Re: [patch 00/11] rseq: Optimize exit to user space

[Thomas Gleixner <tglx@linutronix.de>]

On Mon, Aug 18 2025 at 17:29, Michael Jeanson wrote:

> On 2025-08-18 16:21, Thomas Gleixner wrote:
>> That looks about right. Can you reset the branch to
>> 
>>       commit 85b61b265635 ("rseq: Expose stats")
>> 
>> which is just adding primitive stats on top of the current mainline
>> code, and provide numbers for that too?
>> 
>> That gives you 'notify: , cpuid:, fixup:' numbers, which are not 1:1
>> mappable to the final ones, but that should give some interesting
>> insight.
>
> For amd64 kernel and userspace.
>
> Before:
>
> notify:             12467
> fixup:              12467
> cpuid:              12467
>
>
> After:
>
> notify:         123669528
> fixup:          123669528
> cpuid:          123669528

That's insane compared to this:

> exit:          229294046
> signal:               11
> slowp:              4570
> ids:              615950
> cs:              2493682
> clear:            194637
> fixup:           2299044

You can assume that the number of exits (to user) is roughly the same,
i.e. ~23M, so 12M (> 50%) take the TIF_NOTIFY dance on the way out and
most of them for no good reason.

While with the rework only 4.5K go into the NOTIFY slow path and 2.5M
(10 %) do the critical section evaluation and 600k (~ 3%) update CPU/MM
CID.

No suprise that Jens is seeing this in his profiles...

Thanks,

        tglx