block: don't defer flushes on blk-mq + scheduling

For blk-mq with scheduling, we can potentially end up with ALL
driver tags assigned and sitting on the flush queues. If we
defer because of an inlfight data request, then we can deadlock
if that data request doesn't already have a tag assigned.

This fixes a deadlock with running the xfs/297 xfstest, where
thousands of syncs can cause the drive queue to stall.

Signed-off-by: Jens Axboe <axboe@fb.com>
Reviewed-by: Omar Sandoval <osandov@fb.com>

diff --git a/block/blk-flush.c b/block/blk-flush.c
index 9681625..0d5a9c1 100644 (file)
--- a/block/blk-flush.c
+++ b/block/blk-flush.c
@@ -297,8 +297,14 @@ static bool blk_kick_flush(struct request_queue *q, struct blk_flush_queue *fq)
        if (fq->flush_pending_idx != fq->flush_running_idx || list_empty(pending))
                return false;
 
-       /* C2 and C3 */
+       /* C2 and C3
+        *
+        * For blk-mq + scheduling, we can risk having all driver tags
+        * assigned to empty flushes, and we deadlock if we are expecting
+        * other requests to make progress. Don't defer for that case.
+        */
        if (!list_empty(&fq->flush_data_in_flight) &&
+           !(q->mq_ops && q->elevator) &&
            time_before(jiffies,
                        fq->flush_pending_since + FLUSH_PENDING_TIMEOUT))
                return false;