On Tue, Jul 27, 2021, Ben Gardon wrote: > On Tue, Jul 27, 2021 at 6:06 AM Maxim Levitsky <mlevitsk@xxxxxxxxxx> wrote: > > > > On Thu, 2021-07-22 at 19:06 +0000, Sean Christopherson wrote: > > > The elevated mmu_notifier_count and/or changed mmu_notifier_seq will cause vCPU1 > > > to bail and resume the guest without fixing the #NPF. After acquiring mmu_lock, > > > vCPU1 will see the elevated mmu_notifier_count (if kvm_zap_gfn_range() is about > > > to be called, or just finised) and/or a modified mmu_notifier_seq (after the > > > count was decremented). > > > > > > This is why kvm_zap_gfn_range() needs to take mmu_lock for write. If it's allowed > > > to run in parallel with the page fault handler, there's no guarantee that the > > > correct apic_access_memslot_enabled will be observed. > > > > I understand now. > > > > So, Paolo, Ben Gardon, what do you think. Do you think this approach is feasable? > > Do you agree to revert the usage of the read lock? > > > > I will post a new series using this approach very soon, since I already have > > msot of the code done. > > > > Best regards, > > Maxim Levitsky > > From reading through this thread, it seems like switching from read > lock to write lock is only necessary for a small range of GFNs, (i.e. > the APIC access page) is that correct? For the APICv case, yes, literally a single GFN (the default APIC base). > My initial reaction was that switching kvm_zap_gfn_range back to the > write lock would be terrible for performance, but given its only two > callers, I think it would actually be fine. And more importantly, the two callers are gated by kvm_arch_has_noncoherent_dma() and are very rare flows for the guest (updating MTRRs, toggling CR0.CD). > If you do that though, you should pass shared=false to > kvm_tdp_mmu_zap_gfn_range in that function, so that it knows it's > operating with exclusive access to the MMU lock. Ya, my suggested revert was to drop @shared entirely since kvm_zap_gfn_range() is the only caller that passes @shared=true.
