Re: [PATCH 16/18] KVM: Don't take mmu_lock for range invalidation unless necessary

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

On 31/03/21 23:05, Sean Christopherson wrote:

quoted

quoted

Wouldn't it be incorrect to lock a mutex (e.g. inside*another*  MMU
notifier's invalidate callback) while holding an rwlock_t?  That makes sense
because anybody that's busy waiting in write_lock potentially cannot be
preempted until the other task gets the mutex.  This is a potential
deadlock.

Yes?  I don't think I follow your point though.  Nesting a spinlock or rwlock
inside a rwlock is ok, so long as the locks are always taken in the same order,
i.e. it's never mmu_lock -> mmu_notifier_slots_lock.

*Another* MMU notifier could nest a mutex inside KVM's rwlock.

But... is it correct that the MMU notifier invalidate callbacks are always
called with the mmap_sem taken (sometimes for reading, e.g.
try_to_merge_with_ksm_page->try_to_merge_one_page->write_protect_page)?

No :-(

File-based invalidations through the rmaps do not take mmap_sem.  They get at
the VMAs via the address_space's interval tree, which is protected by its own
i_mmap_rwsem.

E.g. try_to_unmap() -> rmap_walk_file() -> try_to_unmap_one() 

We could take it temporarily in install_memslots, since the MMU notifier's mm
is stored in kvm->mm.

In this case, a pair of kvm_mmu_notifier_lock/unlock functions would be the
best way to abstract it.

Paolo