Re: [PATCH v3 5/7] KVM: x86: Participate in bitmap-based PTE aging

[James Houghton <jthoughton@google.com>]

On Thu, Apr 11, 2024 at 10:28 AM David Matlack <dmatlack@google.com> wrote:
>
> On 2024-04-11 10:08 AM, David Matlack wrote:
> > On 2024-04-01 11:29 PM, James Houghton wrote:
> > > Only handle the TDP MMU case for now. In other cases, if a bitmap was
> > > not provided, fallback to the slowpath that takes mmu_lock, or, if a
> > > bitmap was provided, inform the caller that the bitmap is unreliable.
> > >
> > > Suggested-by: Yu Zhao <yuzhao@google.com>
> > > Signed-off-by: James Houghton <jthoughton@google.com>
> > > ---
> > >  arch/x86/include/asm/kvm_host.h | 14 ++++++++++++++
> > >  arch/x86/kvm/mmu/mmu.c          | 16 ++++++++++++++--
> > >  arch/x86/kvm/mmu/tdp_mmu.c      | 10 +++++++++-
> > >  3 files changed, 37 insertions(+), 3 deletions(-)
> > >
> > > diff --git a/arch/x86/include/asm/kvm_host.h b/arch/x86/include/asm/kvm_host.h
> > > index 3b58e2306621..c30918d0887e 100644
> > > --- a/arch/x86/include/asm/kvm_host.h
> > > +++ b/arch/x86/include/asm/kvm_host.h
> > > @@ -2324,4 +2324,18 @@ int memslot_rmap_alloc(struct kvm_memory_slot *slot, unsigned long npages);
> > >   */
> > >  #define KVM_EXIT_HYPERCALL_MBZ             GENMASK_ULL(31, 1)
> > >
> > > +#define kvm_arch_prepare_bitmap_age kvm_arch_prepare_bitmap_age
> > > +static inline bool kvm_arch_prepare_bitmap_age(struct mmu_notifier *mn)
> > > +{
> > > +   /*
> > > +    * Indicate that we support bitmap-based aging when using the TDP MMU
> > > +    * and the accessed bit is available in the TDP page tables.
> > > +    *
> > > +    * We have no other preparatory work to do here, so we do not need to
> > > +    * redefine kvm_arch_finish_bitmap_age().
> > > +    */
> > > +   return IS_ENABLED(CONFIG_X86_64) && tdp_mmu_enabled
> > > +                                    && shadow_accessed_mask;
> > > +}
> > > +
> > >  #endif /* _ASM_X86_KVM_HOST_H */
> > > diff --git a/arch/x86/kvm/mmu/mmu.c b/arch/x86/kvm/mmu/mmu.c
> > > index 992e651540e8..fae1a75750bb 100644
> > > --- a/arch/x86/kvm/mmu/mmu.c
> > > +++ b/arch/x86/kvm/mmu/mmu.c
> > > @@ -1674,8 +1674,14 @@ bool kvm_age_gfn(struct kvm *kvm, struct kvm_gfn_range *range)
> > >  {
> > >     bool young = false;
> > >
> > > -   if (kvm_memslots_have_rmaps(kvm))
> > > +   if (kvm_memslots_have_rmaps(kvm)) {
> > > +           if (range->lockless) {
> > > +                   kvm_age_set_unreliable(range);
> > > +                   return false;
> > > +           }
> >
> > If a VM has TDP MMU enabled, supports A/D bits, and is using nested
> > virtualization, MGLRU will effectively be blind to all accesses made by
> > the VM.
> >
> > kvm_arch_prepare_bitmap_age() will return true indicating that the
> > bitmap is supported. But then kvm_age_gfn() and kvm_test_age_gfn() will
> > return false immediately and indicate the bitmap is unreliable because a
> > shadow root is allocate. The notfier will then return
> > MMU_NOTIFIER_YOUNG_BITMAP_UNRELIABLE.
> >
> > Looking at the callers, MMU_NOTIFIER_YOUNG_BITMAP_UNRELIABLE is never
> > consumed or used. So I think MGLRU will assume all memory is
> > unaccessed?
> >
> > One way to improve the situation would be to re-order the TDP MMU
> > function first and return young instead of false, so that way MGLRU at
> > least has visibility into accesses made by L1 (and L2 if EPT is disable
> > in L2). But that still means MGLRU is blind to accesses made by L2.
> >
> > What about grabbing the mmu_lock if there's a shadow root allocated and
> > get rid of MMU_NOTIFIER_YOUNG_BITMAP_UNRELIABLE altogether?
> >
> >       if (kvm_memslots_have_rmaps(kvm)) {
> >               write_lock(&kvm->mmu_lock);
> >               young |= kvm_handle_gfn_range(kvm, range, kvm_age_rmap);
> >               write_unlock(&kvm->mmu_lock);
> >       }
> >
> > The TDP MMU walk would still be lockless. KVM only has to take the
> > mmu_lock to collect accesses made by L2.
> >
> > kvm_age_rmap() and kvm_test_age_rmap() will need to become bitmap-aware
> > as well, but that seems relatively simple with the helper functions.
>
> Wait, even simpler, just check kvm_memslots_have_rmaps() in
> kvm_arch_prepare_bitmap_age() and skip the shadow MMU when processing a
> bitmap request.
>
> i.e.
>
> static inline bool kvm_arch_prepare_bitmap_age(struct kvm *kvm, struct mmu_notifier *mn)
> {
>         /*
>          * Indicate that we support bitmap-based aging when using the TDP MMU
>          * and the accessed bit is available in the TDP page tables.
>          *
>          * We have no other preparatory work to do here, so we do not need to
>          * redefine kvm_arch_finish_bitmap_age().
>          */
>         return IS_ENABLED(CONFIG_X86_64)
>                 && tdp_mmu_enabled
>                 && shadow_accessed_mask
>                 && !kvm_memslots_have_rmaps(kvm);
> }
>
> bool kvm_age_gfn(struct kvm *kvm, struct kvm_gfn_range *range)
> {
>         bool young = false;
>
>         if (!range->arg.metadata->bitmap && kvm_memslots_have_rmaps(kvm))
>                 young = kvm_handle_gfn_range(kvm, range, kvm_age_rmap);
>
>         if (tdp_mmu_enabled)
>                 young |= kvm_tdp_mmu_age_gfn_range(kvm, range);
>
>         return young;
> }
>
> bool kvm_test_age_gfn(struct kvm *kvm, struct kvm_gfn_range *range)
> {
>         bool young = false;
>
>         if (!range->arg.metadata->bitmap && kvm_memslots_have_rmaps(kvm))
>                 young = kvm_handle_gfn_range(kvm, range, kvm_test_age_rmap);
>
>         if (tdp_mmu_enabled)
>                 young |= kvm_tdp_mmu_test_age_gfn(kvm, range);
>
>         return young;


Yeah I think this is the right thing to do. Given your other
suggestions (on patch 3), I think this will look something like this
-- let me know if I've misunderstood something:

bool check_rmap = !bitmap && kvm_memslot_have_rmaps(kvm);

if (check_rmap)
  KVM_MMU_LOCK(kvm);

rcu_read_lock(); // perhaps only do this when we don't take the MMU lock?

if (check_rmap)
  kvm_handle_gfn_range(/* ... */ kvm_test_age_rmap)

if (tdp_mmu_enabled)
  kvm_tdp_mmu_test_age_gfn() // modified to be RCU-safe

rcu_read_unlock();
if (check_rmap)
  KVM_MMU_UNLOCK(kvm);

> }
>
> Sure this could race with the creation of a shadow root but so can the
> non-bitmap code.