On Wed, Jan 05, 2022, Chao Peng wrote: > On Tue, Jan 04, 2022 at 05:31:30PM +0000, Sean Christopherson wrote: > > On Fri, Dec 31, 2021, Chao Peng wrote: > > > On Fri, Dec 24, 2021 at 12:13:51PM +0800, Chao Peng wrote: > > > > On Thu, Dec 23, 2021 at 06:06:19PM +0000, Sean Christopherson wrote: > > > > > On Thu, Dec 23, 2021, Chao Peng wrote: > > > > > > This new function establishes the mapping in KVM page tables for a > > > > > > given gfn range. It can be used in the memory fallocate callback for > > > > > > memfd based memory to establish the mapping for KVM secondary MMU when > > > > > > the pages are allocated in the memory backend. > > > > > > > > > > NAK, under no circumstance should KVM install SPTEs in response to allocating > > > > > memory in a file. The correct thing to do is to invalidate the gfn range > > > > > associated with the newly mapped range, i.e. wipe out any shared SPTEs associated > > > > > with the memslot. > > > > > > > > Right, thanks. > > > > > > BTW, I think the current fallocate() callback is just useless as long as > > > we don't want to install KVM SPTEs in response to allocating memory in a > > > file. The invalidation of the shared SPTEs should be notified through > > > mmu_notifier of the shared memory backend, not memfd_notifier of the > > > private memory backend. > > > > No, because the private fd is the final source of truth as to whether or not a > > GPA is private, e.g. userspace may choose to not unmap the shared backing. > > KVM's rule per Paolo's/this proposoal is that a GPA is private if it has a private > > memslot and is present in the private backing store. And the other core rule is > > that KVM must never map both the private and shared variants of a GPA into the > > guest. > > That's true, but I'm wondering if zapping the shared variant can be > handled at the time when the private one gets mapped in the KVM page > fault. No bothering the backing store to dedicate a callback to tell > KVM. Hmm, I don't think that would work for the TDP MMU due to page faults taking mmu_lock for read. E.g. if two vCPUs concurrently fault in both the shared and private variants, a race could exist where the private page fault sees the gfn as private and the shared page fault sees it as shared. In that case, both faults will install a SPTE and KVM would end up running with both variants mapped into the guest. There's also a performance penalty, as KVM would need to walk the shared EPT tree on every private page fault.
