Re: [PATCH 1/4] block: disable the elevator int del_gendisk

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

On Tue, Jun 14, 2022 at 09:48:24AM +0200, Christoph Hellwig wrote:
> The elevator is only used for file system requests, which are stopped in
> del_gendisk.  Move disabling the elevator and freeing the scheduler tags
> to the end of del_gendisk instead of doing that work in disk_release and
> blk_cleanup_queue to avoid a use after free on q->tag_set from
> disk_release as the tag_set might not be alive at that point.
> 
> Move the blk_qos_exit call as well, as it just depends on the elevator
> exit and would be the only reason to keep the not exactly cheap queue
> freeze in disk_release.
> 
> Fixes: e155b0c238b2 ("blk-mq: Use shared tags for shared sbitmap support")
> Reported-by: syzbot+3e3f419f4a7816471838@syzkaller.appspotmail.com
> Signed-off-by: Christoph Hellwig <hch@lst.de>
> Tested-by: syzbot+3e3f419f4a7816471838@syzkaller.appspotmail.com
> ---
>  block/blk-core.c | 13 -------------
>  block/genhd.c    | 39 +++++++++++----------------------------
>  2 files changed, 11 insertions(+), 41 deletions(-)
> 
> diff --git a/block/blk-core.c b/block/blk-core.c
> index 06ff5bbfe8f66..27fb1357ad4b8 100644
> --- a/block/blk-core.c
> +++ b/block/blk-core.c
> @@ -322,19 +322,6 @@ void blk_cleanup_queue(struct request_queue *q)
>  		blk_mq_exit_queue(q);
>  	}
>  
> -	/*
> -	 * In theory, request pool of sched_tags belongs to request queue.
> -	 * However, the current implementation requires tag_set for freeing
> -	 * requests, so free the pool now.
> -	 *
> -	 * Queue has become frozen, there can't be any in-queue requests, so
> -	 * it is safe to free requests now.
> -	 */
> -	mutex_lock(&q->sysfs_lock);
> -	if (q->elevator)
> -		blk_mq_sched_free_rqs(q);
> -	mutex_unlock(&q->sysfs_lock);
> -
>  	/* @q is and will stay empty, shutdown and put */
>  	blk_put_queue(q);
>  }
> diff --git a/block/genhd.c b/block/genhd.c
> index 27205ae47d593..e0675772178b0 100644
> --- a/block/genhd.c
> +++ b/block/genhd.c
> @@ -652,6 +652,17 @@ void del_gendisk(struct gendisk *disk)
>  
>  	blk_sync_queue(q);
>  	blk_flush_integrity();
> +	blk_mq_cancel_work_sync(q);
> +
> +	blk_mq_quiesce_queue(q);

quiesce queue adds a bit long delay in del_gendisk, not sure if this way may
cause regression in big machines with lots of disks.

> +	if (q->elevator) {
> +		mutex_lock(&q->sysfs_lock);
> +		elevator_exit(q);
> +		mutex_unlock(&q->sysfs_lock);
> +	}
> +	rq_qos_exit(q);
> +	blk_mq_unquiesce_queue(q);

Also tearing down elevator here has to be carefully, that means any
elevator reference has to hold rcu read lock or .q_usage_counter,
meantime it has to be checked, otherwise use-after-free may be caused.

Unfortunately, there are some cases which looks not safe, such as,
__blk_mq_update_nr_hw_queues() and blk_mq_has_sqsched().

Another example is bfq_insert_request()<-bfq_insert_requests():

static void bfq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
                               bool at_head)
{
		...
        spin_unlock_irq(&bfqd->lock);

        bfq_update_insert_stats(q, bfqq, idle_timer_disabled,
                                cmd_flags);
}

If last 'rq' is done between unlocking bfqd->lock and calling bfq_update_insert_stats,
del_gendisk() may tear down elevator, and UAF is caused.


Thanks,
Ming