Re: 2.6.20-rc2+: CFQ halving disk throughput.

[Jens Axboe <jens.axboe@oracle.com>]

On Tue, Jan 02 2007, Rene Herman wrote:
> Jens Axboe wrote:
> 
> >On Mon, Jan 01 2007, Andrew Morton wrote:
> 
> >>The patch would appear to need this fix:
> >>
> >>--- a/block/cfq-iosched.c~a
> >>+++ a/block/cfq-iosched.c
> >>@@ -592,7 +592,7 @@ static int cfq_allow_merge(request_queue
> >> 	if (cfqq == RQ_CFQQ(rq))
> >> 		return 1;
> >> 
> >>-	return 1;
> >>+	return 0;
> >> }
> >> 
> >> static inline void
> >>_
> >>
> >>But that might not fix things...
> >
> >Yeah it is, but I don't think it'll fix it (if anything, it'll be more
> >conservative).
> 
> (to possibly save others from trying -- no, doesn't fix any)

As expected. The issue is rq_is_sync(rq) takes the data direction into
account as well, while bio_sync() only checks the sync bit. This should
fix it.

diff --git a/block/cfq-iosched.c b/block/cfq-iosched.c
index 4b4217d..26fb40f 100644
--- a/block/cfq-iosched.c
+++ b/block/cfq-iosched.c
@@ -574,12 +574,14 @@ static int cfq_allow_merge(request_queue
 	struct cfq_data *cfqd = q->elevator->elevator_data;
 	const int rw = bio_data_dir(bio);
 	struct cfq_queue *cfqq;
+	int bio_is_sync;
 	pid_t key;
 
 	/*
 	 * Disallow merge, if bio and rq aren't both sync or async
 	 */
-	if (!!bio_sync(bio) != !!rq_is_sync(rq))
+	bio_is_sync = bio_data_dir(bio) == READ || bio_sync(bio);
+	if (bio_is_sync != !!rq_is_sync(rq))
 		return 0;
 
 	/*
@@ -592,7 +594,7 @@ static int cfq_allow_merge(request_queue
 	if (cfqq == RQ_CFQQ(rq))
 		return 1;
 
-	return 1;
+	return 0;
 }
 

-- 
Jens Axboe