cfq-iosched: don't delay queue kick for a merged request
Jens Axboe [Tue, 14 Apr 2009 12:18:16 +0000 (14:18 +0200)]
"Zhang, Yanmin" <yanmin_zhang@linux.intel.com> reports that commit
b029195dda0129b427c6e579a3bb3ae752da3a93 introduced a regression
of about 50% with sequential threaded read workloads. The test
case is:

tiotest -k0 -k1 -k3 -f 80 -t 32

which starts 32 threads each reading a 80MB file. Twiddle the kick
queue logic so that we do start IO immediately, if it appears to be
a fully merged request. We can't really detect that, so just check
if the request is bigger than a page or not. The assumption is that
since single bio issues will first queue a single request with just
one page attached and then later do merges on that, if we already
have more than a page worth of data in the request, then the request
is most likely good to go.

Verified that this doesn't cause a regression with the test case that
commit b029195dda0129b427c6e579a3bb3ae752da3a93 was fixing. It does not,
we still see maximum sized requests for the queue-then-merge cases.

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

block/cfq-iosched.c

index cba8a5d..56e9d85 100644 (file)
@@ -1903,10 +1903,17 @@ cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
                 * Remember that we saw a request from this process, but
                 * don't start queuing just yet. Otherwise we risk seeing lots
                 * of tiny requests, because we disrupt the normal plugging
-                * and merging.
+                * and merging. If the request is already larger than a single
+                * page, let it rip immediately. For that case we assume that
+                * merging is already done.
                 */
-               if (cfq_cfqq_wait_request(cfqq))
+               if (cfq_cfqq_wait_request(cfqq)) {
+                       if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE) {
+                               del_timer(&cfqd->idle_slice_timer);
+                               blk_start_queueing(cfqd->queue);
+                       }
                        cfq_mark_cfqq_must_dispatch(cfqq);
+               }
        } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
                /*
                 * not the active queue - expire current slice if it is