1 files changed, 65 insertions, 33 deletions

diff --git a/block/blk-rq-qos.c b/block/blk-rq-qos.c
index eb9618cd68ad..654478dfbc20 100644
--- a/block/blk-rq-qos.c
+++ b/block/blk-rq-qos.c
@@ -196,7 +196,6 @@ bool rq_depth_scale_down(struct rq_depth *rqd, bool hard_throttle)
 
 struct rq_qos_wait_data {
 	struct wait_queue_entry wq;
-	struct task_struct *task;
 	struct rq_wait *rqw;
 	acquire_inflight_cb_t *cb;
 	void *private_data;
@@ -218,7 +217,20 @@ static int rq_qos_wake_function(struct wait_queue_entry *curr,
 		return -1;
 
 	data->got_token = true;
-	wake_up_process(data->task);
+	/*
+	 * autoremove_wake_function() removes the wait entry only when it
+	 * actually changed the task state. We want the wait always removed.
+	 * Remove explicitly and use default_wake_function().
+	 */
+	default_wake_function(curr, mode, wake_flags, key);
+	/*
+	 * Note that the order of operations is important as finish_wait()
+	 * tests whether @curr is removed without grabbing the lock. This
+	 * should be the last thing to do to make sure we will not have a
+	 * UAF access to @data. And the semantics of memory barrier in it
+	 * also make sure the waiter will see the latest @data->got_token
+	 * once list_empty_careful() in finish_wait() returns true.
+	 */
 	list_del_init_careful(&curr->entry);
 	return 1;
 }
@@ -244,41 +256,55 @@ void rq_qos_wait(struct rq_wait *rqw, void *private_data,
 		 cleanup_cb_t *cleanup_cb)
 {
 	struct rq_qos_wait_data data = {
-		.wq = {
-			.func	= rq_qos_wake_function,
-			.entry	= LIST_HEAD_INIT(data.wq.entry),
-		},
-		.task = current,
-		.rqw = rqw,
-		.cb = acquire_inflight_cb,
-		.private_data = private_data,
+		.rqw		= rqw,
+		.cb		= acquire_inflight_cb,
+		.private_data	= private_data,
+		.got_token	= false,
 	};
-	bool has_sleeper;
+	bool first_waiter;
 
-	has_sleeper = wq_has_sleeper(&rqw->wait);
-	if (!has_sleeper && acquire_inflight_cb(rqw, private_data))
+	/*
+	 * If there are no waiters in the waiting queue, try to increase the
+	 * inflight counter if we can. Otherwise, prepare for adding ourselves
+	 * to the waiting queue.
+	 */
+	if (!waitqueue_active(&rqw->wait) && acquire_inflight_cb(rqw, private_data))
 		return;
 
-	has_sleeper = !prepare_to_wait_exclusive(&rqw->wait, &data.wq,
+	init_wait_func(&data.wq, rq_qos_wake_function);
+	first_waiter = prepare_to_wait_exclusive(&rqw->wait, &data.wq,
 						 TASK_UNINTERRUPTIBLE);
+	/*
+	 * Make sure there is at least one inflight process; otherwise, waiters
+	 * will never be woken up. Since there may be no inflight process before
+	 * adding ourselves to the waiting queue above, we need to try to
+	 * increase the inflight counter for ourselves. And it is sufficient to
+	 * guarantee that at least the first waiter to enter the waiting queue
+	 * will re-check the waiting condition before going to sleep, thus
+	 * ensuring forward progress.
+	 */
+	if (!data.got_token && first_waiter && acquire_inflight_cb(rqw, private_data)) {
+		finish_wait(&rqw->wait, &data.wq);
+		/*
+		 * We raced with rq_qos_wake_function() getting a token,
+		 * which means we now have two. Put our local token
+		 * and wake anyone else potentially waiting for one.
+		 *
+		 * Enough memory barrier in list_empty_careful() in
+		 * finish_wait() is paired with list_del_init_careful()
+		 * in rq_qos_wake_function() to make sure we will see
+		 * the latest @data->got_token.
+		 */
+		if (data.got_token)
+			cleanup_cb(rqw, private_data);
+		return;
+	}
+
+	/* we are now relying on the waker to increase our inflight counter. */
 	do {
-		/* The memory barrier in set_current_state saves us here. */
 		if (data.got_token)
 			break;
-		if (!has_sleeper && acquire_inflight_cb(rqw, private_data)) {
-			finish_wait(&rqw->wait, &data.wq);
-
-			/*
-			 * We raced with rq_qos_wake_function() getting a token,
-			 * which means we now have two. Put our local token
-			 * and wake anyone else potentially waiting for one.
-			 */
-			if (data.got_token)
-				cleanup_cb(rqw, private_data);
-			return;
-		}
 		io_schedule();
-		has_sleeper = true;
 		set_current_state(TASK_UNINTERRUPTIBLE);
 	} while (1);
 	finish_wait(&rqw->wait, &data.wq);
@@ -292,6 +318,7 @@ void rq_qos_exit(struct request_queue *q)
 		q->rq_qos = rqos->next;
 		rqos->ops->exit(rqos);
 	}
+	blk_queue_flag_clear(QUEUE_FLAG_QOS_ENABLED, q);
 	mutex_unlock(&q->rq_qos_mutex);
 }
 
@@ -299,6 +326,7 @@ int rq_qos_add(struct rq_qos *rqos, struct gendisk *disk, enum rq_qos_id id,
 		const struct rq_qos_ops *ops)
 {
 	struct request_queue *q = disk->queue;
+	unsigned int memflags;
 
 	lockdep_assert_held(&q->rq_qos_mutex);
 
@@ -310,14 +338,15 @@ int rq_qos_add(struct rq_qos *rqos, struct gendisk *disk, enum rq_qos_id id,
 	 * No IO can be in-flight when adding rqos, so freeze queue, which
 	 * is fine since we only support rq_qos for blk-mq queue.
 	 */
-	blk_mq_freeze_queue(q);
+	memflags = blk_mq_freeze_queue(q);
 
 	if (rq_qos_id(q, rqos->id))
 		goto ebusy;
 	rqos->next = q->rq_qos;
 	q->rq_qos = rqos;
+	blk_queue_flag_set(QUEUE_FLAG_QOS_ENABLED, q);
 
-	blk_mq_unfreeze_queue(q);
+	blk_mq_unfreeze_queue(q, memflags);
 
 	if (rqos->ops->debugfs_attrs) {
 		mutex_lock(&q->debugfs_mutex);
@@ -327,7 +356,7 @@ int rq_qos_add(struct rq_qos *rqos, struct gendisk *disk, enum rq_qos_id id,
 
 	return 0;
 ebusy:
-	blk_mq_unfreeze_queue(q);
+	blk_mq_unfreeze_queue(q, memflags);
 	return -EBUSY;
 }
 
@@ -335,17 +364,20 @@ void rq_qos_del(struct rq_qos *rqos)
 {
 	struct request_queue *q = rqos->disk->queue;
 	struct rq_qos **cur;
+	unsigned int memflags;
 
 	lockdep_assert_held(&q->rq_qos_mutex);
 
-	blk_mq_freeze_queue(q);
+	memflags = blk_mq_freeze_queue(q);
 	for (cur = &q->rq_qos; *cur; cur = &(*cur)->next) {
 		if (*cur == rqos) {
 			*cur = rqos->next;
 			break;
 		}
 	}
-	blk_mq_unfreeze_queue(q);
+	if (!q->rq_qos)
+		blk_queue_flag_clear(QUEUE_FLAG_QOS_ENABLED, q);
+	blk_mq_unfreeze_queue(q, memflags);
 
 	mutex_lock(&q->debugfs_mutex);
 	blk_mq_debugfs_unregister_rqos(rqos);