Re: [PATCH V5 0/6] blk-mq: improvement CPU hotplug

On Fri, Jan 31, 2020 at 6:24 PM John Garry [off-list ref] wrote:

quoted

quoted

[  141.976109] Call trace:
[  141.978550]  __switch_to+0xbc/0x218
[  141.982029]  blk_mq_run_work_fn+0x1c/0x28
[  141.986027]  process_one_work+0x1e0/0x358
[  141.990025]  worker_thread+0x40/0x488
[  141.993678]  kthread+0x118/0x120
[  141.996897]  ret_from_fork+0x10/0x18

Hi John,

Thanks for your test!

Hi Ming,

quoted

Could you test the following patchset and only the last one is changed?

https://github.com/ming1/linux/commits/my_for_5.6_block

For SCSI testing, I will ask my colleague Xiang Chen to test when he
returns to work. So I did not see this issue for my SCSI testing for
your original v5, but I was only using 1x as opposed to maybe 20x SAS disks.

BTW, did you test NVMe? For some reason I could not trigger a scenario
where we're draining the outstanding requests for a queue which is being
deactivated - I mean, the queues were always already quiesced.

I run cpu hotplug test on both NVMe and SCSI in KVM, and fio just runs
as expected.

NVMe is often 1:1 mapping, so it might be a bit difficult to trigger
draining in-flight IOs.

Thanks,
Ming Lei