Re: [PATCH] block: restore two stage elevator switch while running nr_hw_queue update

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

On Fri, Jul 18, 2025 at 12:13:25AM +0530, Nilay Shroff wrote:
> 
> 
> On 7/17/25 8:14 PM, Ming Lei wrote:
> 
> >> +static int blk_mq_elv_switch_none(struct request_queue *q,
> >> +		struct xarray *elv_tbl)
> >> +{
> >> +	int ret = 0;
> >> +
> >> +	lockdep_assert_held_write(&q->tag_set->update_nr_hwq_lock);
> >> +
> >> +	/*
> >> +	 * Accessing q->elevator without holding q->elevator_lock is safe here
> >> +	 * because we're called from nr_hw_queue update which is protected by
> >> +	 * set->update_nr_hwq_lock in the writer context. So, scheduler update/
> >> +	 * switch code (which acquires the same lock in the reader context)
> >> +	 * can't run concurrently.
> >> +	 */
> >> +	if (q->elevator) {
> >> +		char *elevator_name = (char *)q->elevator->type->elevator_name;
> >> +
> >> +		ret = xa_insert(elv_tbl, q->id, elevator_name, GFP_KERNEL);
> > 
> > This way isn't memory safe, because elevator module can be reloaded
> > during updating nr_hw_queues. One solution is to build a string<->int mapping
> > table, and store the (int) index of elevator type to xarray.
> > 
> Yes good point! How about avoiding this issue by using __elevator_get() and
> elevator_put()? We can take module reference while switching elevator to 'none'
> and then put the module reference while we switch back elevator.

Yeah, that is another way.

> 
> >> -void elv_update_nr_hw_queues(struct request_queue *q)
> >> -{
> >> -	struct elv_change_ctx ctx = {};
> >> -	int ret = -ENODEV;
> >> -
> >> -	WARN_ON_ONCE(q->mq_freeze_depth == 0);
> >> -
> >> -	mutex_lock(&q->elevator_lock);
> >> -	if (q->elevator && !blk_queue_dying(q) && blk_queue_registered(q)) {
> >> -		ctx.name = q->elevator->type->elevator_name;
> >> -
> >> -		/* force to reattach elevator after nr_hw_queue is updated */
> >> -		ret = elevator_switch(q, &ctx);
> >> -	}
> >> -	mutex_unlock(&q->elevator_lock);
> >> -	blk_mq_unfreeze_queue_nomemrestore(q);
> >> -	if (!ret)
> >> -		WARN_ON_ONCE(elevator_change_done(q, &ctx));
> >> -}
> >> -
> >>  /*
> >>   * Use the default elevator settings. If the chosen elevator initialization
> >>   * fails, fall back to the "none" elevator (no elevator).
> >> diff --git a/block/elevator.h b/block/elevator.h
> >> index a07ce773a38f..440b6e766848 100644
> >> --- a/block/elevator.h
> >> +++ b/block/elevator.h
> >> @@ -85,6 +85,17 @@ struct elevator_type
> >>  	struct list_head list;
> >>  };
> >>  
> >> +/* Holding context data for changing elevator */
> >> +struct elv_change_ctx {
> >> +	const char *name;
> >> +	bool no_uevent;
> >> +
> >> +	/* for unregistering old elevator */
> >> +	struct elevator_queue *old;
> >> +	/* for registering new elevator */
> >> +	struct elevator_queue *new;
> >> +};
> >> +
> > 
> > You may avoid the big chunk of code move for both `elv_change_ctx` and 
> > `elv_update_nr_hw_queues()`, not sure why you did that in a bug fix patch.
> > 
> This is because I’ve reintroduced the blk_mq_elv_switch_none and 
> blk_mq_elv_switch_back functions. I replaced elv_update_nr_hw_queues with
> blk_mq_elv_switch_back(), which was the approach used prior to commit 
> 596dce110b7d ("block: simplify elevator reattachment for updating nr_hw_queues").
> 
> Both blk_mq_elv_switch_none and blk_mq_elv_switch_back are defined in blk-mq.c
> and use the elevator APIs for switching elevators. These APIs — namely 
> elevator_switch and elevator_change_done — rely on the elv_change_ctx structure.
> As a result, I had to move some code around to ensure elv_change_ctx is accessible
> from within blk-mq.c.
> 
> While it would be possible to avoid this code movement by continuing to use 
> elv_update_nr_hw_queues, I felt it was cleaner to restore the two-stage elevator
> switch more fully by reintroducing the blk_mq_elv_switch_none and blk_mq_elv_switch_back 
> APIs, which were used prior to the simplification in the aforementioned commit.
> 
> Do you see any concerns with this code movement? Or is such restructuring generally
> avoided in a bug fix patch?

You still can add blk_mq_elv_switch_none() and blk_mq_elv_switch_back(),
meantime call elv_update_nr_hw_queues() from blk_mq_elv_switch_back() by
passing elevator type name.

If one bug fix is simpler, it may become easier to review & merge in -rc or
backport.


Thanks,
Ming