[dm-devel] [PATCH] block: introduce QUEUE_FLAG_POLL_CAP flag

[Ming Lei]

On Sat, Apr 17, 2021 at 10:06:53PM +0800, JeffleXu wrote:
> 
> 
> On 4/16/21 5:07 PM, Ming Lei wrote:
> > On Fri, Apr 16, 2021 at 04:00:37PM +0800, Jeffle Xu wrote:
> >> Hi,
> >> How about this patch to remove the extra poll_capable() method?
> >>
> >> And the following 'dm: support IO polling for bio-based dm device' needs
> >> following change.
> >>
> >> ```
> >> +       /*
> >> +        * Check for request-based device is remained to
> >> +        * dm_mq_init_request_queue()->blk_mq_init_allocated_queue().
> >> +        * For bio-based device, only set QUEUE_FLAG_POLL when all underlying
> >> +        * devices supporting polling.
> >> +        */
> >> +       if (__table_type_bio_based(t->type)) {
> >> +               if (dm_table_supports_poll(t)) {
> >> +                       blk_queue_flag_set(QUEUE_FLAG_POLL_CAP, q);
> >> +                       blk_queue_flag_set(QUEUE_FLAG_POLL, q);
> >> +               }
> >> +               else {
> >> +                       blk_queue_flag_clear(QUEUE_FLAG_POLL, q);
> >> +                       blk_queue_flag_clear(QUEUE_FLAG_POLL_CAP, q);
> >> +               }
> >> +       }
> >> ```
> > 
> > Frankly speaking, I don't see any value of using QUEUE_FLAG_POLL_CAP for
> > DM, and the result is basically subset of treating DM as always being capable
> > of polling.
> > 
> > Also underlying queue change(either limits or flag) won't be propagated
> > to DM/MD automatically. Strictly speaking it doesn't matter if all underlying
> > queues are capable of supporting polling at the exact time of 'write sysfs/poll',
> > cause any of them may change in future.
> > 
> > So why not start with the simplest approach(always capable of polling)
> > which does meet normal bio based polling requirement?
> > 
> 
> I find one scenario where this issue may matter. Consider the scenario
> where HIPRI bios are submitted to DM device though **all** underlying
> devices has been disabled for polling. In this case, a **valid** cookie
> (pid of current submitting process) is still returned. Then if @spin of
> the following blk_poll() is true, blk_poll() will get stuck in dead loop
> because blk_mq_poll() always returns 0, since previously submitted bios
> are all enqueued into IRQ hw queue.
> 
> Maybe you need to re-remove the bio from the poll context if the
> returned cookie is BLK_QC_T_NONE?

It won't be one issue, see blk_bio_poll_preprocess() which is called
from submit_bio_checks(), so any bio's HIPRI will be cleared if the
queue doesn't support POLL, that code does cover underlying bios.

> 
> 
> Something like:
> 
> -static blk_qc_t __submit_bio_noacct(struct bio *bio)
> +static blk_qc_t __submit_bio_noacct_ctx(struct bio *bio, struct
> io_context *ioc)
>  {
>  	struct bio_list bio_list_on_stack[2];
>  	blk_qc_t ret = BLK_QC_T_NONE;
> @@ -1047,7 +1163,15 @@ static blk_qc_t __submit_bio_noacct(struct bio *bio)
>  		bio_list_on_stack[1] = bio_list_on_stack[0];
>  		bio_list_init(&bio_list_on_stack[0]);
> 
> 		if (ioc && queue_is_mq(q) && (bio->bi_opf & REQ_HIPRI)) {

REQ_HIPRI won't be set for underlying bios which queue doesn't support
poll, so this branch won't be reached. And the submission queue will
be empty, and blk_poll() for DM/MD(bio based queue) checks nothing, but
the polling won't be stopped until the iocb is completed. And this
handling is actually same with current polling on IRQ based queue.

> 			bool queued = blk_bio_poll_prep_submit(ioc, bio);
> 
> 			ret = __submit_bio(bio);
> +			if (queued && !blk_qc_t_valid(ret))
> 				/* TODO:remove bio from poll_context */
> 				
> 				bio_set_private_data(bio, ret);
> 		} else {
> 			ret = __submit_bio(bio);
> 		}
> 
> 
> Then if you'd like fix this in this way, the returned value of
> .submit_bio() of DM/MD also needs to return BLK_QC_T_NONE now. Currently
> .submit_bio() of DM actually returns the cookie of the last split bio
> (to underlying mq deivice).

I am a bit confused, this patch requires .submit_bio() of DM/MD(bio
based queue) to return either 0 or pid of the submission task.


Thanks,
Ming