Date:   Thu, 30 May 2019 12:19:03 +0300
From:   Pavel Begunkov <asml.silence@...il.com>
To:     Omar Sandoval <osandov@...ndov.com>
Cc:     Jens Axboe <axboe@...nel.dk>, linux-block@...r.kernel.org,
        linux-kernel@...r.kernel.org, osandov@...com, ming.lei@...hat.com,
        Hou Tao <houtao1@...wei.com>
Subject: Re: [PATCH 1/1] blk-mq: Fix disabled hybrid polling

On 5/28/2019 9:37 PM, Omar Sandoval wrote:
> On Sat, May 25, 2019 at 04:42:11PM +0300, Pavel Begunkov (Silence) wrote:
>> From: Pavel Begunkov <asml.silence@...il.com>
>>
>> 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.
You're right, it got completely mixed up. It's rather an addition to a
patch from Hou Tao. But considering your comment below, it would be
better just to revert part of changes.


> 
>> __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@...il.com>
>> ---
>>  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.

I asked that in another thread, and it seems we can't reuse this.



> 
>> -	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;
> 
Sure, thanks



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



Download attachment "signature.asc" of type "application/pgp-signature" (834 bytes)

Powered by blists - more mailing lists