Re: 2.6.10-rc2-bk7 - Badness in cfq_put_request at drivers/block/cfq-iosched.c:1402

[Jens Axboe <axboe@suse.de>]

On Wed, Nov 24 2004, Christopher S. Aker wrote:
> > Can you try this simple check to see if it triggers anything?
> >
> > ===== cfq-iosched.c 1.13 vs edited =====
> > --- 1.13/drivers/block/cfq-iosched.c 2004-10-30 01:35:21 +02:00
> > +++ edited/cfq-iosched.c 2004-11-24 14:40:13 +01:00
> > @@ -1389,6 +1389,8 @@
> >   struct cfq_data *cfqd = q->elevator->elevator_data;
> >   struct cfq_rq *crq = RQ_DATA(rq);
> >
> > + WARN_ON(!spin_is_locked(q->queue_lock));
> > +
> >   if (crq) {
> >   struct cfq_queue *cfqq = crq->cfq_queue;
> 
> I'd be happy to, but I won't have a free machine for a couple of days.
> I'll can probably give it a shot during the weekend...

Nevermind, here's a patch to fix it. I was so focused on the decrement
side of things that I forgot to check the increment, pretty silly error
really.

Signed-off-by: Jens Axboe <axboe@suse.de>

===== drivers/block/cfq-iosched.c 1.12 vs edited =====
--- 1.12/drivers/block/cfq-iosched.c	2004-10-28 09:40:02 +02:00
+++ edited/drivers/block/cfq-iosched.c	2004-11-25 19:39:39 +01:00
@@ -1398,10 +1398,7 @@
 		if (crq->io_context)
 			put_io_context(crq->io_context->ioc);
 
-		if (!cfqq->allocated[crq->is_write]) {
-			WARN_ON(1);
-			cfqq->allocated[crq->is_write] = 1;
-		}
+		BUG_ON(!cfqq->allocated[crq->is_write]);
 		cfqq->allocated[crq->is_write]--;
 
 		mempool_free(crq, cfqd->crq_pool);
@@ -1442,6 +1439,7 @@
 	if (cfqq->allocated[rw] >= cfqd->max_queued)
 		goto out_lock;
 
+	cfqq->allocated[rw]++;
 	spin_unlock_irqrestore(q->queue_lock, flags);
 
 	/*
@@ -1465,7 +1463,6 @@
 		crq->in_flight = crq->accounted = crq->is_sync = 0;
 		crq->is_write = rw;
 		rq->elevator_private = crq;
-		cfqq->allocated[rw]++;
 		cfqq->alloc_limit[rw] = 0;
 		return 0;
 	}
@@ -1473,6 +1470,7 @@
 	put_io_context(cic->ioc);
 err:
 	spin_lock_irqsave(q->queue_lock, flags);
+	cfqq->allocated[rw]--;
 	cfq_put_queue(cfqq);
 out_lock:
 	spin_unlock_irqrestore(q->queue_lock, flags);

-- 
Jens Axboe