Revert "cfq: Make use of service count to estimate the rb_key offset"
Jens Axboe [Mon, 30 Nov 2009 08:38:13 +0000 (09:38 +0100)]
This reverts commit 3586e917f2c7df769d173c4ec99554cb40a911e5.

Corrado Zoccolo <czoccolo@gmail.com> correctly points out, that we need
consistency of rb_key offset across groups. This means we cannot properly
use the per-service_tree service count. Revert this change.

Signed-off-by: Jens Axboe <jens.axboe@oracle.com>

block/cfq-iosched.c

index a5de31f..7144649 100644 (file)
@@ -609,15 +609,11 @@ cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
                                      struct cfq_queue *cfqq)
 {
-       struct cfq_rb_root *service_tree;
-
-       service_tree = service_tree_for(cfqq_prio(cfqq), cfqq_type(cfqq), cfqd);
-
        /*
         * just an approximation, should be ok.
         */
-       return  service_tree->count * (cfq_prio_slice(cfqd, 1, 0) -
-                  cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
+       return (cfqd->busy_queues - 1) * (cfq_prio_slice(cfqd, 1, 0) -
+                      cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
 }
 
 /*