On Wed, Jul 19, 2023 at 3:35 PM Yan Zhao [off-list ref] wrote:

On Tue, Jul 04, 2023 at 04:50:50PM +0900, David Stevens wrote:

quoted

@@ -4451,7 +4461,8 @@ static int direct_page_fault(struct kvm_vcpu *vcpu, struct kvm_page_fault *fault

 out_unlock:
      write_unlock(&vcpu->kvm->mmu_lock);
-     kvm_release_pfn_clean(fault->pfn);
+     if (fault->is_refcounted_page)
+             kvm_set_page_accessed(pfn_to_page(fault->pfn));

For a refcounted page, as now KVM puts its ref early in kvm_faultin_pfn(),
should this kvm_set_page_accessed() be placed before unlocking mmu_lock?

Otherwise, if the user unmaps a region (which triggers kvm_unmap_gfn_range()
with mmu_lock holding for write), and release the page, and if the two
steps happen after checking page_count() in kvm_set_page_accessed() and
before mark_page_accessed(), the latter function may mark accessed to a page
that is released or does not belong to current process.

Is it true?

Yes, good catch. During some testing last week, I actually found this
bug thanks to the WARN_ON the first patch in this series added to
kvm_is_ad_tracked_page. I'll fix it in the next revision, after Sean
gets a chance to comment on the series.

Thanks,
David