On Wed, Mar 31, 2021, Paolo Bonzini wrote:

On 26/03/21 03:19, Sean Christopherson wrote:

quoted

+	/*
+	 * Reset the lock used to prevent memslot updates between MMU notifier
+	 * range_start and range_end.  At this point no more MMU notifiers will
+	 * run, but the lock could still be held if KVM's notifier was removed
+	 * between range_start and range_end.  No threads can be waiting on the
+	 * lock as the last reference on KVM has been dropped.  If the lock is
+	 * still held, freeing memslots will deadlock.
+	 */
+	init_rwsem(&kvm->mmu_notifier_slots_lock);

I was going to say that this is nasty, then I noticed that
mmu_notifier_unregister uses SRCU to ensure completion of concurrent calls
to the MMU notifier.  So I guess it's fine, but it's better to point it out:

	/*
	 * At this point no more MMU notifiers will run and pending
	 * calls to range_start have completed, but the lock would
	 * still be held and never released if the MMU notifier was
	 * removed between range_start and range_end.  Since the last
	 * reference to the struct kvm has been dropped, no threads can
	 * be waiting on the lock, but we might still end up taking it
	 * when freeing memslots in kvm_arch_destroy_vm.  Reset the lock
	 * to avoid deadlocks.
	 */

An alternative would be to not take the lock in install_new_memslots() if
kvm->users_count == 0.  It'd be weirder to document, and the conditional locking
would still be quite ugly.  Not sure if that's better than blasting a lock
during destruction?