On Sat, Feb 08, 2025, Yan Zhao wrote: > On Fri, Feb 07, 2025 at 07:03:46AM -0800, Sean Christopherson wrote: > > On Fri, Feb 07, 2025, Yan Zhao wrote: > > > Merge the prefetch check into the is_access_allowed() check to determine a > > > spurious fault. > > > > > > In the TDP MMU, a spurious prefetch fault should also pass the > > > is_access_allowed() check. > > > > How so? > > > > 1. vCPU takes a write-fault on a swapped out page and queues an async #PF > > 2. A different task installs a writable SPTE > > 3. A third task write-protects the SPTE for dirty logging > > 4. Async #PF handler faults in the SPTE, encounters a read-only SPTE for its > > write fault. > > > > KVM shouldn't mark the gfn as dirty in this case. > Hmm, but when we prefetch an entry, if a gfn is not write-tracked, it allows to > mark the gfn as dirty, just like when there's no existing SPTE, a prefetch fault > also marks a gfn as dirty. Yeah, but there's a difference between installing a SPTE and overwriting a SPTE. > If a gfn is write-tracked, make_spte() will not grant write-permission to make > the gfn dirty. > > However, I admit that making the new SPTE as not-accessed again is not desired. > What about below? > > @@ -983,7 +983,7 @@ static int tdp_mmu_map_handle_target_level(struct kvm_vcpu *vcpu, > return RET_PF_RETRY; > > if (is_shadow_present_pte(iter->old_spte) && > - is_access_allowed(fault, iter->old_spte) && > + (fault->prefetch || is_access_allowed(fault, iter->old_spte)) && > is_last_spte(iter->old_spte, iter->level)) > return RET_PF_SPURIOUS; Works for me.
