[PATCH] mm/mmu_notifier: Use hlist_add_head_rcu()

[Jean-Philippe Brucker <jean-philippe.brucker@arm.com>]

Make mmu_notifier_register() safer by issuing a memory barrier before
registering a new notifier. This fixes a theoretical bug on weakly ordered
CPUs. For example, take this simplified use of notifiers by a driver:

	my_struct->mn.ops = &my_ops; /* (1) */
	mmu_notifier_register(&my_struct->mn, mm)
		...
		hlist_add_head(&mn->hlist, &mm->mmu_notifiers); /* (2) */
		...

Once mmu_notifier_register() releases the mm locks, another thread can
invalidate a range:

	mmu_notifier_invalidate_range()
		...
		hlist_for_each_entry_rcu(mn, &mm->mmu_notifiers, hlist) {
			if (mn->ops->invalidate_range)

The read side relies on the data dependency between mn and ops to ensure
that the pointer is properly initialized. But the write side doesn't have
any dependency between (1) and (2), so they could be reordered and the
readers could dereference an invalid mn->ops. mmu_notifier_register() does
take all the mm locks before adding to the hlist, but those have acquire
semantics which isn't sufficient.

By calling hlist_add_head_rcu() instead of hlist_add_head() we update
the hlist using a store-release, ensuring that readers see prior
initialization of my_struct. This situation is better illustated by
litmus test MP+onceassign+derefonce.

Fixes: cddb8a5c14aa ("mmu-notifiers: core")
Signed-off-by: Jean-Philippe Brucker <jean-philippe.brucker@arm.com>
---
 mm/mmu_notifier.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/mm/mmu_notifier.c b/mm/mmu_notifier.c
index 9c884abc7850..9f246c960e65 100644
--- a/mm/mmu_notifier.c
+++ b/mm/mmu_notifier.c
@@ -276,7 +276,7 @@ static int do_mmu_notifier_register(struct mmu_notifier *mn,
 	 * thanks to mm_take_all_locks().
 	 */
 	spin_lock(&mm->mmu_notifier_mm->lock);
-	hlist_add_head(&mn->hlist, &mm->mmu_notifier_mm->list);
+	hlist_add_head_rcu(&mn->hlist, &mm->mmu_notifier_mm->list);
 	spin_unlock(&mm->mmu_notifier_mm->lock);
 
 	mm_drop_all_locks(mm);
-- 
2.21.0