Re: [PATCH 12/15] block: introduce blk-iolatency io controller

[Josef Bacik]

On Wed, Jun 27, 2018 at 01:24:55PM -0600, Jens Axboe wrote:
> On 6/27/18 1:20 PM, Josef Bacik wrote:
> > On Wed, Jun 27, 2018 at 01:06:31PM -0600, Jens Axboe wrote:
> >> On 6/25/18 9:12 AM, Josef Bacik wrote:
> >>> +static void __blkcg_iolatency_throttle(struct rq_qos *rqos,
> >>> +				       struct iolatency_grp *iolat,
> >>> +				       spinlock_t *lock, bool issue_as_root,
> >>> +				       bool use_memdelay)
> >>> +	__releases(lock)
> >>> +	__acquires(lock)
> >>> +{
> >>> +	struct rq_wait *rqw = &iolat->rq_wait;
> >>> +	unsigned use_delay = atomic_read(&lat_to_blkg(iolat)->use_delay);
> >>> +	DEFINE_WAIT(wait);
> >>> +	bool first_block = true;
> >>> +
> >>> +	if (use_delay)
> >>> +		blkcg_schedule_throttle(rqos->q, use_memdelay);
> >>> +
> >>> +	/*
> >>> +	 * To avoid priority inversions we want to just take a slot if we are
> >>> +	 * issuing as root.  If we're being killed off there's no point in
> >>> +	 * delaying things, we may have been killed by OOM so throttling may
> >>> +	 * make recovery take even longer, so just let the IO's through so the
> >>> +	 * task can go away.
> >>> +	 */
> >>> +	if (issue_as_root || fatal_signal_pending(current)) {
> >>> +		atomic_inc(&rqw->inflight);
> >>> +		return;
> >>> +	}
> >>> +
> >>> +	if (iolatency_may_queue(iolat, &wait, first_block))
> >>> +		return;
> >>> +
> >>> +	do {
> >>> +		prepare_to_wait_exclusive(&rqw->wait, &wait,
> >>> +					  TASK_UNINTERRUPTIBLE);
> >>> +
> >>> +		iolatency_may_queue(iolat, &wait, first_block);
> >>> +		first_block = false;
> >>> +
> >>> +		if (lock) {
> >>> +			spin_unlock_irq(lock);
> >>> +			io_schedule();
> >>> +			spin_lock_irq(lock);
> >>> +		} else {
> >>> +			io_schedule();
> >>> +		}
> >>> +	} while (1);
> >>
> >> So how does this wait loop ever exit?
> >>
> > 
> > Sigh, I cleaned this up from what we're using in production and did it poorly,
> > I'll fix it up.  Thanks,
> 
> Also may want to consider NOT using exclusive add if first_block == false, as
> you'll end up at the tail of the waitqueue after sleeping and being denied.
> This is similar to the wbt change I posted last week.
> 

This isn't how it works though.  You aren't removed from the list until you do
finish_wait(), so you don't lose your spot on the list.  We only get added to
the end of the list if

        if (list_empty(&wq_entry->entry))

otherwise nothing changes.

> For may_queue(), your wq_has_sleeper() is also going to be always true
> inside your loop, since you call it after doing the prepare_to_wait()
> which adds you to the queue. That's why wbt does the list checks, but
> it'd be nicer to have a wq_has_other_sleepers() for that. So your
> first iolatency_may_queue() inside the loop will always be false.

Ah yeah that's a good point, I'll go back to using what you had to catch that
case.  Thanks,

Josef