]> git.kernelconcepts.de Git - karo-tx-linux.git/commit
[SCSI] sd: call blk_pm_runtime_init before add_disk
authorAaron Lu <aaron.lu@intel.com>
Thu, 10 Oct 2013 05:22:36 +0000 (13:22 +0800)
committerJames Bottomley <JBottomley@Parallels.com>
Wed, 23 Oct 2013 13:09:18 +0000 (14:09 +0100)
commit10c580e4239df5c3344ca00322eca86ab2de880b
tree81425dadc373bb8dd7299d81c736e78f3bb87c62
parent36008cf118235cee49b6753455f33b6f2c3a7543
[SCSI] sd: call blk_pm_runtime_init before add_disk

Sujit has found a race condition that would make q->nr_pending
unbalanced, it occurs as Sujit explained:

"
sd_probe_async() ->
add_disk() ->
disk_add_event() ->
schedule(disk_events_workfn)
sd_revalidate_disk()
blk_pm_runtime_init()
return;

Let's say the disk_events_workfn() calls sd_check_events() which tries
to send test_unit_ready() and because of sd_revalidate_disk() trying to
send another commands the test_unit_ready() might be re-queued as the
tagged command queuing is disabled.

So the race condition is -

Thread 1    | Thread 2
sd_revalidate_disk()   | sd_check_events()
...nr_pending = 0 as q->dev = NULL| scsi_queue_insert()
blk_runtime_pm_init()   |  blk_pm_requeue_request() ->
  | nr_pending = -1 since
  | q->dev != NULL
"

The problem is, the test_unit_ready request doesn't get counted the
first time it is queued, so the later decrement of q->nr_pending in
blk_pm_requeue_request makes it unbalanced.

Fix this by calling blk_pm_runtime_init before add_disk so that all
requests initiated there will all be counted.

Signed-off-by: Aaron Lu <aaron.lu@intel.com>
Reported-and-tested-by: Sujit Reddy Thumma <sthumma@codeaurora.org>
Cc: stable@vger.kernel.org
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
drivers/scsi/sd.c