Fix scsi_unblock_requests()

In its current implementation, scsi_unblock_requests() simply clears host_self_blocked in the SCSI host struct. This means that either a transaction must complete or a new transaction be issued before the mid-layer will recognize that it can run the queues. There is no guarantee that either of these events will ever happen.

This patch attempts to run the queues manually when the host unblocks. scsi_queue_next_request() verifies all other state to ensure queuing new transactions is safe prior to proceeding.

This patch is part of versions 6.1.10 and 6.1.11 of the aic7xxx driver. Its used, along with scsi_block_requests(), to hold off the mid-layer during the initial bus settle delay without resorting to a busy loop.

--
Justin

diff -u -r -N linux-2.4.3.virgin/drivers/scsi/scsi_lib.c linux-2.4.3/drivers/scsi/scsi_lib.c
@@ -1108,9 +1108,13 @@
  */
 void scsi_unblock_requests(struct Scsi_Host * SHpnt)
 {
+	Scsi_Device *SDloop;
+
 	SHpnt->host_self_blocked = FALSE;
+	/* Now that we are unblocked, try to start the queues. */
+	for (SDloop = SHpnt->host_queue; SDloop; SDloop = SDloop->next)
+		scsi_queue_next_request(&SDloop->request_queue, NULL);
 }
-
 
 /*
  * Function:    scsi_report_bus_reset()

Suppose the queue is unblocked from inside the functions called to process the request. In that situation the old code is correct and your code might introduce other problems

Including recursion ?

The patch seems right apart from checking these details out further.

Re-entrancy is only prevented by holding the io-request lock or in some cases a per-controller or per-controller driver lock. As the comment in scsi_unblock_requests() states, this API assumes that no locks are held. If you hold a lock in calling this routine, you may deadlock.

I suppose a poorly implemented use of this API could tail recurse. The aic7xxx driver calls this routine from a timeout handler so there is no risk of stack explosion.

Well, the patch was written to have minimal impact to an API that has never been used. A more correct solution might be to queue the affected host controller to a queue drained by a bottom-half handler.

--
Justin