[dm-devel] block: avoid unconditionally freeing previously allocated request_queue

[Mike Snitzer]

On Tue, May 25 2010 at 10:37pm -0400,
Kiyoshi Ueda <k-ueda at ct.jp.nec.com> wrote:

> Hi Mike,
> 
> On 05/26/2010 01:34 AM +0900, Mike Snitzer wrote:
> > Mike Snitzer <snitzer at redhat.com> wrote:
> >> Kiyoshi Ueda <k-ueda at ct.jp.nec.com> wrote:
> >>>> +/*
> >>>> + * Fully initialize a request-based queue (->elevator, ->request_fn, etc).
> >>>> + */
> >>>> +static int dm_init_request_based_queue(struct mapped_device *md)
> >>>> +{
> >>>> +	struct request_queue *q = NULL;
> >>>> +
> >>>> +	/* Avoid re-initializing the queue if already fully initialized */
> >>>> +	if (!md->queue->elevator) {
> >>>> +		/* Fully initialize the queue */
> >>>> +		q = blk_init_allocated_queue(md->queue, dm_request_fn, NULL);
> >>>> +		if (!q)
> >>>> +			return 0;
> >>>
> >>> When blk_init_allocated_queue() fails, the block-layer seems not to
> >>> guarantee that the queue is still available.
> >>
> >> Ouch, yes this portion of blk_init_allocated_queue_node() is certainly
> >> problematic:
> >>
> >>         if (blk_init_free_list(q)) {
> >>                 kmem_cache_free(blk_requestq_cachep, q);
> >>                 return NULL;
> >>         }
> 
> Not only that.  The blk_put_queue() in blk_init_allocated_queue_node()
> will also free the queue:
> 
> 	if (!elevator_init(q, NULL)) {
> 		blk_queue_congestion_threshold(q);
> 		return q;
> 	}
> 
> 	blk_put_queue(q);
> 	return NULL;

OK, I'll post v2 that addresses this and we'll see what Jens says.

Mike