io_context sharing - anticipatory changes
authorJens Axboe <jens.axboe@oracle.com>
Mon, 21 Jan 2008 19:03:12 +0000 (20:03 +0100)
committerJens Axboe <jens.axboe@oracle.com>
Mon, 28 Jan 2008 09:50:35 +0000 (10:50 +0100)
changes to anticipatory io scheduler for io_context sharing

Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
block/as-iosched.c

index cb5e53b05c7c6baf6224f36a2c3dee8528c4e116..b201d16a71020f1261a42bc56ac36a9915da16ad 100644 (file)
@@ -170,9 +170,11 @@ static void free_as_io_context(struct as_io_context *aic)
 
 static void as_trim(struct io_context *ioc)
 {
+       spin_lock(&ioc->lock);
        if (ioc->aic)
                free_as_io_context(ioc->aic);
        ioc->aic = NULL;
+       spin_unlock(&ioc->lock);
 }
 
 /* Called when the task exits */
@@ -462,7 +464,9 @@ static void as_antic_timeout(unsigned long data)
        spin_lock_irqsave(q->queue_lock, flags);
        if (ad->antic_status == ANTIC_WAIT_REQ
                        || ad->antic_status == ANTIC_WAIT_NEXT) {
-               struct as_io_context *aic = ad->io_context->aic;
+               struct as_io_context *aic;
+               spin_lock(&ad->io_context->lock);
+               aic = ad->io_context->aic;
 
                ad->antic_status = ANTIC_FINISHED;
                kblockd_schedule_work(&ad->antic_work);
@@ -475,6 +479,7 @@ static void as_antic_timeout(unsigned long data)
                        /* process not "saved" by a cooperating request */
                        ad->exit_no_coop = (7*ad->exit_no_coop + 256)/8;
                }
+               spin_unlock(&ad->io_context->lock);
        }
        spin_unlock_irqrestore(q->queue_lock, flags);
 }
@@ -635,9 +640,11 @@ static int as_can_break_anticipation(struct as_data *ad, struct request *rq)
 
        ioc = ad->io_context;
        BUG_ON(!ioc);
+       spin_lock(&ioc->lock);
 
        if (rq && ioc == RQ_IOC(rq)) {
                /* request from same process */
+               spin_unlock(&ioc->lock);
                return 1;
        }
 
@@ -646,20 +653,25 @@ static int as_can_break_anticipation(struct as_data *ad, struct request *rq)
                 * In this situation status should really be FINISHED,
                 * however the timer hasn't had the chance to run yet.
                 */
+               spin_unlock(&ioc->lock);
                return 1;
        }
 
        aic = ioc->aic;
-       if (!aic)
+       if (!aic) {
+               spin_unlock(&ioc->lock);
                return 0;
+       }
 
        if (atomic_read(&aic->nr_queued) > 0) {
                /* process has more requests queued */
+               spin_unlock(&ioc->lock);
                return 1;
        }
 
        if (atomic_read(&aic->nr_dispatched) > 0) {
                /* process has more requests dispatched */
+               spin_unlock(&ioc->lock);
                return 1;
        }
 
@@ -680,6 +692,7 @@ static int as_can_break_anticipation(struct as_data *ad, struct request *rq)
                }
 
                as_update_iohist(ad, aic, rq);
+               spin_unlock(&ioc->lock);
                return 1;
        }
 
@@ -688,20 +701,27 @@ static int as_can_break_anticipation(struct as_data *ad, struct request *rq)
                if (aic->ttime_samples == 0)
                        ad->exit_prob = (7*ad->exit_prob + 256)/8;
 
-               if (ad->exit_no_coop > 128)
+               if (ad->exit_no_coop > 128) {
+                       spin_unlock(&ioc->lock);
                        return 1;
+               }
        }
 
        if (aic->ttime_samples == 0) {
-               if (ad->new_ttime_mean > ad->antic_expire)
+               if (ad->new_ttime_mean > ad->antic_expire) {
+                       spin_unlock(&ioc->lock);
                        return 1;
-               if (ad->exit_prob * ad->exit_no_coop > 128*256)
+               }
+               if (ad->exit_prob * ad->exit_no_coop > 128*256) {
+                       spin_unlock(&ioc->lock);
                        return 1;
+               }
        } else if (aic->ttime_mean > ad->antic_expire) {
                /* the process thinks too much between requests */
+               spin_unlock(&ioc->lock);
                return 1;
        }
-
+       spin_unlock(&ioc->lock);
        return 0;
 }
 
@@ -1255,7 +1275,9 @@ static void as_merged_requests(struct request_queue *q, struct request *req,
                         * Don't copy here but swap, because when anext is
                         * removed below, it must contain the unused context
                         */
+                       double_spin_lock(&rioc->lock, &nioc->lock, rioc < nioc);
                        swap_io_context(&rioc, &nioc);
+                       double_spin_unlock(&rioc->lock, &nioc->lock, rioc < nioc);
                }
        }