block: Mitigate lock unbalance caused by lock switching
authorAsias He <asias@redhat.com>
Thu, 24 May 2012 15:28:52 +0000 (23:28 +0800)
committerJens Axboe <axboe@kernel.dk>
Fri, 15 Jun 2012 06:46:22 +0000 (08:46 +0200)
Commit 777eb1bf15b8532c396821774bf6451e563438f5 disconnects externally
supplied queue_lock before blk_drain_queue(). Switching the lock would
introduce lock unbalance because theads which have taken the external
lock might unlock the internal lock in the during the queue drain. This
patch mitigate this by disconnecting the lock after the queue draining
since queue draining makes a lot of request_queue users go away.

However, please note, this patch only makes the problem less likely to
happen. Anyone who still holds a ref might try to issue a new request on
a dead queue after the blk_cleanup_queue() finishes draining, the lock
unbalance might still happen in this case.

 =====================================
 [ BUG: bad unlock balance detected! ]
 3.4.0+ #288 Not tainted
 -------------------------------------
 fio/17706 is trying to release lock (&(&q->__queue_lock)->rlock) at:
 [<ffffffff81329372>] blk_queue_bio+0x2a2/0x380
 but there are no more locks to release!

 other info that might help us debug this:
 1 lock held by fio/17706:
  #0:  (&(&vblk->lock)->rlock){......}, at: [<ffffffff81327f1a>]
 get_request_wait+0x19a/0x250

 stack backtrace:
 Pid: 17706, comm: fio Not tainted 3.4.0+ #288
 Call Trace:
  [<ffffffff81329372>] ? blk_queue_bio+0x2a2/0x380
  [<ffffffff810dea49>] print_unlock_inbalance_bug+0xf9/0x100
  [<ffffffff810dfe4f>] lock_release_non_nested+0x1df/0x330
  [<ffffffff811dae24>] ? dio_bio_end_aio+0x34/0xc0
  [<ffffffff811d6935>] ? bio_check_pages_dirty+0x85/0xe0
  [<ffffffff811daea1>] ? dio_bio_end_aio+0xb1/0xc0
  [<ffffffff81329372>] ? blk_queue_bio+0x2a2/0x380
  [<ffffffff81329372>] ? blk_queue_bio+0x2a2/0x380
  [<ffffffff810e0079>] lock_release+0xd9/0x250
  [<ffffffff81a74553>] _raw_spin_unlock_irq+0x23/0x40
  [<ffffffff81329372>] blk_queue_bio+0x2a2/0x380
  [<ffffffff81328faa>] generic_make_request+0xca/0x100
  [<ffffffff81329056>] submit_bio+0x76/0xf0
  [<ffffffff8115470c>] ? set_page_dirty_lock+0x3c/0x60
  [<ffffffff811d69e1>] ? bio_set_pages_dirty+0x51/0x70
  [<ffffffff811dd1a8>] do_blockdev_direct_IO+0xbf8/0xee0
  [<ffffffff811d8620>] ? blkdev_get_block+0x80/0x80
  [<ffffffff811dd4e5>] __blockdev_direct_IO+0x55/0x60
  [<ffffffff811d8620>] ? blkdev_get_block+0x80/0x80
  [<ffffffff811d92e7>] blkdev_direct_IO+0x57/0x60
  [<ffffffff811d8620>] ? blkdev_get_block+0x80/0x80
  [<ffffffff8114c6ae>] generic_file_aio_read+0x70e/0x760
  [<ffffffff810df7c5>] ? __lock_acquire+0x215/0x5a0
  [<ffffffff811e9924>] ? aio_run_iocb+0x54/0x1a0
  [<ffffffff8114bfa0>] ? grab_cache_page_nowait+0xc0/0xc0
  [<ffffffff811e82cc>] aio_rw_vect_retry+0x7c/0x1e0
  [<ffffffff811e8250>] ? aio_fsync+0x30/0x30
  [<ffffffff811e9936>] aio_run_iocb+0x66/0x1a0
  [<ffffffff811ea9b0>] do_io_submit+0x6f0/0xb80
  [<ffffffff8134de2e>] ? trace_hardirqs_on_thunk+0x3a/0x3f
  [<ffffffff811eae50>] sys_io_submit+0x10/0x20
  [<ffffffff81a7c9e9>] system_call_fastpath+0x16/0x1b

Changes since v2: Update commit log to explain how the code is still
                  broken even if we delay the lock switching after the drain.
Changes since v1: Update commit log as Tejun suggested.

Acked-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Asias He <asias@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
block/blk-core.c

index ce7fbf8d85a609e7754148231d8e87b43d635616..93eb3e4f88ce78affc79c5ca6d8b7c7b0759be5a 100644 (file)
@@ -480,7 +480,6 @@ void blk_cleanup_queue(struct request_queue *q)
        /* mark @q DEAD, no new request or merges will be allowed afterwards */
        mutex_lock(&q->sysfs_lock);
        queue_flag_set_unlocked(QUEUE_FLAG_DEAD, q);
-
        spin_lock_irq(lock);
 
        /*
@@ -498,10 +497,6 @@ void blk_cleanup_queue(struct request_queue *q)
        queue_flag_set(QUEUE_FLAG_NOMERGES, q);
        queue_flag_set(QUEUE_FLAG_NOXMERGES, q);
        queue_flag_set(QUEUE_FLAG_DEAD, q);
-
-       if (q->queue_lock != &q->__queue_lock)
-               q->queue_lock = &q->__queue_lock;
-
        spin_unlock_irq(lock);
        mutex_unlock(&q->sysfs_lock);
 
@@ -512,6 +507,11 @@ void blk_cleanup_queue(struct request_queue *q)
        del_timer_sync(&q->backing_dev_info.laptop_mode_wb_timer);
        blk_sync_queue(q);
 
+       spin_lock_irq(lock);
+       if (q->queue_lock != &q->__queue_lock)
+               q->queue_lock = &q->__queue_lock;
+       spin_unlock_irq(lock);
+
        /* @q is and will stay empty, shutdown and put */
        blk_put_queue(q);
 }