On Thu, Apr 04, 2024, David Matlack wrote:
> > I don't love the idea of adding more arch specific MMU behavior (going the wrong
> > direction), but it doesn't seem like an unreasonable approach in this case.
>
> I wonder if this is being overly cautious.
Probably. "Lazy" is another word for it ;-)
> I would expect only more benefit on architectures that more aggressively take
> the mmu_lock on vCPU threads during faults. The more lock acquisition on vCPU
> threads, the more this patch will help reduce vCPU starvation during
> CLEAR_DIRTY_LOG.
>
> Hm, perhaps testing with ept=N (which will use the write-lock for even
> dirty logging faults) would be a way to increase confidence in the
> effect on other architectures?
Turning off the TDP MMU would be more representative, just manually disable the
fast-path, e.g.
diff --git a/arch/x86/kvm/mmu/mmu.c b/arch/x86/kvm/mmu/mmu.c
index 992e651540e8..532c24911f39 100644
--- a/arch/x86/kvm/mmu/mmu.c
+++ b/arch/x86/kvm/mmu/mmu.c
@@ -3371,7 +3371,7 @@ static bool page_fault_can_be_fast(struct kvm_page_fault *fault)
* Note, instruction fetches and writes are mutually exclusive, ignore
* the "exec" flag.
*/
- return fault->write;
+ return false;//fault->write;
}
/*
