Re: [PATCH 1/1] blk-mq: Fix disabled hybrid polling

[Omar Sandoval]

On Sat, May 25, 2019 at 04:42:11PM +0300, Pavel Begunkov (Silence) wrote:
> From: Pavel Begunkov <asml.silence@xxxxxxxxx>
> 
> Commit 4bc6339a583cec650b05 ("block: move blk_stat_add() to
> __blk_mq_end_request()") moved blk_stat_add(), so now it's called after
> blk_update_request(), which zeroes rq->__data_len. Without length,
> blk_stat_add() can't calculate stat bucket and returns error,
> effectively disabling hybrid polling.

I don't see how this patch fixes this problem, am I missing something?
The timing issue seems orthogonal.

> __blk_mq_end_request() is a right place to call blk_stat_add(), as it's
> guaranteed to be called for each request. Yet, calculating time there
> won't provide sufficient accuracy/precision for finer tuned hybrid
> polling, because a path from __blk_mq_complete_request() to
> __blk_mq_end_request() adds unpredictable overhead.
> 
> Add io_end_time_ns field in struct request, save time as soon as
> possible (at __blk_mq_complete_request()) and reuse later.
> 
> Signed-off-by: Pavel Begunkov <asml.silence@xxxxxxxxx>
> ---
>  block/blk-mq.c         | 13 ++++++++++---
>  block/blk-stat.c       |  4 ++--
>  block/blk-stat.h       |  2 +-
>  include/linux/blkdev.h | 11 +++++++++++
>  4 files changed, 24 insertions(+), 6 deletions(-)
> 
> diff --git a/block/blk-mq.c b/block/blk-mq.c
> index 32b8ad3d341b..8f6b6bfe0ccb 100644
> --- a/block/blk-mq.c
> +++ b/block/blk-mq.c
> @@ -330,6 +330,7 @@ static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
>  	else
>  		rq->start_time_ns = 0;
>  	rq->io_start_time_ns = 0;
> +	rq->io_end_time_ns = 0;
>  	rq->nr_phys_segments = 0;
>  #if defined(CONFIG_BLK_DEV_INTEGRITY)
>  	rq->nr_integrity_segments = 0;
> @@ -532,14 +533,17 @@ EXPORT_SYMBOL_GPL(blk_mq_free_request);
>  
>  inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
>  {
> -	u64 now = 0;
> +	u64 now = rq->io_end_time_ns;

Kyber expects the timestamp passed in to kyber_complete_request() to
include the software overhead. iostat should probably include the
software overhead, too. So, we probably won't be able to avoid calling
ktime_get() twice, once for I/O time and one for the end-to-end time.

> -	if (blk_mq_need_time_stamp(rq))
> +	/* called directly bypassing __blk_mq_complete_request */
> +	if (blk_mq_need_time_stamp(rq) && !now) {
>  		now = ktime_get_ns();
> +		rq->io_end_time_ns = now;
> +	}
>  
>  	if (rq->rq_flags & RQF_STATS) {
>  		blk_mq_poll_stats_start(rq->q);
> -		blk_stat_add(rq, now);
> +		blk_stat_add(rq);
>  	}
>  
>  	if (rq->internal_tag != -1)
> @@ -579,6 +583,9 @@ static void __blk_mq_complete_request(struct request *rq)
>  	bool shared = false;
>  	int cpu;
>  
> +	if (blk_mq_need_time_stamp(rq))
> +		rq->io_end_time_ns = ktime_get_ns();
> +
>  	WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
>  	/*
>  	 * Most of single queue controllers, there is only one irq vector
> diff --git a/block/blk-stat.c b/block/blk-stat.c
> index 940f15d600f8..9b9b30927ea8 100644
> --- a/block/blk-stat.c
> +++ b/block/blk-stat.c
> @@ -48,7 +48,7 @@ void blk_rq_stat_add(struct blk_rq_stat *stat, u64 value)
>  	stat->nr_samples++;
>  }
>  
> -void blk_stat_add(struct request *rq, u64 now)
> +void blk_stat_add(struct request *rq)
>  {
>  	struct request_queue *q = rq->q;
>  	struct blk_stat_callback *cb;
> @@ -56,7 +56,7 @@ void blk_stat_add(struct request *rq, u64 now)
>  	int bucket;
>  	u64 value;
>  
> -	value = (now >= rq->io_start_time_ns) ? now - rq->io_start_time_ns : 0;
> +	value = blk_rq_io_time(rq);
>  
>  	blk_throtl_stat_add(rq, value);
>  
> diff --git a/block/blk-stat.h b/block/blk-stat.h
> index 17b47a86eefb..2653818cee36 100644
> --- a/block/blk-stat.h
> +++ b/block/blk-stat.h
> @@ -65,7 +65,7 @@ struct blk_stat_callback {
>  struct blk_queue_stats *blk_alloc_queue_stats(void);
>  void blk_free_queue_stats(struct blk_queue_stats *);
>  
> -void blk_stat_add(struct request *rq, u64 now);
> +void blk_stat_add(struct request *rq);
>  
>  /* record time/size info in request but not add a callback */
>  void blk_stat_enable_accounting(struct request_queue *q);
> diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
> index 592669bcc536..2a8d4b68d707 100644
> --- a/include/linux/blkdev.h
> +++ b/include/linux/blkdev.h
> @@ -198,6 +198,9 @@ struct request {
>  	u64 start_time_ns;
>  	/* Time that I/O was submitted to the device. */
>  	u64 io_start_time_ns;
> +	/* Time that I/O was reported completed by the device. */
> +	u64 io_end_time_ns;
> +
>  
>  #ifdef CONFIG_BLK_WBT
>  	unsigned short wbt_flags;
> @@ -385,6 +388,14 @@ static inline int blkdev_reset_zones_ioctl(struct block_device *bdev,
>  
>  #endif /* CONFIG_BLK_DEV_ZONED */
>  
> +static inline u64 blk_rq_io_time(struct request *rq)
> +{
> +	u64 start = rq->io_start_time_ns;
> +	u64 end = rq->io_end_time_ns;
> +
> +	return (end - start) ? end - start : 0;

I think you meant:

	return end >= start ? end - start : 0;

> +}
> +
>  struct request_queue {
>  	/*
>  	 * Together with queue_head for cacheline sharing
> -- 
> 2.21.0
>