[SCSI] libsas: Reset timer on taskless scsi_cmnds in sas_scsi_timed_out

Every so often, a scsi_cmnd will time out, and the libsas timeout handler
will discover that the scsi_cmnd does not have a sas_task attached to it.
This can happen in two cases: (1) the scsi_cmnd actually made it through
libsas to the HBA and is now going through scsi_done, or (2) the
scsi_cmnd has been held up (host lock, slab alloc, etc) and libsas has
not yet attached a sas_task.  In both cases, it is safe to ask SCSI for
more time to process the command via EH_RESET_TIMER; we cannot blindly
return EH_HANDLED because if (2) happens, we could end up calling
scsi_done while another CPU is heading towards sas_queuecommand, which
causes slab corruption when sas_task_done updates the freed scsi_cmnd.
Signed-off-by: Darrick J. Wong <djwong@us.ibm.com>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>

drivers/scsi/libsas/sas_scsi_host.c

@@ -524,9 +524,13 @@ enum scsi_eh_timer_return sas_scsi_timed_out(struct scsi_cmnd *cmd)
 	unsigned long flags;
 
 	if (!task) {
-		SAS_DPRINTK("command 0x%p, task 0x%p, gone: EH_HANDLED\n",
-			    cmd, task);
-		return EH_HANDLED;
+		cmd->timeout_per_command /= 2;
+		SAS_DPRINTK("command 0x%p, task 0x%p, gone: %s\n",
+			    cmd, task, (cmd->timeout_per_command ?
+			    "EH_RESET_TIMER" : "EH_NOT_HANDLED"));
+		if (!cmd->timeout_per_command)
+			return EH_NOT_HANDLED;
+		return EH_RESET_TIMER;
 	}
 
 	spin_lock_irqsave(&task->task_state_lock, flags);