block cfq: make queue preempt work for queues from different workload
authorShaohua Li <shaohua.li@intel.com>
Fri, 14 Jan 2011 07:41:02 +0000 (08:41 +0100)
committerJens Axboe <jaxboe@fusionio.com>
Fri, 14 Jan 2011 07:41:02 +0000 (08:41 +0100)
I got this:
             fio-874   [007]  2157.724514:   8,32   m   N cfq874 preempt
             fio-874   [007]  2157.724519:   8,32   m   N cfq830 slice expired t=1
             fio-874   [007]  2157.724520:   8,32   m   N cfq830 sl_used=1 disp=0 charge=1 iops=0 sect=0
             fio-874   [007]  2157.724521:   8,32   m   N cfq830 set_active wl_prio:0 wl_type:0
             fio-874   [007]  2157.724522:   8,32   m   N cfq830 Not idling. st->count:1

cfq830 is an async queue, and preempted by a sync queue cfq874. But since we
have cfqg->saved_workload_slice mechanism, the preempt is a nop.
Looks currently our preempt is totally broken if the two queues are not from
the same workload type.
Below patch fixes it. This will might make async queue starvation, but it's
what our old code does before cgroup is added.

Signed-off-by: Shaohua Li <shaohua.li@intel.com>
Signed-off-by: Jens Axboe <jaxboe@fusionio.com>

block/cfq-iosched.c

index 8427697..7bfea53 100644 (file)
@@ -3284,10 +3284,19 @@ cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
  */
 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 {
+       struct cfq_queue *old_cfqq = cfqd->active_queue;
+
        cfq_log_cfqq(cfqd, cfqq, "preempt");
        cfq_slice_expired(cfqd, 1);
 
        /*
+        * workload type is changed, don't save slice, otherwise preempt
+        * doesn't happen
+        */
+       if (cfqq_type(old_cfqq) != cfqq_type(cfqq))
+               cfqq->cfqg->saved_workload_slice = 0;
+
+       /*
         * Put the new queue at the front of the of the current list,
         * so we know that it will be selected next.
         */