lkml.org 
[lkml]   [2014]   [Jun]   [20]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
SubjectRE: scsi-mq
Date


> -----Original Message-----
> From: Bart Van Assche [mailto:bvanassche@acm.org]
> Sent: Wednesday, 18 June, 2014 2:09 AM
> To: Jens Axboe; Christoph Hellwig; James Bottomley
> Cc: Elliott, Robert (Server Storage); linux-scsi@vger.kernel.org; linux-
> kernel@vger.kernel.org
> Subject: Re: scsi-mq
>
...
> Hello Jens,
>
> Fio reports the same queue depth for use_blk_mq=Y (mq below) and
> use_blk_mq=N (sq below), namely ">=64". However, the number of context
> switches differs significantly for the random read-write tests.
>
...
> It seems like with the traditional SCSI mid-layer and block core (sq)
> that the number of context switches does not depend too much on the
> number of I/O operations but that for the multi-queue SCSI core there
> are a little bit more than two context switches per I/O in the
> particular test I ran. The "randrw" script I used for this test takes
> SCSI LUNs as arguments (/dev/sdX) and starts the fio tool as follows:

Some of those context switches might be from scsi_end_request(),
which always schedules the scsi_requeue_run_queue() function via the
requeue_work workqueue for scsi-mq. That causes lots of context
switches from a busy application thread (e.g., fio) to a
kworker thread.

As shown by ftrace:

fio-19340 [005] dNh. 12067.908444: scsi_io_completion <-scsi_finish_command
fio-19340 [005] dNh. 12067.908444: scsi_end_request <-scsi_io_completion
fio-19340 [005] dNh. 12067.908444: blk_update_request <-scsi_end_request
fio-19340 [005] dNh. 12067.908445: blk_account_io_completion <-blk_update_request
fio-19340 [005] dNh. 12067.908445: scsi_mq_free_sgtables <-scsi_end_request
fio-19340 [005] dNh. 12067.908445: scsi_free_sgtable <-scsi_mq_free_sgtables
fio-19340 [005] dNh. 12067.908445: blk_account_io_done <-__blk_mq_end_io
fio-19340 [005] dNh. 12067.908445: blk_mq_free_request <-__blk_mq_end_io
fio-19340 [005] dNh. 12067.908446: blk_mq_map_queue <-blk_mq_free_request
fio-19340 [005] dNh. 12067.908446: blk_mq_put_tag <-__blk_mq_free_request
fio-19340 [005] .N.. 12067.908446: blkdev_direct_IO <-generic_file_direct_write
kworker/5:1H-3207 [005] .... 12067.908448: scsi_requeue_run_queue <-process_one_work
kworker/5:1H-3207 [005] .... 12067.908448: scsi_run_queue <-scsi_requeue_run_queue
kworker/5:1H-3207 [005] .... 12067.908448: blk_mq_start_stopped_hw_queues <-scsi_run_queue
fio-19340 [005] .... 12067.908449: blk_start_plug <-do_blockdev_direct_IO
fio-19340 [005] .... 12067.908449: blkdev_get_block <-do_direct_IO
fio-19340 [005] .... 12067.908450: blk_throtl_bio <-generic_make_request_checks
fio-19340 [005] .... 12067.908450: blk_sq_make_request <-generic_make_request
fio-19340 [005] .... 12067.908450: blk_queue_bounce <-blk_sq_make_request
fio-19340 [005] .... 12067.908450: blk_mq_map_request <-blk_sq_make_request
fio-19340 [005] .... 12067.908451: blk_mq_queue_enter <-blk_mq_map_request
fio-19340 [005] .... 12067.908451: blk_mq_map_queue <-blk_mq_map_request
fio-19340 [005] .... 12067.908451: blk_mq_get_tag <-__blk_mq_alloc_request
fio-19340 [005] .... 12067.908451: blk_mq_bio_to_request <-blk_sq_make_request
fio-19340 [005] .... 12067.908451: blk_rq_bio_prep <-init_request_from_bio
fio-19340 [005] .... 12067.908451: blk_recount_segments <-bio_phys_segments
fio-19340 [005] .... 12067.908452: blk_account_io_start <-blk_mq_bio_to_request
fio-19340 [005] .... 12067.908452: blk_mq_hctx_mark_pending <-__blk_mq_insert_request
fio-19340 [005] .... 12067.908452: blk_mq_run_hw_queue <-blk_sq_make_request
fio-19340 [005] .... 12067.908452: blk_mq_start_request <-__blk_mq_run_hw_queue

In one snapshot just tracing scsi_end_request() and
scsi_request_run_queue(), 30K scsi_end_request() calls yielded
20k scsi_request_run_queue() calls.

In this case, blk_mq_start_stopped_hw_queues() doesn't end up
doing anything since there aren't any stopped queues to restart
(blk_mq_run_hw_queue() gets called a bit later during routine
fio work); the context switch turned out to be a waste of time.
If it did find a stopped queue, then it would call
blk_mq_run_hw_queue() itself.

---
Rob Elliott HP Server Storage



\
 
 \ /
  Last update: 2014-06-21 03:41    [W:0.073 / U:6.696 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site