KVM: Block memslot updates across range_start() and range_end()

We would like to avoid taking mmu_lock for .invalidate_range_{start,end}()
notifications that are unrelated to KVM.  Because mmu_notifier_count
must be modified while holding mmu_lock for write, and must always
be paired across start->end to stay balanced, lock elision must
happen in both or none.  Therefore, in preparation for this change,
this patch prevents memslot updates across range_start() and range_end().

Note, technically flag-only memslot updates could be allowed in parallel,
but stalling a memslot update for a relatively short amount of time is
not a scalability issue, and this is all more than complex enough.

A long note on the locking: a previous version of the patch used an rwsem
to block the memslot update while the MMU notifier run, but this resulted
in the following deadlock involving the pseudo-lock tagged as
"mmu_notifier_invalidate_range_start".

   ======================================================
   WARNING: possible circular locking dependency detected
   5.12.0-rc3+ #6 Tainted: G           OE
   ------------------------------------------------------
   qemu-system-x86/3069 is trying to acquire lock:
   ffffffff9c775ca0 (mmu_notifier_invalidate_range_start){+.+.}-{0:0}, at: __mmu_notifier_invalidate_range_end+0x5/0x190

   but task is already holding lock:
   ffffaff7410a9160 (&kvm->mmu_notifier_slots_lock){.+.+}-{3:3}, at: kvm_mmu_notifier_invalidate_range_start+0x36d/0x4f0 [kvm]

   which lock already depends on the new lock.

This corresponds to the following MMU notifier logic:

    invalidate_range_start
      take pseudo lock
      down_read()           (*)
      release pseudo lock
    invalidate_range_end
      take pseudo lock      (**)
      up_read()
      release pseudo lock

At point (*) we take the mmu_notifiers_slots_lock inside the pseudo lock;
at point (**) we take the pseudo lock inside the mmu_notifiers_slots_lock.

This could cause a deadlock (ignoring for a second that the pseudo lock
is not a lock):

- invalidate_range_start waits on down_read(), because the rwsem is
held by install_new_memslots

- install_new_memslots waits on down_write(), because the rwsem is
held till (another) invalidate_range_end finishes

- invalidate_range_end sits waits on the pseudo lock, held by
invalidate_range_start.

Removing the fairness of the rwsem breaks the cycle (in lockdep terms,
it would change the *shared* rwsem readers into *shared recursive*
readers), so open-code the wait using a readers count and a
spinlock.  This also allows handling blockable and non-blockable
critical section in the same way.

Losing the rwsem fairness does theoretically allow MMU notifiers to
block install_new_memslots forever.  Note that mm/mmu_notifier.c's own
retry scheme in mmu_interval_read_begin also uses wait/wake_up
and is likewise not fair.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

1 files changed, 6 insertions, 0 deletions

diff --git a/Documentation/virt/kvm/locking.rst b/Documentation/virt/kvm/locking.rst
index 35eca377543d..8138201efb09 100644
--- a/Documentation/virt/kvm/locking.rst
+++ b/Documentation/virt/kvm/locking.rst
@@ -21,6 +21,12 @@ The acquisition orders for mutexes are as follows:
   can be taken inside a kvm->srcu read-side critical section,
   while kvm->slots_lock cannot.
 
+- kvm->mn_active_invalidate_count ensures that pairs of
+  invalidate_range_start() and invalidate_range_end() callbacks
+  use the same memslots array.  kvm->slots_lock and kvm->slots_arch_lock
+  are taken on the waiting side in install_new_memslots, so MMU notifiers
+  must not take either kvm->slots_lock or kvm->slots_arch_lock.
+
 On x86:
 
 - vcpu->mutex is taken outside kvm->arch.hyperv.hv_lock