blkio: Wait on sync-noidle queue even if rq_noidle = 1
authorVivek Goyal <vgoyal@redhat.com>
Thu, 3 Dec 2009 17:59:56 +0000 (12:59 -0500)
committerJens Axboe <jens.axboe@oracle.com>
Thu, 3 Dec 2009 18:28:53 +0000 (19:28 +0100)
o rq_noidle() is supposed to tell cfq that do not expect a request after this
  one, hence don't idle. But this does not seem to work very well. For example
  for direct random readers, rq_noidle = 1 but there is next request coming
  after this. Not idling, leads to a group not getting its share even if
  group_isolation=1.

o The right solution for this issue is to scan the higher layers and set
  right flag (WRITE_SYNC or WRITE_ODIRECT). For the time being, this single
  line fix helps. This should not have any significant impact when we are
  not using cgroups. I will later figure out IO paths in higher layer and
  fix it.

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

index 063dcbb714e79d548a3042f167a01e351ba3896b..08b057b1b3b2501ae52ec5d2fe7d6d76161a1e34 100644 (file)
@@ -3314,7 +3314,8 @@ static void cfq_completed_request(struct request_queue *q, struct request *rq)
                         * only if we processed at least one !rq_noidle request
                         */
                        if (cfqd->serving_type == SYNC_WORKLOAD
-                           || cfqd->noidle_tree_requires_idle)
+                           || cfqd->noidle_tree_requires_idle
+                           || cfqq->cfqg->nr_cfqq == 1)
                                cfq_arm_slice_timer(cfqd);
                }
        }