On Thu, Apr 15, 2021 at 09:34:36AM +0800, JeffleXu wrote:
>
>
> On 4/14/21 7:24 PM, Ming Lei wrote:
> > On Wed, Apr 14, 2021 at 04:38:25PM +0800, JeffleXu wrote:
> >>
> >>
> >> On 4/12/21 5:38 PM, Christoph Hellwig wrote:
> >>> On Thu, Apr 01, 2021 at 10:19:26AM +0800, Ming Lei wrote:
> >>>> From: Jeffle Xu <jefflexu@xxxxxxxxxxxxxxxxx>
> >>>>
> >>>> This method can be used to check if bio-based device supports IO polling
> >>>> or not. For mq devices, checking for hw queue in polling mode is
> >>>> adequate, while the sanity check shall be implementation specific for
> >>>> bio-based devices. For example, dm device needs to check if all
> >>>> underlying devices are capable of IO polling.
> >>>>
> >>>> Though bio-based device may have done the sanity check during the
> >>>> device initialization phase, cacheing the result of this sanity check
> >>>> (such as by cacheing in the queue_flags) may not work. Because for dm
> >>>> devices, users could change the state of the underlying devices through
> >>>> '/sys/block/<dev>/io_poll', bypassing the dm device above. In this case,
> >>>> the cached result of the very beginning sanity check could be
> >>>> out-of-date. Thus the sanity check needs to be done every time 'io_poll'
> >>>> is to be modified.
> >>>
> >>> I really don't think thi should be a method, and I really do dislike
> >>> how we have all this "if (is_mq)" junk. Why can't we have a flag on
> >>> the gendisk that signals if the device can support polling that
> >>> is autoamtically set for blk-mq and as-needed by bio based drivers?
> >>
> >> That would consume one more bit of queue->queue_flags.
> >>
> >> Besides, DM/MD is somehow special here that when one of the underlying
> >> devices is disabled polling through '/sys/block/<dev>/io_poll',
> >> currently there's no mechanism notifying the above MD/DM to clear the
> >> previously set queue_flags. Thus the outdated queue_flags still
> >> indicates this DM/MD is capable of polling, while in fact one of the
> >> underlying device has been disabled for polling.
> >
> > Right, just like there isn't queue limit progagation.
> >
> > Another blocker could be that bio based queue doesn't support queue
> > freezing.
>
> Do you mean the queue freezing is called in the following code snippet?
>
> ```
> static ssize_t queue_poll_store(struct request_queue *q, const char
> *page, size_t count)
> {
> ...
> if (poll_on) {
> blk_queue_flag_set(QUEUE_FLAG_POLL, q);
> } else {
> blk_mq_freeze_queue(q);
> blk_queue_flag_clear(QUEUE_FLAG_POLL, q);
> blk_mq_unfreeze_queue(q);
> }
> ```
Yes, if it is a bio based queue. Or bio queued queue(DM, MD or others) may
use freeze_queue to do similar thing.
>
> And I can't understand how bio-based queue doesn't support queue freezing.
>
> ```
> submit_bio_noacct
> __submit_bio_noacct
> bio_queue_enter
> ```
>
> Every time submitting a bio, bio_queue_enter() will be called, and once
> the queue has been frozen, bio_queue_enter() will wait there until the
> queue is unfrozen.
Not like blk-mq, the refcount is just grabbed during submission for bio based
queue. I will research a bit and see if we can extend freeze queue for
covering bio based queue. One trouble is that bio is ended before
freeing request.
>
> >
> >>
> >> Mike had ever suggested that we can trust the queue_flag, and clear the
> >> outdated queue_flags when later the IO submission or polling routine
> >> finally finds that the device is not capable of polling. Currently
> >> submit_bio_checks() will silently clear the REQ_HIPRI flag and still
> >> submit the bio when the device is actually not capable of polling. To
> >> fix the issue, could we break the submission and return an error code in
> >> submit_bio_checks() if the device is not capable of polling when
> >> submitting HIPRI bio?
> >
> > I think we may just leave it alone, if underlying queue becomes not pollable,
> > the bio still can be submitted & completed via IRQ, just not efficient enough.
>
> Yes it still works. I agree if there's no better solution...
>
> And what about the issue Christoph originally concerned? Do we use one
> more flag bit indicating if the queue capable of polling, or the
> poll_capable() method way?
Just wondering why we can't use QUEUE_FLAG_POLL simply? If user wants to
enable it, let's do it for them. And bio driver can start with default poll
state by checking underlying queues.
Thanks,
Ming
