On Tue, Apr 27, 2021 at 11:25 PM Paolo Bonzini <pbonzini@xxxxxxxxxx> wrote:
>
> On 28/04/21 00:36, Ben Gardon wrote:
> > +void kvm_arch_assign_memslots(struct kvm *kvm, int as_id,
> > + struct kvm_memslots *slots)
> > +{
> > + mutex_lock(&kvm->arch.memslot_assignment_lock);
> > + rcu_assign_pointer(kvm->memslots[as_id], slots);
> > + mutex_unlock(&kvm->arch.memslot_assignment_lock);
> > +}
>
> Does the assignment also needs the lock, or only the rmap allocation? I
> would prefer the hook to be something like kvm_arch_setup_new_memslots.
The assignment does need to be under the lock to prevent the following race:
1. Thread 1 (installing a new memslot): Acquires memslot assignment
lock (or perhaps in this case rmap_allocation_lock would be more apt.)
2. Thread 1: Check alloc_memslot_rmaps (it is false)
3. Thread 1: doesn't allocate memslot rmaps for new slot
4. Thread 1: Releases memslot assignment lock
5. Thread 2 (allocating a shadow root): Acquires memslot assignment lock
6. Thread 2: Sets alloc_memslot_rmaps = true
7. Thread 2: Allocates rmaps for all existing slots
8. Thread 2: Releases memslot assignment lock
9. Thread 2: Sets shadow_mmu_active = true
10. Thread 1: Installs the new memslots
11. Thread 3: Null pointer dereference when trying to access rmaps on
the new slot.
Putting the assignment under the lock prevents 5-8 from happening
between 2 and 10.
I'm open to other ideas as far as how to prevent this race though. I
admit this solution is not the most elegant looking.
>
> (Also it is useful to have a comment somewhere explaining why the
> slots_lock does not work. IIUC there would be a deadlock because you'd
> be taking the slots_lock inside an SRCU critical region, while usually
> the slots_lock critical section is the one that includes a
> synchronize_srcu; I should dig that up and document that ordering in
> Documentation/virt/kvm too).
Yeah, sorry about that. I should have added a comment to that effect.
As you suspected, it's because of the slots lock / SRCU deadlock.
Using the slots lock was my original implementation, until the
deadlock issue came up.
I can add comments about that in a v2.
>
> Paolo
>
