backing_dev: fix hung task on sync
Derek Basehore [Thu, 3 Apr 2014 21:46:22 +0000 (14:46 -0700)]
commit 6ca738d60c563d5c6cf6253ee4b8e76fa77b2b9e upstream.

bdi_wakeup_thread_delayed() used the mod_delayed_work() function to
schedule work to writeback dirty inodes.  The problem with this is that
it can delay work that is scheduled for immediate execution, such as the
work from sync_inodes_sb().  This can happen since mod_delayed_work()
can now steal work from a work_queue.  This fixes the problem by using
queue_delayed_work() instead.  This is a regression caused by commit
839a8e8660b6 ("writeback: replace custom worker pool implementation with
unbound workqueue").

The reason that this causes a problem is that laptop-mode will change
the delay, dirty_writeback_centisecs, to 60000 (10 minutes) by default.
In the case that bdi_wakeup_thread_delayed() races with
sync_inodes_sb(), sync will be stopped for 10 minutes and trigger a hung
task.  Even if dirty_writeback_centisecs is not long enough to cause a
hung task, we still don't want to delay sync for that long.

We fix the problem by using queue_delayed_work() when we want to
schedule writeback sometime in future.  This function doesn't change the
timer if it is already armed.

For the same reason, we also change bdi_writeback_workfn() to
immediately queue the work again in the case that the work_list is not
empty.  The same problem can happen if the sync work is run on the
rescue worker.

[ update changelog, add comment, use bdi_wakeup_thread_delayed()]
Signed-off-by: Derek Basehore <>
Reviewed-by: Jan Kara <>
Cc: Alexander Viro <>
Reviewed-by: Tejun Heo <>
Cc: Greg Kroah-Hartman <>
Cc: "Darrick J. Wong" <>
Cc: Derek Basehore <>
Cc: Kees Cook <>
Cc: Benson Leung <>
Cc: Sonny Rao <>
Cc: Luigi Semenzato <>
Cc: Jens Axboe <>
Cc: Dave Chinner <>
Signed-off-by: Andrew Morton <>
Signed-off-by: Linus Torvalds <>
Signed-off-by: Greg Kroah-Hartman <>


index e3ab1e4..f79f641 100644 (file)
@@ -1038,10 +1038,10 @@ void bdi_writeback_workfn(struct work_struct *work)
-       if (!list_empty(&bdi->work_list) ||
-           (wb_has_dirty_io(wb) && dirty_writeback_interval))
-               queue_delayed_work(bdi_wq, &wb->dwork,
-                       msecs_to_jiffies(dirty_writeback_interval * 10));
+       if (!list_empty(&bdi->work_list))
+               mod_delayed_work(bdi_wq, &wb->dwork, 0);
+       else if (wb_has_dirty_io(wb) && dirty_writeback_interval)
+               bdi_wakeup_thread_delayed(bdi);
        current->flags &= ~PF_SWAPWRITE;
index 5025174..95e4374 100644 (file)
@@ -287,13 +287,16 @@ int bdi_has_dirty_io(struct backing_dev_info *bdi)
  * Note, we wouldn't bother setting up the timer, but this function is on the
  * fast-path (used by '__mark_inode_dirty()'), so we save few context switches
  * by delaying the wake-up.
+ *
+ * We have to be careful not to postpone flush work if it is scheduled for
+ * earlier. Thus we use queue_delayed_work().
 void bdi_wakeup_thread_delayed(struct backing_dev_info *bdi)
        unsigned long timeout;
        timeout = msecs_to_jiffies(dirty_writeback_interval * 10);
-       mod_delayed_work(bdi_wq, &bdi->wb.dwork, timeout);
+       queue_delayed_work(bdi_wq, &bdi->wb.dwork, timeout);