Bisect shows it was introduced from the below commit:

commit 2637baed78010eeaae274feb5b99ce90933fadfb
Author: Minwoo Im [off-list ref]
Date:   Wed Apr 21 16:45:04 2021 +0900

    nvme: introduce generic per-namespace chardev


On Thu, Sep 30, 2021 at 6:36 PM Yi Zhang [off-list ref] wrote:

On Thu, Sep 30, 2021 at 3:55 PM Sagi Grimberg [off-list ref] wrote:

quoted

quoted

quoted

quoted

Hello

Below kmemleak was triggered with blktests nvme-tcp on latest
5.15.0-rc3, pls check it.

Please share the test number and the frequency to reproduce this...

Hi
I'm running the full blktests nvme-tcp[1] and it's 100% reproduced.

[1]
# nvme_trtype=tcp ./check nvme/

Yi, this does not happen with nvme_trtype=rdma? It looks like

nvme_trtype=rdma use_siw=1 also can reproduce it.

quoted

we don't get to call cdev_device_del and del_gendisk, which means
we may have a referencing problem...

I'm wandering if this is a regression we can bisect to?

So just run[1] with nvme_core: multipath=Y will trigger it.
[1]
nvme_trtype=tcp ./check nvme/004

Will try bisect it.

-- 
Best Regards,
  Yi Zhang


_______________________________________________
Linux-nvme mailing list
Linux-nvme@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-nvme