cfq-iosched: reduce write depth only if sync was delayed
Corrado Zoccolo [Sun, 6 Dec 2009 10:48:52 +0000 (11:48 +0100)]
The introduction of ramp-up formula for async queue depths has
slowed down dirty page reclaim, by reducing async write performance.
This patch makes sure the formula kicks in only when sync request
was recently delayed.

Signed-off-by: Corrado Zoccolo <czoccolo@gmail.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>

block/cfq-iosched.c

index cfb0b2f..5009af4 100644 (file)
@@ -283,7 +283,7 @@ struct cfq_data {
         */
        struct cfq_queue oom_cfqq;
 
-       unsigned long last_end_sync_rq;
+       unsigned long last_delayed_sync;
 
        /* List of cfq groups being managed on this device*/
        struct hlist_head cfqg_list;
@@ -2264,7 +2264,7 @@ static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
         * based on the last sync IO we serviced
         */
        if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
-               unsigned long last_sync = jiffies - cfqd->last_end_sync_rq;
+               unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
                unsigned int depth;
 
                depth = last_sync / cfqd->cfq_slice[1];
@@ -3273,7 +3273,8 @@ static void cfq_completed_request(struct request_queue *q, struct request *rq)
 
        if (sync) {
                RQ_CIC(rq)->last_end_request = now;
-               cfqd->last_end_sync_rq = now;
+               if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
+                       cfqd->last_delayed_sync = now;
        }
 
        /*
@@ -3711,7 +3712,7 @@ static void *cfq_init_queue(struct request_queue *q)
        cfqd->cfq_latency = 1;
        cfqd->cfq_group_isolation = 0;
        cfqd->hw_tag = -1;
-       cfqd->last_end_sync_rq = jiffies;
+       cfqd->last_delayed_sync = jiffies - HZ;
        INIT_RCU_HEAD(&cfqd->rcu);
        return cfqd;
 }