Message ID | 20230518161949.11203-25-andy.chiu@sifive.com (mailing list archive) |
---|---|
State | Superseded |
Headers | show |
Series | riscv: Add vector ISA support | expand |
Context | Check | Description |
---|---|---|
conchuod/cover_letter | success | Series has a cover letter |
conchuod/tree_selection | success | Guessed tree name to be for-next at HEAD ac9a78681b92 |
conchuod/fixes_present | success | Fixes tag not required for -next series |
conchuod/maintainers_pattern | success | MAINTAINERS pattern errors before the patch: 6 and now 6 |
conchuod/verify_signedoff | success | Signed-off-by tag matches author and committer |
conchuod/kdoc | success | Errors and warnings before: 0 this patch: 0 |
conchuod/build_rv64_clang_allmodconfig | success | Errors and warnings before: 8 this patch: 8 |
conchuod/module_param | success | Was 0 now: 0 |
conchuod/build_rv64_gcc_allmodconfig | success | Errors and warnings before: 8 this patch: 8 |
conchuod/build_rv32_defconfig | success | Build OK |
conchuod/dtb_warn_rv64 | success | Errors and warnings before: 3 this patch: 3 |
conchuod/header_inline | success | No static functions without inline keyword in header files |
conchuod/checkpatch | warning | WARNING: added, moved or deleted file(s), does MAINTAINERS need updating? |
conchuod/build_rv64_nommu_k210_defconfig | success | Build OK |
conchuod/verify_fixes | success | No Fixes tag |
conchuod/build_rv64_nommu_virt_defconfig | success | Build OK |
On 5/18/23 23:19, Andy Chiu wrote: > This patch add a brief documentation of the userspace interface in > regard to the RISC-V Vector extension. > "Add a brief documentation ..." > Signed-off-by: Andy Chiu <andy.chiu@sifive.com> > Reviewed-by: Greentime Hu <greentime.hu@sifive.com> > Reviewed-by: Vincent Chen <vincent.chen@sifive.com> > Co-developed-by: Bagas Sanjaya <bagasdotme@gmail.com> > Signed-off-by: Bagas Sanjaya <bagasdotme@gmail.com> > --- > Changelog v20: > - Drop bit-field repressentation and typos (Björn) > - Fix document styling (Bagas) Anyway, thanks for applying my fixups!
Hi, Le torstaina 18. toukokuuta 2023 19.19.47 EEST, vous avez écrit : > This patch add a brief documentation of the userspace interface in > regard to the RISC-V Vector extension. > > Signed-off-by: Andy Chiu <andy.chiu at sifive.com> > Reviewed-by: Greentime Hu <greentime.hu at sifive.com> > Reviewed-by: Vincent Chen <vincent.chen at sifive.com> > Co-developed-by: Bagas Sanjaya <bagasdotme at gmail.com> > Signed-off-by: Bagas Sanjaya <bagasdotme at gmail.com> > --- > Changelog v20: > - Drop bit-field repressentation and typos (Bj?rn) > - Fix document styling (Bagas) > --- > Documentation/riscv/index.rst | 1 + > Documentation/riscv/vector.rst | 120 +++++++++++++++++++++++++++++++++ > 2 files changed, 121 insertions(+) > create mode 100644 Documentation/riscv/vector.rst > > diff --git a/Documentation/riscv/index.rst b/Documentation/riscv/index.rst > index 175a91db0200..95cf9c1e1da1 100644 > --- a/Documentation/riscv/index.rst > +++ b/Documentation/riscv/index.rst > @@ -10,6 +10,7 @@ RISC-V architecture > hwprobe > patch-acceptance > uabi > + vector > > features > > diff --git a/Documentation/riscv/vector.rst b/Documentation/riscv/vector.rst > new file mode 100644 > index 000000000000..5d37fd212720 > --- /dev/null > +++ b/Documentation/riscv/vector.rst > @@ -0,0 +1,120 @@ > +.. SPDX-License-Identifier: GPL-2.0 > + > +========================================= > +Vector Extension Support for RISC-V Linux > +========================================= > + > +This document briefly outlines the interface provided to userspace by Linux > in +order to support the use of the RISC-V Vector Extension. > + > +1. prctl() Interface > +--------------------- > + > +Two new prctl() calls are added to allow programs to manage the enablement > +status for the use of Vector in userspace: > + > +* prctl(PR_RISCV_V_SET_CONTROL, unsigned long arg) > + > + Sets the Vector enablement status of the calling thread, where the > control > + argument consists of two 2-bit enablement statuses and a bit > for inheritance > + mode. Other threads of the calling process are > unaffected. I somewhat wonder who is/are the intended users of this new prctl(). Are they the run-time (libc)? The main program? Libraries using RVV internally (think OpenSSL, Nettle, FFmpeg, etc)? The init system? Library code doesn't typically know how stacks are allocated and how signal are handled (on alternate or normal stacks), since signal handlers are process-global state. So I figure that libraries should keep off off this one. Conversely, it would be impractical for programs to call a Linux-specific RISC- V-specific in or around their main(). And then libc presumably should not override the configured policy that comes from sysctl or from the parent process. So I guess that that leaves just the init system (in a broad sense) then? In any case, I think the intended use should be clarified with proper usage guidelines. Otherwise, what I bet happens is RVV-capable libraries just blindly invoke the prctl() to "enable RVV", deafeating the purpose of having the prctl() in the first place.
diff --git a/Documentation/riscv/index.rst b/Documentation/riscv/index.rst index 175a91db0200..95cf9c1e1da1 100644 --- a/Documentation/riscv/index.rst +++ b/Documentation/riscv/index.rst @@ -10,6 +10,7 @@ RISC-V architecture hwprobe patch-acceptance uabi + vector features diff --git a/Documentation/riscv/vector.rst b/Documentation/riscv/vector.rst new file mode 100644 index 000000000000..5d37fd212720 --- /dev/null +++ b/Documentation/riscv/vector.rst @@ -0,0 +1,120 @@ +.. SPDX-License-Identifier: GPL-2.0 + +========================================= +Vector Extension Support for RISC-V Linux +========================================= + +This document briefly outlines the interface provided to userspace by Linux in +order to support the use of the RISC-V Vector Extension. + +1. prctl() Interface +--------------------- + +Two new prctl() calls are added to allow programs to manage the enablement +status for the use of Vector in userspace: + +* prctl(PR_RISCV_V_SET_CONTROL, unsigned long arg) + + Sets the Vector enablement status of the calling thread, where the control + argument consists of two 2-bit enablement statuses and a bit for inheritance + mode. Other threads of the calling process are unaffected. + + Enablement status is a tri-state value each occupying 2-bit of space in + the control argument: + + * :c:macro:`PR_RISCV_V_VSTATE_CTRL_DEFAULT`: Use the system-wide default + enablement status on execve(). The system-wide default setting can be + controlled via sysctl interface (see sysctl section below). + + * :c:macro:`PR_RISCV_V_VSTATE_CTRL_ON`: Allow Vector to be run for the + thread. + + * :c:macro:`PR_RISCV_V_VSTATE_CTRL_OFF`: Disallow Vector. Executing Vector + instructions under such condition will trap and casuse the termination of the thread. + + arg: The control argument is a 5-bit value consisting of 3 parts, and + accessed by 3 masks respectively. + + The 3 masks, PR_RISCV_V_VSTATE_CTRL_CUR_MASK, + PR_RISCV_V_VSTATE_CTRL_NEXT_MASK, and PR_RISCV_V_VSTATE_CTRL_INHERIT + represents bit[1:0], bit[3:2], and bit[4]. bit[1:0] accounts for the + enablement status of current thread, and the setting at bit[3:2] takes place + at next execve(). bit[4] defines the inheritance mode of the setting in + bit[3:2]. + + * :c:macro:`PR_RISCV_V_VSTATE_CTRL_CUR_MASK`: bit[1:0]: Account for the + Vector enablement status for the calling thread. The calling thread is + not able to turn off Vector once it has been enabled. The prctl() call + fails with EPERM if the value in this mask is PR_RISCV_V_VSTATE_CTRL_OFF + but the current enablement status is not off. Setting + PR_RISCV_V_VSTATE_CTRL_DEFAULT here takes no effect but to set back + the original enablement status. + + * :c:macro:`PR_RISCV_V_VSTATE_CTRL_NEXT_MASK`: bit[3:2]: Account for the + Vector enablement setting for the calling thread at the next execve() + system call. If PR_RISCV_V_VSTATE_CTRL_DEFAULT is used in this mask, + then the enablement status will be decided by the system-wide + enablement status when execve() happen. + + * :c:macro:`PR_RISCV_V_VSTATE_CTRL_INHERIT`: bit[4]: the inheritance + mode for the setting at PR_RISCV_V_VSTATE_CTRL_NEXT_MASK. If the bit + is set then the following execve() will not clear the setting in both + PR_RISCV_V_VSTATE_CTRL_NEXT_MASK and PR_RISCV_V_VSTATE_CTRL_INHERIT. + This setting persists across changes in the system-wide default value. + + Return value: + * 0 on success; + * EINVAL: Vector not supported, invalid enablement status for current or + next mask; + * EPERM: Turning off Vector in PR_RISCV_V_VSTATE_CTRL_CUR_MASK if Vector + was enabled for the calling thread. + + On success: + * A valid setting for PR_RISCV_V_VSTATE_CTRL_CUR_MASK takes place + immediately. The enablement status specified in + PR_RISCV_V_VSTATE_CTRL_NEXT_MASK happens at the next execve() call, or + all following execve() calls if PR_RISCV_V_VSTATE_CTRL_INHERIT bit is + set. + * Every successful call overwrites a previous setting for the calling + thread. + +* prctl(PR_RISCV_V_GET_CONTROL) + + Gets the same Vector enablement status for the calling thread. Setting for + next execve() call and the inheritance bit are all OR-ed together. + + Return value: + * a nonnegative value on success; + * EINVAL: Vector not supported. + +2. System runtime configuration (sysctl) +----------------------------------------- + +To mitigate the ABI impact of expansion of the signal stack, a +policy mechanism is provided to the administrators, distro maintainers, and +developers to control the default Vector enablement status for userspace +processes in form of sysctl knob: + +* /proc/sys/abi/riscv_v_default_allow + + Writing the text representation of 0 or 1 to this file sets the default + system enablement status for new starting userspace programs. Valid values + are: + + * 0: Do not allow Vector code to be executed as the default for new processes. + * 1: Allow Vector code to be executed as the default for new processes. + + Reading this file returns the current system default enablement status. + + At every execve() call, a new enablement status of the new process is set to + the system default, unless: + + * PR_RISCV_V_VSTATE_CTRL_INHERIT is set for the calling process, and the + setting in PR_RISCV_V_VSTATE_CTRL_NEXT_MASK is not + PR_RISCV_V_VSTATE_CTRL_DEFAULT. Or, + + * The setting in PR_RISCV_V_VSTATE_CTRL_NEXT_MASK is not + PR_RISCV_V_VSTATE_CTRL_DEFAULT. + + Modifying the system default enablement status does not affect the enablement + status of any existing process of thread that do not make an execve() call.