]> git.proxmox.com Git - mirror_ubuntu-disco-kernel.git/commitdiff
cfq-iosched: fix queue depth detection
authorAaron Carroll <aaronc@gelato.unsw.edu.au>
Tue, 26 Aug 2008 13:52:36 +0000 (15:52 +0200)
committerJens Axboe <jens.axboe@oracle.com>
Thu, 9 Oct 2008 06:56:09 +0000 (08:56 +0200)
CFQ's detection of queueing devices assumes a non-queuing device and detects
if the queue depth reaches a certain threshold.  Under some workloads (e.g.
synchronous reads), CFQ effectively forces a unit queue depth, thus defeating
the detection logic.  This leads to poor performance on queuing hardware,
since the idle window remains enabled.

This patch inverts the sense of the logic: assume a queuing-capable device,
and detect if the depth does not exceed the threshold.

Signed-off-by: Aaron Carroll <aaronc@gelato.unsw.edu.au>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
block/cfq-iosched.c

index 5f6fd287c185369e3e7c43981326f73b1bdc9b81..494b6fdcb18340d91395acf6883925a1f7c476e6 100644 (file)
@@ -39,6 +39,7 @@ static int cfq_slice_idle = HZ / 125;
 #define CFQ_MIN_TT             (2)
 
 #define CFQ_SLICE_SCALE                (5)
+#define CFQ_HW_QUEUE_MIN       (5)
 
 #define RQ_CIC(rq)             \
        ((struct cfq_io_context *) (rq)->elevator_private)
@@ -86,7 +87,14 @@ struct cfq_data {
 
        int rq_in_driver;
        int sync_flight;
+
+       /*
+        * queue-depth detection
+        */
+       int rq_queued;
        int hw_tag;
+       int hw_tag_samples;
+       int rq_in_driver_peak;
 
        /*
         * idle window management
@@ -654,15 +662,6 @@ static void cfq_activate_request(struct request_queue *q, struct request *rq)
        cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
                                                cfqd->rq_in_driver);
 
-       /*
-        * If the depth is larger 1, it really could be queueing. But lets
-        * make the mark a little higher - idling could still be good for
-        * low queueing, and a low queueing number could also just indicate
-        * a SCSI mid layer like behaviour where limit+1 is often seen.
-        */
-       if (!cfqd->hw_tag && cfqd->rq_in_driver > 4)
-               cfqd->hw_tag = 1;
-
        cfqd->last_position = rq->hard_sector + rq->hard_nr_sectors;
 }
 
@@ -686,6 +685,7 @@ static void cfq_remove_request(struct request *rq)
        list_del_init(&rq->queuelist);
        cfq_del_rq_rb(rq);
 
+       cfqq->cfqd->rq_queued--;
        if (rq_is_meta(rq)) {
                WARN_ON(!cfqq->meta_pending);
                cfqq->meta_pending--;
@@ -1833,6 +1833,7 @@ cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 {
        struct cfq_io_context *cic = RQ_CIC(rq);
 
+       cfqd->rq_queued++;
        if (rq_is_meta(rq))
                cfqq->meta_pending++;
 
@@ -1880,6 +1881,31 @@ static void cfq_insert_request(struct request_queue *q, struct request *rq)
        cfq_rq_enqueued(cfqd, cfqq, rq);
 }
 
+/*
+ * Update hw_tag based on peak queue depth over 50 samples under
+ * sufficient load.
+ */
+static void cfq_update_hw_tag(struct cfq_data *cfqd)
+{
+       if (cfqd->rq_in_driver > cfqd->rq_in_driver_peak)
+               cfqd->rq_in_driver_peak = cfqd->rq_in_driver;
+
+       if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
+           cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
+               return;
+
+       if (cfqd->hw_tag_samples++ < 50)
+               return;
+
+       if (cfqd->rq_in_driver_peak >= CFQ_HW_QUEUE_MIN)
+               cfqd->hw_tag = 1;
+       else
+               cfqd->hw_tag = 0;
+
+       cfqd->hw_tag_samples = 0;
+       cfqd->rq_in_driver_peak = 0;
+}
+
 static void cfq_completed_request(struct request_queue *q, struct request *rq)
 {
        struct cfq_queue *cfqq = RQ_CFQQ(rq);
@@ -1890,6 +1916,8 @@ static void cfq_completed_request(struct request_queue *q, struct request *rq)
        now = jiffies;
        cfq_log_cfqq(cfqd, cfqq, "complete");
 
+       cfq_update_hw_tag(cfqd);
+
        WARN_ON(!cfqd->rq_in_driver);
        WARN_ON(!cfqq->dispatched);
        cfqd->rq_in_driver--;
@@ -2200,6 +2228,7 @@ static void *cfq_init_queue(struct request_queue *q)
        cfqd->cfq_slice[1] = cfq_slice_sync;
        cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
        cfqd->cfq_slice_idle = cfq_slice_idle;
+       cfqd->hw_tag = 1;
 
        return cfqd;
 }