On Tue, Feb 25, 2025 at 05:53:49PM -0800, Sean Christopherson wrote: > On Tue, Feb 11, 2025, Rick P Edgecombe wrote: > > On Tue, 2025-02-11 at 09:42 +0800, Yan Zhao wrote: > > > > On the fix though, doesn't this remove the coverage of writing to a > > > > region that is in the process of being made RO? I'm thinking about > > > > warnings, etc that may trigger intermittently based on bugs with a race > > > > component. I don't know if we could fix the test and still leave the > > > > write while the "mprotect(PROT_READ) is underway". It seems to be > > > > deliberate. > > > Write before "mprotect(PROT_READ)" has been tested in stage 0. > > > Not sure it's deliberate to test write in the process of being made RO. > > Writing while VMAs are being made RO is 100% intended. The goal is to stress > KVM's interactions with the mmu_notifier, and to verify KVM delivers -EFAULT to > userspace. > > Something isn't quite right in the original analysis. We need to drill down on > that before change anything. > > FWIW, I run this test frequently on large systems and have never observed failures. Could you try adding CONFIG_LOCK_STAT=y? With this config, the failure rate is more than 90% in my SPR non-TDX machine, and 20%~80% in my TDX machine. > Maybe Rick and I should go buy lottery tickets?
