On Tue, Jan 31, 2023 at 9:41 AM David Matlack <dmatlack@xxxxxxxxxx> wrote: > > On Mon, Jan 30, 2023 at 3:49 PM Sean Christopherson <seanjc@xxxxxxxxxx> wrote: > > > > On Mon, Jan 30, 2023, David Matlack wrote: > > > On Wed, Jan 25, 2023 at 5:49 PM Sean Christopherson <seanjc@xxxxxxxxxx> wrote: > > > [...] > > > > --- > > > > arch/x86/kvm/mmu/tdp_mmu.c | 92 ++++++++++---------------------------- > > > > 1 file changed, 24 insertions(+), 68 deletions(-) > > > > > > > > diff --git a/arch/x86/kvm/mmu/tdp_mmu.c b/arch/x86/kvm/mmu/tdp_mmu.c > > > > index bba33aea0fb0..2f78ca43a276 100644 > > > > --- a/arch/x86/kvm/mmu/tdp_mmu.c > > > > +++ b/arch/x86/kvm/mmu/tdp_mmu.c > > > [...] > > > > @@ -1289,8 +1244,7 @@ static bool age_gfn_range(struct kvm *kvm, struct tdp_iter *iter, > > > > new_spte = mark_spte_for_access_track(new_spte); > > > > } > > > > > > > > - tdp_mmu_set_spte_no_acc_track(kvm, iter, new_spte); > > > > - > > > > + kvm_tdp_mmu_write_spte(iter->sptep, iter->old_spte, new_spte, iter->level); > > > > > > This can race with fast_page_fault() setting the W-bit and the CPU > > > setting the D-bit. i.e. This call to kvm_tdp_mmu_write_spte() could > > > clear the W-bit or D-bit. > > > > Ugh, right. Hrm. Duh, I just didn't go far enough. A straight XCHG is silly. > > Except for the access-tracking mess, KVM wants to clear a single bit. Writing > > the entire thing and potentially clobbering bits is wasteful and unnecessarily > > dangerous. And the access-tracking code already needs special handling. > > > > We can just simplify this all by adding a helper to clear a single bit (and > > maybe even use clear_bit() and __clear_bit() if we save the bit number for the > > W/A/D bits and not just the mask). And if it weren't for EPT (different A/D > > locations), we could even use static asserts to restrict the usage to the W/A/D > > bits :-/ Oh well. > > > > E.g. this > > This patch looks good. Vipin can you incorporate this in your next version? On it.
