[SCSI] fc transport: resolve scan vs delete deadlocks
James Smart [Thu, 11 May 2006 17:27:09 +0000 (13:27 -0400)]
In a prior posting to linux-scsi on the fc transport and workq
deadlocks, we noted a second error that did not have a patch:
  http://marc.theaimsgroup.com/?l=linux-scsi&m=114467847711383&w=2
  - There's a deadlock where scsi_remove_target() has to sit behind
    scsi_scan_target() due to contention over the scan_lock().

Subsequently we posted a request for comments about the deadlock:
  http://marc.theaimsgroup.com/?l=linux-scsi&m=114469358829500&w=2

This posting resolves the second error. Here's what we now understand,
and are implementing:

  If the lldd deletes the rport while a scan is active, the sdev's queue
  is blocked which stops the issuing of commands associated with the scan.
  At this point, the scan stalls, and does so with the shost->scan_mutex held.
  If, at this point, if any scan or delete request is made on the host, it
  will stall waiting for the scan_mutex.

  For the FC transport, we queue all delete work to a single workq.
  So, things worked fine when competing with the scan, as long as the
  target blocking the scan was the same target at the top of our delete
  workq, as the delete workq routine always unblocked just prior to
  requesting the delete.  Unfortunately, if the top of our delete workq
  was for a different target, we deadlock.  Additionally, if the target
  blocking scan returned, we were unblocking it in the scan workq routine,
  which really won't execute until the existing stalled scan workq
  completes (e.g. we're re-scheduling it while it is in the midst of its
  execution).

  This patch moves the unblock out of the workq routines and moves it to
  the context that is scheduling the work. This ensures that at some point,
  we will unblock the target that is blocking scan.  Please note, however,
  that the deadlock condition may still occur while it waits for the
  transport to timeout an unblock on a target.  Worst case, this is bounded
  by the transport dev_loss_tmo (default: 30 seconds).

Finally, Michael Reed deserves the credit for the bulk of this patch,
analysis, and it's testing. Thank you for your help.

Note: The request for comments statements about the gross-ness of the
  scan_mutex still stand.

Signed-off-by: Michael Reed <mdr@sgi.com>
Signed-off-by: James Smart <james.smart@emulex.com>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>

drivers/scsi/scsi_transport_fc.c

index f2db7a4..c76e73a 100644 (file)
@@ -1284,7 +1284,9 @@ EXPORT_SYMBOL(fc_release_transport);
  * @work:      Work to queue for execution.
  *
  * Return value:
- *     0 on success / != 0 for error
+ *     1 - work queued for execution
+ *     0 - work is already queued
+ *     -EINVAL - work queue doesn't exist
  **/
 static int
 fc_queue_work(struct Scsi_Host *shost, struct work_struct *work)
@@ -1434,8 +1436,6 @@ fc_starget_delete(void *data)
        struct Scsi_Host *shost = rport_to_shost(rport);
        unsigned long flags;
 
-       scsi_target_unblock(&rport->dev);
-
        spin_lock_irqsave(shost->host_lock, flags);
        if (rport->flags & FC_RPORT_DEVLOSS_PENDING) {
                spin_unlock_irqrestore(shost->host_lock, flags);
@@ -1707,6 +1707,8 @@ fc_remote_port_add(struct Scsi_Host *shost, int channel,
 
                                spin_unlock_irqrestore(shost->host_lock, flags);
 
+                               scsi_target_unblock(&rport->dev);
+
                                return rport;
                        }
                }
@@ -1762,9 +1764,10 @@ fc_remote_port_add(struct Scsi_Host *shost, int channel,
                                /* initiate a scan of the target */
                                rport->flags |= FC_RPORT_SCAN_PENDING;
                                scsi_queue_work(shost, &rport->scan_work);
-                       }
-
-                       spin_unlock_irqrestore(shost->host_lock, flags);
+                               spin_unlock_irqrestore(shost->host_lock, flags);
+                               scsi_target_unblock(&rport->dev);
+                       } else
+                               spin_unlock_irqrestore(shost->host_lock, flags);
 
                        return rport;
                }
@@ -1938,6 +1941,7 @@ fc_remote_port_rolechg(struct fc_rport  *rport, u32 roles)
                rport->flags |= FC_RPORT_SCAN_PENDING;
                scsi_queue_work(shost, &rport->scan_work);
                spin_unlock_irqrestore(shost->host_lock, flags);
+               scsi_target_unblock(&rport->dev);
        }
 }
 EXPORT_SYMBOL(fc_remote_port_rolechg);
@@ -1970,8 +1974,9 @@ fc_timeout_deleted_rport(void  *data)
                dev_printk(KERN_ERR, &rport->dev,
                        "blocked FC remote port time out: no longer"
                        " a FCP target, removing starget\n");
-               fc_queue_work(shost, &rport->stgt_delete_work);
                spin_unlock_irqrestore(shost->host_lock, flags);
+               scsi_target_unblock(&rport->dev);
+               fc_queue_work(shost, &rport->stgt_delete_work);
                return;
        }
 
@@ -2035,17 +2040,15 @@ fc_timeout_deleted_rport(void  *data)
         * went away and didn't come back - we'll remove
         * all attached scsi devices.
         */
-       fc_queue_work(shost, &rport->stgt_delete_work);
-
        spin_unlock_irqrestore(shost->host_lock, flags);
+
+       scsi_target_unblock(&rport->dev);
+       fc_queue_work(shost, &rport->stgt_delete_work);
 }
 
 /**
  * fc_scsi_scan_rport - called to perform a scsi scan on a remote port.
  *
- * Will unblock the target (in case it went away and has now come back),
- * then invoke a scan.
- *
  * @data:      remote port to be scanned.
  **/
 static void
@@ -2057,7 +2060,6 @@ fc_scsi_scan_rport(void *data)
 
        if ((rport->port_state == FC_PORTSTATE_ONLINE) &&
            (rport->roles & FC_RPORT_ROLE_FCP_TARGET)) {
-               scsi_target_unblock(&rport->dev);
                scsi_scan_target(&rport->dev, rport->channel,
                        rport->scsi_target_id, SCAN_WILD_CARD, 1);
        }