scsi: fix scsi_error_handler vs. scsi_host_dev_release race

commit 537b604c upstream.

b9d5c6b7 ("[SCSI] cleanup setting task state in
scsi_error_handler()") has introduced a race between scsi_error_handler
and scsi_host_dev_release resulting in the hang when the device goes
away because scsi_error_handler might miss a wake up:

CPU0					CPU1
scsi_error_handler			scsi_host_dev_release
  					  kthread_stop()
  kthread_should_stop()
    test_bit(KTHREAD_SHOULD_STOP)
					    set_bit(KTHREAD_SHOULD_STOP)
					    wake_up_process()
					    wait_for_completion()

  set_current_state(TASK_INTERRUPTIBLE)
  schedule()

The most straightforward solution seems to be to invert the ordering of
the set_current_state and kthread_should_stop.

The issue has been noticed during reboot test on a 3.0 based kernel but
the current code seems to be affected in the same way.

[jejb: additional comment added]
Reported-and-debugged-by: Mike Mayer <Mike.Meyer@teradata.com>
Signed-off-by: Michal Hocko <mhocko@suse.com>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Reviewed-by: Hannes Reinecke <hare@suse.de>
Signed-off-by: James Bottomley <JBottomley@Odin.com>
Signed-off-by: Kamal Mostafa <kamal@canonical.com>

drivers/scsi/scsi_error.c

@@ -2025,8 +2025,17 @@ int scsi_error_handler(void *data)
 	 * We never actually get interrupted because kthread_run
 	 * disables signal delivery for the created thread.
 	 */
-	while (!kthread_should_stop()) {
+	while (true) {
+		/*
+		 * The sequence in kthread_stop() sets the stop flag first
+		 * then wakes the process.  To avoid missed wakeups, the task
+		 * should always be in a non running state before the stop
+		 * flag is checked
+		 */
 		set_current_state(TASK_INTERRUPTIBLE);
+		if (kthread_should_stop())
+			break;
+
 		if ((shost->host_failed == 0 && shost->host_eh_scheduled == 0) ||
 		    shost->host_failed != shost->host_busy) {
 			SCSI_LOG_ERROR_RECOVERY(1,