Re: [PATCH v3 8/8] xor: Add AVX-512 optimized xor_gen()

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
- [Medium] Missing `vzeroupper` instruction after using AVX-512 (ZMM) registers, leading to severe AVX-to-SSE transition penalties in kernel threads.
--

--- Patch [8]: [PATCH v3 8/8] xor: Add AVX-512 optimized xor_gen() ---
commit 0074b2adffd00d86b2a03e4d30495b21794ecb20
Author: Eric Biggers <[email protected]>

xor: Add AVX-512 optimized xor_gen()

This commit introduces an AVX-512 optimized implementation of xor_gen()
using 512-bit ZMM registers and the vpternlogq instruction. It is enabled
on x86_64 CPUs with AVX512F that do not have the PREFER_YMM feature set.

> diff --git a/lib/raid/xor/x86/xor-avx512.c b/lib/raid/xor/x86/xor-avx512.c
> new file mode 100644
> index 0000000000000..17f57900d8274
> --- /dev/null
> +++ b/lib/raid/xor/x86/xor-avx512.c
[ ... ]
> +static void xor_gen_avx512(void *dest, void **srcs, unsigned int src_cnt,
> +			   unsigned int bytes)
> +{
> +	kernel_fpu_begin();
> +	xor_gen_avx512_inner(dest, srcs, src_cnt, bytes);

[Severity: Medium]
Does this code need a vzeroupper instruction before releasing the FPU to avoid
AVX-to-SSE transition penalties? If a kernel thread modifies the ZMM registers
here and subsequently executes legacy SSE instructions (for example, in kcryptd)
without an intervening user-space context switch to clear the state, the CPU
could incur a severe transition penalty on affected architectures.

While this problem already existed in the xor-avx.c implementation using YMM
registers, the newly introduced xor-avx512.c appears to repeat the defect.

> +	kernel_fpu_end();
> +}

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=8