Re: [PATCH] blk-mq: fix kernel panic during iterating over flush request

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

On Thu, Aug 12, 2021 at 07:47:55PM +0800, Xiaoguang Wang wrote:
> hi,
> 
> 
> > For fixing use-after-free during iterating over requests, we grabbed
> > request's refcount before calling ->fn in commit 2e315dc07df0 ("blk-mq:
> > grab rq->refcount before calling ->fn in blk_mq_tagset_busy_iter").
> > Turns out this way may cause kernel panic when iterating over one flush
> > request:
> > 
> > 1) old flush request's tag is just released, and this tag is reused by
> > one new request, but ->rqs[] isn't updated yet
> > 
> > 2) the flush request can be re-used for submitting one new flush command,
> > so blk_rq_init() is called at the same time
> > 
> > 3) meantime blk_mq_queue_tag_busy_iter() is called, and old flush request
> > is retrieved from ->rqs[tag]; when blk_mq_put_rq_ref() is called,
> > flush_rq->end_io may not be updated yet, so NULL pointer dereference
> > is triggered in blk_mq_put_rq_ref().
> > 
> > Fix the issue by calling refcount_set(&flush_rq->ref, 1) after
> > flush_rq->end_io is set. So far the only other caller of blk_rq_init() is
> > scsi_ioctl_reset() in which the request doesn't enter block IO stack and
> > the request reference count isn't used, so the change is safe.
> > 
> > Fixes: 2e315dc07df0 ("blk-mq: grab rq->refcount before calling ->fn in
> > blk_mq_tagset_busy_iter")
> > Reported-by: "Blank-Burian, Markus, Dr." <blankburian@uni-muenster.de>
> > Tested-by: "Blank-Burian, Markus, Dr." <blankburian@uni-muenster.de>
> > Signed-off-by: Ming Lei <ming.lei@redhat.com>
> > ---
> >   block/blk-core.c  | 1 -
> >   block/blk-flush.c | 8 ++++++++
> >   2 files changed, 8 insertions(+), 1 deletion(-)
> > 
> > diff --git a/block/blk-core.c b/block/blk-core.c
> > index 0874bc2fcdb4..b5098739f72a 100644
> > --- a/block/blk-core.c
> > +++ b/block/blk-core.c
> > @@ -121,7 +121,6 @@ void blk_rq_init(struct request_queue *q, struct request *rq)
> >   	rq->internal_tag = BLK_MQ_NO_TAG;
> >   	rq->start_time_ns = ktime_get_ns();
> >   	rq->part = NULL;
> > -	refcount_set(&rq->ref, 1);
> >   	blk_crypto_rq_set_defaults(rq);
> >   }
> >   EXPORT_SYMBOL(blk_rq_init);
> > diff --git a/block/blk-flush.c b/block/blk-flush.c
> > index 1002f6c58181..4912c8dbb1d8 100644
> > --- a/block/blk-flush.c
> > +++ b/block/blk-flush.c
> > @@ -329,6 +329,14 @@ static void blk_kick_flush(struct request_queue *q, struct blk_flush_queue *fq,
> >   	flush_rq->rq_flags |= RQF_FLUSH_SEQ;
> >   	flush_rq->rq_disk = first_rq->rq_disk;
> >   	flush_rq->end_io = flush_end_io;
> > +	/*
> > +	 * Order WRITE ->end_io and WRITE rq->ref, and its pair is the one
> > +	 * implied in refcount_inc_not_zero() called from
> > +	 * blk_mq_find_and_get_req(), which orders WRITE/READ flush_rq->ref
> > +	 * and READ flush_rq->end_io
> 
> Recently we run into similar panic which is a NULL dereference on
> rq->mq_hctx in is_flush_rq(), we also
> 
> guess there is race bug just what you have fixed.
> 
> But I have one question here, for a blk-mq device, before issuing the first
> flush req, flush_rq->end_io
> 
> is NULL, and for following flush reqs on this blk-mq device,
> flush_rq->end_io won't be NULL. I searched
> 
> the codes and don't find any place that sets flush_rq->end_io to be NULL
> once flush_rq has been completed.

blk_kick_flush():
	blk_rq_init
		memset(rq, 0, sizeof(*rq));


Thanks,
Ming