bfq: Fix warning in bfqq_request_over_limit()

People are occasionally reporting a warning bfqq_request_over_limit()
triggering reporting that BFQ's idea of cgroup hierarchy (and its depth)
does not match what generic blkcg code thinks. This can actually happen
when bfqq gets moved between BFQ groups while bfqq_request_over_limit()
is running. Make sure the code is safe against BFQ queue being moved to
a different BFQ group.

Fixes: 76f1df8 ("bfq: Limit number of requests consumed by each cgroup")
CC: stable@vger.kernel.org
Link: https://lore.kernel.org/all/CAJCQCtTw_2C7ZSz7as5Gvq=OmnDiio=HRkQekqWpKot84sQhFA@mail.gmail.com/
Reported-by: Chris Murphy <lists@colorremedies.com>
Reported-by: "yukuai (C)" <yukuai3@huawei.com>
Signed-off-by: Jan Kara <jack@suse.cz>
Link: https://lore.kernel.org/r/20220407140738.9723-1-jack@suse.cz
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block/bfq-iosched.c

@@ -569,7 +569,7 @@ static bool bfqq_request_over_limit(struct bfq_queue *bfqq, int limit)
 	struct bfq_entity *entity = &bfqq->entity;
 	struct bfq_entity *inline_entities[BFQ_LIMIT_INLINE_DEPTH];
 	struct bfq_entity **entities = inline_entities;
-	int depth, level;
+	int depth, level, alloc_depth = BFQ_LIMIT_INLINE_DEPTH;
 	int class_idx = bfqq->ioprio_class - 1;
 	struct bfq_sched_data *sched_data;
 	unsigned long wsum;
@@ -578,15 +578,21 @@ static bool bfqq_request_over_limit(struct bfq_queue *bfqq, int limit)
 	if (!entity->on_st_or_in_serv)
 		return false;
 
+retry:
+	spin_lock_irq(&bfqd->lock);
 	/* +1 for bfqq entity, root cgroup not included */
 	depth = bfqg_to_blkg(bfqq_group(bfqq))->blkcg->css.cgroup->level + 1;
-	if (depth > BFQ_LIMIT_INLINE_DEPTH) {
+	if (depth > alloc_depth) {
+		spin_unlock_irq(&bfqd->lock);
+		if (entities != inline_entities)
+			kfree(entities);
 		entities = kmalloc_array(depth, sizeof(*entities), GFP_NOIO);
 		if (!entities)
 			return false;
+		alloc_depth = depth;
+		goto retry;
 	}
 
-	spin_lock_irq(&bfqd->lock);
 	sched_data = entity->sched_data;
 	/* Gather our ancestors as we need to traverse them in reverse order */
 	level = 0;