Re: [PATCH v3 4/6] KVM: arm64/mmu: count KVM page table pages in pagetable stats

[Yosry Ahmed]

On Thu, Apr 28, 2022 at 10:45 AM Oliver Upton <oupton@xxxxxxxxxx> wrote:
>
> On Tue, Apr 26, 2022 at 12:27:57PM -0700, Yosry Ahmed wrote:
> > > What page tables do we want to account? KVM on ARM manages several page
> > > tables.
> > >
> > > For regular KVM, the host kernel manages allocations for the hyp stage 1
> > > tables in addition to the stage 2 tables used for a particular VM. The
> > > former is system overhead whereas the latter could be attributed to a
> > > guest VM.
> >
> > Honestly I would love to get your input on this. The main motivation
> > here is to give users insights on the kernel memory usage on their
> > system (or in a cgroup). We currently have NR_PAGETABLE stats for
> > normal kernel page tables (allocated using
> > __pte_alloc_one()/pte_free()), this shows up in /proc/meminfo,
> > /path/to/cgroup/memory.stat, and node stats. The idea is to add
> > NR_SECONDARY_PAGETABLE that should include the memory used for kvm
> > pagetables, which should be a separate category (no overlap). What
> > gets included or not depends on the semantics of KVM and what exactly
> > falls under the category of secondary pagetables from the user's pov.
> >
> > Currently it looks like s2 page table allocations get accounted to
> > kmem of memory control groups (GFP_KERNEL_ACCOUNT), while hyp page
> > table allocations do not (GFP_KERNEL). So we could either follow this
> > and only account s2 page table allocations in the stats, or make hyp
> > allocations use GFP_KERNEL_ACCOUNT as well and add them to the stats.
> > Let me know what you think.
>
> I think it is reasonable to just focus on stage 2 table allocations and
> ignore all else. As Marc pointed out it isn't workable in other
> contexts anyway (pKVM), and keeps the patch tidy too.
>
> GFP_KERNEL_ACCOUNT for hyp allocations wouldn't make sense, as it is
> done at init to build out the system page tables for EL2.

Thanks so much for the insights, will send out v4 according to our discussion.

>
> --
> Thanks,
> Oliver