On Thu, Jul 12, 2018 at 11:29:38AM +0800, Yandong.Zhao wrote:
> From: Yandong Zhao <yandong77...@gmail.com>
>
> It does not matter if the caller of may_use_simd() migrates to
> another cpu after the call, but it is still important that the
> kernel_neon_busy percpu instance that is read matches the cpu the
> task is running on at the time of the read.
>
> This means that raw_cpu_read() is not sufficient. kernel_neon_busy
> may appear true if the caller migrates during the execution of
> raw_cpu_read() and the next task to be scheduled in on the initial
> cpu calls kernel_neon_begin().
>
> This patch replaces raw_cpu_read() with this_cpu_read() to protect
> against this race.
>
> Fixes: cb84d11e1625 ("arm64: neon: Remove support for nested or hardirq
> kernel-mode NEON")
> Acked-by: Ard Biesheuvel <ard.biesheu...@linaro.org>
> Reviewed-by: Dave Martin <dave.mar...@arm.com>
> Reviewed-by: Mark Rutland <mark.rutl...@arm..com>
> Signed-off-by: Yandong Zhao <yandong77...@gmail.com>
> ---
> arch/arm64/include/asm/simd.h | 19 +++++++------------
> 1 file changed, 7 insertions(+), 12 deletions(-)
>
> diff --git a/arch/arm64/include/asm/simd.h b/arch/arm64/include/asm/simd.h
> index fa8b3fe..6495cc5 100644
> --- a/arch/arm64/include/asm/simd.h
> +++ b/arch/arm64/include/asm/simd.h
> @@ -29,20 +29,15 @@
> static __must_check inline bool may_use_simd(void)
> {
> /*
> - * The raw_cpu_read() is racy if called with preemption enabled.
> - * This is not a bug: kernel_neon_busy is only set when
> - * preemption is disabled, so we cannot migrate to another CPU
> - * while it is set, nor can we migrate to a CPU where it is set.
> - * So, if we find it clear on some CPU then we're guaranteed to
> - * find it clear on any CPU we could migrate to.
> - *
> - * If we are in between kernel_neon_begin()...kernel_neon_end(),
> - * the flag will be set, but preemption is also disabled, so we
> - * can't migrate to another CPU and spuriously see it become
> - * false.
> + * kernel_neon_busy is only set while preemption is disabled,
> + * and is clear whenever preemption is enabled. Since
> + * this_cpu_read() is atomic w.r.t. preemption, kernel_neon_busy
> + * cannot change under our feet -- if it's set we cannot be
> + * migrated, and if it's clear we cannot be migrated to a CPU
> + * where it is set.
> */
This new explanation looks fine to me.
[...]
Cheers
---Dave
