Re: [PATCH] blk-mq: Wait for for hctx requests on CPU unplug

[Keith Busch <kbusch@kernel.org>]

On Sat, Apr 06, 2019 at 02:27:10PM -0700, Ming Lei wrote:
> On Fri, Apr 05, 2019 at 05:36:32PM -0600, Keith Busch wrote:
> > On Fri, Apr 5, 2019 at 5:04 PM Jens Axboe <axboe@kernel.dk> wrote:
> > > Looking at current peak testing, I've got around 1.2% in queue enter
> > > and exit. It's definitely not free, hence my question. Probably safe
> > > to assume that we'll double that cycle counter, per IO.
> > 
> > Okay, that's not negligible at all. I don't know of a faster reference
> > than the percpu_ref, but that much overhead would have to rule out
> > having a per hctx counter.
> 
> Or not using any refcount in fast path, how about the following one?

Sure, I don't think we need a high precision completion wait in this path,
so a delay-spin seems okay to me.

 
> diff --git a/block/blk-mq.c b/block/blk-mq.c
> index 3ff3d7b49969..6fe334e12236 100644
> --- a/block/blk-mq.c
> +++ b/block/blk-mq.c
> @@ -2199,6 +2199,23 @@ int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
>  	return -ENOMEM;
>  }
>  
> +static void blk_mq_wait_hctx_become_idle(struct blk_mq_hw_ctx *hctx,
> +		int dead_cpu)
> +{
> +	unsigned long msecs_left = 1000 * 10;
> +
> +	while (msecs_left > 0) {
> +		if (blk_mq_hctx_idle(hctx))
> +			break;
> +		msleep(5);
> +		msecs_left -= 5;
> +	}
> +
> +	if (msecs_left > 0)
> +		printk(KERN_WARNING "requests not completed from "
> +			"CPU %d\n", dead_cpu);
> +}
> +
>  /*
>   * 'cpu' is going away. splice any existing rq_list entries from this
>   * software queue to the hw queue dispatch list, and ensure that it
> @@ -2230,6 +2247,14 @@ static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
>  	spin_unlock(&hctx->lock);
>  
>  	blk_mq_run_hw_queue(hctx, true);
> +
> +	/*
> +	 * Interrupt for this queue will be shutdown, so wait until all
> +	 * requests from this hctx is done or timeout.
> +	 */
> +	if (cpumask_first_and(hctx->cpumask, cpu_online_mask) >= nr_cpu_ids)
> +		blk_mq_wait_hctx_become_idle(hctx, cpu);
> +
>  	return 0;
>  }
>  
> diff --git a/block/blk-mq.h b/block/blk-mq.h
> index d704fc7766f4..935cf8519bf2 100644
> --- a/block/blk-mq.h
> +++ b/block/blk-mq.h
> @@ -240,4 +240,15 @@ static inline void blk_mq_clear_mq_map(struct blk_mq_queue_map *qmap)
>  		qmap->mq_map[cpu] = 0;
>  }
>  
> +static inline bool blk_mq_hctx_idle(struct blk_mq_hw_ctx *hctx)
> +{
> +	struct blk_mq_tags *tags = hctx->sched_tags ?: hctx->tags;
> +
> +	if (!tags)
> +		return true;
> +
> +	return !sbitmap_any_bit_set(&tags->bitmap_tags.sb) &&
> +                       !sbitmap_any_bit_set(&tags->bitmap_tags.sb);
> +}
> +
>  #endif
> 
> Thanks,
> Ming