On 2020-08-19 11:20 a.m., John Garry wrote:
Hi all,
Here is v8 of the patchset.
In this version of the series, we keep the shared sbitmap for driver tags,
and introduce changes to fix up the tag budgeting across request queues.
We also have a change to count requests per-hctx for when an elevator is
enabled, as an optimisation. I also dropped the debugfs changes - more on
that below.
Some performance figures:
Using 12x SAS SSDs on hisi_sas v3 hw. mq-deadline results are included,
but it is not always an appropriate scheduler to use.
Tag depth 4000 (default) 260**
Baseline (v5.9-rc1):
none sched: 2094K IOPS 513K
mq-deadline sched: 2145K IOPS 1336K
Final, host_tagset=0 in LLDD *, ***:
none sched: 2120K IOPS 550K
mq-deadline sched: 2121K IOPS 1309K
Final ***:
none sched: 2132K IOPS 1185
mq-deadline sched: 2145K IOPS 2097
* this is relevant as this is the performance in supporting but not
enabling the feature
** depth=260 is relevant as some point where we are regularly waiting for
tags to be available. Figures were are a bit unstable here.
*** Included "[PATCH V4] scsi: core: only re-run queue in
scsi_end_request() if device queue is busy"
A copy of the patches can be found here:
https://github.com/hisilicon/kernel-dev/tree/private-topic-blk-mq-shared-tags-v8
The hpsa patch depends on:
https://lore.kernel.org/linux-scsi/20200430131904.5847-1-hare@xxxxxxx/
And the smartpqi patch is not to be accepted.
Comments (and testing) welcome, thanks!
I tested this v8 patchset on MKP's 5.10/scsi-queue branch together with
my rewritten sg driver on my laptop and a Ryzen 5 3600 machine. Since I
don't have same hardware, I use the scsi_debug driver as the target:
modprobe scsi_debug dev_size_mb=1024 sector_size=512 add_host=7
per_host_store=1 ndelay=1000 random=1 submit_queues=12
My test is a script which runs these three commands many times with
differing parameters:
sg_mrq_dd iflag=random bs=512 of=/dev/sg8 thr=64 time=2
time to transfer data was 0.312705 secs, 3433.72 MB/sec
2097152+0 records in
2097152+0 records out
sg_mrq_dd bpt=256 thr=64 mrq=36 time=2 if=/dev/sg8 bs=512 of=/dev/sg9
time to transfer data was 0.212090 secs, 5062.67 MB/sec
2097152+0 records in
2097152+0 records out
sg_mrq_dd --verify if=/dev/sg8 of=/dev/sg9 bs=512 bpt=256 thr=64 mrq=36 time=2
Doing verify/cmp rather than copy
time to transfer data was 0.184563 secs, 5817.75 MB/sec
2097152+0 records in
2097152+0 records verified
The above is the output from last section of the my script run on the Ryzen 5.
So the three steps are:
1) produce random data on /dev/sg8
2) copy /dev/sg8 to /dev/sg9
3) verify /dev/sg8 and /dev/sg9 are the same.
The latter step is done with a sequence of READ(/dev/sg8) and
VERIFY(BYTCHK=1 on /dev/sg9). The "mrq" stands for multiple requests (in
one invocation; the bsg driver did that before its write(2) command was
removed.
The SCSI devices on the Ryzen 5 machine are:
# lsscsi -gs
[2:0:0:0] disk IBM-207x HUSMM8020ASS20 J4B6 /dev/sda /dev/sg0 200GB
[2:0:1:0] disk SEAGATE ST200FM0073 0007 /dev/sdb /dev/sg1 200GB
[2:0:2:0] enclosu Areca Te ARC-802801.37.69 0137 - /dev/sg2 -
[3:0:0:0] disk Linux scsi_debug 0190 /dev/sdc /dev/sg3 1.07GB
[4:0:0:0] disk Linux scsi_debug 0190 /dev/sdd /dev/sg4 1.07GB
[5:0:0:0] disk Linux scsi_debug 0190 /dev/sde /dev/sg5 1.07GB
[6:0:0:0] disk Linux scsi_debug 0190 /dev/sdf /dev/sg6 1.07GB
[7:0:0:0] disk Linux scsi_debug 0190 /dev/sdg /dev/sg7 1.07GB
[8:0:0:0] disk Linux scsi_debug 0190 /dev/sdh /dev/sg8 1.07GB
[9:0:0:0] disk Linux scsi_debug 0190 /dev/sdi /dev/sg9 1.07GB
[N:0:1:1] disk WDC WDS250G2B0C-00PXH0__1 /dev/nvme0n1 - 250GB
My script took 17m12 and the highest throughput (on a copy) was 7.5 GB/sec.
Then I reloaded the scsi_debug module, this time with an additional
'host_max_queue=128' parameter. The script run time was 5 seconds shorter
and the maximum throughput was around 7.6 GB/sec. [Average throughput is
around 4 GB/sec.]
For comparison:
# time liburing/examples/io_uring-cp /dev/sdh /dev/sdi
real 0m1.542s
user 0m0.004s
sys 0m1.027s
Umm, that's less then 1 GB/sec. In its defence io_uring-cp is an
extremely simple, single threaded, proof-of-concept copy program,
at least compared to sg_mrq_dd . As used by the sg_mrq_dd the
rewritten sg driver bypasses moving 1 GB to and from the _user_
space while doing the above copy and verify steps.
So:
Tested-by: Douglas Gilbert <dgilbert@xxxxxxxxxxxx>
Differences to v7:
- Add null_blk and scsi_debug support
- Drop debugfs tags patch - it's too difficult to be the same between
hostwide and non-hostwide, as discussed:
https://lore.kernel.org/linux-scsi/1591810159-240929-1-git-send-email-john.garry@xxxxxxxxxx/T/#mb3eb462d8be40273718505989abd12f8228c15fd
And from commit 6bf0eb550452 ("sbitmap: Consider cleared bits in
sbitmap_bitmap_show()"), I guess not many used this anyway...
- Add elevator per-hctx request count for optimisation
- Break up "blk-mq: rename blk_mq_update_tag_set_depth()" into 2x patches
- Pass flags for avoid per-hq queue tags init/free for hostwide tags
- Add Don's reviewed-tag and tested-by tags to appropiate patches
- (@Don, please let me know if issue with how I did this)
- Add "scsi: core: Show nr_hw_queues in sysfs"
- Rework megaraid SAS patch to have module param (Kashyap)
- rebase
V7 is here for more info:
https://lore.kernel.org/linux-scsi/1591810159-240929-1-git-send-email-john.garry@xxxxxxxxxx/T/#t
Hannes Reinecke (5):
blk-mq: Rename blk_mq_update_tag_set_depth()
blk-mq: Free tags in blk_mq_init_tags() upon error
scsi: Add host and host template flag 'host_tagset'
hpsa: enable host_tagset and switch to MQ
smartpqi: enable host tagset
John Garry (10):
blk-mq: Pass flags for tag init/free
blk-mq: Use pointers for blk_mq_tags bitmap tags
blk-mq: Facilitate a shared sbitmap per tagset
blk-mq: Relocate hctx_may_queue()
blk-mq: Record nr_active_requests per queue for when using shared
sbitmap
blk-mq: Record active_queues_shared_sbitmap per tag_set for when using
shared sbitmap
null_blk: Support shared tag bitmap
scsi: core: Show nr_hw_queues in sysfs
scsi: hisi_sas: Switch v3 hw to MQ
scsi: scsi_debug: Support host tagset
Kashyap Desai (2):
blk-mq, elevator: Count requests per hctx to improve performance
scsi: megaraid_sas: Added support for shared host tagset for
cpuhotplug
Ming Lei (1):
blk-mq: Rename BLK_MQ_F_TAG_SHARED as BLK_MQ_F_TAG_QUEUE_SHARED
block/bfq-iosched.c | 9 +-
block/blk-core.c | 2 +
block/blk-mq-debugfs.c | 10 +-
block/blk-mq-sched.c | 13 +-
block/blk-mq-tag.c | 149 ++++++++++++++------
block/blk-mq-tag.h | 56 +++-----
block/blk-mq.c | 81 +++++++----
block/blk-mq.h | 76 +++++++++-
block/kyber-iosched.c | 4 +-
block/mq-deadline.c | 6 +
drivers/block/null_blk_main.c | 6 +
drivers/block/rnbd/rnbd-clt.c | 2 +-
drivers/scsi/hisi_sas/hisi_sas.h | 3 +-
drivers/scsi/hisi_sas/hisi_sas_main.c | 36 ++---
drivers/scsi/hisi_sas/hisi_sas_v3_hw.c | 87 +++++-------
drivers/scsi/hosts.c | 1 +
drivers/scsi/hpsa.c | 44 +-----
drivers/scsi/hpsa.h | 1 -
drivers/scsi/megaraid/megaraid_sas_base.c | 39 +++++
drivers/scsi/megaraid/megaraid_sas_fusion.c | 29 ++--
drivers/scsi/scsi_debug.c | 28 ++--
drivers/scsi/scsi_lib.c | 2 +
drivers/scsi/scsi_sysfs.c | 11 ++
drivers/scsi/smartpqi/smartpqi_init.c | 45 ++++--
include/linux/blk-mq.h | 13 +-
include/linux/blkdev.h | 3 +
include/scsi/scsi_host.h | 9 +-
27 files changed, 484 insertions(+), 281 deletions(-)
