[Drbd-dev] [PATCH for-6.2/block V3 2/2] block: Change the granularity of io ticks from ms to ns

[Gulam Mohamed]

Hi Keith,

   Thanks for reviewing this request. Can you please see my inline comments?

Regards,
Gulam Mohamed.

-----Original Message-----
From: Keith Busch <kbusch at kernel.org> 
Sent: Wednesday, December 21, 2022 10:39 PM
To: Gulam Mohamed <gulam.mohamed at oracle.com>
Cc: linux-block at vger.kernel.org; axboe at kernel.dk; philipp.reisner at linbit.com; lars.ellenberg at linbit.com; christoph.boehmwalder at linbit.com; minchan at kernel.org; ngupta at vflare.org; senozhatsky at chromium.org; colyli at suse.de; kent.overstreet at gmail.com; agk at redhat.com; snitzer at kernel.org; dm-devel at redhat.com; song at kernel.org; dan.j.williams at intel.com; vishal.l.verma at intel.com; dave.jiang at intel.com; ira.weiny at intel.com; Junxiao Bi <junxiao.bi at oracle.com>; Martin Petersen <martin.petersen at oracle.com>; kch at nvidia.com; drbd-dev at lists.linbit.com; linux-kernel at vger.kernel.org; linux-bcache at vger.kernel.org; linux-raid at vger.kernel.org; nvdimm at lists.linux.dev; Konrad Wilk <konrad.wilk at oracle.com>; Joe Jin <joe.jin at oracle.com>; Rajesh Sivaramasubramaniom <rajesh.sivaramasubramaniom at oracle.com>; Shminderjit Singh <shminderjit.singh at oracle.com>
Subject: Re: [PATCH for-6.2/block V3 2/2] block: Change the granularity of io ticks from ms to ns

On Wed, Dec 21, 2022 at 04:05:06AM +0000, Gulam Mohamed wrote:
> +u64  blk_get_iostat_ticks(struct request_queue *q) {
> +       return (blk_queue_precise_io_stat(q) ? ktime_get_ns() : 
> +jiffies); } EXPORT_SYMBOL_GPL(blk_get_iostat_ticks);
> +
>  void update_io_ticks(struct block_device *part, u64 now, bool end)  {
>  	u64 stamp;
> @@ -968,20 +980,26 @@ EXPORT_SYMBOL(bdev_start_io_acct);
>  u64 bio_start_io_acct(struct bio *bio)  {
>  	return bdev_start_io_acct(bio->bi_bdev, bio_sectors(bio),
> -				  bio_op(bio), jiffies);
> +				  bio_op(bio),
> +				  blk_get_iostat_ticks(bio->bi_bdev->bd_queue));
>  }
>  EXPORT_SYMBOL_GPL(bio_start_io_acct);
>  
>  void bdev_end_io_acct(struct block_device *bdev, enum req_op op,
>  		      u64 start_time)
>  {
> +	u64 now;
> +	u64 duration;
> +	struct request_queue *q = bdev_get_queue(bdev);
>  	const int sgrp = op_stat_group(op);
> -	u64 now = READ_ONCE(jiffies);
> -	u64 duration = (unsigned long)now -(unsigned long) start_time;
> +	now = blk_get_iostat_ticks(q);;

I don't think you can rely on the blk_queue_precise_io_stat() flag in the completion side. The user can toggle this with data in flight, which means the completion may use different tick units than the start. I think you'll need to add a flag to the request in the start accounting side to know which method to use for the completion.

[GULAM]: As per my understanding, this may work for a single request_queue implemetation. But this request based accounting, as per my understanding, may be an issue with the Multi-QUEUE as there is a separate queue for each CPU and the time-stamp being recorded for the block device is a global one. Also, the issue you mentioned about the start and end accounting may update the ticks in different 
units for the inflight IOs, may be just for a while. So, even if it works for MQ, I am trying to understand how much is it feasible to do this request-based change for an issue which may be there for just a moment?
So, can you please correct me if I am wrong and explore more on your suggestion so that I can understand properly?

> @@ -951,6 +951,7 @@ ssize_t part_stat_show(struct device *dev,
>  	struct request_queue *q = bdev_get_queue(bdev);
>  	struct disk_stats stat;
>  	unsigned int inflight;
> +	u64 stat_ioticks;
>  
>  	if (queue_is_mq(q))
>  		inflight = blk_mq_in_flight(q, bdev); @@ -959,10 +960,13 @@ ssize_t 
> part_stat_show(struct device *dev,
>  
>  	if (inflight) {
>  		part_stat_lock();
> -		update_io_ticks(bdev, jiffies, true);
> +		update_io_ticks(bdev, blk_get_iostat_ticks(q), true);
>  		part_stat_unlock();
>  	}
>  	part_stat_read_all(bdev, &stat);
> +	stat_ioticks = (blk_queue_precise_io_stat(q) ?
> +				div_u64(stat.io_ticks, NSEC_PER_MSEC) :
> +				jiffies_to_msecs(stat.io_ticks));


With the user able to change the precision at will, I think these io_ticks need to have a consistent unit size. Mixing jiffies and nsecs is going to create garbage stats output. Could existing io_ticks using jiffies be converted to jiffies_to_nsecs() so that you only have one unit to consider?
[GULAM]: I am not sure if this will work as we just multiply with 1000000 to convert jiffies to nano-seconds.