[v1,0/6] riscv: support kernel-mode Vector

Message

Andy Chiu July 15, 2023, 3 p.m. UTC

This series provides support for running Vector code in kernel mode. The
implementation is based on the v12 series of the Vector series, but with
some additions. First, we introduce a mechanism to defer restoring
Vector context for userspace programs (patch 1). This is similar to
arm64 and x86's approaches when dealing with extra userspace register
context. And it is benefitial to both Vector in user and kernel-mode.
Then, patch 2, 3 add the kernel-mode Vector patch from v12 with minor
modifications. At the end of the series, patch 4, 5, 6 add supports for
making kernel-mode Vector code preemptible. We do this by adding
kernel-mode Vector context, and keeping track of the frame where V
context is last valid. We believe that enabling preemption of running V
is a critical path for getting V more generally available in the
kernel-mode. Besides, with status.VS, we can easily tell if
saving/restoring V is required. This reduce the level of cost when
running SIMD in kernel mode as compared to other arches. Other arches
usually do not have a way to tell if extra context is dirty. Thus, if
they also want to support running preemptible code with extra registers,
then they must save/restore extra context at each context switch even if
registers are not dirty.

The series is tested by loading a kernel module on a preemptive kernel.
The module launches multiple kworkers which run Vector operations and
verifies with scalar code. Also, the module provides userspace intefaces
via fops to verify if we can run Vector code on syscall path.

Changes from the vector v12 series (for patch 2, 3):
 - return a failure code when kernel_rvv_begin() fails.
 - Do not immediately restore user's V context.


Andy Chiu (4):
  riscv: sched: defer restoring Vector context for user
  riscv: vector: do not pass task_struct into
    riscv_v_vstate_{save,restore}()
  riscv: vector: allow kernel-mode Vector with preemption
  riscv: vector: enable preemptive kernel-mode Vector to be built

Greentime Hu (2):
  riscv: Add support for kernel mode vector
  riscv: Add vector extension XOR implementation

 arch/riscv/Kconfig                     |  10 ++
 arch/riscv/include/asm/entry-common.h  |  13 ++
 arch/riscv/include/asm/processor.h     |   2 +
 arch/riscv/include/asm/thread_info.h   |   6 +
 arch/riscv/include/asm/vector.h        |  50 +++++--
 arch/riscv/include/asm/xor.h           |  82 +++++++++++
 arch/riscv/kernel/Makefile             |   1 +
 arch/riscv/kernel/asm-offsets.c        |   2 +
 arch/riscv/kernel/entry.S              |  41 ++++++
 arch/riscv/kernel/kernel_mode_vector.c | 180 +++++++++++++++++++++++++
 arch/riscv/kernel/process.c            |  10 +-
 arch/riscv/kernel/ptrace.c             |   2 +-
 arch/riscv/kernel/signal.c             |   4 +-
 arch/riscv/kernel/vector.c             |   5 +-
 arch/riscv/lib/Makefile                |   1 +
 arch/riscv/lib/xor.S                   |  81 +++++++++++
 16 files changed, 473 insertions(+), 17 deletions(-)
 create mode 100644 arch/riscv/include/asm/xor.h
 create mode 100644 arch/riscv/kernel/kernel_mode_vector.c
 create mode 100644 arch/riscv/lib/xor.S

Comments

Heiko Stübner July 16, 2023, 9:26 a.m. UTC | #1

Am Samstag, 15. Juli 2023, 17:00:26 CEST schrieb Andy Chiu:
> This series provides support for running Vector code in kernel mode. The
> implementation is based on the v12 series of the Vector series, but with
> some additions. First, we introduce a mechanism to defer restoring
> Vector context for userspace programs (patch 1). This is similar to
> arm64 and x86's approaches when dealing with extra userspace register
> context. And it is benefitial to both Vector in user and kernel-mode.
> Then, patch 2, 3 add the kernel-mode Vector patch from v12 with minor
> modifications. At the end of the series, patch 4, 5, 6 add supports for
> making kernel-mode Vector code preemptible. We do this by adding
> kernel-mode Vector context, and keeping track of the frame where V
> context is last valid. We believe that enabling preemption of running V
> is a critical path for getting V more generally available in the
> kernel-mode. Besides, with status.VS, we can easily tell if
> saving/restoring V is required. This reduce the level of cost when
> running SIMD in kernel mode as compared to other arches. Other arches
> usually do not have a way to tell if extra context is dirty. Thus, if
> they also want to support running preemptible code with extra registers,
> then they must save/restore extra context at each context switch even if
> registers are not dirty.
> 
> The series is tested by loading a kernel module on a preemptive kernel.
> The module launches multiple kworkers which run Vector operations and
> verifies with scalar code. Also, the module provides userspace intefaces
> via fops to verify if we can run Vector code on syscall path.
> 
> Changes from the vector v12 series (for patch 2, 3):
>  - return a failure code when kernel_rvv_begin() fails.
>  - Do not immediately restore user's V context.

This works nicely with my vector crypto patchset rebased on
top of it:

Tested-by: Heiko Stuebner <[e-mail]>