Re: [PATCH v2] x86: bring back rep movsq for user access on CPUs without ERMS

[Mateusz Guzik]

On 9/3/23, Mateusz Guzik <mjguzik@xxxxxxxxx> wrote:
> On Sun, Sep 03, 2023 at 01:08:03PM -0700, Linus Torvalds wrote:
>> On Sun, 3 Sept 2023 at 11:49, Linus Torvalds
>> <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
>> >
>> > So I have no idea why you claim that "currently they have no choice".
>> > glibc is simply being incredibly stupid, and using newfstatat() for no
>> > good reason.
>>
>> Do you have any good benchmark that shows the effects of this?
>>
>> And if you do, does the attached patch (ENTIRELY UNTESTED!) fix the
>> silly glibc mis-feature?
>>
>
> "real fstat" is syscall(5, fd, &sb).
>
> Sapphire Rapids, will-it-scale, ops/s
>
> stock fstat	5088199
> patched fstat	7625244	(+49%)
> real fstat	8540383	(+67% / +12%)
>
> It dodges lockref et al, but it does not dodge SMAP which accounts for
> the difference.
>

I'm going to ask glibc people what's up with the change later.

I'll note statx does not have a dedicated entry point (I checked again
;)) which elides path lookup. It *did* have a way, but it got removed
in 1e2f82d1e9d1 ("statx: Kill fd-with-NULL-path support in favour of
AT_EMPTY_PATH").

One frequent consumer I know of is Rust.

Would be nice to have a non-hacky way to sort it out.

-- 
Mateusz Guzik <mjguzik gmail.com>