Re: [PATCH v9 00/11] KVM: x86/mmu: Age sptes locklessly

[James Houghton]

On Tue, Feb 18, 2025 at 5:13 PM Sean Christopherson <seanjc@xxxxxxxxxx> wrote:
>
> On Tue, Feb 18, 2025, Maxim Levitsky wrote:
> > On Tue, 2025-02-04 at 00:40 +0000, James Houghton wrote:
> > > By aging sptes locklessly with the TDP MMU and the shadow MMU, neither
> > > vCPUs nor reclaim (mmu_notifier_invalidate_range*) will get stuck
> > > waiting for aging. This contention reduction improves guest performance
> > > and saves a significant amount of Google Cloud's CPU usage, and it has
> > > valuable improvements for ChromeOS, as Yu has mentioned previously[1].
> > >
> > > Please see v8[8] for some performance results using
> > > access_tracking_perf_test patched to use MGLRU.
> > >
> > > Neither access_tracking_perf_test nor mmu_stress_test trigger any
> > > splats (with CONFIG_LOCKDEP=y) with the TDP MMU and with the shadow MMU.
> >
> >
> > Hi, I have a question about this patch series and about the
> > access_tracking_perf_test:
> >
> > Some time ago, I investigated a failure in access_tracking_perf_test which
> > shows up in our CI.
> >
> > The root cause was that 'folio_clear_idle' doesn't clear the idle bit when
> > MGLRU is enabled, and overall I got the impression that MGLRU is not
> > compatible with idle page tracking.
> >
> > I thought that this patch series and the 'mm: multi-gen LRU: Have secondary
> > MMUs participate in MM_WALK' patch series could address this but the test
> > still fails.
> >
> >
> > For the reference the exact problem is:
> >
> > 1. Idle bits for guest memory under test are set via /sys/kernel/mm/page_idle/bitmap
> >
> > 2. Guest dirties memory, which leads to A/D bits being set in the secondary mappings.
> >
> > 3. A NUMA autobalance code write protects the guest memory. KVM in response
> >    evicts the SPTE mappings with A/D bit set, and while doing so tells mm
> >    that pages were accessed using 'folio_mark_accessed' (via kvm_set_page_accessed (*) )
> >    but due to MLGRU the call doesn't clear the idle bit and thus all the traces
> >    of the guest access disappear and the kernel thinks that the page is still idle.
> >
> > I can say that the root cause of this is that folio_mark_accessed doesn't do
> > what it supposed to do.
> >
> > Calling 'folio_clear_idle(folio);' in MLGRU case in folio_mark_accessed()
> > will probably fix this but I don't have enough confidence to say if this is
> > all that is needed to fix this.  If this is the case I can send a patch.
>
> My understanding is that the behavior is deliberate.  Per Yu[1], page_idle/bitmap
> effectively isn't supported by MGLRU.
>
> [1] https://lore.kernel.org/all/CAOUHufZeADNp_y=Ng+acmMMgnTR=ZGFZ7z-m6O47O=CmJauWjw@xxxxxxxxxxxxxx

Yu's suggestion was to look at the generation numbers themselves, and
that is exactly what my patched access_tracking_perf_test does[2]. :)

So I think to make this work with MGLRU, I'll re-post my
access_tracking_perf_test patch, but if MGLRU is enabled, always use
the MGLRU debugfs instead of using page_idle/bitmap. It needs some
cleanup first though.

[2]: https://lore.kernel.org/kvm/20241105184333.2305744-12-jthoughton@xxxxxxxxxx/

> > Any ideas on how to fix all this mess?
>
> The easy answer is to skip the test if MGLRU is in use, or if NUMA balancing is
> enabled.  In a real-world scenario, if the guest is actually accessing the pages
> that get PROT_NONE'd by NUMA balancing, they will be marked accessed when they're
> faulted back in.  There's a window where page_idle/bitmap could be read between
> making the VMA PROT_NONE and re-accessing the page from the guest, but IMO that's
> one of the many flaws of NUMA balancing.
>
> That said, one thing is quite odd.  In the failing case, *half* of the guest pages
> are still idle.  That's quite insane.
>
> Aha!  I wonder if in the failing case, the vCPU gets migrated to a pCPU on a
> different node, and that causes NUMA balancing to go crazy and zap pretty much
> all of guest memory.  If that's what's happening, then a better solution for the
> NUMA balancing issue would be to affine the vCPU to a single NUMA node (or hard
> pin it to a single pCPU?).

+1 to this idea, if this is really what's going on. If NUMA balancing
is only migrating a few pages, the 90% threshold in the test should be
low enough that we tolerate the few pages that were moved.

Or we could just print a warning (instead of fail) if NUMA balancing is enabled.