cfq-iosched: rethink seeky detection for SSDs
Corrado Zoccolo [Sat, 27 Feb 2010 18:45:40 +0000 (19:45 +0100)]
CFQ currently applies the same logic of detecting seeky queues and
grouping them together for rotational disks as well as SSDs.
For SSDs, the time to complete a request doesn't depend on the
request location, but only on the size.
This patch therefore changes the criterion to group queues by
request size in case of SSDs, in order to achieve better fairness.

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

block/cfq-iosched.c

index 3fd8afc..423aee3 100644 (file)
@@ -47,6 +47,7 @@ static const int cfq_hist_divisor = 4;
 #define CFQ_SERVICE_SHIFT       12
 
 #define CFQQ_SEEK_THR          (sector_t)(8 * 100)
+#define CFQQ_SECT_THR_NONROT   (sector_t)(2 * 32)
 #define CFQQ_SEEKY(cfqq)       (hweight32(cfqq->seek_history) > 32/8)
 
 #define RQ_CIC(rq)             \
@@ -2965,6 +2966,7 @@ cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
                       struct request *rq)
 {
        sector_t sdist = 0;
+       sector_t n_sec = blk_rq_sectors(rq);
        if (cfqq->last_request_pos) {
                if (cfqq->last_request_pos < blk_rq_pos(rq))
                        sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
@@ -2973,7 +2975,10 @@ cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
        }
 
        cfqq->seek_history <<= 1;
-       cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
+       if (blk_queue_nonrot(cfqd->queue))
+               cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
+       else
+               cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
 }
 
 /*