Does that still work with your test case? (Ontop of -rc1)
---
From: Noah Goldstein <[email protected]>
Date: Tue, 16 Nov 2021 17:14:21 +0100
Subject: [PATCH] x86/fpu: Correct AVX512 state tracking
Add a separate, local mask for tracking AVX512 usage which does not
include the opmask xfeature set. Opmask registers usage does not cause
frequency throttling so it is a completely unnecessary false positive.
While at it, carve it out into a separate function to keep that
abomination extracted out.
[ bp: Rediff and cleanup ontop of 5.16-rc1. ]
Signed-off-by: Noah Goldstein <[email protected]>
Signed-off-by: Borislav Petkov <[email protected]>
Link: https://lore.kernel.org/r/[email protected]
---
arch/x86/kernel/fpu/core.c | 21 ++++++++++++++-------
1 file changed, 14 insertions(+), 7 deletions(-)
diff --git a/arch/x86/kernel/fpu/core.c b/arch/x86/kernel/fpu/core.c
index 8ea306b1bf8e..dd3777ac0443 100644
--- a/arch/x86/kernel/fpu/core.c
+++ b/arch/x86/kernel/fpu/core.c
@@ -98,6 +98,19 @@ bool irq_fpu_usable(void)
}
EXPORT_SYMBOL(irq_fpu_usable);
+/*
+ * Track AVX512 state use because it is known to slow the max clock
+ * speed of the core.
+ */
+static void update_avx_timestamp(struct fpu *fpu)
+{
+
+#define AVX512_TRACKING_MASK (XFEATURE_MASK_ZMM_Hi256 | XFEATURE_MASK_Hi16_ZMM)
+
+ if (fpu->fpstate->regs.xsave.header.xfeatures & AVX512_TRACKING_MASK)
+ fpu->avx512_timestamp = jiffies;
+}
+
/*
* Save the FPU register state in fpu->fpstate->regs. The register state is
* preserved.
@@ -116,13 +129,7 @@ void save_fpregs_to_fpstate(struct fpu *fpu)
{
if (likely(use_xsave())) {
os_xsave(fpu->fpstate);
-
- /*
- * AVX512 state is tracked here because its use is
- * known to slow the max clock speed of the core.
- */
- if (fpu->fpstate->regs.xsave.header.xfeatures & XFEATURE_MASK_AVX512)
- fpu->avx512_timestamp = jiffies;
+ update_avx_timestamp(fpu);
return;
}
--
2.29.2
--
Regards/Gruss,
Boris.
https://people.kernel.org/tglx/notes-about-netiquette
On Tue, Nov 16, 2021 at 10:21 AM Borislav Petkov <[email protected]> wrote:
>
> Does that still work with your test case? (Ontop of -rc1)
This should work.
>
> ---
> From: Noah Goldstein <[email protected]>
> Date: Tue, 16 Nov 2021 17:14:21 +0100
> Subject: [PATCH] x86/fpu: Correct AVX512 state tracking
>
> Add a separate, local mask for tracking AVX512 usage which does not
> include the opmask xfeature set. Opmask registers usage does not cause
> frequency throttling so it is a completely unnecessary false positive.
>
> While at it, carve it out into a separate function to keep that
> abomination extracted out.
>
> [ bp: Rediff and cleanup ontop of 5.16-rc1. ]
>
> Signed-off-by: Noah Goldstein <[email protected]>
> Signed-off-by: Borislav Petkov <[email protected]>
> Link: https://lore.kernel.org/r/[email protected]
> ---
> arch/x86/kernel/fpu/core.c | 21 ++++++++++++++-------
> 1 file changed, 14 insertions(+), 7 deletions(-)
>
> diff --git a/arch/x86/kernel/fpu/core.c b/arch/x86/kernel/fpu/core.c
> index 8ea306b1bf8e..dd3777ac0443 100644
> --- a/arch/x86/kernel/fpu/core.c
> +++ b/arch/x86/kernel/fpu/core.c
> @@ -98,6 +98,19 @@ bool irq_fpu_usable(void)
> }
> EXPORT_SYMBOL(irq_fpu_usable);
>
> +/*
> + * Track AVX512 state use because it is known to slow the max clock
> + * speed of the core.
> + */
> +static void update_avx_timestamp(struct fpu *fpu)
> +{
> +
> +#define AVX512_TRACKING_MASK (XFEATURE_MASK_ZMM_Hi256 | XFEATURE_MASK_Hi16_ZMM)
> +
> + if (fpu->fpstate->regs.xsave.header.xfeatures & AVX512_TRACKING_MASK)
> + fpu->avx512_timestamp = jiffies;
> +}
> +
> /*
> * Save the FPU register state in fpu->fpstate->regs. The register state is
> * preserved.
> @@ -116,13 +129,7 @@ void save_fpregs_to_fpstate(struct fpu *fpu)
> {
> if (likely(use_xsave())) {
> os_xsave(fpu->fpstate);
> -
> - /*
> - * AVX512 state is tracked here because its use is
> - * known to slow the max clock speed of the core.
> - */
> - if (fpu->fpstate->regs.xsave.header.xfeatures & XFEATURE_MASK_AVX512)
> - fpu->avx512_timestamp = jiffies;
> + update_avx_timestamp(fpu);
> return;
> }
>
> --
> 2.29.2
>
>
> --
> Regards/Gruss,
> Boris.
>
> https://people.kernel.org/tglx/notes-about-netiquette