Re: [PATCH RESEND 1/5] block/blk-rq-qos: add a new helper rq_qos_add_freezed()

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

On Mon, Nov 17, 2025 at 07:39:57PM +0800, Yu Kuai wrote:
> Hi,
> 
> 在 2025/11/17 19:30, Ming Lei 写道:
> > On Mon, Nov 17, 2025 at 04:43:11PM +0530, Nilay Shroff wrote:
> >>
> >> On 11/17/25 4:31 PM, Ming Lei wrote:
> >>> On Sun, Nov 16, 2025 at 12:10:20PM +0800, Yu Kuai wrote:
> >>>> queue should not be freezed under rq_qos_mutex, see example index
> >>>> commit 9730763f4756 ("block: correct locking order for protecting blk-wbt
> >>>> parameters"), which means current implementation of rq_qos_add() is
> >>>> problematic. Add a new helper and prepare to fix this problem in
> >>>> following patches.
> >>>>
> >>>> Signed-off-by: Yu Kuai <yukuai@fnnas.com>
> >>>> ---
> >>>>   block/blk-rq-qos.c | 27 +++++++++++++++++++++++++++
> >>>>   block/blk-rq-qos.h |  2 ++
> >>>>   2 files changed, 29 insertions(+)
> >>>>
> >>>> diff --git a/block/blk-rq-qos.c b/block/blk-rq-qos.c
> >>>> index 654478dfbc20..353397d7e126 100644
> >>>> --- a/block/blk-rq-qos.c
> >>>> +++ b/block/blk-rq-qos.c
> >>>> @@ -322,6 +322,33 @@ void rq_qos_exit(struct request_queue *q)
> >>>>   	mutex_unlock(&q->rq_qos_mutex);
> >>>>   }
> >>>>   
> >>>> +int rq_qos_add_freezed(struct rq_qos *rqos, struct gendisk *disk,
> >>>> +		       enum rq_qos_id id, const struct rq_qos_ops *ops)
> >>>> +{
> >>>> +	struct request_queue *q = disk->queue;
> >>>> +
> >>>> +	WARN_ON_ONCE(q->mq_freeze_depth == 0);
> >>>> +	lockdep_assert_held(&q->rq_qos_mutex);
> >>>> +
> >>>> +	if (rq_qos_id(q, id))
> >>>> +		return -EBUSY;
> >>>> +
> >>>> +	rqos->disk = disk;
> >>>> +	rqos->id = id;
> >>>> +	rqos->ops = ops;
> >>>> +	rqos->next = q->rq_qos;
> >>>> +	q->rq_qos = rqos;
> >>>> +	blk_queue_flag_set(QUEUE_FLAG_QOS_ENABLED, q);
> >>>> +
> >>>> +	if (rqos->ops->debugfs_attrs) {
> >>>> +		mutex_lock(&q->debugfs_mutex);
> >>>> +		blk_mq_debugfs_register_rqos(rqos);
> >>>> +		mutex_unlock(&q->debugfs_mutex);
> >>>> +	}
> >>> It will cause more lockdep splat to let q->debugfs_mutex depend on queue freeze,
> >>>
> >> I think we already have that ->debugfs_mutex dependency on ->freeze_lock.
> >> for instance,
> >>    ioc_qos_write  => freeze-queue
> >>     blk_iocost_init
> >>       rq_qos_add
> > Why is queue freeze needed in above code path?
> >
> > Also blk_iolatency_init()/rq_qos_add() doesn't freeze queue.
> 
> I don't quite understand, rq_qos_add() always require queue freeze, prevent
> deference q->rq_qos from IO path concurrently.
> 
> >
> >> and also,
> >>    queue_wb_lat_store  => freeze-queue
> >>      wbt_init
> >>        rq_qos_add
> > Not all wbt_enable_default()/wbt_init() is called with queue frozen, but Kuai's
> > patchset changes all to freeze queue before registering debugfs entry, people will
> > complain new warning.
> 
> Yes, but the same as above, rq_qos_add() from wbt_init() will always freeze queue
> before this set, so I don't understand why is there new warning?

The in-tree rq_qos_add() registers debugfs after queue is unfreeze, but
your patchset basically moves queue freeze/unfreeze to callsite of rq_qos_add(),
then debugfs register is always done with queue frozen.

Dependency between queue freeze and q->debugfs_mutex is introduced in some
code paths, such as, elevator switch, blk_iolatency_init, ..., this way
will trigger warning because it isn't strange to run into memory
allocation in debugfs_create_*().

> 
> >
> >>> Also blk_mq_debugfs_register_rqos() does _not_ require queue to be frozen,
> >>> and it should be fine to move blk_mq_debugfs_register_rqos() out of queue
> >>> freeze.
> >>>
> >> Yes correct, but I thought this pacthset is meant only to address incorrect
> >> locking order between ->rq_qos_mutex and ->freeze_lock. So do you suggest
> >> also refactoring code to avoid ->debugfs_mutex dependency on ->freeze_lock?
> >> If yes then shouldn't that be handled in a separate patchset?
> > It is fine to fix in that way, but at least regression shouldn't be caused.
> >
> > More importantly we shouldn't add new unnecessary dependency on queue freeze.
> 
> This is correct, I'll work on the v2 set to move debugfs_mutex outside of freeze
> queue, however, as you suggested before we should we should fix this incorrect
> lock order first. How about I make them in a single set?

That is fine, but patches for moving debugfs_mutex should be put before
this patchset, which is always friendly for 'git bisect'.


Thanks,
Ming