[PATCH 1/4] KVM: delete .change_pte MMU notifier callback
Marc Zyngier
maz at kernel.org
Sat Apr 13 02:56:25 PDT 2024
On Fri, 12 Apr 2024 15:54:22 +0100,
Sean Christopherson <seanjc at google.com> wrote:
>
> On Fri, Apr 12, 2024, Marc Zyngier wrote:
> > On Fri, 12 Apr 2024 11:44:09 +0100, Will Deacon <will at kernel.org> wrote:
> > > On Fri, Apr 05, 2024 at 07:58:12AM -0400, Paolo Bonzini wrote:
> > > Also, if you're in the business of hacking the MMU notifier code, it
> > > would be really great to change the .clear_flush_young() callback so
> > > that the architecture could handle the TLB invalidation. At the moment,
> > > the core KVM code invalidates the whole VMID courtesy of 'flush_on_ret'
> > > being set by kvm_handle_hva_range(), whereas we could do a much
> > > lighter-weight and targetted TLBI in the architecture page-table code
> > > when we actually update the ptes for small ranges.
> >
> > Indeed, and I was looking at this earlier this week as it has a pretty
> > devastating effect with NV (it blows the shadow S2 for that VMID, with
> > costly consequences).
> >
> > In general, it feels like the TLB invalidation should stay with the
> > code that deals with the page tables, as it has a pretty good idea of
> > what needs to be invalidated and how -- specially on architectures
> > that have a HW-broadcast facility like arm64.
>
> Would this be roughly on par with an in-line flush on arm64? The simpler, more
> straightforward solution would be to let architectures override flush_on_ret,
> but I would prefer something like the below as x86 can also utilize a range-based
> flush when running as a nested hypervisor.
>
> diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
> index ff0a20565f90..b65116294efe 100644
> --- a/virt/kvm/kvm_main.c
> +++ b/virt/kvm/kvm_main.c
> @@ -601,6 +601,7 @@ static __always_inline kvm_mn_ret_t __kvm_handle_hva_range(struct kvm *kvm,
> struct kvm_gfn_range gfn_range;
> struct kvm_memory_slot *slot;
> struct kvm_memslots *slots;
> + bool need_flush = false;
> int i, idx;
>
> if (WARN_ON_ONCE(range->end <= range->start))
> @@ -653,10 +654,22 @@ static __always_inline kvm_mn_ret_t __kvm_handle_hva_range(struct kvm *kvm,
> break;
> }
> r.ret |= range->handler(kvm, &gfn_range);
> +
> + /*
> + * Use a precise gfn-based TLB flush when possible, as
> + * most mmu_notifier events affect a small-ish range.
> + * Fall back to a full TLB flush if the gfn-based flush
> + * fails, and don't bother trying the gfn-based flush
> + * if a full flush is already pending.
> + */
> + if (range->flush_on_ret && !need_flush && r.ret &&
> + kvm_arch_flush_remote_tlbs_range(kvm, gfn_range.start
> + gfn_range.end - gfn_range.start + 1))
> + need_flush = true;
> }
> }
>
> - if (range->flush_on_ret && r.ret)
> + if (need_flush)
> kvm_flush_remote_tlbs(kvm);
>
> if (r.found_memslot)
I think this works for us on HW that has range invalidation, which
would already be a positive move.
For the lesser HW that isn't range capable, it also gives the
opportunity to perform the iteration ourselves or go for the nuclear
option if the range is larger than some arbitrary constant (though
this is additional work).
But this still considers the whole range as being affected by
range->handler(). It'd be interesting to try and see whether more
precise tracking is (or isn't) generally beneficial.
Thanks,
M.
--
Without deviation from the norm, progress is not possible.
More information about the linux-arm-kernel
mailing list