + let scene = alloc.shared.new_boxed(scene_inner, |inner, ptr| { + Ok(place!( + ptr, + buffer::raw::Scene { + pass_page_count: AtomicU32::new(0), + unk_4: 0, + unk_8: U64(0), + unk_10: U64(0), + user_buffer: inner.user_buffer.gpu_pointer(), + unk_20: 0, + stats: stats_pointer, + total_page_count: AtomicU32::new(0), + unk_30: U64(0), + unk_38: U64(0), + } + )) + })?; + + let mut rebind = false; + + if inner.active_slot.is_none() { + assert_eq!(inner.active_scenes, 0); + + let slot = inner.mgr.0.get(inner.last_token)?; + rebind = slot.changed(); + + mod_pr_debug!("Buffer: assigning slot {} (rebind={})", slot.slot(), rebind); + + inner.last_token = Some(slot.token()); + inner.active_slot = Some(slot); + } + + inner.active_scenes += 1; + + Ok(Scene::ver { + object: scene, + slot: inner.active_slot.as_ref().unwrap().slot(), + rebind, + preempt2_off: inner.preempt1_size, + preempt3_off: inner.preempt1_size + inner.preempt2_size, + meta2_off: meta1_size, + meta3_off: meta1_size + meta2_size, + meta4_off: meta1_size + meta2_size + meta3_size, + }) + } + + /// Increment the buffer manager usage count. Should we done once we know the Scene is ready + /// to be committed and used in commands submitted to the GPU. + pub(crate) fn increment(&self) { + let inner = self.inner.lock(); + inner.info.counter.with(|raw, _inner| { + // We could use fetch_add, but the non-LSE atomic + // sequence Rust produces confuses the hypervisor. + // We have inner locked anyway, so this is not racy. + let v = raw.count.load(Ordering::Relaxed); + raw.count.store(v + 1, Ordering::Relaxed); + }); + } +} + +#[versions(AGX)] +impl Clone for Buffer::ver { + fn clone(&self) -> Self { + Buffer::ver { + inner: self.inner.clone(), + } + } +} + +/// The GPU-global buffer manager, used to allocate and release buffer slots from the pool. +pub(crate) struct BufferManager(slotalloc::SlotAllocator<()>); + +impl BufferManager { + pub(crate) fn new() -> Result { + Ok(BufferManager(slotalloc::SlotAllocator::new( + NUM_BUFFERS, + (), + |_inner, _slot| (), + )?)) + } +} + +impl Clone for BufferManager { + fn clone(&self) -> Self { + BufferManager(self.0.clone()) + } +} diff --git a/drivers/gpu/drm/asahi/channel.rs b/drivers/gpu/drm/asahi/channel.rs new file mode 100644 index 000000000000..0b3c3b65c279 --- /dev/null +++ b/drivers/gpu/drm/asahi/channel.rs @@ -0,0 +1,542 @@ +// SPDX-License-Identifier: GPL-2.0-only OR MIT + +//! GPU ring buffer channels +//! +//! The GPU firmware use a set of ring buffer channels to receive commands from the driver and send +//! it notifications and status messages. +//! +//! These ring buffers mostly follow uniform conventions, so they share the same base +//! implementation. + +use crate::debug::*; +use crate::driver::AsahiDevice; +use crate::fw::channels::*; +use crate::fw::initdata::{raw, ChannelRing}; +use crate::fw::types::*; +use crate::{event, gpu, mem}; +use core::time::Duration; +use kernel::{c_str, delay::coarse_sleep, prelude::*, sync::Arc, time}; + +pub(crate) use crate::fw::channels::PipeType; + +/// A receive (FW->driver) channel. +pub(crate) struct RxChannel +where + for<'a> ::Raw<'a>: Debug + Default + Zeroed, +{ + ring: ChannelRing, + // FIXME: needs feature(generic_const_exprs) + //rptr: [u32; T::SUB_CHANNELS], + rptr: [u32; 6], + count: u32, +} + +impl RxChannel +where + for<'a> ::Raw<'a>: Debug + Default + Zeroed, +{ + /// Allocates a new receive channel with a given message count. + pub(crate) fn new(alloc: &mut gpu::KernelAllocators, count: usize) -> Result> { + Ok(RxChannel { + ring: ChannelRing { + state: alloc.shared.new_default()?, + ring: alloc.shared.array_empty(T::SUB_CHANNELS * count)?, + }, + rptr: Default::default(), + count: count as u32, + }) + } + + /// Receives a message on the specified sub-channel index, optionally leaving in the ring + /// buffer. + /// + /// Returns None if the channel is empty. + fn get_or_peek(&mut self, index: usize, peek: bool) -> Option { + self.ring.state.with(|raw, _inner| { + let wptr = T::wptr(raw, index); + let rptr = &mut self.rptr[index]; + if wptr == *rptr { + None + } else { + let off = self.count as usize * index; + let msg = self.ring.ring[off + *rptr as usize]; + if !peek { + *rptr = (*rptr + 1) % self.count; + T::set_rptr(raw, index, *rptr); + } + Some(msg) + } + }) + } + + /// Receives a message on the specified sub-channel index, and dequeues it from the ring buffer. + /// + /// Returns None if the channel is empty. + pub(crate) fn get(&mut self, index: usize) -> Option { + self.get_or_peek(index, false) + } + + /// Peeks a message on the specified sub-channel index, leaving it in the ring buffer. + /// + /// Returns None if the channel is empty. + pub(crate) fn peek(&mut self, index: usize) -> Option { + self.get_or_peek(index, true) + } +} + +/// A transmit (driver->FW) channel. +pub(crate) struct TxChannel +where + for<'a> ::Raw<'a>: Debug + Default + Zeroed, +{ + ring: ChannelRing, + wptr: u32, + count: u32, +} + +impl TxChannel +where + for<'a> ::Raw<'a>: Debug + Default + Zeroed, +{ + /// Allocates a new cached transmit channel with a given message count. + pub(crate) fn new(alloc: &mut gpu::KernelAllocators, count: usize) -> Result> { + Ok(TxChannel { + ring: ChannelRing { + state: alloc.shared.new_default()?, + ring: alloc.private.array_empty(count)?, + }, + wptr: 0, + count: count as u32, + }) + } + + /// Allocates a new uncached transmit channel with a given message count. + pub(crate) fn new_uncached( + alloc: &mut gpu::KernelAllocators, + count: usize, + ) -> Result> { + Ok(TxChannel { + ring: ChannelRing { + state: alloc.shared.new_default()?, + ring: alloc.shared.array_empty(count)?, + }, + wptr: 0, + count: count as u32, + }) + } + + /// Send a message to the ring, returning a cookie with the ring buffer position. + /// + /// This will poll/block if the ring is full, which we don't really expect to happen. + pub(crate) fn put(&mut self, msg: &U) -> u32 { + self.ring.state.with(|raw, _inner| { + let next_wptr = (self.wptr + 1) % self.count; + let mut rptr = T::rptr(raw); + if next_wptr == rptr { + pr_err!( + "TX ring buffer is full! Waiting... ({}, {})\n", + next_wptr, + rptr + ); + // TODO: block properly on incoming messages? + while next_wptr == rptr { + coarse_sleep(Duration::from_millis(8)); + rptr = T::rptr(raw); + } + } + self.ring.ring[self.wptr as usize] = *msg; + mem::sync(); + T::set_wptr(raw, next_wptr); + self.wptr = next_wptr; + }); + self.wptr + } + + /// Wait for a previously submitted message to be popped off of the ring by the GPU firmware. + /// + /// This busy-loops, and is intended to be used for rare cases when we need to block for + /// completion of a cache management or invalidation operation synchronously (which + /// the firmware normally completes fast enough not to be worth sleeping for). + /// If the poll takes longer than 10ms, this switches to sleeping between polls. + pub(crate) fn wait_for(&mut self, wptr: u32, timeout_ms: u64) -> Result { + const MAX_FAST_POLL: u64 = 10; + let start = time::ktime_get(); + let timeout_fast = start + Duration::from_millis(timeout_ms.min(MAX_FAST_POLL)); + let timeout_slow = start + Duration::from_millis(timeout_ms); + self.ring.state.with(|raw, _inner| { + while time::ktime_get() < timeout_fast { + if T::rptr(raw) == wptr { + return Ok(()); + } + mem::sync(); + } + while time::ktime_get() < timeout_slow { + if T::rptr(raw) == wptr { + return Ok(()); + } + coarse_sleep(Duration::from_millis(5)); + mem::sync(); + } + Err(ETIMEDOUT) + }) + } +} + +/// Device Control channel for global device management commands. +#[versions(AGX)] +pub(crate) struct DeviceControlChannel { + dev: AsahiDevice, + ch: TxChannel, +} + +#[versions(AGX)] +impl DeviceControlChannel::ver { + const COMMAND_TIMEOUT_MS: u64 = 1000; + + /// Allocate a new Device Control channel. + pub(crate) fn new( + dev: &AsahiDevice, + alloc: &mut gpu::KernelAllocators, + ) -> Result { + Ok(DeviceControlChannel::ver { + dev: dev.clone(), + ch: TxChannel::::new(alloc, 0x100)?, + }) + } + + /// Returns the raw `ChannelRing` structure to pass to firmware. + pub(crate) fn to_raw(&self) -> raw::ChannelRing { + self.ch.ring.to_raw() + } + + /// Submits a Device Control command. + pub(crate) fn send(&mut self, msg: &DeviceControlMsg::ver) -> u32 { + cls_dev_dbg!(DeviceControlCh, self.dev, "DeviceControl: {:?}\n", msg); + self.ch.put(msg) + } + + /// Waits for a previously submitted Device Control command to complete. + pub(crate) fn wait_for(&mut self, wptr: u32) -> Result { + self.ch.wait_for(wptr, Self::COMMAND_TIMEOUT_MS) + } +} + +/// Pipe channel to submit WorkQueue execution requests. +#[versions(AGX)] +pub(crate) struct PipeChannel { + dev: AsahiDevice, + ch: TxChannel, +} + +#[versions(AGX)] +impl PipeChannel::ver { + /// Allocate a new Pipe submission channel. + pub(crate) fn new( + dev: &AsahiDevice, + alloc: &mut gpu::KernelAllocators, + ) -> Result { + Ok(PipeChannel::ver { + dev: dev.clone(), + ch: TxChannel::::new(alloc, 0x100)?, + }) + } + + /// Returns the raw `ChannelRing` structure to pass to firmware. + pub(crate) fn to_raw(&self) -> raw::ChannelRing { + self.ch.ring.to_raw() + } + + /// Submits a Pipe kick command to the firmware. + pub(crate) fn send(&mut self, msg: &PipeMsg::ver) { + cls_dev_dbg!(PipeCh, self.dev, "Pipe: {:?}\n", msg); + self.ch.put(msg); + } +} + +/// Firmware Control channel, used for secure cache flush requests. +pub(crate) struct FwCtlChannel { + dev: AsahiDevice, + ch: TxChannel, +} + +impl FwCtlChannel { + const COMMAND_TIMEOUT_MS: u64 = 1000; + + /// Allocate a new Firmware Control channel. + pub(crate) fn new( + dev: &AsahiDevice, + alloc: &mut gpu::KernelAllocators, + ) -> Result { + Ok(FwCtlChannel { + dev: dev.clone(), + ch: TxChannel::::new_uncached(alloc, 0x100)?, + }) + } + + /// Returns the raw `ChannelRing` structure to pass to firmware. + pub(crate) fn to_raw(&self) -> raw::ChannelRing { + self.ch.ring.to_raw() + } + + /// Submits a Firmware Control command to the firmware. + pub(crate) fn send(&mut self, msg: &FwCtlMsg) -> u32 { + cls_dev_dbg!(FwCtlCh, self.dev, "FwCtl: {:?}\n", msg); + self.ch.put(msg) + } + + /// Waits for a previously submitted Firmware Control command to complete. + pub(crate) fn wait_for(&mut self, wptr: u32) -> Result { + self.ch.wait_for(wptr, Self::COMMAND_TIMEOUT_MS) + } +} + +/// Event channel, used to notify the driver of command completions, GPU faults and errors, and +/// other events. +pub(crate) struct EventChannel { + dev: AsahiDevice, + ch: RxChannel, + mgr: Arc, + gpu: Option>, +} + +impl EventChannel { + /// Allocate a new Event channel. + pub(crate) fn new( + dev: &AsahiDevice, + alloc: &mut gpu::KernelAllocators, + mgr: Arc, + ) -> Result { + Ok(EventChannel { + dev: dev.clone(), + ch: RxChannel::::new(alloc, 0x100)?, + mgr, + gpu: None, + }) + } + + /// Registers the managing `Gpu` instance that will handle events on this channel. + pub(crate) fn set_manager(&mut self, gpu: Arc) { + self.gpu = Some(gpu); + } + + /// Returns the raw `ChannelRing` structure to pass to firmware. + pub(crate) fn to_raw(&self) -> raw::ChannelRing { + self.ch.ring.to_raw() + } + + /// Polls for new Event messages on this ring. + pub(crate) fn poll(&mut self) { + while let Some(msg) = self.ch.get(0) { + let tag = unsafe { msg.raw.0 }; + match tag { + 0..=EVENT_MAX => { + let msg = unsafe { msg.msg }; + + cls_dev_dbg!(EventCh, self.dev, "Event: {:?}\n", msg); + match msg { + EventMsg::Fault => match self.gpu.as_ref() { + Some(gpu) => gpu.handle_fault(), + None => { + dev_crit!(self.dev, "EventChannel: No GPU manager available!\n") + } + }, + EventMsg::Timeout { + counter, + event_slot, + .. + } => match self.gpu.as_ref() { + Some(gpu) => gpu.handle_timeout(counter, event_slot), + None => { + dev_crit!(self.dev, "EventChannel: No GPU manager available!\n") + } + }, + EventMsg::Flag { firing, .. } => { + for (i, flags) in firing.iter().enumerate() { + for j in 0..32 { + if flags & (1u32 << j) != 0 { + self.mgr.signal((i * 32 + j) as u32); + } + } + } + } + msg => { + dev_crit!(self.dev, "Unknown event message: {:?}\n", msg); + } + } + } + _ => { + dev_warn!(self.dev, "Unknown event message: {:?}\n", unsafe { + msg.raw + }); + } + } + } + } +} + +/// Firmware Log channel. This one is pretty special, since it has 6 sub-channels (for different log +/// levels), and it also uses a side buffer to actually hold the log messages, only passing around +/// pointers in the main buffer. +pub(crate) struct FwLogChannel { + dev: AsahiDevice, + ch: RxChannel, + payload_buf: GpuArray, +} + +impl FwLogChannel { + const RING_SIZE: usize = 0x100; + const BUF_SIZE: usize = 0x100; + + /// Allocate a new Firmware Log channel. + pub(crate) fn new( + dev: &AsahiDevice, + alloc: &mut gpu::KernelAllocators, + ) -> Result { + Ok(FwLogChannel { + dev: dev.clone(), + ch: RxChannel::::new(alloc, Self::RING_SIZE)?, + payload_buf: alloc + .shared + .array_empty(Self::BUF_SIZE * FwLogChannelState::SUB_CHANNELS)?, + }) + } + + /// Returns the raw `ChannelRing` structure to pass to firmware. + pub(crate) fn to_raw(&self) -> raw::ChannelRing { + self.ch.ring.to_raw() + } + + /// Returns the GPU pointers to the firmware log payload buffer. + pub(crate) fn get_buf(&self) -> GpuWeakPointer<[RawFwLogPayloadMsg]> { + self.payload_buf.weak_pointer() + } + + /// Polls for new log messages on all sub-rings. + pub(crate) fn poll(&mut self) { + for i in 0..=FwLogChannelState::SUB_CHANNELS - 1 { + while let Some(msg) = self.ch.peek(i) { + cls_dev_dbg!(FwLogCh, self.dev, "FwLog{}: {:?}\n", i, msg); + if msg.msg_type != 2 { + dev_warn!(self.dev, "Unknown FWLog{} message: {:?}\n", i, msg); + self.ch.get(i); + continue; + } + if msg.msg_index.0 as usize >= Self::BUF_SIZE { + dev_warn!( + self.dev, + "FWLog{} message index out of bounds: {:?}\n", + i, + msg + ); + self.ch.get(i); + continue; + } + let index = Self::BUF_SIZE * i + msg.msg_index.0 as usize; + let payload = &self.payload_buf.as_slice()[index]; + if payload.msg_type != 3 { + dev_warn!(self.dev, "Unknown FWLog{} payload: {:?}\n", i, payload); + self.ch.get(i); + continue; + } + let msg = if let Some(end) = payload.msg.iter().position(|&r| r == 0) { + CStr::from_bytes_with_nul(&(*payload.msg)[..end + 1]) + .unwrap_or(c_str!("cstr_err")) + } else { + dev_warn!( + self.dev, + "FWLog{} payload not NUL-terminated: {:?}\n", + i, + payload + ); + self.ch.get(i); + continue; + }; + match i { + 0 => dev_dbg!(self.dev, "FWLog: {}\n", msg), + 1 => dev_info!(self.dev, "FWLog: {}\n", msg), + 2 => dev_notice!(self.dev, "FWLog: {}\n", msg), + 3 => dev_warn!(self.dev, "FWLog: {}\n", msg), + 4 => dev_err!(self.dev, "FWLog: {}\n", msg), + 5 => dev_crit!(self.dev, "FWLog: {}\n", msg), + _ => (), + }; + self.ch.get(i); + } + } + } +} + +pub(crate) struct KTraceChannel { + dev: AsahiDevice, + ch: RxChannel, +} + +/// KTrace channel, used to receive detailed execution trace markers from the firmware. +/// We currently disable this in initdata, so no messages are expected here at this time. +impl KTraceChannel { + /// Allocate a new KTrace channel. + pub(crate) fn new( + dev: &AsahiDevice, + alloc: &mut gpu::KernelAllocators, + ) -> Result { + Ok(KTraceChannel { + dev: dev.clone(), + ch: RxChannel::::new(alloc, 0x200)?, + }) + } + + /// Returns the raw `ChannelRing` structure to pass to firmware. + pub(crate) fn to_raw(&self) -> raw::ChannelRing { + self.ch.ring.to_raw() + } + + /// Polls for new KTrace messages on this ring. + pub(crate) fn poll(&mut self) { + while let Some(msg) = self.ch.get(0) { + cls_dev_dbg!(KTraceCh, self.dev, "KTrace: {:?}\n", msg); + } + } +} + +/// Statistics channel, reporting power-related statistics to the driver. +/// Not really implemented other than debug logs yet... +#[versions(AGX)] +pub(crate) struct StatsChannel { + dev: AsahiDevice, + ch: RxChannel, +} + +#[versions(AGX)] +impl StatsChannel::ver { + /// Allocate a new Statistics channel. + pub(crate) fn new( + dev: &AsahiDevice, + alloc: &mut gpu::KernelAllocators, + ) -> Result { + Ok(StatsChannel::ver { + dev: dev.clone(), + ch: RxChannel::::new(alloc, 0x100)?, + }) + } + + /// Returns the raw `ChannelRing` structure to pass to firmware. + pub(crate) fn to_raw(&self) -> raw::ChannelRing { + self.ch.ring.to_raw() + } + + /// Polls for new statistics messages on this ring. + pub(crate) fn poll(&mut self) { + while let Some(msg) = self.ch.get(0) { + let tag = unsafe { msg.raw.0 }; + match tag { + 0..=STATS_MAX::ver => { + let msg = unsafe { msg.msg }; + cls_dev_dbg!(StatsCh, self.dev, "Stats: {:?}\n", msg); + } + _ => { + pr_warn!("Unknown stats message: {:?}\n", unsafe { msg.raw }); + } + } + } + } +} diff --git a/drivers/gpu/drm/asahi/debug.rs b/drivers/gpu/drm/asahi/debug.rs new file mode 100644 index 000000000000..2f3a70e04cfd --- /dev/null +++ b/drivers/gpu/drm/asahi/debug.rs @@ -0,0 +1,129 @@ +// SPDX-License-Identifier: GPL-2.0-only OR MIT +#![allow(dead_code)] + +//! Debug enable/disable flags and convenience macros + +#[allow(unused_imports)] +pub(crate) use super::{cls_dev_dbg, cls_pr_debug, debug, mod_dev_dbg, mod_pr_debug}; +use core::sync::atomic::{AtomicU64, Ordering}; + +static DEBUG_FLAGS: AtomicU64 = AtomicU64::new(0); + +/// Debug flag bit indices +pub(crate) enum DebugFlags { + // 0-3: Memory-related debug + Mmu = 0, + Alloc = 1, + Gem = 2, + Object = 3, + + // 4-7: Firmware objects and resources + Event = 4, + Buffer = 5, + WorkQueue = 6, + + // 8-13: DRM interface, rendering, compute, GPU globals + Gpu = 8, + File = 9, + Queue = 10, + Render = 11, + Compute = 12, + + // 14-15: Misc stats + MemStats = 14, + TVBStats = 15, + + // 16-22: Channels + FwLogCh = 16, + KTraceCh = 17, + StatsCh = 18, + EventCh = 19, + PipeCh = 20, + DeviceControlCh = 21, + FwCtlCh = 22, + + // 32-35: Allocator debugging + FillAllocations = 32, + DebugAllocations = 33, + DetectOverflows = 34, + ForceCPUMaps = 35, + + // 36-: Behavior flags + ConservativeTlbi = 36, + KeepGpuPowered = 37, + WaitForPowerOff = 38, + NoGpuRecovery = 39, + DisableClustering = 40, + + // 48-: Misc + Debug0 = 48, + Debug1 = 49, + Debug2 = 50, + Debug3 = 51, + Debug4 = 52, + Debug5 = 53, + Debug6 = 54, + Debug7 = 55, +} + +/// Update the cached global debug flags from the module parameter +pub(crate) fn update_debug_flags() { + let flags = { + let lock = crate::THIS_MODULE.kernel_param_lock(); + *crate::debug_flags.read(&lock) + }; + + DEBUG_FLAGS.store(flags, Ordering::Relaxed); +} + +/// Check whether debug is enabled for a given flag +#[inline(always)] +pub(crate) fn debug_enabled(flag: DebugFlags) -> bool { + DEBUG_FLAGS.load(Ordering::Relaxed) & 1 << (flag as usize) != 0 +} + +/// Run some code only if debug is enabled for the calling module +#[macro_export] +macro_rules! debug { + ($($arg:tt)*) => { + if $crate::debug::debug_enabled(DEBUG_CLASS) { + $($arg)* + } + }; +} + +/// pr_info!() if debug is enabled for the calling module +#[macro_export] +macro_rules! mod_pr_debug ( + ($($arg:tt)*) => ( + $crate::debug! { ::kernel::pr_info! ( $($arg)* ); } + ) +); + +/// dev_info!() if debug is enabled for the calling module +#[macro_export] +macro_rules! mod_dev_dbg ( + ($($arg:tt)*) => ( + $crate::debug! { ::kernel::dev_info! ( $($arg)* ); } + ) +); + +/// pr_info!() if debug is enabled for a specific module +#[macro_export] +macro_rules! cls_pr_debug ( + ($cls:ident, $($arg:tt)*) => ( + if $crate::debug::debug_enabled($crate::debug::DebugFlags::$cls) { + ::kernel::pr_info! ( $($arg)* ); + } + ) +); + +/// dev_info!() if debug is enabled for a specific module +#[macro_export] +macro_rules! cls_dev_dbg ( + ($cls:ident, $($arg:tt)*) => ( + if $crate::debug::debug_enabled($crate::debug::DebugFlags::$cls) { + ::kernel::dev_info! ( $($arg)* ); + } + ) +); diff --git a/drivers/gpu/drm/asahi/driver.rs b/drivers/gpu/drm/asahi/driver.rs new file mode 100644 index 000000000000..d49d8b1934a4 --- /dev/null +++ b/drivers/gpu/drm/asahi/driver.rs @@ -0,0 +1,166 @@ +// SPDX-License-Identifier: GPL-2.0-only OR MIT + +//! Top-level GPU driver implementation. + +use kernel::{ + c_str, device, drm, drm::drv, drm::ioctl, error::Result, of, platform, prelude::*, sync::Arc, +}; + +use crate::{debug, file, gem, gpu, hw, regs}; + +use kernel::device::RawDevice; +use kernel::macros::vtable; + +/// Driver metadata +const INFO: drv::DriverInfo = drv::DriverInfo { + major: 0, + minor: 0, + patchlevel: 0, + name: c_str!("asahi"), + desc: c_str!("Apple AGX Graphics"), + date: c_str!("20220831"), +}; + +/// Device data for the driver registration. +/// +/// Holds a reference to the top-level `GpuManager` object. +pub(crate) struct AsahiData { + pub(crate) dev: device::Device, + pub(crate) gpu: Arc, +} + +/// Convenience type alias for the `device::Data` type for this driver. +type DeviceData = device::Data, regs::Resources, AsahiData>; + +/// Empty struct representing this driver. +pub(crate) struct AsahiDriver; + +/// Convenience type alias for the DRM device type for this driver. +pub(crate) type AsahiDevice = kernel::drm::device::Device; + +/// DRM Driver implementation for `AsahiDriver`. +#[vtable] +impl drv::Driver for AsahiDriver { + /// Our `DeviceData` type, reference-counted + type Data = Arc; + /// Our `File` type. + type File = file::File; + /// Our `Object` type. + type Object = gem::Object; + + const INFO: drv::DriverInfo = INFO; + const FEATURES: u32 = + drv::FEAT_GEM | drv::FEAT_RENDER | drv::FEAT_SYNCOBJ | drv::FEAT_SYNCOBJ_TIMELINE; + + kernel::declare_drm_ioctls! { + (ASAHI_GET_PARAMS, drm_asahi_get_params, + ioctl::RENDER_ALLOW, file::File::get_params), + (ASAHI_VM_CREATE, drm_asahi_vm_create, + ioctl::AUTH | ioctl::RENDER_ALLOW, file::File::vm_create), + (ASAHI_VM_DESTROY, drm_asahi_vm_destroy, + ioctl::AUTH | ioctl::RENDER_ALLOW, file::File::vm_destroy), + (ASAHI_GEM_CREATE, drm_asahi_gem_create, + ioctl::AUTH | ioctl::RENDER_ALLOW, file::File::gem_create), + (ASAHI_GEM_MMAP_OFFSET, drm_asahi_gem_mmap_offset, + ioctl::AUTH | ioctl::RENDER_ALLOW, file::File::gem_mmap_offset), + (ASAHI_GEM_BIND, drm_asahi_gem_bind, + ioctl::AUTH | ioctl::RENDER_ALLOW, file::File::gem_bind), + (ASAHI_QUEUE_CREATE, drm_asahi_queue_create, + ioctl::AUTH | ioctl::RENDER_ALLOW, file::File::queue_create), + (ASAHI_QUEUE_DESTROY, drm_asahi_queue_destroy, + ioctl::AUTH | ioctl::RENDER_ALLOW, file::File::queue_destroy), + (ASAHI_SUBMIT, drm_asahi_submit, + ioctl::AUTH | ioctl::RENDER_ALLOW, file::File::submit), + } +} + +// OF Device ID table. +kernel::define_of_id_table! {ASAHI_ID_TABLE, &'static hw::HwConfig, [ + (of::DeviceId::Compatible(b"apple,agx-t8103"), Some(&hw::t8103::HWCONFIG)), + (of::DeviceId::Compatible(b"apple,agx-t8112"), Some(&hw::t8112::HWCONFIG)), + (of::DeviceId::Compatible(b"apple,agx-t6000"), Some(&hw::t600x::HWCONFIG_T6000)), + (of::DeviceId::Compatible(b"apple,agx-t6001"), Some(&hw::t600x::HWCONFIG_T6001)), + (of::DeviceId::Compatible(b"apple,agx-t6002"), Some(&hw::t600x::HWCONFIG_T6002)), +]} + +/// Platform Driver implementation for `AsahiDriver`. +impl platform::Driver for AsahiDriver { + /// Our `DeviceData` type, reference-counted + type Data = Arc; + /// Data associated with each hardware ID. + type IdInfo = &'static hw::HwConfig; + + // Assign the above OF ID table to this driver. + kernel::driver_of_id_table!(ASAHI_ID_TABLE); + + /// Device probe function. + fn probe( + pdev: &mut platform::Device, + id_info: Option<&Self::IdInfo>, + ) -> Result> { + debug::update_debug_flags(); + + let dev = device::Device::from_dev(pdev); + + dev_info!(dev, "Probing...\n"); + + let cfg = id_info.ok_or(ENODEV)?; + + pdev.set_dma_masks((1 << cfg.uat_oas) - 1)?; + + let res = regs::Resources::new(pdev)?; + + // Initialize misc MMIO + res.init_mmio()?; + + // Start the coprocessor CPU, so UAT can initialize the handoff + res.start_cpu()?; + + let node = dev.of_node().ok_or(EIO)?; + let compat: Vec = node.get_property(c_str!("apple,firmware-compat"))?; + + let reg = drm::drv::Registration::::new(&dev)?; + let gpu = match (cfg.gpu_gen, compat.as_slice()) { + (hw::GpuGen::G13, &[12, 3, 0]) => { + gpu::GpuManagerG13V12_3::new(reg.device(), &res, cfg)? as Arc + } + (hw::GpuGen::G13, &[13, 2, 0]) => { + gpu::GpuManagerG13V13_2::new(reg.device(), &res, cfg)? as Arc + } + (hw::GpuGen::G14, &[12, 4, 0]) => { + gpu::GpuManagerG14V12_4::new(reg.device(), &res, cfg)? as Arc + } + (hw::GpuGen::G14, &[13, 2, 0]) => { + gpu::GpuManagerG14V13_2::new(reg.device(), &res, cfg)? as Arc + } + _ => { + dev_info!( + dev, + "Unsupported GPU/firmware combination ({:?}, {:?})\n", + cfg.gpu_gen, + compat + ); + return Err(ENODEV); + } + }; + + let data = + kernel::new_device_data!(reg, res, AsahiData { dev, gpu }, "Asahi::Registrations")?; + + let data = Arc::::from(data); + + data.gpu.init()?; + + kernel::drm_device_register!( + data.registrations().ok_or(ENXIO)?.as_pinned_mut(), + data.clone(), + 0 + )?; + + dev_info!(data.dev, "Probed!\n"); + Ok(data) + } +} + +// Export the OF ID table as a module ID table, to make modpost/autoloading work. +kernel::module_of_id_table!(MOD_TABLE, ASAHI_ID_TABLE); diff --git a/drivers/gpu/drm/asahi/event.rs b/drivers/gpu/drm/asahi/event.rs new file mode 100644 index 000000000000..ccf00e4104be --- /dev/null +++ b/drivers/gpu/drm/asahi/event.rs @@ -0,0 +1,229 @@ +// SPDX-License-Identifier: GPL-2.0-only OR MIT + +//! GPU event manager +//! +//! The GPU firmware manages work completion by using event objects (Apple calls them "stamps"), +//! which are monotonically incrementing counters. There are a fixed number of objects, and +//! they are managed with a `SlotAllocator`. +//! +//! This module manages the set of available events and lets users compute expected values. +//! It also manages signaling owners when the GPU firmware reports that an event fired. + +use crate::debug::*; +use crate::fw::types::*; +use crate::{gpu, slotalloc, workqueue}; +use core::cmp; +use core::sync::atomic::Ordering; +use kernel::prelude::*; +use kernel::sync::Arc; + +const DEBUG_CLASS: DebugFlags = DebugFlags::Event; + +/// Number of events managed by the firmware. +const NUM_EVENTS: u32 = 128; + +/// Inner data associated with a given event slot. +pub(crate) struct EventInner { + /// CPU pointer to the driver notification event stamp + stamp: *const AtomicU32, + /// GPU pointer to the driver notification event stamp + gpu_stamp: GpuWeakPointer, + /// GPU pointer to the firmware-internal event stamp + gpu_fw_stamp: GpuWeakPointer, +} + +/// SAFETY: The event slots are safe to send across threads. +unsafe impl Send for EventInner {} + +/// Alias for an event token, which allows requesting the same event. +pub(crate) type Token = slotalloc::SlotToken; +/// Alias for an allocated `Event` that has a slot. +pub(crate) type Event = slotalloc::Guard; + +/// Represents a given stamp value for an event. +#[derive(Eq, PartialEq, Copy, Clone, Debug)] +#[repr(transparent)] +pub(crate) struct EventValue(u32); + +impl EventValue { + /// Returns the `EventValue` that succeeds this one. + pub(crate) fn next(&self) -> EventValue { + EventValue(self.0.wrapping_add(0x100)) + } + + /// Increments this `EventValue` in place. + pub(crate) fn increment(&mut self) { + self.0 = self.0.wrapping_add(0x100); + } + + /* Not used + /// Increments this `EventValue` in place by a certain count. + pub(crate) fn add(&mut self, val: u32) { + self.0 = self + .0 + .wrapping_add(val.checked_mul(0x100).expect("Adding too many events")); + } + */ + + /// Increments this `EventValue` in place by a certain count. + pub(crate) fn sub(&mut self, val: u32) { + self.0 = self + .0 + .wrapping_sub(val.checked_mul(0x100).expect("Subtracting too many events")); + } + + /// Computes the delta between this event and another event. + pub(crate) fn delta(&self, other: &EventValue) -> i32 { + (self.0.wrapping_sub(other.0) as i32) >> 8 + } +} + +impl PartialOrd for EventValue { + fn partial_cmp(&self, other: &Self) -> Option { + Some(self.cmp(other)) + } +} + +impl Ord for EventValue { + fn cmp(&self, other: &Self) -> cmp::Ordering { + self.delta(other).cmp(&0) + } +} + +impl EventInner { + /// Returns the GPU pointer to the driver notification stamp + pub(crate) fn stamp_pointer(&self) -> GpuWeakPointer { + self.gpu_stamp + } + + /// Returns the GPU pointer to the firmware internal stamp + pub(crate) fn fw_stamp_pointer(&self) -> GpuWeakPointer { + self.gpu_fw_stamp + } + + /// Fetches the current event value from shared memory + pub(crate) fn current(&self) -> EventValue { + // SAFETY: The pointer is always valid as constructed in + // EventManager below, and outside users cannot construct + // new EventInners, nor move or copy them, and Guards as + // returned by the SlotAllocator hold a reference to the + // SlotAllocator containing the EventManagerInner, which + // keeps the GpuObject the stamp is contained within alive. + EventValue(unsafe { &*self.stamp }.load(Ordering::Acquire)) + } +} + +impl slotalloc::SlotItem for EventInner { + type Data = EventManagerInner; + + fn release(&mut self, data: &mut Self::Data, slot: u32) { + mod_pr_debug!("EventManager: Released slot {}\n", slot); + data.owners[slot as usize] = None; + } +} + +/// Inner data for the event manager, to be protected by the SlotAllocator lock. +pub(crate) struct EventManagerInner { + stamps: GpuArray, + fw_stamps: GpuArray, + // Note: Use dyn to avoid having to version this entire module. + owners: Vec>>, +} + +/// Top-level EventManager object. +pub(crate) struct EventManager { + alloc: slotalloc::SlotAllocator, +} + +impl EventManager { + /// Create a new EventManager. + #[inline(never)] + pub(crate) fn new(alloc: &mut gpu::KernelAllocators) -> Result { + let mut owners = Vec::new(); + for _i in 0..(NUM_EVENTS as usize) { + owners.try_push(None)?; + } + let inner = EventManagerInner { + stamps: alloc.shared.array_empty(NUM_EVENTS as usize)?, + fw_stamps: alloc.private.array_empty(NUM_EVENTS as usize)?, + owners, + }; + + Ok(EventManager { + alloc: slotalloc::SlotAllocator::new( + NUM_EVENTS, + inner, + |inner: &mut EventManagerInner, slot| EventInner { + stamp: &inner.stamps[slot as usize].0, + gpu_stamp: inner.stamps.weak_item_pointer(slot as usize), + gpu_fw_stamp: inner.fw_stamps.weak_item_pointer(slot as usize), + }, + )?, + }) + } + + /// Gets a free `Event`, optionally trying to reuse the last one allocated by this caller. + pub(crate) fn get( + &self, + token: Option, + owner: Arc, + ) -> Result { + let ev = self.alloc.get_inner(token, |inner, ev| { + mod_pr_debug!( + "EventManager: Registered owner {:p} on slot {}\n", + &*owner, + ev.slot() + ); + inner.owners[ev.slot() as usize] = Some(owner); + Ok(()) + })?; + Ok(ev) + } + + /// Signals an event by slot, indicating completion (of one or more commands). + pub(crate) fn signal(&self, slot: u32) { + match self + .alloc + .with_inner(|inner| inner.owners[slot as usize].as_ref().cloned()) + { + Some(owner) => { + owner.signal(); + } + None => { + mod_pr_debug!("EventManager: Received event for empty slot {}\n", slot); + } + } + } + + /// Marks the owner of an event as having lost its work due to a GPU error. + pub(crate) fn mark_error(&self, slot: u32, wait_value: u32, error: workqueue::WorkError) { + match self + .alloc + .with_inner(|inner| inner.owners[slot as usize].as_ref().cloned()) + { + Some(owner) => { + owner.mark_error(EventValue(wait_value), error); + } + None => { + pr_err!("Received error for empty slot {}\n", slot); + } + } + } + + /// Fail all commands, used when the GPU crashes. + pub(crate) fn fail_all(&self, error: workqueue::WorkError) { + let mut owners: Vec> = Vec::new(); + + self.alloc.with_inner(|inner| { + for wq in inner.owners.iter().filter_map(|o| o.as_ref()).cloned() { + if owners.try_push(wq).is_err() { + pr_err!("Failed to signal failure to WorkQueue\n"); + } + } + }); + + for wq in owners { + wq.fail_all(error); + } + } +} diff --git a/drivers/gpu/drm/asahi/file.rs b/drivers/gpu/drm/asahi/file.rs new file mode 100644 index 000000000000..5d47feb30134 --- /dev/null +++ b/drivers/gpu/drm/asahi/file.rs @@ -0,0 +1,718 @@ +// SPDX-License-Identifier: GPL-2.0-only OR MIT +#![allow(clippy::unusual_byte_groupings)] + +//! File implementation, which represents a single DRM client. +//! +//! This is in charge of managing the resources associated with one GPU client, including an +//! arbitrary number of submission queues and Vm objects, and reporting hardware/driver +//! information to userspace and accepting submissions. + +use crate::debug::*; +use crate::driver::AsahiDevice; +use crate::{alloc, buffer, driver, gem, mmu, queue}; +use core::mem::MaybeUninit; +use kernel::dma_fence::RawDmaFence; +use kernel::drm::gem::BaseObject; +use kernel::io_buffer::{IoBufferReader, IoBufferWriter}; +use kernel::prelude::*; +use kernel::sync::{smutex::Mutex, Arc}; +use kernel::user_ptr::UserSlicePtr; +use kernel::{bindings, dma_fence, drm, xarray}; + +const DEBUG_CLASS: DebugFlags = DebugFlags::File; + +const MAX_SYNCS_PER_SUBMISSION: u32 = 64; +const MAX_COMMANDS_PER_SUBMISSION: u32 = 64; +pub(crate) const MAX_COMMANDS_IN_FLIGHT: u32 = 1024; + +/// A client instance of an `mmu::Vm` address space. +struct Vm { + ualloc: Arc>, + ualloc_priv: Arc>, + vm: mmu::Vm, + dummy_obj: gem::ObjectRef, +} + +impl Drop for Vm { + fn drop(&mut self) { + // Mappings create a reference loop, make sure to break it. + self.dummy_obj.drop_vm_mappings(self.vm.id()); + } +} + +/// Sync object from userspace. +pub(crate) struct SyncItem { + pub(crate) syncobj: drm::syncobj::SyncObj, + pub(crate) fence: Option, + pub(crate) chain_fence: Option, + pub(crate) timeline_value: u64, +} + +impl SyncItem { + fn parse_one(file: &DrmFile, data: bindings::drm_asahi_sync, out: bool) -> Result { + if data.extensions != 0 { + return Err(EINVAL); + } + + match data.sync_type { + bindings::drm_asahi_sync_type_DRM_ASAHI_SYNC_SYNCOBJ => { + if data.timeline_value != 0 { + return Err(EINVAL); + } + let syncobj = drm::syncobj::SyncObj::lookup_handle(file, data.handle)?; + + Ok(SyncItem { + fence: if out { + None + } else { + Some(syncobj.fence_get().ok_or(EINVAL)?) + }, + syncobj, + chain_fence: None, + timeline_value: data.timeline_value, + }) + } + bindings::drm_asahi_sync_type_DRM_ASAHI_SYNC_TIMELINE_SYNCOBJ => { + let syncobj = drm::syncobj::SyncObj::lookup_handle(file, data.handle)?; + let fence = if out { + None + } else { + Some( + syncobj + .fence_get() + .ok_or(EINVAL)? + .chain_find_seqno(data.timeline_value)?, + ) + }; + + Ok(SyncItem { + fence, + syncobj, + chain_fence: if out { + Some(dma_fence::FenceChain::new()?) + } else { + None + }, + timeline_value: data.timeline_value, + }) + } + _ => Err(EINVAL), + } + } + + fn parse_array(file: &DrmFile, ptr: u64, count: u32, out: bool) -> Result> { + let mut vec = Vec::try_with_capacity(count as usize)?; + + const STRIDE: usize = core::mem::size_of::(); + let size = STRIDE * count as usize; + + // SAFETY: We only read this once, so there are no TOCTOU issues. + let mut reader = unsafe { UserSlicePtr::new(ptr as usize as *mut _, size).reader() }; + + for _i in 0..count { + let mut sync: MaybeUninit = MaybeUninit::uninit(); + + // SAFETY: The size of `sync` is STRIDE + unsafe { reader.read_raw(sync.as_mut_ptr() as *mut u8, STRIDE)? }; + + // SAFETY: All bit patterns in the struct are valid + let sync = unsafe { sync.assume_init() }; + + vec.try_push(SyncItem::parse_one(file, sync, out)?)?; + } + + Ok(vec) + } +} + +/// State associated with a client. +pub(crate) struct File { + id: u64, + vms: xarray::XArray>, + queues: xarray::XArray>>>, +} + +/// Convenience type alias for our DRM `File` type. +pub(crate) type DrmFile = drm::file::File; + +/// Start address of the 32-bit USC address space. +const VM_SHADER_START: u64 = 0x11_00000000; +/// End address of the 32-bit USC address space. +const VM_SHADER_END: u64 = 0x11_ffffffff; +/// Start address of the general user mapping region. +const VM_USER_START: u64 = 0x20_00000000; +/// End address of the general user mapping region. +const VM_USER_END: u64 = 0x5f_ffffffff; + +/// Start address of the kernel-managed GPU-only mapping region. +const VM_DRV_GPU_START: u64 = 0x60_00000000; +/// End address of the kernel-managed GPU-only mapping region. +const VM_DRV_GPU_END: u64 = 0x60_ffffffff; +/// Start address of the kernel-managed GPU/FW shared mapping region. +const VM_DRV_GPUFW_START: u64 = 0x61_00000000; +/// End address of the kernel-managed GPU/FW shared mapping region. +const VM_DRV_GPUFW_END: u64 = 0x61_ffffffff; +/// Address of a special dummy page? +const VM_UNK_PAGE: u64 = 0x6f_ffff8000; + +impl drm::file::DriverFile for File { + type Driver = driver::AsahiDriver; + + /// Create a new `File` instance for a fresh client. + fn open(device: &AsahiDevice) -> Result> { + debug::update_debug_flags(); + + let gpu = &device.data().gpu; + let id = gpu.ids().file.next(); + + mod_dev_dbg!(device, "[File {}]: DRM device opened\n", id); + Ok(Box::try_new(Self { + id, + vms: xarray::XArray::new(xarray::flags::ALLOC1)?, + queues: xarray::XArray::new(xarray::flags::ALLOC1)?, + })?) + } +} + +impl File { + /// IOCTL: get_param: Get a driver parameter value. + pub(crate) fn get_params( + device: &AsahiDevice, + data: &mut bindings::drm_asahi_get_params, + file: &DrmFile, + ) -> Result { + mod_dev_dbg!(device, "[File {}]: IOCTL: get_params\n", file.id); + + let gpu = &device.data().gpu; + + if data.extensions != 0 || data.param_group != 0 || data.pad != 0 { + return Err(EINVAL); + } + + let mut params = bindings::drm_asahi_params_global { + unstable_uabi_version: bindings::DRM_ASAHI_UNSTABLE_UABI_VERSION, + pad0: 0, + + feat_compat: gpu.get_cfg().gpu_feat_compat, + feat_incompat: gpu.get_cfg().gpu_feat_incompat, + + gpu_generation: gpu.get_dyncfg().id.gpu_gen as u32, + gpu_variant: gpu.get_dyncfg().id.gpu_variant as u32, + gpu_revision: gpu.get_dyncfg().id.gpu_rev as u32, + chip_id: gpu.get_cfg().chip_id, + + num_dies: gpu.get_dyncfg().id.max_dies, + num_clusters_total: gpu.get_dyncfg().id.num_clusters, + num_cores_per_cluster: gpu.get_dyncfg().id.num_cores, + num_frags_per_cluster: gpu.get_dyncfg().id.num_frags, + num_gps_per_cluster: gpu.get_dyncfg().id.num_gps, + num_cores_total_active: gpu.get_dyncfg().id.total_active_cores, + core_masks: [0; bindings::DRM_ASAHI_MAX_CLUSTERS as usize], + + vm_page_size: mmu::UAT_PGSZ as u32, + pad1: 0, + vm_user_start: VM_USER_START, + vm_user_end: VM_USER_END, + vm_shader_start: VM_SHADER_START, + vm_shader_end: VM_SHADER_END, + + max_syncs_per_submission: MAX_SYNCS_PER_SUBMISSION, + max_commands_per_submission: MAX_COMMANDS_PER_SUBMISSION, + max_commands_in_flight: MAX_COMMANDS_IN_FLIGHT, + max_attachments: crate::microseq::MAX_ATTACHMENTS as u32, + + timer_frequency_hz: gpu.get_cfg().base_clock_hz, + min_frequency_khz: gpu.get_dyncfg().pwr.min_frequency_khz(), + max_frequency_khz: gpu.get_dyncfg().pwr.max_frequency_khz(), + max_power_mw: gpu.get_dyncfg().pwr.max_power_mw, + + result_render_size: core::mem::size_of::() as u32, + result_compute_size: core::mem::size_of::() as u32, + }; + + for (i, mask) in gpu.get_dyncfg().id.core_masks.iter().enumerate() { + *(params.core_masks.get_mut(i).ok_or(EIO)?) = (*mask).try_into()?; + } + + let size = + core::mem::size_of::().min(data.size.try_into()?); + + // SAFETY: We only write to this userptr once, so there are no TOCTOU issues. + let mut params_writer = + unsafe { UserSlicePtr::new(data.pointer as usize as *mut _, size).writer() }; + + // SAFETY: `size` is at most the sizeof of `params` + unsafe { params_writer.write_raw(¶ms as *const _ as *const u8, size)? }; + + Ok(0) + } + + /// IOCTL: vm_create: Create a new `Vm`. + pub(crate) fn vm_create( + device: &AsahiDevice, + data: &mut bindings::drm_asahi_vm_create, + file: &DrmFile, + ) -> Result { + if data.extensions != 0 { + return Err(EINVAL); + } + + let gpu = &device.data().gpu; + let file_id = file.id; + let vm = gpu.new_vm(file_id)?; + + let resv = file.vms.reserve()?; + let id: u32 = resv.index().try_into()?; + + mod_dev_dbg!(device, "[File {} VM {}]: VM Create\n", file_id, id); + mod_dev_dbg!( + device, + "[File {} VM {}]: Creating allocators\n", + file_id, + id + ); + let ualloc = Arc::try_new(Mutex::new(alloc::DefaultAllocator::new( + device, + &vm, + VM_DRV_GPU_START, + VM_DRV_GPU_END, + buffer::PAGE_SIZE, + mmu::PROT_GPU_SHARED_RW, + 512 * 1024, + true, + fmt!("File {} VM {} GPU Shared", file_id, id), + false, + )?))?; + let ualloc_priv = Arc::try_new(Mutex::new(alloc::DefaultAllocator::new( + device, + &vm, + VM_DRV_GPUFW_START, + VM_DRV_GPUFW_END, + buffer::PAGE_SIZE, + mmu::PROT_GPU_FW_PRIV_RW, + 64 * 1024, + true, + fmt!("File {} VM {} GPU FW Private", file_id, id), + false, + )?))?; + + mod_dev_dbg!( + device, + "[File {} VM {}]: Creating dummy object\n", + file_id, + id + ); + let mut dummy_obj = gem::new_kernel_object(device, 0x4000)?; + dummy_obj.vmap()?.as_mut_slice().fill(0); + dummy_obj.map_at(&vm, VM_UNK_PAGE, mmu::PROT_GPU_SHARED_RW, true)?; + + mod_dev_dbg!(device, "[File {} VM {}]: VM created\n", file_id, id); + resv.store(Box::try_new(Vm { + ualloc, + ualloc_priv, + vm, + dummy_obj, + })?)?; + + data.vm_id = id; + + Ok(0) + } + + /// IOCTL: vm_destroy: Destroy a `Vm`. + pub(crate) fn vm_destroy( + _device: &AsahiDevice, + data: &mut bindings::drm_asahi_vm_destroy, + file: &DrmFile, + ) -> Result { + if data.extensions != 0 { + return Err(EINVAL); + } + + if file.vms.remove(data.vm_id as usize).is_none() { + Err(ENOENT) + } else { + Ok(0) + } + } + + /// IOCTL: gem_create: Create a new GEM object. + pub(crate) fn gem_create( + device: &AsahiDevice, + data: &mut bindings::drm_asahi_gem_create, + file: &DrmFile, + ) -> Result { + mod_dev_dbg!( + device, + "[File {}]: IOCTL: gem_create size={:#x?}\n", + file.id, + data.size + ); + + if data.extensions != 0 + || (data.flags & !(bindings::ASAHI_GEM_WRITEBACK | bindings::ASAHI_GEM_VM_PRIVATE)) != 0 + || (data.flags & bindings::ASAHI_GEM_VM_PRIVATE == 0 && data.vm_id != 0) + { + return Err(EINVAL); + } + + let vm_id = if data.flags & bindings::ASAHI_GEM_VM_PRIVATE != 0 { + Some(file.vms.get(data.vm_id.try_into()?).ok_or(ENOENT)?.vm.id()) + } else { + None + }; + + let bo = gem::new_object(device, data.size.try_into()?, data.flags, vm_id)?; + + let handle = bo.gem.create_handle(file)?; + data.handle = handle; + + mod_dev_dbg!( + device, + "[File {}]: IOCTL: gem_create size={:#x} handle={:#x?}\n", + file.id, + data.size, + data.handle + ); + + Ok(0) + } + + /// IOCTL: gem_mmap_offset: Assign an mmap offset to a GEM object. + pub(crate) fn gem_mmap_offset( + device: &AsahiDevice, + data: &mut bindings::drm_asahi_gem_mmap_offset, + file: &DrmFile, + ) -> Result { + mod_dev_dbg!( + device, + "[File {}]: IOCTL: gem_mmap_offset handle={:#x?}\n", + file.id, + data.handle + ); + + if data.extensions != 0 || data.flags != 0 { + return Err(EINVAL); + } + + let bo = gem::lookup_handle(file, data.handle)?; + data.offset = bo.gem.create_mmap_offset()?; + Ok(0) + } + + /// IOCTL: gem_bind: Map or unmap a GEM object into a Vm. + pub(crate) fn gem_bind( + device: &AsahiDevice, + data: &mut bindings::drm_asahi_gem_bind, + file: &DrmFile, + ) -> Result { + mod_dev_dbg!( + device, + "[File {} VM {}]: IOCTL: gem_bind op={:?} handle={:#x?} flags={:#x?} {:#x?}:{:#x?} -> {:#x?}\n", + file.id, + data.op, + data.vm_id, + data.handle, + data.flags, + data.offset, + data.range, + data.addr + ); + + if data.extensions != 0 { + return Err(EINVAL); + } + + match data.op { + bindings::drm_asahi_bind_op_ASAHI_BIND_OP_BIND => Self::do_gem_bind(device, data, file), + bindings::drm_asahi_bind_op_ASAHI_BIND_OP_UNBIND => Err(ENOTSUPP), + bindings::drm_asahi_bind_op_ASAHI_BIND_OP_UNBIND_ALL => { + Self::do_gem_unbind_all(device, data, file) + } + _ => Err(EINVAL), + } + } + + pub(crate) fn do_gem_bind( + _device: &AsahiDevice, + data: &mut bindings::drm_asahi_gem_bind, + file: &DrmFile, + ) -> Result { + if data.offset != 0 { + return Err(EINVAL); // Not supported yet + } + + if (data.addr | data.range) as usize & mmu::UAT_PGMSK != 0 { + return Err(EINVAL); // Must be page aligned + } + + if (data.flags & !(bindings::ASAHI_BIND_READ | bindings::ASAHI_BIND_WRITE)) != 0 { + return Err(EINVAL); + } + + let mut bo = gem::lookup_handle(file, data.handle)?; + + if data.range != bo.size().try_into()? { + return Err(EINVAL); // Not supported yet + } + + let start = data.addr; + let end = data.addr + data.range - 1; + + if (VM_SHADER_START..=VM_SHADER_END).contains(&start) { + if !(VM_SHADER_START..=VM_SHADER_END).contains(&end) { + return Err(EINVAL); // Invalid map range + } + } else if (VM_USER_START..=VM_USER_END).contains(&start) { + if !(VM_USER_START..=VM_USER_END).contains(&end) { + return Err(EINVAL); // Invalid map range + } + } else { + return Err(EINVAL); // Invalid map range + } + + // Just in case + if end >= VM_DRV_GPU_START { + return Err(EINVAL); + } + + let prot = if data.flags & bindings::ASAHI_BIND_READ != 0 { + if data.flags & bindings::ASAHI_BIND_WRITE != 0 { + mmu::PROT_GPU_SHARED_RW + } else { + mmu::PROT_GPU_SHARED_RO + } + } else if data.flags & bindings::ASAHI_BIND_WRITE != 0 { + mmu::PROT_GPU_SHARED_WO + } else { + return Err(EINVAL); // Must specify one of ASAHI_BIND_{READ,WRITE} + }; + + // Clone it immediately so we aren't holding the XArray lock + let vm = file + .vms + .get(data.vm_id.try_into()?) + .ok_or(ENOENT)? + .vm + .clone(); + + bo.map_at(&vm, start, prot, true)?; + + Ok(0) + } + + pub(crate) fn do_gem_unbind_all( + _device: &AsahiDevice, + data: &mut bindings::drm_asahi_gem_bind, + file: &DrmFile, + ) -> Result { + if data.flags != 0 || data.offset != 0 || data.range != 0 || data.addr != 0 { + return Err(EINVAL); + } + + let mut bo = gem::lookup_handle(file, data.handle)?; + + if data.vm_id == 0 { + bo.drop_file_mappings(file.id); + } else { + let vm_id = file.vms.get(data.vm_id.try_into()?).ok_or(ENOENT)?.vm.id(); + bo.drop_vm_mappings(vm_id); + } + + Ok(0) + } + + /// IOCTL: queue_create: Create a new command submission queue of a given type. + pub(crate) fn queue_create( + device: &AsahiDevice, + data: &mut bindings::drm_asahi_queue_create, + file: &DrmFile, + ) -> Result { + let file_id = file.id; + + mod_dev_dbg!( + device, + "[File {} VM {}]: Creating queue caps={:?} prio={:?} flags={:#x?}\n", + file_id, + data.vm_id, + data.queue_caps, + data.priority, + data.flags, + ); + + if data.extensions != 0 + || data.flags != 0 + || data.priority > 3 + || data.queue_caps == 0 + || (data.queue_caps + & !(bindings::drm_asahi_queue_cap_DRM_ASAHI_QUEUE_CAP_RENDER + | bindings::drm_asahi_queue_cap_DRM_ASAHI_QUEUE_CAP_BLIT + | bindings::drm_asahi_queue_cap_DRM_ASAHI_QUEUE_CAP_COMPUTE)) + != 0 + { + return Err(EINVAL); + } + + let resv = file.queues.reserve()?; + let file_vm = file.vms.get(data.vm_id.try_into()?).ok_or(ENOENT)?; + let vm = file_vm.vm.clone(); + let ualloc = file_vm.ualloc.clone(); + let ualloc_priv = file_vm.ualloc_priv.clone(); + // Drop the vms lock eagerly + core::mem::drop(file_vm); + + let queue = + device + .data() + .gpu + .new_queue(vm, ualloc, ualloc_priv, data.priority, data.queue_caps)?; + + data.queue_id = resv.index().try_into()?; + resv.store(Arc::try_new(Mutex::new(queue))?)?; + + Ok(0) + } + + /// IOCTL: queue_destroy: Destroy a command submission queue. + pub(crate) fn queue_destroy( + _device: &AsahiDevice, + data: &mut bindings::drm_asahi_queue_destroy, + file: &DrmFile, + ) -> Result { + if data.extensions != 0 { + return Err(EINVAL); + } + + if file.queues.remove(data.queue_id as usize).is_none() { + Err(ENOENT) + } else { + Ok(0) + } + } + + /// IOCTL: submit: Submit GPU work to a command submission queue. + pub(crate) fn submit( + device: &AsahiDevice, + data: &mut bindings::drm_asahi_submit, + file: &DrmFile, + ) -> Result { + if data.extensions != 0 + || data.flags != 0 + || data.in_sync_count > MAX_SYNCS_PER_SUBMISSION + || data.out_sync_count > MAX_SYNCS_PER_SUBMISSION + || data.command_count > MAX_COMMANDS_PER_SUBMISSION + { + return Err(EINVAL); + } + + debug::update_debug_flags(); + + let gpu = &device.data().gpu; + gpu.update_globals(); + + // Upgrade to Arc to drop the XArray lock early + let queue: Arc>> = file + .queues + .get(data.queue_id.try_into()?) + .ok_or(ENOENT)? + .borrow() + .into(); + + let id = gpu.ids().submission.next(); + mod_dev_dbg!( + device, + "[File {} Queue {}]: IOCTL: submit (submission ID: {})\n", + file.id, + data.queue_id, + id + ); + + mod_dev_dbg!( + device, + "[File {} Queue {}]: IOCTL: submit({}): Parsing in_syncs\n", + file.id, + data.queue_id, + id + ); + let in_syncs = SyncItem::parse_array(file, data.in_syncs, data.in_sync_count, false)?; + mod_dev_dbg!( + device, + "[File {} Queue {}]: IOCTL: submit({}): Parsing out_syncs\n", + file.id, + data.queue_id, + id + ); + let out_syncs = SyncItem::parse_array(file, data.out_syncs, data.out_sync_count, true)?; + + let result_buf = if data.result_handle != 0 { + mod_dev_dbg!( + device, + "[File {} Queue {}]: IOCTL: submit({}): Looking up result_handle {}\n", + file.id, + data.queue_id, + id, + data.result_handle + ); + Some(gem::lookup_handle(file, data.result_handle)?) + } else { + None + }; + + mod_dev_dbg!( + device, + "[File {} Queue {}]: IOCTL: submit({}): Parsing commands\n", + file.id, + data.queue_id, + id + ); + let mut commands = Vec::try_with_capacity(data.command_count as usize)?; + + const STRIDE: usize = core::mem::size_of::(); + let size = STRIDE * data.command_count as usize; + + // SAFETY: We only read this once, so there are no TOCTOU issues. + let mut reader = + unsafe { UserSlicePtr::new(data.commands as usize as *mut _, size).reader() }; + + for _i in 0..data.command_count { + let mut cmd: MaybeUninit = MaybeUninit::uninit(); + + // SAFETY: The size of `sync` is STRIDE + unsafe { reader.read_raw(cmd.as_mut_ptr() as *mut u8, STRIDE)? }; + + // SAFETY: All bit patterns in the struct are valid + commands.try_push(unsafe { cmd.assume_init() })?; + } + + let ret = queue + .lock() + .submit(id, in_syncs, out_syncs, result_buf, commands); + + match ret { + Err(ERESTARTSYS) => Err(ERESTARTSYS), + Err(e) => { + dev_info!( + device, + "[File {} Queue {}]: IOCTL: submit failed! (submission ID: {} err: {:?})\n", + file.id, + data.queue_id, + id, + e + ); + Err(e) + } + Ok(_) => Ok(0), + } + } + + /// Returns the unique file ID for this `File`. + pub(crate) fn file_id(&self) -> u64 { + self.id + } +} + +impl Drop for File { + fn drop(&mut self) { + mod_pr_debug!("[File {}]: Closing...\n", self.id); + } +} diff --git a/drivers/gpu/drm/asahi/float.rs b/drivers/gpu/drm/asahi/float.rs new file mode 100644 index 000000000000..e73b4b628cf9 --- /dev/null +++ b/drivers/gpu/drm/asahi/float.rs @@ -0,0 +1,381 @@ +// SPDX-License-Identifier: GPL-2.0-only OR MIT + +//! Basic soft floating-point support +//! +//! The GPU firmware requires a large number of power-related configuration values, many of which +//! are IEEE 754 32-bit floating point values. These values change not only between GPU/SoC +//! variants, but also between specific hardware platforms using these SoCs, so they must be +//! derived from device tree properties. There are many redundant values computed from the same +//! inputs with simple add/sub/mul/div calculations, plus a few values that are actually specific +//! to each individual device depending on its binning and fused voltage configuration, so it +//! doesn't make sense to store the final values to be passed to the firmware in the device tree. +//! +//! Therefore, we need a way to perform floating-point calculations in the kernel. +//! +//! Using the actual FPU from kernel mode is asking for trouble, since there is no way to bound +//! the execution of FPU instructions to a controlled section of code without outright putting it +//! in its own compilation unit, which is quite painful for Rust. Since these calculations only +//! have to happen at initialization time and there is no need for performance, let's use a simple +//! software float implementation instead. +//! +//! This implementation makes no attempt to be fully IEEE754 compliant, but it's good enough and +//! gives bit-identical results to macOS in the vast majority of cases, with one or two exceptions +//! related to slightly non-compliant rounding. + +use core::ops; +use kernel::{of, prelude::*}; + +/// An IEEE754-compatible floating point number implemented in software. +#[derive(Default, Debug, Copy, Clone)] +pub(crate) struct F32(u32); + +#[derive(Default, Debug, Copy, Clone)] +struct F32U { + sign: bool, + exp: i32, + frac: i64, +} + +impl F32 { + /// Convert a raw 32-bit representation into an F32 + pub(crate) const fn from_bits(u: u32) -> F32 { + F32(u) + } + + // Convert a `f32` value into an F32 + // + // This must ONLY be used in const context. Use the `f32!{}` macro to do it safely. + #[doc(hidden)] + pub(crate) const fn from_f32(v: f32) -> F32 { + F32(unsafe { core::mem::transmute(v) }) + } + + // Convert an F32 into a `f32` value + // + // For testing only. + #[doc(hidden)] + #[cfg(test)] + pub(crate) fn to_f32(self) -> f32 { + f32::from_bits(self.0) + } + + const fn unpack(&self) -> F32U { + F32U { + sign: self.0 & (1 << 31) != 0, + exp: ((self.0 >> 23) & 0xff) as i32 - 127, + frac: (((self.0 & 0x7fffff) | 0x800000) as i64) << 9, + } + .norm() + } +} + +/// Safely construct an `F32` out of a constant floating-point value. +/// +/// This ensures that the conversion happens in const context, so no floating point operations are +/// emitted. +#[macro_export] +macro_rules! f32 { + ([$($val:expr),*]) => {{ + [$(f32!($val)),*] + }}; + ($val:expr) => {{ + const _K: $crate::float::F32 = $crate::float::F32::from_f32($val); + _K + }}; +} + +impl ops::Neg for F32 { + type Output = F32; + + fn neg(self) -> F32 { + F32(self.0 ^ (1 << 31)) + } +} + +impl ops::Add for F32 { + type Output = F32; + + fn add(self, rhs: F32) -> F32 { + self.unpack().add(rhs.unpack()).pack() + } +} + +impl ops::Sub for F32 { + type Output = F32; + + fn sub(self, rhs: F32) -> F32 { + self.unpack().add((-rhs).unpack()).pack() + } +} + +impl ops::Mul for F32 { + type Output = F32; + + fn mul(self, rhs: F32) -> F32 { + self.unpack().mul(rhs.unpack()).pack() + } +} + +impl ops::Div for F32 { + type Output = F32; + + fn div(self, rhs: F32) -> F32 { + self.unpack().div(rhs.unpack()).pack() + } +} + +macro_rules! from_ints { + ($u:ty, $i:ty) => { + impl From<$i> for F32 { + fn from(v: $i) -> F32 { + F32U::from_i64(v as i64).pack() + } + } + impl From<$u> for F32 { + fn from(v: $u) -> F32 { + F32U::from_u64(v as u64).pack() + } + } + }; +} + +from_ints!(u8, i8); +from_ints!(u16, i16); +from_ints!(u32, i32); +from_ints!(u64, i64); + +impl F32U { + const INFINITY: F32U = f32!(f32::INFINITY).unpack(); + const NEG_INFINITY: F32U = f32!(f32::NEG_INFINITY).unpack(); + + fn from_i64(v: i64) -> F32U { + F32U { + sign: v < 0, + exp: 32, + frac: v.abs(), + } + .norm() + } + + fn from_u64(mut v: u64) -> F32U { + let mut exp = 32; + if v >= (1 << 63) { + exp = 31; + v >>= 1; + } + F32U { + sign: false, + exp, + frac: v as i64, + } + .norm() + } + + fn shr(&mut self, shift: i32) { + if shift > 63 { + self.exp = 0; + self.frac = 0; + } else { + self.frac >>= shift; + } + } + + fn align(a: &mut F32U, b: &mut F32U) { + if a.exp > b.exp { + b.shr(a.exp - b.exp); + b.exp = a.exp; + } else { + a.shr(b.exp - a.exp); + a.exp = b.exp; + } + } + + fn mul(self, other: F32U) -> F32U { + F32U { + sign: self.sign != other.sign, + exp: self.exp + other.exp, + frac: ((self.frac >> 8) * (other.frac >> 8)) >> 16, + } + } + + fn div(self, other: F32U) -> F32U { + if other.frac == 0 || self.is_inf() { + if self.sign { + F32U::NEG_INFINITY + } else { + F32U::INFINITY + } + } else { + F32U { + sign: self.sign != other.sign, + exp: self.exp - other.exp, + frac: ((self.frac << 24) / (other.frac >> 8)), + } + } + } + + fn add(mut self, mut other: F32U) -> F32U { + F32U::align(&mut self, &mut other); + if self.sign == other.sign { + self.frac += other.frac; + } else { + self.frac -= other.frac; + } + if self.frac < 0 { + self.sign = !self.sign; + self.frac = -self.frac; + } + self + } + + const fn norm(mut self) -> F32U { + let lz = self.frac.leading_zeros() as i32; + if lz > 31 { + self.frac <<= lz - 31; + self.exp -= lz - 31; + } else if lz < 31 { + self.frac >>= 31 - lz; + self.exp += 31 - lz; + } + + if self.is_zero() { + return F32U { + sign: self.sign, + frac: 0, + exp: 0, + }; + } + self + } + + const fn is_zero(&self) -> bool { + self.frac == 0 || self.exp < -126 + } + + const fn is_inf(&self) -> bool { + self.exp > 127 + } + + const fn pack(mut self) -> F32 { + self = self.norm(); + if !self.is_zero() { + self.frac += 0x100; + self = self.norm(); + } + + if self.is_inf() { + if self.sign { + return f32!(f32::NEG_INFINITY); + } else { + return f32!(f32::INFINITY); + } + } else if self.is_zero() { + if self.sign { + return f32!(-0.0); + } else { + return f32!(0.0); + } + } + + F32(if self.sign { 1u32 << 31 } else { 0u32 } + | ((self.exp + 127) as u32) << 23 + | ((self.frac >> 9) & 0x7fffff) as u32) + } +} + +impl<'a> TryFrom> for F32 { + type Error = Error; + + fn try_from(p: of::Property<'_>) -> core::result::Result { + let bits: u32 = p.try_into()?; + Ok(F32::from_bits(bits)) + } +} + +impl of::PropertyUnit for F32 { + const UNIT_SIZE: usize = 4; + + fn from_bytes(data: &[u8]) -> Result { + Ok(F32::from_bits(::from_bytes(data)?)) + } +} + +// TODO: Make this an actual test and figure out how to make it run. +#[cfg(test)] +mod tests { + #[test] + fn test_all() { + fn add(a: f32, b: f32) { + println!( + "{} + {} = {} {}", + a, + b, + (F32::from_f32(a) + F32::from_f32(b)).to_f32(), + a + b + ); + } + fn sub(a: f32, b: f32) { + println!( + "{} - {} = {} {}", + a, + b, + (F32::from_f32(a) - F32::from_f32(b)).to_f32(), + a - b + ); + } + fn mul(a: f32, b: f32) { + println!( + "{} * {} = {} {}", + a, + b, + (F32::from_f32(a) * F32::from_f32(b)).to_f32(), + a * b + ); + } + fn div(a: f32, b: f32) { + println!( + "{} / {} = {} {}", + a, + b, + (F32::from_f32(a) / F32::from_f32(b)).to_f32(), + a / b + ); + } + + fn test(a: f32, b: f32) { + add(a, b); + sub(a, b); + mul(a, b); + div(a, b); + } + + test(1.123, 7.567); + test(1.123, 1.456); + test(7.567, 1.123); + test(1.123, -7.567); + test(1.123, -1.456); + test(7.567, -1.123); + test(-1.123, -7.567); + test(-1.123, -1.456); + test(-7.567, -1.123); + test(1000.123, 0.001); + test(1000.123, 0.0000001); + test(0.0012, 1000.123); + test(0.0000001, 1000.123); + test(0., 0.); + test(0., 1.); + test(1., 0.); + test(1., 1.); + test(2., f32::INFINITY); + test(2., f32::NEG_INFINITY); + test(f32::INFINITY, 2.); + test(f32::NEG_INFINITY, 2.); + test(f32::NEG_INFINITY, 2.); + test(f32::MAX, 2.); + test(f32::MIN, 2.); + test(f32::MIN_POSITIVE, 2.); + test(2., f32::MAX); + test(2., f32::MIN); + test(2., f32::MIN_POSITIVE); + } +} diff --git a/drivers/gpu/drm/asahi/fw/buffer.rs b/drivers/gpu/drm/asahi/fw/buffer.rs new file mode 100644 index 000000000000..a8a467879518 --- /dev/null +++ b/drivers/gpu/drm/asahi/fw/buffer.rs @@ -0,0 +1,170 @@ +// SPDX-License-Identifier: GPL-2.0-only OR MIT + +//! GPU tiled vertex buffer control firmware structures + +use super::types::*; +use super::workqueue; +use crate::{default_zeroed, no_debug, trivial_gpustruct}; +use kernel::sync::Arc; + +pub(crate) mod raw { + use super::*; + + #[derive(Debug)] + #[repr(C)] + pub(crate) struct BlockControl { + pub(crate) total: AtomicU32, + pub(crate) wptr: AtomicU32, + pub(crate) unk: AtomicU32, + pub(crate) pad: Pad<0x34>, + } + default_zeroed!(BlockControl); + + #[derive(Debug)] + #[repr(C)] + pub(crate) struct Counter { + pub(crate) count: AtomicU32, + __pad: Pad<0x3c>, + } + default_zeroed!(Counter); + + #[derive(Debug, Default)] + #[repr(C)] + pub(crate) struct Stats { + pub(crate) max_pages: AtomicU32, + pub(crate) max_b: AtomicU32, + pub(crate) overflow_count: AtomicU32, + pub(crate) gpu_c: AtomicU32, + pub(crate) __pad0: Pad<0x10>, + pub(crate) reset: AtomicU32, + pub(crate) __pad1: Pad<0x1c>, + } + + #[versions(AGX)] + #[derive(Debug)] + #[repr(C)] + pub(crate) struct Info<'a> { + pub(crate) gpu_counter: u32, + pub(crate) unk_4: u32, + pub(crate) last_id: i32, + pub(crate) cur_id: i32, + pub(crate) unk_10: u32, + pub(crate) gpu_counter2: u32, + pub(crate) unk_18: u32, + + #[ver(V < V13_0B4)] + pub(crate) unk_1c: u32, + + pub(crate) page_list: GpuPointer<'a, &'a [u32]>, + pub(crate) page_list_size: u32, + pub(crate) page_count: AtomicU32, + pub(crate) max_blocks: u32, + pub(crate) block_count: AtomicU32, + pub(crate) unk_38: u32, + pub(crate) block_list: GpuPointer<'a, &'a [u32]>, + pub(crate) block_ctl: GpuPointer<'a, super::BlockControl>, + pub(crate) last_page: AtomicU32, + pub(crate) gpu_page_ptr1: u32, + pub(crate) gpu_page_ptr2: u32, + pub(crate) unk_58: u32, + pub(crate) block_size: u32, + pub(crate) unk_60: U64, + pub(crate) counter: GpuPointer<'a, super::Counter>, + pub(crate) unk_70: u32, + pub(crate) unk_74: u32, + pub(crate) unk_78: u32, + pub(crate) unk_7c: u32, + pub(crate) unk_80: u32, + pub(crate) max_pages: u32, + pub(crate) max_pages_nomemless: u32, + pub(crate) unk_8c: u32, + pub(crate) unk_90: Array<0x30, u8>, + } + + #[derive(Debug)] + #[repr(C)] + pub(crate) struct Scene<'a> { + pub(crate) pass_page_count: AtomicU32, + pub(crate) unk_4: u32, + pub(crate) unk_8: U64, + pub(crate) unk_10: U64, + pub(crate) user_buffer: GpuPointer<'a, &'a [u8]>, + pub(crate) unk_20: u32, + pub(crate) stats: GpuWeakPointer, + pub(crate) total_page_count: AtomicU32, + pub(crate) unk_30: U64, // pad + pub(crate) unk_38: U64, // pad + } + + #[versions(AGX)] + #[derive(Debug)] + #[repr(C)] + pub(crate) struct InitBuffer<'a> { + pub(crate) tag: workqueue::CommandType, + pub(crate) vm_slot: u32, + pub(crate) buffer_slot: u32, + pub(crate) unk_c: u32, + pub(crate) block_count: u32, + pub(crate) buffer: GpuPointer<'a, super::Info::ver>, + pub(crate) stamp_value: EventValue, + } +} + +trivial_gpustruct!(BlockControl); +trivial_gpustruct!(Counter); +trivial_gpustruct!(Stats); + +#[versions(AGX)] +#[derive(Debug)] +pub(crate) struct Info { + pub(crate) block_ctl: GpuObject, + pub(crate) counter: GpuObject, + pub(crate) page_list: GpuArray, + pub(crate) block_list: GpuArray, +} + +#[versions(AGX)] +impl GpuStruct for Info::ver { + type Raw<'a> = raw::Info::ver<'a>; +} + +pub(crate) struct ClusterBuffers { + pub(crate) tilemaps: GpuArray, + pub(crate) meta: GpuArray, +} + +#[versions(AGX)] +pub(crate) struct Scene { + pub(crate) user_buffer: GpuArray, + pub(crate) buffer: crate::buffer::Buffer::ver, + pub(crate) tvb_heapmeta: GpuArray, + pub(crate) tvb_tilemap: GpuArray, + pub(crate) tpc: Arc>, + pub(crate) clustering: Option, + pub(crate) preempt_buf: GpuArray, + pub(crate) seq_buf: GpuArray, +} + +#[versions(AGX)] +no_debug!(Scene::ver); + +#[versions(AGX)] +impl GpuStruct for Scene::ver { + type Raw<'a> = raw::Scene<'a>; +} + +#[versions(AGX)] +pub(crate) struct InitBuffer { + pub(crate) scene: Arc, +} + +#[versions(AGX)] +no_debug!(InitBuffer::ver); + +#[versions(AGX)] +impl workqueue::Command for InitBuffer::ver {} + +#[versions(AGX)] +impl GpuStruct for InitBuffer::ver { + type Raw<'a> = raw::InitBuffer::ver<'a>; +} diff --git a/drivers/gpu/drm/asahi/fw/channels.rs b/drivers/gpu/drm/asahi/fw/channels.rs new file mode 100644 index 000000000000..db5ac9a3ded5 --- /dev/null +++ b/drivers/gpu/drm/asahi/fw/channels.rs @@ -0,0 +1,385 @@ +// SPDX-License-Identifier: GPL-2.0-only OR MIT + +//! GPU communication channel firmware structures (ring buffers) + +use super::types::*; +use crate::default_zeroed; +use core::sync::atomic::Ordering; + +pub(crate) mod raw { + use super::*; + + #[derive(Debug)] + #[repr(C)] + pub(crate) struct ChannelState<'a> { + pub(crate) read_ptr: AtomicU32, + __pad0: Pad<0x1c>, + pub(crate) write_ptr: AtomicU32, + __pad1: Pad<0xc>, + _p: PhantomData<&'a ()>, + } + default_zeroed!(<'a>, ChannelState<'a>); + + #[derive(Debug)] + #[repr(C)] + pub(crate) struct FwCtlChannelState<'a> { + pub(crate) read_ptr: AtomicU32, + __pad0: Pad<0xc>, + pub(crate) write_ptr: AtomicU32, + __pad1: Pad<0xc>, + _p: PhantomData<&'a ()>, + } + default_zeroed!(<'a>, FwCtlChannelState<'a>); +} + +pub(crate) trait RxChannelState: GpuStruct + Debug + Default +where + for<'a> ::Raw<'a>: Default + Zeroed, +{ + const SUB_CHANNELS: usize; + + fn wptr(raw: &Self::Raw<'_>, index: usize) -> u32; + fn set_rptr(raw: &Self::Raw<'_>, index: usize, rptr: u32); +} + +#[derive(Debug, Default)] +pub(crate) struct ChannelState {} + +impl GpuStruct for ChannelState { + type Raw<'a> = raw::ChannelState<'a>; +} + +impl RxChannelState for ChannelState { + const SUB_CHANNELS: usize = 1; + + fn wptr(raw: &Self::Raw<'_>, _index: usize) -> u32 { + raw.write_ptr.load(Ordering::Acquire) + } + + fn set_rptr(raw: &Self::Raw<'_>, _index: usize, rptr: u32) { + raw.read_ptr.store(rptr, Ordering::Release); + } +} + +#[derive(Debug, Default)] +pub(crate) struct FwLogChannelState {} + +impl GpuStruct for FwLogChannelState { + type Raw<'a> = Array<6, raw::ChannelState<'a>>; +} + +impl RxChannelState for FwLogChannelState { + const SUB_CHANNELS: usize = 6; + + fn wptr(raw: &Self::Raw<'_>, index: usize) -> u32 { + raw[index].write_ptr.load(Ordering::Acquire) + } + + fn set_rptr(raw: &Self::Raw<'_>, index: usize, rptr: u32) { + raw[index].read_ptr.store(rptr, Ordering::Release); + } +} + +#[derive(Debug, Default)] +pub(crate) struct FwCtlChannelState {} + +impl GpuStruct for FwCtlChannelState { + type Raw<'a> = raw::FwCtlChannelState<'a>; +} + +pub(crate) trait TxChannelState: GpuStruct + Debug + Default { + fn rptr(raw: &Self::Raw<'_>) -> u32; + fn set_wptr(raw: &Self::Raw<'_>, wptr: u32); +} + +impl TxChannelState for ChannelState { + fn rptr(raw: &Self::Raw<'_>) -> u32 { + raw.read_ptr.load(Ordering::Acquire) + } + + fn set_wptr(raw: &Self::Raw<'_>, wptr: u32) { + raw.write_ptr.store(wptr, Ordering::Release); + } +} + +impl TxChannelState for FwCtlChannelState { + fn rptr(raw: &Self::Raw<'_>) -> u32 { + raw.read_ptr.load(Ordering::Acquire) + } + + fn set_wptr(raw: &Self::Raw<'_>, wptr: u32) { + raw.write_ptr.store(wptr, Ordering::Release); + } +} + +#[derive(Debug, Copy, Clone, Default)] +#[repr(u32)] +pub(crate) enum PipeType { + #[default] + Vertex = 0, + Fragment = 1, + Compute = 2, +} + +#[versions(AGX)] +#[derive(Debug, Copy, Clone, Default)] +#[repr(C)] +pub(crate) struct RunWorkQueueMsg { + pub(crate) pipe_type: PipeType, + pub(crate) work_queue: Option>, + pub(crate) wptr: u32, + pub(crate) event_slot: u32, + pub(crate) is_new: bool, + #[ver(V >= V13_2 && G >= G14)] + pub(crate) __pad: Pad<0x2b>, + #[ver(V < V13_2 || G < G14)] + pub(crate) __pad: Pad<0x1b>, +} + +#[versions(AGX)] +pub(crate) type PipeMsg = RunWorkQueueMsg::ver; + +#[versions(AGX)] +pub(crate) const DEVICECONTROL_SZ: usize = { + #[ver(V < V13_2 || G < G14)] + { + 0x2c + } + #[ver(V >= V13_2 && G >= G14)] + { + 0x3c + } +}; + +// TODO: clean up when arbitrary_enum_discriminant is stable +// https://github.com/rust-lang/rust/issues/60553 + +#[versions(AGX)] +#[derive(Debug, Copy, Clone)] +#[repr(C, u32)] +#[allow(dead_code)] +pub(crate) enum DeviceControlMsg { + Unk00(Array), + Unk01(Array), + Unk02(Array), + Unk03(Array), + Unk04(Array), + Unk05(Array), + Unk06(Array), + Unk07(Array), + Unk08(Array), + Unk09(Array), + Unk0a(Array), + Unk0b(Array), + Unk0c(Array), + Unk0d(Array), + Unk0e(Array), + Unk0f(Array), + Unk10(Array), + Unk11(Array), + Unk12(Array), + Unk13(Array), + Unk14(Array), + Unk15(Array), + Unk16(Array), + DestroyContext { + unk_4: u32, + ctx_23: u8, + __pad0: Pad<3>, + unk_c: u32, + unk_10: u32, + ctx_0: u8, + ctx_1: u8, + ctx_4: u8, + __pad1: Pad<1>, + unk_18: u32, + gpu_context: Option>, + __pad2: Pad<{ DEVICECONTROL_SZ::ver - 0x20 }>, + }, + Unk18(Array), + Initialize(Pad), +} + +#[versions(AGX)] +default_zeroed!(DeviceControlMsg::ver); + +#[derive(Copy, Clone, Default, Debug)] +#[repr(C)] +#[allow(dead_code)] +pub(crate) struct FwCtlMsg { + pub(crate) addr: U64, + pub(crate) unk_8: u32, + pub(crate) slot: u32, + pub(crate) page_count: u16, + pub(crate) unk_12: u16, +} + +pub(crate) const EVENT_SZ: usize = 0x34; + +#[derive(Debug, Copy, Clone)] +#[repr(C, u32)] +#[allow(dead_code)] +pub(crate) enum EventMsg { + Fault, + Flag { + firing: [u32; 4], + unk_14: u16, + }, + Unk2(Array), + Unk3(Array), + Timeout { + counter: u32, + unk_8: u32, + event_slot: u32, + }, // Max discriminant: 0x4 +} + +pub(crate) const EVENT_MAX: u32 = 0x4; + +#[derive(Copy, Clone)] +#[repr(C)] +pub(crate) union RawEventMsg { + pub(crate) raw: (u32, Array), + pub(crate) msg: EventMsg, +} + +default_zeroed!(RawEventMsg); + +#[derive(Debug, Copy, Clone, Default)] +#[repr(C)] +pub(crate) struct RawFwLogMsg { + pub(crate) msg_type: u32, + __pad0: u32, + pub(crate) msg_index: U64, + __pad1: Pad<0x28>, +} + +#[derive(Debug, Copy, Clone, Default)] +#[repr(C)] +pub(crate) struct RawFwLogPayloadMsg { + pub(crate) msg_type: u32, + pub(crate) seq_no: u32, + pub(crate) timestamp: U64, + pub(crate) msg: Array<0xc8, u8>, +} + +#[derive(Debug, Copy, Clone, Default)] +#[repr(C)] +pub(crate) struct RawKTraceMsg { + pub(crate) msg_type: u32, + pub(crate) timestamp: U64, + pub(crate) args: Array<4, U64>, + pub(crate) code: u8, + pub(crate) channel: u8, + __pad: Pad<1>, + pub(crate) thread: u8, + pub(crate) unk_flag: U64, +} + +#[versions(AGX)] +pub(crate) const STATS_SZ: usize = { + #[ver(V < V13_0B4)] + { + 0x2c + } + #[ver(V >= V13_0B4)] + { + 0x3c + } +}; + +#[versions(AGX)] +#[derive(Debug, Copy, Clone)] +#[repr(C, u32)] +#[allow(dead_code)] +pub(crate) enum StatsMsg { + Power { + // 0x00 + __pad: Pad<0x18>, + power: U64, + }, + Unk1(Array<{ STATS_SZ::ver }, u8>), + PowerOn { + // 0x02 + off_time: U64, + }, + PowerOff { + // 0x03 + on_time: U64, + }, + Utilization { + // 0x04 + timestamp: U64, + util1: u32, + util2: u32, + util3: u32, + util4: u32, + }, + Unk5(Array<{ STATS_SZ::ver }, u8>), + Unk6(Array<{ STATS_SZ::ver }, u8>), + Unk7(Array<{ STATS_SZ::ver }, u8>), + Unk8(Array<{ STATS_SZ::ver }, u8>), + AvgPower { + // 0x09 + active_cs: U64, + unk2: u32, + unk3: u32, + unk4: u32, + avg_power: u32, + }, + Temperature { + // 0x0a + __pad: Pad<0x8>, + raw_value: u32, + scale: u32, + tmin: u32, + tmax: u32, + }, + PowerState { + // 0x0b + timestamp: U64, + last_busy_ts: U64, + active: u32, + poweroff: u32, + unk1: u32, + pstate: u32, + unk2: u32, + unk3: u32, + }, + FwBusy { + // 0x0c + timestamp: U64, + busy: u32, + }, + PState { + // 0x0d + __pad: Pad<0x8>, + ps_min: u32, + unk1: u32, + ps_max: u32, + unk2: u32, + }, + TempSensor { + // 0x0e + __pad: Pad<0x4>, + sensor_id: u32, + raw_value: u32, + scale: u32, + tmin: u32, + tmax: u32, + }, // Max discriminant: 0xe +} + +#[versions(AGX)] +pub(crate) const STATS_MAX: u32 = 0xe; + +#[versions(AGX)] +#[derive(Copy, Clone)] +#[repr(C)] +pub(crate) union RawStatsMsg { + pub(crate) raw: (u32, Array<{ STATS_SZ::ver }, u8>), + pub(crate) msg: StatsMsg::ver, +} + +#[versions(AGX)] +default_zeroed!(RawStatsMsg::ver); diff --git a/drivers/gpu/drm/asahi/fw/compute.rs b/drivers/gpu/drm/asahi/fw/compute.rs new file mode 100644 index 000000000000..0dbcd77c5e3e --- /dev/null +++ b/drivers/gpu/drm/asahi/fw/compute.rs @@ -0,0 +1,107 @@ +// SPDX-License-Identifier: GPL-2.0-only OR MIT + +//! GPU compute job firmware structures + +use super::types::*; +use super::{event, job, workqueue}; +use crate::{microseq, mmu}; +use kernel::sync::Arc; + +pub(crate) mod raw { + use super::*; + + #[derive(Debug)] + #[repr(C)] + pub(crate) struct JobParameters1<'a> { + pub(crate) preempt_buf1: GpuPointer<'a, &'a [u8]>, + pub(crate) encoder: U64, + pub(crate) preempt_buf2: GpuPointer<'a, &'a [u8]>, + pub(crate) preempt_buf3: GpuPointer<'a, &'a [u8]>, + pub(crate) preempt_buf4: GpuPointer<'a, &'a [u8]>, + pub(crate) preempt_buf5: GpuPointer<'a, &'a [u8]>, + pub(crate) pipeline_base: U64, + pub(crate) unk_38: U64, + pub(crate) unk_40: u32, + pub(crate) unk_44: u32, + pub(crate) compute_layout_addr: U64, + pub(crate) unk_50: u32, + pub(crate) unk_54: u32, + pub(crate) unk_58: u32, + pub(crate) unk_5c: u32, + pub(crate) iogpu_unk_40: u32, + } + + #[versions(AGX)] + #[derive(Debug)] + #[repr(C)] + pub(crate) struct JobParameters2<'a> { + #[ver(V >= V13_0B4)] + pub(crate) unk_0_0: u32, + pub(crate) unk_0: Array<0x24, u8>, + pub(crate) preempt_buf1: GpuPointer<'a, &'a [u8]>, + pub(crate) encoder_end: U64, + pub(crate) unk_34: Array<0x28, u8>, + #[ver(V < V13_0B4)] + pub(crate) unk_5c: u32, + } + + #[versions(AGX)] + #[derive(Debug)] + #[repr(C)] + pub(crate) struct RunCompute<'a> { + pub(crate) tag: workqueue::CommandType, + + #[ver(V >= V13_0B4)] + pub(crate) counter: U64, + + pub(crate) unk_4: u32, + pub(crate) vm_slot: u32, + pub(crate) notifier: GpuPointer<'a, event::Notifier::ver>, + pub(crate) unk_pointee: Array<0x54, u8>, + pub(crate) job_params1: JobParameters1<'a>, + pub(crate) unk_b8: Array<0x11c, u8>, + pub(crate) microsequence: GpuPointer<'a, &'a [u8]>, + pub(crate) microsequence_size: u32, + pub(crate) job_params2: JobParameters2::ver<'a>, + pub(crate) encoder_params: job::raw::EncoderParams<'a>, + pub(crate) meta: job::raw::JobMeta, + pub(crate) cur_ts: U64, + pub(crate) start_ts: Option>, + pub(crate) end_ts: Option>, + pub(crate) unk_2c0: u32, + pub(crate) unk_2c4: u32, + pub(crate) unk_2c8: u32, + pub(crate) unk_2cc: u32, + pub(crate) client_sequence: u8, + pub(crate) pad_2d1: Array<3, u8>, + pub(crate) unk_2d4: u32, + pub(crate) unk_2d8: u8, + #[ver(V >= V13_0B4)] + pub(crate) unk_ts: U64, + #[ver(V >= V13_0B4)] + pub(crate) unk_2e1: Array<0x1c, u8>, + #[ver(V >= V13_0B4)] + pub(crate) unk_flag: U32, + #[ver(V >= V13_0B4)] + pub(crate) unk_pad: Array<0x10, u8>, + } +} + +#[versions(AGX)] +#[derive(Debug)] +pub(crate) struct RunCompute { + pub(crate) notifier: Arc>, + pub(crate) preempt_buf: GpuArray, + pub(crate) seq_buf: GpuArray, + pub(crate) micro_seq: microseq::MicroSequence, + pub(crate) vm_bind: mmu::VmBind, + pub(crate) timestamps: Arc>, +} + +#[versions(AGX)] +impl GpuStruct for RunCompute::ver { + type Raw<'a> = raw::RunCompute::ver<'a>; +} + +#[versions(AGX)] +impl workqueue::Command for RunCompute::ver {} diff --git a/drivers/gpu/drm/asahi/fw/event.rs b/drivers/gpu/drm/asahi/fw/event.rs new file mode 100644 index 000000000000..fbf65ab6d976 --- /dev/null +++ b/drivers/gpu/drm/asahi/fw/event.rs @@ -0,0 +1,100 @@ +// SPDX-License-Identifier: GPL-2.0-only OR MIT + +//! GPU events control structures & stamps + +use super::types::*; +use crate::{default_zeroed, trivial_gpustruct}; +use core::sync::atomic::Ordering; + +pub(crate) mod raw { + use super::*; + + #[derive(Debug, Clone, Copy, Default)] + #[repr(C)] + pub(crate) struct LinkedListHead { + pub(crate) prev: Option>, + pub(crate) next: Option>, + } + + #[derive(Debug, Clone, Copy)] + #[repr(C)] + pub(crate) struct NotifierList { + pub(crate) list_head: LinkedListHead, + pub(crate) unkptr_10: U64, + } + default_zeroed!(NotifierList); + + #[versions(AGX)] + #[derive(Debug, Clone, Copy)] + #[repr(C)] + pub(crate) struct NotifierState { + unk_14: u32, + unk_18: U64, + unk_20: u32, + vm_slot: u32, + has_vtx: u32, + pstamp_vtx: Array<4, U64>, + has_frag: u32, + pstamp_frag: Array<4, U64>, + has_comp: u32, + pstamp_comp: Array<4, U64>, + #[ver(G >= G14 && V < V13_0B4)] + unk_98_g14_0: Array<0x14, u8>, + in_list: u32, + list_head: LinkedListHead, + #[ver(G >= G14 && V < V13_0B4)] + unk_a8_g14_0: Pad<4>, + #[ver(V >= V13_0B4)] + pub(crate) unk_buf: Array<0x8, u8>, // Init to all-ff + } + + #[versions(AGX)] + impl Default for NotifierState::ver { + fn default() -> Self { + #[allow(unused_mut)] + let mut s: Self = unsafe { core::mem::zeroed() }; + #[ver(V >= V13_0B4)] + s.unk_buf = Array::new([0xff; 0x8]); + s + } + } + + #[derive(Debug)] + #[repr(transparent)] + pub(crate) struct Threshold(AtomicU64); + default_zeroed!(Threshold); + + impl Threshold { + pub(crate) fn increment(&self) { + // We could use fetch_add, but the non-LSE atomic + // sequence Rust produces confuses the hypervisor. + let v = self.0.load(Ordering::Relaxed); + self.0.store(v + 1, Ordering::Relaxed); + } + } + + #[versions(AGX)] + #[derive(Debug)] + #[repr(C)] + pub(crate) struct Notifier<'a> { + pub(crate) threshold: GpuPointer<'a, super::Threshold>, + pub(crate) generation: AtomicU32, + pub(crate) cur_count: AtomicU32, + pub(crate) unk_10: AtomicU32, + pub(crate) state: NotifierState::ver, + } +} + +trivial_gpustruct!(Threshold); +trivial_gpustruct!(NotifierList); + +#[versions(AGX)] +#[derive(Debug)] +pub(crate) struct Notifier { + pub(crate) threshold: GpuObject, +} + +#[versions(AGX)] +impl GpuStruct for Notifier::ver { + type Raw<'a> = raw::Notifier::ver<'a>; +} diff --git a/drivers/gpu/drm/asahi/fw/fragment.rs b/drivers/gpu/drm/asahi/fw/fragment.rs new file mode 100644 index 000000000000..eca275efb967 --- /dev/null +++ b/drivers/gpu/drm/asahi/fw/fragment.rs @@ -0,0 +1,276 @@ +// SPDX-License-Identifier: GPL-2.0-only OR MIT + +//! GPU fragment job firmware structures + +use super::types::*; +use super::{event, job, workqueue}; +use crate::{buffer, fw, microseq, mmu}; +use kernel::sync::Arc; + +pub(crate) mod raw { + use super::*; + + #[derive(Debug, Clone, Copy)] + #[repr(C)] + pub(crate) struct ClearPipelineBinding { + pub(crate) pipeline_bind: U64, + pub(crate) address: U64, + } + + #[derive(Debug, Clone, Copy, Default)] + #[repr(C)] + pub(crate) struct StorePipelineBinding { + pub(crate) unk_0: U64, + pub(crate) unk_8: u32, + pub(crate) pipeline_bind: u32, + pub(crate) unk_10: u32, + pub(crate) address: u32, + pub(crate) unk_18: u32, + pub(crate) unk_1c_padding: u32, + } + + impl StorePipelineBinding { + pub(crate) fn new(pipeline_bind: u32, address: u32) -> StorePipelineBinding { + StorePipelineBinding { + pipeline_bind, + address, + ..Default::default() + } + } + } + + #[derive(Debug)] + #[repr(C)] + pub(crate) struct ArrayAddr { + pub(crate) ptr: U64, + pub(crate) unk_padding: U64, + } + + #[versions(AGX)] + #[derive(Debug, Clone, Copy)] + #[repr(C)] + pub(crate) struct AuxFBInfo { + pub(crate) iogpu_unk_214: u32, + pub(crate) unk2: u32, + pub(crate) width: u32, + pub(crate) height: u32, + + #[ver(V >= V13_0B4)] + pub(crate) unk3: U64, + } + + #[versions(AGX)] + #[derive(Debug)] + #[repr(C)] + pub(crate) struct JobParameters1<'a> { + pub(crate) utile_config: u32, + pub(crate) unk_4: u32, + pub(crate) clear_pipeline: ClearPipelineBinding, + pub(crate) ppp_multisamplectl: U64, + pub(crate) scissor_array: U64, + pub(crate) depth_bias_array: U64, + pub(crate) aux_fb_info: AuxFBInfo::ver, + pub(crate) depth_dimensions: U64, + pub(crate) visibility_result_buffer: U64, + pub(crate) zls_ctrl: U64, + + #[ver(G >= G14)] + pub(crate) unk_58_g14_0: U64, + #[ver(G >= G14)] + pub(crate) unk_58_g14_8: U64, + + pub(crate) depth_buffer_ptr1: U64, + pub(crate) depth_buffer_ptr2: U64, + pub(crate) stencil_buffer_ptr1: U64, + pub(crate) stencil_buffer_ptr2: U64, + + #[ver(G >= G14)] + pub(crate) unk_68_g14_0: Array<0x20, u8>, + + pub(crate) unk_78: Array<0x4, U64>, + pub(crate) depth_meta_buffer_ptr1: U64, + pub(crate) unk_a0: U64, + pub(crate) depth_meta_buffer_ptr2: U64, + pub(crate) unk_b0: U64, + pub(crate) stencil_meta_buffer_ptr1: U64, + pub(crate) unk_c0: U64, + pub(crate) stencil_meta_buffer_ptr2: U64, + pub(crate) unk_d0: U64, + pub(crate) tvb_tilemap: GpuPointer<'a, &'a [u8]>, + pub(crate) tvb_heapmeta: GpuPointer<'a, &'a [u8]>, + pub(crate) mtile_stride_dwords: U64, + pub(crate) tvb_heapmeta_2: GpuPointer<'a, &'a [u8]>, + pub(crate) tile_config: U64, + pub(crate) aux_fb: GpuPointer<'a, &'a [u8]>, + pub(crate) unk_108: Array<0x6, U64>, + pub(crate) pipeline_base: U64, + pub(crate) unk_140: U64, + pub(crate) unk_148: U64, + pub(crate) unk_150: U64, + pub(crate) unk_158: U64, + pub(crate) unk_160: U64, + + #[ver(G < G14)] + pub(crate) unk_168_padding: Array<0x1d8, u8>, + #[ver(G >= G14)] + pub(crate) unk_168_padding: Array<0x1a8, u8>, + #[ver(V < V13_0B4)] + pub(crate) __pad0: Pad<0x8>, + } + + #[derive(Debug)] + #[repr(C)] + pub(crate) struct JobParameters2 { + pub(crate) store_pipeline_bind: u32, + pub(crate) store_pipeline_addr: u32, + pub(crate) unk_8: u32, + pub(crate) unk_c: u32, + pub(crate) merge_upper_x: F32, + pub(crate) merge_upper_y: F32, + pub(crate) unk_18: U64, + pub(crate) utiles_per_mtile_y: u16, + pub(crate) utiles_per_mtile_x: u16, + pub(crate) unk_24: u32, + pub(crate) tile_counts: u32, + pub(crate) iogpu_unk_212: u32, + pub(crate) isp_bgobjdepth: u32, + pub(crate) isp_bgobjvals: u32, + pub(crate) unk_38: u32, + pub(crate) unk_3c: u32, + pub(crate) unk_40: u32, + } + + #[versions(AGX)] + #[derive(Debug)] + #[repr(C)] + pub(crate) struct JobParameters3 { + pub(crate) unk_44_padding: Array<0xac, u8>, + pub(crate) depth_bias_array: ArrayAddr, + pub(crate) scissor_array: ArrayAddr, + pub(crate) visibility_result_buffer: U64, + pub(crate) unk_118: U64, + pub(crate) unk_120: Array<0x25, U64>, + pub(crate) unk_reload_pipeline: ClearPipelineBinding, + pub(crate) unk_258: U64, + pub(crate) unk_260: U64, + pub(crate) unk_268: U64, + pub(crate) unk_270: U64, + pub(crate) reload_pipeline: ClearPipelineBinding, + pub(crate) zls_ctrl: U64, + pub(crate) unk_290: U64, + pub(crate) depth_buffer_ptr1: U64, + pub(crate) unk_2a0: U64, + pub(crate) unk_2a8: U64, + pub(crate) depth_buffer_ptr2: U64, + pub(crate) depth_buffer_ptr3: U64, + pub(crate) depth_meta_buffer_ptr3: U64, + pub(crate) stencil_buffer_ptr1: U64, + pub(crate) unk_2d0: U64, + pub(crate) unk_2d8: U64, + pub(crate) stencil_buffer_ptr2: U64, + pub(crate) stencil_buffer_ptr3: U64, + pub(crate) stencil_meta_buffer_ptr3: U64, + pub(crate) unk_2f8: Array<2, U64>, + pub(crate) iogpu_unk_212: u32, + pub(crate) unk_30c: u32, + pub(crate) aux_fb_info: AuxFBInfo::ver, + pub(crate) unk_320_padding: Array<0x10, u8>, + pub(crate) unk_partial_store_pipeline: StorePipelineBinding, + pub(crate) partial_store_pipeline: StorePipelineBinding, + pub(crate) isp_bgobjdepth: u32, + pub(crate) isp_bgobjvals: u32, + pub(crate) iogpu_unk_49: u32, + pub(crate) unk_37c: u32, + pub(crate) unk_380: U64, + pub(crate) unk_388: U64, + + #[ver(V >= V13_0B4)] + pub(crate) unk_390_0: U64, + + pub(crate) depth_dimensions: U64, + } + + #[versions(AGX)] + #[derive(Debug)] + #[repr(C)] + pub(crate) struct RunFragment<'a> { + pub(crate) tag: workqueue::CommandType, + + #[ver(V >= V13_0B4)] + pub(crate) counter: U64, + + pub(crate) vm_slot: u32, + pub(crate) unk_8: u32, + pub(crate) microsequence: GpuPointer<'a, &'a [u8]>, + pub(crate) microsequence_size: u32, + pub(crate) notifier: GpuPointer<'a, event::Notifier::ver>, + pub(crate) buffer: GpuPointer<'a, fw::buffer::Info::ver>, + pub(crate) scene: GpuPointer<'a, fw::buffer::Scene::ver>, + pub(crate) unk_buffer_buf: GpuWeakPointer<[u8]>, + pub(crate) tvb_tilemap: GpuPointer<'a, &'a [u8]>, + pub(crate) ppp_multisamplectl: U64, + pub(crate) samples: u32, + pub(crate) tiles_per_mtile_y: u16, + pub(crate) tiles_per_mtile_x: u16, + pub(crate) unk_50: U64, + pub(crate) unk_58: U64, + pub(crate) merge_upper_x: F32, + pub(crate) merge_upper_y: F32, + pub(crate) unk_68: U64, + pub(crate) tile_count: U64, + pub(crate) job_params1: JobParameters1::ver<'a>, + pub(crate) job_params2: JobParameters2, + pub(crate) job_params3: JobParameters3::ver, + pub(crate) unk_758_flag: u32, + pub(crate) unk_75c_flag: u32, + pub(crate) unk_buf: Array<0x110, u8>, + pub(crate) busy_flag: u32, + pub(crate) tvb_overflow_count: u32, + pub(crate) unk_878: u32, + pub(crate) encoder_params: job::raw::EncoderParams<'a>, + pub(crate) process_empty_tiles: u32, + pub(crate) no_clear_pipeline_textures: u32, + pub(crate) unk_param: u32, + pub(crate) unk_pointee: u32, + pub(crate) meta: job::raw::JobMeta, + pub(crate) unk_after_meta: u32, + pub(crate) unk_buf_0: U64, + pub(crate) unk_buf_8: U64, + pub(crate) unk_buf_10: U64, + pub(crate) cur_ts: U64, + pub(crate) start_ts: Option>, + pub(crate) end_ts: Option>, + pub(crate) unk_914: u32, + pub(crate) unk_918: U64, + pub(crate) unk_920: u32, + pub(crate) client_sequence: u8, + pub(crate) pad_925: Array<3, u8>, + pub(crate) unk_928: u32, + pub(crate) unk_92c: u8, + + #[ver(V >= V13_0B4)] + pub(crate) unk_ts: U64, + + #[ver(V >= V13_0B4)] + pub(crate) unk_92d_8: Array<0x1b, u8>, + } +} + +#[versions(AGX)] +#[derive(Debug)] +pub(crate) struct RunFragment { + pub(crate) notifier: Arc>, + pub(crate) scene: Arc, + pub(crate) micro_seq: microseq::MicroSequence, + pub(crate) vm_bind: mmu::VmBind, + pub(crate) aux_fb: GpuArray, + pub(crate) timestamps: Arc>, +} + +#[versions(AGX)] +impl GpuStruct for RunFragment::ver { + type Raw<'a> = raw::RunFragment::ver<'a>; +} + +#[versions(AGX)] +impl workqueue::Command for RunFragment::ver {} diff --git a/drivers/gpu/drm/asahi/fw/initdata.rs b/drivers/gpu/drm/asahi/fw/initdata.rs new file mode 100644 index 000000000000..44de0c1cccf3 --- /dev/null +++ b/drivers/gpu/drm/asahi/fw/initdata.rs @@ -0,0 +1,1264 @@ +// SPDX-License-Identifier: GPL-2.0-only OR MIT + +//! GPU initialization / global structures + +use super::channels; +use super::types::*; +use crate::{default_zeroed, no_debug, trivial_gpustruct}; + +pub(crate) mod raw { + use super::*; + + #[derive(Debug, Default)] + #[repr(C)] + pub(crate) struct ChannelRing { + pub(crate) state: Option>, + pub(crate) ring: Option>, + } + + #[versions(AGX)] + #[derive(Debug)] + #[repr(C)] + pub(crate) struct PipeChannels { + pub(crate) vtx: ChannelRing, + pub(crate) frag: ChannelRing, + pub(crate) comp: ChannelRing, + } + #[versions(AGX)] + default_zeroed!(PipeChannels::ver); + + #[derive(Debug, Default)] + #[repr(C)] + pub(crate) struct FwStatusFlags { + pub(crate) halt_count: AtomicU32, + __pad0: Pad<0xc>, + pub(crate) halted: AtomicU32, + __pad1: Pad<0xc>, + pub(crate) resume: AtomicU32, + __pad2: Pad<0xc>, + pub(crate) unk_40: u32, + __pad3: Pad<0xc>, + pub(crate) unk_ctr: u32, + __pad4: Pad<0xc>, + pub(crate) unk_60: u32, + __pad5: Pad<0xc>, + pub(crate) unk_70: u32, + __pad6: Pad<0xc>, + } + + #[derive(Debug, Default)] + #[repr(C)] + pub(crate) struct FwStatus { + pub(crate) fwctl_channel: ChannelRing, + pub(crate) flags: FwStatusFlags, + } + + #[derive(Debug)] + #[repr(C)] + pub(crate) struct HwDataShared1 { + pub(crate) table: Array<16, i32>, + pub(crate) unk_44: Array<0x60, u8>, + pub(crate) unk_a4: u32, + pub(crate) unk_a8: u32, + } + default_zeroed!(HwDataShared1); + + #[derive(Debug, Default)] + #[repr(C)] + pub(crate) struct HwDataShared2Curve { + pub(crate) unk_0: u32, + pub(crate) unk_4: u32, + pub(crate) t1: Array<16, i16>, + pub(crate) t2: Array<16, i16>, + pub(crate) t3: Array<8, Array<16, i32>>, + } + + #[derive(Debug, Default)] + #[repr(C)] + pub(crate) struct HwDataShared2T8112 { + pub(crate) unk_0: Array<5, u32>, + pub(crate) unk_14: u32, + pub(crate) unk_18: Array<8, u32>, + pub(crate) curve1: HwDataShared2Curve, + pub(crate) curve2: HwDataShared2Curve, + } + + #[derive(Debug)] + #[repr(C)] + pub(crate) struct HwDataShared2 { + pub(crate) table: Array<10, i32>, + pub(crate) unk_28: Array<0x10, u8>, + pub(crate) t8112: HwDataShared2T8112, + pub(crate) unk_500: u32, + pub(crate) unk_504: u32, + pub(crate) unk_508: u32, + pub(crate) unk_50c: u32, + pub(crate) unk_510: u32, + } + default_zeroed!(HwDataShared2); + + #[derive(Debug)] + #[repr(C)] + pub(crate) struct HwDataShared3 { + pub(crate) unk_0: u32, + pub(crate) unk_4: u32, + pub(crate) unk_8: u32, + pub(crate) table: Array<16, u32>, + pub(crate) unk_4c: u32, + } + default_zeroed!(HwDataShared3); + + #[derive(Debug)] + #[repr(C)] + pub(crate) struct HwDataA130Extra { + pub(crate) unk_0: Array<0x38, u8>, + pub(crate) unk_38: u32, + pub(crate) unk_3c: u32, + pub(crate) unk_40: u32, + pub(crate) unk_44: u32, + pub(crate) unk_48: u32, + pub(crate) unk_4c: u32, + pub(crate) unk_50: u32, + pub(crate) unk_54: u32, + pub(crate) unk_58: u32, + pub(crate) unk_5c: u32, + pub(crate) unk_60: F32, + pub(crate) unk_64: F32, + pub(crate) unk_68: F32, + pub(crate) unk_6c: F32, + pub(crate) unk_70: F32, + pub(crate) unk_74: F32, + pub(crate) unk_78: F32, + pub(crate) unk_7c: F32, + pub(crate) unk_80: F32, + pub(crate) unk_84: F32, + pub(crate) unk_88: u32, + pub(crate) unk_8c: u32, + pub(crate) max_pstate_scaled_1: u32, + pub(crate) unk_94: u32, + pub(crate) unk_98: u32, + pub(crate) unk_9c: F32, + pub(crate) unk_a0: u32, + pub(crate) unk_a4: u32, + pub(crate) unk_a8: u32, + pub(crate) unk_ac: u32, + pub(crate) unk_b0: u32, + pub(crate) unk_b4: u32, + pub(crate) unk_b8: u32, + pub(crate) unk_bc: u32, + pub(crate) unk_c0: u32, + pub(crate) unk_c4: F32, + pub(crate) unk_c8: Array<0x4c, u8>, + pub(crate) unk_114: F32, + pub(crate) unk_118: u32, + pub(crate) unk_11c: u32, + pub(crate) unk_120: u32, + pub(crate) unk_124: u32, + pub(crate) max_pstate_scaled_2: u32, + pub(crate) unk_12c: Array<0x8c, u8>, + } + default_zeroed!(HwDataA130Extra); + + #[derive(Default)] + #[repr(C)] + pub(crate) struct T81xxData { + pub(crate) unk_d8c: u32, + pub(crate) unk_d90: u32, + pub(crate) unk_d94: u32, + pub(crate) unk_d98: u32, + pub(crate) unk_d9c: F32, + pub(crate) unk_da0: u32, + pub(crate) unk_da4: F32, + pub(crate) unk_da8: u32, + pub(crate) unk_dac: F32, + pub(crate) unk_db0: u32, + pub(crate) unk_db4: u32, + pub(crate) unk_db8: F32, + pub(crate) unk_dbc: F32, + pub(crate) unk_dc0: u32, + pub(crate) unk_dc4: u32, + pub(crate) unk_dc8: u32, + pub(crate) max_pstate_scaled: u32, + } + + #[versions(AGX)] + #[derive(Default, Copy, Clone)] + #[repr(C)] + pub(crate) struct PowerZone { + pub(crate) val: F32, + pub(crate) target: u32, + pub(crate) target_off: u32, + pub(crate) filter_tc_x4: u32, + pub(crate) filter_tc_xperiod: u32, + #[ver(V >= V13_0B4)] + pub(crate) unk_10: u32, + #[ver(V >= V13_0B4)] + pub(crate) unk_14: u32, + pub(crate) filter_a_neg: F32, + pub(crate) filter_a: F32, + pub(crate) pad: u32, + } + + #[versions(AGX)] + #[repr(C)] + pub(crate) struct HwDataA { + pub(crate) unk_0: u32, + pub(crate) clocks_per_period: u32, + + #[ver(V >= V13_0B4)] + pub(crate) clocks_per_period_2: u32, + + pub(crate) unk_8: u32, + pub(crate) pwr_status: AtomicU32, + pub(crate) unk_10: F32, + pub(crate) unk_14: u32, + pub(crate) unk_18: u32, + pub(crate) unk_1c: u32, + pub(crate) unk_20: u32, + pub(crate) unk_24: u32, + pub(crate) actual_pstate: u32, + pub(crate) tgt_pstate: u32, + pub(crate) unk_30: u32, + pub(crate) cur_pstate: u32, + pub(crate) unk_38: u32, + + #[ver(V >= V13_0B4)] + pub(crate) unk_3c_0: u32, + + pub(crate) base_pstate_scaled: u32, + pub(crate) unk_40: u32, + pub(crate) max_pstate_scaled: u32, + pub(crate) unk_48: u32, + pub(crate) min_pstate_scaled: u32, + pub(crate) freq_mhz: F32, + pub(crate) unk_54: Array<0x20, u8>, + + #[ver(V >= V13_0B4)] + pub(crate) unk_74_0: u32, + + pub(crate) sram_k: Array<0x10, F32>, + pub(crate) unk_b4: Array<0x100, u8>, + pub(crate) unk_1b4: u32, + pub(crate) temp_c: u32, + pub(crate) avg_power_mw: u32, + pub(crate) update_ts: U64, + pub(crate) unk_1c8: u32, + pub(crate) unk_1cc: Array<0x478, u8>, + pub(crate) pad_644: Pad<0x8>, + pub(crate) unk_64c: u32, + pub(crate) unk_650: u32, + pub(crate) pad_654: u32, + pub(crate) pwr_filter_a_neg: F32, + pub(crate) pad_65c: u32, + pub(crate) pwr_filter_a: F32, + pub(crate) pad_664: u32, + pub(crate) pwr_integral_gain: F32, + pub(crate) pad_66c: u32, + pub(crate) pwr_integral_min_clamp: F32, + pub(crate) max_power_1: F32, + pub(crate) pwr_proportional_gain: F32, + pub(crate) pad_67c: u32, + pub(crate) pwr_pstate_related_k: F32, + pub(crate) pwr_pstate_max_dc_offset: i32, + pub(crate) unk_688: u32, + pub(crate) max_pstate_scaled_2: u32, + pub(crate) pad_690: u32, + pub(crate) unk_694: u32, + pub(crate) max_power_2: u32, + pub(crate) pad_69c: Pad<0x18>, + pub(crate) unk_6b4: u32, + + #[ver(V >= V13_0B4)] + pub(crate) unk_6b8_0: Array<0x10, u8>, + + pub(crate) max_pstate_scaled_3: u32, + pub(crate) unk_6bc: u32, + pub(crate) pad_6c0: Pad<0x14>, + pub(crate) ppm_filter_tc_periods_x4: u32, + pub(crate) unk_6d8: u32, + pub(crate) pad_6dc: u32, + pub(crate) ppm_filter_a_neg: F32, + pub(crate) pad_6e4: u32, + pub(crate) ppm_filter_a: F32, + pub(crate) pad_6ec: u32, + pub(crate) ppm_ki_dt: F32, + pub(crate) pad_6f4: u32, + pub(crate) pwr_integral_min_clamp_2: u32, + pub(crate) unk_6fc: F32, + pub(crate) ppm_kp: F32, + pub(crate) pad_704: u32, + pub(crate) unk_708: u32, + pub(crate) pwr_min_duty_cycle: u32, + pub(crate) max_pstate_scaled_4: u32, + pub(crate) unk_714: u32, + pub(crate) pad_718: u32, + pub(crate) unk_71c: F32, + pub(crate) max_power_3: u32, + pub(crate) cur_power_mw_2: u32, + pub(crate) ppm_filter_tc_ms: u32, + pub(crate) unk_72c: u32, + + #[ver(V >= V13_0B4)] + pub(crate) ppm_filter_tc_clks: u32, + + #[ver(V >= V13_0B4)] + pub(crate) unk_730_4: u32, + + #[ver(V >= V13_0B4)] + pub(crate) unk_730_8: u32, + + #[ver(V >= V13_0B4)] + pub(crate) unk_730_c: u32, + + pub(crate) unk_730: F32, + pub(crate) unk_734: u32, + pub(crate) unk_738: u32, + pub(crate) unk_73c: u32, + pub(crate) unk_740: u32, + pub(crate) unk_744: u32, + pub(crate) unk_748: Array<0x4, F32>, + pub(crate) unk_758: u32, + pub(crate) perf_tgt_utilization: u32, + pub(crate) pad_760: u32, + pub(crate) perf_boost_min_util: u32, + pub(crate) perf_boost_ce_step: u32, + pub(crate) perf_reset_iters: u32, + pub(crate) pad_770: u32, + pub(crate) unk_774: u32, + pub(crate) unk_778: u32, + pub(crate) perf_filter_drop_threshold: u32, + pub(crate) perf_filter_a_neg: F32, + pub(crate) perf_filter_a2_neg: F32, + pub(crate) perf_filter_a: F32, + pub(crate) perf_filter_a2: F32, + pub(crate) perf_ki: F32, + pub(crate) perf_ki2: F32, + pub(crate) perf_integral_min_clamp: F32, + pub(crate) unk_79c: F32, + pub(crate) perf_kp: F32, + pub(crate) perf_kp2: F32, + pub(crate) boost_state_unk_k: F32, + pub(crate) base_pstate_scaled_2: u32, + pub(crate) max_pstate_scaled_5: u32, + pub(crate) base_pstate_scaled_3: u32, + pub(crate) pad_7b8: u32, + pub(crate) perf_cur_utilization: F32, + pub(crate) perf_tgt_utilization_2: u32, + pub(crate) pad_7c4: Pad<0x18>, + pub(crate) unk_7dc: u32, + + #[ver(V >= V13_0B4)] + pub(crate) unk_7e0_0: Array<0x10, u8>, + + pub(crate) base_pstate_scaled_4: u32, + pub(crate) pad_7e4: u32, + pub(crate) unk_7e8: Array<0x14, u8>, + pub(crate) unk_7fc: F32, + pub(crate) pwr_min_duty_cycle_2: F32, + pub(crate) max_pstate_scaled_6: F32, + pub(crate) max_freq_mhz: u32, + pub(crate) pad_80c: u32, + pub(crate) unk_810: u32, + pub(crate) pad_814: u32, + pub(crate) pwr_min_duty_cycle_3: u32, + pub(crate) unk_81c: u32, + pub(crate) pad_820: u32, + pub(crate) min_pstate_scaled_4: F32, + pub(crate) max_pstate_scaled_7: u32, + pub(crate) unk_82c: u32, + pub(crate) unk_alpha_neg: F32, + pub(crate) unk_alpha: F32, + pub(crate) unk_838: u32, + pub(crate) unk_83c: u32, + pub(crate) pad_840: Pad<0x2c>, + pub(crate) unk_86c: u32, + pub(crate) fast_die0_sensor_mask: U64, + pub(crate) fast_die0_release_temp_cc: u32, + pub(crate) unk_87c: i32, + pub(crate) unk_880: u32, + pub(crate) unk_884: u32, + pub(crate) pad_888: u32, + pub(crate) unk_88c: u32, + pub(crate) pad_890: u32, + pub(crate) unk_894: F32, + pub(crate) pad_898: u32, + pub(crate) fast_die0_ki_dt: F32, + pub(crate) pad_8a0: u32, + pub(crate) unk_8a4: u32, + pub(crate) unk_8a8: F32, + pub(crate) fast_die0_kp: F32, + pub(crate) pad_8b0: u32, + pub(crate) unk_8b4: u32, + pub(crate) pwr_min_duty_cycle_4: u32, + pub(crate) max_pstate_scaled_8: u32, + pub(crate) max_pstate_scaled_9: u32, + pub(crate) fast_die0_prop_tgt_delta: u32, + pub(crate) unk_8c8: u32, + pub(crate) unk_8cc: u32, + pub(crate) pad_8d0: Pad<0x14>, + + #[ver(V >= V13_0B4)] + pub(crate) unk_8e4_0: Array<0x10, u8>, + + pub(crate) unk_8e4: u32, + pub(crate) unk_8e8: u32, + pub(crate) max_pstate_scaled_10: u32, + pub(crate) unk_8f0: u32, + pub(crate) unk_8f4: u32, + pub(crate) pad_8f8: u32, + pub(crate) pad_8fc: u32, + pub(crate) unk_900: Array<0x24, u8>, + pub(crate) unk_coef_a1: Array<8, Array<8, F32>>, + pub(crate) unk_coef_a2: Array<8, Array<8, F32>>, + pub(crate) pad_b24: Pad<0x70>, + pub(crate) max_pstate_scaled_11: u32, + pub(crate) freq_with_off: u32, + pub(crate) unk_b9c: u32, + pub(crate) unk_ba0: U64, + pub(crate) unk_ba8: U64, + pub(crate) unk_bb0: u32, + pub(crate) unk_bb4: u32, + pub(crate) pad_bb8: Pad<0x74>, + pub(crate) unk_c2c: u32, + pub(crate) power_zone_count: u32, + pub(crate) max_power_4: u32, + pub(crate) max_power_5: u32, + pub(crate) max_power_6: u32, + pub(crate) unk_c40: u32, + pub(crate) unk_c44: F32, + pub(crate) avg_power_target_filter_a_neg: F32, + pub(crate) avg_power_target_filter_a: F32, + pub(crate) avg_power_target_filter_tc_x4: u32, + pub(crate) avg_power_target_filter_tc_xperiod: u32, + + #[ver(V >= V13_0B4)] + pub(crate) avg_power_target_filter_tc_clks: u32, + + #[ver(V >= V13_0B4)] + pub(crate) unk_c58_4: u32, + + pub(crate) power_zones: Array<5, PowerZone::ver>, + pub(crate) avg_power_filter_tc_periods_x4: u32, + pub(crate) unk_cfc: u32, + pub(crate) unk_d00: u32, + pub(crate) avg_power_filter_a_neg: F32, + pub(crate) unk_d08: u32, + pub(crate) avg_power_filter_a: F32, + pub(crate) unk_d10: u32, + pub(crate) avg_power_ki_dt: F32, + pub(crate) unk_d18: u32, + pub(crate) unk_d1c: u32, + pub(crate) unk_d20: F32, + pub(crate) avg_power_kp: F32, + pub(crate) unk_d28: u32, + pub(crate) unk_d2c: u32, + pub(crate) avg_power_min_duty_cycle: u32, + pub(crate) max_pstate_scaled_12: u32, + pub(crate) max_pstate_scaled_13: u32, + pub(crate) unk_d3c: u32, + pub(crate) max_power_7: F32, + pub(crate) max_power_8: u32, + pub(crate) unk_d48: u32, + pub(crate) avg_power_filter_tc_ms: u32, + pub(crate) unk_d50: u32, + + #[ver(V >= V13_0B4)] + pub(crate) avg_power_filter_tc_clks: u32, + + #[ver(V >= V13_0B4)] + pub(crate) unk_d54_4: Array<0xc, u8>, + + pub(crate) unk_d54: Array<0x10, u8>, + pub(crate) max_pstate_scaled_14: u32, + pub(crate) unk_d68: Array<0x24, u8>, + + pub(crate) t81xx_data: T81xxData, + + pub(crate) unk_dd0: Array<0x40, u8>, + + #[ver(V >= V13_2)] + pub(crate) unk_e10_pad: Array<0x10, u8>, + + #[ver(V >= V13_0B4)] + pub(crate) unk_e10_0: HwDataA130Extra, + + pub(crate) unk_e10: Array<0xc, u8>, + pub(crate) fast_die0_sensor_mask_2: U64, + pub(crate) unk_e24: u32, + pub(crate) unk_e28: u32, + pub(crate) unk_e2c: Pad<0x1c>, + pub(crate) unk_coef_b1: Array<8, Array<8, F32>>, + pub(crate) unk_coef_b2: Array<8, Array<8, F32>>, + pub(crate) pad_1048: Pad<0x5e4>, + pub(crate) fast_die0_sensor_mask_alt: U64, + #[ver(V < V13_0B4)] + pub(crate) fast_die0_sensor_present: U64, + + pub(crate) unk_163c: u32, + + pub(crate) unk_1640: Array<0x2000, u8>, + pub(crate) unk_3640: u32, + pub(crate) unk_3644: u32, + pub(crate) hws1: HwDataShared1, + + #[ver(V >= V13_0B4)] + pub(crate) unk_pad1: Pad<0x20>, + + pub(crate) hws2: HwDataShared2, + pub(crate) unk_3c04: u32, + pub(crate) hws3: HwDataShared3, + pub(crate) unk_3c58: Array<0x3c, u8>, + pub(crate) unk_3c94: u32, + pub(crate) unk_3c98: U64, + pub(crate) unk_3ca0: U64, + pub(crate) unk_3ca8: U64, + pub(crate) unk_3cb0: U64, + pub(crate) ts_last_idle: U64, + pub(crate) ts_last_poweron: U64, + pub(crate) ts_last_poweroff: U64, + pub(crate) unk_3cd0: U64, + pub(crate) unk_3cd8: U64, + + #[ver(V >= V13_0B4)] + pub(crate) unk_3ce0_0: u32, + + pub(crate) unk_3ce0: u32, + pub(crate) unk_3ce4: u32, + pub(crate) unk_3ce8: u32, + pub(crate) unk_3cec: u32, + pub(crate) unk_3cf0: u32, + pub(crate) core_leak_coef: Array<8, F32>, + pub(crate) sram_leak_coef: Array<8, F32>, + pub(crate) unk_3d34: Array<0x38, u8>, + + #[ver(V >= V13_0B4)] + pub(crate) unk_3d6c: Array<0x38, u8>, + } + #[versions(AGX)] + default_zeroed!(HwDataA::ver); + #[versions(AGX)] + no_debug!(HwDataA::ver); + + #[derive(Debug, Default, Clone, Copy)] + #[repr(C)] + pub(crate) struct IOMapping { + pub(crate) phys_addr: U64, + pub(crate) virt_addr: U64, + pub(crate) size: u32, + pub(crate) range_size: u32, + pub(crate) readwrite: U64, + } + + #[versions(AGX)] + #[derive(Debug)] + #[repr(C)] + pub(crate) struct HwDataB { + #[ver(V < V13_0B4)] + pub(crate) unk_0: U64, + + pub(crate) unk_8: U64, + + #[ver(V < V13_0B4)] + pub(crate) unk_10: U64, + + pub(crate) unk_18: U64, + pub(crate) unk_20: U64, + pub(crate) unk_28: U64, + pub(crate) unk_30: U64, + pub(crate) unkptr_38: U64, + pub(crate) pad_40: Pad<0x20>, + + #[ver(V < V13_0B4)] + pub(crate) yuv_matrices: Array<0xf, Array<3, Array<4, i16>>>, + + #[ver(V >= V13_0B4)] + pub(crate) yuv_matrices: Array<0x3f, Array<3, Array<4, i16>>>, + + pub(crate) pad_1c8: Pad<0x8>, + pub(crate) io_mappings: Array<0x14, IOMapping>, + + #[ver(V >= V13_0B4)] + pub(crate) unk_450_0: Array<0x68, u8>, + + pub(crate) chip_id: u32, + pub(crate) unk_454: u32, + pub(crate) unk_458: u32, + pub(crate) unk_45c: u32, + pub(crate) unk_460: u32, + pub(crate) unk_464: u32, + pub(crate) unk_468: u32, + pub(crate) unk_46c: u32, + pub(crate) unk_470: u32, + pub(crate) unk_474: u32, + pub(crate) unk_478: u32, + pub(crate) unk_47c: u32, + pub(crate) unk_480: u32, + pub(crate) unk_484: u32, + pub(crate) unk_488: u32, + pub(crate) unk_48c: u32, + pub(crate) base_clock_khz: u32, + pub(crate) power_sample_period: u32, + pub(crate) pad_498: Pad<0x4>, + pub(crate) unk_49c: u32, + pub(crate) unk_4a0: u32, + pub(crate) unk_4a4: u32, + pub(crate) pad_4a8: Pad<0x4>, + pub(crate) unk_4ac: u32, + pub(crate) pad_4b0: Pad<0x8>, + pub(crate) unk_4b8: u32, + pub(crate) unk_4bc: Array<0x4, u8>, + pub(crate) unk_4c0: u32, + pub(crate) unk_4c4: u32, + pub(crate) unk_4c8: u32, + pub(crate) unk_4cc: u32, + pub(crate) unk_4d0: u32, + pub(crate) unk_4d4: u32, + pub(crate) unk_4d8: Array<0x4, u8>, + pub(crate) unk_4dc: u32, + pub(crate) unk_4e0: U64, + pub(crate) unk_4e8: u32, + pub(crate) unk_4ec: u32, + pub(crate) unk_4f0: u32, + pub(crate) unk_4f4: u32, + pub(crate) unk_4f8: u32, + pub(crate) unk_4fc: u32, + pub(crate) unk_500: u32, + + #[ver(V >= V13_0B4)] + pub(crate) unk_504_0: u32, + + pub(crate) unk_504: u32, + pub(crate) unk_508: u32, + pub(crate) unk_50c: u32, + pub(crate) unk_510: u32, + pub(crate) unk_514: u32, + pub(crate) unk_518: u32, + pub(crate) unk_51c: u32, + pub(crate) unk_520: u32, + pub(crate) unk_524: u32, + pub(crate) unk_528: u32, + pub(crate) unk_52c: u32, + pub(crate) unk_530: u32, + pub(crate) unk_534: u32, + pub(crate) unk_538: u32, + + #[ver(V >= V13_0B4)] + pub(crate) unk_53c_0: u32, + + pub(crate) num_frags: u32, + pub(crate) unk_540: u32, + pub(crate) unk_544: u32, + pub(crate) unk_548: u32, + pub(crate) unk_54c: u32, + pub(crate) unk_550: u32, + pub(crate) unk_554: u32, + pub(crate) uat_ttb_base: U64, + pub(crate) gpu_core_id: u32, + pub(crate) gpu_rev_id: u32, + pub(crate) num_cores: u32, + pub(crate) max_pstate: u32, + + #[ver(V < V13_0B4)] + pub(crate) num_pstates: u32, + + pub(crate) frequencies: Array<0x10, u32>, + pub(crate) voltages: Array<0x10, [u32; 0x8]>, + pub(crate) voltages_sram: Array<0x10, [u32; 0x8]>, + pub(crate) sram_k: Array<0x10, F32>, + pub(crate) unk_9f4: Array<0x10, u32>, + pub(crate) rel_max_powers: Array<0x10, u32>, + pub(crate) rel_boost_freqs: Array<0x10, u32>, + + #[ver(V < V13_0B4)] + pub(crate) min_sram_volt: u32, + + #[ver(V < V13_0B4)] + pub(crate) unk_ab8: u32, + + #[ver(V < V13_0B4)] + pub(crate) unk_abc: u32, + + #[ver(V < V13_0B4)] + pub(crate) unk_ac0: u32, + + #[ver(V >= V13_0B4)] + pub(crate) unk_ac4_0: Array<0x1f0, u8>, + + pub(crate) pad_ac4: Pad<0x8>, + pub(crate) unk_acc: u32, + pub(crate) unk_ad0: u32, + pub(crate) pad_ad4: Pad<0x10>, + pub(crate) unk_ae4: Array<0x4, u32>, + pub(crate) pad_af4: Pad<0x4>, + pub(crate) unk_af8: u32, + pub(crate) pad_afc: Pad<0x8>, + pub(crate) unk_b04: u32, + pub(crate) unk_b08: u32, + pub(crate) unk_b0c: u32, + pub(crate) unk_b10: u32, + pub(crate) pad_b14: Pad<0x8>, + pub(crate) unk_b1c: u32, + pub(crate) unk_b20: u32, + pub(crate) unk_b24: u32, + pub(crate) unk_b28: u32, + pub(crate) unk_b2c: u32, + pub(crate) unk_b30: u32, + pub(crate) unk_b34: u32, + + #[ver(V >= V13_0B4)] + pub(crate) unk_b38_0: u32, + + #[ver(V >= V13_0B4)] + pub(crate) unk_b38_4: u32, + + pub(crate) unk_b38: Array<0xc, u32>, + pub(crate) unk_b68: u32, + + #[ver(V >= V13_0B4)] + pub(crate) unk_b6c: Array<0xd0, u8>, + + #[ver(V >= V13_0B4)] + pub(crate) unk_c3c: u32, + } + #[versions(AGX)] + default_zeroed!(HwDataB::ver); + + #[derive(Debug, Clone, Copy)] + #[repr(C, packed)] + pub(crate) struct GpuQueueStatsVtx { + pub(crate) busy: u32, + pub(crate) unk_4: u32, + pub(crate) cur_cmdqueue: U64, + pub(crate) cur_count: u32, + pub(crate) unk_14: u32, + } + default_zeroed!(GpuQueueStatsVtx); + + #[versions(AGX)] + #[derive(Debug, Default, Clone, Copy)] + #[repr(C, packed)] + pub(crate) struct GpuStatsVtx { + pub(crate) unk_4: u32, + pub(crate) queues: Array<0x4, GpuQueueStatsVtx>, + pub(crate) unk_68: Array<0x8, u8>, + pub(crate) unk_70: u32, + pub(crate) unk_74: u32, + pub(crate) unk_timestamp: U64, + pub(crate) unk_80: Array<0x40, u8>, + } + + #[derive(Debug, Default, Clone, Copy)] + #[repr(C, packed)] + pub(crate) struct GpuQueueStatsFrag { + pub(crate) busy: u32, + pub(crate) cur_cmdqueue: U64, + pub(crate) unk_c: u32, + pub(crate) unk_10: u32, + pub(crate) unk_14: Array<0x14, u8>, + } + + #[versions(AGX)] + #[derive(Debug)] + #[repr(C)] + pub(crate) struct GpuStatsFrag { + pub(crate) unk_0: Array<0x18, u8>, + pub(crate) queues: Array<0x4, GpuQueueStatsFrag>, + pub(crate) unk_d0: Array<0x38, u8>, + pub(crate) tvb_overflows_1: u32, + pub(crate) tvb_overflows_2: u32, + pub(crate) unk_f8: u32, + pub(crate) unk_fc: u32, + pub(crate) cur_stamp_id: i32, + pub(crate) unk_104: Array<0x14, u8>, + pub(crate) unk_118: i32, + pub(crate) unk_11c: u32, + pub(crate) unk_120: u32, + pub(crate) unk_124: u32, + pub(crate) unk_128: u32, + pub(crate) unk_12c: u32, + pub(crate) unk_timestamp: U64, + pub(crate) unk_134: Array<0x8c, u8>, + } + #[versions(AGX)] + default_zeroed!(GpuStatsFrag::ver); + + #[versions(AGX)] + #[derive(Debug)] + #[repr(C)] + pub(crate) struct GpuGlobalStatsVtx { + pub(crate) total_cmds: u32, + pub(crate) stats: GpuStatsVtx::ver, + #[ver(V >= V13_0B4)] + pub(crate) unk_pad: Array<0x5c4, u8>, + } + #[versions(AGX)] + default_zeroed!(GpuGlobalStatsVtx::ver); + + #[versions(AGX)] + #[derive(Debug)] + #[repr(C)] + pub(crate) struct GpuGlobalStatsFrag { + pub(crate) total_cmds: u32, + pub(crate) unk_4: u32, + pub(crate) stats: GpuStatsFrag::ver, + #[ver(V >= V13_0B4)] + pub(crate) unk_pad: Array<0x580, u8>, + } + #[versions(AGX)] + default_zeroed!(GpuGlobalStatsFrag::ver); + + #[derive(Debug)] + #[repr(C)] + pub(crate) struct GpuStatsComp { + pub(crate) unk: Array<0x140, u8>, + } + default_zeroed!(GpuStatsComp); + + #[derive(Debug)] + #[repr(C)] + pub(crate) struct RuntimeScratch { + pub(crate) unk_280: Array<0x6800, u8>, + pub(crate) unk_6a80: u32, + pub(crate) gpu_idle: u32, + pub(crate) unkpad_6a88: Pad<0x14>, + pub(crate) unk_6a9c: u32, + pub(crate) unk_ctr0: u32, + pub(crate) unk_ctr1: u32, + pub(crate) unk_6aa8: u32, + pub(crate) unk_6aac: u32, + pub(crate) unk_ctr2: u32, + pub(crate) unk_6ab4: u32, + pub(crate) unk_6ab8: u32, + pub(crate) unk_6abc: u32, + pub(crate) unk_6ac0: u32, + pub(crate) unk_6ac4: u32, + pub(crate) unk_ctr3: u32, + pub(crate) unk_6acc: u32, + pub(crate) unk_6ad0: u32, + pub(crate) unk_6ad4: u32, + pub(crate) unk_6ad8: u32, + pub(crate) unk_6adc: u32, + pub(crate) unk_6ae0: u32, + pub(crate) unk_6ae4: u32, + pub(crate) unk_6ae8: u32, + pub(crate) unk_6aec: u32, + pub(crate) unk_6af0: u32, + pub(crate) unk_ctr4: u32, + pub(crate) unk_ctr5: u32, + pub(crate) unk_6afc: u32, + pub(crate) pad_6b00: Pad<0x38>, + pub(crate) unk_6b38: u32, + pub(crate) pad_6b3c: Pad<0x84>, + } + default_zeroed!(RuntimeScratch); + + pub(crate) type BufferMgrCtl = Array<4, u32>; + + #[versions(AGX)] + #[repr(C)] + pub(crate) struct RuntimePointers<'a> { + pub(crate) pipes: Array<4, PipeChannels::ver>, + + pub(crate) device_control: + ChannelRing, + pub(crate) event: ChannelRing, + pub(crate) fw_log: ChannelRing, + pub(crate) ktrace: ChannelRing, + pub(crate) stats: ChannelRing, + + pub(crate) __pad0: Pad<0x50>, + pub(crate) unk_160: U64, + pub(crate) unk_168: U64, + pub(crate) stats_vtx: GpuPointer<'a, super::GpuGlobalStatsVtx::ver>, + pub(crate) stats_frag: GpuPointer<'a, super::GpuGlobalStatsFrag::ver>, + pub(crate) stats_comp: GpuPointer<'a, super::GpuStatsComp>, + pub(crate) hwdata_a: GpuPointer<'a, super::HwDataA::ver>, + pub(crate) unkptr_190: GpuPointer<'a, &'a [u8]>, + pub(crate) unkptr_198: GpuPointer<'a, &'a [u8]>, + pub(crate) hwdata_b: GpuPointer<'a, super::HwDataB::ver>, + pub(crate) hwdata_b_2: GpuPointer<'a, super::HwDataB::ver>, + pub(crate) fwlog_buf: Option>, + pub(crate) unkptr_1b8: GpuPointer<'a, &'a [u8]>, + pub(crate) unkptr_1c0: GpuPointer<'a, &'a [u8]>, + pub(crate) unkptr_1c8: GpuPointer<'a, &'a [u8]>, + pub(crate) unk_1d0: u32, + pub(crate) unk_1d4: u32, + pub(crate) unk_1d8: Array<0x3c, u8>, + pub(crate) buffer_mgr_ctl: GpuPointer<'a, &'a [BufferMgrCtl]>, + pub(crate) buffer_mgr_ctl_2: GpuPointer<'a, &'a [BufferMgrCtl]>, + pub(crate) __pad1: Pad<0x5c>, + pub(crate) gpu_scratch: RuntimeScratch, + } + #[versions(AGX)] + no_debug!(RuntimePointers::ver<'_>); + + #[derive(Debug)] + #[repr(C)] + pub(crate) struct PendingStamp { + pub(crate) info: AtomicU32, + pub(crate) wait_value: AtomicU32, + } + default_zeroed!(PendingStamp); + + #[derive(Debug, Clone, Copy)] + #[repr(C, packed)] + pub(crate) struct FaultInfo { + pub(crate) unk_0: u32, + pub(crate) unk_4: u32, + pub(crate) queue_uuid: u32, + pub(crate) unk_c: u32, + pub(crate) unk_10: u32, + pub(crate) unk_14: u32, + } + default_zeroed!(FaultInfo); + + #[versions(AGX)] + #[derive(Debug, Clone, Copy)] + #[repr(C, packed)] + pub(crate) struct GlobalsSub { + pub(crate) unk_54: u16, + pub(crate) unk_56: u16, + pub(crate) unk_58: u16, + pub(crate) unk_5a: U32, + pub(crate) unk_5e: U32, + pub(crate) unk_62: U32, + + #[ver(V >= V13_0B4)] + pub(crate) unk_66_0: Array<0xc, u8>, + + pub(crate) unk_66: U32, + pub(crate) unk_6a: Array<0x16, u8>, + } + #[versions(AGX)] + default_zeroed!(GlobalsSub::ver); + + #[derive(Debug, Clone, Copy)] + #[repr(C)] + pub(crate) struct PowerZoneGlobal { + pub(crate) target: u32, + pub(crate) target_off: u32, + pub(crate) filter_tc: u32, + } + default_zeroed!(PowerZoneGlobal); + + #[versions(AGX)] + #[derive(Debug)] + #[repr(C)] + pub(crate) struct Globals { + pub(crate) ktrace_enable: u32, + pub(crate) unk_4: Array<0x20, u8>, + + #[ver(V >= V13_2)] + pub(crate) unk_24_0: u32, + + pub(crate) unk_24: u32, + + #[ver(V >= V13_0B4)] + pub(crate) unk_28_0: u32, + + pub(crate) unk_28: u32, + + #[ver(V >= V13_0B4)] + pub(crate) unk_2c_0: u32, + + pub(crate) unk_2c: u32, + pub(crate) unk_30: u32, + pub(crate) unk_34: u32, + pub(crate) unk_38: Array<0x1c, u8>, + + pub(crate) sub: GlobalsSub::ver, + + pub(crate) unk_80: Array<0xf80, u8>, + pub(crate) unk_1000: Array<0x7000, u8>, + pub(crate) unk_8000: Array<0x900, u8>, + + #[ver(V >= V13_0B4 && V < V13_2)] + pub(crate) unk_8900_0: u32, + + pub(crate) unk_8900: u32, + pub(crate) pending_submissions: AtomicU32, + pub(crate) max_power: u32, + pub(crate) max_pstate_scaled: u32, + pub(crate) max_pstate_scaled_2: u32, + pub(crate) unk_8914: u32, + pub(crate) unk_8918: u32, + pub(crate) max_pstate_scaled_3: u32, + pub(crate) unk_8920: u32, + pub(crate) power_zone_count: u32, + pub(crate) avg_power_filter_tc_periods: u32, + pub(crate) avg_power_ki_dt: F32, + pub(crate) avg_power_kp: F32, + pub(crate) avg_power_min_duty_cycle: u32, + pub(crate) avg_power_target_filter_tc: u32, + pub(crate) power_zones: Array<5, PowerZoneGlobal>, + pub(crate) unk_8978: Array<0x44, u8>, + + #[ver(V >= V13_0B4)] + pub(crate) unk_89bc_0: Array<0x3c, u8>, + + pub(crate) unk_89bc: u32, + pub(crate) fast_die0_release_temp: u32, + pub(crate) unk_89c4: i32, + pub(crate) fast_die0_prop_tgt_delta: u32, + pub(crate) fast_die0_kp: F32, + pub(crate) fast_die0_ki_dt: F32, + pub(crate) unk_89d4: Array<0xc, u8>, + pub(crate) unk_89e0: u32, + pub(crate) max_power_2: u32, + pub(crate) ppm_kp: F32, + pub(crate) ppm_ki_dt: F32, + pub(crate) unk_89f0: u32, + + #[ver(V >= V13_0B4)] + pub(crate) unk_89f4_0: Array<0x8, u8>, + + #[ver(V >= V13_0B4)] + pub(crate) unk_89f4_8: u32, + + #[ver(V >= V13_0B4)] + pub(crate) unk_89f4_c: Array<0x50, u8>, + + pub(crate) unk_89f4: u32, + pub(crate) hws1: HwDataShared1, + pub(crate) hws2: HwDataShared2, + + #[ver(V >= V13_0B4)] + pub(crate) unk_hws2_0: Array<0x28, u8>, + + pub(crate) hws3: HwDataShared3, + pub(crate) unk_9004: Array<8, u8>, + pub(crate) unk_900c: u32, + + #[ver(V >= V13_0B4)] + pub(crate) unk_9010_0: u32, + + #[ver(V >= V13_0B4)] + pub(crate) unk_9010_4: Array<0x14, u8>, + + pub(crate) unk_9010: Array<0x2c, u8>, + pub(crate) unk_903c: u32, + pub(crate) unk_9040: Array<0xc0, u8>, + pub(crate) unk_9100: Array<0x6f00, u8>, + pub(crate) unk_10000: Array<0xe50, u8>, + pub(crate) unk_10e50: u32, + pub(crate) unk_10e54: Array<0x2c, u8>, + pub(crate) fault_control: u32, + pub(crate) do_init: u32, + pub(crate) unk_10e88: Array<0x188, u8>, + pub(crate) idle_ts: U64, + pub(crate) idle_unk: U64, + pub(crate) unk_11020: u32, + pub(crate) unk_11024: u32, + pub(crate) unk_11028: u32, + + #[ver(V >= V13_0B4)] + pub(crate) unk_1102c_0: u32, + + #[ver(V >= V13_0B4)] + pub(crate) unk_1102c_4: u32, + + #[ver(V >= V13_0B4)] + pub(crate) unk_1102c_8: u32, + + #[ver(V >= V13_0B4)] + pub(crate) unk_1102c_c: u32, + + #[ver(V >= V13_0B4)] + pub(crate) unk_1102c_10: u32, + + pub(crate) unk_1102c: u32, + pub(crate) idle_off_delay_ms: AtomicU32, + pub(crate) fender_idle_off_delay_ms: u32, + pub(crate) fw_early_wake_timeout_ms: u32, + pub(crate) pending_stamps: Array<0x110, PendingStamp>, + pub(crate) unk_117bc: u32, + pub(crate) fault_info: FaultInfo, + pub(crate) counter: u32, + pub(crate) unk_118dc: u32, + + #[ver(V >= V13_0B4)] + pub(crate) unk_118e0_0: Array<0x9c, u8>, + + pub(crate) unk_118e0: u32, + + #[ver(V >= V13_0B4)] + pub(crate) unk_118e4_0: u32, + + pub(crate) unk_118e4: u32, + pub(crate) unk_118e8: u32, + pub(crate) unk_118ec: Array<0x15, u8>, + pub(crate) unk_11901: Array<0x43f, u8>, + + #[ver(V >= V13_0B4)] + pub(crate) unk_11d40: Array<0x19c, u8>, + + #[ver(V >= V13_0B4)] + pub(crate) unk_11edc: u32, + + #[ver(V >= V13_0B4)] + pub(crate) unk_11ee0: Array<0x1c, u8>, + + #[ver(V >= V13_0B4)] + pub(crate) unk_11efc: u32, + } + #[versions(AGX)] + default_zeroed!(Globals::ver); + + #[derive(Debug, Default, Clone, Copy)] + #[repr(C, packed)] + pub(crate) struct UatLevelInfo { + pub(crate) unk_3: u8, + pub(crate) unk_1: u8, + pub(crate) unk_2: u8, + pub(crate) index_shift: u8, + pub(crate) num_entries: u16, + pub(crate) unk_4: u16, + pub(crate) unk_8: U64, + pub(crate) unk_10: U64, + pub(crate) index_mask: U64, + } + + #[versions(AGX)] + #[derive(Debug)] + #[repr(C)] + pub(crate) struct InitData<'a> { + #[ver(V >= V13_0B4)] + pub(crate) ver_info: Array<0x4, u16>, + + pub(crate) unk_buf: GpuPointer<'a, &'a [u8]>, + pub(crate) unk_8: u32, + pub(crate) unk_c: u32, + pub(crate) runtime_pointers: GpuPointer<'a, super::RuntimePointers::ver>, + pub(crate) globals: GpuPointer<'a, super::Globals::ver>, + pub(crate) fw_status: GpuPointer<'a, super::FwStatus>, + pub(crate) uat_page_size: u16, + pub(crate) uat_page_bits: u8, + pub(crate) uat_num_levels: u8, + pub(crate) uat_level_info: Array<0x3, UatLevelInfo>, + pub(crate) __pad0: Pad<0x14>, + pub(crate) host_mapped_fw_allocations: u32, + pub(crate) unk_ac: u32, + pub(crate) unk_b0: u32, + pub(crate) unk_b4: u32, + pub(crate) unk_b8: u32, + } +} + +#[derive(Debug)] +pub(crate) struct ChannelRing +where + for<'a> ::Raw<'a>: Debug, +{ + pub(crate) state: GpuObject, + pub(crate) ring: GpuArray, +} + +impl ChannelRing +where + for<'a> ::Raw<'a>: Debug, +{ + pub(crate) fn to_raw(&self) -> raw::ChannelRing { + raw::ChannelRing { + state: Some(self.state.weak_pointer()), + ring: Some(self.ring.weak_pointer()), + } + } +} + +trivial_gpustruct!(FwStatus); + +#[versions(AGX)] +#[derive(Debug, Default)] +pub(crate) struct GpuGlobalStatsVtx {} + +#[versions(AGX)] +impl GpuStruct for GpuGlobalStatsVtx::ver { + type Raw<'a> = raw::GpuGlobalStatsVtx::ver; +} + +#[versions(AGX)] +#[derive(Debug, Default)] +pub(crate) struct GpuGlobalStatsFrag {} + +#[versions(AGX)] +impl GpuStruct for GpuGlobalStatsFrag::ver { + type Raw<'a> = raw::GpuGlobalStatsFrag::ver; +} + +#[derive(Debug, Default)] +pub(crate) struct GpuStatsComp {} + +impl GpuStruct for GpuStatsComp { + type Raw<'a> = raw::GpuStatsComp; +} + +#[versions(AGX)] +#[derive(Debug, Default)] +pub(crate) struct HwDataA {} + +#[versions(AGX)] +impl GpuStruct for HwDataA::ver { + type Raw<'a> = raw::HwDataA::ver; +} + +#[versions(AGX)] +#[derive(Debug, Default)] +pub(crate) struct HwDataB {} + +#[versions(AGX)] +impl GpuStruct for HwDataB::ver { + type Raw<'a> = raw::HwDataB::ver; +} + +#[versions(AGX)] +#[derive(Debug)] +pub(crate) struct Stats { + pub(crate) vtx: GpuObject, + pub(crate) frag: GpuObject, + pub(crate) comp: GpuObject, +} + +#[versions(AGX)] +#[derive(Debug)] +pub(crate) struct RuntimePointers { + pub(crate) stats: Stats::ver, + + pub(crate) hwdata_a: GpuObject, + pub(crate) unkptr_190: GpuArray, + pub(crate) unkptr_198: GpuArray, + pub(crate) hwdata_b: GpuObject, + + pub(crate) unkptr_1b8: GpuArray, + pub(crate) unkptr_1c0: GpuArray, + pub(crate) unkptr_1c8: GpuArray, + + pub(crate) buffer_mgr_ctl: GpuArray, +} + +#[versions(AGX)] +impl GpuStruct for RuntimePointers::ver { + type Raw<'a> = raw::RuntimePointers::ver<'a>; +} + +#[versions(AGX)] +#[derive(Debug, Default)] +pub(crate) struct Globals {} + +#[versions(AGX)] +impl GpuStruct for Globals::ver { + type Raw<'a> = raw::Globals::ver; +} + +#[versions(AGX)] +#[derive(Debug)] +pub(crate) struct InitData { + pub(crate) unk_buf: GpuArray, + pub(crate) runtime_pointers: GpuObject, + pub(crate) globals: GpuObject, + pub(crate) fw_status: GpuObject, +} + +#[versions(AGX)] +impl GpuStruct for InitData::ver { + type Raw<'a> = raw::InitData::ver<'a>; +} diff --git a/drivers/gpu/drm/asahi/fw/job.rs b/drivers/gpu/drm/asahi/fw/job.rs new file mode 100644 index 000000000000..a0bbf67b1b1d --- /dev/null +++ b/drivers/gpu/drm/asahi/fw/job.rs @@ -0,0 +1,56 @@ +// SPDX-License-Identifier: GPL-2.0-only OR MIT + +//! Common GPU job firmware structures + +use super::types::*; +use crate::{default_zeroed, trivial_gpustruct}; + +pub(crate) mod raw { + use super::*; + + #[derive(Debug, Clone, Copy)] + #[repr(C)] + pub(crate) struct JobMeta { + pub(crate) unk_4: u32, + pub(crate) stamp: GpuWeakPointer, + pub(crate) fw_stamp: GpuWeakPointer, + pub(crate) stamp_value: EventValue, + pub(crate) stamp_slot: u32, + pub(crate) evctl_index: u32, + pub(crate) flush_stamps: u32, + pub(crate) uuid: u32, + pub(crate) cmd_seq: u32, + } + + #[derive(Debug)] + #[repr(C)] + pub(crate) struct EncoderParams<'a> { + pub(crate) unk_8: u32, + pub(crate) unk_c: u32, + pub(crate) unk_10: u32, + pub(crate) encoder_id: u32, + pub(crate) unk_18: u32, + pub(crate) iogpu_compute_unk44: u32, + pub(crate) seq_buffer: GpuPointer<'a, &'a [u64]>, + pub(crate) unk_28: U64, + } + + #[derive(Debug)] + #[repr(C)] + pub(crate) struct JobTimestamps { + pub(crate) start: AtomicU64, + pub(crate) end: AtomicU64, + } + default_zeroed!(JobTimestamps); + + #[derive(Debug)] + #[repr(C)] + pub(crate) struct RenderTimestamps { + pub(crate) vtx: JobTimestamps, + pub(crate) frag: JobTimestamps, + } + default_zeroed!(RenderTimestamps); +} + +trivial_gpustruct!(JobTimestamps); +trivial_gpustruct!(RenderTimestamps); diff --git a/drivers/gpu/drm/asahi/fw/microseq.rs b/drivers/gpu/drm/asahi/fw/microseq.rs new file mode 100644 index 000000000000..8deea3fb9914 --- /dev/null +++ b/drivers/gpu/drm/asahi/fw/microseq.rs @@ -0,0 +1,384 @@ +// SPDX-License-Identifier: GPL-2.0-only OR MIT + +//! GPU firmware microsequence operations + +use super::types::*; +use super::{buffer, compute, fragment, initdata, vertex, workqueue}; +use crate::default_zeroed; + +pub(crate) trait Operation {} + +#[derive(Debug, Copy, Clone)] +#[repr(u32)] +enum OpCode { + WaitForIdle = 0x01, + RetireStamp = 0x18, + #[allow(dead_code)] + Timestamp = 0x19, + StartVertex = 0x22, + FinalizeVertex = 0x23, + StartFragment = 0x24, + FinalizeFragment = 0x25, + StartCompute = 0x29, + FinalizeCompute = 0x2a, +} + +#[derive(Debug, Copy, Clone)] +#[repr(u32)] +pub(crate) enum Pipe { + Vertex = 1 << 0, + Fragment = 1 << 8, + Compute = 1 << 15, +} + +pub(crate) const MAX_ATTACHMENTS: usize = 16; + +#[derive(Debug, Clone, Copy)] +#[repr(C)] +pub(crate) struct Attachment { + pub(crate) address: U64, + pub(crate) size: u32, + pub(crate) unk_c: u16, + pub(crate) unk_e: u16, +} +default_zeroed!(Attachment); + +#[derive(Debug, Clone, Copy, Default)] +#[repr(C)] +pub(crate) struct Attachments { + pub(crate) list: Array, + pub(crate) count: u32, +} + +#[derive(Debug, Copy, Clone)] +#[repr(transparent)] +pub(crate) struct OpHeader(u32); + +impl OpHeader { + const fn new(opcode: OpCode) -> OpHeader { + OpHeader(opcode as u32) + } + const fn with_args(opcode: OpCode, args: u32) -> OpHeader { + OpHeader(opcode as u32 | args) + } +} + +macro_rules! simple_op { + ($name:ident) => { + #[derive(Debug, Copy, Clone)] + pub(crate) struct $name(OpHeader); + + impl $name { + pub(crate) const HEADER: $name = $name(OpHeader::new(OpCode::$name)); + } + }; +} + +pub(crate) mod op { + use super::*; + + simple_op!(StartVertex); + simple_op!(FinalizeVertex); + simple_op!(StartFragment); + simple_op!(FinalizeFragment); + simple_op!(StartCompute); + simple_op!(FinalizeCompute); + + #[derive(Debug, Copy, Clone)] + pub(crate) struct RetireStamp(OpHeader); + impl RetireStamp { + pub(crate) const HEADER: RetireStamp = + RetireStamp(OpHeader::with_args(OpCode::RetireStamp, 0x40000000)); + } + + #[derive(Debug, Copy, Clone)] + pub(crate) struct WaitForIdle(OpHeader); + impl WaitForIdle { + pub(crate) const fn new(pipe: Pipe) -> WaitForIdle { + WaitForIdle(OpHeader::with_args(OpCode::WaitForIdle, (pipe as u32) << 8)) + } + } + + #[derive(Debug, Copy, Clone)] + pub(crate) struct Timestamp(OpHeader); + impl Timestamp { + #[allow(dead_code)] + pub(crate) const fn new(flag: bool) -> Timestamp { + Timestamp(OpHeader::with_args(OpCode::Timestamp, (flag as u32) << 31)) + } + } +} + +#[derive(Debug)] +#[repr(C)] +pub(crate) struct WaitForIdle { + pub(crate) header: op::WaitForIdle, +} + +impl Operation for WaitForIdle {} + +#[derive(Debug)] +#[repr(C)] +pub(crate) struct RetireStamp { + pub(crate) header: op::RetireStamp, +} + +impl Operation for RetireStamp {} + +#[versions(AGX)] +#[derive(Debug)] +#[repr(C)] +pub(crate) struct Timestamp<'a> { + pub(crate) header: op::Timestamp, + pub(crate) cur_ts: GpuWeakPointer, + pub(crate) start_ts: GpuWeakPointer>>, + pub(crate) update_ts: GpuWeakPointer>>, + pub(crate) work_queue: GpuWeakPointer, + pub(crate) unk_24: U64, + + #[ver(V >= V13_0B4)] + pub(crate) unk_ts: GpuWeakPointer, + + pub(crate) uuid: u32, + pub(crate) unk_30_padding: u32, +} + +#[versions(AGX)] +impl<'a> Operation for Timestamp::ver<'a> {} + +#[versions(AGX)] +#[derive(Debug)] +#[repr(C)] +pub(crate) struct StartVertex<'a> { + pub(crate) header: op::StartVertex, + pub(crate) tiling_params: GpuWeakPointer, + pub(crate) job_params1: GpuWeakPointer>, + pub(crate) buffer: GpuWeakPointer, + pub(crate) scene: GpuWeakPointer, + pub(crate) stats: GpuWeakPointer, + pub(crate) work_queue: GpuWeakPointer, + pub(crate) vm_slot: u32, + pub(crate) unk_38: u32, + pub(crate) event_generation: u32, + pub(crate) buffer_slot: u32, + pub(crate) unk_44: u32, + pub(crate) cmd_seq: U64, + pub(crate) unk_50: u32, + pub(crate) unk_pointer: GpuWeakPointer, + pub(crate) unk_job_buf: GpuWeakPointer, + pub(crate) unk_64: u32, + pub(crate) unk_68: u32, + pub(crate) uuid: u32, + pub(crate) unk_70: u32, + pub(crate) unk_74: Array<0x1d, U64>, + pub(crate) unk_15c: u32, + pub(crate) unk_160: U64, + pub(crate) unk_168: u32, + pub(crate) unk_16c: u32, + pub(crate) unk_170: U64, + + #[ver(V >= V13_0B4)] + pub(crate) counter: U64, + + #[ver(V >= V13_0B4)] + pub(crate) notifier_buf: GpuWeakPointer>, + + pub(crate) unk_178: u32, +} + +#[versions(AGX)] +impl<'a> Operation for StartVertex::ver<'a> {} + +#[versions(AGX)] +#[derive(Debug)] +#[repr(C)] +pub(crate) struct FinalizeVertex { + pub(crate) header: op::FinalizeVertex, + pub(crate) scene: GpuWeakPointer, + pub(crate) buffer: GpuWeakPointer, + pub(crate) stats: GpuWeakPointer, + pub(crate) work_queue: GpuWeakPointer, + pub(crate) vm_slot: u32, + pub(crate) unk_28: u32, + pub(crate) unk_pointer: GpuWeakPointer, + pub(crate) unk_34: u32, + pub(crate) uuid: u32, + pub(crate) fw_stamp: GpuWeakPointer, + pub(crate) stamp_value: EventValue, + pub(crate) unk_48: U64, + pub(crate) unk_50: u32, + pub(crate) unk_54: u32, + pub(crate) unk_58: U64, + pub(crate) unk_60: u32, + pub(crate) unk_64: u32, + pub(crate) unk_68: u32, + + #[ver(G >= G14 && V < V13_0B4)] + pub(crate) unk_68_g14: U64, + + pub(crate) restart_branch_offset: i32, + pub(crate) unk_70: u32, + + #[ver(V >= V13_0B4)] + pub(crate) unk_74: Array<0x10, u8>, +} + +#[versions(AGX)] +impl Operation for FinalizeVertex::ver {} + +#[versions(AGX)] +#[derive(Debug)] +#[repr(C)] +pub(crate) struct StartFragment<'a> { + pub(crate) header: op::StartFragment, + pub(crate) job_params2: GpuWeakPointer, + pub(crate) job_params1: GpuWeakPointer>, + pub(crate) scene: GpuPointer<'a, buffer::Scene::ver>, + pub(crate) stats: GpuWeakPointer, + pub(crate) busy_flag: GpuWeakPointer, + pub(crate) tvb_overflow_count: GpuWeakPointer, + pub(crate) unk_pointer: GpuWeakPointer, + pub(crate) work_queue: GpuWeakPointer, + pub(crate) work_item: GpuWeakPointer, + pub(crate) vm_slot: u32, + pub(crate) unk_50: u32, + pub(crate) event_generation: u32, + pub(crate) buffer_slot: u32, + pub(crate) unk_5c: u32, + pub(crate) cmd_seq: U64, + pub(crate) unk_68: u32, + pub(crate) unk_758_flag: GpuWeakPointer, + pub(crate) unk_job_buf: GpuWeakPointer, + pub(crate) unk_7c: u32, + pub(crate) unk_80: u32, + pub(crate) unk_84: u32, + pub(crate) uuid: u32, + pub(crate) attachments: Attachments, + pub(crate) unk_190: u32, + + #[ver(V >= V13_0B4)] + pub(crate) counter: U64, + + #[ver(V >= V13_0B4)] + pub(crate) notifier_buf: GpuWeakPointer>, +} + +#[versions(AGX)] +impl<'a> Operation for StartFragment::ver<'a> {} + +#[versions(AGX)] +#[derive(Debug)] +#[repr(C)] +pub(crate) struct FinalizeFragment { + pub(crate) header: op::FinalizeFragment, + pub(crate) uuid: u32, + pub(crate) unk_8: u32, + pub(crate) fw_stamp: GpuWeakPointer, + pub(crate) stamp_value: EventValue, + pub(crate) unk_18: u32, + pub(crate) scene: GpuWeakPointer, + pub(crate) buffer: GpuWeakPointer, + pub(crate) unk_2c: U64, + pub(crate) stats: GpuWeakPointer, + pub(crate) unk_pointer: GpuWeakPointer, + pub(crate) busy_flag: GpuWeakPointer, + pub(crate) work_queue: GpuWeakPointer, + pub(crate) work_item: GpuWeakPointer, + pub(crate) vm_slot: u32, + pub(crate) unk_60: u32, + pub(crate) unk_758_flag: GpuWeakPointer, + pub(crate) unk_6c: U64, + pub(crate) unk_74: U64, + pub(crate) unk_7c: U64, + pub(crate) unk_84: U64, + pub(crate) unk_8c: U64, + + #[ver(G == G14 && V < V13_0B4)] + pub(crate) unk_8c_g14: U64, + + pub(crate) restart_branch_offset: i32, + pub(crate) unk_98: u32, + + #[ver(V >= V13_0B4)] + pub(crate) unk_9c: Array<0x10, u8>, +} + +#[versions(AGX)] +impl Operation for FinalizeFragment::ver {} + +#[versions(AGX)] +#[derive(Debug)] +#[repr(C)] +pub(crate) struct StartCompute<'a> { + pub(crate) header: op::StartCompute, + pub(crate) unk_pointer: GpuWeakPointer>, + pub(crate) job_params1: GpuWeakPointer>, + pub(crate) stats: GpuWeakPointer, + pub(crate) work_queue: GpuWeakPointer, + pub(crate) vm_slot: u32, + pub(crate) unk_28: u32, + pub(crate) event_generation: u32, + pub(crate) cmd_seq: U64, + pub(crate) unk_38: u32, + pub(crate) job_params2: GpuWeakPointer>, + pub(crate) unk_44: u32, + pub(crate) uuid: u32, + pub(crate) attachments: Attachments, + pub(crate) padding: u32, + + #[ver(V >= V13_0B4)] + pub(crate) unk_flag: GpuWeakPointer, + + #[ver(V >= V13_0B4)] + pub(crate) counter: U64, + + #[ver(V >= V13_0B4)] + pub(crate) notifier_buf: GpuWeakPointer>, +} + +#[versions(AGX)] +impl<'a> Operation for StartCompute::ver<'a> {} + +#[versions(AGX)] +#[derive(Debug)] +#[repr(C)] +pub(crate) struct FinalizeCompute<'a> { + pub(crate) header: op::FinalizeCompute, + pub(crate) stats: GpuWeakPointer, + pub(crate) work_queue: GpuWeakPointer, + pub(crate) vm_slot: u32, + #[ver(V < V13_0B4)] + pub(crate) unk_18: u32, + pub(crate) job_params2: GpuWeakPointer>, + pub(crate) unk_24: u32, + pub(crate) uuid: u32, + pub(crate) fw_stamp: GpuWeakPointer, + pub(crate) stamp_value: EventValue, + pub(crate) unk_38: u32, + pub(crate) unk_3c: u32, + pub(crate) unk_40: u32, + pub(crate) unk_44: u32, + pub(crate) unk_48: u32, + pub(crate) unk_4c: u32, + pub(crate) unk_50: u32, + pub(crate) unk_54: u32, + pub(crate) unk_58: u32, + + #[ver(G == G14 && V < V13_0B4)] + pub(crate) unk_5c_g14: U64, + + pub(crate) restart_branch_offset: i32, + pub(crate) unk_60: u32, + + #[ver(V >= V13_0B4)] + pub(crate) unk_64: Array<0xd, u8>, + + #[ver(V >= V13_0B4)] + pub(crate) unk_flag: GpuWeakPointer, + + #[ver(V >= V13_0B4)] + pub(crate) unk_79: Array<0x7, u8>, +} + +#[versions(AGX)] +impl<'a> Operation for FinalizeCompute::ver<'a> {} diff --git a/drivers/gpu/drm/asahi/fw/mod.rs b/drivers/gpu/drm/asahi/fw/mod.rs new file mode 100644 index 000000000000..a5649aa20d3a --- /dev/null +++ b/drivers/gpu/drm/asahi/fw/mod.rs @@ -0,0 +1,15 @@ +// SPDX-License-Identifier: GPL-2.0-only OR MIT + +//! Firmware structures for Apple AGX GPUs + +pub(crate) mod buffer; +pub(crate) mod channels; +pub(crate) mod compute; +pub(crate) mod event; +pub(crate) mod fragment; +pub(crate) mod initdata; +pub(crate) mod job; +pub(crate) mod microseq; +pub(crate) mod types; +pub(crate) mod vertex; +pub(crate) mod workqueue; diff --git a/drivers/gpu/drm/asahi/fw/types.rs b/drivers/gpu/drm/asahi/fw/types.rs new file mode 100644 index 000000000000..c1a07be1e047 --- /dev/null +++ b/drivers/gpu/drm/asahi/fw/types.rs @@ -0,0 +1,233 @@ +// SPDX-License-Identifier: GPL-2.0-only OR MIT + +//! Common types for firmware structure definitions + +use crate::{alloc, object}; +use core::fmt; +use core::ops::{Deref, DerefMut, Index, IndexMut}; + +pub(crate) use crate::event::EventValue; +pub(crate) use crate::object::{GpuPointer, GpuStruct, GpuWeakPointer}; +pub(crate) use crate::{f32, float::F32}; + +pub(crate) use ::alloc::boxed::Box; +pub(crate) use core::fmt::Debug; +pub(crate) use core::marker::PhantomData; +pub(crate) use core::sync::atomic::{AtomicI32, AtomicU32, AtomicU64}; +pub(crate) use kernel::macros::versions; + +// Make the trait visible +pub(crate) use crate::alloc::Allocator as _Allocator; + +/// General allocator type used for the driver +pub(crate) type Allocator = alloc::DefaultAllocator; + +/// General GpuObject type used for the driver +pub(crate) type GpuObject = + object::GpuObject>; + +/// General GpuArray type used for the driver +pub(crate) type GpuArray = object::GpuArray>; + +/// General GpuOnlyArray type used for the driver +pub(crate) type GpuOnlyArray = + object::GpuOnlyArray>; + +/// A stamp slot that is shared between firmware and the driver. +#[derive(Debug, Default)] +#[repr(transparent)] +pub(crate) struct Stamp(pub(crate) AtomicU32); + +/// A stamp slot that is for private firmware use. +/// +/// This is a separate type to guard against pointer type confusion. +#[derive(Debug, Default)] +#[repr(transparent)] +pub(crate) struct FwStamp(pub(crate) AtomicU32); + +/// An unaligned u64 type. +/// +/// This is useful to avoid having to pack firmware structures entirely, since that is incompatible +/// with `#[derive(Debug)]` and atomics. +#[derive(Copy, Clone, Default)] +#[repr(C, packed(1))] +pub(crate) struct U64(pub(crate) u64); + +unsafe impl Zeroed for U64 {} + +impl fmt::Debug for U64 { + fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result { + let v = self.0; + f.write_fmt(format_args!("{:#x}", v)) + } +} + +/// An unaligned u32 type. +/// +/// This is useful to avoid having to pack firmware structures entirely, since that is incompatible +/// with `#[derive(Debug)]` and atomics. +#[derive(Copy, Clone, Default)] +#[repr(C, packed(1))] +pub(crate) struct U32(pub(crate) u32); + +unsafe impl Zeroed for U32 {} + +impl fmt::Debug for U32 { + fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result { + let v = self.0; + f.write_fmt(format_args!("{:#x}", v)) + } +} + +unsafe impl Zeroed for u8 {} +unsafe impl Zeroed for u16 {} +unsafe impl Zeroed for u32 {} +unsafe impl Zeroed for u64 {} +unsafe impl Zeroed for i8 {} +unsafe impl Zeroed for i16 {} +unsafe impl Zeroed for i32 {} +unsafe impl Zeroed for i64 {} + +/// Create a dummy `Debug` implementation, for when we need it but it's too painful to write by +/// hand or not very useful. +#[macro_export] +macro_rules! no_debug { + ($type:ty) => { + impl Debug for $type { + fn fmt(&self, f: &mut core::fmt::Formatter<'_>) -> core::fmt::Result { + write!(f, "...") + } + } + }; +} + +/// Types which can be safely initialized with an all-zero bit pattern. +/// +/// See: https://github.com/rust-lang/rfcs/issues/2626 +/// +/// # Safety +/// +/// This trait must only be implemented if a type only contains primitive types which can be +/// zero-initialized, FFI structs intended to be zero-initialized, or other types which impl Zeroed. +pub(crate) unsafe trait Zeroed: Default { + fn zeroed() -> Self { + // SAFETY: The user is responsible for ensuring this is safe. + unsafe { core::mem::zeroed() } + } +} + +/// Implement Zeroed for a given type (and Default along with it). +/// +/// # Safety +/// +/// This macro must only be used if a type only contains primitive types which can be +/// zero-initialized, FFI structs intended to be zero-initialized, or other types which impl Zeroed. +#[macro_export] +macro_rules! default_zeroed { + (<$($lt:lifetime),*>, $type:ty) => { + impl<$($lt),*> Default for $type { + fn default() -> $type { + Zeroed::zeroed() + } + } + // SAFETY: The user is responsible for ensuring this is safe. + unsafe impl<$($lt),*> Zeroed for $type {} + }; + ($type:ty) => { + impl Default for $type { + fn default() -> $type { + Zeroed::zeroed() + } + } + // SAFETY: The user is responsible for ensuring this is safe. + unsafe impl Zeroed for $type {} + }; +} + +/// A convenience type for a number of padding bytes. Hidden from Debug formatting. +#[derive(Copy, Clone)] +#[repr(C, packed)] +pub(crate) struct Pad([u8; N]); + +/// SAFETY: Primitive type, safe to zero-init. +unsafe impl Zeroed for Pad {} + +impl Default for Pad { + fn default() -> Self { + Zeroed::zeroed() + } +} + +impl fmt::Debug for Pad { + fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result { + f.write_fmt(format_args!("")) + } +} + +/// A convenience type for a fixed-sized array with Default/Zeroed impls. +#[derive(Copy, Clone)] +#[repr(C)] +pub(crate) struct Array([T; N]); + +impl Array { + pub(crate) fn new(data: [T; N]) -> Self { + Self(data) + } +} + +// SAFETY: Arrays of Zeroed values can be safely Zeroed. +unsafe impl Zeroed for Array {} + +impl Default for Array { + fn default() -> Self { + Zeroed::zeroed() + } +} + +impl Index for Array { + type Output = T; + + fn index(&self, index: usize) -> &Self::Output { + &self.0[index] + } +} + +impl IndexMut for Array { + fn index_mut(&mut self, index: usize) -> &mut Self::Output { + &mut self.0[index] + } +} + +impl Deref for Array { + type Target = [T; N]; + + fn deref(&self) -> &Self::Target { + &self.0 + } +} + +impl DerefMut for Array { + fn deref_mut(&mut self) -> &mut Self::Target { + &mut self.0 + } +} + +impl fmt::Debug for Array { + fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result { + self.0.fmt(f) + } +} + +/// Convenience macro to define an identically-named trivial GpuStruct with no inner fields for a +/// given raw type name. +#[macro_export] +macro_rules! trivial_gpustruct { + ($type:ident) => { + #[derive(Debug, Default)] + pub(crate) struct $type {} + + impl GpuStruct for $type { + type Raw<'a> = raw::$type; + } + }; +} diff --git a/drivers/gpu/drm/asahi/fw/vertex.rs b/drivers/gpu/drm/asahi/fw/vertex.rs new file mode 100644 index 000000000000..959a0913e693 --- /dev/null +++ b/drivers/gpu/drm/asahi/fw/vertex.rs @@ -0,0 +1,177 @@ +// SPDX-License-Identifier: GPL-2.0-only OR MIT + +//! GPU vertex job firmware structures + +use super::types::*; +use super::{event, job, workqueue}; +use crate::{buffer, fw, microseq, mmu}; +use kernel::sync::Arc; + +pub(crate) mod raw { + use super::*; + + #[derive(Debug, Default, Copy, Clone)] + #[repr(C)] + pub(crate) struct TilingParameters { + pub(crate) rgn_size: u32, + pub(crate) unk_4: u32, + pub(crate) ppp_ctrl: u32, + pub(crate) x_max: u16, + pub(crate) y_max: u16, + pub(crate) te_screen: u32, + pub(crate) te_mtile1: u32, + pub(crate) te_mtile2: u32, + pub(crate) tiles_per_mtile: u32, + pub(crate) tpc_stride: u32, + pub(crate) unk_24: u32, + pub(crate) unk_28: u32, + } + + #[versions(AGX)] + #[derive(Debug)] + #[repr(C)] + pub(crate) struct JobParameters1<'a> { + pub(crate) unk_0: U64, + pub(crate) unk_8: F32, + pub(crate) unk_c: F32, + pub(crate) tvb_tilemap: GpuPointer<'a, &'a [u8]>, + #[ver(G < G14)] + pub(crate) tvb_cluster_tilemaps: Option>, + pub(crate) tpc: GpuPointer<'a, &'a [u8]>, + pub(crate) tvb_heapmeta: GpuPointer<'a, &'a [u8]>, + pub(crate) iogpu_unk_54: u32, + pub(crate) iogpu_unk_55: u32, + pub(crate) iogpu_unk_56: U64, + #[ver(G < G14)] + pub(crate) tvb_cluster_meta1: Option>, + pub(crate) utile_config: u32, + pub(crate) unk_4c: u32, + pub(crate) ppp_multisamplectl: U64, + pub(crate) tvb_heapmeta_2: GpuPointer<'a, &'a [u8]>, + #[ver(G < G14)] + pub(crate) unk_60: U64, + #[ver(G < G14)] + pub(crate) core_mask: Array<2, u32>, + pub(crate) preempt_buf1: GpuPointer<'a, &'a [u8]>, + pub(crate) preempt_buf2: GpuPointer<'a, &'a [u8]>, + pub(crate) unk_80: U64, + pub(crate) preempt_buf3: GpuPointer<'a, &'a [u8]>, + pub(crate) encoder_addr: U64, + #[ver(G < G14)] + pub(crate) tvb_cluster_meta2: Option>, + #[ver(G < G14)] + pub(crate) tvb_cluster_meta3: Option>, + #[ver(G < G14)] + pub(crate) tiling_control: u32, + #[ver(G < G14)] + pub(crate) unk_ac: u32, + pub(crate) unk_b0: Array<6, U64>, + pub(crate) pipeline_base: U64, + #[ver(G < G14)] + pub(crate) tvb_cluster_meta4: Option>, + #[ver(G < G14)] + pub(crate) unk_f0: U64, + pub(crate) unk_f8: U64, + pub(crate) unk_100: Array<3, U64>, + pub(crate) unk_118: u32, + #[ver(G >= G14)] + pub(crate) __pad: Pad<{ 8 * 9 }>, + } + + #[derive(Debug)] + #[repr(C)] + pub(crate) struct JobParameters2<'a> { + pub(crate) unk_480: Array<4, u32>, + pub(crate) unk_498: U64, + pub(crate) unk_4a0: u32, + pub(crate) preempt_buf1: GpuPointer<'a, &'a [u8]>, + pub(crate) unk_4ac: u32, + pub(crate) unk_4b0: U64, + pub(crate) unk_4b8: u32, + pub(crate) unk_4bc: U64, + pub(crate) unk_4c4_padding: Array<0x48, u8>, + pub(crate) unk_50c: u32, + pub(crate) unk_510: U64, + pub(crate) unk_518: U64, + pub(crate) unk_520: U64, + } + + #[versions(AGX)] + #[derive(Debug)] + #[repr(C)] + pub(crate) struct RunVertex<'a> { + pub(crate) tag: workqueue::CommandType, + + #[ver(V >= V13_0B4)] + pub(crate) counter: U64, + + pub(crate) vm_slot: u32, + pub(crate) unk_8: u32, + pub(crate) notifier: GpuPointer<'a, event::Notifier::ver>, + pub(crate) buffer_slot: u32, + pub(crate) unk_1c: u32, + pub(crate) buffer: GpuPointer<'a, fw::buffer::Info::ver>, + pub(crate) scene: GpuPointer<'a, fw::buffer::Scene::ver>, + pub(crate) unk_buffer_buf: GpuWeakPointer<[u8]>, + pub(crate) unk_34: u32, + pub(crate) job_params1: JobParameters1::ver<'a>, + pub(crate) unk_154: Array<0x268, u8>, + pub(crate) tiling_params: TilingParameters, + pub(crate) unk_3e8: Array<0x74, u8>, + pub(crate) tpc: GpuPointer<'a, &'a [u8]>, + pub(crate) tpc_size: U64, + pub(crate) microsequence: GpuPointer<'a, &'a [u8]>, + pub(crate) microsequence_size: u32, + pub(crate) fragment_stamp_slot: u32, + pub(crate) fragment_stamp_value: EventValue, + pub(crate) unk_pointee: u32, + pub(crate) unk_pad: u32, + pub(crate) job_params2: JobParameters2<'a>, + pub(crate) encoder_params: job::raw::EncoderParams<'a>, + pub(crate) unk_55c: u32, + pub(crate) unk_560: u32, + pub(crate) memoryless_rts_used: u32, + pub(crate) unk_568: u32, + pub(crate) unk_56c: u32, + pub(crate) meta: job::raw::JobMeta, + pub(crate) unk_after_meta: u32, + pub(crate) unk_buf_0: U64, + pub(crate) unk_buf_8: U64, + pub(crate) unk_buf_10: U64, + pub(crate) cur_ts: U64, + pub(crate) start_ts: Option>, + pub(crate) end_ts: Option>, + pub(crate) unk_5c4: u32, + pub(crate) unk_5c8: u32, + pub(crate) unk_5cc: u32, + pub(crate) unk_5d0: u32, + pub(crate) client_sequence: u8, + pub(crate) pad_5d5: Array<3, u8>, + pub(crate) unk_5d8: u32, + pub(crate) unk_5dc: u8, + + #[ver(V >= V13_0B4)] + pub(crate) unk_ts: U64, + + #[ver(V >= V13_0B4)] + pub(crate) unk_5dd_8: Array<0x1b, u8>, + } +} + +#[versions(AGX)] +#[derive(Debug)] +pub(crate) struct RunVertex { + pub(crate) notifier: Arc>, + pub(crate) scene: Arc, + pub(crate) micro_seq: microseq::MicroSequence, + pub(crate) vm_bind: mmu::VmBind, + pub(crate) timestamps: Arc>, +} + +#[versions(AGX)] +impl GpuStruct for RunVertex::ver { + type Raw<'a> = raw::RunVertex::ver<'a>; +} + +#[versions(AGX)] +impl workqueue::Command for RunVertex::ver {} diff --git a/drivers/gpu/drm/asahi/fw/workqueue.rs b/drivers/gpu/drm/asahi/fw/workqueue.rs new file mode 100644 index 000000000000..e81025b6c014 --- /dev/null +++ b/drivers/gpu/drm/asahi/fw/workqueue.rs @@ -0,0 +1,168 @@ +// SPDX-License-Identifier: GPL-2.0-only OR MIT + +//! GPU work queue firmware structes + +use super::event; +use super::types::*; +use crate::event::EventValue; +use crate::{default_zeroed, trivial_gpustruct}; +use kernel::sync::Arc; + +#[derive(Debug)] +#[repr(u32)] +pub(crate) enum CommandType { + RunVertex = 0, + RunFragment = 1, + #[allow(dead_code)] + RunBlitter = 2, + RunCompute = 3, + Barrier = 4, + InitBuffer = 6, +} + +pub(crate) trait Command: GpuStruct + Send + Sync {} + +pub(crate) mod raw { + use super::*; + + #[derive(Debug)] + #[repr(C)] + pub(crate) struct Barrier { + pub(crate) tag: CommandType, + pub(crate) wait_stamp: GpuWeakPointer, + pub(crate) wait_value: EventValue, + pub(crate) wait_slot: u32, + pub(crate) stamp_self: EventValue, + pub(crate) uuid: u32, + pub(crate) unk: u32, + } + + #[derive(Debug, Clone, Copy)] + #[repr(C)] + pub(crate) struct GpuContextData { + pub(crate) unk_0: u8, + pub(crate) unk_1: u8, + unk_2: Array<0x2, u8>, + pub(crate) unk_4: u8, + pub(crate) unk_5: u8, + unk_6: Array<0x18, u8>, + pub(crate) unk_1e: u8, + pub(crate) unk_1f: u8, + unk_20: Array<0x3, u8>, + pub(crate) unk_23: u8, + unk_24: Array<0x1c, u8>, + } + + impl Default for GpuContextData { + fn default() -> Self { + Self { + unk_0: 0xff, + unk_1: 0xff, + unk_2: Default::default(), + unk_4: 0, + unk_5: 1, + unk_6: Default::default(), + unk_1e: 0xff, + unk_1f: 0, + unk_20: Default::default(), + unk_23: 2, + unk_24: Default::default(), + } + } + } + + #[derive(Debug)] + #[repr(C)] + pub(crate) struct RingState { + pub(crate) gpu_doneptr: AtomicU32, + __pad0: Pad<0xc>, + pub(crate) unk_10: AtomicU32, + __pad1: Pad<0xc>, + pub(crate) unk_20: AtomicU32, + __pad2: Pad<0xc>, + pub(crate) gpu_rptr: AtomicU32, + __pad3: Pad<0xc>, + pub(crate) cpu_wptr: AtomicU32, + __pad4: Pad<0xc>, + pub(crate) rb_size: u32, + __pad5: Pad<0xc>, + // This isn't part of the structure, but it's here as a + // debugging hack so we can inspect what ring position + // the driver considered complete and freeable. + pub(crate) cpu_freeptr: AtomicU32, + __pad6: Pad<0xc>, + } + default_zeroed!(RingState); + + #[derive(Debug, Clone, Copy)] + #[repr(C)] + pub(crate) struct Priority(u32, u32, U64, u32, u32, u32); + + pub(crate) const PRIORITY: [Priority; 4] = [ + Priority(0, 0, U64(0xffff_ffff_ffff_0000), 1, 0, 1), + Priority(1, 1, U64(0xffff_ffff_0000_0000), 0, 0, 0), + Priority(2, 2, U64(0xffff_0000_0000_0000), 0, 0, 2), + Priority(3, 3, U64(0x0000_0000_0000_0000), 0, 0, 3), + ]; + + impl Default for Priority { + fn default() -> Priority { + PRIORITY[2] + } + } + + #[versions(AGX)] + #[derive(Debug)] + #[repr(C)] + pub(crate) struct QueueInfo<'a> { + pub(crate) state: GpuPointer<'a, super::RingState>, + pub(crate) ring: GpuPointer<'a, &'a [u64]>, + pub(crate) notifier_list: GpuPointer<'a, event::NotifierList>, + pub(crate) gpu_buf: GpuPointer<'a, &'a [u8]>, + pub(crate) gpu_rptr1: AtomicU32, + pub(crate) gpu_rptr2: AtomicU32, + pub(crate) gpu_rptr3: AtomicU32, + pub(crate) event_id: AtomicI32, + pub(crate) priority: Priority, + pub(crate) unk_4c: i32, + pub(crate) uuid: u32, + pub(crate) unk_54: i32, + pub(crate) unk_58: U64, + pub(crate) busy: AtomicU32, + pub(crate) __pad: Pad<0x20>, + pub(crate) unk_84_state: AtomicU32, + pub(crate) unk_88: u32, + pub(crate) unk_8c: u32, + pub(crate) unk_90: u32, + pub(crate) unk_94: u32, + pub(crate) pending: AtomicU32, + pub(crate) unk_9c: u32, + #[ver(V >= V13_2)] + pub(crate) unk_a0_0: u32, + pub(crate) gpu_context: GpuPointer<'a, super::GpuContextData>, + pub(crate) unk_a8: U64, + #[ver(V >= V13_2)] + pub(crate) unk_b0: u32, + } +} + +trivial_gpustruct!(Barrier); +trivial_gpustruct!(GpuContextData); +trivial_gpustruct!(RingState); + +impl Command for Barrier {} + +#[versions(AGX)] +#[derive(Debug)] +pub(crate) struct QueueInfo { + pub(crate) state: GpuObject, + pub(crate) ring: GpuArray, + pub(crate) gpu_buf: GpuArray, + pub(crate) notifier_list: Arc>, + pub(crate) gpu_context: Arc, +} + +#[versions(AGX)] +impl GpuStruct for QueueInfo::ver { + type Raw<'a> = raw::QueueInfo::ver<'a>; +} diff --git a/drivers/gpu/drm/asahi/gem.rs b/drivers/gpu/drm/asahi/gem.rs new file mode 100644 index 000000000000..b334bebb0e8e --- /dev/null +++ b/drivers/gpu/drm/asahi/gem.rs @@ -0,0 +1,301 @@ +// SPDX-License-Identifier: GPL-2.0-only OR MIT + +//! Asahi driver GEM object implementation +//! +//! Basic wrappers and adaptations between generic GEM shmem objects and this driver's +//! view of what a GPU buffer object is. It is in charge of keeping track of all mappings for +//! each GEM object so we can remove them when a client (File) or a Vm are destroyed, as well as +//! implementing RTKit buffers on top of GEM objects for firmware use. + +use kernel::{ + bindings, + drm::{gem, gem::shmem}, + error::Result, + prelude::*, + soc::apple::rtkit, + sync::smutex::Mutex, +}; + +use kernel::drm::gem::BaseObject; + +use core::sync::atomic::{AtomicU64, Ordering}; + +use crate::debug::*; +use crate::driver::AsahiDevice; +use crate::file::DrmFile; + +const DEBUG_CLASS: DebugFlags = DebugFlags::Gem; + +/// Represents the inner data of a GEM object for this driver. +pub(crate) struct DriverObject { + /// Whether this is a kernel-created object. + kernel: bool, + /// Object creation flags. + flags: u32, + /// VM ID for VM-private objects. + vm_id: Option, + /// Locked list of mapping tuples: (file_id, vm_id, mapping) + mappings: Mutex>, + /// ID for debug + id: u64, +} + +/// Type alias for the shmem GEM object type for this driver. +pub(crate) type Object = shmem::Object; + +/// Type alias for the SGTable type for this driver. +pub(crate) type SGTable = shmem::SGTable; + +/// A shared reference to a GEM object for this driver. +pub(crate) struct ObjectRef { + /// The underlying GEM object reference + pub(crate) gem: gem::ObjectRef>, + /// The kernel-side VMap of this object, if needed + vmap: Option>, +} + +static GEM_ID: AtomicU64 = AtomicU64::new(0); + +impl DriverObject { + /// Drop all object mappings for a given file ID. + /// + /// Used on file close. + fn drop_file_mappings(&self, file_id: u64) { + let mut mappings = self.mappings.lock(); + for (index, (mapped_fid, _mapped_vmid, _mapping)) in mappings.iter().enumerate() { + if *mapped_fid == file_id { + mappings.swap_remove(index); + return; + } + } + } + + /// Drop all object mappings for a given VM ID. + /// + /// Used on VM destroy. + fn drop_vm_mappings(&self, vm_id: u64) { + let mut mappings = self.mappings.lock(); + for (index, (_mapped_fid, mapped_vmid, _mapping)) in mappings.iter().enumerate() { + if *mapped_vmid == vm_id { + mappings.swap_remove(index); + return; + } + } + } +} + +impl ObjectRef { + /// Create a new wrapper for a raw GEM object reference. + pub(crate) fn new(gem: gem::ObjectRef>) -> ObjectRef { + ObjectRef { gem, vmap: None } + } + + /// Return the `VMap` for this object, creating it if necessary. + pub(crate) fn vmap(&mut self) -> Result<&mut shmem::VMap> { + if self.vmap.is_none() { + self.vmap = Some(self.gem.vmap()?); + } + Ok(self.vmap.as_mut().unwrap()) + } + + /// Return the IOVA of this object at which it is mapped in a given `Vm` identified by its ID, + /// if it is mapped in that `Vm`. + pub(crate) fn iova(&self, vm_id: u64) -> Option { + let mappings = self.gem.mappings.lock(); + for (_mapped_fid, mapped_vmid, mapping) in mappings.iter() { + if *mapped_vmid == vm_id { + return Some(mapping.iova()); + } + } + + None + } + + /// Returns the size of an object in bytes + pub(crate) fn size(&self) -> usize { + self.gem.size() + } + + /// Maps an object into a given `Vm` at any free address. + /// + /// Returns Err(EBUSY) if there is already a mapping. + pub(crate) fn map_into(&mut self, vm: &crate::mmu::Vm) -> Result { + let vm_id = vm.id(); + + if self.gem.vm_id.is_some() && self.gem.vm_id != Some(vm_id) { + return Err(EINVAL); + } + + let mut mappings = self.gem.mappings.lock(); + for (_mapped_fid, mapped_vmid, _mapping) in mappings.iter() { + if *mapped_vmid == vm_id { + return Err(EBUSY); + } + } + + let sgt = self.gem.sg_table()?; + let new_mapping = vm.map(self.gem.size(), sgt)?; + + let iova = new_mapping.iova(); + mappings.try_push((vm.file_id(), vm_id, new_mapping))?; + Ok(iova) + } + + /// Maps an object into a given `Vm` at any free address within a given range. + /// + /// Returns Err(EBUSY) if there is already a mapping. + pub(crate) fn map_into_range( + &mut self, + vm: &crate::mmu::Vm, + start: u64, + end: u64, + alignment: u64, + prot: u32, + guard: bool, + ) -> Result { + let vm_id = vm.id(); + + if self.gem.vm_id.is_some() && self.gem.vm_id != Some(vm_id) { + return Err(EINVAL); + } + + let mut mappings = self.gem.mappings.lock(); + for (_mapped_fid, mapped_vmid, _mapping) in mappings.iter() { + if *mapped_vmid == vm_id { + return Err(EBUSY); + } + } + + let sgt = self.gem.sg_table()?; + let new_mapping = + vm.map_in_range(self.gem.size(), sgt, alignment, start, end, prot, guard)?; + + let iova = new_mapping.iova(); + mappings.try_push((vm.file_id(), vm_id, new_mapping))?; + Ok(iova) + } + + /// Maps an object into a given `Vm` at a specific address. + /// + /// Returns Err(EBUSY) if there is already a mapping. + /// Returns Err(ENOSPC) if the requested address is already busy. + pub(crate) fn map_at( + &mut self, + vm: &crate::mmu::Vm, + addr: u64, + prot: u32, + guard: bool, + ) -> Result { + let vm_id = vm.id(); + + if self.gem.vm_id.is_some() && self.gem.vm_id != Some(vm_id) { + return Err(EINVAL); + } + + let mut mappings = self.gem.mappings.lock(); + for (_mapped_fid, mapped_vmid, _mapping) in mappings.iter() { + if *mapped_vmid == vm_id { + return Err(EBUSY); + } + } + + let sgt = self.gem.sg_table()?; + let new_mapping = vm.map_at(addr, self.gem.size(), sgt, prot, guard)?; + + let iova = new_mapping.iova(); + assert!(iova == addr as usize); + mappings.try_push((vm.file_id(), vm_id, new_mapping))?; + Ok(()) + } + + /// Drop all mappings for this object owned by a given `Vm` identified by its ID. + pub(crate) fn drop_vm_mappings(&mut self, vm_id: u64) { + self.gem.drop_vm_mappings(vm_id); + } + + /// Drop all mappings for this object owned by a given `File` identified by its ID. + pub(crate) fn drop_file_mappings(&mut self, file_id: u64) { + self.gem.drop_file_mappings(file_id); + } +} + +/// Create a new kernel-owned GEM object. +pub(crate) fn new_kernel_object(dev: &AsahiDevice, size: usize) -> Result { + let mut gem = shmem::Object::::new(dev, size)?; + gem.kernel = true; + gem.flags = 0; + + gem.set_exportable(false); + + mod_pr_debug!("DriverObject new kernel object id={}\n", gem.id); + Ok(ObjectRef::new(gem.into_ref())) +} + +/// Create a new user-owned GEM object with the given flags. +pub(crate) fn new_object( + dev: &AsahiDevice, + size: usize, + flags: u32, + vm_id: Option, +) -> Result { + let mut gem = shmem::Object::::new(dev, size)?; + gem.kernel = false; + gem.flags = flags; + gem.vm_id = vm_id; + + gem.set_exportable(vm_id.is_none()); + gem.set_wc(flags & bindings::ASAHI_GEM_WRITEBACK == 0); + + mod_pr_debug!( + "DriverObject new user object: vm_id={:?} id={}\n", + vm_id, + gem.id + ); + Ok(ObjectRef::new(gem.into_ref())) +} + +/// Look up a GEM object handle for a `File` and return an `ObjectRef` for it. +pub(crate) fn lookup_handle(file: &DrmFile, handle: u32) -> Result { + Ok(ObjectRef::new(shmem::Object::lookup_handle(file, handle)?)) +} + +impl gem::BaseDriverObject for DriverObject { + /// Callback to create the inner data of a GEM object + fn new(_dev: &AsahiDevice, _size: usize) -> Result { + let id = GEM_ID.fetch_add(1, Ordering::Relaxed); + mod_pr_debug!("DriverObject::new id={}\n", id); + Ok(DriverObject { + kernel: false, + flags: 0, + vm_id: None, + mappings: Mutex::new(Vec::new()), + id, + }) + } + + /// Callback to drop all mappings for a GEM object owned by a given `File` + fn close(obj: &Object, file: &DrmFile) { + mod_pr_debug!("DriverObject::close vm_id={:?} id={}\n", obj.vm_id, obj.id); + obj.drop_file_mappings(file.file_id()); + } +} + +impl Drop for DriverObject { + fn drop(&mut self) { + mod_pr_debug!("DriverObject::drop vm_id={:?} id={}\n", self.vm_id, self.id); + } +} + +impl shmem::DriverObject for DriverObject { + type Driver = crate::driver::AsahiDriver; +} + +impl rtkit::Buffer for ObjectRef { + fn iova(&self) -> Result { + self.iova(0).ok_or(EIO) + } + fn buf(&mut self) -> Result<&mut [u8]> { + let vmap = self.vmap.as_mut().ok_or(ENOMEM)?; + Ok(vmap.as_mut_slice()) + } +} diff --git a/drivers/gpu/drm/asahi/gpu.rs b/drivers/gpu/drm/asahi/gpu.rs new file mode 100644 index 000000000000..1384b52d11a2 --- /dev/null +++ b/drivers/gpu/drm/asahi/gpu.rs @@ -0,0 +1,1088 @@ +// SPDX-License-Identifier: GPL-2.0-only OR MIT + +//! Top-level GPU manager +//! +//! This module is the root of all GPU firmware management for a given driver instance. It is +//! responsible for initialization, owning the top-level managers (events, UAT, etc.), and +//! communicating with the raw RtKit endpoints to send and receive messages to/from the GPU +//! firmware. +//! +//! It is also the point where diverging driver firmware/GPU variants (using the versions macro) +//! are unified, so that the top level of the driver itself (in `driver`) does not have to concern +//! itself with version dependence. + +use core::any::Any; +use core::sync::atomic::{AtomicBool, AtomicU64, Ordering}; +use core::time::Duration; + +use kernel::{ + delay::coarse_sleep, + error::code::*, + macros::versions, + prelude::*, + soc::apple::rtkit, + sync::{smutex::Mutex, Arc, Guard, UniqueArc}, + time, + types::ForeignOwnable, +}; + +use crate::alloc::Allocator; +use crate::box_in_place; +use crate::debug::*; +use crate::driver::AsahiDevice; +use crate::fw::channels::PipeType; +use crate::fw::types::U64; +use crate::{ + alloc, buffer, channel, event, fw, gem, hw, initdata, mem, mmu, queue, regs, workqueue, +}; + +const DEBUG_CLASS: DebugFlags = DebugFlags::Gpu; + +/// Firmware endpoint for init & incoming notifications. +const EP_FIRMWARE: u8 = 0x20; + +/// Doorbell endpoint for work/message submissions. +const EP_DOORBELL: u8 = 0x21; + +/// Initialize the GPU firmware. +const MSG_INIT: u64 = 0x81 << 48; +const INIT_DATA_MASK: u64 = (1 << 44) - 1; + +/// TX channel doorbell. +const MSG_TX_DOORBELL: u64 = 0x83 << 48; +/// Firmware control channel doorbell. +const MSG_FWCTL: u64 = 0x84 << 48; +// /// Halt the firmware (?). +// const MSG_HALT: u64 = 0x85 << 48; + +/// Receive channel doorbell notification. +const MSG_RX_DOORBELL: u64 = 0x42 << 48; + +/// Doorbell number for firmware kicks/wakeups. +const DOORBELL_KICKFW: u64 = 0x10; +/// Doorbell number for device control channel kicks. +const DOORBELL_DEVCTRL: u64 = 0x11; + +// Upper kernel half VA address ranges. +/// Private (cached) firmware structure VA range base. +const IOVA_KERN_PRIV_BASE: u64 = 0xffffffa000000000; +/// Private (cached) firmware structure VA range top. +const IOVA_KERN_PRIV_TOP: u64 = 0xffffffa7ffffffff; +/// Shared (uncached) firmware structure VA range base. +const IOVA_KERN_SHARED_BASE: u64 = 0xffffffa800000000; +/// Shared (uncached) firmware structure VA range top. +const IOVA_KERN_SHARED_TOP: u64 = 0xffffffa9ffffffff; +/// Shared (uncached) read-only firmware structure VA range base. +const IOVA_KERN_SHARED_RO_BASE: u64 = 0xffffffaa00000000; +/// Shared (uncached) read-only firmware structure VA range top. +const IOVA_KERN_SHARED_RO_TOP: u64 = 0xffffffabffffffff; +/// GPU/FW shared structure VA range base. +const IOVA_KERN_GPU_BASE: u64 = 0xffffffaf00000000; +/// GPU/FW shared structure VA range top. +const IOVA_KERN_GPU_TOP: u64 = 0xffffffafffffffff; + +/// Timeout for entering the halt state after a fault or request. +const HALT_ENTER_TIMEOUT_MS: u64 = 100; + +/// Global allocators used for kernel-half structures. +pub(crate) struct KernelAllocators { + pub(crate) private: alloc::DefaultAllocator, + pub(crate) shared: alloc::DefaultAllocator, + pub(crate) shared_ro: alloc::DefaultAllocator, + pub(crate) gpu: alloc::DefaultAllocator, +} + +/// Receive (GPU->driver) ring buffer channels. +#[versions(AGX)] +struct RxChannels { + event: channel::EventChannel, + fw_log: channel::FwLogChannel, + ktrace: channel::KTraceChannel, + stats: channel::StatsChannel::ver, +} + +/// GPU work submission pipe channels (driver->GPU). +#[versions(AGX)] +struct PipeChannels { + pub(crate) vtx: Vec>, + pub(crate) frag: Vec>, + pub(crate) comp: Vec>, +} + +/// Misc command transmit (driver->GPU) channels. +#[versions(AGX)] +struct TxChannels { + pub(crate) device_control: channel::DeviceControlChannel::ver, +} + +/// Number of work submission pipes per type, one for each priority level. +const NUM_PIPES: usize = 4; + +/// A generic monotonically incrementing ID used to uniquely identify object instances within the +/// driver. +pub(crate) struct ID(AtomicU64); + +impl ID { + /// Create a new ID counter with a given value. + fn new(val: u64) -> ID { + ID(AtomicU64::new(val)) + } + + /// Fetch the next unique ID. + pub(crate) fn next(&self) -> u64 { + self.0.fetch_add(1, Ordering::Relaxed) + } +} + +impl Default for ID { + /// IDs default to starting at 2, as 0/1 are considered reserved for the system. + fn default() -> Self { + Self::new(2) + } +} + +/// A guard representing one active submission on the GPU. When dropped, decrements the active +/// submission count. +pub(crate) struct OpGuard(Arc); + +impl Drop for OpGuard { + fn drop(&mut self) { + self.0.end_op(); + } +} + +/// Set of global sequence IDs used in the driver. +#[derive(Default)] +pub(crate) struct SequenceIDs { + /// `File` instance ID. + pub(crate) file: ID, + /// `Vm` instance ID. + pub(crate) vm: ID, + /// Submission instance ID. + pub(crate) submission: ID, + /// `Queue` instance ID. + pub(crate) queue: ID, +} + +/// Top-level GPU manager that owns all the global state relevant to the driver instance. +#[versions(AGX)] +pub(crate) struct GpuManager { + dev: AsahiDevice, + cfg: &'static hw::HwConfig, + dyncfg: Box, + pub(crate) initdata: Box>, + uat: Box, + crashed: AtomicBool, + alloc: Mutex, + io_mappings: Vec, + rtkit: Mutex>>>, + rx_channels: Mutex>, + tx_channels: Mutex>, + fwctl_channel: Mutex>, + pipes: PipeChannels::ver, + event_manager: Arc, + buffer_mgr: buffer::BufferManager, + ids: SequenceIDs, +} + +/// Trait used to abstract the firmware/GPU-dependent variants of the GpuManager. +pub(crate) trait GpuManager: Send + Sync { + /// Cast as an Any type. + fn as_any(&self) -> &dyn Any; + /// Cast Arc as an Any type. + fn arc_as_any(self: Arc) -> Arc; + /// Initialize the GPU. + fn init(&self) -> Result; + /// Update the GPU globals from global info + /// + /// TODO: Unclear what can and cannot be updated like this. + fn update_globals(&self); + /// Get a reference to the KernelAllocators. + fn alloc(&self) -> Guard<'_, Mutex>; + /// Create a new `Vm` given a unique `File` ID. + fn new_vm(&self, file_id: u64) -> Result; + /// Bind a `Vm` to an available slot and return the `VmBind`. + fn bind_vm(&self, vm: &mmu::Vm) -> Result; + /// Create a new user command queue. + fn new_queue( + &self, + vm: mmu::Vm, + ualloc: Arc>, + ualloc_priv: Arc>, + priority: u32, + caps: u32, + ) -> Result>; + /// Return a reference to the global `SequenceIDs` instance. + fn ids(&self) -> &SequenceIDs; + /// Kick the firmware (wake it up if asleep). + /// + /// This should be useful to reduce latency on work submission, so we can ask the firmware to + /// wake up while we do some preparatory work for the work submission. + fn kick_firmware(&self) -> Result; + /// Invalidate a GPU scheduler context. Must be called before the relevant structures are freed. + fn invalidate_context( + &self, + context: &fw::types::GpuObject, + ) -> Result; + /// Flush the entire firmware cache. + /// + /// TODO: Does this actually work? + fn flush_fw_cache(&self) -> Result; + /// Handle a GPU work timeout event. + fn handle_timeout(&self, counter: u32, event_slot: u32); + /// Handle a GPU fault event. + fn handle_fault(&self); + /// Wait for the GPU to become idle and power off. + fn wait_for_poweroff(&self, timeout: usize) -> Result; + /// Send a firmware control command (secure cache flush). + fn fwctl(&self, msg: fw::channels::FwCtlMsg) -> Result; + /// Get the static GPU configuration for this SoC. + fn get_cfg(&self) -> &'static hw::HwConfig; + /// Get the dynamic GPU configuration for this SoC. + fn get_dyncfg(&self) -> &hw::DynConfig; +} + +/// Private generic trait for functions that don't need to escape this module. +trait GpuManagerPriv { + /// Decrement the pending submission counter. + fn end_op(&self); +} + +#[versions(AGX)] +#[vtable] +impl rtkit::Operations for GpuManager::ver { + type Data = Arc; + type Buffer = gem::ObjectRef; + + fn recv_message(data: ::Borrowed<'_>, ep: u8, msg: u64) { + let dev = &data.dev; + //dev_info!(dev, "RtKit message: {:#x}:{:#x}\n", ep, msg); + + if ep != EP_FIRMWARE || msg != MSG_RX_DOORBELL { + dev_err!(dev, "Unknown message: {:#x}:{:#x}\n", ep, msg); + return; + } + + let mut ch = data.rx_channels.lock(); + + ch.fw_log.poll(); + ch.ktrace.poll(); + ch.stats.poll(); + ch.event.poll(); + } + + fn crashed(data: ::Borrowed<'_>) { + let dev = &data.dev; + dev_err!(dev, "GPU firmware crashed, failing all jobs\n"); + + data.crashed.store(true, Ordering::Relaxed); + data.event_manager.fail_all(workqueue::WorkError::NoDevice); + } + + fn shmem_alloc( + data: ::Borrowed<'_>, + size: usize, + ) -> Result { + let dev = &data.dev; + mod_dev_dbg!(dev, "shmem_alloc() {:#x} bytes\n", size); + + let mut obj = gem::new_kernel_object(dev, size)?; + obj.vmap()?; + let iova = obj.map_into(data.uat.kernel_vm())?; + mod_dev_dbg!(dev, "shmem_alloc() -> VA {:#x}\n", iova); + Ok(obj) + } +} + +#[versions(AGX)] +impl GpuManager::ver { + /// Create a new GpuManager of this version/GPU combination. + #[inline(never)] + pub(crate) fn new( + dev: &AsahiDevice, + res: ®s::Resources, + cfg: &'static hw::HwConfig, + ) -> Result> { + let uat = Self::make_uat(dev, cfg)?; + let dyncfg = Self::make_dyncfg(dev, res, cfg, &uat)?; + + let mut alloc = KernelAllocators { + private: alloc::DefaultAllocator::new( + dev, + uat.kernel_vm(), + IOVA_KERN_PRIV_BASE, + IOVA_KERN_PRIV_TOP, + 0x80, + mmu::PROT_FW_PRIV_RW, + 1024 * 1024, + true, + fmt!("Kernel Private"), + true, + )?, + shared: alloc::DefaultAllocator::new( + dev, + uat.kernel_vm(), + IOVA_KERN_SHARED_BASE, + IOVA_KERN_SHARED_TOP, + 0x80, + mmu::PROT_FW_SHARED_RW, + 1024 * 1024, + true, + fmt!("Kernel Shared"), + false, + )?, + shared_ro: alloc::DefaultAllocator::new( + dev, + uat.kernel_vm(), + IOVA_KERN_SHARED_RO_BASE, + IOVA_KERN_SHARED_RO_TOP, + 0x80, + mmu::PROT_FW_SHARED_RO, + 64 * 1024, + true, + fmt!("Kernel RO Shared"), + false, + )?, + gpu: alloc::DefaultAllocator::new( + dev, + uat.kernel_vm(), + IOVA_KERN_GPU_BASE, + IOVA_KERN_GPU_TOP, + 0x80, + mmu::PROT_GPU_FW_SHARED_RW, + 64 * 1024, + true, + fmt!("Kernel GPU Shared"), + false, + )?, + }; + + let event_manager = Self::make_event_manager(&mut alloc)?; + let initdata = Self::make_initdata(cfg, &dyncfg, &mut alloc)?; + let mut mgr = Self::make_mgr(dev, cfg, dyncfg, uat, alloc, event_manager, initdata)?; + + { + let fwctl = mgr.fwctl_channel.lock(); + let p_fwctl = fwctl.to_raw(); + core::mem::drop(fwctl); + + mgr.initdata.fw_status.with_mut(|raw, _inner| { + raw.fwctl_channel = p_fwctl; + }); + } + + { + let txc = mgr.tx_channels.lock(); + let p_device_control = txc.device_control.to_raw(); + core::mem::drop(txc); + + let rxc = mgr.rx_channels.lock(); + let p_event = rxc.event.to_raw(); + let p_fw_log = rxc.fw_log.to_raw(); + let p_ktrace = rxc.ktrace.to_raw(); + let p_stats = rxc.stats.to_raw(); + let p_fwlog_buf = rxc.fw_log.get_buf(); + core::mem::drop(rxc); + + mgr.initdata.runtime_pointers.with_mut(|raw, _inner| { + raw.device_control = p_device_control; + raw.event = p_event; + raw.fw_log = p_fw_log; + raw.ktrace = p_ktrace; + raw.stats = p_stats; + raw.fwlog_buf = Some(p_fwlog_buf); + }); + } + + let mut p_pipes: Vec = Vec::new(); + + for ((v, f), c) in mgr + .pipes + .vtx + .iter() + .zip(&mgr.pipes.frag) + .zip(&mgr.pipes.comp) + { + p_pipes.try_push(fw::initdata::raw::PipeChannels::ver { + vtx: v.lock().to_raw(), + frag: f.lock().to_raw(), + comp: c.lock().to_raw(), + })?; + } + + mgr.initdata.runtime_pointers.with_mut(|raw, _inner| { + for (i, p) in p_pipes.into_iter().enumerate() { + raw.pipes[i].vtx = p.vtx; + raw.pipes[i].frag = p.frag; + raw.pipes[i].comp = p.comp; + } + }); + + for (i, map) in cfg.io_mappings.iter().enumerate() { + if let Some(map) = map.as_ref() { + mgr.iomap(i, map)?; + } + } + + let mgr = Arc::from(mgr); + + let rtkit = Box::try_new(rtkit::RtKit::::new( + dev, + None, + 0, + mgr.clone(), + )?)?; + + *mgr.rtkit.lock() = Some(rtkit); + + { + let mut rxc = mgr.rx_channels.lock(); + rxc.event.set_manager(mgr.clone()); + } + + Ok(mgr) + } + + /// Build the entire GPU InitData structure tree and return it as a boxed GpuObject. + fn make_initdata( + cfg: &'static hw::HwConfig, + dyncfg: &hw::DynConfig, + alloc: &mut KernelAllocators, + ) -> Result>> { + let mut builder = initdata::InitDataBuilder::ver::new(alloc, cfg, dyncfg); + builder.build() + } + + /// Create a fresh boxed Uat instance. + /// + /// Force disable inlining to avoid blowing up the stack. + #[inline(never)] + fn make_uat(dev: &AsahiDevice, cfg: &'static hw::HwConfig) -> Result> { + Ok(Box::try_new(mmu::Uat::new(dev, cfg)?)?) + } + + /// Actually create the final GpuManager instance, as a UniqueArc. + /// + /// Force disable inlining to avoid blowing up the stack. + #[inline(never)] + fn make_mgr( + dev: &AsahiDevice, + cfg: &'static hw::HwConfig, + dyncfg: Box, + uat: Box, + mut alloc: KernelAllocators, + event_manager: Arc, + initdata: Box>, + ) -> Result> { + let mut pipes = PipeChannels::ver { + vtx: Vec::new(), + frag: Vec::new(), + comp: Vec::new(), + }; + + for _i in 0..=NUM_PIPES - 1 { + pipes + .vtx + .try_push(Mutex::new(channel::PipeChannel::ver::new(dev, &mut alloc)?))?; + pipes + .frag + .try_push(Mutex::new(channel::PipeChannel::ver::new(dev, &mut alloc)?))?; + pipes + .comp + .try_push(Mutex::new(channel::PipeChannel::ver::new(dev, &mut alloc)?))?; + } + + UniqueArc::try_new(GpuManager::ver { + dev: dev.clone(), + cfg, + dyncfg, + initdata, + uat, + io_mappings: Vec::new(), + rtkit: Mutex::new(None), + crashed: AtomicBool::new(false), + rx_channels: Mutex::new(box_in_place!(RxChannels::ver { + event: channel::EventChannel::new(dev, &mut alloc, event_manager.clone())?, + fw_log: channel::FwLogChannel::new(dev, &mut alloc)?, + ktrace: channel::KTraceChannel::new(dev, &mut alloc)?, + stats: channel::StatsChannel::ver::new(dev, &mut alloc)?, + })?), + tx_channels: Mutex::new(Box::try_new(TxChannels::ver { + device_control: channel::DeviceControlChannel::ver::new(dev, &mut alloc)?, + })?), + fwctl_channel: Mutex::new(Box::try_new(channel::FwCtlChannel::new(dev, &mut alloc)?)?), + pipes, + event_manager, + buffer_mgr: buffer::BufferManager::new()?, + alloc: Mutex::new(alloc), + ids: Default::default(), + }) + } + + /// Fetch and validate the GPU dynamic configuration from the device tree and hardware. + /// + /// Force disable inlining to avoid blowing up the stack. + #[inline(never)] + fn make_dyncfg( + dev: &AsahiDevice, + res: ®s::Resources, + cfg: &'static hw::HwConfig, + uat: &mmu::Uat, + ) -> Result> { + let gpu_id = res.get_gpu_id()?; + + dev_info!(dev, "GPU Information:\n"); + dev_info!( + dev, + " Type: {:?}{:?}\n", + gpu_id.gpu_gen, + gpu_id.gpu_variant + ); + dev_info!(dev, " Max dies: {}\n", gpu_id.max_dies); + dev_info!(dev, " Clusters: {}\n", gpu_id.num_clusters); + dev_info!( + dev, + " Cores: {} ({})\n", + gpu_id.num_cores, + gpu_id.num_cores * gpu_id.num_clusters + ); + dev_info!( + dev, + " Frags: {} ({})\n", + gpu_id.num_frags, + gpu_id.num_frags * gpu_id.num_clusters + ); + dev_info!( + dev, + " GPs: {} ({})\n", + gpu_id.num_gps, + gpu_id.num_gps * gpu_id.num_clusters + ); + dev_info!(dev, " Core masks: {:#x?}\n", gpu_id.core_masks); + dev_info!(dev, " Active cores: {}\n", gpu_id.total_active_cores); + + dev_info!(dev, "Getting configuration from device tree...\n"); + let pwr_cfg = hw::PwrConfig::load(dev, cfg)?; + dev_info!(dev, "Dynamic configuration fetched\n"); + + if gpu_id.gpu_gen != cfg.gpu_gen || gpu_id.gpu_variant != cfg.gpu_variant { + dev_err!( + dev, + "GPU type mismatch (expected {:?}{:?}, found {:?}{:?})\n", + cfg.gpu_gen, + cfg.gpu_variant, + gpu_id.gpu_gen, + gpu_id.gpu_variant + ); + return Err(EIO); + } + if gpu_id.num_clusters > cfg.max_num_clusters { + dev_err!( + dev, + "Too many clusters ({} > {})\n", + gpu_id.num_clusters, + cfg.max_num_clusters + ); + return Err(EIO); + } + if gpu_id.num_cores > cfg.max_num_cores { + dev_err!( + dev, + "Too many cores ({} > {})\n", + gpu_id.num_cores, + cfg.max_num_cores + ); + return Err(EIO); + } + if gpu_id.num_frags > cfg.max_num_frags { + dev_err!( + dev, + "Too many frags ({} > {})\n", + gpu_id.num_frags, + cfg.max_num_frags + ); + return Err(EIO); + } + if gpu_id.num_gps > cfg.max_num_gps { + dev_err!( + dev, + "Too many GPs ({} > {})\n", + gpu_id.num_gps, + cfg.max_num_gps + ); + return Err(EIO); + } + + Ok(Box::try_new(hw::DynConfig { + pwr: pwr_cfg, + uat_ttb_base: uat.ttb_base(), + id: gpu_id, + })?) + } + + /// Create the global GPU event manager, and return an `Arc<>` to it. + fn make_event_manager(alloc: &mut KernelAllocators) -> Result> { + Arc::try_new(event::EventManager::new(alloc)?) + } + + /// Create a new MMIO mapping and add it to the mappings list in initdata at the specified + /// index. + fn iomap(&mut self, index: usize, map: &hw::IOMapping) -> Result { + let off = map.base & mmu::UAT_PGMSK; + let base = map.base - off; + let end = (map.base + map.size + mmu::UAT_PGMSK) & !mmu::UAT_PGMSK; + let mapping = self + .uat + .kernel_vm() + .map_io(base, end - base, map.writable)?; + + self.initdata.runtime_pointers.hwdata_b.with_mut(|raw, _| { + raw.io_mappings[index] = fw::initdata::raw::IOMapping { + phys_addr: U64(map.base as u64), + virt_addr: U64((mapping.iova() + off) as u64), + size: map.size as u32, + range_size: map.range_size as u32, + readwrite: U64(map.writable as u64), + }; + }); + + self.io_mappings.try_push(mapping)?; + Ok(()) + } + + /// Mark work associated with currently in-progress event slots as failed, after a fault or + /// timeout. + fn mark_pending_events(&self, culprit_slot: Option, error: workqueue::WorkError) { + dev_err!(self.dev, " Pending events:\n"); + + self.initdata.globals.with(|raw, _inner| { + for i in raw.pending_stamps.iter() { + let info = i.info.load(Ordering::Relaxed); + let wait_value = i.wait_value.load(Ordering::Relaxed); + + if info & 1 != 0 { + let slot = info >> 3; + let flags = info & 0x7; + dev_err!( + self.dev, + " [{}] flags={} value={:#x}\n", + slot, + flags, + wait_value + ); + let error = if culprit_slot.is_some() && culprit_slot != Some(slot) { + workqueue::WorkError::Killed + } else { + error + }; + self.event_manager.mark_error(slot, wait_value, error); + i.info.store(0, Ordering::Relaxed); + i.wait_value.store(0, Ordering::Relaxed); + } + } + }); + } + + /// Fetch the GPU MMU fault information from the hardware registers. + fn get_fault_info(&self) -> Option { + let data = self.dev.data(); + + let res = match data.resources() { + Some(res) => res, + None => { + dev_err!(self.dev, " Failed to acquire resources\n"); + return None; + } + }; + + let info = res.get_fault_info(); + if info.is_some() { + dev_err!(self.dev, " Fault info: {:#x?}\n", info.as_ref().unwrap()); + } + info + } + + /// Resume the GPU firmware after it halts (due to a timeout, fault, or request). + fn recover(&self) { + self.initdata.fw_status.with(|raw, _inner| { + let halt_count = raw.flags.halt_count.load(Ordering::Relaxed); + let mut halted = raw.flags.halted.load(Ordering::Relaxed); + dev_err!(self.dev, " Halt count: {}\n", halt_count); + dev_err!(self.dev, " Halted: {}\n", halted); + + if halted == 0 { + let timeout = time::ktime_get() + Duration::from_millis(HALT_ENTER_TIMEOUT_MS); + while time::ktime_get() < timeout { + halted = raw.flags.halted.load(Ordering::Relaxed); + if halted != 0 { + break; + } + mem::sync(); + } + halted = raw.flags.halted.load(Ordering::Relaxed); + } + + if debug_enabled(DebugFlags::NoGpuRecovery) { + dev_crit!(self.dev, " GPU recovery is disabled, wedging forever!\n"); + } else if halted != 0 { + dev_err!(self.dev, " Attempting recovery...\n"); + raw.flags.halted.store(0, Ordering::SeqCst); + raw.flags.resume.store(1, Ordering::SeqCst); + } else { + dev_err!(self.dev, " Cannot recover.\n"); + } + }); + } + + /// Return the packed GPU enabled core masks. + // Only used for some versions + #[allow(dead_code)] + pub(crate) fn core_masks_packed(&self) -> &[u32] { + self.dyncfg.id.core_masks_packed.as_slice() + } + + /// Kick a submission pipe for a submitted job to tell the firmware to start processing it. + pub(crate) fn run_job(&self, job: workqueue::JobSubmission::ver<'_>) -> Result { + mod_dev_dbg!(self.dev, "GPU: run_job\n"); + + let pipe_type = job.pipe_type(); + mod_dev_dbg!(self.dev, "GPU: run_job: pipe_type={:?}\n", pipe_type); + + let pipes = match pipe_type { + PipeType::Vertex => &self.pipes.vtx, + PipeType::Fragment => &self.pipes.frag, + PipeType::Compute => &self.pipes.comp, + }; + + let index: usize = job.priority() as usize; + let mut pipe = pipes.get(index).ok_or(EIO)?.lock(); + + mod_dev_dbg!(self.dev, "GPU: run_job: run()\n"); + job.run(&mut pipe); + mod_dev_dbg!(self.dev, "GPU: run_job: ring doorbell\n"); + + let mut guard = self.rtkit.lock(); + let rtk = guard.as_mut().unwrap(); + rtk.send_message( + EP_DOORBELL, + MSG_TX_DOORBELL | pipe_type as u64 | ((index as u64) << 2), + )?; + mod_dev_dbg!(self.dev, "GPU: run_job: done\n"); + + Ok(()) + } + + pub(crate) fn is_crashed(&self) -> bool { + self.crashed.load(Ordering::Relaxed) + } + + pub(crate) fn start_op(self: &Arc) -> Result { + if self.is_crashed() { + return Err(ENODEV); + } + + let val = self + .initdata + .globals + .with(|raw, _inner| raw.pending_submissions.fetch_add(1, Ordering::Acquire)); + + mod_dev_dbg!(self.dev, "OP start (pending: {})\n", val + 1); + self.kick_firmware()?; + Ok(OpGuard(self.clone())) + } +} + +#[versions(AGX)] +impl GpuManager for GpuManager::ver { + fn as_any(&self) -> &dyn Any { + self + } + + fn arc_as_any(self: Arc) -> Arc { + self as Arc + } + + fn init(&self) -> Result { + self.tx_channels.lock().device_control.send( + &fw::channels::DeviceControlMsg::ver::Initialize(Default::default()), + ); + + let initdata = self.initdata.gpu_va().get(); + let mut guard = self.rtkit.lock(); + let rtk = guard.as_mut().unwrap(); + + rtk.boot()?; + rtk.start_endpoint(EP_FIRMWARE)?; + rtk.start_endpoint(EP_DOORBELL)?; + rtk.send_message(EP_FIRMWARE, MSG_INIT | (initdata & INIT_DATA_MASK))?; + rtk.send_message(EP_DOORBELL, MSG_TX_DOORBELL | DOORBELL_DEVCTRL)?; + core::mem::drop(guard); + + self.kick_firmware()?; + Ok(()) + } + + fn update_globals(&self) { + let mut timeout: u32 = 2; + if debug_enabled(DebugFlags::WaitForPowerOff) { + timeout = 0; + } else if debug_enabled(DebugFlags::KeepGpuPowered) { + timeout = 5000; + } + + self.initdata.globals.with(|raw, _inner| { + raw.idle_off_delay_ms.store(timeout, Ordering::Relaxed); + }); + } + + fn alloc(&self) -> Guard<'_, Mutex> { + let mut guard = self.alloc.lock(); + let (garbage_count, garbage_bytes) = guard.private.garbage(); + if garbage_bytes > 1024 * 1024 { + mod_dev_dbg!( + self.dev, + "Collecting kalloc garbage ({} objects, {} bytes)\n", + garbage_count, + garbage_bytes + ); + if self.flush_fw_cache().is_err() { + dev_err!(self.dev, "Failed to flush FW cache\n"); + } else { + guard.private.collect_garbage(garbage_count); + } + } + + guard + } + + fn new_vm(&self, file_id: u64) -> Result { + self.uat.new_vm(self.ids.vm.next(), file_id) + } + + fn bind_vm(&self, vm: &mmu::Vm) -> Result { + self.uat.bind(vm) + } + + fn new_queue( + &self, + vm: mmu::Vm, + ualloc: Arc>, + ualloc_priv: Arc>, + priority: u32, + caps: u32, + ) -> Result> { + let mut kalloc = self.alloc(); + let id = self.ids.queue.next(); + Ok(Box::try_new(queue::Queue::ver::new( + &self.dev, + vm, + &mut kalloc, + ualloc, + ualloc_priv, + self.event_manager.clone(), + &self.buffer_mgr, + id, + priority, + caps, + )?)?) + } + + fn kick_firmware(&self) -> Result { + if self.is_crashed() { + return Err(ENODEV); + } + + let mut guard = self.rtkit.lock(); + let rtk = guard.as_mut().unwrap(); + rtk.send_message(EP_DOORBELL, MSG_TX_DOORBELL | DOORBELL_KICKFW)?; + + Ok(()) + } + + fn invalidate_context( + &self, + context: &fw::types::GpuObject, + ) -> Result { + mod_dev_dbg!( + self.dev, + "Invalidating GPU context @ {:?}\n", + context.weak_pointer() + ); + + if self.is_crashed() { + return Err(ENODEV); + } + + let mut guard = self.alloc.lock(); + let (garbage_count, _) = guard.private.garbage(); + + let dc = context.with( + |raw, _inner| fw::channels::DeviceControlMsg::ver::DestroyContext { + unk_4: 0, + ctx_23: raw.unk_23, + __pad0: Default::default(), + unk_c: 0, + unk_10: 0, + ctx_0: raw.unk_0, + ctx_1: raw.unk_1, + ctx_4: raw.unk_4, + __pad1: Default::default(), + unk_18: 0, + gpu_context: Some(context.weak_pointer()), + __pad2: Default::default(), + }, + ); + + mod_dev_dbg!(self.dev, "Context invalidation command: {:?}\n", &dc); + + let mut txch = self.tx_channels.lock(); + + let token = txch.device_control.send(&dc); + + { + let mut guard = self.rtkit.lock(); + let rtk = guard.as_mut().unwrap(); + rtk.send_message(EP_DOORBELL, MSG_TX_DOORBELL | DOORBELL_DEVCTRL)?; + } + + txch.device_control.wait_for(token)?; + + mod_dev_dbg!( + self.dev, + "GPU context invalidated: {:?}\n", + context.weak_pointer() + ); + + // The invalidation does a cache flush, so it is okay to collect garbage + guard.private.collect_garbage(garbage_count); + + Ok(()) + } + + fn flush_fw_cache(&self) -> Result { + mod_dev_dbg!(self.dev, "Flushing coprocessor data cache\n"); + + if self.is_crashed() { + return Err(ENODEV); + } + + // ctx_0 == 0xff or ctx_1 == 0xff cause no effect on context, + // but this command does a full cache flush too, so abuse it + // for that. + + let dc = fw::channels::DeviceControlMsg::ver::DestroyContext { + unk_4: 0, + ctx_23: 0, + __pad0: Default::default(), + unk_c: 0, + unk_10: 0, + ctx_0: 0xff, + ctx_1: 0xff, + ctx_4: 0, + __pad1: Default::default(), + unk_18: 0, + gpu_context: None, + __pad2: Default::default(), + }; + + let mut txch = self.tx_channels.lock(); + + let token = txch.device_control.send(&dc); + { + let mut guard = self.rtkit.lock(); + let rtk = guard.as_mut().unwrap(); + rtk.send_message(EP_DOORBELL, MSG_TX_DOORBELL | DOORBELL_DEVCTRL)?; + } + + txch.device_control.wait_for(token)?; + Ok(()) + } + + fn ids(&self) -> &SequenceIDs { + &self.ids + } + + fn handle_timeout(&self, counter: u32, event_slot: u32) { + dev_err!(self.dev, " (\\________/) \n"); + dev_err!(self.dev, " | | \n"); + dev_err!(self.dev, "'.| \\ , / |.'\n"); + dev_err!(self.dev, "--| / (( \\ |--\n"); + dev_err!(self.dev, ".'| _-_- |'.\n"); + dev_err!(self.dev, " |________| \n"); + dev_err!(self.dev, "** GPU timeout nya~!!!!! **\n"); + dev_err!(self.dev, " Event slot: {}\n", event_slot); + dev_err!(self.dev, " Timeout count: {}\n", counter); + + // If we have fault info, consider it a fault. + let error = match self.get_fault_info() { + Some(info) => workqueue::WorkError::Fault(info), + None => workqueue::WorkError::Timeout, + }; + self.mark_pending_events(Some(event_slot), error); + self.recover(); + } + + fn handle_fault(&self) { + dev_err!(self.dev, " (\\________/) \n"); + dev_err!(self.dev, " | | \n"); + dev_err!(self.dev, "'.| \\ , / |.'\n"); + dev_err!(self.dev, "--| / (( \\ |--\n"); + dev_err!(self.dev, ".'| _-_- |'.\n"); + dev_err!(self.dev, " |________| \n"); + dev_err!(self.dev, "GPU fault nya~!!!!!\n"); + let error = match self.get_fault_info() { + Some(info) => workqueue::WorkError::Fault(info), + None => workqueue::WorkError::Unknown, + }; + self.mark_pending_events(None, error); + self.recover(); + } + + fn wait_for_poweroff(&self, timeout: usize) -> Result { + self.initdata.runtime_pointers.hwdata_a.with(|raw, _inner| { + for _i in 0..timeout { + if raw.pwr_status.load(Ordering::Relaxed) == 4 { + return Ok(()); + } + coarse_sleep(Duration::from_millis(1)); + } + Err(ETIMEDOUT) + }) + } + + fn fwctl(&self, msg: fw::channels::FwCtlMsg) -> Result { + if self.is_crashed() { + return Err(ENODEV); + } + + let mut fwctl = self.fwctl_channel.lock(); + let token = fwctl.send(&msg); + { + let mut guard = self.rtkit.lock(); + let rtk = guard.as_mut().unwrap(); + rtk.send_message(EP_DOORBELL, MSG_FWCTL)?; + } + fwctl.wait_for(token)?; + Ok(()) + } + + fn get_cfg(&self) -> &'static hw::HwConfig { + self.cfg + } + + fn get_dyncfg(&self) -> &hw::DynConfig { + &self.dyncfg + } +} + +#[versions(AGX)] +impl GpuManagerPriv for GpuManager::ver { + fn end_op(&self) { + let val = self + .initdata + .globals + .with(|raw, _inner| raw.pending_submissions.fetch_sub(1, Ordering::Release)); + + mod_dev_dbg!(self.dev, "OP end (pending: {})\n", val - 1); + } +} diff --git a/drivers/gpu/drm/asahi/hw/mod.rs b/drivers/gpu/drm/asahi/hw/mod.rs new file mode 100644 index 000000000000..a92bb70aeae8 --- /dev/null +++ b/drivers/gpu/drm/asahi/hw/mod.rs @@ -0,0 +1,522 @@ +// SPDX-License-Identifier: GPL-2.0-only OR MIT + +//! Per-SoC hardware configuration structures +//! +//! This module contains the definitions used to store per-GPU and per-SoC configuration data. + +use crate::driver::AsahiDevice; +use crate::fw::types::*; +use alloc::vec::Vec; +use kernel::c_str; +use kernel::device::RawDevice; +use kernel::prelude::*; + +const MAX_POWERZONES: usize = 5; + +pub(crate) mod t600x; +pub(crate) mod t8103; +pub(crate) mod t8112; + +/// GPU generation enumeration. Note: Part of the UABI. +#[derive(Debug, PartialEq, Copy, Clone)] +#[repr(u32)] +pub(crate) enum GpuGen { + G13 = 13, + G14 = 14, +} + +/// GPU variant enumeration. Note: Part of the UABI. +#[derive(Debug, PartialEq, Copy, Clone)] +#[repr(u32)] +pub(crate) enum GpuVariant { + P = 'P' as u32, + G = 'G' as u32, + S = 'S' as u32, + C = 'C' as u32, + D = 'D' as u32, +} + +/// GPU revision enumeration. Note: Part of the UABI. +#[derive(Debug, PartialEq, Copy, Clone)] +#[repr(u32)] +pub(crate) enum GpuRevision { + A0 = 0x00, + A1 = 0x01, + B0 = 0x10, + B1 = 0x11, + C0 = 0x20, + C1 = 0x21, +} + +/// GPU core type enumeration. Note: Part of the firmware ABI. +#[derive(Debug, Copy, Clone)] +#[repr(u32)] +pub(crate) enum GpuCore { + // Unknown = 0, + // G5P = 1, + // G5G = 2, + // G9P = 3, + // G9G = 4, + // G10P = 5, + // G11P = 6, + // G11M = 7, + // G11G = 8, + // G12P = 9, + // G13P = 10, + G13G = 11, + G13S = 12, + G13C = 13, + // G14P = 14, + G14G = 15, +} + +/// GPU revision ID. Note: Part of the firmware ABI. +#[derive(Debug, PartialEq, Copy, Clone)] +#[repr(u32)] +pub(crate) enum GpuRevisionID { + // Unknown = 0, + A0 = 1, + A1 = 2, + B0 = 3, + B1 = 4, + C0 = 5, + C1 = 6, +} + +/// GPU driver/hardware features, from the UABI. +pub(crate) mod feat { + /// Backwards-compatible features. + pub(crate) mod compat {} + + /// Backwards-incompatible features. + pub(crate) mod incompat { + use kernel::bindings; + + /// Hardware requires Z/S compression to be mandatorily enabled. + pub(crate) const MANDATORY_ZS_COMPRESSION: u64 = + bindings::drm_asahi_feat_incompat_DRM_ASAHI_FEAT_MANDATORY_ZS_COMPRESSION as u64; + } +} + +/// A single performance state of the GPU. +#[derive(Debug)] +pub(crate) struct PState { + /// Voltage in millivolts, per GPU cluster. + pub(crate) volt_mv: Vec, + /// Frequency in hertz. + pub(crate) freq_hz: u32, + /// Maximum power consumption of the GPU at this pstate, in milliwatts. + pub(crate) pwr_mw: u32, +} + +/// A power zone definition (we have no idea what this is but Apple puts them in the DT). +#[allow(missing_docs)] +#[derive(Debug, Copy, Clone)] +pub(crate) struct PowerZone { + pub(crate) target: u32, + pub(crate) target_offset: u32, + pub(crate) filter_tc: u32, +} + +/// An MMIO mapping used by the firmware. +#[derive(Debug, Copy, Clone)] +pub(crate) struct IOMapping { + /// Base physical address of the mapping. + pub(crate) base: usize, + /// Size of the mapping. + pub(crate) size: usize, + /// Range size of the mapping (for arrays?) + pub(crate) range_size: usize, + /// Whether the mapping should be writable. + pub(crate) writable: bool, +} + +impl IOMapping { + /// Convenience constructor for a new IOMapping. + pub(crate) const fn new( + base: usize, + size: usize, + range_size: usize, + writable: bool, + ) -> IOMapping { + IOMapping { + base, + size, + range_size, + writable, + } + } +} + +/// Unknown HwConfigA fields that vary from SoC to SoC. +#[allow(missing_docs)] +#[derive(Debug, Copy, Clone)] +pub(crate) struct HwConfigA { + pub(crate) unk_87c: i32, + pub(crate) unk_8cc: u32, + pub(crate) unk_e24: u32, +} + +/// Unknown HwConfigB fields that vary from SoC to SoC. +#[allow(missing_docs)] +#[derive(Debug, Copy, Clone)] +pub(crate) struct HwConfigB { + pub(crate) unk_4e0: u64, + pub(crate) unk_534: u32, + pub(crate) unk_ab8: u32, + pub(crate) unk_abc: u32, + pub(crate) unk_b30: u32, +} + +/// Render command configs that vary from SoC to SoC. +#[derive(Debug, Copy, Clone)] +pub(crate) struct HwRenderConfig { + /// Vertex/tiling-related configuration register (lsb: disable clustering) + pub(crate) tiling_control: u32, +} + +/// Static hardware configuration for a given SoC model. +#[derive(Debug)] +pub(crate) struct HwConfig { + /// Chip ID in hex format (e.g. 0x8103 for t8103). + pub(crate) chip_id: u32, + /// GPU generation. + pub(crate) gpu_gen: GpuGen, + /// GPU variant type. + pub(crate) gpu_variant: GpuVariant, + /// GPU core type ID (as known by the firmware). + pub(crate) gpu_core: GpuCore, + /// Compatible feature bitmask for this GPU. + pub(crate) gpu_feat_compat: u64, + /// Incompatible feature bitmask for this GPU. + pub(crate) gpu_feat_incompat: u64, + + /// Base clock used used for timekeeping. + pub(crate) base_clock_hz: u32, + /// Output address space for the UAT on this SoC. + pub(crate) uat_oas: usize, + /// Maximum number of clusters on this SoC. + pub(crate) max_num_clusters: u32, + /// Maximum number of cores per cluster for this GPU. + pub(crate) max_num_cores: u32, + /// Maximum number of frags per cluster for this GPU. + pub(crate) max_num_frags: u32, + /// Maximum number of GPs per cluster for this GPU. + pub(crate) max_num_gps: u32, + + /// Required size of the first preemption buffer. + pub(crate) preempt1_size: usize, + /// Required size of the second preemption buffer. + pub(crate) preempt2_size: usize, + /// Required size of the third preemption buffer. + pub(crate) preempt3_size: usize, + + /// Rendering-relevant configuration. + pub(crate) render: HwRenderConfig, + + /// Misc HWDataA field values. + pub(crate) da: HwConfigA, + /// Misc HWDataB field values. + pub(crate) db: HwConfigB, + /// HwDataShared1.table. + pub(crate) shared1_tab: &'static [i32], + /// HwDataShared1.unk_a4. + pub(crate) shared1_a4: u32, + /// HwDataShared2.table. + pub(crate) shared2_tab: &'static [i32], + /// HwDataShared2.unk_508. + pub(crate) shared2_unk_508: u32, + /// Constant related to SRAM voltages. + pub(crate) sram_k: F32, + /// Unknown per-cluster coefficients 1. + pub(crate) unk_coef_a: &'static [&'static [F32]], + /// Unknown per-cluster coefficients 2. + pub(crate) unk_coef_b: &'static [&'static [F32]], + /// Unknown table in Global struct. + pub(crate) global_tab: Option<&'static [u8]>, + + /// Temperature sensor list (8 bits per sensor). + pub(crate) fast_die0_sensor_mask: u64, + /// Temperature sensor list (alternate). + pub(crate) fast_die0_sensor_mask_alt: u64, + /// Temperature sensor present bitmask. + pub(crate) fast_die0_sensor_present: u32, + /// Required MMIO mappings for this GPU/firmware. + pub(crate) io_mappings: &'static [Option], +} + +/// Dynamic (fetched from hardware/DT) configuration. +#[derive(Debug)] +pub(crate) struct DynConfig { + /// Base physical address of the UAT TTB (from DT reserved memory region). + pub(crate) uat_ttb_base: u64, + /// GPU ID configuration read from hardware. + pub(crate) id: GpuIdConfig, + /// Power calibration configuration for this specific chip/device. + pub(crate) pwr: PwrConfig, +} + +/// Specific GPU ID configuration fetched from SGX MMIO registers. +#[derive(Debug)] +pub(crate) struct GpuIdConfig { + /// GPU generation (should match static config). + pub(crate) gpu_gen: GpuGen, + /// GPU variant type (should match static config). + pub(crate) gpu_variant: GpuVariant, + /// GPU silicon revision. + pub(crate) gpu_rev: GpuRevision, + /// GPU silicon revision ID (firmware enum). + pub(crate) gpu_rev_id: GpuRevisionID, + /// Maximum number of dies supported. + pub(crate) max_dies: u32, + /// Total number of GPU clusters. + pub(crate) num_clusters: u32, + /// Maximum number of GPU cores per cluster. + pub(crate) num_cores: u32, + /// Number of frags per cluster. + pub(crate) num_frags: u32, + /// Number of GPs per cluster. + pub(crate) num_gps: u32, + /// Total number of active cores for the whole GPU. + pub(crate) total_active_cores: u32, + /// Mask of active cores per cluster. + pub(crate) core_masks: Vec, + /// Packed mask of all active cores. + pub(crate) core_masks_packed: Vec, +} + +/// Configurable GPU power settings from the device tree. +#[derive(Debug)] +pub(crate) struct PwrConfig { + /// GPU performance state list. + pub(crate) perf_states: Vec, + /// GPU power zone list. + pub(crate) power_zones: Vec, + + /// Core leakage coefficient per cluster. + pub(crate) core_leak_coef: Vec, + /// SRAM leakage coefficient per cluster. + pub(crate) sram_leak_coef: Vec, + + /// Maximum total power of the GPU in milliwatts. + pub(crate) max_power_mw: u32, + /// Maximum frequency of the GPU in megahertz. + pub(crate) max_freq_mhz: u32, + + /// Minimum performance state to start at. + pub(crate) perf_base_pstate: u32, + /// Maximum enabled performance state. + pub(crate) perf_max_pstate: u32, + + /// Minimum voltage for the SRAM power domain in microvolts. + pub(crate) min_sram_microvolt: u32, + + // Most of these fields are just named after Apple ADT property names and we don't fully + // understand them. They configure various power-related PID loops and filters. + /// Average power filter time constant in milliseconds. + pub(crate) avg_power_filter_tc_ms: u32, + /// Average power filter PID integral gain? + pub(crate) avg_power_ki_only: F32, + /// Average power filter PID proportional gain? + pub(crate) avg_power_kp: F32, + pub(crate) avg_power_min_duty_cycle: u32, + /// Average power target filter time constant in periods. + pub(crate) avg_power_target_filter_tc: u32, + /// "Fast die0" (temperature?) PID integral gain. + pub(crate) fast_die0_integral_gain: F32, + /// "Fast die0" (temperature?) PID proportional gain. + pub(crate) fast_die0_proportional_gain: F32, + pub(crate) fast_die0_prop_tgt_delta: u32, + pub(crate) fast_die0_release_temp: u32, + /// Delay from the fender (?) becoming idle to powerdown + pub(crate) fender_idle_off_delay_ms: u32, + /// Timeout from firmware early wake to sleep if no work was submitted (?) + pub(crate) fw_early_wake_timeout_ms: u32, + /// Delay from the GPU becoming idle to powerdown + pub(crate) idle_off_delay_ms: u32, + /// Percent? + pub(crate) perf_boost_ce_step: u32, + /// Minimum utilization before performance state is increased in %. + pub(crate) perf_boost_min_util: u32, + pub(crate) perf_filter_drop_threshold: u32, + /// Performance PID filter time constant? (periods?) + pub(crate) perf_filter_time_constant: u32, + /// Performance PID filter time constant 2? (periods?) + pub(crate) perf_filter_time_constant2: u32, + /// Performance PID integral gain. + pub(crate) perf_integral_gain: F32, + /// Performance PID integral gain 2 (?). + pub(crate) perf_integral_gain2: F32, + pub(crate) perf_integral_min_clamp: u32, + /// Performance PID proportional gain. + pub(crate) perf_proportional_gain: F32, + /// Performance PID proportional gain 2 (?). + pub(crate) perf_proportional_gain2: F32, + pub(crate) perf_reset_iters: u32, + /// Target GPU utilization for the performance controller in %. + pub(crate) perf_tgt_utilization: u32, + /// Power sampling period in milliseconds. + pub(crate) power_sample_period: u32, + /// PPM (?) filter time constant in milliseconds. + pub(crate) ppm_filter_time_constant_ms: u32, + /// PPM (?) filter PID integral gain. + pub(crate) ppm_ki: F32, + /// PPM (?) filter PID proportional gain. + pub(crate) ppm_kp: F32, + /// Power consumption filter time constant (periods?) + pub(crate) pwr_filter_time_constant: u32, + /// Power consumption filter PID integral gain. + pub(crate) pwr_integral_gain: F32, + pub(crate) pwr_integral_min_clamp: u32, + pub(crate) pwr_min_duty_cycle: u32, + pub(crate) pwr_proportional_gain: F32, +} + +impl PwrConfig { + /// Load the GPU power configuration from the device tree. + pub(crate) fn load(dev: &AsahiDevice, cfg: &HwConfig) -> Result { + let mut perf_states = Vec::new(); + + let node = dev.of_node().ok_or(EIO)?; + let opps = node + .parse_phandle(c_str!("operating-points-v2"), 0) + .ok_or(EIO)?; + + let mut max_power_mw: u32 = 0; + let mut max_freq_mhz: u32 = 0; + + macro_rules! prop { + ($prop:expr, $default:expr) => {{ + node.get_opt_property(c_str!($prop)) + .map_err(|e| { + dev_err!(dev, "Error reading property {}: {:?}\n", $prop, e); + e + })? + .unwrap_or($default) + }}; + ($prop:expr) => {{ + node.get_property(c_str!($prop)).map_err(|e| { + dev_err!(dev, "Error reading property {}: {:?}\n", $prop, e); + e + })? + }}; + } + + for opp in opps.children() { + let freq_hz: u64 = opp.get_property(c_str!("opp-hz"))?; + let mut volt_uv: Vec = opp.get_property(c_str!("opp-microvolt"))?; + let pwr_uw: u32 = opp.get_property(c_str!("opp-microwatt"))?; + + if volt_uv.len() != cfg.max_num_clusters as usize { + dev_err!( + dev, + "Invalid opp-microvolt length (expected {}, got {})\n", + cfg.max_num_clusters, + volt_uv.len() + ); + return Err(EINVAL); + } + + volt_uv.iter_mut().for_each(|a| *a /= 1000); + let volt_mv = volt_uv; + + let pwr_mw = pwr_uw / 1000; + max_power_mw = max_power_mw.max(pwr_mw); + + let freq_mhz: u32 = (freq_hz / 1_000_000).try_into()?; + max_freq_mhz = max_freq_mhz.max(freq_mhz); + + perf_states.try_push(PState { + freq_hz: freq_hz.try_into()?, + volt_mv, + pwr_mw, + })?; + } + + let pz_data = prop!("apple,power-zones", Vec::new()); + + if pz_data.len() > 3 * MAX_POWERZONES || pz_data.len() % 3 != 0 { + dev_err!(dev, "Invalid apple,power-zones value\n"); + return Err(EINVAL); + } + + let pz_count = pz_data.len() / 3; + let mut power_zones = Vec::new(); + for i in (0..pz_count).step_by(3) { + power_zones.try_push(PowerZone { + target: pz_data[i], + target_offset: pz_data[i + 1], + filter_tc: pz_data[i + 2], + })?; + } + + let core_leak_coef: Vec = prop!("apple,core-leak-coef"); + let sram_leak_coef: Vec = prop!("apple,sram-leak-coef"); + + if core_leak_coef.len() != cfg.max_num_clusters as usize { + dev_err!(dev, "Invalid apple,core-leak-coef\n"); + return Err(EINVAL); + } + if sram_leak_coef.len() != cfg.max_num_clusters as usize { + dev_err!(dev, "Invalid apple,sram_leak_coef\n"); + return Err(EINVAL); + } + + Ok(PwrConfig { + core_leak_coef, + sram_leak_coef, + + max_power_mw, + max_freq_mhz, + + perf_base_pstate: prop!("apple,perf-base-pstate", 1), + perf_max_pstate: perf_states.len() as u32 - 1, + min_sram_microvolt: prop!("apple,min-sram-microvolt"), + + avg_power_filter_tc_ms: prop!("apple,avg-power-filter-tc-ms"), + avg_power_ki_only: prop!("apple,avg-power-ki-only"), + avg_power_kp: prop!("apple,avg-power-kp"), + avg_power_min_duty_cycle: prop!("apple,avg-power-min-duty-cycle"), + avg_power_target_filter_tc: prop!("apple,avg-power-target-filter-tc"), + fast_die0_integral_gain: prop!("apple,fast-die0-integral-gain"), + fast_die0_proportional_gain: prop!("apple,fast-die0-proportional-gain"), + fast_die0_prop_tgt_delta: prop!("apple,fast-die0-prop-tgt-delta", 0), + fast_die0_release_temp: prop!("apple,fast-die0-release-temp", 80), + fender_idle_off_delay_ms: prop!("apple,fender-idle-off-delay-ms", 40), + fw_early_wake_timeout_ms: prop!("apple,fw-early-wake-timeout-ms", 5), + idle_off_delay_ms: prop!("apple,idle-off-delay-ms", 2), + perf_boost_ce_step: prop!("apple,perf-boost-ce-step", 25), + perf_boost_min_util: prop!("apple,perf-boost-min-util", 100), + perf_filter_drop_threshold: prop!("apple,perf-filter-drop-threshold"), + perf_filter_time_constant2: prop!("apple,perf-filter-time-constant2"), + perf_filter_time_constant: prop!("apple,perf-filter-time-constant"), + perf_integral_gain2: prop!("apple,perf-integral-gain2"), + perf_integral_gain: prop!("apple,perf-integral-gain", f32!(7.8956833)), + perf_integral_min_clamp: prop!("apple,perf-integral-min-clamp"), + perf_proportional_gain2: prop!("apple,perf-proportional-gain2"), + perf_proportional_gain: prop!("apple,perf-proportional-gain", f32!(14.707963)), + perf_reset_iters: prop!("apple,perf-reset-iters", 6), + perf_tgt_utilization: prop!("apple,perf-tgt-utilization"), + power_sample_period: prop!("apple,power-sample-period"), + ppm_filter_time_constant_ms: prop!("apple,ppm-filter-time-constant-ms"), + ppm_ki: prop!("apple,ppm-ki"), + ppm_kp: prop!("apple,ppm-kp"), + pwr_filter_time_constant: prop!("apple,pwr-filter-time-constant", 313), + pwr_integral_gain: prop!("apple,pwr-integral-gain", f32!(0.0202129)), + pwr_integral_min_clamp: prop!("apple,pwr-integral-min-clamp", 0), + pwr_min_duty_cycle: prop!("apple,pwr-min-duty-cycle"), + pwr_proportional_gain: prop!("apple,pwr-proportional-gain", f32!(5.2831855)), + + perf_states, + power_zones, + }) + } + + pub(crate) fn min_frequency_khz(&self) -> u32 { + self.perf_states[self.perf_base_pstate as usize].freq_hz / 1000 + } + + pub(crate) fn max_frequency_khz(&self) -> u32 { + self.perf_states[self.perf_max_pstate as usize].freq_hz / 1000 + } +} diff --git a/drivers/gpu/drm/asahi/hw/t600x.rs b/drivers/gpu/drm/asahi/hw/t600x.rs new file mode 100644 index 000000000000..8a8267a7e18a --- /dev/null +++ b/drivers/gpu/drm/asahi/hw/t600x.rs @@ -0,0 +1,140 @@ +// SPDX-License-Identifier: GPL-2.0-only OR MIT + +//! Hardware configuration for t600x (M1 Pro/Max/Ultra) platforms. + +use crate::f32; + +use super::*; + +const fn iomaps(mcc_count: usize, has_die1: bool) -> [Option; 20] { + [ + Some(IOMapping::new(0x404d00000, 0x1c000, 0x1c000, true)), // Fender + Some(IOMapping::new(0x20e100000, 0x4000, 0x4000, false)), // AICTimer + Some(IOMapping::new(0x28e104000, 0x4000, 0x4000, true)), // AICSWInt + Some(IOMapping::new(0x404000000, 0x20000, 0x20000, true)), // RGX + None, // UVD + None, // unused + None, // DisplayUnderrunWA + Some(IOMapping::new(0x28e494000, 0x1000, 0x1000, false)), // AnalogTempSensorControllerRegs + None, // PMPDoorbell + Some(IOMapping::new(0x404d80000, 0x8000, 0x8000, true)), // MetrologySensorRegs + Some(IOMapping::new(0x204d61000, 0x1000, 0x1000, true)), // GMGIFAFRegs + Some(IOMapping::new( + 0x200000000, + mcc_count * 0xd8000, + 0xd6400, + true, + )), // MCache registers + None, // AICBankedRegisters + None, // PMGRScratch + Some(IOMapping::new(0x2643c4000, 0x1000, 0x1000, true)), // NIA Special agent idle register die 0 + if has_die1 { + // NIA Special agent idle register die 1 + Some(IOMapping::new(0x22643c4000, 0x1000, 0x1000, true)) + } else { + None + }, + None, // CRE registers + None, // Streaming codec registers + Some(IOMapping::new(0x28e3d0000, 0x1000, 0x1000, true)), // ? + Some(IOMapping::new(0x28e3c0000, 0x1000, 0x1000, false)), // ? + ] +} + +pub(crate) const HWCONFIG_T6002: super::HwConfig = HwConfig { + chip_id: 0x6002, + gpu_gen: GpuGen::G13, + gpu_variant: GpuVariant::D, + gpu_core: GpuCore::G13C, + gpu_feat_compat: 0, + gpu_feat_incompat: feat::incompat::MANDATORY_ZS_COMPRESSION, + + base_clock_hz: 24_000_000, + uat_oas: 42, + max_num_clusters: 8, + max_num_cores: 8, + max_num_frags: 8, + max_num_gps: 4, + + preempt1_size: 0x540, + preempt2_size: 0x280, + preempt3_size: 0x20, + + render: HwRenderConfig { + tiling_control: 0xa540, + }, + + da: HwConfigA { + unk_87c: 900, + unk_8cc: 11000, + unk_e24: 125, + }, + db: HwConfigB { + unk_4e0: 4, + unk_534: 1, + unk_ab8: 0x2084, + unk_abc: 0x80, + unk_b30: 0, + }, + shared1_tab: &[ + 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, + 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, + ], + shared1_a4: 0xffff, + shared2_tab: &[-1, -1, -1, -1, 0x2aa, 0xaaa, -1, -1, 0, 0], + shared2_unk_508: 0xcc00001, + sram_k: f32!(1.02), + unk_coef_a: &[ + &f32!([9.838]), + &f32!([9.819]), + &f32!([9.826]), + &f32!([9.799]), + &f32!([9.799]), + &f32!([9.826]), + &f32!([9.819]), + &f32!([9.838]), + ], + unk_coef_b: &[ + &f32!([13.0]), + &f32!([13.0]), + &f32!([13.0]), + &f32!([13.0]), + &f32!([13.0]), + &f32!([13.0]), + &f32!([13.0]), + &f32!([13.0]), + ], + global_tab: Some(&[ + 0, 1, 2, 1, 1, 90, 75, 1, 1, 1, 2, 90, 75, 1, 1, 1, 1, 90, 75, 1, 1, + ]), + fast_die0_sensor_mask: 0x8080808080808080, + fast_die0_sensor_mask_alt: 0x9090909090909090, + fast_die0_sensor_present: 0xff, + io_mappings: &iomaps(16, true), +}; + +pub(crate) const HWCONFIG_T6001: super::HwConfig = HwConfig { + chip_id: 0x6001, + gpu_variant: GpuVariant::C, + gpu_core: GpuCore::G13C, + + max_num_clusters: 4, + fast_die0_sensor_mask: 0x80808080, + fast_die0_sensor_mask_alt: 0x90909090, + fast_die0_sensor_present: 0x0f, + io_mappings: &iomaps(8, false), + ..HWCONFIG_T6002 +}; + +pub(crate) const HWCONFIG_T6000: super::HwConfig = HwConfig { + chip_id: 0x6000, + gpu_variant: GpuVariant::S, + gpu_core: GpuCore::G13S, + + max_num_clusters: 2, + fast_die0_sensor_mask: 0x8080, + fast_die0_sensor_mask_alt: 0x9090, + fast_die0_sensor_present: 0x03, + io_mappings: &iomaps(4, false), + ..HWCONFIG_T6001 +}; diff --git a/drivers/gpu/drm/asahi/hw/t8103.rs b/drivers/gpu/drm/asahi/hw/t8103.rs new file mode 100644 index 000000000000..3d38b088a0f5 --- /dev/null +++ b/drivers/gpu/drm/asahi/hw/t8103.rs @@ -0,0 +1,80 @@ +// SPDX-License-Identifier: GPL-2.0-only OR MIT + +//! Hardware configuration for t8103 platforms (M1). + +use crate::f32; + +use super::*; + +pub(crate) const HWCONFIG: super::HwConfig = HwConfig { + chip_id: 0x8103, + gpu_gen: GpuGen::G13, + gpu_variant: GpuVariant::G, + gpu_core: GpuCore::G13G, + gpu_feat_compat: 0, + gpu_feat_incompat: 0, + + base_clock_hz: 24_000_000, + uat_oas: 40, + max_num_clusters: 1, + max_num_cores: 8, + max_num_frags: 8, + max_num_gps: 4, + + preempt1_size: 0x540, + preempt2_size: 0x280, + preempt3_size: 0x20, + + render: HwRenderConfig { + // bit 0: disable clustering (always) + tiling_control: 0xa041, + }, + + da: HwConfigA { + unk_87c: -220, + unk_8cc: 9880, + unk_e24: 112, + }, + db: HwConfigB { + unk_4e0: 0, + unk_534: 0, + unk_ab8: 0x48, + unk_abc: 0x8, + unk_b30: 0, + }, + shared1_tab: &[ + -1, 0x7282, 0x50ea, 0x370a, 0x25be, 0x1c1f, 0x16fb, -1, -1, -1, -1, -1, -1, -1, -1, -1, + ], + shared1_a4: 0xffff, + shared2_tab: &[0x800, 0x1555, -1, -1, -1, -1, -1, -1, 0, 0], + shared2_unk_508: 0xc00007, + sram_k: f32!(1.02), + unk_coef_a: &[], + unk_coef_b: &[], + global_tab: None, + fast_die0_sensor_mask: 0x12, + fast_die0_sensor_mask_alt: 0x12, + fast_die0_sensor_present: 0x01, + io_mappings: &[ + Some(IOMapping::new(0x204d00000, 0x1c000, 0x1c000, true)), // Fender + Some(IOMapping::new(0x20e100000, 0x4000, 0x4000, false)), // AICTimer + Some(IOMapping::new(0x23b104000, 0x4000, 0x4000, true)), // AICSWInt + Some(IOMapping::new(0x204000000, 0x20000, 0x20000, true)), // RGX + None, // UVD + None, // unused + None, // DisplayUnderrunWA + Some(IOMapping::new(0x23b2e8000, 0x1000, 0x1000, false)), // AnalogTempSensorControllerRegs + Some(IOMapping::new(0x23bc00000, 0x1000, 0x1000, true)), // PMPDoorbell + Some(IOMapping::new(0x204d80000, 0x5000, 0x5000, true)), // MetrologySensorRegs + Some(IOMapping::new(0x204d61000, 0x1000, 0x1000, true)), // GMGIFAFRegs + Some(IOMapping::new(0x200000000, 0xd6400, 0xd6400, true)), // MCache registers + None, // AICBankedRegisters + Some(IOMapping::new(0x23b738000, 0x1000, 0x1000, true)), // PMGRScratch + None, // NIA Special agent idle register die 0 + None, // NIA Special agent idle register die 1 + None, // CRE registers + None, // Streaming codec registers + None, // + None, // + ], +}; diff --git a/drivers/gpu/drm/asahi/hw/t8112.rs b/drivers/gpu/drm/asahi/hw/t8112.rs new file mode 100644 index 000000000000..5624dca130be --- /dev/null +++ b/drivers/gpu/drm/asahi/hw/t8112.rs @@ -0,0 +1,82 @@ +// SPDX-License-Identifier: GPL-2.0-only OR MIT + +//! Hardware configuration for t8112 platforms (M2). + +use crate::f32; + +use super::*; + +pub(crate) const HWCONFIG: super::HwConfig = HwConfig { + chip_id: 0x8112, + gpu_gen: GpuGen::G14, + gpu_variant: GpuVariant::G, + gpu_core: GpuCore::G14G, + gpu_feat_compat: 0, + gpu_feat_incompat: 0, + + base_clock_hz: 24_000_000, + uat_oas: 40, + max_num_clusters: 1, + max_num_cores: 10, + max_num_frags: 10, + max_num_gps: 4, + + preempt1_size: 0x540, + preempt2_size: 0x280, + preempt3_size: 0x20, + + render: HwRenderConfig { + // TODO: this is unused here, may be present in newer FW + tiling_control: 0xa041, + }, + + da: HwConfigA { + unk_87c: 900, + unk_8cc: 11000, + unk_e24: 125, + }, + db: HwConfigB { + unk_4e0: 4, + unk_534: 0, + unk_ab8: 0x2048, + unk_abc: 0x4000, + unk_b30: 1, + }, + shared1_tab: &[ + 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, + 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, + ], + shared1_a4: 0, + shared2_tab: &[-1, -1, -1, -1, -1, -1, -1, -1, 0xaa5aa, 0], + shared2_unk_508: 0xc00000, + sram_k: f32!(1.02), + // 13.2: last coef changed from 6.6 to 5.3, assuming that was a fix we can backport + unk_coef_a: &[&f32!([0.0, 0.0, 0.0, 0.0, 5.3, 0.0, 5.3, /*6.6*/ 5.3])], + unk_coef_b: &[&f32!([0.0, 0.0, 0.0, 0.0, 5.3, 0.0, 5.3, /*6.6*/ 5.3])], + global_tab: None, + fast_die0_sensor_mask: 0x6800, + fast_die0_sensor_mask_alt: 0x6800, + fast_die0_sensor_present: 0x02, + io_mappings: &[ + Some(IOMapping::new(0x204d00000, 0x14000, 0x14000, true)), // Fender + Some(IOMapping::new(0x20e100000, 0x4000, 0x4000, false)), // AICTimer + Some(IOMapping::new(0x23b0c4000, 0x4000, 0x4000, true)), // AICSWInt + Some(IOMapping::new(0x204000000, 0x20000, 0x20000, true)), // RGX + None, // UVD + None, // unused + None, // DisplayUnderrunWA + Some(IOMapping::new(0x23b2c0000, 0x1000, 0x1000, false)), // AnalogTempSensorControllerRegs + None, // PMPDoorbell + Some(IOMapping::new(0x204d80000, 0x8000, 0x8000, true)), // MetrologySensorRegs + Some(IOMapping::new(0x204d61000, 0x1000, 0x1000, true)), // GMGIFAFRegs + Some(IOMapping::new(0x200000000, 0xd6400, 0xd6400, true)), // MCache registers + None, // AICBankedRegisters + None, // PMGRScratch + None, // NIA Special agent idle register die 0 + None, // NIA Special agent idle register die 1 + Some(IOMapping::new(0x204e00000, 0x10000, 0x10000, true)), // CRE registers + Some(IOMapping::new(0x27d050000, 0x4000, 0x4000, true)), // Streaming codec registers + Some(IOMapping::new(0x23b3d0000, 0x1000, 0x1000, true)), // + Some(IOMapping::new(0x23b3c0000, 0x1000, 0x1000, true)), // + ], +}; diff --git a/drivers/gpu/drm/asahi/initdata.rs b/drivers/gpu/drm/asahi/initdata.rs new file mode 100644 index 000000000000..472c42169130 --- /dev/null +++ b/drivers/gpu/drm/asahi/initdata.rs @@ -0,0 +1,777 @@ +// SPDX-License-Identifier: GPL-2.0-only OR MIT +#![allow(clippy::unusual_byte_groupings)] + +//! GPU initialization data builder. +//! +//! The root of all interaction between the GPU firmware and the host driver is a complex set of +//! nested structures that we call InitData. This includes both GPU hardware/firmware configuration +//! and the pointers to the ring buffers and global data fields that are used for communication at +//! runtime. +//! +//! Many of these structures are poorly understood, so there are lots of hardcoded unknown values +//! derived from observing the InitData structures that macOS generates. + +use crate::fw::initdata::*; +use crate::fw::types::*; +use crate::{box_in_place, f32, place}; +use crate::{gpu, hw, mmu}; +use kernel::error::Result; +use kernel::macros::versions; + +/// Builder helper for the global GPU InitData. +#[versions(AGX)] +pub(crate) struct InitDataBuilder<'a> { + alloc: &'a mut gpu::KernelAllocators, + cfg: &'a hw::HwConfig, + dyncfg: &'a hw::DynConfig, +} + +#[versions(AGX)] +impl<'a> InitDataBuilder::ver<'a> { + /// Create a new InitData builder + pub(crate) fn new( + alloc: &'a mut gpu::KernelAllocators, + cfg: &'a hw::HwConfig, + dyncfg: &'a hw::DynConfig, + ) -> InitDataBuilder::ver<'a> { + InitDataBuilder::ver { alloc, cfg, dyncfg } + } + + /// Create the HwDataShared1 structure, which is used in two places in InitData. + #[inline(never)] + fn hw_shared1(cfg: &hw::HwConfig) -> raw::HwDataShared1 { + let mut ret = raw::HwDataShared1 { + unk_a4: cfg.shared1_a4, + ..Default::default() + }; + for (i, val) in cfg.shared1_tab.iter().enumerate() { + ret.table[i] = *val; + } + ret + } + + fn init_curve( + curve: &mut raw::HwDataShared2Curve, + unk_0: u32, + unk_4: u32, + t1: &[i16], + t2: &[i16], + t3: &[&[i32]], + ) { + curve.unk_0 = unk_0; + curve.unk_4 = unk_4; + (*curve.t1)[..t1.len()].copy_from_slice(t1); + (*curve.t1)[t1.len()..].fill(t1[0]); + (*curve.t2)[..t2.len()].copy_from_slice(t2); + (*curve.t2)[t2.len()..].fill(t2[0]); + for (i, a) in curve.t3.iter_mut().enumerate() { + a.fill(0x3ffffff); + if i < t3.len() { + let b = t3[i]; + (**a)[..b.len()].copy_from_slice(b); + } + } + } + + /// Create the HwDataShared2 structure, which is used in two places in InitData. + #[inline(never)] + fn hw_shared2(cfg: &hw::HwConfig) -> Result> { + let mut ret = box_in_place!(raw::HwDataShared2 { + unk_28: Array::new([0xff; 16]), + t8112: Default::default(), + unk_508: cfg.shared2_unk_508, + ..Default::default() + })?; + + for (i, val) in cfg.shared2_tab.iter().enumerate() { + ret.table[i] = *val; + } + + if cfg.chip_id == 0x8112 { + ret.t8112.unk_14 = 0x6000000; + Self::init_curve(&mut ret.t8112.curve1, 0, 0x20000000, &[-1], &[0x0f07], &[]); + Self::init_curve( + &mut ret.t8112.curve2, + 7, + 0x80000000, + &[-1, 25740, 17429, 12550, 9597, 7910, 6657, 5881, 5421], + &[ + 0x0f07, 0x04c0, 0x06c0, 0x08c0, 0x0ac0, 0x0c40, 0x0dc0, 0x0ec0, 0x0f80, + ], + &[ + &[0x3ffffff, 107, 101, 94, 87, 82, 77, 73, 71], + &[ + 0x3ffffff, 38240, 36251, 33562, 31368, 29379, 27693, 26211, 25370, + ], + &[ + 0x3ffffff, 123933, 117485, 108771, 101661, 95217, 89751, 84948, 82222, + ], + ], + ); + } + + Ok(ret) + } + + /// Create the HwDataShared3 structure, which is used in two places in InitData. + #[inline(never)] + fn hw_shared3(cfg: &hw::HwConfig) -> Result> { + let mut ret = box_in_place!(raw::HwDataShared3 { + ..Default::default() + })?; + + if cfg.chip_id == 0x8112 { + ret.unk_0 = 1; + ret.unk_4 = 500; + ret.unk_8 = 5; + ret.table.copy_from_slice(&[ + 10700, 10700, 10700, 10700, 10700, 6000, 1000, 1000, 1000, 10700, 10700, 10700, + 10700, 10700, 10700, 10700, + ]); + ret.unk_4c = 1; + } + + Ok(ret) + } + + /// Create an unknown T81xx-specific data structure. + fn t81xx_data(dyncfg: &'a hw::DynConfig) -> raw::T81xxData { + raw::T81xxData { + unk_d8c: 0x80000000, + unk_d90: 4, + unk_d9c: f32!(0.6), + unk_da4: f32!(0.4), + unk_dac: f32!(0.38552), + unk_db8: f32!(65536.0), + unk_dbc: f32!(13.56), + max_pstate_scaled: 100 * dyncfg.pwr.perf_max_pstate, + ..Default::default() + } + } + + /// Create the HwDataA structure. This mostly contains power-related configuration. + #[inline(never)] + fn hwdata_a(&mut self) -> Result> { + self.alloc + .private + .new_inplace(Default::default(), |_inner, ptr| { + let pwr = &self.dyncfg.pwr; + let period_ms = pwr.power_sample_period; + let period_s = F32::from(period_ms) / f32!(1000.0); + let ppm_filter_tc_periods = pwr.ppm_filter_time_constant_ms / period_ms; + #[ver(V >= V13_0B4)] + let ppm_filter_tc_ms_rounded = ppm_filter_tc_periods * period_ms; + let ppm_filter_a = f32!(1.0) / ppm_filter_tc_periods.into(); + let perf_filter_a = f32!(1.0) / pwr.perf_filter_time_constant.into(); + let perf_filter_a2 = f32!(1.0) / pwr.perf_filter_time_constant2.into(); + let avg_power_target_filter_a = f32!(1.0) / pwr.avg_power_target_filter_tc.into(); + let avg_power_filter_tc_periods = pwr.avg_power_filter_tc_ms / period_ms; + #[ver(V >= V13_0B4)] + let avg_power_filter_tc_ms_rounded = avg_power_filter_tc_periods * period_ms; + let avg_power_filter_a = f32!(1.0) / avg_power_filter_tc_periods.into(); + let pwr_filter_a = f32!(1.0) / pwr.pwr_filter_time_constant.into(); + + let base_ps = pwr.perf_base_pstate; + let base_ps_scaled = 100 * base_ps; + let max_ps = pwr.perf_max_pstate; + let max_ps_scaled = 100 * max_ps; + let boost_ps_count = max_ps - base_ps; + + let base_clock_khz = self.cfg.base_clock_hz / 1000; + let clocks_per_period = base_clock_khz * period_ms; + + let raw = place!( + ptr, + raw::HwDataA::ver { + clocks_per_period: clocks_per_period, + #[ver(V >= V13_0B4)] + clocks_per_period_2: clocks_per_period, + pwr_status: AtomicU32::new(4), + unk_10: f32!(1.0), + actual_pstate: 1, + tgt_pstate: 1, + base_pstate_scaled: base_ps_scaled, + unk_40: 1, + max_pstate_scaled: max_ps_scaled, + min_pstate_scaled: 100, + unk_64c: 625, + pwr_filter_a_neg: f32!(1.0) - pwr_filter_a, + pwr_filter_a: pwr_filter_a, + pwr_integral_gain: pwr.pwr_integral_gain, + pwr_integral_min_clamp: pwr.pwr_integral_min_clamp.into(), + max_power_1: pwr.max_power_mw.into(), + pwr_proportional_gain: pwr.pwr_proportional_gain, + pwr_pstate_related_k: -F32::from(max_ps_scaled) / pwr.max_power_mw.into(), + pwr_pstate_max_dc_offset: pwr.pwr_min_duty_cycle as i32 + - max_ps_scaled as i32, + max_pstate_scaled_2: max_ps_scaled, + max_power_2: pwr.max_power_mw, + max_pstate_scaled_3: max_ps_scaled, + ppm_filter_tc_periods_x4: ppm_filter_tc_periods * 4, + ppm_filter_a_neg: f32!(1.0) - ppm_filter_a, + ppm_filter_a: ppm_filter_a, + ppm_ki_dt: pwr.ppm_ki * period_s, + unk_6fc: f32!(65536.0), + ppm_kp: pwr.ppm_kp, + pwr_min_duty_cycle: pwr.pwr_min_duty_cycle, + max_pstate_scaled_4: max_ps_scaled, + unk_71c: f32!(0.0), + max_power_3: pwr.max_power_mw, + cur_power_mw_2: 0x0, + ppm_filter_tc_ms: pwr.ppm_filter_time_constant_ms, + #[ver(V >= V13_0B4)] + ppm_filter_tc_clks: ppm_filter_tc_ms_rounded * base_clock_khz, + perf_tgt_utilization: pwr.perf_tgt_utilization, + perf_boost_min_util: pwr.perf_boost_min_util, + perf_boost_ce_step: pwr.perf_boost_ce_step, + perf_reset_iters: pwr.perf_reset_iters, + unk_774: 6, + unk_778: 1, + perf_filter_drop_threshold: pwr.perf_filter_drop_threshold, + perf_filter_a_neg: f32!(1.0) - perf_filter_a, + perf_filter_a2_neg: f32!(1.0) - perf_filter_a2, + perf_filter_a: perf_filter_a, + perf_filter_a2: perf_filter_a2, + perf_ki: pwr.perf_integral_gain, + perf_ki2: pwr.perf_integral_gain2, + perf_integral_min_clamp: pwr.perf_integral_min_clamp.into(), + unk_79c: f32!(95.0), + perf_kp: pwr.perf_proportional_gain, + perf_kp2: pwr.perf_proportional_gain2, + boost_state_unk_k: F32::from(boost_ps_count) / f32!(0.95), + base_pstate_scaled_2: base_ps_scaled, + max_pstate_scaled_5: max_ps_scaled, + base_pstate_scaled_3: base_ps_scaled, + perf_tgt_utilization_2: pwr.perf_tgt_utilization, + base_pstate_scaled_4: base_ps_scaled, + unk_7fc: f32!(65536.0), + pwr_min_duty_cycle_2: pwr.pwr_min_duty_cycle.into(), + max_pstate_scaled_6: max_ps_scaled.into(), + max_freq_mhz: pwr.max_freq_mhz, + pwr_min_duty_cycle_3: pwr.pwr_min_duty_cycle, + min_pstate_scaled_4: f32!(100.0), + max_pstate_scaled_7: max_ps_scaled, + unk_alpha_neg: f32!(0.8), + unk_alpha: f32!(0.2), + fast_die0_sensor_mask: U64(self.cfg.fast_die0_sensor_mask), + fast_die0_release_temp_cc: 100 * pwr.fast_die0_release_temp, + unk_87c: self.cfg.da.unk_87c, + unk_880: 0x4, + unk_894: f32!(1.0), + + fast_die0_ki_dt: pwr.fast_die0_integral_gain * period_s, + unk_8a8: f32!(65536.0), + fast_die0_kp: pwr.fast_die0_proportional_gain, + pwr_min_duty_cycle_4: pwr.pwr_min_duty_cycle, + max_pstate_scaled_8: max_ps_scaled, + max_pstate_scaled_9: max_ps_scaled, + fast_die0_prop_tgt_delta: 100 * pwr.fast_die0_prop_tgt_delta, + unk_8cc: self.cfg.da.unk_8cc, + max_pstate_scaled_10: max_ps_scaled, + max_pstate_scaled_11: max_ps_scaled, + unk_c2c: 1, + power_zone_count: pwr.power_zones.len() as u32, + max_power_4: pwr.max_power_mw, + max_power_5: pwr.max_power_mw, + max_power_6: pwr.max_power_mw, + avg_power_target_filter_a_neg: f32!(1.0) - avg_power_target_filter_a, + avg_power_target_filter_a: avg_power_target_filter_a, + avg_power_target_filter_tc_x4: 4 * pwr.avg_power_target_filter_tc, + avg_power_target_filter_tc_xperiod: period_ms + * pwr.avg_power_target_filter_tc, + #[ver(V >= V13_0B4)] + avg_power_target_filter_tc_clks: period_ms + * pwr.avg_power_target_filter_tc + * base_clock_khz, + avg_power_filter_tc_periods_x4: 4 * avg_power_filter_tc_periods, + avg_power_filter_a_neg: f32!(1.0) - avg_power_filter_a, + avg_power_filter_a: avg_power_filter_a, + avg_power_ki_dt: pwr.avg_power_ki_only * period_s, + unk_d20: f32!(65536.0), + avg_power_kp: pwr.avg_power_kp, + avg_power_min_duty_cycle: pwr.avg_power_min_duty_cycle, + max_pstate_scaled_12: max_ps_scaled, + max_pstate_scaled_13: max_ps_scaled, + max_power_7: pwr.max_power_mw.into(), + max_power_8: pwr.max_power_mw, + avg_power_filter_tc_ms: pwr.avg_power_filter_tc_ms, + #[ver(V >= V13_0B4)] + avg_power_filter_tc_clks: avg_power_filter_tc_ms_rounded * base_clock_khz, + max_pstate_scaled_14: max_ps_scaled, + t81xx_data: match self.cfg.chip_id { + 0x8103 | 0x8112 => Self::t81xx_data(self.dyncfg), + _ => Default::default(), + }, + #[ver(V >= V13_0B4)] + unk_e10_0: raw::HwDataA130Extra { + unk_38: 4, + unk_3c: 8000, + unk_40: 2500, + unk_48: 0xffffffff, + unk_4c: 50, + unk_54: 50, + unk_58: 0x1, + unk_60: f32!(0.8888889), + unk_64: f32!(0.6666667), + unk_68: f32!(0.11111111), + unk_6c: f32!(0.33333333), + unk_70: f32!(-0.4), + unk_74: f32!(-0.8), + unk_7c: f32!(65536.0), + unk_80: f32!(-5.0), + unk_84: f32!(-10.0), + unk_8c: 40, + max_pstate_scaled_1: max_ps_scaled, + unk_9c: f32!(8000.0), + unk_a0: 1400, + unk_a8: 72, + unk_ac: 24, + unk_b0: 1728000, + unk_b8: 576000, + unk_c4: f32!(65536.0), + unk_114: f32!(65536.0), + unk_124: 40, + max_pstate_scaled_2: max_ps_scaled, + ..Default::default() + }, + fast_die0_sensor_mask_2: U64(self.cfg.fast_die0_sensor_mask), + unk_e24: self.cfg.da.unk_e24, + unk_e28: 1, + fast_die0_sensor_mask_alt: U64(self.cfg.fast_die0_sensor_mask_alt), + #[ver(V < V13_0B4)] + fast_die0_sensor_present: U64(self.cfg.fast_die0_sensor_present as u64), + unk_163c: 1, + unk_3644: 0, + hws1: Self::hw_shared1(self.cfg), + hws2: *Self::hw_shared2(self.cfg)?, + hws3: *Self::hw_shared3(self.cfg)?, + unk_3ce8: 1, + ..Default::default() + } + ); + + for i in 0..self.dyncfg.pwr.perf_states.len() { + raw.sram_k[i] = self.cfg.sram_k; + } + + for (i, coef) in pwr.core_leak_coef.iter().enumerate() { + raw.core_leak_coef[i] = *coef; + } + + for (i, coef) in pwr.sram_leak_coef.iter().enumerate() { + raw.sram_leak_coef[i] = *coef; + } + + for i in 0..self.dyncfg.id.num_clusters as usize { + if let Some(coef_a) = self.cfg.unk_coef_a.get(i) { + (*raw.unk_coef_a1[i])[..coef_a.len()].copy_from_slice(coef_a); + (*raw.unk_coef_a2[i])[..coef_a.len()].copy_from_slice(coef_a); + } + if let Some(coef_b) = self.cfg.unk_coef_b.get(i) { + (*raw.unk_coef_b1[i])[..coef_b.len()].copy_from_slice(coef_b); + (*raw.unk_coef_b2[i])[..coef_b.len()].copy_from_slice(coef_b); + } + } + + for (i, pz) in pwr.power_zones.iter().enumerate() { + raw.power_zones[i].target = pz.target; + raw.power_zones[i].target_off = pz.target - pz.target_offset; + raw.power_zones[i].filter_tc_x4 = 4 * pz.filter_tc; + raw.power_zones[i].filter_tc_xperiod = period_ms * pz.filter_tc; + let filter_a = f32!(1.0) / pz.filter_tc.into(); + raw.power_zones[i].filter_a = filter_a; + raw.power_zones[i].filter_a_neg = f32!(1.0) - filter_a; + #[ver(V >= V13_0B4)] + raw.power_zones[i].unk_10 = 1320000000; + } + + Ok(raw) + }) + } + + /// Create the HwDataB structure. This mostly contains GPU-related configuration. + #[inline(never)] + fn hwdata_b(&mut self) -> Result> { + self.alloc + .private + .new_inplace(Default::default(), |_inner, ptr| { + let raw = place!( + ptr, + raw::HwDataB::ver { + // Userspace VA map related + #[ver(V < V13_0B4)] + unk_0: U64(0x13_00000000), + unk_8: U64(0x14_00000000), + #[ver(V < V13_0B4)] + unk_10: U64(0x1_00000000), + unk_18: U64(0xffc00000), + unk_20: U64(0x11_00000000), + unk_28: U64(0x11_00000000), + // userspace address? + unk_30: U64(0x6f_ffff8000), + // unmapped? + unkptr_38: U64(0xffffffa0_11800000), + // TODO: yuv matrices + chip_id: self.cfg.chip_id, + unk_454: 0x1, + unk_458: 0x1, + unk_460: 0x1, + unk_464: 0x1, + unk_468: 0x1, + unk_47c: 0x1, + unk_484: 0x1, + unk_48c: 0x1, + base_clock_khz: self.cfg.base_clock_hz / 1000, + power_sample_period: self.dyncfg.pwr.power_sample_period, + unk_49c: 0x1, + unk_4a0: 0x1, + unk_4a4: 0x1, + unk_4c0: 0x1f, + unk_4e0: U64(self.cfg.db.unk_4e0), + unk_4f0: 0x1, + unk_4f4: 0x1, + unk_504: 0x31, + unk_524: 0x1, // use_secure_cache_flush + unk_534: self.cfg.db.unk_534, + num_frags: self.dyncfg.id.num_frags * self.dyncfg.id.num_clusters, + unk_554: 0x1, + uat_ttb_base: U64(self.dyncfg.uat_ttb_base), + gpu_core_id: self.cfg.gpu_core as u32, + gpu_rev_id: self.dyncfg.id.gpu_rev_id as u32, + num_cores: self.dyncfg.id.num_cores * self.dyncfg.id.num_clusters, + max_pstate: self.dyncfg.pwr.perf_states.len() as u32 - 1, + #[ver(V < V13_0B4)] + num_pstates: self.dyncfg.pwr.perf_states.len() as u32, + #[ver(V < V13_0B4)] + min_sram_volt: self.dyncfg.pwr.min_sram_microvolt / 1000, + #[ver(V < V13_0B4)] + unk_ab8: self.cfg.db.unk_ab8, + #[ver(V < V13_0B4)] + unk_abc: self.cfg.db.unk_abc, + #[ver(V < V13_0B4)] + unk_ac0: 0x1020, + + #[ver(V >= V13_0B4)] + unk_ae4: Array::new([0x0, 0x3, 0x7, 0x7]), + #[ver(V < V13_0B4)] + unk_ae4: Array::new([0x0, 0xf, 0x3f, 0x3f]), + unk_b10: 0x1, + unk_b24: 0x1, + unk_b28: 0x1, + unk_b2c: 0x1, + unk_b30: self.cfg.db.unk_b30, + #[ver(V >= V13_0B4)] + unk_b38_0: 1, + #[ver(V >= V13_0B4)] + unk_b38_4: 1, + unk_b38: Array::new([0xffffffff; 12]), + #[ver(V >= V13_0B4)] + unk_c3c: 0x19, + ..Default::default() + } + ); + + let base_ps = self.dyncfg.pwr.perf_base_pstate as usize; + let max_ps = self.dyncfg.pwr.perf_max_pstate as usize; + let base_freq = self.dyncfg.pwr.perf_states[base_ps].freq_hz; + let max_freq = self.dyncfg.pwr.perf_states[max_ps].freq_hz; + + for (i, ps) in self.dyncfg.pwr.perf_states.iter().enumerate() { + raw.frequencies[i] = ps.freq_hz / 1000000; + for (j, mv) in ps.volt_mv.iter().enumerate() { + let sram_mv = (*mv).max(self.dyncfg.pwr.min_sram_microvolt / 1000); + raw.voltages[i][j] = *mv; + raw.voltages_sram[i][j] = sram_mv; + } + raw.sram_k[i] = self.cfg.sram_k; + raw.rel_max_powers[i] = ps.pwr_mw * 100 / self.dyncfg.pwr.max_power_mw; + raw.rel_boost_freqs[i] = if i > base_ps { + (ps.freq_hz - base_freq) / ((max_freq - base_freq) / 100) + } else { + 0 + }; + } + + Ok(raw) + }) + } + + /// Create the Globals structure, which contains global firmware config including more power + /// configuration data and globals used to exchange state between the firmware and driver. + #[inline(never)] + fn globals(&mut self) -> Result> { + self.alloc + .shared + .new_inplace(Default::default(), |_inner, ptr| { + let pwr = &self.dyncfg.pwr; + let period_ms = pwr.power_sample_period; + let period_s = F32::from(period_ms) / f32!(1000.0); + let avg_power_filter_tc_periods = pwr.avg_power_filter_tc_ms / period_ms; + + let max_ps = pwr.perf_max_pstate; + let max_ps_scaled = 100 * max_ps; + + let raw = place!( + ptr, + raw::Globals::ver { + //ktrace_enable: 0xffffffff, + ktrace_enable: 0, + #[ver(V >= V13_2)] + unk_24_0: 3000, + unk_24: 0, + #[ver(V >= V13_0B4)] + unk_28_0: 0, // debug + unk_28: 1, + #[ver(V >= V13_0B4)] + unk_2c_0: 0, + unk_2c: 1, + unk_30: 0, + unk_34: 120, + sub: raw::GlobalsSub::ver { + unk_54: 0xffff, + unk_56: 40, + unk_58: 0xffff, + unk_5e: U32(1), + unk_66: U32(1), + ..Default::default() + }, + unk_8900: 1, + pending_submissions: AtomicU32::new(0), + max_power: pwr.max_power_mw, + max_pstate_scaled: max_ps_scaled, + max_pstate_scaled_2: max_ps_scaled, + max_pstate_scaled_3: max_ps_scaled, + power_zone_count: pwr.power_zones.len() as u32, + avg_power_filter_tc_periods: avg_power_filter_tc_periods, + avg_power_ki_dt: pwr.avg_power_ki_only * period_s, + avg_power_kp: pwr.avg_power_kp, + avg_power_min_duty_cycle: pwr.avg_power_min_duty_cycle, + avg_power_target_filter_tc: pwr.avg_power_target_filter_tc, + unk_89bc: self.cfg.da.unk_8cc, + fast_die0_release_temp: 100 * pwr.fast_die0_release_temp, + unk_89c4: self.cfg.da.unk_87c, + fast_die0_prop_tgt_delta: 100 * pwr.fast_die0_prop_tgt_delta, + fast_die0_kp: pwr.fast_die0_proportional_gain, + fast_die0_ki_dt: pwr.fast_die0_integral_gain * period_s, + unk_89e0: 1, + max_power_2: pwr.max_power_mw, + ppm_kp: pwr.ppm_kp, + ppm_ki_dt: pwr.ppm_ki * period_s, + #[ver(V >= V13_0B4)] + unk_89f4_8: 1, + unk_89f4: 0, + hws1: Self::hw_shared1(self.cfg), + hws2: *Self::hw_shared2(self.cfg)?, + hws3: *Self::hw_shared3(self.cfg)?, + unk_900c: 1, + #[ver(V >= V13_0B4)] + unk_9010_0: 1, + #[ver(V >= V13_0B4)] + unk_903c: 1, + #[ver(V < V13_0B4)] + unk_903c: 0, + fault_control: *crate::fault_control.read(), + do_init: 1, + unk_11020: 40, + unk_11024: 10, + unk_11028: 250, + #[ver(V >= V13_0B4)] + unk_1102c_0: 1, + #[ver(V >= V13_0B4)] + unk_1102c_4: 1, + #[ver(V >= V13_0B4)] + unk_1102c_8: 100, + #[ver(V >= V13_0B4)] + unk_1102c_c: 1, + idle_off_delay_ms: AtomicU32::new(pwr.idle_off_delay_ms), + fender_idle_off_delay_ms: pwr.fender_idle_off_delay_ms, + fw_early_wake_timeout_ms: pwr.fw_early_wake_timeout_ms, + unk_118e0: 40, + #[ver(V >= V13_0B4)] + unk_118e4_0: 50, + #[ver(V >= V13_0B4)] + unk_11edc: 0, + #[ver(V >= V13_0B4)] + unk_11efc: 0, + ..Default::default() + } + ); + + for (i, pz) in pwr.power_zones.iter().enumerate() { + raw.power_zones[i].target = pz.target; + raw.power_zones[i].target_off = pz.target - pz.target_offset; + raw.power_zones[i].filter_tc = pz.filter_tc; + } + + if let Some(tab) = self.cfg.global_tab.as_ref() { + for (i, x) in tab.iter().enumerate() { + raw.unk_118ec[i] = *x; + } + raw.unk_118e8 = 1; + } + + Ok(raw) + }) + } + + /// Create the RuntimePointers structure, which contains pointers to most of the other + /// structures including the ring buffer channels, statistics structures, and HwDataA/HwDataB. + #[inline(never)] + fn runtime_pointers(&mut self) -> Result> { + let hwa = self.hwdata_a()?; + let hwb = self.hwdata_b()?; + + let pointers: Box = box_in_place!(RuntimePointers::ver { + stats: Stats::ver { + vtx: self.alloc.private.new_default::()?, + frag: self.alloc.private.new_inplace( + Default::default(), + |_inner, ptr: &mut MaybeUninit| { + Ok(place!( + ptr, + raw::GpuGlobalStatsFrag::ver { + stats: raw::GpuStatsFrag::ver { + cur_stamp_id: -1, + unk_118: -1, + ..Default::default() + }, + ..Default::default() + } + )) + }, + )?, + comp: self.alloc.private.new_default::()?, + }, + + hwdata_a: hwa, + unkptr_190: self.alloc.private.array_empty(0x80)?, + unkptr_198: self.alloc.private.array_empty(0xc0)?, + hwdata_b: hwb, + + unkptr_1b8: self.alloc.private.array_empty(0x1000)?, + unkptr_1c0: self.alloc.private.array_empty(0x300)?, + unkptr_1c8: self.alloc.private.array_empty(0x1000)?, + + buffer_mgr_ctl: self.alloc.gpu.array_empty(127)?, + })?; + + self.alloc.private.new_boxed(pointers, |inner, ptr| { + Ok(place!( + ptr, + raw::RuntimePointers::ver { + pipes: Default::default(), + device_control: Default::default(), + event: Default::default(), + fw_log: Default::default(), + ktrace: Default::default(), + stats: Default::default(), + + stats_vtx: inner.stats.vtx.gpu_pointer(), + stats_frag: inner.stats.frag.gpu_pointer(), + stats_comp: inner.stats.comp.gpu_pointer(), + + hwdata_a: inner.hwdata_a.gpu_pointer(), + unkptr_190: inner.unkptr_190.gpu_pointer(), + unkptr_198: inner.unkptr_198.gpu_pointer(), + hwdata_b: inner.hwdata_b.gpu_pointer(), + hwdata_b_2: inner.hwdata_b.gpu_pointer(), + + fwlog_buf: None, + + unkptr_1b8: inner.unkptr_1b8.gpu_pointer(), + unkptr_1c0: inner.unkptr_1c0.gpu_pointer(), + unkptr_1c8: inner.unkptr_1c8.gpu_pointer(), + + buffer_mgr_ctl: inner.buffer_mgr_ctl.gpu_pointer(), + buffer_mgr_ctl_2: inner.buffer_mgr_ctl.gpu_pointer(), + + __pad0: Default::default(), + unk_160: U64(0), + unk_168: U64(0), + unk_1d0: 0, + unk_1d4: 0, + unk_1d8: Default::default(), + + __pad1: Default::default(), + gpu_scratch: raw::RuntimeScratch { + unk_6b38: 0xff, + ..Default::default() + }, + } + )) + }) + } + + /// Create the FwStatus structure, which is used to coordinate the firmware halt state between + /// the firmware and the driver. + #[inline(never)] + fn fw_status(&mut self) -> Result> { + self.alloc + .shared + .new_object(Default::default(), |_inner| Default::default()) + } + + /// Create one UatLevelInfo structure, which describes one level of translation for the UAT MMU. + #[inline(never)] + fn uat_level_info( + cfg: &hw::HwConfig, + index_shift: usize, + num_entries: usize, + ) -> raw::UatLevelInfo { + raw::UatLevelInfo { + index_shift: index_shift as _, + unk_1: 14, + unk_2: 14, + unk_3: 8, + unk_4: 0x4000, + num_entries: num_entries as _, + unk_8: U64(1), + unk_10: U64(((1u64 << cfg.uat_oas) - 1) & !(mmu::UAT_PGMSK as u64)), + index_mask: U64(((num_entries - 1) << index_shift) as u64), + } + } + + /// Build the top-level InitData object. + #[inline(never)] + pub(crate) fn build(&mut self) -> Result>> { + let inner: Box = box_in_place!(InitData::ver { + unk_buf: self.alloc.shared_ro.array_empty(0x4000)?, + runtime_pointers: self.runtime_pointers()?, + globals: self.globals()?, + fw_status: self.fw_status()?, + })?; + + Ok(Box::try_new(self.alloc.shared_ro.new_boxed( + inner, + |inner, ptr| { + Ok(place!( + ptr, + raw::InitData::ver { + #[ver(V >= V13_0B4)] + ver_info: Array::new([1, 1, 16, 1]), + unk_buf: inner.unk_buf.gpu_pointer(), + unk_8: 0, + unk_c: 0, + runtime_pointers: inner.runtime_pointers.gpu_pointer(), + globals: inner.globals.gpu_pointer(), + fw_status: inner.fw_status.gpu_pointer(), + uat_page_size: 0x4000, + uat_page_bits: 14, + uat_num_levels: 3, + uat_level_info: Array::new([ + Self::uat_level_info(self.cfg, 36, 8), + Self::uat_level_info(self.cfg, 25, 2048), + Self::uat_level_info(self.cfg, 14, 2048), + ]), + __pad0: Default::default(), + host_mapped_fw_allocations: 1, + unk_ac: 0, + unk_b0: 0, + unk_b4: 0, + unk_b8: 0, + } + )) + }, + )?)?) + } +} diff --git a/drivers/gpu/drm/asahi/mem.rs b/drivers/gpu/drm/asahi/mem.rs new file mode 100644 index 000000000000..491d4f8a4016 --- /dev/null +++ b/drivers/gpu/drm/asahi/mem.rs @@ -0,0 +1,133 @@ +// SPDX-License-Identifier: GPL-2.0-only OR MIT + +//! ARM64 low level memory operations. +//! +//! This GPU uses CPU-side `tlbi` outer-shareable instructions to manage its TLBs. +//! Yes, really. Even though the VA address spaces are unrelated. +//! +//! Right now we pick our own ASIDs and don't coordinate with the CPU. This might result +//! in needless TLB shootdowns on the CPU side... TODO: fix this. + +use core::arch::asm; +use core::cmp::min; + +use crate::debug::*; +use crate::mmu; + +type Asid = u8; + +/// Invalidate the entire GPU TLB. +#[inline(always)] +pub(crate) fn tlbi_all() { + unsafe { + asm!(".arch armv8.4-a", "tlbi vmalle1os",); + } +} + +/// Invalidate all TLB entries for a given ASID. +#[inline(always)] +pub(crate) fn tlbi_asid(asid: Asid) { + if debug_enabled(DebugFlags::ConservativeTlbi) { + tlbi_all(); + sync(); + return; + } + + unsafe { + asm!( + ".arch armv8.4-a", + "tlbi aside1os, {x}", + x = in(reg) ((asid as u64) << 48) + ); + } +} + +/// Invalidate a single page for a given ASID. +#[inline(always)] +pub(crate) fn tlbi_page(asid: Asid, va: usize) { + if debug_enabled(DebugFlags::ConservativeTlbi) { + tlbi_all(); + sync(); + return; + } + + let val: u64 = ((asid as u64) << 48) | ((va as u64 >> 12) & 0xffffffffffc); + unsafe { + asm!( + ".arch armv8.4-a", + "tlbi vae1os, {x}", + x = in(reg) val + ); + } +} + +/// Invalidate a range of pages for a given ASID. +#[inline(always)] +pub(crate) fn tlbi_range(asid: Asid, va: usize, len: usize) { + if debug_enabled(DebugFlags::ConservativeTlbi) { + tlbi_all(); + sync(); + return; + } + + if len == 0 { + return; + } + + let start_pg = va >> mmu::UAT_PGBIT; + let end_pg = (va + len + mmu::UAT_PGMSK) >> mmu::UAT_PGBIT; + + let mut val: u64 = ((asid as u64) << 48) | (2 << 46) | (start_pg as u64 & 0x1fffffffff); + let pages = end_pg - start_pg; + + if pages == 1 { + tlbi_page(asid, va); + return; + } + + // Page count is always in units of 2 + let num = ((pages + 1) >> 1) as u64; + // base: 5 bits + // exp: 2 bits + // pages = (base + 1) << (5 * exp + 1) + // 0:00000 -> 2 pages = 2 << 0 + // 0:11111 -> 32 * 2 pages = 2 << 5 + // 1:00000 -> 1 * 32 * 2 pages = 2 << 5 + // 1:11111 -> 32 * 32 * 2 pages = 2 << 10 + // 2:00000 -> 1 * 32 * 32 * 2 pages = 2 << 10 + // 2:11111 -> 32 * 32 * 32 * 2 pages = 2 << 15 + // 3:00000 -> 1 * 32 * 32 * 32 * 2 pages = 2 << 15 + // 3:11111 -> 32 * 32 * 32 * 32 * 2 pages = 2 << 20 + let exp = min(3, (64 - num.leading_zeros()) / 5); + let bits = 5 * exp; + let mut base = (num + (1 << bits) - 1) >> bits; + + val |= (exp as u64) << 44; + + while base > 32 { + unsafe { + asm!( + ".arch armv8.4-a", + "tlbi rvae1os, {x}", + x = in(reg) val | (31 << 39) + ); + } + base -= 32; + } + + unsafe { + asm!( + ".arch armv8.4-a", + "tlbi rvae1os, {x}", + x = in(reg) val | ((base - 1) << 39) + ); + } +} + +/// Issue a memory barrier (`dsb sy`). +#[inline(always)] +pub(crate) fn sync() { + unsafe { + asm!("dsb sy"); + } +} diff --git a/drivers/gpu/drm/asahi/microseq.rs b/drivers/gpu/drm/asahi/microseq.rs new file mode 100644 index 000000000000..dca94ebc53a1 --- /dev/null +++ b/drivers/gpu/drm/asahi/microseq.rs @@ -0,0 +1,61 @@ +// SPDX-License-Identifier: GPL-2.0-only OR MIT + +//! GPU Micro operation sequence builder +//! +//! As part of a single job submisssion to the GPU, the GPU firmware interprets a sequence of +//! commands that we call a "microsequence". These are responsible for setting up the job execution, +//! timestamping the process, waiting for completion, tearing up any resources, and signaling +//! completion to the driver via the event stamp mechanism. +//! +//! Although the microsequences used by the macOS driver are usually quite uniform and simple, the +//! firmware actually implements enough operations to make this interpreter Turing-complete (!). +//! Most of those aren't implemented yet, since we don't need them, but they could come in handy in +//! the future to do strange things or work around firmware bugs... +//! +//! This module simply implements a collection of microsequence operations that can be appended to +//! and later concatenated into one buffer, ready for firmware execution. + +use crate::fw::microseq; +pub(crate) use crate::fw::microseq::*; +use crate::fw::types::*; +use kernel::prelude::*; + +/// MicroSequence object type, which is just an opaque byte array. +pub(crate) type MicroSequence = GpuArray; + +/// MicroSequence builder. +pub(crate) struct Builder { + ops: Vec, +} + +impl Builder { + /// Create a new Builder object + pub(crate) fn new() -> Builder { + Builder { ops: Vec::new() } + } + + /// Get the relative offset from the current pointer to a given target offset. + /// + /// Used for relative jumps. + pub(crate) fn offset_to(&self, target: i32) -> i32 { + target - self.ops.len() as i32 + } + + /// Add an operation to the end of the sequence. + pub(crate) fn add(&mut self, op: T) -> Result { + let off = self.ops.len(); + let p: *const T = &op; + let p: *const u8 = p as *const u8; + let s: &[u8] = unsafe { core::slice::from_raw_parts(p, core::mem::size_of::()) }; + self.ops.try_extend_from_slice(s)?; + Ok(off as i32) + } + + /// Collect all submitted operations into a finalized GPU object. + pub(crate) fn build(self, alloc: &mut Allocator) -> Result { + let mut array = alloc.array_empty::(self.ops.len())?; + + array.as_mut_slice().clone_from_slice(self.ops.as_slice()); + Ok(array) + } +} diff --git a/drivers/gpu/drm/asahi/mmu.rs b/drivers/gpu/drm/asahi/mmu.rs new file mode 100644 index 000000000000..226ca0b7c1d7 --- /dev/null +++ b/drivers/gpu/drm/asahi/mmu.rs @@ -0,0 +1,1249 @@ +// SPDX-License-Identifier: GPL-2.0-only OR MIT + +//! GPU UAT (MMU) management +//! +//! AGX GPUs use an MMU called the UAT, which is largely compatible with the ARM64 page table +//! format. This module manages the global MMU structures, including a shared handoff structure +//! that is used to coordinate VM management operations with the firmware, the TTBAT which points +//! to currently active GPU VM contexts, as well as the individual `Vm` operations to map and +//! unmap buffer objects into a single user or kernel address space. +//! +//! The actual page table management is delegated to the common kernel `io_pgtable` code. + +use core::fmt::Debug; +use core::mem::size_of; +use core::ptr::{addr_of_mut, NonNull}; +use core::sync::atomic::{fence, AtomicU32, AtomicU64, AtomicU8, Ordering}; +use core::time::Duration; + +use kernel::{ + bindings, c_str, delay, device, + drm::mm, + error::{to_result, Result}, + io_pgtable, + io_pgtable::{prot, AppleUAT, IoPageTable}, + prelude::*, + sync::{smutex::Mutex, Guard}, + sync::{Arc, LockClassKey, UniqueArc}, + time, + types::ForeignOwnable, +}; + +use crate::debug::*; +use crate::no_debug; +use crate::{driver, fw, gem, hw, mem, slotalloc}; + +const DEBUG_CLASS: DebugFlags = DebugFlags::Mmu; + +/// PPL magic number for the handoff region +const PPL_MAGIC: u64 = 0x4b1d000000000002; + +/// Number of supported context entries in the TTBAT +const UAT_NUM_CTX: usize = 64; +/// First context available for users +const UAT_USER_CTX_START: usize = 1; +/// Number of available user contexts +const UAT_USER_CTX: usize = UAT_NUM_CTX - UAT_USER_CTX_START; + +/// Number of bits in a page offset. +pub(crate) const UAT_PGBIT: usize = 14; +/// UAT page size. +pub(crate) const UAT_PGSZ: usize = 1 << UAT_PGBIT; +/// UAT page offset mask. +pub(crate) const UAT_PGMSK: usize = UAT_PGSZ - 1; + +type Pte = AtomicU64; + +/// Number of PTEs per page. +const UAT_NPTE: usize = UAT_PGSZ / size_of::(); + +/// UAT input address space (user) +pub(crate) const UAT_IAS: usize = 39; +/// "Fake" kernel UAT input address space (one page level lower) +pub(crate) const UAT_IAS_KERN: usize = 36; + +/// Lower/user base VA +const IOVA_USER_BASE: usize = UAT_PGSZ; +/// Lower/user top VA +const IOVA_USER_TOP: usize = (1 << UAT_IAS) - 1; +/// Upper/kernel base VA +// const IOVA_TTBR1_BASE: usize = 0xffffff8000000000; +/// Driver-managed kernel base VA +const IOVA_KERN_BASE: usize = 0xffffffa000000000; +/// Driver-managed kernel top VA +const IOVA_KERN_TOP: usize = 0xffffffafffffffff; + +const TTBR_VALID: u64 = 0x1; // BIT(0) +const TTBR_ASID_SHIFT: usize = 48; + +const PTE_TABLE: u64 = 0x3; // BIT(0) | BIT(1) + +// Mapping protection types + +// Note: prot::CACHE means "cache coherency", which for UAT means *uncached*, +// since uncached mappings from the GFX ASC side are cache coherent with the AP cache. +// Not having that flag means *cached noncoherent*. + +/// Firmware MMIO R/W +pub(crate) const PROT_FW_MMIO_RW: u32 = + prot::PRIV | prot::READ | prot::WRITE | prot::CACHE | prot::MMIO; +/// Firmware MMIO R/O +pub(crate) const PROT_FW_MMIO_RO: u32 = prot::PRIV | prot::READ | prot::CACHE | prot::MMIO; +/// Firmware shared (uncached) RW +pub(crate) const PROT_FW_SHARED_RW: u32 = prot::PRIV | prot::READ | prot::WRITE | prot::CACHE; +/// Firmware shared (uncached) RO +pub(crate) const PROT_FW_SHARED_RO: u32 = prot::PRIV | prot::READ | prot::CACHE; +/// Firmware private (cached) RW +pub(crate) const PROT_FW_PRIV_RW: u32 = prot::PRIV | prot::READ | prot::WRITE; +/* +/// Firmware private (cached) RO +pub(crate) const PROT_FW_PRIV_RO: u32 = prot::PRIV | prot::READ; +*/ +/// Firmware/GPU shared (uncached) RW +pub(crate) const PROT_GPU_FW_SHARED_RW: u32 = prot::READ | prot::WRITE | prot::CACHE; +/// Firmware/GPU shared (private) RW +pub(crate) const PROT_GPU_FW_PRIV_RW: u32 = prot::READ | prot::WRITE; +/// GPU shared/coherent RW +pub(crate) const PROT_GPU_SHARED_RW: u32 = prot::READ | prot::WRITE | prot::CACHE | prot::NOEXEC; +/// GPU shared/coherent RO +pub(crate) const PROT_GPU_SHARED_RO: u32 = prot::READ | prot::CACHE | prot::NOEXEC; +/// GPU shared/coherent WO +pub(crate) const PROT_GPU_SHARED_WO: u32 = prot::WRITE | prot::CACHE | prot::NOEXEC; +/* +/// GPU private/noncoherent RW +pub(crate) const PROT_GPU_PRIV_RW: u32 = prot::READ | prot::WRITE | prot::NOEXEC; +/// GPU private/noncoherent RO +pub(crate) const PROT_GPU_PRIV_RO: u32 = prot::READ | prot::NOEXEC; +*/ + +type PhysAddr = bindings::phys_addr_t; + +/// A pre-allocated memory region for UAT management +struct UatRegion { + base: PhysAddr, + map: NonNull, +} + +/// It's safe to share UAT region records across threads. +unsafe impl Send for UatRegion {} +unsafe impl Sync for UatRegion {} + +/// Handoff region flush info structure +#[repr(C)] +struct FlushInfo { + state: AtomicU64, + addr: AtomicU64, + size: AtomicU64, +} + +/// UAT Handoff region layout +#[repr(C)] +struct Handoff { + magic_ap: AtomicU64, + magic_fw: AtomicU64, + + lock_ap: AtomicU8, + lock_fw: AtomicU8, + // Implicit padding: 2 bytes + turn: AtomicU32, + cur_slot: AtomicU32, + // Implicit padding: 4 bytes + flush: [FlushInfo; UAT_NUM_CTX + 1], + + unk2: AtomicU8, + // Implicit padding: 7 bytes + unk3: AtomicU64, +} + +const HANDOFF_SIZE: usize = size_of::(); + +/// One VM slot in the TTBAT +#[repr(C)] +struct SlotTTBS { + ttb0: AtomicU64, + ttb1: AtomicU64, +} + +const SLOTS_SIZE: usize = UAT_NUM_CTX * size_of::(); + +// We need at least page 0 (ttb0) +const PAGETABLES_SIZE: usize = UAT_PGSZ; + +/// Inner data for a Vm instance. This is reference-counted by the outer Vm object. +struct VmInner { + dev: driver::AsahiDevice, + is_kernel: bool, + min_va: usize, + max_va: usize, + page_table: AppleUAT, + mm: mm::Allocator<(), MappingInner>, + uat_inner: Arc, + active_users: usize, + binding: Option>, + bind_token: Option, + id: u64, +} + +impl VmInner { + /// Returns the slot index, if this VM is bound. + fn slot(&self) -> Option { + if self.is_kernel { + // The GFX ASC does not care about the ASID. Pick an arbitrary one. + // TODO: This needs to be a persistently reserved ASID once we integrate + // with the ARM64 kernel ASID machinery to avoid overlap. + Some(0) + } else { + // We don't check whether we lost the slot, which could cause unnecessary + // invalidations against another Vm. However, this situation should be very + // rare (e.g. a Vm lost its slot, which means 63 other Vms bound in the + // interim, and then it gets killed / drops its mappings without doing any + // final rendering). Anything doing active maps/unmaps is probably also + // rendering and therefore likely bound. + self.bind_token + .as_ref() + .map(|token| (token.last_slot() + UAT_USER_CTX_START as u32)) + } + } + + /// Returns the translation table base for this Vm + fn ttb(&self) -> u64 { + self.page_table.cfg().ttbr + } + + /// Map an IOVA to the shifted address the underlying io_pgtable uses. + fn map_iova(&self, iova: usize, size: usize) -> Result { + if iova < self.min_va || (iova + size - 1) > self.max_va { + Err(EINVAL) + } else if self.is_kernel { + Ok(iova - self.min_va) + } else { + Ok(iova) + } + } + + /// Map a contiguous range of virtual->physical pages. + fn map_pages( + &mut self, + mut iova: usize, + mut paddr: usize, + pgsize: usize, + pgcount: usize, + prot: u32, + ) -> Result { + let mut left = pgcount; + while left > 0 { + let mapped_iova = self.map_iova(iova, pgsize * left)?; + let mapped = self + .page_table + .map_pages(mapped_iova, paddr, pgsize, left, prot)?; + assert!(mapped <= left * pgsize); + + left -= mapped / pgsize; + paddr += mapped; + iova += mapped; + } + Ok(pgcount * pgsize) + } + + /// Unmap a contiguous range of pages. + fn unmap_pages(&mut self, mut iova: usize, pgsize: usize, pgcount: usize) -> Result { + let mut left = pgcount; + while left > 0 { + let mapped_iova = self.map_iova(iova, pgsize * left)?; + let unmapped = self.page_table.unmap_pages(mapped_iova, pgsize, left); + assert!(unmapped <= left * pgsize); + + left -= unmapped / pgsize; + iova += unmapped; + } + + Ok(pgcount * pgsize) + } + + /// Map an `mm::Node` representing an mapping in VA space. + fn map_node(&mut self, node: &mm::Node<(), MappingInner>, prot: u32) -> Result { + let mut iova = node.start() as usize; + let sgt = node.sgt.as_ref().ok_or(EINVAL)?; + + for range in sgt.iter() { + let addr = range.dma_address(); + let len = range.dma_len(); + + if (addr | len | iova) & UAT_PGMSK != 0 { + dev_err!( + self.dev, + "MMU: Mapping {:#x}:{:#x} -> {:#x} is not page-aligned\n", + addr, + len, + iova + ); + return Err(EINVAL); + } + + mod_dev_dbg!( + self.dev, + "MMU: map: {:#x}:{:#x} -> {:#x}\n", + addr, + len, + iova + ); + + self.map_pages(iova, addr, UAT_PGSZ, len >> UAT_PGBIT, prot)?; + + iova += len; + } + Ok(()) + } +} + +/// Shared reference to a virtual memory address space ([`Vm`]). +#[derive(Clone)] +pub(crate) struct Vm { + id: u64, + file_id: u64, + inner: Arc>, +} +no_debug!(Vm); + +/// Slot data for a [`Vm`] slot (nothing, we only care about the indices). +pub(crate) struct SlotInner(); + +impl slotalloc::SlotItem for SlotInner { + type Data = (); +} + +/// Represents a single user of a binding of a [`Vm`] to a slot. +/// +/// The number of users is counted, and the slot will be freed when it drops to 0. +#[derive(Debug)] +pub(crate) struct VmBind(Vm, u32); + +impl VmBind { + /// Returns the slot that this `Vm` is bound to. + pub(crate) fn slot(&self) -> u32 { + self.1 + } +} + +impl Drop for VmBind { + fn drop(&mut self) { + let mut inner = self.0.inner.lock(); + + assert_ne!(inner.active_users, 0); + inner.active_users -= 1; + mod_pr_debug!("MMU: slot {} active users {}\n", self.1, inner.active_users); + if inner.active_users == 0 { + inner.binding = None; + } + } +} + +impl Clone for VmBind { + fn clone(&self) -> VmBind { + let mut inner = self.0.inner.lock(); + + inner.active_users += 1; + mod_pr_debug!("MMU: slot {} active users {}\n", self.1, inner.active_users); + VmBind(self.0.clone(), self.1) + } +} + +/// Inner data required for an object mapping into a [`Vm`]. +pub(crate) struct MappingInner { + owner: Arc>, + uat_inner: Arc, + prot: u32, + mapped_size: usize, + sgt: Option, +} + +/// An object mapping into a [`Vm`], which reserves the address range from use by other mappings. +pub(crate) struct Mapping(mm::Node<(), MappingInner>); + +impl Mapping { + /// Returns the IOVA base of this mapping + pub(crate) fn iova(&self) -> usize { + self.0.start() as usize + } + + /// Returns the size of this mapping in bytes + pub(crate) fn size(&self) -> usize { + self.0.mapped_size + } + + /// Remap a cached mapping as uncached, then synchronously flush that range of VAs from the + /// coprocessor cache. This is required to safely unmap cached/private mappings. + fn remap_uncached_and_flush(&mut self) { + let mut owner = self.0.owner.lock(); + mod_dev_dbg!( + owner.dev, + "MMU: remap as uncached {:#x}:{:#x}\n", + self.iova(), + self.size() + ); + + // The IOMMU API does not allow us to remap things in-place... + // just do an unmap and map again for now. + // Do not try to unmap guard page (-1) + if owner + .unmap_pages(self.iova(), UAT_PGSZ, self.size() >> UAT_PGBIT) + .is_err() + { + dev_err!( + owner.dev, + "MMU: unmap for remap {:#x}:{:#x} failed\n", + self.iova(), + self.size() + ); + } + + let prot = self.0.prot | prot::CACHE; + if owner.map_node(&self.0, prot).is_err() { + dev_err!( + owner.dev, + "MMU: remap {:#x}:{:#x} failed\n", + self.iova(), + self.size() + ); + } + + // If we don't have (and have never had) a VM slot, just return + let slot = match owner.slot() { + None => return, + Some(slot) => slot, + }; + + let flush_slot = if owner.is_kernel { + // If this is a kernel mapping, always flush on index 64 + UAT_NUM_CTX as u32 + } else { + // Otherwise, check if this slot is the active one, otherwise return + // Also check that we actually own this slot + let ttb = owner.ttb() | TTBR_VALID | (slot as u64) << TTBR_ASID_SHIFT; + + let uat_inner = self.0.uat_inner.lock(); + uat_inner.handoff().lock(); + let cur_slot = uat_inner.handoff().current_slot(); + let ttb_cur = uat_inner.ttbs()[slot as usize].ttb0.load(Ordering::Relaxed); + uat_inner.handoff().unlock(); + if cur_slot == Some(slot) && ttb_cur == ttb { + slot + } else { + return; + } + }; + + // FIXME: There is a race here, though it'll probably never happen in practice. + // In theory, it's possible for the ASC to finish using our slot, whatever command + // it was processing to complete, the slot to be lost to another context, and the ASC + // to begin using it again with a different page table, thus faulting when it gets a + // flush request here. In practice, the chance of this happening is probably vanishingly + // small, as all 62 other slots would have to be recycled or in use before that slot can + // be reused, and the ASC using user contexts at all is very rare. + + // Still, the locking around UAT/Handoff/TTBs should probably be redesigned to better + // model the interactions with the firmware and avoid these races. + // Possibly TTB changes should be tied to slot locks: + + // Flush: + // - Can early check handoff here (no need to lock). + // If user slot and it doesn't match the active ASC slot, + // we can elide the flush as the ASC guarantees it flushes + // TLBs/caches when it switches context. We just need a + // barrier to ensure ordering. + // - Lock TTB slot + // - If user ctx: + // - Lock handoff AP-side + // - Lock handoff dekker + // - Check TTB & handoff cur ctx + // - Perform flush if necessary + // - This implies taking the fwring lock + // + // TTB change: + // - lock TTB slot + // - lock handoff AP-side + // - lock handoff dekker + // change TTB + + // Lock this flush slot, and write the range to it + let flush = self.0.uat_inner.lock_flush(flush_slot); + let pages = self.size() >> UAT_PGBIT; + flush.begin_flush(self.iova() as u64, self.size() as u64); + if pages >= 0x10000 { + dev_err!(owner.dev, "MMU: Flush too big ({:#x} pages))\n", pages); + } + + let cmd = fw::channels::FwCtlMsg { + addr: fw::types::U64(self.iova() as u64), + unk_8: 0, + slot: flush_slot, + page_count: pages as u16, + unk_12: 2, // ? + }; + + // Tell the firmware to do a cache flush + if let Err(e) = owner.dev.data().gpu.fwctl(cmd) { + dev_err!( + owner.dev, + "MMU: ASC cache flush {:#x}:{:#x} failed (err: {:?})\n", + self.iova(), + self.size(), + e + ); + } + + // Finish the flush + flush.end_flush(); + + // Slot is unlocked here + } +} + +impl Drop for Mapping { + fn drop(&mut self) { + // This is the main unmap function for UAT mappings. + // The sequence of operations here is finicky, due to the interaction + // between cached GFX ASC mappings and the page tables. These mappings + // always have to be flushed from the cache before being unmapped. + + // For uncached mappings, just unmapping and flushing the TLB is sufficient. + + // For cached mappings, this is the required sequence: + // 1. Remap it as uncached + // 2. Flush the TLB range + // 3. If kernel VA mapping OR user VA mapping and handoff.current_slot() == slot: + // a. Take a lock for this slot + // b. Write the flush range to the right context slot in handoff area + // c. Issue a cache invalidation request via FwCtl queue + // d. Poll for completion via queue + // e. Check for completion flag in the handoff area + // f. Drop the lock + // 4. Unmap + // 5. Flush the TLB range again + + // prot::CACHE means "cache coherent" which means *uncached* here. + if self.0.prot & prot::CACHE == 0 { + self.remap_uncached_and_flush(); + } + + let mut owner = self.0.owner.lock(); + mod_dev_dbg!( + owner.dev, + "MMU: unmap {:#x}:{:#x}\n", + self.iova(), + self.size() + ); + + if owner + .unmap_pages(self.iova(), UAT_PGSZ, self.size() >> UAT_PGBIT) + .is_err() + { + dev_err!( + owner.dev, + "MMU: unmap {:#x}:{:#x} failed\n", + self.iova(), + self.size() + ); + } + + if let Some(asid) = owner.slot() { + mem::tlbi_range(asid as u8, self.iova(), self.size()); + mod_dev_dbg!( + owner.dev, + "MMU: flush range: asid={:#x} start={:#x} len={:#x}\n", + asid, + self.iova(), + self.size() + ); + mem::sync(); + } + } +} + +/// Shared UAT global data structures +struct UatShared { + handoff_rgn: UatRegion, + ttbs_rgn: UatRegion, +} + +impl UatShared { + /// Returns the handoff region area + fn handoff(&self) -> &Handoff { + // SAFETY: pointer is non-null per the type invariant + unsafe { (self.handoff_rgn.map.as_ptr() as *mut Handoff).as_ref() }.unwrap() + } + + /// Returns the TTBAT area + fn ttbs(&self) -> &[SlotTTBS; UAT_NUM_CTX] { + // SAFETY: pointer is non-null per the type invariant + unsafe { (self.ttbs_rgn.map.as_ptr() as *mut [SlotTTBS; UAT_NUM_CTX]).as_ref() }.unwrap() + } +} + +// SAFETY: Nothing here is unsafe to send across threads. +unsafe impl Send for UatShared {} + +/// Inner data for the top-level UAT instance. +struct UatInner { + shared: Mutex, + handoff_flush: [Mutex; UAT_NUM_CTX + 1], +} + +impl UatInner { + /// Take the lock on the shared data and return the guard. + fn lock(&self) -> Guard<'_, Mutex> { + self.shared.lock() + } + + /// Take a lock on a handoff flush slot and return the guard. + fn lock_flush(&self, slot: u32) -> Guard<'_, Mutex> { + self.handoff_flush[slot as usize].lock() + } +} + +/// Top-level UAT manager object +pub(crate) struct Uat { + dev: driver::AsahiDevice, + cfg: &'static hw::HwConfig, + pagetables_rgn: UatRegion, + + inner: Arc, + slots: slotalloc::SlotAllocator, + + kernel_vm: Vm, + _kernel_lower_vm: Vm, +} + +impl Drop for UatRegion { + fn drop(&mut self) { + // SAFETY: the pointer is valid by the type invariant + unsafe { bindings::memunmap(self.map.as_ptr()) }; + } +} + +impl Handoff { + /// Lock the handoff region from firmware access + fn lock(&self) { + self.lock_ap.store(1, Ordering::Relaxed); + fence(Ordering::SeqCst); + + while self.lock_fw.load(Ordering::Relaxed) != 0 { + if self.turn.load(Ordering::Relaxed) != 0 { + self.lock_ap.store(0, Ordering::Relaxed); + while self.turn.load(Ordering::Relaxed) != 0 {} + self.lock_ap.store(1, Ordering::Relaxed); + fence(Ordering::SeqCst); + } + } + fence(Ordering::Acquire); + } + + /// Unlock the handoff region, allowing firmware access + fn unlock(&self) { + self.turn.store(1, Ordering::Relaxed); + self.lock_ap.store(0, Ordering::Release); + } + + /// Returns the current Vm slot mapped by the firmware for lower/unprivileged access, if any. + fn current_slot(&self) -> Option { + let slot = self.cur_slot.load(Ordering::Relaxed); + if slot == 0 || slot == u32::MAX { + None + } else { + Some(slot) + } + } + + /// Initialize the handoff region + fn init(&self) -> Result { + self.magic_ap.store(PPL_MAGIC, Ordering::Relaxed); + self.cur_slot.store(0, Ordering::Relaxed); + self.unk3.store(0, Ordering::Relaxed); + fence(Ordering::SeqCst); + + let timeout = time::ktime_get() + Duration::from_millis(1000); + + self.lock(); + while time::ktime_get() < timeout { + if self.magic_fw.load(Ordering::Relaxed) == PPL_MAGIC { + break; + } else { + self.unlock(); + delay::coarse_sleep(Duration::from_millis(10)); + self.lock(); + } + } + + if self.magic_fw.load(Ordering::Relaxed) != PPL_MAGIC { + self.unlock(); + pr_err!("Handoff: Failed to initialize (firmware not running?)\n"); + return Err(EIO); + } + + self.unlock(); + + for i in 0..=UAT_NUM_CTX { + self.flush[i].state.store(0, Ordering::Relaxed); + self.flush[i].addr.store(0, Ordering::Relaxed); + self.flush[i].size.store(0, Ordering::Relaxed); + } + fence(Ordering::SeqCst); + Ok(()) + } +} + +/// Represents a single flush info slot in the handoff region. +/// +/// # Invariants +/// The pointer is valid and there is no aliasing HandoffFlush instance. +struct HandoffFlush(*const FlushInfo); + +// SAFETY: These pointers are safe to send across threads. +unsafe impl Send for HandoffFlush {} + +impl HandoffFlush { + /// Set up a flush operation for the coprocessor + fn begin_flush(&self, start: u64, size: u64) { + let flush = unsafe { self.0.as_ref().unwrap() }; + + let state = flush.state.load(Ordering::Relaxed); + if state != 0 { + pr_err!("Handoff: expected flush state 0, got {}\n", state); + } + flush.addr.store(start, Ordering::Relaxed); + flush.size.store(size, Ordering::Relaxed); + flush.state.store(1, Ordering::Relaxed); + } + + /// Complete a flush operation for the coprocessor + fn end_flush(&self) { + let flush = unsafe { self.0.as_ref().unwrap() }; + let state = flush.state.load(Ordering::Relaxed); + if state != 2 { + pr_err!("Handoff: expected flush state 2, got {}\n", state); + } + flush.state.store(0, Ordering::Relaxed); + } +} + +// We do not implement FlushOps, since we flush manually in this module after +// page table operations. Just provide dummy implementations. +impl io_pgtable::FlushOps for Uat { + type Data = (); + + fn tlb_flush_all(_data: ::Borrowed<'_>) {} + fn tlb_flush_walk( + _data: ::Borrowed<'_>, + _iova: usize, + _size: usize, + _granule: usize, + ) { + } + fn tlb_add_page( + _data: ::Borrowed<'_>, + _iova: usize, + _granule: usize, + ) { + } +} + +static LOCK_KEY: LockClassKey = LockClassKey::new(); + +impl Vm { + /// Create a new virtual memory address space + fn new( + dev: driver::AsahiDevice, + uat_inner: Arc, + cfg: &'static hw::HwConfig, + is_kernel: bool, + id: u64, + file_id: u64, + ) -> Result { + let page_table = AppleUAT::new( + &dev, + io_pgtable::Config { + pgsize_bitmap: UAT_PGSZ, + ias: if is_kernel { UAT_IAS_KERN } else { UAT_IAS }, + oas: cfg.uat_oas, + coherent_walk: true, + quirks: 0, + }, + (), + )?; + let min_va = if is_kernel { + IOVA_KERN_BASE + } else { + IOVA_USER_BASE + }; + let max_va = if is_kernel { + IOVA_KERN_TOP + } else { + IOVA_USER_TOP + }; + + let mm = mm::Allocator::new( + min_va as u64, + (max_va - min_va + 1) as u64, + (), + c_str!("asahi Vm"), + &LOCK_KEY, + )?; + + Ok(Vm { + id, + file_id, + inner: Arc::try_new(Mutex::new(VmInner { + dev, + min_va, + max_va, + is_kernel, + page_table, + mm, + uat_inner, + binding: None, + bind_token: None, + active_users: 0, + id, + }))?, + }) + } + + /// Get the translation table base for this Vm + fn ttb(&self) -> u64 { + self.inner.lock().ttb() + } + + /// Map a GEM object (using its `SGTable`) into this Vm at a free address. + pub(crate) fn map(&self, size: usize, sgt: gem::SGTable) -> Result { + let mut inner = self.inner.lock(); + + let uat_inner = inner.uat_inner.clone(); + let node = inner.mm.insert_node( + MappingInner { + owner: self.inner.clone(), + uat_inner, + prot: PROT_FW_SHARED_RW, + sgt: Some(sgt), + mapped_size: size, + }, + (size + UAT_PGSZ) as u64, // Add guard page + )?; + + inner.map_node(&node, PROT_FW_SHARED_RW)?; + Ok(Mapping(node)) + } + + /// Map a GEM object (using its `SGTable`) into this Vm at a free address in a given range. + #[allow(clippy::too_many_arguments)] + pub(crate) fn map_in_range( + &self, + size: usize, + sgt: gem::SGTable, + alignment: u64, + start: u64, + end: u64, + prot: u32, + guard: bool, + ) -> Result { + let mut inner = self.inner.lock(); + + let uat_inner = inner.uat_inner.clone(); + let node = inner.mm.insert_node_in_range( + MappingInner { + owner: self.inner.clone(), + uat_inner, + prot, + sgt: Some(sgt), + mapped_size: size, + }, + (size + if guard { UAT_PGSZ } else { 0 }) as u64, // Add guard page + alignment, + 0, + start, + end, + mm::InsertMode::Best, + )?; + + inner.map_node(&node, prot)?; + Ok(Mapping(node)) + } + + /// Map a GEM object (using its `SGTable`) into this Vm at a specific address. + #[allow(clippy::too_many_arguments)] + pub(crate) fn map_at( + &self, + addr: u64, + size: usize, + sgt: gem::SGTable, + prot: u32, + guard: bool, + ) -> Result { + let mut inner = self.inner.lock(); + + let uat_inner = inner.uat_inner.clone(); + let node = inner.mm.reserve_node( + MappingInner { + owner: self.inner.clone(), + uat_inner, + prot, + sgt: Some(sgt), + mapped_size: size, + }, + addr, + (size + if guard { UAT_PGSZ } else { 0 }) as u64, // Add guard page + 0, + )?; + + inner.map_node(&node, prot)?; + Ok(Mapping(node)) + } + + /// Add a direct MMIO mapping to this Vm at a free address. + pub(crate) fn map_io(&self, phys: usize, size: usize, rw: bool) -> Result { + let prot = if rw { PROT_FW_MMIO_RW } else { PROT_FW_MMIO_RO }; + let mut inner = self.inner.lock(); + + let uat_inner = inner.uat_inner.clone(); + let node = inner.mm.insert_node( + MappingInner { + owner: self.inner.clone(), + uat_inner, + prot, + sgt: None, + mapped_size: size, + }, + (size + UAT_PGSZ) as u64, // Add guard page + )?; + + let iova = node.start() as usize; + + if (phys | size | iova) & UAT_PGMSK != 0 { + dev_err!( + inner.dev, + "MMU: Mapping {:#x}:{:#x} -> {:#x} is not page-aligned\n", + phys, + size, + iova + ); + return Err(EINVAL); + } + + dev_info!( + inner.dev, + "MMU: IO map: {:#x}:{:#x} -> {:#x}\n", + phys, + size, + iova + ); + + inner.map_pages(iova, phys, UAT_PGSZ, size >> UAT_PGBIT, prot)?; + + Ok(Mapping(node)) + } + + /// Returns the unique ID of this Vm + pub(crate) fn id(&self) -> u64 { + self.id + } + + /// Returns the unique File ID of the owner of this Vm + pub(crate) fn file_id(&self) -> u64 { + self.file_id + } +} + +impl Drop for VmInner { + fn drop(&mut self) { + assert_eq!(self.active_users, 0); + + mod_pr_debug!( + "VmInner::Drop [{}]: bind_token={:?}\n", + self.id, + self.bind_token + ); + + // Make sure this VM is not mapped to a TTB if it was + if let Some(token) = self.bind_token.take() { + let idx = (token.last_slot() as usize) + UAT_USER_CTX_START; + let ttb = self.ttb() | TTBR_VALID | (idx as u64) << TTBR_ASID_SHIFT; + + let uat_inner = self.uat_inner.lock(); + uat_inner.handoff().lock(); + let handoff_cur = uat_inner.handoff().current_slot(); + let ttb_cur = uat_inner.ttbs()[idx].ttb0.load(Ordering::SeqCst); + let inval = ttb_cur == ttb; + if inval { + if handoff_cur == Some(idx as u32) { + pr_err!( + "VmInner::drop owning slot {}, but it is currently in use by the ASC?\n", + idx + ); + } + uat_inner.ttbs()[idx].ttb0.store(0, Ordering::SeqCst); + } + uat_inner.handoff().unlock(); + core::mem::drop(uat_inner); + + // In principle we dropped all the Mappings already, but we might as + // well play it safe and invalidate the whole ASID. + if inval { + mod_pr_debug!( + "VmInner::Drop [{}]: need inval for ASID {:#x}\n", + self.id, + idx + ); + mem::tlbi_asid(idx as u8); + mem::sync(); + } + } + } +} + +impl Uat { + /// Map a bootloader-preallocated memory region + fn map_region( + dev: &dyn device::RawDevice, + name: &CStr, + size: usize, + cached: bool, + ) -> Result { + let rdev = dev.raw_device(); + + let mut res = core::mem::MaybeUninit::::uninit(); + + let res = unsafe { + let idx = bindings::of_property_match_string( + (*rdev).of_node, + c_str!("memory-region-names").as_char_ptr(), + name.as_char_ptr(), + ); + to_result(idx)?; + + let np = bindings::of_parse_phandle( + (*rdev).of_node, + c_str!("memory-region").as_char_ptr(), + idx, + ); + if np.is_null() { + dev_err!(dev, "Missing {} region\n", name); + return Err(EINVAL); + } + let ret = bindings::of_address_to_resource(np, 0, res.as_mut_ptr()); + bindings::of_node_put(np); + + if ret < 0 { + dev_err!(dev, "Failed to get {} region\n", name); + to_result(ret)? + } + + res.assume_init() + }; + + let rgn_size: usize = unsafe { bindings::resource_size(&res) } as usize; + + if size > rgn_size { + dev_err!( + dev, + "Region {} is too small (expected {}, got {})\n", + name, + size, + rgn_size + ); + return Err(ENOMEM); + } + + let flags = if cached { + bindings::MEMREMAP_WB + } else { + bindings::MEMREMAP_WC + }; + let map = unsafe { bindings::memremap(res.start, rgn_size, flags.into()) }; + let map = NonNull::new(map); + + match map { + None => { + dev_err!(dev, "Failed to remap {} region\n", name); + Err(ENOMEM) + } + Some(map) => Ok(UatRegion { + base: res.start, + map, + }), + } + } + + /// Returns a view into the root kernel (upper half) page table + fn kpt0(&self) -> &[Pte; UAT_NPTE] { + // SAFETY: pointer is non-null per the type invariant + unsafe { (self.pagetables_rgn.map.as_ptr() as *mut [Pte; UAT_NPTE]).as_ref() }.unwrap() + } + + /// Returns a reference to the global kernel (upper half) `Vm` + pub(crate) fn kernel_vm(&self) -> &Vm { + &self.kernel_vm + } + + /// Returns the base physical address of the TTBAT region. + pub(crate) fn ttb_base(&self) -> u64 { + let inner = self.inner.lock(); + + inner.ttbs_rgn.base + } + + /// Binds a `Vm` to a slot, preferring the last used one. + pub(crate) fn bind(&self, vm: &Vm) -> Result { + let mut inner = vm.inner.lock(); + + if inner.binding.is_none() { + assert_eq!(inner.active_users, 0); + + let slot = self.slots.get(inner.bind_token)?; + if slot.changed() { + mod_pr_debug!("Vm Bind [{}]: bind_token={:?}\n", vm.id, slot.token(),); + let idx = (slot.slot() as usize) + UAT_USER_CTX_START; + let ttb = inner.ttb() | TTBR_VALID | (idx as u64) << TTBR_ASID_SHIFT; + + let uat_inner = self.inner.lock(); + let ttbs = uat_inner.ttbs(); + uat_inner.handoff().lock(); + if uat_inner.handoff().current_slot() == Some(idx as u32) { + pr_err!( + "Vm::bind to slot {}, but it is currently in use by the ASC?\n", + idx + ); + } + ttbs[idx].ttb0.store(ttb, Ordering::Relaxed); + ttbs[idx].ttb1.store(0, Ordering::Relaxed); + uat_inner.handoff().unlock(); + core::mem::drop(uat_inner); + + // Make sure all TLB entries from the previous owner of this ASID are gone + mem::tlbi_asid(idx as u8); + mem::sync(); + } + + inner.bind_token = Some(slot.token()); + inner.binding = Some(slot); + } + + inner.active_users += 1; + + let slot = inner.binding.as_ref().unwrap().slot() + UAT_USER_CTX_START as u32; + mod_pr_debug!("MMU: slot {} active users {}\n", slot, inner.active_users); + Ok(VmBind(vm.clone(), slot)) + } + + /// Creates a new `Vm` linked to this UAT. + pub(crate) fn new_vm(&self, id: u64, file_id: u64) -> Result { + Vm::new( + self.dev.clone(), + self.inner.clone(), + self.cfg, + false, + id, + file_id, + ) + } + + /// Creates the reference-counted inner data for a new `Uat` instance. + #[inline(never)] + fn make_inner(dev: &driver::AsahiDevice) -> Result> { + let handoff_rgn = Self::map_region(dev, c_str!("handoff"), HANDOFF_SIZE, false)?; + let ttbs_rgn = Self::map_region(dev, c_str!("ttbs"), SLOTS_SIZE, false)?; + + dev_info!(dev, "MMU: Initializing kernel page table\n"); + + let mut inner = UniqueArc::::try_new_uninit()?; + let ptr = inner.as_mut_ptr(); + + Ok(unsafe { + let handoff = &(handoff_rgn.map.as_ptr() as *mut Handoff).as_ref().unwrap(); + + for i in 0..UAT_NUM_CTX + 1 { + addr_of_mut!((*ptr).handoff_flush[i]) + .write(Mutex::new(HandoffFlush(&handoff.flush[i]))); + } + + addr_of_mut!((*ptr).shared).write(Mutex::new(UatShared { + handoff_rgn, + ttbs_rgn, + })); + + inner.assume_init() + } + .into()) + } + + /// Creates a new `Uat` instance given the relevant hardware config. + #[inline(never)] + pub(crate) fn new(dev: &driver::AsahiDevice, cfg: &'static hw::HwConfig) -> Result { + dev_info!(dev, "MMU: Initializing...\n"); + + let inner = Self::make_inner(dev)?; + + let pagetables_rgn = Self::map_region(dev, c_str!("pagetables"), PAGETABLES_SIZE, true)?; + + dev_info!(dev, "MMU: Creating kernel page tables\n"); + let kernel_lower_vm = Vm::new(dev.clone(), inner.clone(), cfg, false, 1, 0)?; + let kernel_vm = Vm::new(dev.clone(), inner.clone(), cfg, true, 0, 0)?; + + dev_info!(dev, "MMU: Kernel page tables created\n"); + + let ttb0 = kernel_lower_vm.ttb(); + let ttb1 = kernel_vm.ttb(); + + let uat = Self { + dev: dev.clone(), + cfg, + pagetables_rgn, + kernel_vm, + _kernel_lower_vm: kernel_lower_vm, + inner, + slots: slotalloc::SlotAllocator::new(UAT_USER_CTX as u32, (), |_inner, _slot| { + SlotInner() + })?, + }; + + let inner = uat.inner.lock(); + + inner.handoff().init()?; + + dev_info!(dev, "MMU: Initializing TTBs\n"); + + inner.handoff().lock(); + + let ttbs = inner.ttbs(); + + ttbs[0].ttb0.store(ttb0 | TTBR_VALID, Ordering::Relaxed); + ttbs[0] + .ttb1 + .store(uat.pagetables_rgn.base | TTBR_VALID, Ordering::Relaxed); + + for ctx in &ttbs[1..] { + ctx.ttb0.store(0, Ordering::Relaxed); + ctx.ttb1.store(0, Ordering::Relaxed); + } + + inner.handoff().unlock(); + + core::mem::drop(inner); + + uat.kpt0()[2].store(ttb1 | PTE_TABLE, Ordering::Relaxed); + + dev_info!(dev, "MMU: initialized\n"); + + Ok(uat) + } +} + +impl Drop for Uat { + fn drop(&mut self) { + // Unmap what we mapped + self.kpt0()[2].store(0, Ordering::Relaxed); + + // Make sure we flush the TLBs + fence(Ordering::SeqCst); + mem::tlbi_all(); + mem::sync(); + } +} diff --git a/drivers/gpu/drm/asahi/object.rs b/drivers/gpu/drm/asahi/object.rs new file mode 100644 index 000000000000..449899b88181 --- /dev/null +++ b/drivers/gpu/drm/asahi/object.rs @@ -0,0 +1,704 @@ +// SPDX-License-Identifier: GPL-2.0-only OR MIT + +//! Asahi GPU object model +//! +//! The AGX GPU includes a coprocessor that uses a large number of shared memory structures to +//! communicate with the driver. These structures contain GPU VA pointers to each other, which are +//! directly dereferenced by the firmware and are expected to always be valid for the usage +//! lifetime of the containing struct (which is an implicit contract, not explicitly managed). +//! Any faults cause an unrecoverable firmware crash, requiring a full system reboot. +//! +//! In order to manage this complexity safely, we implement a GPU object model using Rust's type +//! system to enforce GPU object lifetime relationships. GPU objects represent an allocated piece +//! of memory of a given type, mapped to the GPU (and usually also the CPU). On the CPU side, +//! these objects are associated with a pure Rust structure that contains the objects it depends +//! on (or references to them). This allows us to map Rust lifetimes into the GPU object model +//! system. Then, GPU VA pointers also inherit those lifetimes, which means the Rust borrow checker +//! can ensure that all pointers are assigned an address that is guaranteed to outlive the GPU +//! object it points to. +//! +//! Since the firmware object model does have self-referencing pointers (and there is of course no +//! underlying revocability mechanism to make it safe), we must have an escape hatch. GPU pointers +//! can be weak pointers, which do not enforce lifetimes. In those cases, it is the user's +//! responsibility to ensure that lifetime requirements are met. +//! +//! In other words, the model is necessarily leaky and there is no way to fully map Rust safety to +//! GPU firmware object safety. The goal of the model is to make it easy to model the lifetimes of +//! GPU objects and have the compiler help in avoiding mistakes, rather than to guarantee safety +//! 100% of the time as would be the case for CPU-side Rust code. + +// TODO: There is a fundamental soundness issue with sharing memory with the GPU (that even affects +// C code too). Since the GPU is free to mutate that memory at any time, normal reference invariants +// cannot be enforced on the CPU side. For example, the compiler could perform an optimization that +// assumes that a given memory location does not change between two reads, and causes UB otherwise, +// and then the GPU could mutate that memory out from under the CPU. +// +// For cases where we *expect* this to happen, we use atomic types, which avoid this issue. However, +// doing so for every single field of every type is a non-starter. Right now, there seems to be no +// good solution for this that does not come with significant performance or ergonomics downsides. +// +// In *practice* we are almost always only writing GPU memory, and only reading from atomics, so the +// chances of this actually triggering UB (e.g. a security issue that can be triggered from the GPU +// side) due to a compiler optimization are very slim. +// +// Further discussion: https://github.com/rust-lang/unsafe-code-guidelines/issues/152 + +use kernel::{error::code::*, prelude::*}; + +use alloc::boxed::Box; +use core::fmt; +use core::fmt::Debug; +use core::fmt::Formatter; +use core::marker::PhantomData; +use core::mem::MaybeUninit; +use core::num::NonZeroU64; +use core::ops::{Deref, DerefMut, Index, IndexMut}; +use core::{mem, ptr, slice}; + +use crate::alloc::Allocation; +use crate::debug::*; +use crate::fw::types::Zeroed; + +const DEBUG_CLASS: DebugFlags = DebugFlags::Object; + +/// A GPU-side strong pointer, which is a 64-bit non-zero VA with an associated lifetime. +/// +/// In rare cases these pointers are not aligned, so this is `packed(1)`. +#[repr(C, packed(1))] +pub(crate) struct GpuPointer<'a, T: ?Sized>(NonZeroU64, PhantomData<&'a T>); + +impl<'a, T: ?Sized> GpuPointer<'a, T> { + /// Logical OR the pointer with an arbitrary `u64`. This is used when GPU struct fields contain + /// misc flag fields in the upper bits. The lifetime is retained. This is GPU-unsafe in + /// principle, but we assert that only non-implemented address bits are touched, which is safe + /// for pointers used by the GPU (not by firmware). + pub(crate) fn or(&self, other: u64) -> GpuPointer<'a, T> { + // This will fail for kernel-half pointers, which should not be ORed. + assert_eq!(self.0.get() & other, 0); + // Assert that we only touch the high bits. + assert_eq!(other & 0xffffffffff, 0); + GpuPointer(self.0 | other, PhantomData) + } + + /// Add an arbitrary offset to the pointer. This is not safe (from the GPU perspective), and + /// should only be used via the `inner_ptr` macro to get pointers to inner fields, hence we mark + /// it `unsafe` to discourage direct use. + // NOTE: The third argument is a type inference hack. + pub(crate) unsafe fn offset(&self, off: usize, _: *const U) -> GpuPointer<'a, U> { + GpuPointer::<'a, U>( + NonZeroU64::new(self.0.get() + (off as u64)).unwrap(), + PhantomData, + ) + } +} + +impl<'a, T: ?Sized> Debug for GpuPointer<'a, T> { + fn fmt(&self, f: &mut Formatter<'_>) -> fmt::Result { + let val = self.0; + f.write_fmt(format_args!("{:#x} ({})", val, core::any::type_name::())) + } +} + +/// Take a pointer to a sub-field within a structure pointed to by a GpuPointer, keeping the +/// lifetime. +#[macro_export] +macro_rules! inner_ptr { + ($gpuva:expr, $($f:tt)*) => ({ + // This mirrors kernel::offset_of(), except we use type inference to avoid having to know + // the type of the pointer explicitly. + fn uninit_from<'a, T: GpuStruct>(_: GpuPointer<'a, T>) -> core::mem::MaybeUninit> { + core::mem::MaybeUninit::uninit() + } + let tmp = uninit_from($gpuva); + let outer = tmp.as_ptr(); + // SAFETY: The pointer is valid and aligned, just not initialised; `addr_of` ensures that + // we don't actually read from `outer` (which would be UB) nor create an intermediate + // reference. + let p: *const _ = unsafe { core::ptr::addr_of!((*outer).$($f)*) }; + let inner = p as *const u8; + // SAFETY: The two pointers are within the same allocation block. + let off = unsafe { inner.offset_from(outer as *const u8) }; + // SAFETY: The resulting pointer is guaranteed to point to valid memory within the outer + // object. + unsafe { $gpuva.offset(off.try_into().unwrap(), p) } + }) +} + +/// A GPU-side weak pointer, which is a 64-bit non-zero VA with no lifetime. +/// +/// In rare cases these pointers are not aligned, so this is `packed(1)`. +#[repr(C, packed(1))] +pub(crate) struct GpuWeakPointer(NonZeroU64, PhantomData<*const T>); + +/// SAFETY: GPU weak pointers are always safe to share between threads. +unsafe impl Send for GpuWeakPointer {} +unsafe impl Sync for GpuWeakPointer {} + +// Weak pointers can be copied/cloned regardless of their target type. +impl Copy for GpuWeakPointer {} + +impl Clone for GpuWeakPointer { + fn clone(&self) -> Self { + *self + } +} + +impl GpuWeakPointer { + /// Add an arbitrary offset to the pointer. This is not safe (from the GPU perspective), and + /// should only be used via the `inner_ptr` macro to get pointers to inner fields, hence we mark + /// it `unsafe` to discourage direct use. + // NOTE: The third argument is a type inference hack. + pub(crate) unsafe fn offset(&self, off: usize, _: *const U) -> GpuWeakPointer { + GpuWeakPointer::( + NonZeroU64::new(self.0.get() + (off as u64)).unwrap(), + PhantomData, + ) + } + + /// Upgrade a weak pointer into a strong pointer. This is not considered safe from the GPU + /// perspective. + pub(crate) unsafe fn upgrade<'a>(&self) -> GpuPointer<'a, T> { + GpuPointer(self.0, PhantomData) + } +} + +impl Debug for GpuWeakPointer { + fn fmt(&self, f: &mut Formatter<'_>) -> fmt::Result { + let val = self.0; + f.write_fmt(format_args!("{:#x} ({})", val, core::any::type_name::())) + } +} + +/// Take a pointer to a sub-field within a structure pointed to by a GpuWeakPointer. +#[macro_export] +macro_rules! inner_weak_ptr { + ($gpuva:expr, $($f:tt)*) => ({ + // See inner_ptr() + fn uninit_from(_: GpuWeakPointer) -> core::mem::MaybeUninit> { + core::mem::MaybeUninit::uninit() + } + let tmp = uninit_from($gpuva); + let outer = tmp.as_ptr(); + // SAFETY: The pointer is valid and aligned, just not initialised; `addr_of` ensures that + // we don't actually read from `outer` (which would be UB) nor create an intermediate + // reference. + let p: *const _ = unsafe { core::ptr::addr_of!((*outer).$($f)*) }; + let inner = p as *const u8; + // SAFETY: The two pointers are within the same allocation block. + let off = unsafe { inner.offset_from(outer as *const u8) }; + // SAFETY: The resulting pointer is guaranteed to point to valid memory within the outer + // object. + unsafe { $gpuva.offset(off.try_into().unwrap(), p) } + }) +} + +/// Types that implement this trait represent a GPU structure from the CPU side. +/// +/// The `Raw` type represents the actual raw structure definition on the GPU side. +/// +/// Types implementing [`GpuStruct`] must have fields owning any objects (or strong references +/// to them) that GPU pointers in the `Raw` structure point to. This mechanism is used to enforce +/// lifetimes. +pub(crate) trait GpuStruct: 'static { + /// The type of the GPU-side structure definition representing the firmware struct layout. + type Raw<'a>; +} + +/// An instance of a GPU object in memory. +/// +/// # Invariants +/// `raw` must point to a valid mapping of the `T::Raw` type associated with the `alloc` allocation. +/// `gpu_ptr` must be the GPU address of the same object. +pub(crate) struct GpuObject> { + raw: *mut T::Raw<'static>, + alloc: U, + gpu_ptr: GpuWeakPointer, + inner: Box, +} + +impl> GpuObject { + /// Create a new GpuObject given an allocator and the inner data (a type implementing + /// GpuStruct). + /// + /// The caller passes a closure that constructs the `T::Raw` type given a reference to the + /// `GpuStruct`. This is the mechanism used to enforce lifetimes. + pub(crate) fn new( + alloc: U, + inner: T, + callback: impl for<'a> FnOnce(&'a T) -> T::Raw<'a>, + ) -> Result { + let size = mem::size_of::>(); + if size > 0x1000 { + dev_crit!( + alloc.device(), + "Allocating {} of size {:#x}, with new, please use new_boxed!\n", + core::any::type_name::(), + size + ); + } + if alloc.size() < size { + return Err(ENOMEM); + } + let gpu_ptr = + GpuWeakPointer::(NonZeroU64::new(alloc.gpu_ptr()).ok_or(EINVAL)?, PhantomData); + mod_dev_dbg!( + alloc.device(), + "Allocating {} @ {:#x}\n", + core::any::type_name::(), + alloc.gpu_ptr() + ); + let p = alloc.ptr().ok_or(EINVAL)?.as_ptr() as *mut T::Raw<'static>; + let mut raw = callback(&inner); + // SAFETY: `p` is guaranteed to be valid per the Allocation invariant, and the type is + // identical to the type of `raw` other than the lifetime. + unsafe { p.copy_from(&mut raw as *mut _ as *mut u8 as *mut _, 1) }; + mem::forget(raw); + Ok(Self { + raw: p, + gpu_ptr, + alloc, + inner: Box::try_new(inner)?, + }) + } + + /// Create a new GpuObject given an allocator and the boxed inner data (a type implementing + /// GpuStruct). + /// + /// The caller passes a closure that initializes the `T::Raw` type given a reference to the + /// `GpuStruct` and a `MaybeUninit`. This is intended to be used with the place!() + /// macro to avoid constructing the whole `T::Raw` object on the stack. + pub(crate) fn new_boxed( + alloc: U, + inner: Box, + callback: impl for<'a> FnOnce( + &'a T, + &'a mut MaybeUninit>, + ) -> Result<&'a mut T::Raw<'a>>, + ) -> Result { + if alloc.size() < mem::size_of::>() { + return Err(ENOMEM); + } + let gpu_ptr = + GpuWeakPointer::(NonZeroU64::new(alloc.gpu_ptr()).ok_or(EINVAL)?, PhantomData); + mod_dev_dbg!( + alloc.device(), + "Allocating {} @ {:#x}\n", + core::any::type_name::(), + alloc.gpu_ptr() + ); + let p = alloc.ptr().ok_or(EINVAL)?.as_ptr() as *mut MaybeUninit>; + // SAFETY: `p` is guaranteed to be valid per the Allocation invariant. + let raw = callback(&inner, unsafe { &mut *p })?; + if p as *mut T::Raw<'_> != raw as *mut _ { + dev_err!( + alloc.device(), + "Allocation callback returned a mismatched reference ({})\n", + core::any::type_name::(), + ); + return Err(EINVAL); + } + Ok(Self { + raw: p as *mut u8 as *mut T::Raw<'static>, + gpu_ptr, + alloc, + inner, + }) + } + + /// Create a new GpuObject given an allocator and the inner data (a type implementing + /// GpuStruct). + /// + /// The caller passes a closure that initializes the `T::Raw` type given a reference to the + /// `GpuStruct` and a `MaybeUninit`. This is intended to be used with the place!() + /// macro to avoid constructing the whole `T::Raw` object on the stack. + pub(crate) fn new_inplace( + alloc: U, + inner: T, + callback: impl for<'a> FnOnce( + &'a T, + &'a mut MaybeUninit>, + ) -> Result<&'a mut T::Raw<'a>>, + ) -> Result { + GpuObject::::new_boxed(alloc, Box::try_new(inner)?, callback) + } + + /// Create a new GpuObject given an allocator, with callback-based initialization. + /// + /// This is used when the construction of the `T` type requires knowing the GPU VA address of + /// the structure that is being constructed ahead of time. The first callback constructs a + /// `Box` given the pointer to the about-to-be-initialized GPU structure, and the second + /// callback initializes that structure as in `new_boxed`. + pub(crate) fn new_prealloc( + alloc: U, + inner_cb: impl FnOnce(GpuWeakPointer) -> Result>, + raw_cb: impl for<'a> FnOnce( + &'a T, + &'a mut MaybeUninit>, + ) -> Result<&'a mut T::Raw<'a>>, + ) -> Result { + if alloc.size() < mem::size_of::>() { + return Err(ENOMEM); + } + let gpu_ptr = + GpuWeakPointer::(NonZeroU64::new(alloc.gpu_ptr()).ok_or(EINVAL)?, PhantomData); + mod_dev_dbg!( + alloc.device(), + "Allocating {} @ {:#x}\n", + core::any::type_name::(), + alloc.gpu_ptr() + ); + let inner = inner_cb(gpu_ptr)?; + let p = alloc.ptr().ok_or(EINVAL)?.as_ptr() as *mut MaybeUninit>; + // SAFETY: `p` is guaranteed to be valid per the Allocation invariant. + let raw = raw_cb(&*inner, unsafe { &mut *p })?; + if p as *mut T::Raw<'_> != raw as *mut _ { + dev_err!( + alloc.device(), + "Allocation callback returned a mismatched reference ({})\n", + core::any::type_name::(), + ); + return Err(EINVAL); + } + Ok(Self { + raw: p as *mut u8 as *mut T::Raw<'static>, + gpu_ptr, + alloc, + inner, + }) + } + + /// Returns the GPU VA of this object (as a raw [`NonZeroU64`]) + pub(crate) fn gpu_va(&self) -> NonZeroU64 { + self.gpu_ptr.0 + } + + /// Returns a strong GPU pointer to this object, with a lifetime. + pub(crate) fn gpu_pointer(&self) -> GpuPointer<'_, T> { + GpuPointer(self.gpu_ptr.0, PhantomData) + } + + /// Returns a weak GPU pointer to this object, with no lifetime. + pub(crate) fn weak_pointer(&self) -> GpuWeakPointer { + GpuWeakPointer(self.gpu_ptr.0, PhantomData) + } + + /// Perform a mutation to the inner `Raw` data given a user-supplied callback. + /// + /// The callback gets a mutable reference to the `GpuStruct` type. + pub(crate) fn with_mut( + &mut self, + callback: impl for<'a> FnOnce(&'a mut ::Raw<'a>, &'a mut T) -> RetVal, + ) -> RetVal { + // SAFETY: `self.raw` is valid per the type invariant, and the second half is just + // converting lifetimes. + unsafe { callback(&mut *self.raw, &mut *(&mut *self.inner as *mut _)) } + } + + /// Access the inner `Raw` data given a user-supplied callback. + /// + /// The callback gets a reference to the `GpuStruct` type. + pub(crate) fn with( + &self, + callback: impl for<'a> FnOnce(&'a ::Raw<'a>, &'a T) -> RetVal, + ) -> RetVal { + // SAFETY: `self.raw` is valid per the type invariant, and the second half is just + // converting lifetimes. + unsafe { callback(&*self.raw, &*(&*self.inner as *const _)) } + } +} + +impl> Deref for GpuObject { + type Target = T; + + fn deref(&self) -> &Self::Target { + &self.inner + } +} + +impl> DerefMut for GpuObject { + fn deref_mut(&mut self) -> &mut Self::Target { + &mut self.inner + } +} + +impl> Debug for GpuObject +where + ::Raw<'static>: Debug, +{ + fn fmt(&self, f: &mut Formatter<'_>) -> fmt::Result { + f.debug_struct(core::any::type_name::()) + // SAFETY: `self.raw` is valid per the type invariant. + .field("raw", &format_args!("{:#X?}", unsafe { &*self.raw })) + .field("inner", &format_args!("{:#X?}", &self.inner)) + .field("alloc", &format_args!("{:?}", &self.alloc)) + .finish() + } +} + +impl> GpuObject +where + for<'a> ::Raw<'a>: Default + Zeroed, +{ + /// Create a new GpuObject with default data. `T` must implement `Default` and `T::Raw` must + /// implement `Zeroed`, since the GPU-side memory is initialized by zeroing. + pub(crate) fn new_default(alloc: U) -> Result { + GpuObject::::new_inplace(alloc, Default::default(), |_inner, raw| { + // SAFETY: `raw` is valid here, and `T::Raw` implements `Zeroed`. + Ok(unsafe { + ptr::write_bytes(raw, 0, 1); + (*raw).assume_init_mut() + }) + }) + } +} + +impl> Drop for GpuObject { + fn drop(&mut self) { + mod_dev_dbg!( + self.alloc.device(), + "Dropping {} @ {:?}\n", + core::any::type_name::(), + self.gpu_pointer() + ); + } +} + +// SAFETY: GpuObjects are Send as long as the GpuStruct itself is Send +unsafe impl> Send for GpuObject {} +// SAFETY: GpuObjects are Send as long as the GpuStruct itself is Send +unsafe impl> Sync for GpuObject {} + +/// Trait used to erase the type of a GpuObject, used when we need to keep a list of heterogenous +/// objects around. +pub(crate) trait OpaqueGpuObject: Send + Sync { + fn gpu_va(&self) -> NonZeroU64; +} + +impl> OpaqueGpuObject for GpuObject { + fn gpu_va(&self) -> NonZeroU64 { + Self::gpu_va(self) + } +} + +/// An array of raw GPU objects that is only accessible to the GPU (no CPU-side mapping required). +/// +/// This must necessarily be uninitialized as far as the GPU is concerned, so it cannot be used +/// when initialization is required. +/// +/// # Invariants +/// +/// `alloc` is valid and at least as large as `len` times the size of one `T`. +/// `gpu_ptr` is valid and points to the allocation start. +pub(crate) struct GpuOnlyArray> { + len: usize, + alloc: U, + gpu_ptr: NonZeroU64, + _p: PhantomData, +} + +impl> GpuOnlyArray { + /// Allocate a new GPU-only array with the given length. + pub(crate) fn new(alloc: U, count: usize) -> Result> { + let bytes = count * mem::size_of::(); + let gpu_ptr = NonZeroU64::new(alloc.gpu_ptr()).ok_or(EINVAL)?; + if alloc.size() < bytes { + return Err(ENOMEM); + } + Ok(Self { + len: count, + alloc, + gpu_ptr, + _p: PhantomData, + }) + } + + /// Returns the GPU VA of this arraw (as a raw [`NonZeroU64`]) + pub(crate) fn gpu_va(&self) -> NonZeroU64 { + self.gpu_ptr + } + + /// Returns a strong GPU pointer to this array, with a lifetime. + pub(crate) fn gpu_pointer(&self) -> GpuPointer<'_, &'_ [T]> { + GpuPointer(self.gpu_ptr, PhantomData) + } + + /// Returns a weak GPU pointer to this array, with no lifetime. + pub(crate) fn weak_pointer(&self) -> GpuWeakPointer<[T]> { + GpuWeakPointer(self.gpu_ptr, PhantomData) + } + + /// Returns a pointer to an offset within the array (as a subslice). + pub(crate) fn gpu_offset_pointer(&self, offset: usize) -> GpuPointer<'_, &'_ [T]> { + if offset > self.len { + panic!("Index {} out of bounds (len: {})", offset, self.len); + } + GpuPointer( + NonZeroU64::new(self.gpu_ptr.get() + (offset * mem::size_of::()) as u64).unwrap(), + PhantomData, + ) + } + + /* Not used yet + /// Returns a weak pointer to an offset within the array (as a subslice). + pub(crate) fn weak_offset_pointer(&self, offset: usize) -> GpuWeakPointer<[T]> { + if offset > self.len { + panic!("Index {} out of bounds (len: {})", offset, self.len); + } + GpuWeakPointer( + NonZeroU64::new(self.gpu_ptr.get() + (offset * mem::size_of::()) as u64).unwrap(), + PhantomData, + ) + } + + /// Returns a pointer to an element within the array. + pub(crate) fn gpu_item_pointer(&self, index: usize) -> GpuPointer<'_, &'_ T> { + if index >= self.len { + panic!("Index {} out of bounds (len: {})", index, self.len); + } + GpuPointer( + NonZeroU64::new(self.gpu_ptr.get() + (index * mem::size_of::()) as u64).unwrap(), + PhantomData, + ) + } + */ + + /// Returns a weak pointer to an element within the array. + pub(crate) fn weak_item_pointer(&self, index: usize) -> GpuWeakPointer { + if index >= self.len { + panic!("Index {} out of bounds (len: {})", index, self.len); + } + GpuWeakPointer( + NonZeroU64::new(self.gpu_ptr.get() + (index * mem::size_of::()) as u64).unwrap(), + PhantomData, + ) + } + + /// Returns the length of the array. + pub(crate) fn len(&self) -> usize { + self.len + } +} + +impl> Debug for GpuOnlyArray { + fn fmt(&self, f: &mut Formatter<'_>) -> fmt::Result { + f.debug_struct(core::any::type_name::()) + .field("len", &format_args!("{:#X?}", self.len())) + .finish() + } +} + +impl> Drop for GpuOnlyArray { + fn drop(&mut self) { + mod_dev_dbg!( + self.alloc.device(), + "Dropping {} @ {:?}\n", + core::any::type_name::(), + self.gpu_pointer() + ); + } +} + +/// An array of raw GPU objects that is also CPU-accessible. +/// +/// # Invariants +/// +/// `raw` is valid and points to the CPU-side view of the array (which must have one). +pub(crate) struct GpuArray> { + raw: *mut T, + array: GpuOnlyArray, +} + +/* Not used yet +impl> GpuArray { + /// Allocate a new GPU array, copying the contents from a slice. + pub(crate) fn new(alloc: U, data: &[T]) -> Result> { + let p = alloc.ptr().ok_or(EINVAL)?.as_ptr(); + let inner = GpuOnlyArray::new(alloc, data.len())?; + // SAFETY: `p` is valid per the Allocation type invariant, and GpuOnlyArray guarantees + // that its size is at least as large as `data.len()`. + unsafe { ptr::copy(data.as_ptr(), p, data.len()) }; + Ok(Self { + raw: p, + array: inner, + }) + } +} +*/ + +impl> GpuArray { + /// Allocate a new GPU array, initializing each element to its default. + pub(crate) fn empty(alloc: U, count: usize) -> Result> { + let p = alloc.ptr().ok_or(EINVAL)?.as_ptr() as *mut T; + let inner = GpuOnlyArray::new(alloc, count)?; + let mut pi = p; + for _i in 0..count { + // SAFETY: `pi` is valid per the Allocation type invariant, and GpuOnlyArray guarantees + // that it can never iterate beyond the buffer length. + unsafe { + pi.write(Default::default()); + pi = pi.add(1); + } + } + Ok(Self { + raw: p, + array: inner, + }) + } +} + +impl> GpuArray { + /// Get a slice view of the array contents. + pub(crate) fn as_slice(&self) -> &[T] { + // SAFETY: self.raw / self.len are valid per the type invariant + unsafe { slice::from_raw_parts(self.raw, self.len) } + } + + /// Get a mutable slice view of the array contents. + pub(crate) fn as_mut_slice(&mut self) -> &mut [T] { + // SAFETY: self.raw / self.len are valid per the type invariant + unsafe { slice::from_raw_parts_mut(self.raw, self.len) } + } +} + +impl> Deref for GpuArray { + type Target = GpuOnlyArray; + + fn deref(&self) -> &GpuOnlyArray { + &self.array + } +} + +impl> Index for GpuArray { + type Output = T; + + fn index(&self, index: usize) -> &T { + if index >= self.len { + panic!("Index {} out of bounds (len: {})", index, self.len); + } + // SAFETY: This is bounds checked above + unsafe { &*(self.raw.add(index)) } + } +} + +impl> IndexMut for GpuArray { + fn index_mut(&mut self, index: usize) -> &mut T { + if index >= self.len { + panic!("Index {} out of bounds (len: {})", index, self.len); + } + // SAFETY: This is bounds checked above + unsafe { &mut *(self.raw.add(index)) } + } +} + +// SAFETY: GpuArray are Send as long as the contained type itself is Send +unsafe impl> Send for GpuArray {} +// SAFETY: GpuArray are Sync as long as the contained type itself is Sync +unsafe impl> Sync for GpuArray {} + +impl> Debug for GpuArray { + fn fmt(&self, f: &mut Formatter<'_>) -> fmt::Result { + f.debug_struct(core::any::type_name::()) + .field("array", &format_args!("{:#X?}", self.as_slice())) + .finish() + } +} diff --git a/drivers/gpu/drm/asahi/place.rs b/drivers/gpu/drm/asahi/place.rs new file mode 100644 index 000000000000..40c51f4fab8d --- /dev/null +++ b/drivers/gpu/drm/asahi/place.rs @@ -0,0 +1,343 @@ +// SPDX-License-Identifier: Apache-2.0 OR MIT + +//! "Placement new" macro +//! +//! This cursed abomination of a declarative macro is used to emulate a "placement new" feature, +//! which allows initializing objects directly in a user-provided memory region without first +//! going through the stack. +//! +//! This driver needs to manage several large GPU objects of a fixed layout. Linux kernel stacks are +//! very small, so it is impossible to create these objects on the stack. While the compiler can +//! sometimes optimize away the stack copy and directly instantiate in target memory, this is not +//! guaranteed and not reliable. Therefore, we need some mechanism to ergonomically initialize +//! complex structures directly in a pre-allocated piece of memory. +//! +//! This issue also affects some driver-internal structs which are large/complex enough to overflow +//! the stack. While this can be solved by breaking them up into pieces and using `Box` more +//! liberally, this has performance implications and still isn't very nice. This macro can also be +//! used to solve this issue. +//! +//! # Further reading +//! https://github.com/rust-lang/rust/issues/27779#issuecomment-378416911 +//! https://internals.rust-lang.org/t/removal-of-all-unstable-placement-features/7223 + +/// Initialize a `MaybeUninit` in-place, without constructing the value on the stack first. +/// +/// This macro is analogous to `MaybeUninit::write()`. In other words, +/// `place!(foo, bar)` is equivalent to `MaybeUninit::write(foo, bar)`, except that `bar` is not +/// constructed first, but rather its fields (if it is a structure constructor) are copied one by +/// one into the correct location in the `MaybeUninit`. +/// +/// The macro supports most Rust initialization syntax including type paths, generic arguments, +/// and nested structures. Nested structures are themselves initialized in-place field by field. +/// `..Default::default()` is supported, but this macro converts it to `..Zeroed::zeroed()`, as it +/// initializes those structs by zero-initializing the underlying memory. Usage of +/// `..Default::default()` with a type not implementing `Zeroed` will result in a compile error. +/// +/// Usage: +/// ``` +/// let mut buf = MaybeUninit::uninit(); +/// let mut_ref = place!(&mut buf, MyStruct { +/// b: true, +/// s: String::from("works"), +/// i: str::parse::("123").unwrap(), +/// v: vec![String::from("works")], +/// x: foo::MyOtherCoolStruct { +/// a: false, +/// b: String::from("Hello, world!"), +/// }, +/// y: foo::MyOtherCoolStruct { +/// a: false, +/// b: String::from("Hello, world!"), +/// }, +/// z: foo::MyCoolGenericStruct:: { +/// a: false, +/// b: String::from("Hello, world!"), +/// }, +/// }; +/// // `mut_ref` is now a mutable reference to the `buf`, which is now safely initialized. +/// ``` +/// +/// Based on https://crates.io/crates/place by DianaNites, with contributions by Joshua Barretto. +#[macro_export] +macro_rules! place { + // Top-level struct + (@STRUCT $ptr:ident, _TOP, $typ:path, {$($typ_init:tt)*} { $($fields:tt)* }) => {{ + place!(@STRUCT_ZERO $ptr, {$($typ_init)*} { $($fields)* }); + place!(@STRUCT_CHECK $ptr, {$($typ_init)*} { $($fields)* } { + place!(@FIELDS $ptr, $($fields)*); + }); + }}; + // Nested structure + (@STRUCT $ptr:ident, $f_struct:ident, $typ:path, {$($typ_init:tt)*} { $($fields:tt)* }) => {{ + use core::ptr::addr_of_mut; + let buf = unsafe { addr_of_mut!((*$ptr).$f_struct) }; + place!(@STRUCT_ZERO buf, {$($typ_init)*} { $($fields)* }); + place!(@STRUCT_CHECK $ptr, {$($typ_init)*} { $($fields)* } { + place!(@FIELDS buf, $($fields)*); + }); + }}; + + // Zero-initialize structure if the initializer ends in ..default::Default() + (@STRUCT_ZERO $ptr:ident, {$($typ_init:tt)*} { $($f:ident $(: $v:expr)?),* $(,)? }) => {}; + (@STRUCT_ZERO $ptr:ident, {$($typ_init:tt)*} { $($($f:ident $(: $v:expr)?),*,)? ..Default::default() }) => {{ + // Check that the structure actually implements Zeroed + const _: () = { + fn _check_default() { + let _ = $($typ_init)* { + ..Zeroed::zeroed() + }; + } + }; + use core::ptr; + unsafe { ptr::write_bytes($ptr, 0, 1) }; + + }}; + + // Check that all fields are specified + (@STRUCT_CHECK $ptr:ident, {$($typ_init:tt)*} { $($($f:ident $(: $v:expr)?),*,)? ..Default::default() } {$($body:tt)*}) => { + if false { + #[allow(clippy::redundant_field_names)] + let _x = $($typ_init)* { + $($( + $f $(: $v)? + ),* + ,)? + ..Zeroed::zeroed() + }; + } else { + {$($body)*} + } + }; + (@STRUCT_CHECK $ptr:ident, {$($typ_init:tt)*} { $($f:ident $(: $v:expr)?),* $(,)? } {$($body:tt)*}) => { + if false { + #[allow(clippy::redundant_field_names)] + let _x = $($typ_init)* { + $( + $f $(: $v)? + ),* + }; + } else { + {$($body)*} + } + }; + // Top-level scalar + (@SCALAR $ptr:ident, _TOP, $val:expr) => { + let tmp = $val; + unsafe { $ptr.write(tmp); } + }; + // Regular field + (@SCALAR $ptr:ident, $f:ident, $val:expr) => {{ + use core::ptr::addr_of_mut; + let tmp = $val; + unsafe { addr_of_mut!((*$ptr).$f).write(tmp); } + }}; + // Type-like name followed by braces is a nested structure + (@PARTIAL $ptr:ident, $f:ident, {$($head:tt)*}, {{ $($fields:tt)* } $($tail:tt)*}) => { + place!(@STRUCT $ptr, $f, $($head)*, {$($head)*} { $($fields)* }); + place!(@FIELDS $ptr $($tail)*) + }; + // Type-like name followed by ::ident, append to head + (@PARTIAL $ptr:ident, $f:ident, {$($head:tt)*}, {::$id:ident $($tail:tt)*}) => { + place!(@PARTIAL $ptr, $f, {$($head)* :: $id}, {$($tail)*}); + }; + // Type-like name followed by ::, append to head + (@PARTIAL $ptr:ident, $f:ident, {$($head:tt)*}, {::<$($gen:ty),*> $($tail:tt)*}) => { + place!(@PARTIAL $ptr, $f, {$($head)* :: <$($gen),*>}, {$($tail)*}); + }; + // Type-like name followed by ::<'lifetime>, append to head + (@PARTIAL $ptr:ident, $f:ident, {$($head:tt)*}, {::<$li:lifetime> $($tail:tt)*}) => { + place!(@PARTIAL $ptr, $f, {$($head)* :: <$li>}, {$($tail)*}); + }; + // Anything else, parse it as an expression + (@PARTIAL $ptr:ident, $f:ident, {$($head:tt)*}, {$($tail:tt)*}) => { + place!(@EXPR $ptr, $f, $($head)* $($tail)*) + }; + // Expression followed by more fields + (@EXPR $ptr:ident, $f:ident, $val:expr, $($tail:tt)*) => { + place!(@SCALAR $ptr, $f, $val); + place!(@FIELDS $ptr, $($tail)*) + }; + // Last field expression, without a trailing comma + (@EXPR $ptr:ident, $f:ident, $val:expr) => { + place!(@SCALAR $ptr, $f, $val); + }; + // Field with a value starting with an ident, start incremental type parsing + (@FIELDS $ptr:ident, $f:ident : $id:ident $($tail:tt)*) => { + place!(@PARTIAL $ptr, $f, {$id}, {$($tail)*}); + }; + // Same, but starting with ::ident + (@FIELDS $ptr:ident, $f:ident : ::$id:ident $($tail:tt)*) => { + place!(@PARTIAL $ptr, $f, {::$id}, {$($tail)*}); + }; + // Otherwise, parse it as an expression + (@FIELDS $ptr:ident, $f:ident : $($tail:tt)*) => { + place!(@EXPR $ptr, $f, $($tail)*) + }; + // Default terminating case + (@FIELDS $ptr:ident, ..Default::default() ) => {}; + // Terminating case + (@FIELDS $ptr:ident $(,)? ) => {}; + ( + $buf:expr, + $($val:tt)* + ) => {{ + use core::mem::MaybeUninit; + // Ensures types are correct + let obj: &mut MaybeUninit<_> = $buf; + let top_ptr = obj.as_mut_ptr(); + place!(@FIELDS top_ptr, _TOP: $($val)*); + // SAFETY: All fields have been initialized above + // The compiler ensures that all fields were used, all types were correct, + // and that size and alignment are correct. + unsafe { obj.assume_init_mut() } + }}; +} + +/// Helper macro to get the struct type part of a struct initialization expression. +#[macro_export] +#[doc(hidden)] +macro_rules! get_type { + ($t:ty { $($val:tt)* }) => { + $t + }; +} + +/// Like `Box::try_new(...)`, but with in-place initialization. +#[macro_export] +macro_rules! box_in_place { + ($($val:tt)*) => {{ + use $crate::place; + let b = Box::<$crate::get_type!($($val)*)>::try_new_uninit(); + match b { + Ok(mut p) => { + place!((&mut *p), $($val)*); + Ok(unsafe { p.assume_init() }) + } + Err(e) => Err(e) + } + }}; +} + +// TODO: figure out how to make this run +#[cfg(test)] +mod tests { + use super::*; + use core::mem::MaybeUninit; + + #[derive(Debug, PartialEq)] + struct MyCoolStruct { + b: bool, + s: String, + i: i32, + v: Vec, + x: MyOtherCoolStruct, + y: MyOtherCoolStruct, + z: foo::MyCoolGenericStruct, + } + + #[derive(Debug, PartialEq)] + struct MyDefaultStruct { + b: bool, + i: i32, + j: i16, + } + default_zeroed!(MyDefaultStruct); + + mod foo { + #[derive(Debug, PartialEq)] + pub struct MyOtherCoolStruct { + pub a: bool, + pub b: String, + } + #[derive(Debug, PartialEq)] + pub struct MyCoolGenericStruct { + pub a: T, + pub b: U, + } + } + + use foo::MyOtherCoolStruct; + + #[test] + fn test_initialized() { + let mut buf: MaybeUninit = MaybeUninit::uninit(); + + let x: &mut MyCoolStruct = place!( + &mut buf, + MyCoolStruct { + b: true, + s: String::from("works"), + i: str::parse::("123").unwrap(), + v: vec![String::from("works")], + x: MyOtherCoolStruct { + a: false, + b: String::from("Hello, world!"), + }, + y: foo::MyOtherCoolStruct { + a: false, + b: String::from("Hello, world!"), + }, + z: foo::MyCoolGenericStruct:: { + a: false, + b: String::from("Hello, world!"), + } + } + ); + //dbg!(x); + + assert_eq!( + x, + &MyCoolStruct { + b: true, + s: String::from("works"), + i: str::parse::("123").unwrap(), + v: vec![String::from("works")], + x: foo::MyOtherCoolStruct { + a: false, + b: String::from("Hello, world!"), + }, + y: foo::MyOtherCoolStruct { + a: false, + b: String::from("Hello, world!"), + }, + z: foo::MyCoolGenericStruct:: { + a: false, + b: String::from("Hello, world!"), + }, + }, + ); + } + + #[test] + fn test_default() { + let mut buf: MaybeUninit = MaybeUninit::uninit(); + + let x: &mut MyDefaultStruct = place!( + &mut buf, + MyDefaultStruct { + b: true, + i: 1, + ..Default::default() + } + ); + + assert_eq!( + x, + &MyDefaultStruct { + b: true, + i: 1, + j: 0, + }, + ); + } + + #[test] + fn test_scalar() { + let mut buf: MaybeUninit = MaybeUninit::uninit(); + + let x: &mut u32 = place!(&mut buf, 1234); + + assert_eq!(x, &mut 1234u32); + } +} diff --git a/drivers/gpu/drm/asahi/queue/common.rs b/drivers/gpu/drm/asahi/queue/common.rs new file mode 100644 index 000000000000..127b4ccc6eca --- /dev/null +++ b/drivers/gpu/drm/asahi/queue/common.rs @@ -0,0 +1,52 @@ +// SPDX-License-Identifier: GPL-2.0-only OR MIT + +//! Common queue functionality. +//! +//! Shared helpers used by the submission logic for multiple command types. + +use crate::fw::microseq; +use crate::fw::types::*; + +use kernel::bindings; +use kernel::io_buffer::IoBufferReader; +use kernel::prelude::*; +use kernel::user_ptr::UserSlicePtr; + +use core::mem::MaybeUninit; + +pub(super) fn build_attachments(pointer: u64, count: u32) -> Result { + if count as usize > microseq::MAX_ATTACHMENTS { + return Err(EINVAL); + } + + const STRIDE: usize = core::mem::size_of::(); + let size = STRIDE * count as usize; + + // SAFETY: We only read this once, so there are no TOCTOU issues. + let mut reader = unsafe { UserSlicePtr::new(pointer as usize as *mut _, size).reader() }; + + let mut attachments: microseq::Attachments = Default::default(); + + for i in 0..count { + let mut att: MaybeUninit = MaybeUninit::uninit(); + + // SAFETY: The size of `att` is STRIDE + unsafe { reader.read_raw(att.as_mut_ptr() as *mut u8, STRIDE)? }; + + // SAFETY: All bit patterns in the struct are valid + let att = unsafe { att.assume_init() }; + + let cache_lines = (att.size + 127) >> 7; + let order = 1; + attachments.list[i as usize] = microseq::Attachment { + address: U64(att.pointer), + size: cache_lines, + unk_c: 0x17, + unk_e: order, + }; + + attachments.count += 1; + } + + Ok(attachments) +} diff --git a/drivers/gpu/drm/asahi/queue/compute.rs b/drivers/gpu/drm/asahi/queue/compute.rs new file mode 100644 index 000000000000..6590382c75af --- /dev/null +++ b/drivers/gpu/drm/asahi/queue/compute.rs @@ -0,0 +1,371 @@ +// SPDX-License-Identifier: GPL-2.0-only OR MIT +#![allow(clippy::unusual_byte_groupings)] + +//! Compute work queue. +//! +//! A compute queue consists of one underlying WorkQueue. +//! This module is in charge of creating all of the firmware structures required to submit compute +//! work to the GPU, based on the userspace command buffer. + +use super::common; +use crate::alloc::Allocator; +use crate::debug::*; +use crate::fw::types::*; +use crate::gpu::GpuManager; +use crate::{box_in_place, inner_ptr, inner_weak_ptr, place}; +use crate::{fw, gpu, microseq}; +use core::mem::MaybeUninit; +use core::sync::atomic::Ordering; +use kernel::bindings; +use kernel::dma_fence::RawDmaFence; +use kernel::drm::sched::Job; +use kernel::io_buffer::IoBufferReader; +use kernel::prelude::*; +use kernel::sync::Arc; +use kernel::user_ptr::UserSlicePtr; + +const DEBUG_CLASS: DebugFlags = DebugFlags::Compute; + +#[versions(AGX)] +impl super::Queue::ver { + /// Submit work to a compute queue. + pub(super) fn submit_compute( + &self, + job: &mut Job, + cmd: &bindings::drm_asahi_command, + result_writer: Option, + id: u64, + flush_stamps: bool, + ) -> Result { + if cmd.cmd_type != bindings::drm_asahi_cmd_type_DRM_ASAHI_CMD_COMPUTE { + return Err(EINVAL); + } + + let dev = self.dev.data(); + let gpu = match dev.gpu.as_any().downcast_ref::() { + Some(gpu) => gpu, + None => { + dev_crit!(self.dev, "GpuManager mismatched with Queue!\n"); + return Err(EIO); + } + }; + + let mut alloc = gpu.alloc(); + let kalloc = &mut *alloc; + + mod_dev_dbg!(self.dev, "[Submission {}] Compute!\n", id); + + let mut cmdbuf_reader = unsafe { + UserSlicePtr::new( + cmd.cmd_buffer as usize as *mut _, + core::mem::size_of::(), + ) + .reader() + }; + + let mut cmdbuf: MaybeUninit = MaybeUninit::uninit(); + unsafe { + cmdbuf_reader.read_raw( + cmdbuf.as_mut_ptr() as *mut u8, + core::mem::size_of::(), + )?; + } + let cmdbuf = unsafe { cmdbuf.assume_init() }; + + if cmdbuf.flags != 0 { + return Err(EINVAL); + } + + // This sequence number increases per new client/VM? assigned to some slot, + // but it's unclear *which* slot... + let slot_client_seq: u8 = (self.id & 0xff) as u8; + + let vm_bind = job.vm_bind.clone(); + + mod_dev_dbg!( + self.dev, + "[Submission {}] VM slot = {}\n", + id, + vm_bind.slot() + ); + + let notifier = self.notifier.clone(); + + let fence = job.fence.clone(); + let comp_job = job.get_comp()?; + let ev_comp = comp_job.event_info(); + + // TODO: Is this the same on all GPUs? Is this really for preemption? + let preempt_size = 0x7fa0; + let preempt2_off = 0x7f80; + let preempt3_off = 0x7f88; + let preempt4_off = 0x7f90; + let preempt5_off = 0x7f98; + + let preempt_buf = self.ualloc.lock().array_empty(preempt_size)?; + + let mut seq_buf = self.ualloc.lock().array_empty(0x800)?; + for i in 1..0x400 { + seq_buf[i] = (i + 1) as u64; + } + + mod_dev_dbg!( + self.dev, + "[Submission {}] Event #{} {:#x?} -> {:#x?}\n", + id, + ev_comp.slot, + ev_comp.value, + ev_comp.value.next(), + ); + + let timestamps = Arc::try_new(kalloc.shared.new_default::()?)?; + + let uuid = cmdbuf.cmd_id; + + let unk3 = debug_enabled(debug::DebugFlags::Debug3); + + mod_dev_dbg!(self.dev, "[Submission {}] UUID = {:#x?}\n", id, uuid); + + // TODO: check + #[ver(V >= V13_0B4)] + let count = self.counter.fetch_add(1, Ordering::Relaxed); + + let comp = GpuObject::new_prealloc( + kalloc.private.alloc_object()?, + |ptr: GpuWeakPointer| { + let mut builder = microseq::Builder::new(); + + let stats = gpu.initdata.runtime_pointers.stats.comp.weak_pointer(); + + let start_comp = builder.add(microseq::StartCompute::ver { + header: microseq::op::StartCompute::HEADER, + unk_pointer: inner_weak_ptr!(ptr, unk_pointee), + job_params1: inner_weak_ptr!(ptr, job_params1), + stats, + work_queue: ev_comp.info_ptr, + vm_slot: vm_bind.slot(), + unk_28: 0x1, + event_generation: self.id as u32, + cmd_seq: U64(ev_comp.cmd_seq), + unk_38: 0x0, + job_params2: inner_weak_ptr!(ptr, job_params2), + unk_44: 0x0, + uuid, + attachments: common::build_attachments( + cmdbuf.attachments, + cmdbuf.attachment_count, + )?, + padding: Default::default(), + #[ver(V >= V13_0B4)] + unk_flag: inner_weak_ptr!(ptr, unk_flag), + #[ver(V >= V13_0B4)] + counter: U64(count), + #[ver(V >= V13_0B4)] + notifier_buf: inner_weak_ptr!(notifier.weak_pointer(), state.unk_buf), + })?; + + if result_writer.is_some() { + builder.add(microseq::Timestamp::ver { + header: microseq::op::Timestamp::new(true), + cur_ts: inner_weak_ptr!(ptr, cur_ts), + start_ts: inner_weak_ptr!(ptr, start_ts), + update_ts: inner_weak_ptr!(ptr, start_ts), + work_queue: ev_comp.info_ptr, + unk_24: U64(0), + #[ver(V >= V13_0B4)] + unk_ts: inner_weak_ptr!(ptr, unk_ts), + uuid, + unk_30_padding: 0, + })?; + } + + builder.add(microseq::WaitForIdle { + header: microseq::op::WaitForIdle::new(microseq::Pipe::Compute), + })?; + + if result_writer.is_some() { + builder.add(microseq::Timestamp::ver { + header: microseq::op::Timestamp::new(false), + cur_ts: inner_weak_ptr!(ptr, cur_ts), + start_ts: inner_weak_ptr!(ptr, start_ts), + update_ts: inner_weak_ptr!(ptr, end_ts), + work_queue: ev_comp.info_ptr, + unk_24: U64(0), + #[ver(V >= V13_0B4)] + unk_ts: inner_weak_ptr!(ptr, unk_ts), + uuid, + unk_30_padding: 0, + })?; + } + + let off = builder.offset_to(start_comp); + builder.add(microseq::FinalizeCompute::ver { + header: microseq::op::FinalizeCompute::HEADER, + stats, + work_queue: ev_comp.info_ptr, + vm_slot: vm_bind.slot(), + #[ver(V < V13_0B4)] + unk_18: 0, + job_params2: inner_weak_ptr!(ptr, job_params2), + unk_24: 0, + uuid, + fw_stamp: ev_comp.fw_stamp_pointer, + stamp_value: ev_comp.value.next(), + unk_38: 0, + unk_3c: 0, + unk_40: 0, + unk_44: 0, + unk_48: 0, + unk_4c: 0, + unk_50: 0, + unk_54: 0, + unk_58: 0, + #[ver(G == G14 && V < V13_0B4)] + unk_5c_g14: U64(0), + restart_branch_offset: off, + unk_60: unk3.into(), + #[ver(V >= V13_0B4)] + unk_64: Default::default(), + #[ver(V >= V13_0B4)] + unk_flag: inner_weak_ptr!(ptr, unk_flag), + #[ver(V >= V13_0B4)] + unk_79: Default::default(), + })?; + + builder.add(microseq::RetireStamp { + header: microseq::op::RetireStamp::HEADER, + })?; + + Ok(box_in_place!(fw::compute::RunCompute::ver { + notifier: notifier.clone(), + preempt_buf: preempt_buf, + seq_buf: seq_buf, + micro_seq: builder.build(&mut kalloc.private)?, + vm_bind: vm_bind.clone(), + timestamps: timestamps.clone(), + })?) + }, + |inner, ptr| { + Ok(place!( + ptr, + fw::compute::raw::RunCompute::ver { + tag: fw::workqueue::CommandType::RunCompute, + #[ver(V >= V13_0B4)] + counter: U64(count), + unk_4: 0, + vm_slot: vm_bind.slot(), + notifier: inner.notifier.gpu_pointer(), + unk_pointee: Default::default(), + job_params1: fw::compute::raw::JobParameters1 { + preempt_buf1: inner.preempt_buf.gpu_pointer(), + encoder: U64(cmdbuf.encoder_ptr), + // buf2-5 Only if internal program is used + preempt_buf2: inner.preempt_buf.gpu_offset_pointer(preempt2_off), + preempt_buf3: inner.preempt_buf.gpu_offset_pointer(preempt3_off), + preempt_buf4: inner.preempt_buf.gpu_offset_pointer(preempt4_off), + preempt_buf5: inner.preempt_buf.gpu_offset_pointer(preempt5_off), + pipeline_base: U64(0x11_00000000), + unk_38: U64(0x8c60), + unk_40: cmdbuf.ctx_switch_prog, // Internal program addr | 1 + unk_44: 0, + compute_layout_addr: U64(cmdbuf.buffer_descriptor), // Only if internal program used + unk_50: cmdbuf.buffer_descriptor_size, // 0x40 if internal program used + unk_54: 0, + unk_58: 1, + unk_5c: 0, + iogpu_unk_40: cmdbuf.iogpu_unk_40, // 0x1c if internal program used + }, + unk_b8: Default::default(), + microsequence: inner.micro_seq.gpu_pointer(), + microsequence_size: inner.micro_seq.len() as u32, + job_params2: fw::compute::raw::JobParameters2::ver { + #[ver(V >= V13_0B4)] + unk_0_0: 0, + unk_0: Default::default(), + preempt_buf1: inner.preempt_buf.gpu_pointer(), + encoder_end: U64(cmdbuf.encoder_end), + unk_34: Default::default(), + #[ver(V < V13_0B4)] + unk_5c: 0, + }, + encoder_params: fw::job::raw::EncoderParams { + unk_8: 0x0, // fixed + unk_c: 0x0, // fixed + unk_10: 0x0, // fixed + encoder_id: cmdbuf.encoder_id, + unk_18: 0x0, // fixed + iogpu_compute_unk44: cmdbuf.iogpu_unk_44, + seq_buffer: inner.seq_buf.gpu_pointer(), + unk_28: U64(0x0), // fixed + }, + meta: fw::job::raw::JobMeta { + unk_4: 0, + stamp: ev_comp.stamp_pointer, + fw_stamp: ev_comp.fw_stamp_pointer, + stamp_value: ev_comp.value.next(), + stamp_slot: ev_comp.slot, + evctl_index: 0, // fixed + flush_stamps: flush_stamps as u32, + uuid: uuid, + cmd_seq: ev_comp.cmd_seq as u32, + }, + cur_ts: U64(0), + start_ts: Some(inner_ptr!(inner.timestamps.gpu_pointer(), start)), + end_ts: Some(inner_ptr!(inner.timestamps.gpu_pointer(), end)), + unk_2c0: 0, + unk_2c4: 0, + unk_2c8: 0, + unk_2cc: 0, + client_sequence: slot_client_seq, + pad_2d1: Default::default(), + unk_2d4: 0, + unk_2d8: 0, + #[ver(V >= V13_0B4)] + unk_ts: U64(0), + #[ver(V >= V13_0B4)] + unk_2e1: Default::default(), + #[ver(V >= V13_0B4)] + unk_flag: U32(0), + #[ver(V >= V13_0B4)] + unk_pad: Default::default(), + } + )) + }, + )?; + + core::mem::drop(alloc); + + fence.add_command(); + comp_job.add_cb(comp, vm_bind.slot(), move |cmd, error| { + if let Some(err) = error { + fence.set_error(err.into()) + } + if let Some(mut rw) = result_writer { + let mut result: bindings::drm_asahi_result_compute = Default::default(); + + cmd.timestamps.with(|raw, _inner| { + result.ts_start = raw.start.load(Ordering::Relaxed); + result.ts_end = raw.end.load(Ordering::Relaxed); + }); + + if let Some(err) = error { + result.info = err.into(); + } else { + result.info.status = bindings::drm_asahi_status_DRM_ASAHI_STATUS_COMPLETE; + } + + rw.write(result); + } + + fence.command_complete(); + })?; + + notifier.threshold.with(|raw, _inner| { + raw.increment(); + }); + + comp_job.next_seq(); + + Ok(()) + } +} diff --git a/drivers/gpu/drm/asahi/queue/mod.rs b/drivers/gpu/drm/asahi/queue/mod.rs new file mode 100644 index 000000000000..15988af33cf3 --- /dev/null +++ b/drivers/gpu/drm/asahi/queue/mod.rs @@ -0,0 +1,725 @@ +// SPDX-License-Identifier: GPL-2.0-only OR MIT + +//! Submission queue management +//! +//! This module implements the userspace view of submission queues and the logic to map userspace +//! submissions to firmware queues. + +use kernel::dma_fence::*; +use kernel::prelude::*; +use kernel::{ + bindings, c_str, dma_fence, + drm::gem::shmem::VMap, + drm::sched, + macros::versions, + sync::{smutex::Mutex, Arc}, +}; + +use crate::alloc::Allocator; +use crate::debug::*; +use crate::driver::AsahiDevice; +use crate::fw::types::*; +use crate::gpu::GpuManager; +use crate::{alloc, buffer, channel, event, file, fw, gem, gpu, mmu, workqueue}; +use crate::{inner_weak_ptr, place}; + +use core::mem::MaybeUninit; +use core::sync::atomic::{AtomicU64, Ordering}; + +const DEBUG_CLASS: DebugFlags = DebugFlags::Queue; + +const WQ_SIZE: u32 = 0x500; + +mod common; +mod compute; +mod render; + +/// Trait implemented by all versioned queues. +pub(crate) trait Queue: Send + Sync { + fn submit( + &mut self, + id: u64, + in_syncs: Vec, + out_syncs: Vec, + result_buf: Option, + commands: Vec, + ) -> Result; +} + +#[versions(AGX)] +struct SubQueue { + wq: Arc, +} + +#[versions(AGX)] +impl SubQueue::ver { + fn new_job(&mut self) -> SubQueueJob::ver { + SubQueueJob::ver { + wq: self.wq.clone(), + job: None, + } + } +} + +#[versions(AGX)] +struct SubQueueJob { + wq: Arc, + job: Option, +} + +#[versions(AGX)] +impl SubQueueJob::ver { + fn get(&mut self) -> Result<&mut workqueue::Job::ver> { + if self.job.is_none() { + mod_pr_debug!("SubQueueJob: Creating {:?} job\n", self.wq.pipe_type()); + self.job.replace(self.wq.new_job()?); + } + Ok(self.job.as_mut().expect("expected a Job")) + } + + fn commit(&mut self) -> Result { + match self.job.as_mut() { + Some(job) => job.commit(), + None => Ok(()), + } + } + + fn can_submit(&self) -> bool { + match self.job.as_ref() { + None => true, + Some(job) => job.can_submit(), + } + } +} + +#[versions(AGX)] +pub(crate) struct Queue { + dev: AsahiDevice, + _sched: sched::Scheduler, + entity: sched::Entity, + vm: mmu::Vm, + ualloc: Arc>, + q_vtx: Option, + q_frag: Option, + q_comp: Option, + buffer: Option>, + gpu_context: Arc, + notifier_list: Arc>, + notifier: Arc>, + id: u64, + fence_ctx: FenceContexts, + #[ver(V >= V13_0B4)] + counter: AtomicU64, +} + +#[versions(AGX)] +#[derive(Default)] +pub(crate) struct JobFence { + id: u64, + pending: AtomicU64, +} + +#[versions(AGX)] +impl JobFence::ver { + fn add_command(self: &FenceObject) { + self.pending.fetch_add(1, Ordering::Relaxed); + } + + fn command_complete(self: &FenceObject) { + let remain = self.pending.fetch_sub(1, Ordering::Relaxed) - 1; + mod_pr_debug!( + "JobFence[{}]: Command complete (remain: {})\n", + self.id, + remain + ); + if remain == 0 { + mod_pr_debug!("JobFence[{}]: Signaling\n", self.id); + if self.signal().is_err() { + pr_err!("JobFence[{}]: Fence signal failed\n", self.id); + } + } + } +} + +#[versions(AGX)] +#[vtable] +impl dma_fence::FenceOps for JobFence::ver { + const USE_64BIT_SEQNO: bool = true; + + fn get_driver_name<'a>(self: &'a FenceObject) -> &'a CStr { + c_str!("asahi") + } + fn get_timeline_name<'a>(self: &'a FenceObject) -> &'a CStr { + c_str!("queue") + } +} + +#[versions(AGX)] +pub(crate) struct QueueJob { + dev: AsahiDevice, + vm_bind: mmu::VmBind, + op_guard: Option, + sj_vtx: Option, + sj_frag: Option, + sj_comp: Option, + fence: UserFence, + did_run: bool, + id: u64, +} + +#[versions(AGX)] +impl QueueJob::ver { + fn get_vtx(&mut self) -> Result<&mut workqueue::Job::ver> { + self.sj_vtx.as_mut().ok_or(EINVAL)?.get() + } + fn get_frag(&mut self) -> Result<&mut workqueue::Job::ver> { + self.sj_frag.as_mut().ok_or(EINVAL)?.get() + } + fn get_comp(&mut self) -> Result<&mut workqueue::Job::ver> { + self.sj_comp.as_mut().ok_or(EINVAL)?.get() + } + + fn commit(&mut self) -> Result { + mod_dev_dbg!(self.dev, "QueueJob: Committing\n"); + + self.sj_vtx.as_mut().map(|a| a.commit()).unwrap_or(Ok(()))?; + self.sj_frag + .as_mut() + .map(|a| a.commit()) + .unwrap_or(Ok(()))?; + self.sj_comp.as_mut().map(|a| a.commit()).unwrap_or(Ok(())) + } +} + +#[versions(AGX)] +impl sched::JobImpl for QueueJob::ver { + fn can_run(job: &mut sched::Job) -> bool { + mod_dev_dbg!(job.dev, "QueueJob {}: Checking runnability\n", job.id); + + if let Some(sj) = job.sj_vtx.as_ref() { + if !sj.can_submit() { + mod_dev_dbg!( + job.dev, + "QueueJob {}: Blocking due to vertex queue full\n", + job.id + ); + return false; + } + } + if let Some(sj) = job.sj_frag.as_ref() { + if !sj.can_submit() { + mod_dev_dbg!( + job.dev, + "QueueJob {}: Blocking due to fragment queue full\n", + job.id + ); + return false; + } + } + if let Some(sj) = job.sj_comp.as_ref() { + if !sj.can_submit() { + mod_dev_dbg!( + job.dev, + "QueueJob {}: Blocking due to compute queue full\n", + job.id + ); + return false; + } + } + true + } + + #[allow(unused_assignments)] + fn run(job: &mut sched::Job) -> Result> { + mod_dev_dbg!(job.dev, "QueueJob {}: Running Job\n", job.id); + + let dev = job.dev.data(); + let gpu = match dev + .gpu + .clone() + .arc_as_any() + .downcast::() + { + Ok(gpu) => gpu, + Err(_) => { + dev_crit!(job.dev, "GpuManager mismatched with QueueJob!\n"); + return Err(EIO); + } + }; + + if job.op_guard.is_none() { + job.op_guard = Some(gpu.start_op()?); + } + + // First submit all the commands for each queue. This can fail. + + let mut frag_job = None; + let mut frag_sub = None; + if let Some(sj) = job.sj_frag.as_mut() { + frag_job = sj.job.take(); + if let Some(wqjob) = frag_job.as_mut() { + mod_dev_dbg!(job.dev, "QueueJob {}: Submit fragment\n", job.id); + frag_sub = Some(wqjob.submit()?); + } + } + + let mut vtx_job = None; + let mut vtx_sub = None; + if let Some(sj) = job.sj_vtx.as_mut() { + vtx_job = sj.job.take(); + if let Some(wqjob) = vtx_job.as_mut() { + mod_dev_dbg!(job.dev, "QueueJob {}: Submit vertex\n", job.id); + vtx_sub = Some(wqjob.submit()?); + } + } + + let mut comp_job = None; + let mut comp_sub = None; + if let Some(sj) = job.sj_comp.as_mut() { + comp_job = sj.job.take(); + if let Some(wqjob) = comp_job.as_mut() { + mod_dev_dbg!(job.dev, "QueueJob {}: Submit compute\n", job.id); + comp_sub = Some(wqjob.submit()?); + } + } + + // Now we fully commit to running the job + mod_dev_dbg!(job.dev, "QueueJob {}: Run fragment\n", job.id); + frag_sub.map(|a| gpu.run_job(a)).transpose()?; + + mod_dev_dbg!(job.dev, "QueueJob {}: Run vertex\n", job.id); + vtx_sub.map(|a| gpu.run_job(a)).transpose()?; + + mod_dev_dbg!(job.dev, "QueueJob {}: Run compute\n", job.id); + comp_sub.map(|a| gpu.run_job(a)).transpose()?; + + mod_dev_dbg!(job.dev, "QueueJob {}: Drop compute job\n", job.id); + core::mem::drop(comp_job); + mod_dev_dbg!(job.dev, "QueueJob {}: Drop vertex job\n", job.id); + core::mem::drop(vtx_job); + mod_dev_dbg!(job.dev, "QueueJob {}: Drop fragment job\n", job.id); + core::mem::drop(frag_job); + + job.did_run = true; + + Ok(Some(Fence::from_fence(&job.fence))) + } + + fn timed_out(job: &mut sched::Job) -> sched::Status { + // FIXME: Handle timeouts properly + dev_err!( + job.dev, + "QueueJob {}: Job timed out on the DRM scheduler, things will probably break (ran: {})\n", + job.id, job.did_run + ); + sched::Status::NoDevice + } +} + +#[versions(AGX)] +impl Drop for QueueJob::ver { + fn drop(&mut self) { + mod_dev_dbg!(self.dev, "QueueJob {}: Dropping\n", self.id); + } +} + +struct ResultWriter { + vmap: VMap, + offset: usize, + len: usize, +} + +impl ResultWriter { + fn write(&mut self, mut value: T) { + let p: *mut u8 = &mut value as *mut _ as *mut u8; + // SAFETY: We know `p` points to a type T of that size, and UAPI types must have + // no padding and all bit patterns valid. + let slice = unsafe { core::slice::from_raw_parts_mut(p, core::mem::size_of::()) }; + let len = slice.len().min(self.len); + self.vmap.as_mut_slice()[self.offset..self.offset + len].copy_from_slice(&slice[..len]); + } +} + +static QUEUE_NAME: &CStr = c_str!("asahi_fence"); +static QUEUE_CLASS_KEY: kernel::sync::LockClassKey = kernel::sync::LockClassKey::new(); + +#[versions(AGX)] +impl Queue::ver { + /// Create a new user queue. + #[allow(clippy::too_many_arguments)] + pub(crate) fn new( + dev: &AsahiDevice, + vm: mmu::Vm, + alloc: &mut gpu::KernelAllocators, + ualloc: Arc>, + ualloc_priv: Arc>, + event_manager: Arc, + mgr: &buffer::BufferManager, + id: u64, + priority: u32, + caps: u32, + ) -> Result { + mod_dev_dbg!(dev, "[Queue {}] Creating queue\n", id); + + let data = dev.data(); + + let mut notifier_list = alloc.private.new_default::()?; + + let self_ptr = notifier_list.weak_pointer(); + notifier_list.with_mut(|raw, _inner| { + raw.list_head.next = Some(inner_weak_ptr!(self_ptr, list_head)); + }); + + let threshold = alloc.shared.new_default::()?; + + let notifier: Arc> = + Arc::try_new(alloc.private.new_inplace( + fw::event::Notifier::ver { threshold }, + |inner, ptr: &mut MaybeUninit>| { + Ok(place!( + ptr, + fw::event::raw::Notifier::ver { + threshold: inner.threshold.gpu_pointer(), + generation: AtomicU32::new(id as u32), + cur_count: AtomicU32::new(0), + unk_10: AtomicU32::new(0x50), + state: Default::default() + } + )) + }, + )?)?; + + let sched = sched::Scheduler::new(dev, WQ_SIZE, 0, 100000, c_str!("asahi_sched"))?; + // Priorities are handled by the AGX scheduler, there is no meaning within a + // per-queue scheduler. + let entity = sched::Entity::new(&sched, sched::Priority::Normal)?; + + let mut ret = Queue::ver { + dev: dev.clone(), + _sched: sched, + entity, + vm, + ualloc, + q_vtx: None, + q_frag: None, + q_comp: None, + buffer: None, + gpu_context: Arc::try_new(workqueue::GpuContext::new(dev, alloc)?)?, + notifier_list: Arc::try_new(notifier_list)?, + notifier, + id, + fence_ctx: FenceContexts::new(1, QUEUE_NAME, &QUEUE_CLASS_KEY)?, + #[ver(V >= V13_0B4)] + counter: AtomicU64::new(0), + }; + + // Rendering structures + if caps & bindings::drm_asahi_queue_cap_DRM_ASAHI_QUEUE_CAP_RENDER != 0 { + let buffer = + buffer::Buffer::ver::new(&*data.gpu, alloc, ret.ualloc.clone(), ualloc_priv, mgr)?; + let tvb_blocks = { + let lock = crate::THIS_MODULE.kernel_param_lock(); + *crate::initial_tvb_size.read(&lock) + }; + + buffer.ensure_blocks(tvb_blocks)?; + + ret.buffer = Some(Mutex::new(buffer)); + ret.q_vtx = Some(SubQueue::ver { + wq: workqueue::WorkQueue::ver::new( + alloc, + event_manager.clone(), + ret.gpu_context.clone(), + ret.notifier_list.clone(), + channel::PipeType::Vertex, + id, + priority, + WQ_SIZE, + )?, + }); + } + + // Rendering & blit structures + if caps + & (bindings::drm_asahi_queue_cap_DRM_ASAHI_QUEUE_CAP_RENDER + | bindings::drm_asahi_queue_cap_DRM_ASAHI_QUEUE_CAP_BLIT) + != 0 + { + ret.q_frag = Some(SubQueue::ver { + wq: workqueue::WorkQueue::ver::new( + alloc, + event_manager.clone(), + ret.gpu_context.clone(), + ret.notifier_list.clone(), + channel::PipeType::Fragment, + id, + priority, + WQ_SIZE, + )?, + }); + } + + // Compute structures + if caps & bindings::drm_asahi_queue_cap_DRM_ASAHI_QUEUE_CAP_COMPUTE != 0 { + ret.q_comp = Some(SubQueue::ver { + wq: workqueue::WorkQueue::ver::new( + alloc, + event_manager, + ret.gpu_context.clone(), + ret.notifier_list.clone(), + channel::PipeType::Compute, + id, + priority, + WQ_SIZE, + )?, + }); + } + + mod_dev_dbg!(dev, "[Queue {}] Queue created\n", id); + Ok(ret) + } +} + +const SQ_RENDER: usize = bindings::drm_asahi_subqueue_DRM_ASAHI_SUBQUEUE_RENDER as usize; +const SQ_COMPUTE: usize = bindings::drm_asahi_subqueue_DRM_ASAHI_SUBQUEUE_COMPUTE as usize; +const SQ_COUNT: usize = bindings::drm_asahi_subqueue_DRM_ASAHI_SUBQUEUE_COUNT as usize; + +#[versions(AGX)] +impl Queue for Queue::ver { + fn submit( + &mut self, + id: u64, + in_syncs: Vec, + out_syncs: Vec, + result_buf: Option, + commands: Vec, + ) -> Result { + let dev = self.dev.data(); + let gpu = match dev + .gpu + .clone() + .arc_as_any() + .downcast::() + { + Ok(gpu) => gpu, + Err(_) => { + dev_crit!(self.dev, "GpuManager mismatched with JobImpl!\n"); + return Err(EIO); + } + }; + + mod_dev_dbg!(self.dev, "[Submission {}] Submit job\n", id); + + if gpu.is_crashed() { + dev_err!( + self.dev, + "[Submission {}] GPU is crashed, cannot submit\n", + id + ); + return Err(ENODEV); + } + + // Empty submissions are not legal + if commands.is_empty() { + return Err(EINVAL); + } + + let op_guard = if !in_syncs.is_empty() { + Some(gpu.start_op()?) + } else { + None + }; + + let mut events: [Vec>; SQ_COUNT] = + Default::default(); + + events[SQ_RENDER].try_push(self.q_frag.as_ref().and_then(|a| a.wq.event_info()))?; + events[SQ_COMPUTE].try_push(self.q_comp.as_ref().and_then(|a| a.wq.event_info()))?; + + let vm_bind = gpu.bind_vm(&self.vm)?; + let vm_slot = vm_bind.slot(); + + mod_dev_dbg!(self.dev, "[Submission {}] Creating job\n", id); + let mut job = self.entity.new_job(QueueJob::ver { + dev: self.dev.clone(), + vm_bind, + op_guard, + sj_vtx: self.q_vtx.as_mut().map(|a| a.new_job()), + sj_frag: self.q_frag.as_mut().map(|a| a.new_job()), + sj_comp: self.q_comp.as_mut().map(|a| a.new_job()), + fence: self + .fence_ctx + .new_fence::( + 0, + JobFence::ver { + id, + pending: Default::default(), + }, + )? + .into(), + did_run: false, + id, + })?; + + mod_dev_dbg!( + self.dev, + "[Submission {}] Adding {} in_syncs\n", + id, + in_syncs.len() + ); + for sync in in_syncs { + job.add_dependency(sync.fence.expect("in_sync missing fence"))?; + } + + let mut last_render = None; + let mut last_compute = None; + + for (i, cmd) in commands.iter().enumerate() { + match cmd.cmd_type { + bindings::drm_asahi_cmd_type_DRM_ASAHI_CMD_RENDER => last_render = Some(i), + bindings::drm_asahi_cmd_type_DRM_ASAHI_CMD_COMPUTE => last_compute = Some(i), + _ => return Err(EINVAL), + } + } + + mod_dev_dbg!( + self.dev, + "[Submission {}] Submitting {} commands\n", + id, + commands.len() + ); + for (i, cmd) in commands.into_iter().enumerate() { + for (queue_idx, index) in cmd.barriers.iter().enumerate() { + if *index == bindings::DRM_ASAHI_BARRIER_NONE as u32 { + continue; + } + if let Some(event) = events[queue_idx].get(*index as usize).ok_or(EINVAL)? { + let mut alloc = gpu.alloc(); + let queue_job = match cmd.cmd_type { + bindings::drm_asahi_cmd_type_DRM_ASAHI_CMD_RENDER => job.get_vtx()?, + bindings::drm_asahi_cmd_type_DRM_ASAHI_CMD_COMPUTE => job.get_comp()?, + _ => return Err(EINVAL), + }; + mod_dev_dbg!(self.dev, "[Submission {}] Create Explicit Barrier\n", id); + let barrier: GpuObject = alloc.private.new_inplace( + Default::default(), + |_inner, ptr: &mut MaybeUninit| { + Ok(place!( + ptr, + fw::workqueue::raw::Barrier { + tag: fw::workqueue::CommandType::Barrier, + wait_stamp: event.fw_stamp_pointer, + wait_value: event.value, + wait_slot: event.slot, + stamp_self: queue_job.event_info().value.next(), + uuid: 0xffffbbbb, + unk: 0, + } + )) + }, + )?; + mod_dev_dbg!(self.dev, "[Submission {}] Add Explicit Barrier\n", id); + queue_job.add(barrier, vm_slot)?; + } else { + assert!(*index == 0); + } + } + + let result_writer = match result_buf.as_ref() { + None => { + if cmd.result_offset != 0 || cmd.result_size != 0 { + return Err(EINVAL); + } + None + } + Some(buf) => { + if cmd.result_size != 0 { + if cmd + .result_offset + .checked_add(cmd.result_size) + .ok_or(EINVAL)? + > buf.size() as u64 + { + return Err(EINVAL); + } + Some(ResultWriter { + vmap: buf.gem.vmap()?, + offset: cmd.result_offset.try_into()?, + len: cmd.result_size.try_into()?, + }) + } else { + None + } + } + }; + + match cmd.cmd_type { + bindings::drm_asahi_cmd_type_DRM_ASAHI_CMD_RENDER => { + self.submit_render( + &mut job, + &cmd, + result_writer, + id, + last_render.unwrap() == i, + )?; + events[SQ_RENDER].try_push(Some( + job.sj_frag + .as_ref() + .expect("No frag queue?") + .job + .as_ref() + .expect("No frag job?") + .event_info(), + ))?; + } + bindings::drm_asahi_cmd_type_DRM_ASAHI_CMD_COMPUTE => { + self.submit_compute( + &mut job, + &cmd, + result_writer, + id, + last_compute.unwrap() == i, + )?; + events[SQ_COMPUTE].try_push(Some( + job.sj_comp + .as_ref() + .expect("No comp queue?") + .job + .as_ref() + .expect("No comp job?") + .event_info(), + ))?; + } + _ => return Err(EINVAL), + } + } + + mod_dev_dbg!(self.dev, "Queue: Committing job\n"); + job.commit()?; + + mod_dev_dbg!(self.dev, "Queue: Arming job\n"); + let job = job.arm(); + let out_fence = job.fences().finished(); + mod_dev_dbg!(self.dev, "Queue: Pushing job\n"); + job.push(); + + mod_dev_dbg!(self.dev, "Queue: Adding {} out_syncs\n", out_syncs.len()); + for mut sync in out_syncs { + if let Some(chain) = sync.chain_fence.take() { + sync.syncobj + .add_point(chain, &out_fence, sync.timeline_value); + } else { + sync.syncobj.replace_fence(Some(&out_fence)); + } + } + + Ok(()) + } +} + +#[versions(AGX)] +impl Drop for Queue::ver { + fn drop(&mut self) { + mod_dev_dbg!(self.dev, "[Queue {}] Dropping queue\n", self.id); + } +} diff --git a/drivers/gpu/drm/asahi/queue/render.rs b/drivers/gpu/drm/asahi/queue/render.rs new file mode 100644 index 000000000000..318c952df020 --- /dev/null +++ b/drivers/gpu/drm/asahi/queue/render.rs @@ -0,0 +1,1173 @@ +// SPDX-License-Identifier: GPL-2.0-only OR MIT +#![allow(clippy::unusual_byte_groupings)] + +//! Render work queue. +//! +//! A render queue consists of two underlying WorkQueues, one for vertex and one for fragment work. +//! This module is in charge of creating all of the firmware structures required to submit 3D +//! rendering work to the GPU, based on the userspace command buffer. + +use super::common; +use crate::alloc::Allocator; +use crate::debug::*; +use crate::fw::types::*; +use crate::gpu::GpuManager; +use crate::util::*; +use crate::workqueue::WorkError; +use crate::{box_in_place, inner_ptr, inner_weak_ptr, place}; +use crate::{buffer, fw, gpu, microseq, workqueue}; +use core::mem::MaybeUninit; +use core::sync::atomic::Ordering; +use kernel::bindings; +use kernel::dma_fence::RawDmaFence; +use kernel::drm::sched::Job; +use kernel::io_buffer::IoBufferReader; +use kernel::prelude::*; +use kernel::sync::{smutex::Mutex, Arc}; +use kernel::user_ptr::UserSlicePtr; + +const DEBUG_CLASS: DebugFlags = DebugFlags::Render; + +/// Tiling/Vertex control bit to disable using more than one GPU cluster. This results in decreased +/// throughput but also less latency, which is probably desirable for light vertex loads where the +/// overhead of clustering/merging would exceed the time it takes to just run the job on one +/// cluster. +const TILECTL_DISABLE_CLUSTERING: u32 = 1u32 << 0; + +struct RenderResult { + result: bindings::drm_asahi_result_render, + vtx_complete: bool, + frag_complete: bool, + vtx_error: Option, + frag_error: Option, + writer: super::ResultWriter, +} + +impl RenderResult { + fn commit(&mut self) { + if !self.vtx_complete || !self.frag_complete { + return; + } + + let mut error = self.vtx_error.take(); + if let Some(frag_error) = self.frag_error.take() { + if error.is_none() || error == Some(WorkError::Killed) { + error = Some(frag_error); + } + } + + if let Some(err) = error { + self.result.info = err.into(); + } else { + self.result.info.status = bindings::drm_asahi_status_DRM_ASAHI_STATUS_COMPLETE; + } + + self.writer.write(self.result); + } +} + +#[versions(AGX)] +impl super::Queue::ver { + /// Get the appropriate tiling parameters for a given userspace command buffer. + fn get_tiling_params( + cmdbuf: &bindings::drm_asahi_cmd_render, + num_clusters: u32, + ) -> Result { + let width: u32 = cmdbuf.fb_width; + let height: u32 = cmdbuf.fb_height; + let layers: u32 = cmdbuf.layers; + + if width > 65536 || height > 65536 { + return Err(EINVAL); + } + + if layers == 0 || layers > 2048 { + return Err(EINVAL); + } + + let tile_width = 32u32; + let tile_height = 32u32; + + let utile_width = cmdbuf.utile_width; + let utile_height = cmdbuf.utile_height; + + match (utile_width, utile_height) { + (32, 32) | (32, 16) | (16, 16) => (), + _ => return Err(EINVAL), + }; + + let utiles_per_tile_x = tile_width / utile_width; + let utiles_per_tile_y = tile_height / utile_height; + + let utiles_per_tile = utiles_per_tile_x * utiles_per_tile_y; + + let tiles_x = (width + tile_width - 1) / tile_width; + let tiles_y = (height + tile_height - 1) / tile_height; + let tiles = tiles_x * tiles_y; + + let mtiles_x = 4u32; + let mtiles_y = 4u32; + let mtiles = mtiles_x * mtiles_y; + + // TODO: *samples + let tiles_per_mtile_x = align(div_ceil(tiles_x, mtiles_x), 4); + let tiles_per_mtile_y = align(div_ceil(tiles_y, mtiles_y), 4); + let tiles_per_mtile = tiles_per_mtile_x * tiles_per_mtile_y; + + let mtile_x1 = tiles_per_mtile_x; + let mtile_x2 = 2 * tiles_per_mtile_x; + let mtile_x3 = 3 * tiles_per_mtile_x; + + let mtile_y1 = tiles_per_mtile_y; + let mtile_y2 = 2 * tiles_per_mtile_y; + let mtile_y3 = 3 * tiles_per_mtile_y; + + let rgn_entry_size = 5; + // Macrotile stride in 32-bit words + let rgn_size = align(rgn_entry_size * tiles_per_mtile * utiles_per_tile, 4) / 4; + let tilemap_size = (4 * rgn_size * mtiles * layers) as usize; + + let tpc_entry_size = 8; + // TPC stride in 32-bit words + let tpc_mtile_stride = tpc_entry_size * utiles_per_tile * tiles_per_mtile / 4; + let tpc_size = (num_clusters * (4 * tpc_mtile_stride * mtiles) * layers) as usize; + + // No idea where this comes from, but it fits what macOS does... + // TODO: layers? + let meta1_blocks = if num_clusters > 1 { + div_ceil(align(tiles_x, 2) * align(tiles_y, 4), 0x1980) + } else { + 0 + }; + + let min_tvb_blocks = + div_ceil(tiles_x * tiles_y, 128).max(if num_clusters > 1 { 9 } else { 8 }) as usize; + + // Sometimes clustering seems to use twice the cluster tilemap count + // and twice the meta4 size. TODO: Is this random or can we calculate + // it somehow??? Does it go higher??? + let cluster_factor = 2; + + Ok(buffer::TileInfo { + tiles_x, + tiles_y, + tiles, + utile_width, + utile_height, + //mtiles_x, + //mtiles_y, + tiles_per_mtile_x, + tiles_per_mtile_y, + //tiles_per_mtile, + utiles_per_mtile_x: tiles_per_mtile_x * utiles_per_tile_x, + utiles_per_mtile_y: tiles_per_mtile_y * utiles_per_tile_y, + //utiles_per_mtile: tiles_per_mtile * utiles_per_tile, + tilemap_size, + tpc_size, + meta1_blocks, + min_tvb_blocks, + cluster_factor, + params: fw::vertex::raw::TilingParameters { + rgn_size, + unk_4: 0x88, + ppp_ctrl: cmdbuf.ppp_ctrl, + x_max: (width - 1) as u16, + y_max: (height - 1) as u16, + te_screen: ((tiles_y - 1) << 12) | (tiles_x - 1), + te_mtile1: mtile_x3 | (mtile_x2 << 9) | (mtile_x1 << 18), + te_mtile2: mtile_y3 | (mtile_y2 << 9) | (mtile_y1 << 18), + tiles_per_mtile, + tpc_stride: tpc_mtile_stride, + unk_24: 0x100, + unk_28: if layers > 1 { + 0xe000 | (layers - 1) + } else { + 0x8000 + }, + }, + }) + } + + /// Submit work to a render queue. + pub(super) fn submit_render( + &self, + job: &mut Job, + cmd: &bindings::drm_asahi_command, + result_writer: Option, + id: u64, + flush_stamps: bool, + ) -> Result { + if cmd.cmd_type != bindings::drm_asahi_cmd_type_DRM_ASAHI_CMD_RENDER { + return Err(EINVAL); + } + + mod_dev_dbg!(self.dev, "[Submission {}] Render!\n", id); + + let mut cmdbuf_reader = unsafe { + UserSlicePtr::new( + cmd.cmd_buffer as usize as *mut _, + core::mem::size_of::(), + ) + .reader() + }; + + let mut cmdbuf: MaybeUninit = MaybeUninit::uninit(); + unsafe { + cmdbuf_reader.read_raw( + cmdbuf.as_mut_ptr() as *mut u8, + core::mem::size_of::(), + )?; + } + let cmdbuf = unsafe { cmdbuf.assume_init() }; + + if cmdbuf.flags + & !(bindings::ASAHI_RENDER_NO_CLEAR_PIPELINE_TEXTURES + | bindings::ASAHI_RENDER_SET_WHEN_RELOADING_Z_OR_S + | bindings::ASAHI_RENDER_MEMORYLESS_RTS_USED + | bindings::ASAHI_RENDER_PROCESS_EMPTY_TILES + | bindings::ASAHI_RENDER_NO_VERTEX_CLUSTERING) as u64 + != 0 + { + return Err(EINVAL); + } + + if cmdbuf.flags & bindings::ASAHI_RENDER_MEMORYLESS_RTS_USED as u64 != 0 { + // Not supported yet + return Err(EINVAL); + } + + if cmdbuf.fb_width == 0 + || cmdbuf.fb_height == 0 + || cmdbuf.fb_width > 16384 + || cmdbuf.fb_height > 16384 + { + mod_dev_dbg!( + self.dev, + "[Submission {}] Invalid dimensions {}x{}\n", + id, + cmdbuf.fb_width, + cmdbuf.fb_height + ); + return Err(EINVAL); + } + + let dev = self.dev.data(); + let gpu = match dev.gpu.as_any().downcast_ref::() { + Some(gpu) => gpu, + None => { + dev_crit!(self.dev, "GpuManager mismatched with Queue!\n"); + return Err(EIO); + } + }; + + let nclusters = gpu.get_dyncfg().id.num_clusters; + + // Can be set to false to disable clustering (for simpler jobs), but then the + // core masks below should be adjusted to cover a single rolling cluster. + let mut clustering = nclusters > 1; + + if debug_enabled(debug::DebugFlags::DisableClustering) + || cmdbuf.flags & bindings::ASAHI_RENDER_NO_VERTEX_CLUSTERING as u64 != 0 + { + clustering = false; + } + + #[ver(G < G14)] + let tiling_control = { + let render_cfg = gpu.get_cfg().render; + let mut tiling_control = render_cfg.tiling_control; + + if !clustering { + tiling_control |= TILECTL_DISABLE_CLUSTERING; + } + tiling_control + }; + + let mut alloc = gpu.alloc(); + let kalloc = &mut *alloc; + + // This sequence number increases per new client/VM? assigned to some slot, + // but it's unclear *which* slot... + let slot_client_seq: u8 = (self.id & 0xff) as u8; + + let tile_info = Self::get_tiling_params(&cmdbuf, if clustering { nclusters } else { 1 })?; + + let buffer = self.buffer.as_ref().ok_or(EINVAL)?.lock(); + + let scene = Arc::try_new(buffer.new_scene(kalloc, &tile_info)?)?; + + let notifier = self.notifier.clone(); + + let tvb_autogrown = buffer.auto_grow()?; + if tvb_autogrown { + let new_size = buffer.block_count() as usize; + cls_dev_dbg!( + TVBStats, + &self.dev, + "[Submission {}] TVB grew to {} bytes ({} blocks) due to overflows\n", + id, + new_size * buffer::BLOCK_SIZE, + new_size, + ); + } + + let tvb_grown = buffer.ensure_blocks(tile_info.min_tvb_blocks)?; + if tvb_grown { + cls_dev_dbg!( + TVBStats, + &self.dev, + "[Submission {}] TVB grew to {} bytes ({} blocks) due to dimensions ({}x{})\n", + id, + tile_info.min_tvb_blocks * buffer::BLOCK_SIZE, + tile_info.min_tvb_blocks, + cmdbuf.fb_width, + cmdbuf.fb_height + ); + } + + let vm_bind = job.vm_bind.clone(); + + mod_dev_dbg!( + self.dev, + "[Submission {}] VM slot = {}\n", + id, + vm_bind.slot() + ); + + let ev_vtx = job.get_vtx()?.event_info(); + let ev_frag = job.get_frag()?.event_info(); + + mod_dev_dbg!( + self.dev, + "[Submission {}] Vert event #{} -> {:#x?}\n", + id, + ev_vtx.slot, + ev_vtx.value.next(), + ); + mod_dev_dbg!( + self.dev, + "[Submission {}] Frag event #{} -> {:#x?}\n", + id, + ev_frag.slot, + ev_frag.value.next(), + ); + + let uuid_3d = cmdbuf.cmd_3d_id; + let uuid_ta = cmdbuf.cmd_ta_id; + + mod_dev_dbg!( + self.dev, + "[Submission {}] Vert UUID = {:#x?}\n", + id, + uuid_ta + ); + mod_dev_dbg!( + self.dev, + "[Submission {}] Frag UUID = {:#x?}\n", + id, + uuid_3d + ); + + let fence = job.fence.clone(); + let frag_job = job.get_frag()?; + + mod_dev_dbg!(self.dev, "[Submission {}] Create Barrier\n", id); + let barrier: GpuObject = kalloc.private.new_inplace( + Default::default(), + |_inner, ptr: &mut MaybeUninit| { + Ok(place!( + ptr, + fw::workqueue::raw::Barrier { + tag: fw::workqueue::CommandType::Barrier, + wait_stamp: ev_vtx.fw_stamp_pointer, + wait_value: ev_vtx.value.next(), + wait_slot: ev_vtx.slot, + stamp_self: ev_frag.value.next(), + uuid: uuid_3d, + unk: 0, + } + )) + }, + )?; + + mod_dev_dbg!(self.dev, "[Submission {}] Add Barrier\n", id); + frag_job.add(barrier, vm_bind.slot())?; + + let timestamps = Arc::try_new(kalloc.shared.new_default::()?)?; + + let unk1 = debug_enabled(debug::DebugFlags::Debug1); + let unk2 = debug_enabled(debug::DebugFlags::Debug2); + let unk3 = debug_enabled(debug::DebugFlags::Debug3); + + let mut tile_config: u64 = 0; + if !unk1 { + tile_config |= 0x280; + } + if cmdbuf.layers > 1 { + tile_config |= 1; + } + if cmdbuf.flags & bindings::ASAHI_RENDER_PROCESS_EMPTY_TILES as u64 != 0 { + tile_config |= 0x10000; + } + + let mut utile_config = + ((tile_info.utile_width / 16) << 12) | ((tile_info.utile_height / 16) << 14); + utile_config |= match cmdbuf.samples { + 1 => 0, + 2 => 1, + 4 => 2, + _ => return Err(EINVAL), + }; + + let frag_result = result_writer + .map(|writer| { + let mut result = RenderResult { + result: Default::default(), + vtx_complete: false, + frag_complete: false, + vtx_error: None, + frag_error: None, + writer, + }; + + if tvb_autogrown { + result.result.flags |= bindings::DRM_ASAHI_RESULT_RENDER_TVB_GROW_OVF as u64; + } + if tvb_grown { + result.result.flags |= bindings::DRM_ASAHI_RESULT_RENDER_TVB_GROW_MIN as u64; + } + result.result.tvb_size_bytes = buffer.size() as u64; + + Arc::try_new(Mutex::new(result)) + }) + .transpose()?; + + let vtx_result = frag_result.clone(); + + // TODO: check + #[ver(V >= V13_0B4)] + let count_frag = self.counter.fetch_add(2, Ordering::Relaxed); + #[ver(V >= V13_0B4)] + let count_vtx = count_frag + 1; + + mod_dev_dbg!(self.dev, "[Submission {}] Create Frag\n", id); + let frag = GpuObject::new_prealloc( + kalloc.private.alloc_object()?, + |ptr: GpuWeakPointer| { + let mut builder = microseq::Builder::new(); + + let stats = inner_weak_ptr!( + gpu.initdata.runtime_pointers.stats.frag.weak_pointer(), + stats + ); + + let start_frag = builder.add(microseq::StartFragment::ver { + header: microseq::op::StartFragment::HEADER, + job_params2: inner_weak_ptr!(ptr, job_params2), + job_params1: inner_weak_ptr!(ptr, job_params1), + scene: scene.gpu_pointer(), + stats, + busy_flag: inner_weak_ptr!(ptr, busy_flag), + tvb_overflow_count: inner_weak_ptr!(ptr, tvb_overflow_count), + unk_pointer: inner_weak_ptr!(ptr, unk_pointee), + work_queue: ev_frag.info_ptr, + work_item: ptr, + vm_slot: vm_bind.slot(), + unk_50: 0x1, // fixed + event_generation: self.id as u32, + buffer_slot: scene.slot(), + unk_5c: 0, + cmd_seq: U64(ev_frag.cmd_seq), + unk_68: 0, + unk_758_flag: inner_weak_ptr!(ptr, unk_758_flag), + unk_job_buf: inner_weak_ptr!(ptr, unk_buf_0), + unk_7c: 0, + unk_80: 0, + unk_84: 0, + uuid: uuid_3d, + attachments: common::build_attachments( + cmdbuf.attachments, + cmdbuf.attachment_count, + )?, + unk_190: 0, + #[ver(V >= V13_0B4)] + counter: U64(count_frag), + #[ver(V >= V13_0B4)] + notifier_buf: inner_weak_ptr!(notifier.weak_pointer(), state.unk_buf), + })?; + + if frag_result.is_some() { + builder.add(microseq::Timestamp::ver { + header: microseq::op::Timestamp::new(true), + cur_ts: inner_weak_ptr!(ptr, cur_ts), + start_ts: inner_weak_ptr!(ptr, start_ts), + update_ts: inner_weak_ptr!(ptr, start_ts), + work_queue: ev_frag.info_ptr, + unk_24: U64(0), + #[ver(V >= V13_0B4)] + unk_ts: inner_weak_ptr!(ptr, unk_ts), + uuid: uuid_3d, + unk_30_padding: 0, + })?; + } + + builder.add(microseq::WaitForIdle { + header: microseq::op::WaitForIdle::new(microseq::Pipe::Fragment), + })?; + + if frag_result.is_some() { + builder.add(microseq::Timestamp::ver { + header: microseq::op::Timestamp::new(false), + cur_ts: inner_weak_ptr!(ptr, cur_ts), + start_ts: inner_weak_ptr!(ptr, start_ts), + update_ts: inner_weak_ptr!(ptr, end_ts), + work_queue: ev_frag.info_ptr, + unk_24: U64(0), + #[ver(V >= V13_0B4)] + unk_ts: inner_weak_ptr!(ptr, unk_ts), + uuid: uuid_3d, + unk_30_padding: 0, + })?; + } + + let off = builder.offset_to(start_frag); + builder.add(microseq::FinalizeFragment::ver { + header: microseq::op::FinalizeFragment::HEADER, + uuid: uuid_3d, + unk_8: 0, + fw_stamp: ev_frag.fw_stamp_pointer, + stamp_value: ev_frag.value.next(), + unk_18: 0, + scene: scene.weak_pointer(), + buffer: scene.weak_buffer_pointer(), + unk_2c: U64(1), + stats, + unk_pointer: inner_weak_ptr!(ptr, unk_pointee), + busy_flag: inner_weak_ptr!(ptr, busy_flag), + work_queue: ev_frag.info_ptr, + work_item: ptr, + vm_slot: vm_bind.slot(), + unk_60: 0, + unk_758_flag: inner_weak_ptr!(ptr, unk_758_flag), + unk_6c: U64(0), + unk_74: U64(0), + unk_7c: U64(0), + unk_84: U64(0), + unk_8c: U64(0), + #[ver(G == G14 && V < V13_0B4)] + unk_8c_g14: U64(0), + restart_branch_offset: off, + unk_98: unk3.into(), + #[ver(V >= V13_0B4)] + unk_9c: Default::default(), + })?; + + builder.add(microseq::RetireStamp { + header: microseq::op::RetireStamp::HEADER, + })?; + + Ok(box_in_place!(fw::fragment::RunFragment::ver { + notifier: notifier.clone(), + scene: scene.clone(), + micro_seq: builder.build(&mut kalloc.private)?, + vm_bind: vm_bind.clone(), + aux_fb: self.ualloc.lock().array_empty(0x8000)?, + timestamps: timestamps.clone(), + })?) + }, + |inner, ptr| { + let aux_fb_info = fw::fragment::raw::AuxFBInfo::ver { + iogpu_unk_214: cmdbuf.iogpu_unk_214, + unk2: 0, + width: cmdbuf.fb_width, + height: cmdbuf.fb_height, + #[ver(V >= V13_0B4)] + unk3: U64(0x100000), + }; + + Ok(place!( + ptr, + fw::fragment::raw::RunFragment::ver { + tag: fw::workqueue::CommandType::RunFragment, + #[ver(V >= V13_0B4)] + counter: U64(count_frag), + vm_slot: vm_bind.slot(), + unk_8: 0, + microsequence: inner.micro_seq.gpu_pointer(), + microsequence_size: inner.micro_seq.len() as u32, + notifier: inner.notifier.gpu_pointer(), + buffer: inner.scene.buffer_pointer(), + scene: inner.scene.gpu_pointer(), + unk_buffer_buf: inner.scene.kernel_buffer_pointer(), + tvb_tilemap: inner.scene.tvb_tilemap_pointer(), + ppp_multisamplectl: U64(cmdbuf.ppp_multisamplectl), + samples: cmdbuf.samples, + tiles_per_mtile_y: tile_info.tiles_per_mtile_y as u16, + tiles_per_mtile_x: tile_info.tiles_per_mtile_x as u16, + unk_50: U64(0), + unk_58: U64(0), + merge_upper_x: F32::from_bits(cmdbuf.merge_upper_x), + merge_upper_y: F32::from_bits(cmdbuf.merge_upper_y), + unk_68: U64(0), + tile_count: U64(tile_info.tiles as u64), + job_params1: fw::fragment::raw::JobParameters1::ver { + utile_config: utile_config, + unk_4: 0, + clear_pipeline: fw::fragment::raw::ClearPipelineBinding { + pipeline_bind: U64(cmdbuf.load_pipeline_bind as u64), + address: U64(cmdbuf.load_pipeline as u64), + }, + ppp_multisamplectl: U64(cmdbuf.ppp_multisamplectl), + scissor_array: U64(cmdbuf.scissor_array), + depth_bias_array: U64(cmdbuf.depth_bias_array), + aux_fb_info: aux_fb_info, + depth_dimensions: U64(cmdbuf.depth_dimensions as u64), + visibility_result_buffer: U64(cmdbuf.visibility_result_buffer), + zls_ctrl: U64(cmdbuf.zls_ctrl), + #[ver(G >= G14)] + unk_58_g14_0: U64(0x4040404), + #[ver(G >= G14)] + unk_58_g14_8: U64(0), + depth_buffer_ptr1: U64(cmdbuf.depth_buffer_1), + depth_buffer_ptr2: U64(cmdbuf.depth_buffer_2), + stencil_buffer_ptr1: U64(cmdbuf.stencil_buffer_1), + stencil_buffer_ptr2: U64(cmdbuf.stencil_buffer_2), + #[ver(G >= G14)] + unk_68_g14_0: Default::default(), + unk_78: Default::default(), + depth_meta_buffer_ptr1: U64(cmdbuf.depth_meta_buffer_1), + unk_a0: Default::default(), + depth_meta_buffer_ptr2: U64(cmdbuf.depth_meta_buffer_2), + unk_b0: Default::default(), + stencil_meta_buffer_ptr1: U64(cmdbuf.stencil_meta_buffer_1), + unk_c0: Default::default(), + stencil_meta_buffer_ptr2: U64(cmdbuf.stencil_meta_buffer_2), + unk_d0: Default::default(), + tvb_tilemap: inner.scene.tvb_tilemap_pointer(), + tvb_heapmeta: inner.scene.tvb_heapmeta_pointer(), + mtile_stride_dwords: U64((4 * tile_info.params.rgn_size as u64) << 24), + tvb_heapmeta_2: inner.scene.tvb_heapmeta_pointer(), + tile_config: U64(tile_config), + aux_fb: inner.aux_fb.gpu_pointer(), + unk_108: Default::default(), + pipeline_base: U64(0x11_00000000), + unk_140: U64(0x8c60), + unk_148: U64(0x0), + unk_150: U64(0x0), + unk_158: U64(0x1c), + unk_160: U64(0), + unk_168_padding: Default::default(), + #[ver(V < V13_0B4)] + __pad0: Default::default(), + }, + job_params2: fw::fragment::raw::JobParameters2 { + store_pipeline_bind: cmdbuf.store_pipeline_bind, + store_pipeline_addr: cmdbuf.store_pipeline, + unk_8: 0x0, + unk_c: 0x0, + merge_upper_x: F32::from_bits(cmdbuf.merge_upper_x), + merge_upper_y: F32::from_bits(cmdbuf.merge_upper_y), + unk_18: U64(0x0), + utiles_per_mtile_y: tile_info.utiles_per_mtile_y as u16, + utiles_per_mtile_x: tile_info.utiles_per_mtile_x as u16, + unk_24: 0x0, + tile_counts: ((tile_info.tiles_y - 1) << 12) | (tile_info.tiles_x - 1), + iogpu_unk_212: cmdbuf.iogpu_unk_212, + isp_bgobjdepth: cmdbuf.isp_bgobjdepth, + // TODO: does this flag need to be exposed to userspace? + isp_bgobjvals: cmdbuf.isp_bgobjvals | 0x400, + unk_38: 0x0, + unk_3c: 0x1, + unk_40: 0, + }, + job_params3: fw::fragment::raw::JobParameters3::ver { + unk_44_padding: Default::default(), + depth_bias_array: fw::fragment::raw::ArrayAddr { + ptr: U64(cmdbuf.depth_bias_array), + unk_padding: U64(0), + }, + scissor_array: fw::fragment::raw::ArrayAddr { + ptr: U64(cmdbuf.scissor_array), + unk_padding: U64(0), + }, + visibility_result_buffer: U64(cmdbuf.visibility_result_buffer), + unk_118: U64(0x0), + unk_120: Default::default(), + unk_reload_pipeline: fw::fragment::raw::ClearPipelineBinding { + pipeline_bind: U64(cmdbuf.partial_reload_pipeline_bind as u64), + address: U64(cmdbuf.partial_reload_pipeline as u64), + }, + unk_258: U64(0), + unk_260: U64(0), + unk_268: U64(0), + unk_270: U64(0), + reload_pipeline: fw::fragment::raw::ClearPipelineBinding { + pipeline_bind: U64(cmdbuf.partial_reload_pipeline_bind as u64), + address: U64(cmdbuf.partial_reload_pipeline as u64), + }, + zls_ctrl: U64(cmdbuf.zls_ctrl), + unk_290: U64(0x0), + depth_buffer_ptr1: U64(cmdbuf.depth_buffer_1), + unk_2a0: U64(0x0), + unk_2a8: U64(0x0), + depth_buffer_ptr2: U64(cmdbuf.depth_buffer_2), + depth_buffer_ptr3: U64(cmdbuf.depth_buffer_3), + depth_meta_buffer_ptr3: U64(cmdbuf.depth_meta_buffer_3), + stencil_buffer_ptr1: U64(cmdbuf.stencil_buffer_1), + unk_2d0: U64(0x0), + unk_2d8: U64(0x0), + stencil_buffer_ptr2: U64(cmdbuf.stencil_buffer_2), + stencil_buffer_ptr3: U64(cmdbuf.stencil_buffer_3), + stencil_meta_buffer_ptr3: U64(cmdbuf.stencil_meta_buffer_3), + unk_2f8: Default::default(), + iogpu_unk_212: cmdbuf.iogpu_unk_212, + unk_30c: 0x0, + aux_fb_info: aux_fb_info, + unk_320_padding: Default::default(), + unk_partial_store_pipeline: + fw::fragment::raw::StorePipelineBinding::new( + cmdbuf.partial_store_pipeline_bind, + cmdbuf.partial_store_pipeline + ), + partial_store_pipeline: fw::fragment::raw::StorePipelineBinding::new( + cmdbuf.partial_store_pipeline_bind, + cmdbuf.partial_store_pipeline + ), + isp_bgobjdepth: cmdbuf.isp_bgobjdepth, + isp_bgobjvals: cmdbuf.isp_bgobjvals, + iogpu_unk_49: cmdbuf.iogpu_unk_49, + unk_37c: 0x0, + unk_380: U64(0x0), + unk_388: U64(0x0), + #[ver(V >= V13_0B4)] + unk_390_0: U64(0x0), + depth_dimensions: U64(cmdbuf.depth_dimensions as u64), + }, + unk_758_flag: 0, + unk_75c_flag: 0, + unk_buf: Default::default(), + busy_flag: 0, + tvb_overflow_count: 0, + unk_878: 0, + encoder_params: fw::job::raw::EncoderParams { + unk_8: (cmdbuf.flags + & bindings::ASAHI_RENDER_SET_WHEN_RELOADING_Z_OR_S as u64 + != 0) as u32, + unk_c: 0x0, // fixed + unk_10: 0x0, // fixed + encoder_id: cmdbuf.encoder_id, + unk_18: 0x0, // fixed + iogpu_compute_unk44: 0xffffffff, + seq_buffer: inner.scene.seq_buf_pointer(), + unk_28: U64(0x0), // fixed + }, + process_empty_tiles: (cmdbuf.flags + & bindings::ASAHI_RENDER_PROCESS_EMPTY_TILES as u64 + != 0) as u32, + no_clear_pipeline_textures: (cmdbuf.flags + & bindings::ASAHI_RENDER_NO_CLEAR_PIPELINE_TEXTURES as u64 + != 0) as u32, + unk_param: unk2.into(), // 1 for boot stuff? + unk_pointee: 0, + meta: fw::job::raw::JobMeta { + unk_4: 0, + stamp: ev_frag.stamp_pointer, + fw_stamp: ev_frag.fw_stamp_pointer, + stamp_value: ev_frag.value.next(), + stamp_slot: ev_frag.slot, + evctl_index: 0, // fixed + flush_stamps: flush_stamps as u32, + uuid: uuid_3d, + cmd_seq: ev_frag.cmd_seq as u32, + }, + unk_after_meta: unk1.into(), + unk_buf_0: U64(0), + unk_buf_8: U64(0), + unk_buf_10: U64(1), + cur_ts: U64(0), + start_ts: Some(inner_ptr!(inner.timestamps.gpu_pointer(), frag.start)), + end_ts: Some(inner_ptr!(inner.timestamps.gpu_pointer(), frag.end)), + unk_914: 0, + unk_918: U64(0), + unk_920: 0, + client_sequence: slot_client_seq, + pad_925: Default::default(), + unk_928: 0, + unk_92c: 0, + #[ver(V >= V13_0B4)] + unk_ts: U64(0), + #[ver(V >= V13_0B4)] + unk_92d_8: Default::default(), + } + )) + }, + )?; + + mod_dev_dbg!(self.dev, "[Submission {}] Add Frag\n", id); + fence.add_command(); + + frag_job.add_cb(frag, vm_bind.slot(), move |cmd, error| { + if let Some(err) = error { + fence.set_error(err.into()); + } + if let Some(mut res) = frag_result.as_ref().map(|a| a.lock()) { + cmd.timestamps.with(|raw, _inner| { + res.result.fragment_ts_start = raw.frag.start.load(Ordering::Relaxed); + res.result.fragment_ts_end = raw.frag.end.load(Ordering::Relaxed); + }); + cmd.with(|raw, _inner| { + res.result.num_tvb_overflows = raw.tvb_overflow_count; + }); + res.frag_error = error; + res.frag_complete = true; + res.commit(); + } + fence.command_complete(); + })?; + + let fence = job.fence.clone(); + let vtx_job = job.get_vtx()?; + + if scene.rebind() || tvb_grown || tvb_autogrown { + mod_dev_dbg!(self.dev, "[Submission {}] Create Bind Buffer\n", id); + let bind_buffer = kalloc.private.new_inplace( + fw::buffer::InitBuffer::ver { + scene: scene.clone(), + }, + |inner, ptr: &mut MaybeUninit>| { + Ok(place!( + ptr, + fw::buffer::raw::InitBuffer::ver { + tag: fw::workqueue::CommandType::InitBuffer, + vm_slot: vm_bind.slot(), + buffer_slot: inner.scene.slot(), + unk_c: 0, + block_count: buffer.block_count(), + buffer: inner.scene.buffer_pointer(), + stamp_value: ev_vtx.value.next(), + } + )) + }, + )?; + + mod_dev_dbg!(self.dev, "[Submission {}] Add Bind Buffer\n", id); + vtx_job.add(bind_buffer, vm_bind.slot())?; + } + + mod_dev_dbg!(self.dev, "[Submission {}] Create Vertex\n", id); + let vtx = GpuObject::new_prealloc( + kalloc.private.alloc_object()?, + |ptr: GpuWeakPointer| { + let mut builder = microseq::Builder::new(); + + let stats = inner_weak_ptr!( + gpu.initdata.runtime_pointers.stats.vtx.weak_pointer(), + stats + ); + + let start_vtx = builder.add(microseq::StartVertex::ver { + header: microseq::op::StartVertex::HEADER, + tiling_params: inner_weak_ptr!(ptr, tiling_params), + job_params1: inner_weak_ptr!(ptr, job_params1), + buffer: scene.weak_buffer_pointer(), + scene: scene.weak_pointer(), + stats, + work_queue: ev_vtx.info_ptr, + vm_slot: vm_bind.slot(), + unk_38: 1, // fixed + event_generation: self.id as u32, + buffer_slot: scene.slot(), + unk_44: 0, + cmd_seq: U64(ev_vtx.cmd_seq), + unk_50: 0, + unk_pointer: inner_weak_ptr!(ptr, unk_pointee), + unk_job_buf: inner_weak_ptr!(ptr, unk_buf_0), + unk_64: 0x0, // fixed + unk_68: unk1.into(), + uuid: uuid_ta, + unk_70: 0x0, // fixed + unk_74: Default::default(), // fixed + unk_15c: 0x0, // fixed + unk_160: U64(0x0), // fixed + unk_168: 0x0, // fixed + unk_16c: 0x0, // fixed + unk_170: U64(0x0), // fixed + #[ver(V >= V13_0B4)] + counter: U64(count_vtx), + #[ver(V >= V13_0B4)] + notifier_buf: inner_weak_ptr!(notifier.weak_pointer(), state.unk_buf), + unk_178: 0x0, // padding? + })?; + + if vtx_result.is_some() { + builder.add(microseq::Timestamp::ver { + header: microseq::op::Timestamp::new(true), + cur_ts: inner_weak_ptr!(ptr, cur_ts), + start_ts: inner_weak_ptr!(ptr, start_ts), + update_ts: inner_weak_ptr!(ptr, start_ts), + work_queue: ev_vtx.info_ptr, + unk_24: U64(0), + #[ver(V >= V13_0B4)] + unk_ts: inner_weak_ptr!(ptr, unk_ts), + uuid: uuid_ta, + unk_30_padding: 0, + })?; + } + + builder.add(microseq::WaitForIdle { + header: microseq::op::WaitForIdle::new(microseq::Pipe::Vertex), + })?; + + if vtx_result.is_some() { + builder.add(microseq::Timestamp::ver { + header: microseq::op::Timestamp::new(false), + cur_ts: inner_weak_ptr!(ptr, cur_ts), + start_ts: inner_weak_ptr!(ptr, start_ts), + update_ts: inner_weak_ptr!(ptr, end_ts), + work_queue: ev_vtx.info_ptr, + unk_24: U64(0), + #[ver(V >= V13_0B4)] + unk_ts: inner_weak_ptr!(ptr, unk_ts), + uuid: uuid_ta, + unk_30_padding: 0, + })?; + } + + let off = builder.offset_to(start_vtx); + builder.add(microseq::FinalizeVertex::ver { + header: microseq::op::FinalizeVertex::HEADER, + scene: scene.weak_pointer(), + buffer: scene.weak_buffer_pointer(), + stats, + work_queue: ev_vtx.info_ptr, + vm_slot: vm_bind.slot(), + unk_28: 0x0, // fixed + unk_pointer: inner_weak_ptr!(ptr, unk_pointee), + unk_34: 0x0, // fixed + uuid: uuid_ta, + fw_stamp: ev_vtx.fw_stamp_pointer, + stamp_value: ev_vtx.value.next(), + unk_48: U64(0x0), // fixed + unk_50: 0x0, // fixed + unk_54: 0x0, // fixed + unk_58: U64(0x0), // fixed + unk_60: 0x0, // fixed + unk_64: 0x0, // fixed + unk_68: 0x0, // fixed + #[ver(G >= G14 && V < V13_0B4)] + unk_68_g14: U64(0), + restart_branch_offset: off, + unk_70: 0x0, // fixed + #[ver(V >= V13_0B4)] + unk_74: Default::default(), // Ventura + })?; + + builder.add(microseq::RetireStamp { + header: microseq::op::RetireStamp::HEADER, + })?; + + Ok(box_in_place!(fw::vertex::RunVertex::ver { + notifier: notifier, + scene: scene.clone(), + micro_seq: builder.build(&mut kalloc.private)?, + vm_bind: vm_bind.clone(), + timestamps: timestamps, + })?) + }, + |inner, ptr| { + #[ver(G < G14)] + let core_masks = gpu.core_masks_packed(); + Ok(place!( + ptr, + fw::vertex::raw::RunVertex::ver { + tag: fw::workqueue::CommandType::RunVertex, + #[ver(V >= V13_0B4)] + counter: U64(count_vtx), + vm_slot: vm_bind.slot(), + unk_8: 0, + notifier: inner.notifier.gpu_pointer(), + buffer_slot: inner.scene.slot(), + unk_1c: 0, + buffer: inner.scene.buffer_pointer(), + scene: inner.scene.gpu_pointer(), + unk_buffer_buf: inner.scene.kernel_buffer_pointer(), + unk_34: 0, + job_params1: fw::vertex::raw::JobParameters1::ver { + unk_0: U64(if unk1 { 0 } else { 0x200 }), // sometimes 0 + unk_8: f32!(1e-20), // fixed + unk_c: f32!(1e-20), // fixed + tvb_tilemap: inner.scene.tvb_tilemap_pointer(), + #[ver(G < G14)] + tvb_cluster_tilemaps: inner.scene.cluster_tilemaps_pointer(), + tpc: inner.scene.tpc_pointer(), + tvb_heapmeta: inner + .scene + .tvb_heapmeta_pointer() + .or(0x8000_0000_0000_0000), + iogpu_unk_54: 0x6b0003, // fixed + iogpu_unk_55: 0x3a0012, // fixed + iogpu_unk_56: U64(0x1), // fixed + #[ver(G < G14)] + tvb_cluster_meta1: inner + .scene + .meta_1_pointer() + .map(|x| x.or((tile_info.meta1_blocks as u64) << 50)), + utile_config: utile_config, + unk_4c: 0, + ppp_multisamplectl: U64(cmdbuf.ppp_multisamplectl), // fixed + tvb_heapmeta_2: inner.scene.tvb_heapmeta_pointer(), + #[ver(G < G14)] + unk_60: U64(0x0), // fixed + #[ver(G < G14)] + core_mask: Array::new([ + *core_masks.first().unwrap_or(&0), + *core_masks.get(1).unwrap_or(&0), + ]), + preempt_buf1: inner.scene.preempt_buf_1_pointer(), + preempt_buf2: inner.scene.preempt_buf_2_pointer(), + unk_80: U64(0x1), // fixed + preempt_buf3: inner + .scene + .preempt_buf_3_pointer() + .or(0x4_0000_0000_0000), // check + encoder_addr: U64(cmdbuf.encoder_ptr), + #[ver(G < G14)] + tvb_cluster_meta2: inner.scene.meta_2_pointer(), + #[ver(G < G14)] + tvb_cluster_meta3: inner.scene.meta_3_pointer(), + #[ver(G < G14)] + tiling_control: tiling_control, + #[ver(G < G14)] + unk_ac: Default::default(), // fixed + unk_b0: Default::default(), // fixed + pipeline_base: U64(0x11_00000000), + #[ver(G < G14)] + tvb_cluster_meta4: inner + .scene + .meta_4_pointer() + .map(|x| x.or(0x3000_0000_0000_0000)), + #[ver(G < G14)] + unk_f0: U64(0x1c + align(tile_info.meta1_blocks, 4) as u64), + unk_f8: U64(0x8c60), // fixed + unk_100: Default::default(), // fixed + unk_118: 0x1c, // fixed + #[ver(G >= G14)] + __pad: Default::default(), + }, + unk_154: Default::default(), + tiling_params: tile_info.params, + unk_3e8: Default::default(), + tpc: inner.scene.tpc_pointer(), + tpc_size: U64(tile_info.tpc_size as u64), + microsequence: inner.micro_seq.gpu_pointer(), + microsequence_size: inner.micro_seq.len() as u32, + fragment_stamp_slot: ev_frag.slot, + fragment_stamp_value: ev_frag.value.next(), + unk_pointee: 0, + unk_pad: 0, + job_params2: fw::vertex::raw::JobParameters2 { + unk_480: Default::default(), // fixed + unk_498: U64(0x0), // fixed + unk_4a0: 0x0, // fixed + preempt_buf1: inner.scene.preempt_buf_1_pointer(), + unk_4ac: 0x0, // fixed + unk_4b0: U64(0x0), // fixed + unk_4b8: 0x0, // fixed + unk_4bc: U64(0x0), // fixed + unk_4c4_padding: Default::default(), + unk_50c: 0x0, // fixed + unk_510: U64(0x0), // fixed + unk_518: U64(0x0), // fixed + unk_520: U64(0x0), // fixed + }, + encoder_params: fw::job::raw::EncoderParams { + unk_8: 0x0, // fixed + unk_c: 0x0, // fixed + unk_10: 0x0, // fixed + encoder_id: cmdbuf.encoder_id, + unk_18: 0x0, // fixed + iogpu_compute_unk44: 0xffffffff, + seq_buffer: inner.scene.seq_buf_pointer(), + unk_28: U64(0x0), // fixed + }, + unk_55c: 0, + unk_560: 0, + memoryless_rts_used: (cmdbuf.flags + & bindings::ASAHI_RENDER_MEMORYLESS_RTS_USED as u64 + != 0) as u32, + unk_568: 0, + unk_56c: 0, + meta: fw::job::raw::JobMeta { + unk_4: 0, + stamp: ev_vtx.stamp_pointer, + fw_stamp: ev_vtx.fw_stamp_pointer, + stamp_value: ev_vtx.value.next(), + stamp_slot: ev_vtx.slot, + evctl_index: 0, // fixed + flush_stamps: flush_stamps as u32, + uuid: uuid_ta, + cmd_seq: ev_vtx.cmd_seq as u32, + }, + unk_after_meta: unk1.into(), + unk_buf_0: U64(0), + unk_buf_8: U64(0), + unk_buf_10: U64(0), + cur_ts: U64(0), + start_ts: Some(inner_ptr!(inner.timestamps.gpu_pointer(), vtx.start)), + end_ts: Some(inner_ptr!(inner.timestamps.gpu_pointer(), vtx.end)), + unk_5c4: 0, + unk_5c8: 0, + unk_5cc: 0, + unk_5d0: 0, + client_sequence: slot_client_seq, + pad_5d5: Default::default(), + unk_5d8: 0, + unk_5dc: 0, + #[ver(V >= V13_0B4)] + unk_ts: U64(0), + #[ver(V >= V13_0B4)] + unk_5dd_8: Default::default(), + } + )) + }, + )?; + + core::mem::drop(alloc); + + mod_dev_dbg!(self.dev, "[Submission {}] Add Vertex\n", id); + fence.add_command(); + vtx_job.add_cb(vtx, vm_bind.slot(), move |cmd, error| { + if let Some(err) = error { + fence.set_error(err.into()) + } + if let Some(mut res) = vtx_result.as_ref().map(|a| a.lock()) { + cmd.timestamps.with(|raw, _inner| { + res.result.vertex_ts_start = raw.vtx.start.load(Ordering::Relaxed); + res.result.vertex_ts_end = raw.vtx.end.load(Ordering::Relaxed); + }); + res.result.tvb_usage_bytes = cmd.scene.used_bytes() as u64; + if cmd.scene.overflowed() { + res.result.flags |= bindings::DRM_ASAHI_RESULT_RENDER_TVB_OVERFLOWED as u64; + } + res.vtx_error = error; + res.vtx_complete = true; + res.commit(); + } + fence.command_complete(); + })?; + + mod_dev_dbg!(self.dev, "[Submission {}] Increment counters\n", id); + self.notifier.threshold.with(|raw, _inner| { + raw.increment(); + raw.increment(); + }); + + // TODO: handle rollbacks, move to job submit? + buffer.increment(); + + job.get_vtx()?.next_seq(); + job.get_frag()?.next_seq(); + + Ok(()) + } +} diff --git a/drivers/gpu/drm/asahi/regs.rs b/drivers/gpu/drm/asahi/regs.rs new file mode 100644 index 000000000000..019d7214793d --- /dev/null +++ b/drivers/gpu/drm/asahi/regs.rs @@ -0,0 +1,387 @@ +// SPDX-License-Identifier: GPL-2.0-only OR MIT + +//! GPU MMIO register abstraction +//! +//! Since the vast majority of the interactions with the GPU are brokered through the firmware, +//! there is very little need to interact directly with GPU MMIO register. This module abstracts +//! the few operations that require that, mainly reading the MMU fault status, reading GPU ID +//! information, and starting the GPU firmware coprocessor. + +use crate::hw; +use kernel::{device, io_mem::IoMem, platform, prelude::*}; + +/// Size of the ASC control MMIO region. +pub(crate) const ASC_CTL_SIZE: usize = 0x4000; + +/// Size of the SGX MMIO region. +pub(crate) const SGX_SIZE: usize = 0x1000000; + +const CPU_CONTROL: usize = 0x44; +const CPU_RUN: u32 = 0x1 << 4; // BIT(4) + +const FAULT_INFO: usize = 0x17030; + +const ID_VERSION: usize = 0xd04000; +const ID_UNK08: usize = 0xd04008; +const ID_COUNTS_1: usize = 0xd04010; +const ID_COUNTS_2: usize = 0xd04014; +const ID_UNK18: usize = 0xd04018; +const ID_CLUSTERS: usize = 0xd0401c; + +const CORE_MASK_0: usize = 0xd01500; +const CORE_MASK_1: usize = 0xd01514; + +/// Enum representing the unit that caused an MMU fault. +#[allow(non_camel_case_types)] +#[allow(clippy::upper_case_acronyms)] +#[derive(Copy, Clone, Debug, Eq, PartialEq)] +pub(crate) enum FaultUnit { + /// Decompress / pixel fetch + DCMP(u8), + /// USC L1 Cache (device loads/stores) + UL1C(u8), + /// Compress / pixel store + CMP(u8), + GSL1(u8), + IAP(u8), + VCE(u8), + /// Tiling Engine + TE(u8), + RAS(u8), + /// Vertex Data Master + VDM(u8), + PPP(u8), + /// ISP Parameter Fetch + IPF(u8), + IPF_CPF(u8), + VF(u8), + VF_CPF(u8), + /// Depth/Stencil load/store + ZLS(u8), + + /// Parameter Management + dPM, + /// Compute Data Master + dCDM_KS(u8), + dIPP, + dIPP_CS, + // Vertex Data Master + dVDM_CSD, + dVDM_SSD, + dVDM_ILF, + dVDM_ILD, + dRDE(u8), + FC, + GSL2, + + /// Graphics L2 Cache Control? + GL2CC_META(u8), + GL2CC_MB, + + /// Parameter Management + gPM_SP(u8), + /// Vertex Data Master - CSD + gVDM_CSD_SP(u8), + gVDM_SSD_SP(u8), + gVDM_ILF_SP(u8), + gVDM_TFP_SP(u8), + gVDM_MMB_SP(u8), + /// Compute Data Master + gCDM_CS_KS0_SP(u8), + gCDM_CS_KS1_SP(u8), + gCDM_CS_KS2_SP(u8), + gCDM_KS0_SP(u8), + gCDM_KS1_SP(u8), + gCDM_KS2_SP(u8), + gIPP_SP(u8), + gIPP_CS_SP(u8), + gRDE0_SP(u8), + gRDE1_SP(u8), + + Unknown(u8), +} + +/// Reason for an MMU fault. +#[derive(Copy, Clone, Debug, Eq, PartialEq)] +pub(crate) enum FaultReason { + Unmapped, + AfFault, + WriteOnly, + ReadOnly, + NoAccess, + Unknown(u8), +} + +/// Collection of information about an MMU fault. +#[derive(Copy, Clone, Debug, Eq, PartialEq)] +pub(crate) struct FaultInfo { + pub(crate) address: u64, + pub(crate) sideband: u8, + pub(crate) vm_slot: u32, + pub(crate) unit_code: u8, + pub(crate) unit: FaultUnit, + pub(crate) level: u8, + pub(crate) unk_5: u8, + pub(crate) read: bool, + pub(crate) reason: FaultReason, +} + +/// Device resources for this GPU instance. +pub(crate) struct Resources { + dev: device::Device, + asc: IoMem, + sgx: IoMem, +} + +impl Resources { + /// Map the required resources given our platform device. + pub(crate) fn new(pdev: &mut platform::Device) -> Result { + // TODO: add device abstraction to ioremap by name + let asc_res = unsafe { pdev.ioremap_resource(0)? }; + let sgx_res = unsafe { pdev.ioremap_resource(1)? }; + + Ok(Resources { + // SAFETY: This device does DMA via the UAT IOMMU. + dev: device::Device::from_dev(pdev), + asc: asc_res, + sgx: sgx_res, + }) + } + + fn sgx_read32(&self, off: usize) -> u32 { + self.sgx.readl_relaxed(off) + } + + /* Not yet used + fn sgx_write32(&self, off: usize, val: u32) { + self.sgx.writel_relaxed(val, off) + } + */ + + fn sgx_read64(&self, off: usize) -> u64 { + self.sgx.readq_relaxed(off) + } + + /* Not yet used + fn sgx_write64(&self, off: usize, val: u64) { + self.sgx.writeq_relaxed(val, off) + } + */ + + /// Initialize the MMIO registers for the GPU. + pub(crate) fn init_mmio(&self) -> Result { + // Nothing to do for now... + + Ok(()) + } + + /// Start the ASC coprocessor CPU. + pub(crate) fn start_cpu(&self) -> Result { + let val = self.asc.readl_relaxed(CPU_CONTROL); + + self.asc.writel_relaxed(val | CPU_RUN, CPU_CONTROL); + + Ok(()) + } + + /// Get the GPU identification info from registers. + /// + /// See [`hw::GpuIdConfig`] for the result. + pub(crate) fn get_gpu_id(&self) -> Result { + let id_version = self.sgx_read32(ID_VERSION); + let id_unk08 = self.sgx_read32(ID_UNK08); + let id_counts_1 = self.sgx_read32(ID_COUNTS_1); + let id_counts_2 = self.sgx_read32(ID_COUNTS_2); + let id_unk18 = self.sgx_read32(ID_UNK18); + let id_clusters = self.sgx_read32(ID_CLUSTERS); + + dev_info!( + self.dev, + "GPU ID registers: {:#x} {:#x} {:#x} {:#x} {:#x} {:#x}\n", + id_version, + id_unk08, + id_counts_1, + id_counts_2, + id_unk18, + id_clusters + ); + + let core_mask_0 = self.sgx_read32(CORE_MASK_0); + let core_mask_1 = self.sgx_read32(CORE_MASK_1); + let mut core_mask = (core_mask_0 as u64) | ((core_mask_1 as u64) << 32); + + dev_info!(self.dev, "Core mask: {:#x}\n", core_mask); + + let num_clusters = (id_clusters >> 12) & 0xff; + let num_cores = id_counts_1 & 0xff; + + if num_cores * num_clusters > 64 { + dev_err!( + self.dev, + "Too many total cores ({} x {} > 64)\n", + num_clusters, + num_cores + ); + return Err(ENODEV); + } + + let mut core_masks = Vec::new(); + let mut total_active_cores: u32 = 0; + + let max_core_mask = (1u64 << num_cores) - 1; + for _i in 0..num_clusters { + let mask = core_mask & max_core_mask; + core_masks.try_push(mask as u32)?; + core_mask >>= num_cores; + total_active_cores += mask.count_ones(); + } + let mut core_masks_packed = Vec::new(); + core_masks_packed.try_push(core_mask_0)?; + if core_mask_1 != 0 { + core_masks_packed.try_push(core_mask_1)?; + } + + if core_mask != 0 { + dev_err!(self.dev, "Leftover core mask: {:#x}\n", core_mask); + return Err(EIO); + } + + let (gpu_rev, gpu_rev_id) = match (id_version >> 8) & 0xff { + 0x00 => (hw::GpuRevision::A0, hw::GpuRevisionID::A0), + 0x01 => (hw::GpuRevision::A1, hw::GpuRevisionID::A1), + 0x10 => (hw::GpuRevision::B0, hw::GpuRevisionID::B0), + 0x11 => (hw::GpuRevision::B1, hw::GpuRevisionID::B1), + 0x20 => (hw::GpuRevision::C0, hw::GpuRevisionID::C0), + 0x21 => (hw::GpuRevision::C1, hw::GpuRevisionID::C1), + a => { + dev_err!(self.dev, "Unknown GPU revision {}\n", a); + return Err(ENODEV); + } + }; + + Ok(hw::GpuIdConfig { + gpu_gen: match (id_version >> 24) & 0xff { + 4 => hw::GpuGen::G13, + 5 => hw::GpuGen::G14, + a => { + dev_err!(self.dev, "Unknown GPU generation {}\n", a); + return Err(ENODEV); + } + }, + gpu_variant: match (id_version >> 16) & 0xff { + 1 => hw::GpuVariant::P, // Guess + 2 => hw::GpuVariant::G, + 3 => hw::GpuVariant::S, + 4 => { + if num_clusters > 4 { + hw::GpuVariant::D + } else { + hw::GpuVariant::C + } + } + a => { + dev_err!(self.dev, "Unknown GPU variant {}\n", a); + return Err(ENODEV); + } + }, + gpu_rev, + gpu_rev_id, + max_dies: (id_clusters >> 20) & 0xf, + num_clusters, + num_cores, + num_frags: (id_counts_1 >> 8) & 0xff, + num_gps: (id_counts_2 >> 16) & 0xff, + total_active_cores, + core_masks, + core_masks_packed, + }) + } + + /// Get the fault information from the MMU status register, if one occurred. + pub(crate) fn get_fault_info(&self) -> Option { + let fault_info = self.sgx_read64(FAULT_INFO); + + if fault_info & 1 == 0 { + return None; + } + + let unit_code = ((fault_info >> 9) & 0xff) as u8; + let unit = match unit_code { + 0x00..=0x9f => match unit_code & 0xf { + 0x0 => FaultUnit::DCMP(unit_code >> 4), + 0x1 => FaultUnit::UL1C(unit_code >> 4), + 0x2 => FaultUnit::CMP(unit_code >> 4), + 0x3 => FaultUnit::GSL1(unit_code >> 4), + 0x4 => FaultUnit::IAP(unit_code >> 4), + 0x5 => FaultUnit::VCE(unit_code >> 4), + 0x6 => FaultUnit::TE(unit_code >> 4), + 0x7 => FaultUnit::RAS(unit_code >> 4), + 0x8 => FaultUnit::VDM(unit_code >> 4), + 0x9 => FaultUnit::PPP(unit_code >> 4), + 0xa => FaultUnit::IPF(unit_code >> 4), + 0xb => FaultUnit::IPF_CPF(unit_code >> 4), + 0xc => FaultUnit::VF(unit_code >> 4), + 0xd => FaultUnit::VF_CPF(unit_code >> 4), + 0xe => FaultUnit::ZLS(unit_code >> 4), + _ => FaultUnit::Unknown(unit_code), + }, + 0xa1 => FaultUnit::dPM, + 0xa2 => FaultUnit::dCDM_KS(0), + 0xa3 => FaultUnit::dCDM_KS(1), + 0xa4 => FaultUnit::dCDM_KS(2), + 0xa5 => FaultUnit::dIPP, + 0xa6 => FaultUnit::dIPP_CS, + 0xa7 => FaultUnit::dVDM_CSD, + 0xa8 => FaultUnit::dVDM_SSD, + 0xa9 => FaultUnit::dVDM_ILF, + 0xaa => FaultUnit::dVDM_ILD, + 0xab => FaultUnit::dRDE(0), + 0xac => FaultUnit::dRDE(1), + 0xad => FaultUnit::FC, + 0xae => FaultUnit::GSL2, + 0xb0..=0xb7 => FaultUnit::GL2CC_META(unit_code & 0xf), + 0xb8 => FaultUnit::GL2CC_MB, + 0xe0..=0xff => match unit_code & 0xf { + 0x0 => FaultUnit::gPM_SP((unit_code >> 4) & 1), + 0x1 => FaultUnit::gVDM_CSD_SP((unit_code >> 4) & 1), + 0x2 => FaultUnit::gVDM_SSD_SP((unit_code >> 4) & 1), + 0x3 => FaultUnit::gVDM_ILF_SP((unit_code >> 4) & 1), + 0x4 => FaultUnit::gVDM_TFP_SP((unit_code >> 4) & 1), + 0x5 => FaultUnit::gVDM_MMB_SP((unit_code >> 4) & 1), + 0x6 => FaultUnit::gCDM_CS_KS0_SP((unit_code >> 4) & 1), + 0x7 => FaultUnit::gCDM_CS_KS1_SP((unit_code >> 4) & 1), + 0x8 => FaultUnit::gCDM_CS_KS2_SP((unit_code >> 4) & 1), + 0x9 => FaultUnit::gCDM_KS0_SP((unit_code >> 4) & 1), + 0xa => FaultUnit::gCDM_KS1_SP((unit_code >> 4) & 1), + 0xb => FaultUnit::gCDM_KS2_SP((unit_code >> 4) & 1), + 0xc => FaultUnit::gIPP_SP((unit_code >> 4) & 1), + 0xd => FaultUnit::gIPP_CS_SP((unit_code >> 4) & 1), + 0xe => FaultUnit::gRDE0_SP((unit_code >> 4) & 1), + 0xf => FaultUnit::gRDE1_SP((unit_code >> 4) & 1), + _ => FaultUnit::Unknown(unit_code), + }, + _ => FaultUnit::Unknown(unit_code), + }; + + let reason = match (fault_info >> 1) & 0x7 { + 0 => FaultReason::Unmapped, + 1 => FaultReason::AfFault, + 2 => FaultReason::WriteOnly, + 3 => FaultReason::ReadOnly, + 4 => FaultReason::NoAccess, + a => FaultReason::Unknown(a as u8), + }; + + Some(FaultInfo { + address: (fault_info >> 30) << 6, + sideband: ((fault_info >> 23) & 0x7f) as u8, + vm_slot: ((fault_info >> 17) & 0x3f) as u32, + unit_code, + unit, + level: ((fault_info >> 7) & 3) as u8, + unk_5: ((fault_info >> 5) & 3) as u8, + read: (fault_info & (1 << 4)) != 0, + reason, + }) + } +} diff --git a/drivers/gpu/drm/asahi/slotalloc.rs b/drivers/gpu/drm/asahi/slotalloc.rs new file mode 100644 index 000000000000..6493111643fe --- /dev/null +++ b/drivers/gpu/drm/asahi/slotalloc.rs @@ -0,0 +1,292 @@ +// SPDX-License-Identifier: GPL-2.0-only OR MIT + +//! Generic slot allocator +//! +//! This is a simple allocator to manage fixed-size pools of GPU resources that are transiently +//! required during command execution. Each item resides in a "slot" at a given index. Users borrow +//! and return free items from the available pool. +//! +//! Allocations are "sticky", and return a token that callers can use to request the same slot +//! again later. This allows slots to be lazily invalidated, so that multiple uses by the same user +//! avoid any actual cleanup work. +//! +//! The allocation policy is currently a simple LRU mechanism, doing a full linear scan over the +//! slots when no token was previously provided. This is probably good enough, since in the absence +//! of serious system contention most allocation requests will be immediately fulfilled from the +//! previous slot without doing an LRU scan. + +use core::ops::{Deref, DerefMut}; +use kernel::{ + error::{code::*, Result}, + prelude::*, + sync::{Arc, CondVar, Mutex, UniqueArc}, +}; + +/// Trait representing a single item within a slot. +pub(crate) trait SlotItem { + /// Arbitrary user data associated with the SlotAllocator. + type Data; + + /// Called eagerly when this item is released back into the available pool. + fn release(&mut self, _data: &mut Self::Data, _slot: u32) {} +} + +/// Trivial implementation for users which do not require any slot data nor any allocator data. +impl SlotItem for () { + type Data = (); +} + +/// Represents a current or previous allocation of an item from a slot. Users keep `SlotToken`s +/// around across allocations to request that, if possible, the same slot be reused. +#[derive(Copy, Clone, Debug)] +pub(crate) struct SlotToken { + time: u64, + slot: u32, +} + +impl SlotToken { + /// Returns the slot index that this token represents a past assignment to. + pub(crate) fn last_slot(&self) -> u32 { + self.slot + } +} + +/// A guard representing active ownership of a slot. +pub(crate) struct Guard { + item: Option, + changed: bool, + token: SlotToken, + alloc: Arc>, +} + +impl Guard { + /// Returns the active slot owned by this `Guard`. + pub(crate) fn slot(&self) -> u32 { + self.token.slot + } + + /// Returns `true` if the slot changed since the last allocation (or no `SlotToken` was + /// provided), or `false` if the previously allocated slot was successfully re-acquired with + /// no other users in the interim. + pub(crate) fn changed(&self) -> bool { + self.changed + } + + /// Returns a `SlotToken` that can be used to re-request the same slot at a later time, after + /// this `Guard` is dropped. + pub(crate) fn token(&self) -> SlotToken { + self.token + } +} + +impl Deref for Guard { + type Target = T; + + fn deref(&self) -> &Self::Target { + self.item.as_ref().expect("SlotItem Guard lost our item!") + } +} + +impl DerefMut for Guard { + fn deref_mut(&mut self) -> &mut Self::Target { + self.item.as_mut().expect("SlotItem Guard lost our item!") + } +} + +/// A slot item that is currently free. +struct Entry { + item: T, + get_time: u64, + drop_time: u64, +} + +/// Inner data for the `SlotAllocator`, protected by a `Mutex`. +struct SlotAllocatorInner { + data: T::Data, + slots: Vec>>, + get_count: u64, + drop_count: u64, +} + +/// A single slot allocator instance. +struct SlotAllocatorOuter { + inner: Mutex>, + cond: CondVar, +} + +/// A shared reference to a slot allocator instance. +pub(crate) struct SlotAllocator(Arc>); + +impl SlotAllocator { + /// Creates a new `SlotAllocator`, with a fixed number of slots and arbitrary associated data. + /// + /// The caller provides a constructor callback which takes a reference to the `T::Data` and + /// creates a single slot. This is called during construction to create all the initial + /// items, which then live the lifetime of the `SlotAllocator`. + pub(crate) fn new( + num_slots: u32, + mut data: T::Data, + mut constructor: impl FnMut(&mut T::Data, u32) -> T, + ) -> Result> { + let mut slots = Vec::try_with_capacity(num_slots as usize)?; + + for i in 0..num_slots { + slots + .try_push(Some(Entry { + item: constructor(&mut data, i), + get_time: 0, + drop_time: 0, + })) + .expect("try_push() failed after reservation"); + } + + let inner = SlotAllocatorInner { + data, + slots, + get_count: 0, + drop_count: 0, + }; + + let mut alloc = Pin::from(UniqueArc::try_new(SlotAllocatorOuter { + // SAFETY: `condvar_init!` is called below. + cond: unsafe { CondVar::new() }, + // SAFETY: `mutex_init!` is called below. + inner: unsafe { Mutex::new(inner) }, + })?); + + // SAFETY: `cond` is pinned when `alloc` is. + let pinned = unsafe { alloc.as_mut().map_unchecked_mut(|s| &mut s.cond) }; + kernel::condvar_init!(pinned, "SlotAllocator::cond"); + + // SAFETY: `inner` is pinned when `alloc` is. + let pinned = unsafe { alloc.as_mut().map_unchecked_mut(|s| &mut s.inner) }; + kernel::mutex_init!(pinned, "SlotAllocator::inner"); + + Ok(SlotAllocator(alloc.into())) + } + + /// Calls a callback on the inner data associated with this allocator, taking the lock. + pub(crate) fn with_inner(&self, cb: impl FnOnce(&mut T::Data) -> RetVal) -> RetVal { + let mut inner = self.0.inner.lock(); + cb(&mut inner.data) + } + + /// Gets a fresh slot, optionally reusing a previous allocation if a `SlotToken` is provided. + /// + /// Blocks if no slots are free. + pub(crate) fn get(&self, token: Option) -> Result> { + self.get_inner(token, |_a, _b| Ok(())) + } + + /// Gets a fresh slot, optionally reusing a previous allocation if a `SlotToken` is provided. + /// + /// Blocks if no slots are free. + /// + /// This version allows the caller to pass in a callback that gets a mutable reference to the + /// user data for the allocator and the freshly acquired slot, which is called before the + /// allocator lock is released. This can be used to perform bookkeeping associated with + /// specific slots (such as tracking their current owner). + pub(crate) fn get_inner( + &self, + token: Option, + cb: impl FnOnce(&mut T::Data, &mut Guard) -> Result<()>, + ) -> Result> { + let mut inner = self.0.inner.lock(); + + if let Some(token) = token { + let slot = &mut inner.slots[token.slot as usize]; + if slot.is_some() { + let count = slot.as_ref().unwrap().get_time; + if count == token.time { + let mut guard = Guard { + item: Some(slot.take().unwrap().item), + token, + changed: false, + alloc: self.0.clone(), + }; + cb(&mut inner.data, &mut guard)?; + return Ok(guard); + } + } + } + + let mut first = true; + let slot = loop { + let mut oldest_time = u64::MAX; + let mut oldest_slot = 0u32; + + for (i, slot) in inner.slots.iter().enumerate() { + if let Some(slot) = slot.as_ref() { + if slot.drop_time < oldest_time { + oldest_slot = i as u32; + oldest_time = slot.drop_time; + } + } + } + + if oldest_time == u64::MAX { + if first { + pr_warn!( + "{}: out of slots, blocking\n", + core::any::type_name::() + ); + } + first = false; + if self.0.cond.wait(&mut inner) { + return Err(ERESTARTSYS); + } + } else { + break oldest_slot; + } + }; + + inner.get_count += 1; + + let item = inner.slots[slot as usize] + .take() + .expect("Someone stole our slot?") + .item; + + let mut guard = Guard { + item: Some(item), + changed: true, + token: SlotToken { + time: inner.get_count, + slot, + }, + alloc: self.0.clone(), + }; + + cb(&mut inner.data, &mut guard)?; + Ok(guard) + } +} + +impl Clone for SlotAllocator { + fn clone(&self) -> Self { + SlotAllocator(self.0.clone()) + } +} + +impl Drop for Guard { + fn drop(&mut self) { + let mut inner = self.alloc.inner.lock(); + if inner.slots[self.token.slot as usize].is_some() { + pr_crit!( + "{}: tried to return an item into a full slot ({})\n", + core::any::type_name::(), + self.token.slot + ); + } else { + inner.drop_count += 1; + let mut item = self.item.take().expect("Guard lost its item"); + item.release(&mut inner.data, self.token.slot); + inner.slots[self.token.slot as usize] = Some(Entry { + item, + get_time: self.token.time, + drop_time: inner.drop_count, + }); + self.alloc.cond.notify_one(); + } + } +} diff --git a/drivers/gpu/drm/asahi/util.rs b/drivers/gpu/drm/asahi/util.rs new file mode 100644 index 000000000000..8d1a37f17cd8 --- /dev/null +++ b/drivers/gpu/drm/asahi/util.rs @@ -0,0 +1,44 @@ +// SPDX-License-Identifier: GPL-2.0-only OR MIT + +//! Miscellaneous utility functions + +use core::ops::{Add, BitAnd, Div, Not, Sub}; + +/// Aligns an integer type to a power of two. +pub(crate) fn align(a: T, b: T) -> T +where + T: Copy + + Default + + BitAnd + + Not + + Add + + Sub + + Div + + core::cmp::PartialEq, +{ + let def: T = Default::default(); + #[allow(clippy::eq_op)] + let one: T = !def / !def; + + assert!((b & (b - one)) == def); + + (a + b - one) & !(b - one) +} + +/// Integer division rounding up. +pub(crate) fn div_ceil(a: T, b: T) -> T +where + T: Copy + + Default + + BitAnd + + Not + + Add + + Sub + + Div, +{ + let def: T = Default::default(); + #[allow(clippy::eq_op)] + let one: T = !def / !def; + + (a + b - one) / b +} diff --git a/drivers/gpu/drm/asahi/workqueue.rs b/drivers/gpu/drm/asahi/workqueue.rs new file mode 100644 index 000000000000..ce1d1f89e48e --- /dev/null +++ b/drivers/gpu/drm/asahi/workqueue.rs @@ -0,0 +1,880 @@ +// SPDX-License-Identifier: GPL-2.0-only OR MIT + +//! GPU command execution queues +//! +//! The AGX GPU firmware schedules GPU work commands out of work queues, which are ring buffers of +//! pointers to work commands. There can be an arbitrary number of work queues. Work queues have an +//! associated type (vertex, fragment, or compute) and may only contain generic commands or commands +//! specific to that type. +//! +//! This module manages queueing work commands into a work queue and submitting them for execution +//! by the firmware. An active work queue needs an event to signal completion of its work, which is +//! owned by what we call a batch. This event then notifies the work queue when work is completed, +//! and that triggers freeing of all resources associated with that work. An idle work queue gives +//! up its associated event. + +use crate::debug::*; +use crate::fw::channels::PipeType; +use crate::fw::types::*; +use crate::fw::workqueue::*; +use crate::object::OpaqueGpuObject; +use crate::regs::FaultReason; +use crate::{box_in_place, no_debug, place}; +use crate::{channel, driver, event, fw, gpu, object, regs}; +use core::num::NonZeroU64; +use core::sync::atomic::Ordering; +use kernel::{ + bindings, + error::code::*, + prelude::*, + sync::{Arc, Guard, Mutex, UniqueArc}, +}; + +const DEBUG_CLASS: DebugFlags = DebugFlags::WorkQueue; + +const MAX_JOB_SLOTS: u32 = 127; + +/// An enum of possible errors that might cause a piece of work to fail execution. +#[derive(Copy, Clone, Debug, PartialEq, Eq)] +pub(crate) enum WorkError { + /// GPU timeout (command execution took too long). + Timeout, + /// GPU MMU fault (invalid access). + Fault(regs::FaultInfo), + /// Work failed due to an error caused by other concurrent GPU work. + Killed, + /// The GPU crashed. + NoDevice, + /// Unknown reason. + Unknown, +} + +impl From for bindings::drm_asahi_result_info { + fn from(err: WorkError) -> Self { + match err { + WorkError::Fault(info) => Self { + status: bindings::drm_asahi_status_DRM_ASAHI_STATUS_FAULT, + fault_type: match info.reason { + FaultReason::Unmapped => bindings::drm_asahi_fault_DRM_ASAHI_FAULT_UNMAPPED, + FaultReason::AfFault => bindings::drm_asahi_fault_DRM_ASAHI_FAULT_AF_FAULT, + FaultReason::WriteOnly => bindings::drm_asahi_fault_DRM_ASAHI_FAULT_WRITE_ONLY, + FaultReason::ReadOnly => bindings::drm_asahi_fault_DRM_ASAHI_FAULT_READ_ONLY, + FaultReason::NoAccess => bindings::drm_asahi_fault_DRM_ASAHI_FAULT_NO_ACCESS, + FaultReason::Unknown(_) => bindings::drm_asahi_fault_DRM_ASAHI_FAULT_UNKNOWN, + }, + unit: info.unit_code.into(), + sideband: info.sideband.into(), + level: info.level, + extra: info.unk_5.into(), + is_read: info.read as u8, + pad: 0, + address: info.address, + }, + a => Self { + status: match a { + WorkError::Timeout => bindings::drm_asahi_status_DRM_ASAHI_STATUS_TIMEOUT, + WorkError::Killed => bindings::drm_asahi_status_DRM_ASAHI_STATUS_KILLED, + WorkError::NoDevice => bindings::drm_asahi_status_DRM_ASAHI_STATUS_NO_DEVICE, + _ => bindings::drm_asahi_status_DRM_ASAHI_STATUS_UNKNOWN_ERROR, + }, + ..Default::default() + }, + } + } +} + +impl From for kernel::error::Error { + fn from(err: WorkError) -> Self { + match err { + WorkError::Timeout => ETIMEDOUT, + // Not EFAULT because that's for userspace faults + WorkError::Fault(_) => EIO, + WorkError::Unknown => ENODATA, + WorkError::Killed => ECANCELED, + WorkError::NoDevice => ENODEV, + } + } +} + +/// A GPU context tracking structure, which must be explicitly invalidated when dropped. +pub(crate) struct GpuContext { + dev: driver::AsahiDevice, + data: GpuObject, +} +no_debug!(GpuContext); + +impl GpuContext { + /// Allocate a new GPU context. + pub(crate) fn new( + dev: &driver::AsahiDevice, + alloc: &mut gpu::KernelAllocators, + ) -> Result { + Ok(GpuContext { + dev: dev.clone(), + data: alloc + .shared + .new_object(Default::default(), |_inner| Default::default())?, + }) + } + + /// Returns the GPU pointer to the inner GPU context data structure. + pub(crate) fn gpu_pointer(&self) -> GpuPointer<'_, fw::workqueue::GpuContextData> { + self.data.gpu_pointer() + } +} + +impl Drop for GpuContext { + fn drop(&mut self) { + mod_dev_dbg!(self.dev, "GpuContext: Invalidating GPU context\n"); + let dev = self.dev.data(); + if dev.gpu.invalidate_context(&self.data).is_err() { + dev_err!(self.dev, "GpuContext: Failed to invalidate GPU context!\n"); + } + } +} + +struct SubmittedWork +where + O: OpaqueGpuObject, + C: FnOnce(O, Option) + Send + Sync + 'static, +{ + object: O, + value: EventValue, + error: Option, + wptr: u32, + vm_slot: u32, + callback: C, +} + +trait GenSubmittedWork: Send + Sync { + fn gpu_va(&self) -> NonZeroU64; + fn value(&self) -> event::EventValue; + fn wptr(&self) -> u32; + fn set_wptr(&mut self, wptr: u32); + fn mark_error(&mut self, error: WorkError); + fn complete(self: Box); +} + +impl) + Send + Sync> GenSubmittedWork + for SubmittedWork +{ + fn gpu_va(&self) -> NonZeroU64 { + self.object.gpu_va() + } + + fn value(&self) -> event::EventValue { + self.value + } + + fn wptr(&self) -> u32 { + self.wptr + } + + fn set_wptr(&mut self, wptr: u32) { + self.wptr = wptr; + } + + fn complete(self: Box) { + let SubmittedWork { + object, + value: _, + error, + wptr: _, + vm_slot: _, + callback, + } = *self; + + callback(object, error); + } + + fn mark_error(&mut self, error: WorkError) { + mod_pr_debug!("WorkQueue: Command at value {:#x?} failed\n", self.value); + self.error = Some(match error { + WorkError::Fault(info) if info.vm_slot != self.vm_slot => WorkError::Killed, + err => err, + }); + } +} + +/// Inner data for managing a single work queue. +#[versions(AGX)] +struct WorkQueueInner { + event_manager: Arc, + info: GpuObject, + new: bool, + pipe_type: PipeType, + size: u32, + wptr: u32, + pending: Vec>, + last_token: Option, + pending_jobs: usize, + last_submitted: Option, + last_completed: Option, + event: Option<(event::Event, event::EventValue)>, + priority: u32, + commit_seq: u64, + submit_seq: u64, +} + +/// An instance of a work queue. +#[versions(AGX)] +pub(crate) struct WorkQueue { + info_pointer: GpuWeakPointer, + inner: Mutex, +} + +#[versions(AGX)] +impl WorkQueueInner::ver { + /// Return the GPU done pointer, representing how many work items have been completed by the + /// GPU. + fn doneptr(&self) -> u32 { + self.info + .state + .with(|raw, _inner| raw.gpu_doneptr.load(Ordering::Acquire)) + } +} + +#[versions(AGX)] +#[derive(Copy, Clone)] +pub(crate) struct QueueEventInfo { + pub(crate) stamp_pointer: GpuWeakPointer, + pub(crate) fw_stamp_pointer: GpuWeakPointer, + pub(crate) slot: u32, + pub(crate) value: event::EventValue, + pub(crate) cmd_seq: u64, + pub(crate) info_ptr: GpuWeakPointer, +} + +#[versions(AGX)] +pub(crate) struct Job { + wq: Arc, + event_info: QueueEventInfo::ver, + start_value: EventValue, + pending: Vec>, + committed: bool, + submitted: bool, + event_count: usize, +} + +#[versions(AGX)] +pub(crate) struct JobSubmission<'a> { + inner: Option>>, + wptr: u32, + event_count: usize, + command_count: usize, +} + +#[versions(AGX)] +impl Job::ver { + pub(crate) fn event_info(&self) -> QueueEventInfo::ver { + let mut info = self.event_info; + info.cmd_seq += self.event_count as u64; + + info + } + + pub(crate) fn next_seq(&mut self) { + self.event_count += 1; + self.event_info.value.increment(); + } + + pub(crate) fn add( + &mut self, + command: O, + vm_slot: u32, + ) -> Result { + self.add_cb(command, vm_slot, |_, _| {}) + } + + pub(crate) fn add_cb( + &mut self, + command: O, + vm_slot: u32, + callback: impl FnOnce(O, Option) + Sync + Send + 'static, + ) -> Result { + if self.committed { + pr_err!("WorkQueue: Tried to mutate committed Job\n"); + return Err(EINVAL); + } + + self.pending.try_push(Box::try_new(SubmittedWork::<_, _> { + object: command, + value: self.event_info.value.next(), + error: None, + callback, + wptr: 0, + vm_slot, + })?)?; + + Ok(()) + } + + pub(crate) fn commit(&mut self) -> Result { + if self.committed { + pr_err!("WorkQueue: Tried to commit committed Job\n"); + return Err(EINVAL); + } + + if self.pending.is_empty() { + pr_err!("WorkQueue: Job::commit() with no commands\n"); + return Err(EINVAL); + } + + let mut inner = self.wq.inner.lock(); + + let ev = inner.event.as_mut().expect("WorkQueue: Job lost its event"); + + if ev.1 != self.start_value { + pr_err!( + "WorkQueue: Job::commit() out of order (event slot {} {:?} != {:?}\n", + ev.0.slot(), + ev.1, + self.start_value + ); + return Err(EINVAL); + } + + ev.1 = self.event_info.value; + inner.commit_seq += self.pending.len() as u64; + self.committed = true; + + Ok(()) + } + + pub(crate) fn can_submit(&self) -> bool { + self.wq.free_slots() > self.event_count && self.wq.free_space() > self.pending.len() + } + + pub(crate) fn submit(&mut self) -> Result> { + if !self.committed { + pr_err!("WorkQueue: Tried to submit uncommitted Job\n"); + return Err(EINVAL); + } + + if self.submitted { + pr_err!("WorkQueue: Tried to submit Job twice\n"); + return Err(EINVAL); + } + + if self.pending.is_empty() { + pr_err!("WorkQueue: Job::submit() with no commands\n"); + return Err(EINVAL); + } + + let mut inner = self.wq.inner.lock(); + + if inner.submit_seq != self.event_info.cmd_seq { + pr_err!( + "WorkQueue: Job::submit() out of order (submit_seq {} != {})\n", + inner.submit_seq, + self.event_info.cmd_seq + ); + return Err(EINVAL); + } + + if inner.commit_seq < (self.event_info.cmd_seq + self.pending.len() as u64) { + pr_err!( + "WorkQueue: Job::submit() out of order (commit_seq {} != {})\n", + inner.commit_seq, + (self.event_info.cmd_seq + self.pending.len() as u64) + ); + return Err(EINVAL); + } + + let mut wptr = inner.wptr; + let command_count = self.pending.len(); + + if inner.free_space() <= command_count { + pr_err!("WorkQueue: Job does not fit in ring buffer\n"); + return Err(EBUSY); + } + + inner.pending.try_reserve(command_count)?; + + inner.last_submitted = inner.event.as_ref().map(|e| e.1); + + for mut command in self.pending.drain(..) { + command.set_wptr(wptr); + + let next_wptr = (wptr + 1) % inner.size; + assert!(inner.doneptr() != next_wptr); + inner.info.ring[wptr as usize] = command.gpu_va().get(); + wptr = next_wptr; + + // Cannot fail, since we did a try_reserve(1) above + inner + .pending + .try_push(command) + .expect("try_push() failed after try_reserve()"); + } + + self.submitted = true; + + Ok(JobSubmission::ver { + inner: Some(inner), + wptr, + command_count, + event_count: self.event_count, + }) + } +} + +#[versions(AGX)] +impl<'a> JobSubmission::ver<'a> { + pub(crate) fn run(mut self, channel: &mut channel::PipeChannel::ver) { + let command_count = self.command_count; + let mut inner = self.inner.take().expect("No inner?"); + let wptr = self.wptr; + core::mem::forget(self); + + inner + .info + .state + .with(|raw, _inner| raw.cpu_wptr.store(wptr, Ordering::Release)); + + inner.wptr = wptr; + + let event = inner.event.as_mut().expect("JobSubmission lost its event"); + + let event_slot = event.0.slot(); + + let msg = fw::channels::RunWorkQueueMsg::ver { + pipe_type: inner.pipe_type, + work_queue: Some(inner.info.weak_pointer()), + wptr: inner.wptr, + event_slot, + is_new: inner.new, + __pad: Default::default(), + }; + channel.send(&msg); + inner.new = false; + + inner.submit_seq += command_count as u64; + } + + pub(crate) fn pipe_type(&self) -> PipeType { + self.inner.as_ref().expect("No inner?").pipe_type + } + + pub(crate) fn priority(&self) -> u32 { + self.inner.as_ref().expect("No inner?").priority + } +} + +#[versions(AGX)] +impl Drop for Job::ver { + fn drop(&mut self) { + mod_pr_debug!("WorkQueue: Dropping Job\n"); + let mut inner = self.wq.inner.lock(); + + if self.committed && !self.submitted { + let pipe_type = inner.pipe_type; + let event = inner.event.as_mut().expect("Job lost its event"); + mod_pr_debug!( + "WorkQueue({:?}): Roll back {} events (slot {} val {:#x?}) and {} commands\n", + pipe_type, + self.event_count, + event.0.slot(), + event.1, + self.pending.len() + ); + event.1.sub(self.event_count as u32); + inner.commit_seq -= self.pending.len() as u64; + } + + inner.pending_jobs -= 1; + + if inner.pending.is_empty() && inner.pending_jobs == 0 { + mod_pr_debug!("WorkQueue({:?}): Dropping event\n", inner.pipe_type); + inner.event = None; + inner.last_submitted = None; + inner.last_completed = None; + } + mod_pr_debug!("WorkQueue({:?}): Dropped Job\n", inner.pipe_type); + } +} + +#[versions(AGX)] +impl<'a> Drop for JobSubmission::ver<'a> { + fn drop(&mut self) { + let inner = self.inner.as_mut().expect("No inner?"); + mod_pr_debug!("WorkQueue({:?}): Dropping JobSubmission\n", inner.pipe_type); + + let new_len = inner.pending.len() - self.command_count; + inner.pending.truncate(new_len); + + let pipe_type = inner.pipe_type; + let event = inner.event.as_mut().expect("JobSubmission lost its event"); + mod_pr_debug!( + "WorkQueue({:?}): Roll back {} events (slot {} val {:#x?}) and {} commands\n", + pipe_type, + self.event_count, + event.0.slot(), + event.1, + self.command_count + ); + event.1.sub(self.event_count as u32); + inner.commit_seq -= self.command_count as u64; + mod_pr_debug!("WorkQueue({:?}): Dropped JobSubmission\n", inner.pipe_type); + } +} + +#[versions(AGX)] +impl WorkQueueInner::ver { + /// Return the number of free entries in the workqueue + pub(crate) fn free_space(&self) -> usize { + self.size as usize - self.pending.len() - 1 + } + + pub(crate) fn free_slots(&self) -> usize { + let busy_slots = if let Some(ls) = self.last_submitted { + let lc = self + .last_completed + .expect("last_submitted but not completed?"); + ls.delta(&lc) + } else { + 0 + }; + + ((MAX_JOB_SLOTS as i32) - busy_slots).max(0) as usize + } +} + +#[versions(AGX)] +impl WorkQueue::ver { + /// Create a new WorkQueue of a given type and priority. + #[allow(clippy::too_many_arguments)] + pub(crate) fn new( + alloc: &mut gpu::KernelAllocators, + event_manager: Arc, + gpu_context: Arc, + notifier_list: Arc>, + pipe_type: PipeType, + id: u64, + priority: u32, + size: u32, + ) -> Result> { + let mut info = box_in_place!(QueueInfo::ver { + state: alloc.shared.new_default::()?, + ring: alloc.shared.array_empty(size as usize)?, + gpu_buf: alloc.private.array_empty(0x2c18)?, + notifier_list: notifier_list, + gpu_context: gpu_context, + })?; + + info.state.with_mut(|raw, _inner| { + raw.rb_size = size; + }); + + let inner = WorkQueueInner::ver { + event_manager, + info: alloc.private.new_boxed(info, |inner, ptr| { + Ok(place!( + ptr, + raw::QueueInfo::ver { + state: inner.state.gpu_pointer(), + ring: inner.ring.gpu_pointer(), + notifier_list: inner.notifier_list.gpu_pointer(), + gpu_buf: inner.gpu_buf.gpu_pointer(), + gpu_rptr1: Default::default(), + gpu_rptr2: Default::default(), + gpu_rptr3: Default::default(), + event_id: AtomicI32::new(-1), + priority: *raw::PRIORITY.get(priority as usize).ok_or(EINVAL)?, + unk_4c: -1, + uuid: id as u32, + unk_54: -1, + unk_58: Default::default(), + busy: Default::default(), + __pad: Default::default(), + unk_84_state: Default::default(), + unk_88: 0, + unk_8c: 0, + unk_90: 0, + unk_94: 0, + pending: Default::default(), + unk_9c: 0, + #[ver(V >= V13_2)] + unk_a0_0: 0, + gpu_context: inner.gpu_context.gpu_pointer(), + unk_a8: Default::default(), + #[ver(V >= V13_2)] + unk_b0: 0, + } + )) + })?, + new: true, + pipe_type, + size, + wptr: 0, + pending: Vec::new(), + last_token: None, + event: None, + priority, + pending_jobs: 0, + commit_seq: 0, + submit_seq: 0, + last_completed: None, + last_submitted: None, + }; + + let mut queue = Pin::from(UniqueArc::try_new(Self { + info_pointer: inner.info.weak_pointer(), + // SAFETY: `mutex_init!` is called below. + inner: unsafe { Mutex::new(inner) }, + })?); + + // SAFETY: `inner` is pinned when `queue` is. + let pinned = unsafe { queue.as_mut().map_unchecked_mut(|s| &mut s.inner) }; + match pipe_type { + PipeType::Vertex => kernel::mutex_init!(pinned, "WorkQueue::inner (Vertex)"), + PipeType::Fragment => kernel::mutex_init!(pinned, "WorkQueue::inner (Fragment)"), + PipeType::Compute => kernel::mutex_init!(pinned, "WorkQueue::inner (Compute)"), + } + + Ok(queue.into()) + } + + pub(crate) fn event_info(&self) -> Option { + let inner = self.inner.lock(); + + inner.event.as_ref().map(|ev| QueueEventInfo::ver { + stamp_pointer: ev.0.stamp_pointer(), + fw_stamp_pointer: ev.0.fw_stamp_pointer(), + slot: ev.0.slot(), + value: ev.1, + cmd_seq: inner.commit_seq, + info_ptr: self.info_pointer, + }) + } + + pub(crate) fn new_job(self: &Arc) -> Result { + let mut inner = self.inner.lock(); + + if inner.event.is_none() { + mod_pr_debug!("WorkQueue({:?}): Grabbing event\n", inner.pipe_type); + let event = inner.event_manager.get(inner.last_token, self.clone())?; + let cur = event.current(); + inner.last_token = Some(event.token()); + mod_pr_debug!( + "WorkQueue({:?}): Grabbed event slot {}: {:#x?}\n", + inner.pipe_type, + event.slot(), + cur + ); + inner.event = Some((event, cur)); + inner.last_submitted = Some(cur); + inner.last_completed = Some(cur); + } + + inner.pending_jobs += 1; + + let ev = &inner.event.as_ref().unwrap(); + + mod_pr_debug!("WorkQueue({:?}): New job\n", inner.pipe_type); + Ok(Job::ver { + wq: self.clone(), + event_info: QueueEventInfo::ver { + stamp_pointer: ev.0.stamp_pointer(), + fw_stamp_pointer: ev.0.fw_stamp_pointer(), + slot: ev.0.slot(), + value: ev.1, + cmd_seq: inner.commit_seq, + info_ptr: self.info_pointer, + }, + start_value: ev.1, + pending: Vec::new(), + event_count: 0, + committed: false, + submitted: false, + }) + } + + /// Return the number of free entries in the workqueue + pub(crate) fn free_space(&self) -> usize { + self.inner.lock().free_space() + } + + /// Return the number of free job slots in the workqueue + pub(crate) fn free_slots(&self) -> usize { + self.inner.lock().free_slots() + } + + pub(crate) fn pipe_type(&self) -> PipeType { + self.inner.lock().pipe_type + } +} + +/// Trait used to erase the version-specific type of WorkQueues, to avoid leaking +/// version-specificity into the event module. +pub(crate) trait WorkQueue { + fn signal(&self) -> bool; + fn mark_error(&self, value: event::EventValue, error: WorkError); + fn fail_all(&self, error: WorkError); +} + +#[versions(AGX)] +impl WorkQueue for WorkQueue::ver { + /// Signal a workqueue that some work was completed. + /// + /// This will check the event stamp value to find out exactly how many commands were processed. + fn signal(&self) -> bool { + let mut inner = self.inner.lock(); + let event = inner.event.as_ref(); + let value = match event { + None => { + pr_err!("WorkQueue: signal() called but no event?\n"); + return true; + } + Some(event) => event.0.current(), + }; + + inner.last_completed = Some(value); + + mod_pr_debug!( + "WorkQueue({:?}): Signaling event {:?} value {:#x?}\n", + inner.pipe_type, + inner.last_token, + value + ); + + let mut completed_commands: usize = 0; + + for cmd in inner.pending.iter() { + if cmd.value() <= value { + mod_pr_debug!( + "WorkQueue({:?}): Command at value {:#x?} complete\n", + inner.pipe_type, + cmd.value() + ); + completed_commands += 1; + } else { + break; + } + } + + if completed_commands == 0 { + return inner.pending.is_empty(); + } + + let mut completed = Vec::new(); + + if completed.try_reserve(completed_commands).is_err() { + pr_crit!( + "WorkQueue({:?}): Failed to allocated space for {} completed commands\n", + inner.pipe_type, + completed_commands + ); + } + + let pipe_type = inner.pipe_type; + + for cmd in inner.pending.drain(..completed_commands) { + if completed.try_push(cmd).is_err() { + pr_crit!( + "WorkQueue({:?}): Failed to signal a completed command\n", + pipe_type, + ); + } + } + + mod_pr_debug!( + "WorkQueue({:?}): Completed {} commands\n", + inner.pipe_type, + completed_commands + ); + + if let Some(i) = completed.last() { + inner + .info + .state + .with(|raw, _inner| raw.cpu_freeptr.store(i.wptr(), Ordering::Release)); + } + + let empty = inner.pending.is_empty(); + if empty && inner.pending_jobs == 0 { + inner.event = None; + inner.last_submitted = None; + inner.last_completed = None; + } + + core::mem::drop(inner); + + for cmd in completed { + cmd.complete(); + } + + empty + } + + /// Mark this queue's work up to a certain stamp value as having failed. + fn mark_error(&self, value: event::EventValue, error: WorkError) { + // If anything is marked completed, we can consider it successful + // at this point, even if we didn't get the signal event yet. + self.signal(); + + let mut inner = self.inner.lock(); + + if inner.event.is_none() { + pr_err!("WorkQueue: signal_fault() called but no event?\n"); + return; + } + + mod_pr_debug!( + "WorkQueue({:?}): Signaling fault for event {:?} at value {:#x?}\n", + inner.pipe_type, + inner.last_token, + value + ); + + for cmd in inner.pending.iter_mut() { + if cmd.value() <= value { + cmd.mark_error(error); + } else { + break; + } + } + } + + /// Mark all of this queue's work as having failed, and complete it. + fn fail_all(&self, error: WorkError) { + // If anything is marked completed, we can consider it successful + // at this point, even if we didn't get the signal event yet. + self.signal(); + + let mut inner = self.inner.lock(); + + if inner.event.is_none() { + pr_err!("WorkQueue: fail_all() called but no event?\n"); + return; + } + + mod_pr_debug!( + "WorkQueue({:?}): Failing all jobs {:?}\n", + inner.pipe_type, + error + ); + + let mut cmds = Vec::new(); + + core::mem::swap(&mut inner.pending, &mut cmds); + + if inner.pending_jobs == 0 { + inner.event = None; + } + + core::mem::drop(inner); + + for mut cmd in cmds { + cmd.mark_error(error); + cmd.complete(); + } + } +} + +#[versions(AGX)] +impl Drop for WorkQueue::ver { + fn drop(&mut self) { + mod_pr_debug!("WorkQueue({:?}): Dropping\n", self.inner.lock().pipe_type); + } +}