On Tue, 27 Feb 2024 at 09:20, Kent Overstreet <kent.overstreet@xxxxxxxxx> wrote:
>
> fio \
> --group_reporting \
[..]
Ok, I have to say that my kernel profile looks very nice with that
patch. It all looks almost perfect, and the main big spike in the
profile is the inevitable "clac/stac" pair around the user copy
because my old threadripper just ends up not doing that very well.
Of course, I disable all the mitigations for my profiling runs,
because not doing that is just too depressing and any real kernel
issues that we can do something about get hidden by the noise
introduced by the mitigations. So my profiles are entirely artificial
in that sense (I tell myself that it's what future CPU's will look
like and that it's ok. Then I cry myself to sleep).
I'm not going to push that patch of mine - I still suspect that doing
this all under RCU with page faults optimistically disabled would be a
much nicer model and work for bigger areas - but I don't hate the
patch either, and if somebody else wants to push it, I'm ok with it.
I can't actually think of any real load that does a lot of small
reads. Willy's odd customer not-withstanding.
Linus
