scsi: convert to blk_delay_queue()
authorJens Axboe <jaxboe@fusionio.com>
Fri, 16 Apr 2010 19:13:15 +0000 (21:13 +0200)
committerJens Axboe <jaxboe@fusionio.com>
Thu, 10 Mar 2011 07:45:54 +0000 (08:45 +0100)
It was always abuse to reuse the plugging infrastructure for this,
convert it to the (new) real API for delaying queueing a bit. A
default delay of 3 msec is defined, to match the previous
behaviour.

Signed-off-by: Jens Axboe <jaxboe@fusionio.com>
drivers/scsi/scsi_lib.c

index 9045c52..5a0ae7a 100644 (file)
@@ -67,6 +67,13 @@ static struct scsi_host_sg_pool scsi_sg_pools[] = {
 
 struct kmem_cache *scsi_sdb_cache;
 
+/*
+ * When to reinvoke queueing after a resource shortage. It's 3 msecs to
+ * not change behaviour from the previous unplug mechanism, experimentation
+ * may prove this needs changing.
+ */
+#define SCSI_QUEUE_DELAY       3
+
 static void scsi_run_queue(struct request_queue *q);
 
 /*
@@ -149,14 +156,7 @@ static int __scsi_queue_insert(struct scsi_cmnd *cmd, int reason, int unbusy)
        /*
         * Requeue this command.  It will go before all other commands
         * that are already in the queue.
-        *
-        * NOTE: there is magic here about the way the queue is plugged if
-        * we have no outstanding commands.
-        * 
-        * Although we *don't* plug the queue, we call the request
-        * function.  The SCSI request function detects the blocked condition
-        * and plugs the queue appropriately.
-         */
+        */
        spin_lock_irqsave(q->queue_lock, flags);
        blk_requeue_request(q, cmd->request);
        spin_unlock_irqrestore(q->queue_lock, flags);
@@ -1194,11 +1194,11 @@ int scsi_prep_return(struct request_queue *q, struct request *req, int ret)
        case BLKPREP_DEFER:
                /*
                 * If we defer, the blk_peek_request() returns NULL, but the
-                * queue must be restarted, so we plug here if no returning
-                * command will automatically do that.
+                * queue must be restarted, so we schedule a callback to happen
+                * shortly.
                 */
                if (sdev->device_busy == 0)
-                       blk_plug_device(q);
+                       blk_delay_queue(q, SCSI_QUEUE_DELAY);
                break;
        default:
                req->cmd_flags |= REQ_DONTPREP;
@@ -1237,7 +1237,7 @@ static inline int scsi_dev_queue_ready(struct request_queue *q,
                                   sdev_printk(KERN_INFO, sdev,
                                   "unblocking device at zero depth\n"));
                } else {
-                       blk_plug_device(q);
+                       blk_delay_queue(q, SCSI_QUEUE_DELAY);
                        return 0;
                }
        }
@@ -1467,7 +1467,7 @@ static void scsi_request_fn(struct request_queue *q)
         * the host is no longer able to accept any more requests.
         */
        shost = sdev->host;
-       while (!blk_queue_plugged(q)) {
+       for (;;) {
                int rtn;
                /*
                 * get next queueable request.  We do this early to make sure
@@ -1546,15 +1546,8 @@ static void scsi_request_fn(struct request_queue *q)
                 */
                rtn = scsi_dispatch_cmd(cmd);
                spin_lock_irq(q->queue_lock);
-               if(rtn) {
-                       /* we're refusing the command; because of
-                        * the way locks get dropped, we need to 
-                        * check here if plugging is required */
-                       if(sdev->device_busy == 0)
-                               blk_plug_device(q);
-
-                       break;
-               }
+               if (rtn)
+                       goto out_delay;
        }
 
        goto out;
@@ -1573,9 +1566,10 @@ static void scsi_request_fn(struct request_queue *q)
        spin_lock_irq(q->queue_lock);
        blk_requeue_request(q, req);
        sdev->device_busy--;
-       if(sdev->device_busy == 0)
-               blk_plug_device(q);
- out:
+out_delay:
+       if (sdev->device_busy == 0)
+               blk_delay_queue(q, SCSI_QUEUE_DELAY);
+out:
        /* must be careful here...if we trigger the ->remove() function
         * we cannot be holding the q lock */
        spin_unlock_irq(q->queue_lock);