[PATCH v2 2/2] block: refactor rq_qos_wait()

[Muchun Song <songmuchun@bytedance.com>]

When rq_qos_wait() is first introduced, it is easy to understand. But
with some bug fixes applied, it is not easy for newcomers to understand
the whole logic under those fixes. In this patch, rq_qos_wait() is
refactored and more comments are added for better understanding. There
are 3 points for the improvement:

1) Use waitqueue_active() instead of wq_has_sleeper() to eliminate
   unnecessary memory barrier in wq_has_sleeper() which is supposed
   to be used in waker side. In this case, we do need the barrier.
   So use the cheaper one to locklessly test for waiters on the queue.

2) Remove acquire_inflight_cb() logic for the first waiter out of the
   while loop to make the code clear.

3) Add more comments to explain how to sync with different waiters and
   the waker.

Signed-off-by: Muchun Song <songmuchun@bytedance.com>
Reviewed-by: Yu Kuai <yukuai3@huawei.com>
---
v2:
 - Introduce init_wait_func() in a seprate patch (Yu Kuai).

 block/blk-rq-qos.c | 68 ++++++++++++++++++++++++++++++++--------------
 1 file changed, 47 insertions(+), 21 deletions(-)

diff --git a/block/blk-rq-qos.c b/block/blk-rq-qos.c
index 858ce69c04ece..5d995d389eaf5 100644
--- a/block/blk-rq-qos.c
+++ b/block/blk-rq-qos.c
@@ -223,6 +223,14 @@ static int rq_qos_wake_function(struct wait_queue_entry *curr,
 	 * 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;
 }
@@ -248,37 +256,55 @@ void rq_qos_wait(struct rq_wait *rqw, void *private_data,
 		 cleanup_cb_t *cleanup_cb)
 {
 	struct rq_qos_wait_data data = {
-		.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;
 
 	init_wait_func(&data.wq, rq_qos_wake_function);
-	has_sleeper = !prepare_to_wait_exclusive(&rqw->wait, &data.wq,
+	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_task_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);
-- 
2.20.1