On Tue, Jul 20, 2021 at 11:05:29AM +0200, Oleksandr Natalenko wrote:

Hello, Ming.

On pondělí 19. července 2021 8:27:29 CEST Oleksandr Natalenko wrote:

quoted

On pondělí 19. července 2021 3:40:40 CEST Ming Lei wrote:

quoted

On Sat, Jul 17, 2021 at 02:35:14PM +0200, Oleksandr Natalenko wrote:

quoted

On sobota 17. července 2021 14:19:59 CEST Oleksandr Natalenko wrote:

quoted

On sobota 17. července 2021 14:11:05 CEST Oleksandr Natalenko wrote:

quoted

On sobota 17. července 2021 11:35:32 CEST Ming Lei wrote:

quoted

Maybe you need to check if the build is OK, I can't reproduce it
in
my
VM, and BFQ is still builtin:

[root@ktest-01 ~]# uname -a
Linux ktest-01 5.14.0-rc1+ #52 SMP Fri Jul 16 18:56:36 CST 2021
x86_64
x86_64 x86_64 GNU/Linux [root@ktest-01 ~]# cat
/sys/block/nvme0n1/queue/scheduler
[none] mq-deadline kyber bfq

I don't think this is an issue with the build… BTW, with
`initcall_debug`:

[    0.902555] calling  bfq_init+0x0/0x8b @ 1
[    0.903448] initcall bfq_init+0x0/0x8b returned -28 after 507
usecs

-ENOSPC? Why? Also re-tested with the latest git tip, same result
:(.

OK, one extra pr_info, and I see this:

[    0.871180] blkcg_policy_register: BLKCG_MAX_POLS too small
[    0.871612] blkcg_policy_register: -28

What does it mean please :)? The value seems to be hard-coded:

include/linux/blkdev.h
60:#define BLKCG_MAX_POLS               5

OK, after increasing this to 6 I've got my BFQ back. Please see [1].

[1]
https://lore.kernel.org/linux-block/20210717123328.945810-1-oleksandr@na
t
alenko.name/

OK, after you fixed the issue in blkcg_policy_register(), can you
reproduce the discard issue on v5.14-rc1 with BFQ applied? If yes,
can you test the patch I posted previously?

Yes, the issue is reproducible with both v5.13.2 and v5.14-rc1. I haven't
managed to reproduce it with v5.13.2+your patch. Now I will build v5.14-
rc2+your patch and test further.

I'm still hammering v5.14-rc2 + your patch, and I cannot reproduce the issue. 
Given I do not have a reliable reproducer (I'm just firing up the kernel build, 
and the issue pops up eventually, sooner or later, but usually within a couple 
of first tries), for how long I should hammer it for your fix to be considered 
proven?

You mentioned that the issue is reproducible with v5.14-rc, that means
it can be always reproduced in limited time(suppose it is A). If the issue
can't be reproduced any more after applying the patch in long enough time B(B >> A),
we can think it is fixed by the patch.

For example, if A is one hour, we can set B as 5*A or bigger to simulate
the long enough time.


Thanks,
Ming


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