On Tue, Jan 26, 2021, Sean Christopherson wrote: > On Tue, Jan 26, 2021, Paolo Bonzini wrote: > > On 21/01/21 22:32, Sean Christopherson wrote: > > > Coming back to this series, I wonder if the RCU approach is truly necessary to > > > get the desired scalability. If both zap_collapsible_sptes() and NX huge page > > > recovery zap_only_ leaf SPTEs, then the only path that can actually unlink a > > > shadow page while holding the lock for read is the page fault path that installs > > > a huge page over an existing shadow page. > > > > > > Assuming the above analysis is correct, I think it's worth exploring alternatives > > > to using RCU to defer freeing the SP memory, e.g. promoting to a write lock in > > > the specific case of overwriting a SP (though that may not exist for rwlocks), > > > or maybe something entirely different? > > > > You can do the deferred freeing with a short write-side critical section to > > ensure all readers have terminated. > > Hmm, the most obvious downside I see is that the zap_collapsible_sptes() case > will not scale as well as the RCU approach. E.g. the lock may be heavily > contested when refaulting all of guest memory to (re)install huge pages after a > failed migration. > > Though I wonder, could we do something even more clever for that particular > case? And I suppose it would apply to NX huge pages as well. Instead of > zapping the leaf PTEs and letting the fault handler install the huge page, do an > in-place promotion when dirty logging is disabled. That could all be done under > the read lock, and with Paolo's method for deferred free on the back end. That > way only the thread doing the memslot update would take mmu_lock for write, and > only once per memslot update. Oh, and we could even skip the remote TLB flush in that case since the GPA->HPA translation is unchanged.
