Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
MIME-Version: 1.0
References: <20210920053951.4093668-1-goldstein.w.n@gmail.com>
 <CAFUsyfJWXJEc1879uX_YOJg7cvie0gitv+bu83DiUXiE74uQww@mail.gmail.com>
 <CAFUsyfJQ8_=QnZjZ2zS45wwB6AaRgL4JO8_sT1a_rXUsfh_1WQ@mail.gmail.com> <YWfqEIaCKOwSWGHF@zn.tnic>
In-Reply-To: <YWfqEIaCKOwSWGHF@zn.tnic>
From:   Noah Goldstein <goldstein.w.n@gmail.com>
Date:   Thu, 14 Oct 2021 10:49:25 -0500
Message-ID: <CAFUsyfLf5a6-L7f2AjLx3j8+qbG7EU9iLPJCTA0+UoOUg3C61A@mail.gmail.com>
Subject: Re: [PATCH v1] x86/fpu: Remove opmask state from avx512_timestamp check
To:     Borislav Petkov <bp@alien8.de>
Cc:     tglx@linutronix.de, mingo@redhat.com, X86 ML <x86@kernel.org>,
        hpa@zytor.com, Andy Lutomirski <luto@kernel.org>,
        open list <linux-kernel@vger.kernel.org>
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Thu, Oct 14, 2021 at 3:28 AM Borislav Petkov <bp@alien8.de> wrote:
>
> On Wed, Oct 13, 2021 at 05:36:14PM -0500, Noah Goldstein wrote:
> > Ping2
>
> Why?


This left me confused and annoyed for several hours because
the avx512 usage information on my applications didn't make
sense. Figured worth a patch for posterity.

>
> The original patch which added this abomination:
>
> 2f7726f95557 ("x86/fpu: Track AVX-512 usage of tasks")
>
> says already:
>
>  the tracking mechanism is imprecise and can theoretically miss AVX-512
>  usage during context switch. But it has been measured to be precise
>  enough to be useful under real-world workloads like tensorflow and
>  linpack.
>
>  If higher precision is required, suggest user space tools to use the
>  PMU-based mechanisms in combination.

This patch isn't about higher precision / false negatives.
It's about false positives.

>
> and as you've noticed, the high 16 regs would cause a false positive
> too.
>
> So what is the actual real-life use case for this?

No big project. There is a case for avoiding the extended avx512
registers (reg16..31) for various reasons (context switches,
code size, instruction limitations) that don't apply to mask registers.

Irrelevant of the still existing flaws, it makes the output more accurate.

Is there a cost to the change I am not seeing?

If not it seems to be a tradeoff between more accurate vs less accurate.
So why not take the former?

>
> --
> Regards/Gruss,
>     Boris.
>
> https://people.kernel.org/tglx/notes-about-netiquette