blk-mq: fix waitqueue_active without memory barrier in block/blk-mq-tag.c
authorKosuke Tatsukawa <tatsu@ab.jp.nec.com>
Fri, 9 Oct 2015 00:35:38 +0000 (00:35 +0000)
committerJens Axboe <axboe@fb.com>
Fri, 9 Oct 2015 16:52:46 +0000 (10:52 -0600)
commit8ee1b7b9d8761a7b1a0f37a6e596b5315d27d2f2
tree909a288ceefca515fe3a03eb183c1a2c5e3044d3
parentfd48ca384960c831d4fae4636282e422e140ca7c
blk-mq: fix waitqueue_active without memory barrier in block/blk-mq-tag.c

blk_mq_tag_update_depth() seems to be missing a memory barrier which
might cause the waker to not notice the waiter and fail to send a
wake_up as in the following figure.

blk_mq_tag_update_depth bt_get
------------------------------------------------------------------------
if (waitqueue_active(&bs->wait))
/* The CPU might reorder the test for
   the waitqueue up here, before
   prior writes complete */
prepare_to_wait(&bs->wait, &wait,
  TASK_UNINTERRUPTIBLE);
tag = __bt_get(hctx, bt, last_tag,
  tags);
/* Value set in bt_update_count not
   visible yet */
bt_update_count(&tags->bitmap_tags, tdepth);
/* blk_mq_tag_wakeup_all(tags, false); */
 bt = &tags->bitmap_tags;
 wake_index = atomic_read(&bt->wake_index);
...
io_schedule();
------------------------------------------------------------------------

This patch adds the missing memory barrier.

I found this issue when I was looking through the linux source code
for places calling waitqueue_active() before wake_up*(), but without
preceding memory barriers, after sending a patch to fix a similar
issue in drivers/tty/n_tty.c  (Details about the original issue can be
found here: https://lkml.org/lkml/2015/9/28/849).

Signed-off-by: Kosuke Tatsukawa <tatsu@ab.jp.nec.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
block/blk-mq-tag.c