Re: regression/bisected commit 773688a6cb24b0b3c2ba40354d883348a2befa38 make my system completely unusable under high load

[Marco Elver <elver@google.com>]

On Tue, 20 Feb 2024 at 19:51, Marco Elver <elver@google.com> wrote:
>
> On Tue, 20 Feb 2024 at 19:16, Vlastimil Babka <vbabka@suse.cz> wrote:
> >
> > On 2/20/24 18:30, Andrew Morton wrote:
> > > On Tue, 20 Feb 2024 10:37:03 +0500 Mikhail Gavrilov <mikhail.v.gavrilov@gmail.com> wrote:
> > >
> > >> On Tue, Feb 20, 2024 at 4:50 AM Vlastimil Babka <vbabka@suse.cz> wrote:
> > >> > >
> > >> > > I'm all confused.
> > >> > >
> > >> > > 4434a56ec209 ("stackdepot: make fast paths lock-less again") was
> > >> > > mainlined for v6.8-rc3.
> > >> >
> > >> > Uh sorry, I just trusted the info that it's not merged and didn't verify
> > >> > it myself. Yeah, I can see it is there.
> > >> >
> > >>
> > >> Wait, I am talk about these two patches which is not merged yet:
> > >> [PATCH v2 1/2] stackdepot: use variable size records for non-evictable entries
> > >> [PATCH v2 2/2] kasan: revert eviction of stack traces in generic mode
> > >> https://lore.kernel.org/linux-mm/20240129100708.39460-1-elver@google.com/
> > >
> > > A can move those into the 6.8-rc hotfixes queue, and it appears a
> > > cc:stable will not be required.
> > >
> > > However I'm not seeing anything in the changelogs to indicate that
> > > we're fixing a dramatic performance regression, nor why that
> > > regressions is occurring.
>
> It's primarily fixing a regression of memory usage overhead for
> stackdepot users in general. Performance is mostly fixed, but patch
> 2/2 ("kasan: revert eviction of stack traces in generic mode") also
> helps with KASAN performance because entries that were being
> repeatedly evicted-then-reallocated are just allocated once and with
> increasing system uptime the slow path will be taken much less.
>
> > We also seem have an unhappy bot with the 2/2 patch :/ although it's not yet
> > clear if it's a genuine issue.
> >
> > https://lore.kernel.org/all/202402201506.b7e4b9b6-oliver.sang@intel.com/

This was confirmed to be a non-bug by RCU devs.

> While it would be nice if 6.8 would not regress over 6.7 (performance
> is mostly fixed, memory usage is not), waiting for confirmation what
> the rcutorture issue from the bot is about might be good.
>
> Mikhail: since you are testing mainline, in about 4 weeks the fixes
> should then reach 6.9-rc in the next merge window. Until then, if it's
> not too difficult for you, you can apply those 2 patches in your own
> tree.

There are more issues that are fixed by "[PATCH v2 1/2] stackdepot:
use variable size records for non-evictable entries". See
https://lore.kernel.org/all/ZdxYXQdZDuuhcqiv@elver.google.com/

This will eventually reach stable, but it might be good to reconsider
mainlining it earlier.

Thanks,
-- Marco