On (21/10/21 11:48), Sergey Senozhatsky wrote: > > We are using TDP. And somehow I never see (literally never) async PFs. > It's always either hva_to_pfn_fast() or hva_to_pfn_slow() or > __direct_map() from tdp_page_fault(). Hmm, and tdp_page_fault()->fast_page_fault() always fails on !is_access_allowed(error_code, new_spte), it never handles the faults. And I see some ->mmu_lock contention: spin_lock(&vcpu->kvm->mmu_lock); __direct_map(); spin_unlock(&vcpu->kvm->mmu_lock); So it might be that we setup guest memory wrongly and never get advantages of TPD and fast page faults?
