Re: [RFC PATCH] blk-mq: Clean up references when freeing rqs

[Ming Lei <ming.lei@redhat.com>]

On Wed, Dec 02, 2020 at 11:18:31AM +0000, John Garry wrote:
> On 02/12/2020 03:31, Ming Lei wrote:
> > On Tue, Dec 01, 2020 at 09:02:18PM +0800, John Garry wrote:
> > > It has been reported many times that a use-after-free can be intermittently
> > > found when iterating busy requests:
> > > 
> > > - https://lore.kernel.org/linux-block/8376443a-ec1b-0cef-8244-ed584b96fa96@huawei.com/
> > > - https://lore.kernel.org/linux-block/5c3ac5af-ed81-11e4-fee3-f92175f14daf@acm.org/T/#m6c1ac11540522716f645d004e2a5a13c9f218908
> > > - https://lore.kernel.org/linux-block/04e2f9e8-79fa-f1cb-ab23-4a15bf3f64cc@kernel.dk/
> > > 
> > > The issue is that when we switch scheduler or change queue nr_requests,
> > > the driver tagset may keep references to the stale requests.
> > > 
> > > As a solution, clean up any references to those requests in the driver
> > > tagset when freeing. This is done with a cmpxchg to make safe any race
> > > with setting the driver tagset request from another queue.
> > > 
> > > Signed-off-by: John Garry <john.garry@huawei.com>
> > > --
> > > Set as RFC as I need to test more. And not sure on solution method, as
> > > Bart had another idea.
> > > 
> > > diff --git a/block/blk-mq-sched.c b/block/blk-mq-sched.c
> > > index d1eafe2c045c..9b042c7036b3 100644
> > > --- a/block/blk-mq-sched.c
> > > +++ b/block/blk-mq-sched.c
> > > @@ -621,7 +621,7 @@ void blk_mq_sched_free_requests(struct request_queue *q)
> > >   	queue_for_each_hw_ctx(q, hctx, i) {
> > >   		if (hctx->sched_tags)
> > > -			blk_mq_free_rqs(q->tag_set, hctx->sched_tags, i);
> > > +			blk_mq_free_rqs_ext(q->tag_set, hctx->sched_tags, i, hctx->tags);
> > >   	}
> > >   }
> > > diff --git a/block/blk-mq-tag.c b/block/blk-mq-tag.c
> > > index 9c92053e704d..562db72e7d79 100644
> > > --- a/block/blk-mq-tag.c
> > > +++ b/block/blk-mq-tag.c
> > > @@ -576,7 +576,7 @@ int blk_mq_tag_update_depth(struct blk_mq_hw_ctx *hctx,
> > >   			return -ENOMEM;
> > >   		}
> > > -		blk_mq_free_rqs(set, *tagsptr, hctx->queue_num);
> > > +		blk_mq_free_rqs_ext(set, *tagsptr, hctx->queue_num, hctx->tags);
> > >   		blk_mq_free_rq_map(*tagsptr, flags);
> > >   		*tagsptr = new;
> > >   	} else {
> > > diff --git a/block/blk-mq.c b/block/blk-mq.c
> > > index 55bcee5dc032..f3aad695cd25 100644
> > > --- a/block/blk-mq.c
> > > +++ b/block/blk-mq.c
> > > @@ -2271,8 +2271,8 @@ blk_qc_t blk_mq_submit_bio(struct bio *bio)
> > >   	return BLK_QC_T_NONE;
> > >   }
> > > -void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
> > > -		     unsigned int hctx_idx)
> > > +void blk_mq_free_rqs_ext(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
> > > +		     unsigned int hctx_idx, struct blk_mq_tags *references)
> > >   {
> > >   	struct page *page;
> > > @@ -2281,10 +2281,13 @@ void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
> > >   		for (i = 0; i < tags->nr_tags; i++) {
> > >   			struct request *rq = tags->static_rqs[i];
> > > +			int j;
> > >   			if (!rq)
> > >   				continue;
> > >   			set->ops->exit_request(set, rq, hctx_idx);
> > > +			for (j = 0; references && j < references->nr_tags; j++)
> > > +				cmpxchg(&references->rqs[j], rq, 0);
> > 
> > Seems you didn't address the comment in the following link:
> > 
> > 	https://lore.kernel.org/linux-block/10331543-9e45-ae63-8cdb-17e5a2a3b7ef@huawei.com/
> > 
> > The request to be freed may still be refered in another path, such as blk_mq_queue_tag_busy_iter
> > or blk_mq_tagset_busy_iter(), and cmpxchg() doesn't drain/wait for other refers.
> > 
> 
> Hi Ming,
> 
> Yeah, so I said that was another problem which you mentioned there, which
> I'm not addressing, but I don't think that I'm making thing worse here.

The thing is that this patch does not fix the issue completely.

> 
> So AFAICS, the blk-mq/sched code doesn't wait for any "readers" to be
> finished, such as those running blk_mq_queue_tag_busy_iter or
> blk_mq_tagset_busy_iter() in another context.
> 
> So how about the idea of introducing some synchronization primitive, such as
> semaphore, which those "readers" must grab and release at start and end (of
> iter), to ensure the requests are not freed during the iteration?

It looks good, however devil is in details, please make into patch for
review.


thanks,
Ming