Revert "ide: improve handling of Power Management requests"
authorDavid S. Miller <davem@davemloft.net>
Mon, 6 Jul 2009 19:39:27 +0000 (12:39 -0700)
committerDavid S. Miller <davem@davemloft.net>
Mon, 6 Jul 2009 19:39:27 +0000 (12:39 -0700)
This reverts commit a1317f714af7aed60ddc182d0122477cbe36ee9b.

drivers/ide/ide-io.c

index d5f3c77..db96138 100644 (file)
@@ -466,14 +466,10 @@ void do_ide_request(struct request_queue *q)
 
        if (!ide_lock_port(hwif)) {
                ide_hwif_t *prev_port;
+
+               WARN_ON_ONCE(hwif->rq);
 repeat:
                prev_port = hwif->host->cur_port;
-
-               if (drive->dev_flags & IDE_DFLAG_BLOCKED)
-                       rq = hwif->rq;
-               else
-                       WARN_ON_ONCE(hwif->rq);
-
                if (drive->dev_flags & IDE_DFLAG_SLEEPING &&
                    time_after(drive->sleep, jiffies)) {
                        ide_unlock_port(hwif);
@@ -500,29 +496,43 @@ repeat:
                hwif->cur_dev = drive;
                drive->dev_flags &= ~(IDE_DFLAG_SLEEPING | IDE_DFLAG_PARKED);
 
-               if (rq == NULL) {
-                       spin_unlock_irq(&hwif->lock);
-                       spin_lock_irq(q->queue_lock);
-                       /*
-                        * we know that the queue isn't empty, but this can
-                        * happen if ->prep_rq_fn() decides to kill a request
-                        */
+               spin_unlock_irq(&hwif->lock);
+               spin_lock_irq(q->queue_lock);
+               /*
+                * we know that the queue isn't empty, but this can happen
+                * if the q->prep_rq_fn() decides to kill a request
+                */
+               if (!rq)
                        rq = blk_fetch_request(drive->queue);
-                       spin_unlock_irq(q->queue_lock);
-                       spin_lock_irq(&hwif->lock);
 
-                       if (rq == NULL) {
-                               ide_unlock_port(hwif);
-                               goto out;
-                       }
+               spin_unlock_irq(q->queue_lock);
+               spin_lock_irq(&hwif->lock);
+
+               if (!rq) {
+                       ide_unlock_port(hwif);
+                       goto out;
                }
 
                /*
                 * Sanity: don't accept a request that isn't a PM request
-                * if we are currently power managed.
+                * if we are currently power managed. This is very important as
+                * blk_stop_queue() doesn't prevent the blk_fetch_request()
+                * above to return us whatever is in the queue. Since we call
+                * ide_do_request() ourselves, we end up taking requests while
+                * the queue is blocked...
+                * 
+                * We let requests forced at head of queue with ide-preempt
+                * though. I hope that doesn't happen too much, hopefully not
+                * unless the subdriver triggers such a thing in its own PM
+                * state machine.
                 */
-               BUG_ON((drive->dev_flags & IDE_DFLAG_BLOCKED) &&
-                      blk_pm_request(rq) == 0);
+               if ((drive->dev_flags & IDE_DFLAG_BLOCKED) &&
+                   blk_pm_request(rq) == 0 &&
+                   (rq->cmd_flags & REQ_PREEMPT) == 0) {
+                       /* there should be no pending command at this point */
+                       ide_unlock_port(hwif);
+                       goto plug_device;
+               }
 
                hwif->rq = rq;