lkml.org 
[lkml]   [2021]   [Dec]   [21]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [PATCH RFT] blk-mq: optimize queue tag busy iter for shared_tags
From
Date
Hi Kashyap,

> This is for current/5.17. This patch is meaningfully only on top of [1].
>
> [1] " blk-mq: Use shared tags for shared sbitmap support" Commit -
> e155b0c238b20f0a866f4334d292656665836c8a
>

But your change seems effectively the same as in
https://lore.kernel.org/all/1638794990-137490-4-git-send-email-john.garry@huawei.com/,
which is now merged in Jens' 5.17 queue:

https://git.kernel.org/pub/scm/linux/kernel/git/axboe/linux-block.git/commit/?h=for-5.17/block&id=fea9f92f1748083cb82049ed503be30c3d3a9b69

> While doing additional testing for [1], I noticed some performance issue.
> Along with the performance issue, I noticed CPU lockup as well. Lockup
> trace -
>
> _raw_spin_lock_irqsave+0x42/0x50
> blk_mq_find_and_get_req+0x20/0xa0
> bt_iter+0x2d/0x80
> blk_mq_queue_tag_busy_iter+0x1aa/0x2f0
> ? blk_mq_complete_request+0x30/0x30
> ? blk_mq_complete_request+0x30/0x30
> ? __schedule+0x360/0x850
> blk_mq_timeout_work+0x5e/0x120
> process_one_work+0x1a8/0x380
> worker_thread+0x30/0x380
> ? wq_calc_node_cpumask.isra.30+0x100/0x100
> kthread+0x167/0x190
> ? set_kthread_struct+0x40/0x40
> ret_from_fork+0x22/0x30
>
> It is a generic performance issue if driver use " shost->host_tagset = 1".
> In fact, I found that [1] is useful to fix performance issue and provided
> this additional patch.
>
> I changed my setup to have 64 scsi_devices (earlier I just kept 16 or 24
> drives, so did not noticed this issue). Performance/cpu lockup issue is not
> due to [1].
> More number of scsi device, hardware context per host and high queue depth
> will increase the chances of lockup and performance drop.
>
> Do you think, it is good to have changes in 5.16 + stable ?
> I don't know if this patch will create any side effect. Can you review and
> let me know your feedback. ?
>

Can you test my merged change again for this scenario?

I will also note that I mentioned previously that
blk_mq_queue_tag_busy_iter() was not optimum for shared sbitmap, i.e.
before shared tags, but no one said performance was bad for shared sbitmap.

Thanks,
John

\
 
 \ /
  Last update: 2021-12-21 16:21    [W:0.056 / U:3.100 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site