[PATCH 6/7] mempool: fix a wakeup race when sleeping for elements

[Christoph Hellwig]

Waiters always need to re-check their condition after adding themselves
to the waitqueue, as otherwise they might miss a wakeup.  Check
for elements in the pool and use them before going to sleep.

The workaround mentioned was probably due to this, but seems genuinely
useful for other reasons, so keep it and update the comment describing
it.

Signed-off-by: Christoph Hellwig <hch@xxxxxx>
---
 mm/mempool.c | 13 +++++++++----
 1 file changed, 9 insertions(+), 4 deletions(-)

diff --git a/mm/mempool.c b/mm/mempool.c
index 850362f4ca7a..8cf3b5705b7f 100644
--- a/mm/mempool.c
+++ b/mm/mempool.c
@@ -388,6 +388,7 @@ static void *mempool_alloc_from_pool(struct mempool *pool, gfp_t gfp_mask)
 	spin_lock_irqsave(&pool->lock, flags);
 	if (unlikely(!pool->curr_nr))
 		goto fail;
+alloc:
 	element = remove_element(pool);
 	spin_unlock_irqrestore(&pool->lock, flags);
 
@@ -406,13 +407,17 @@ static void *mempool_alloc_from_pool(struct mempool *pool, gfp_t gfp_mask)
 		DEFINE_WAIT(wait);
 
 		prepare_to_wait(&pool->wait, &wait, TASK_UNINTERRUPTIBLE);
+		if (pool->curr_nr) {
+			finish_wait(&pool->wait, &wait);
+			goto alloc;
+		}
 		spin_unlock_irqrestore(&pool->lock, flags);
 
 		/*
-		 * Wait for someone else to return an element to @pool.
-		 *
-		 * FIXME: this should be io_schedule().  The timeout is there as
-		 * a workaround for some DM problems in 2.6.18.
+		 * Wait for someone else to return an element to @pool, but wake
+		 * up occasionally as memory pressure might have reduced even
+		 * and the normal allocation in alloc_fn could succeed even if
+		 * no element was returned.
 		 */
 		io_schedule_timeout(5 * HZ);
 		finish_wait(&pool->wait, &wait);
-- 
2.47.3