ide: use per-device request queue locks (v2)

* Move hack for flush requests from choose_drive() to do_ide_request().

* Add ide_plug_device() helper and convert core IDE code from using
  per-hwgroup lock as a request lock to use the ->queue_lock instead.

* Remove no longer needed:
  - choose_drive() function
  - WAKEUP() macro
  - 'sleeping' flag from ide_hwif_t
  - 'service_{start,time}' fields from ide_drive_t

This patch results in much simpler and more maintainable code
(besides being a scalability improvement).

v2:
* Fixes/improvements based on review from Elias:
  - take as many requests off the queue as possible
  - remove now redundant BUG_ON()

Cc: Elias Oltmanns <eo@nebensachen.de>
Signed-off-by: Bartlomiej Zolnierkiewicz <bzolnier@gmail.com>

4 files changed, 77 insertions, 157 deletions

diff --git a/drivers/ide/ide-io.c b/drivers/ide/ide-io.c
index ab480042757a..bb3248abf47d 100644
--- a/drivers/ide/ide-io.c
+++ b/drivers/ide/ide-io.c
@@ -667,85 +667,10 @@ void ide_stall_queue (ide_drive_t *drive, unsigned long timeout)
         drive->sleep = timeout + jiffies;
         drive->dev_flags |= IDE_DFLAG_SLEEPING;
 }
-
 EXPORT_SYMBOL(ide_stall_queue);
 
-#define WAKEUP(drive)   ((drive)->service_start + 2 * (drive)->service_time)
-
-/**
- *      choose_drive            -       select a drive to service
- *      @hwgroup: hardware group to select on
- *
- *      choose_drive() selects the next drive which will be serviced.
- *      This is necessary because the IDE layer can't issue commands
- *      to both drives on the same cable, unlike SCSI.
- */
- 
-static inline ide_drive_t *choose_drive (ide_hwgroup_t *hwgroup)
-{
-        ide_drive_t *drive, *best;
-
-repeat: 
-        best = NULL;
-        drive = hwgroup->drive;
-
-        /*
-         * drive is doing pre-flush, ordered write, post-flush sequence. even
-         * though that is 3 requests, it must be seen as a single transaction.
-         * we must not preempt this drive until that is complete
-         */
-        if (blk_queue_flushing(drive->queue)) {
-                /*
-                 * small race where queue could get replugged during
-                 * the 3-request flush cycle, just yank the plug since
-                 * we want it to finish asap
-                 */
-                blk_remove_plug(drive->queue);
-                return drive;
-        }
-
-        do {
-                u8 dev_s = !!(drive->dev_flags & IDE_DFLAG_SLEEPING);
-                u8 best_s = (best && !!(best->dev_flags & IDE_DFLAG_SLEEPING));
-
-                if ((dev_s == 0 || time_after_eq(jiffies, drive->sleep)) &&
-                    !elv_queue_empty(drive->queue)) {
-                        if (best == NULL ||
-                            (dev_s && (best_s == 0 || time_before(drive->sleep, best->sleep))) ||
-                            (best_s == 0 && time_before(WAKEUP(drive), WAKEUP(best)))) {
-                                if (!blk_queue_plugged(drive->queue))
-                                        best = drive;
-                        }
-                }
-        } while ((drive = drive->next) != hwgroup->drive);
-
-        if (best && (best->dev_flags & IDE_DFLAG_NICE1) &&
-            (best->dev_flags & IDE_DFLAG_SLEEPING) == 0 &&
-            best != hwgroup->drive && best->service_time > WAIT_MIN_SLEEP) {
-                long t = (signed long)(WAKEUP(best) - jiffies);
-                if (t >= WAIT_MIN_SLEEP) {
-                /*
-                 * We *may* have some time to spare, but first let's see if
-                 * someone can potentially benefit from our nice mood today..
-                 */
-                        drive = best->next;
-                        do {
-                                if ((drive->dev_flags & IDE_DFLAG_SLEEPING) == 0
-                                 && time_before(jiffies - best->service_time, WAKEUP(drive))
-                                 && time_before(WAKEUP(drive), jiffies + t))
-                                {
-                                        ide_stall_queue(best, min_t(long, t, 10 * WAIT_MIN_SLEEP));
-                                        goto repeat;
-                                }
-                        } while ((drive = drive->next) != best);
-                }
-        }
-        return best;
-}
-
 /*
  * Issue a new request to a drive from hwgroup
- * Caller must have already done spin_lock_irqsave(&hwgroup->lock, ..);
  *
  * A hwgroup is a serialized group of IDE interfaces.  Usually there is
  * exactly one hwif (interface) per hwgroup, but buggy controllers (eg. CMD640)
@@ -757,8 +682,7 @@ repeat:
  * possibly along with many other devices.  This is especially common in
  * PCI-based systems with off-board IDE controller cards.
  *
- * The IDE driver uses a per-hwgroup spinlock to protect
- * access to the request queues, and to protect the hwgroup->busy flag.
+ * The IDE driver uses a per-hwgroup lock to protect the hwgroup->busy flag.
  *
  * The first thread into the driver for a particular hwgroup sets the
  * hwgroup->busy flag to indicate that this hwgroup is now active,
@@ -780,61 +704,38 @@ repeat:
  */
 void do_ide_request(struct request_queue *q)
 {
-        ide_drive_t     *orig_drive = q->queuedata;
-        ide_hwgroup_t   *hwgroup = orig_drive->hwif->hwgroup;
-        ide_drive_t     *drive;
-        ide_hwif_t      *hwif;
+        ide_drive_t     *drive = q->queuedata;
+        ide_hwif_t      *hwif = drive->hwif;
+        ide_hwgroup_t   *hwgroup = hwif->hwgroup;
         struct request  *rq;
         ide_startstop_t startstop;
 
-        /* caller must own hwgroup->lock */
-        BUG_ON(!irqs_disabled());
-
-        while (!ide_lock_hwgroup(hwgroup)) {
-                drive = choose_drive(hwgroup);
-                if (drive == NULL) {
-                        int sleeping = 0;
-                        unsigned long sleep = 0; /* shut up, gcc */
-                        hwgroup->rq = NULL;
-                        drive = hwgroup->drive;
-                        do {
-                                if ((drive->dev_flags & IDE_DFLAG_SLEEPING) &&
-                                    (sleeping == 0 ||
-                                     time_before(drive->sleep, sleep))) {
-                                        sleeping = 1;
-                                        sleep = drive->sleep;
-                                }
-                        } while ((drive = drive->next) != hwgroup->drive);
-                        if (sleeping) {
+        /*
+         * drive is doing pre-flush, ordered write, post-flush sequence. even
+         * though that is 3 requests, it must be seen as a single transaction.
+         * we must not preempt this drive until that is complete
+         */
+        if (blk_queue_flushing(q))
                 /*
-                 * Take a short snooze, and then wake up this hwgroup again.
-                 * This gives other hwgroups on the same a chance to
-                 * play fairly with us, just in case there are big differences
-                 * in relative throughputs.. don't want to hog the cpu too much.
+                 * small race where queue could get replugged during
+                 * the 3-request flush cycle, just yank the plug since
+                 * we want it to finish asap
                  */
-                                if (time_before(sleep, jiffies + WAIT_MIN_SLEEP))
-                                        sleep = jiffies + WAIT_MIN_SLEEP;
-#if 1
-                                if (timer_pending(&hwgroup->timer))
-                                        printk(KERN_CRIT "ide_set_handler: timer already active\n");
-#endif
-                                /* so that ide_timer_expiry knows what to do */
-                                hwgroup->sleeping = 1;
-                                hwgroup->req_gen_timer = hwgroup->req_gen;
-                                mod_timer(&hwgroup->timer, sleep);
-                                /* we purposely leave hwgroup locked
-                                 * while sleeping */
-                        } else
-                                ide_unlock_hwgroup(hwgroup);
+                blk_remove_plug(q);
 
-                        /* no more work for this hwgroup (for now) */
-                        goto plug_device;
-                }
+        spin_unlock_irq(q->queue_lock);
+        spin_lock_irq(&hwgroup->lock);
 
-                if (drive != orig_drive)
-                        goto plug_device;
+        if (!ide_lock_hwgroup(hwgroup)) {
+repeat:
+                hwgroup->rq = NULL;
 
-                hwif = drive->hwif;
+                if (drive->dev_flags & IDE_DFLAG_SLEEPING) {
+                        if (time_before(drive->sleep, jiffies)) {
+                                ide_unlock_hwgroup(hwgroup);
+                                goto plug_device;
+                        }
+                }
 
                 if (hwif != hwgroup->hwif) {
                         /*
@@ -847,16 +748,20 @@ void do_ide_request(struct request_queue *q)
                 hwgroup->hwif = hwif;
                 hwgroup->drive = drive;
                 drive->dev_flags &= ~(IDE_DFLAG_SLEEPING | IDE_DFLAG_PARKED);
-                drive->service_start = jiffies;
 
+                spin_unlock_irq(&hwgroup->lock);
+                spin_lock_irq(q->queue_lock);
                 /*
                  * we know that the queue isn't empty, but this can happen