On Wed, Mar 31, 2021, Paolo Bonzini wrote: > On 26/03/21 03:19, Sean Christopherson wrote: > > + /* > > + * 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, Heh, I still think it's 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. > */ > > That said, the easiest way to avoid this would be to always update > mmu_notifier_count. Updating mmu_notifier_count requires taking mmu_lock, which would defeat the purpose of these shenanigans. I think it could be made atomic, since mmu_lock would be taken before the elevated count _must_ be visible, but that would break the mmu_notifier_range_{start,end} optimization that was recently added. Or did I completely misunderstand what you're suggesting? > I don't mind the rwsem, but at least I suggest that you > split the patch in two---the first one keeping the mmu_notifier_count update > unconditional, and the second one introducing the rwsem and the on_lock > function kvm_inc_notifier_count. Please document the new lock in > Documentation/virt/kvm/locking.rst too. Note, will update docs. > Also, related to the first part of the series, perhaps you could structure > the series in a slightly different way: > > 1) introduce the HVA walking API in common code, complete with on_lock and > patch 15, so that you can use on_lock to increase mmu_notifier_seq > > 2) then migrate all architectures including x86 to the new API > > IOW, first half of patch 10 and all of patch 15; then the second half of > patch 10; then patches 11-14. > > > +#if defined(CONFIG_MMU_NOTIFIER) && defined(KVM_ARCH_WANT_MMU_NOTIFIER) > > + down_write(&kvm->mmu_notifier_slots_lock); > > +#endif > > rcu_assign_pointer(kvm->memslots[as_id], slots); > > +#if defined(CONFIG_MMU_NOTIFIER) && defined(KVM_ARCH_WANT_MMU_NOTIFIER) > > + up_write(&kvm->mmu_notifier_slots_lock); > > +#endif > > Please do this unconditionally, the cost is minimal if the rwsem is not > contended (as is the case if the architecture doesn't use MMU notifiers at > all). It's not the cost, it's that mmu_notifier_slots_lock doesn't exist. That's an easily solved problem, but then the lock wouldn't be initialized since kvm_init_mmu_notifier() is a nop. That's again easy to solve, but IMO would look rather weird. I guess the counter argument is that __kvm_memslots() wouldn't need #ifdeffery. These are the to ideas I've come up with: Option 1: static int kvm_init_mmu_notifier(struct kvm *kvm) { init_rwsem(&kvm->mmu_notifier_slots_lock); #if defined(CONFIG_MMU_NOTIFIER) && defined(KVM_ARCH_WANT_MMU_NOTIFIER) kvm->mmu_notifier.ops = &kvm_mmu_notifier_ops; return mmu_notifier_register(&kvm->mmu_notifier, current->mm); #else return 0; #endif } Option 2: kvm_mmu_notifier_lock(kvm); rcu_assign_pointer(kvm->memslots[as_id], slots); kvm_mmu_notifier_unlock(kvm);