[<prev] [next>] [thread-next>] [day] [month] [year] [list]
Message-ID: <1635852455-39935-1-git-send-email-john.garry@huawei.com>
Date: Tue, 2 Nov 2021 19:27:32 +0800
From: John Garry <john.garry@...wei.com>
To: <axboe@...nel.dk>
CC: <linux-block@...r.kernel.org>, <linux-kernel@...r.kernel.org>,
<ming.lei@...hat.com>, <kashyap.desai@...adcom.com>,
<hare@...e.de>, "John Garry" <john.garry@...wei.com>
Subject: [PATCH RFT 0/3] blk-mq: Optimise blk_mq_queue_tag_busy_iter() for shared tags
In [0], Kashyap reports high CPU usage for blk_mq_queue_tag_busy_iter()
and callees for shared tags.
Indeed blk_mq_queue_tag_busy_iter() would be less optimum for moving to
shared tags, but it was not optimum previously.
So I would like this series tested, and also to know what is triggering
blk_mq_queue_tag_busy_iter() from userspace to cause such high CPU
loading.
As suggested by Ming, reading /proc/diskstats in a while true loop
can trigger blk_mq_queue_tag_busy_iter(); I do so in a test with 2x
separate consoles, and here are the results:
v5.15
blk_mq_queue_tag_busy_iter() 6.2%
part_stat_read_all() 6.7
pre-v5.16 (Linus' master branch @ commit bfc484fe6abb)
blk_mq_queue_tag_busy_iter() 4.5%
part_stat_read_all() 6.2
pre-v5.16+this series
blk_mq_queue_tag_busy_iter() not shown in top users
part_stat_read_all() 7.5%
These results are from perf top, on a system with 7x
disks, with hisi_sas which has 16x HW queues.
[0] https://lore.kernel.org/linux-block/e4e92abbe9d52bcba6b8cc6c91c442cc@mail.gmail.com/
John Garry (3):
blk-mq: Drop busy_iter_fn blk_mq_hw_ctx argument
blk-mq: Delete busy_iter_fn
blk-mq: Optimise blk_mq_queue_tag_busy_iter() for shared tags
block/blk-mq-tag.c | 58 +++++++++++++++++++++++++++---------------
block/blk-mq-tag.h | 2 +-
block/blk-mq.c | 17 ++++++-------
include/linux/blk-mq.h | 2 --
4 files changed, 47 insertions(+), 32 deletions(-)
--
2.17.1
Powered by blists - more mailing lists