workqueue: fix ghost PENDING flag while doing MQ IO
authorRoman Pen <roman.penyaev@profitbricks.com>
Tue, 26 Apr 2016 11:15:35 +0000 (13:15 +0200)
committerTejun Heo <tj@kernel.org>
Tue, 26 Apr 2016 15:23:22 +0000 (11:23 -0400)
The bug in a workqueue leads to a stalled IO request in MQ ctx->rq_list
with the following backtrace:

[  601.347452] INFO: task kworker/u129:5:1636 blocked for more than 120 seconds.
[  601.347574]       Tainted: G           O    4.4.5-1-storage+ #6
[  601.347651] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
[  601.348142] kworker/u129:5  D ffff880803077988     0  1636      2 0x00000000
[  601.348519] Workqueue: ibnbd_server_fileio_wq ibnbd_dev_file_submit_io_worker [ibnbd_server]
[  601.348999]  ffff880803077988 ffff88080466b900 ffff8808033f9c80 ffff880803078000
[  601.349662]  ffff880807c95000 7fffffffffffffff ffffffff815b0920 ffff880803077ad0
[  601.350333]  ffff8808030779a0 ffffffff815b01d5 0000000000000000 ffff880803077a38
[  601.350965] Call Trace:
[  601.351203]  [<ffffffff815b0920>] ? bit_wait+0x60/0x60
[  601.351444]  [<ffffffff815b01d5>] schedule+0x35/0x80
[  601.351709]  [<ffffffff815b2dd2>] schedule_timeout+0x192/0x230
[  601.351958]  [<ffffffff812d43f7>] ? blk_flush_plug_list+0xc7/0x220
[  601.352208]  [<ffffffff810bd737>] ? ktime_get+0x37/0xa0
[  601.352446]  [<ffffffff815b0920>] ? bit_wait+0x60/0x60
[  601.352688]  [<ffffffff815af784>] io_schedule_timeout+0xa4/0x110
[  601.352951]  [<ffffffff815b3a4e>] ? _raw_spin_unlock_irqrestore+0xe/0x10
[  601.353196]  [<ffffffff815b093b>] bit_wait_io+0x1b/0x70
[  601.353440]  [<ffffffff815b056d>] __wait_on_bit+0x5d/0x90
[  601.353689]  [<ffffffff81127bd0>] wait_on_page_bit+0xc0/0xd0
[  601.353958]  [<ffffffff81096db0>] ? autoremove_wake_function+0x40/0x40
[  601.354200]  [<ffffffff81127cc4>] __filemap_fdatawait_range+0xe4/0x140
[  601.354441]  [<ffffffff81127d34>] filemap_fdatawait_range+0x14/0x30
[  601.354688]  [<ffffffff81129a9f>] filemap_write_and_wait_range+0x3f/0x70
[  601.354932]  [<ffffffff811ced3b>] blkdev_fsync+0x1b/0x50
[  601.355193]  [<ffffffff811c82d9>] vfs_fsync_range+0x49/0xa0
[  601.355432]  [<ffffffff811cf45a>] blkdev_write_iter+0xca/0x100
[  601.355679]  [<ffffffff81197b1a>] __vfs_write+0xaa/0xe0
[  601.355925]  [<ffffffff81198379>] vfs_write+0xa9/0x1a0
[  601.356164]  [<ffffffff811c59d8>] kernel_write+0x38/0x50

The underlying device is a null_blk, with default parameters:

  queue_mode    = MQ
  submit_queues = 1

Verification that nullb0 has something inflight:

root@pserver8:~# cat /sys/block/nullb0/inflight
       0        1
root@pserver8:~# find /sys/block/nullb0/mq/0/cpu* -name rq_list -print -exec cat {} \;
...
/sys/block/nullb0/mq/0/cpu2/rq_list
CTX pending:
        ffff8838038e2400
...

During debug it became clear that stalled request is always inserted in
the rq_list from the following path:

   save_stack_trace_tsk + 34
   blk_mq_insert_requests + 231
   blk_mq_flush_plug_list + 281
   blk_flush_plug_list + 199
   wait_on_page_bit + 192
   __filemap_fdatawait_range + 228
   filemap_fdatawait_range + 20
   filemap_write_and_wait_range + 63
   blkdev_fsync + 27
   vfs_fsync_range + 73
   blkdev_write_iter + 202
   __vfs_write + 170
   vfs_write + 169
   kernel_write + 56

So blk_flush_plug_list() was called with from_schedule == true.

If from_schedule is true, that means that finally blk_mq_insert_requests()
offloads execution of __blk_mq_run_hw_queue() and uses kblockd workqueue,
i.e. it calls kblockd_schedule_delayed_work_on().

That means, that we race with another CPU, which is about to execute
__blk_mq_run_hw_queue() work.

Further debugging shows the following traces from different CPUs:

  CPU#0                                  CPU#1
  ----------------------------------     -------------------------------
  reqeust A inserted
  STORE hctx->ctx_map[0] bit marked
  kblockd_schedule...() returns 1
  <schedule to kblockd workqueue>
                                         request B inserted
                                         STORE hctx->ctx_map[1] bit marked
                                         kblockd_schedule...() returns 0
  *** WORK PENDING bit is cleared ***
  flush_busy_ctxs() is executed, but
  bit 1, set by CPU#1, is not observed

As a result request B pended forever.

This behaviour can be explained by speculative LOAD of hctx->ctx_map on
CPU#0, which is reordered with clear of PENDING bit and executed _before_
actual STORE of bit 1 on CPU#1.

The proper fix is an explicit full barrier <mfence>, which guarantees
that clear of PENDING bit is to be executed before all possible
speculative LOADS or STORES inside actual work function.

Signed-off-by: Roman Pen <roman.penyaev@profitbricks.com>
Cc: Gioh Kim <gi-oh.kim@profitbricks.com>
Cc: Michael Wang <yun.wang@profitbricks.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: linux-block@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Cc: stable@vger.kernel.org
Signed-off-by: Tejun Heo <tj@kernel.org>
kernel/workqueue.c

index 16e13d8628a301c61e2178df18756ba7483d4941..801a698564d49b7baa615a8c090f6737f2abaf05 100644 (file)
@@ -666,6 +666,35 @@ static void set_work_pool_and_clear_pending(struct work_struct *work,
         */
        smp_wmb();
        set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
+       /*
+        * The following mb guarantees that previous clear of a PENDING bit
+        * will not be reordered with any speculative LOADS or STORES from
+        * work->current_func, which is executed afterwards.  This possible
+        * reordering can lead to a missed execution on attempt to qeueue
+        * the same @work.  E.g. consider this case:
+        *
+        *   CPU#0                         CPU#1
+        *   ----------------------------  --------------------------------
+        *
+        * 1  STORE event_indicated
+        * 2  queue_work_on() {
+        * 3    test_and_set_bit(PENDING)
+        * 4 }                             set_..._and_clear_pending() {
+        * 5                                 set_work_data() # clear bit
+        * 6                                 smp_mb()
+        * 7                               work->current_func() {
+        * 8                                  LOAD event_indicated
+        *                                 }
+        *
+        * Without an explicit full barrier speculative LOAD on line 8 can
+        * be executed before CPU#0 does STORE on line 1.  If that happens,
+        * CPU#0 observes the PENDING bit is still set and new execution of
+        * a @work is not queued in a hope, that CPU#1 will eventually
+        * finish the queued @work.  Meanwhile CPU#1 does not see
+        * event_indicated is set, because speculative LOAD was executed
+        * before actual STORE.
+        */
+       smp_mb();
 }
 
 static void clear_work_data(struct work_struct *work)