cfq-iosched: account for slice over/under time
authorJens Axboe <jens.axboe@oracle.com>
Fri, 19 Jan 2007 00:56:49 +0000 (11:56 +1100)
committerJens Axboe <axboe@carl.home.kernel.dk>
Sun, 11 Feb 2007 22:14:45 +0000 (23:14 +0100)
If a slice uses less than it is entitled to (or perhaps more), include
that in the decision on how much time to give it the next time it
gets serviced.

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

index d44402a4c5cd99b5eec4d3fc8f9ca8c6531a2615..039b38cf805c51cc012cc732cde8e48ed67ea1ee 100644 (file)
@@ -147,8 +147,8 @@ struct cfq_queue {
        struct list_head fifo;
 
        unsigned long slice_end;
-       unsigned long slice_left;
        unsigned long service_last;
+       long slice_resid;
 
        /* number of requests that are on the dispatch list */
        int on_dispatch[2];
@@ -251,6 +251,14 @@ static inline void
 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 {
        cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
+       cfqq->slice_end += cfqq->slice_resid;
+
+       /*
+        * Don't carry over residual for more than one slice, we only want
+        * to slightly correct the fairness. Carrying over forever would
+        * easily introduce oscillations.
+        */
+       cfqq->slice_resid = 0;
 }
 
 /*
@@ -667,7 +675,6 @@ __cfq_set_active_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
                del_timer(&cfqd->idle_class_timer);
 
                cfqq->slice_end = 0;
-               cfqq->slice_left = 0;
                cfq_clear_cfqq_must_alloc_slice(cfqq);
                cfq_clear_cfqq_fifo_expire(cfqq);
                cfq_mark_cfqq_slice_new(cfqq);
@@ -683,8 +690,6 @@ static void
 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
                    int preempted)
 {
-       unsigned long now = jiffies;
-
        if (cfq_cfqq_wait_request(cfqq))
                del_timer(&cfqd->idle_slice_timer);
 
@@ -699,10 +704,8 @@ __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
         * store what was left of this slice, if the queue idled out
         * or was preempted
         */
-       if (cfq_slice_used(cfqq))
-               cfqq->slice_left = cfqq->slice_end - now;
-       else
-               cfqq->slice_left = 0;
+       if (!cfq_cfqq_slice_new(cfqq))
+               cfqq->slice_resid = cfqq->slice_end - jiffies;
 
        cfq_resort_rr_list(cfqq, preempted);
 
@@ -1364,10 +1367,7 @@ retry:
                hlist_add_head(&cfqq->cfq_hash, &cfqd->cfq_hash[hashval]);
                atomic_set(&cfqq->ref, 0);
                cfqq->cfqd = cfqd;
-               /*
-                * set ->slice_left to allow preemption for a new process
-                */
-               cfqq->slice_left = 2 * cfqd->cfq_slice_idle;
+
                cfq_mark_cfqq_idle_window(cfqq);
                cfq_mark_cfqq_prio_changed(cfqq);
                cfq_mark_cfqq_queue_new(cfqq);
@@ -1585,11 +1585,6 @@ cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
                return 1;
        if (!cfq_cfqq_wait_request(new_cfqq))
                return 0;
-       /*
-        * if it doesn't have slice left, forget it
-        */
-       if (new_cfqq->slice_left < cfqd->cfq_slice_idle)
-               return 0;
        /*
         * if the new request is sync, but the currently running queue is
         * not, let the sync request have priority.
@@ -1614,9 +1609,6 @@ static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 {
        cfq_slice_expired(cfqd, 1);
 
-       if (!cfqq->slice_left)
-               cfqq->slice_left = cfq_prio_to_slice(cfqd, cfqq) / 2;
-
        /*
         * Put the new queue at the front of the of the current list,
         * so we know that it will be selected next.