| Message ID | 20240928085727.56883-1-phil@philjordan.eu (mailing list archive) |
|---|---|
| Headers | show |
| Series | macOS PV Graphics and new vmapple machine type | expand |
Phil Dennis-Jordan <phil@philjordan.eu> writes: > (Apologies to anyone who has received more than one version of this > series of emails; my git-send-email was misconfigured and this is > a new attempt.) > > This patch set introduces a new ARM and macOS HVF specific machine type > called "vmapple", as well as a family of display devices based on the > ParavirtualizedGraphics.framework in macOS. One of the display adapter > variants, apple-gfx-vmapple, is required for the new machine type, while > apple-gfx-pci can be used to enable 3D graphics acceleration with x86-64 > macOS guest OSes. > <snip> > > There are currently a few limitations to this which aren't intrinsic, > just imperfect emulation of the VZF, but it's good enough to be just > about usable for some purposes: > > * macOS 12 guests only. Versions 13+ currently fail during early boot. > * macOS 11+ arm64 hosts only, with hvf accel. (Perhaps some differences > between Apple M series CPUs and TCG's aarch64 implementation? macOS > hosts only because ParavirtualizedGraphics.framework is a black box > implementing most of the logic behind the apple-gfx device.) We don't currently have TCG CPU models for the Apple Silicon processors. They are not too hard to add (basically setting the correct ID register bits, c.f. aarch64_neoverse_n1_initfn for an example). However that would only cover Aarch64 architectural features. We do no modelling of the extra instructions that Apple added (although in theory that should only be run in Apples own ML libraries).
On Thu, 3 Oct 2024 at 10:06, Alex Bennée <alex.bennee@linaro.org> wrote: > > > There are currently a few limitations to this which aren't intrinsic, > > just imperfect emulation of the VZF, but it's good enough to be just > > about usable for some purposes: > > > > * macOS 12 guests only. Versions 13+ currently fail during early boot. > > * macOS 11+ arm64 hosts only, with hvf accel. (Perhaps some differences > > between Apple M series CPUs and TCG's aarch64 implementation? macOS > > hosts only because ParavirtualizedGraphics.framework is a black box > > implementing most of the logic behind the apple-gfx device.) > > We don't currently have TCG CPU models for the Apple Silicon processors. > They are not too hard to add (basically setting the correct ID register > bits, c.f. aarch64_neoverse_n1_initfn for an example). However that > would only cover Aarch64 architectural features. We do no modelling of > the extra instructions that Apple added (although in theory that should > only be run in Apples own ML libraries). > This really isn't my area of expertise, and I don't see myself attempting to make it work with TCG. Given that the OS only boots with the PV graphics device, you can only really use this machine type on a macOS host, so there aren't many reasons to use TCG over HVF. I suppose it might make debugging the myriad other rough edges easier!
(Apologies to anyone who has received more than one version of this series of emails; my git-send-email was misconfigured and this is a new attempt.) This patch set introduces a new ARM and macOS HVF specific machine type called "vmapple", as well as a family of display devices based on the ParavirtualizedGraphics.framework in macOS. One of the display adapter variants, apple-gfx-vmapple, is required for the new machine type, while apple-gfx-pci can be used to enable 3D graphics acceleration with x86-64 macOS guest OSes. Previous versions of this patch set were submitted semi-separately: the original vmapple patch set by Alexander Graf included a monolithic implementation of apple-gfx-vmapple. I subsequently reviewed and reworked the latter to support the PCI variant of the device as well and submitted the result in isolation. As requested in subsequent review, I have now recombined this with the original vmapple patch set, which I have updated and improved in a few ways as well. The vmapple machine type approximates the configuration in macOS's own Virtualization.framework when running arm64 macOS guests. In addition to generic components such as a GICv3 and an XHCI USB controller, it includes nonstandard extensions to the virtio block device, a special "hardware" aes engine, a configuration device, a pvpanic variant, a "backdoor" interface, and of course the apple-gfx paravirtualised display adapter. There are currently a few limitations to this which aren't intrinsic, just imperfect emulation of the VZF, but it's good enough to be just about usable for some purposes: * macOS 12 guests only. Versions 13+ currently fail during early boot. * macOS 11+ arm64 hosts only, with hvf accel. (Perhaps some differences between Apple M series CPUs and TCG's aarch64 implementation? macOS hosts only because ParavirtualizedGraphics.framework is a black box implementing most of the logic behind the apple-gfx device.) * PCI devices use legacy IRQs, not MSI/MSI-X. As far as I can tell, we'd need to include the GICv3 ITS, but it's unclear to me what exactly needs wiring up. * Due to lack of MSI(-X), event delivery from USB devices to the guest macOS isn't working correctly. My current conclusion is that the OS's XHCI driver simply was never designed to work with legacy IRQs. The upshot is that keyboard and mouse/tablet input is very laggy. The solution would be to implement MSI(-X) support or figure out how to make hcd-xhci-sysbus work with the macOS guest, if at all possible. (EHCI and UHCI/OHCI controllers are not an option as the VMAPPLE guest kernel does not include drivers for these.) * The guest OS must first be provisioned using Virtualization.framework; the disk images can subsequently be used in Qemu. (See docs.) The apple-gfx device can be used independently from the vmapple machine type, at least in the PCI variant. It mainly targets x86-64 macOS guests from version 11 on, but also includes a UEFI bootrom for basic framebuffer mode. macOS 11 is also required on the host side, as well as a GPU that supports the Metal API. On the guest side, this provides 3D acceleration/GPGPU support with a baseline Metal feature set, irrespective of the host GPU's feature set. A few limitations in the current integration: * Although it works fine with TCG, it does not work correctly cross-architecture: x86-64 guests on arm64 hosts appear to make some boot progress, but rendering is corrupted. I suspect incompatible texture memory layouts; I have no idea if this is fixable. * ParavirtualizedGraphics.framework and the guest driver support multi-headed configurations. The current Qemu integration always connects precisely 1 display. * State serialisation and deserialisation is currently not implemented, though supported in principle by the framework. Both apple-gfx variants thus set up a migration blocker. * Rendering efficiency could be better. The GPU-rendered guest framebuffer is copied to system memory and uses Qemu's usual CPU-based drawing. For maximum efficiency, the Metal texture containing the guest framebuffer could be drawn directly to a Metal view in the host window, staying on the GPU. (Similar to the OpenGL/virgl render path on other platforms.) --- v2 -> v3: * Merged the apple-gfx and vmapple patchsets. * Squashed a bunch of later apple-gfx patches into the main one. (dGPU support, queried MMIO area size, host GPU picking logic.) * Rebased on latest upstream, fixing any breakages due to internal Qemu API changes. * apple-gfx: Switched to re-entrant MMIO. This is supported by the underlying framework and simplifies the MMIO forwarding code which was previously different on x86-64 vs aarch64. * vmapple: Fixes for minor bugs and comments from the last round of review. * vmapple aes, conf, apple-gfx: Switched reset methods to implement the ResettableClass base's interface. * vmapple: switched from virtio-hid to an XHCI USB controller and USB mouse and tablet devices. macOS does not provide drivers for virtio HID devices, at least not in version 12's vmapple kernel. So input now sort of works (interrupt issues) rather than not at all. Use network-based remote access to the guest OS as a work-around. Alexander Graf (9): hw: Add vmapple subdir hw/misc/pvpanic: Add MMIO interface hvf: arm: Ignore writes to CNTP_CTL_EL0 gpex: Allow more than 4 legacy IRQs hw/vmapple/aes: Introduce aes engine hw/vmapple/bdif: Introduce vmapple backdoor interface hw/vmapple/cfg: Introduce vmapple cfg region hw/vmapple/virtio-blk: Add support for apple virtio-blk hw/vmapple/vmapple: Add vmapple machine type Phil Dennis-Jordan (5): hw/display/apple-gfx: Introduce ParavirtualizedGraphics.Framework support hw/display/apple-gfx: Adds PCI implementation ui/cocoa: Adds non-app runloop on main thread mode hw/display/apple-gfx: Adds configurable mode list MAINTAINERS: Add myself as maintainer for apple-gfx, reviewer for HVF MAINTAINERS | 15 + docs/system/arm/vmapple.rst | 63 +++ docs/system/target-arm.rst | 1 + hw/Kconfig | 1 + hw/arm/sbsa-ref.c | 2 +- hw/arm/virt.c | 2 +- hw/block/virtio-blk.c | 19 +- hw/display/Kconfig | 14 + hw/display/apple-gfx-pci.m | 179 +++++++++ hw/display/apple-gfx-vmapple.m | 215 ++++++++++ hw/display/apple-gfx.h | 72 ++++ hw/display/apple-gfx.m | 668 ++++++++++++++++++++++++++++++++ hw/display/meson.build | 3 + hw/display/trace-events | 26 ++ hw/i386/microvm.c | 2 +- hw/loongarch/virt.c | 2 +- hw/meson.build | 1 + hw/mips/loongson3_virt.c | 2 +- hw/misc/Kconfig | 4 + hw/misc/meson.build | 1 + hw/misc/pvpanic-mmio.c | 61 +++ hw/openrisc/virt.c | 12 +- hw/pci-host/gpex.c | 36 +- hw/riscv/virt.c | 12 +- hw/vmapple/Kconfig | 32 ++ hw/vmapple/aes.c | 584 ++++++++++++++++++++++++++++ hw/vmapple/bdif.c | 245 ++++++++++++ hw/vmapple/cfg.c | 106 +++++ hw/vmapple/meson.build | 5 + hw/vmapple/trace-events | 26 ++ hw/vmapple/trace.h | 1 + hw/vmapple/virtio-blk.c | 212 ++++++++++ hw/vmapple/vmapple.c | 661 +++++++++++++++++++++++++++++++ hw/xtensa/virt.c | 2 +- include/hw/misc/pvpanic.h | 1 + include/hw/pci-host/gpex.h | 7 +- include/hw/pci/pci_ids.h | 1 + include/hw/virtio/virtio-blk.h | 12 +- include/hw/vmapple/bdif.h | 31 ++ include/hw/vmapple/cfg.h | 68 ++++ include/hw/vmapple/virtio-blk.h | 39 ++ include/qemu-main.h | 2 + meson.build | 5 + target/arm/hvf/hvf.c | 9 + ui/cocoa.m | 15 +- 45 files changed, 3443 insertions(+), 34 deletions(-) create mode 100644 docs/system/arm/vmapple.rst create mode 100644 hw/display/apple-gfx-pci.m create mode 100644 hw/display/apple-gfx-vmapple.m create mode 100644 hw/display/apple-gfx.h create mode 100644 hw/display/apple-gfx.m create mode 100644 hw/misc/pvpanic-mmio.c create mode 100644 hw/vmapple/Kconfig create mode 100644 hw/vmapple/aes.c create mode 100644 hw/vmapple/bdif.c create mode 100644 hw/vmapple/cfg.c create mode 100644 hw/vmapple/meson.build create mode 100644 hw/vmapple/trace-events create mode 100644 hw/vmapple/trace.h create mode 100644 hw/vmapple/virtio-blk.c create mode 100644 hw/vmapple/vmapple.c create mode 100644 include/hw/vmapple/bdif.h create mode 100644 include/hw/vmapple/cfg.h create mode 100644 include/hw/vmapple/virtio-blk.h