> On Jan 17, 2022, at 10:50 AM, Bruce Fields <bfields@xxxxxxxxxxxx> wrote:
>
> On Sat, Jan 15, 2022 at 07:46:06PM +0000, Chuck Lever III wrote:
>>
>>> On Jan 15, 2022, at 3:14 AM, Jonathan Woithe <jwoithe@xxxxxxxxxx> wrote:
>>>
>>> Hi Chuck
>>>
>>> Thanks for your response.
>>>
>>> On Fri, Jan 14, 2022 at 03:18:01PM +0000, Chuck Lever III wrote:
>>>>> Recently we migrated an NFS server from a 32-bit environment running
>>>>> kernel 4.14.128 to a 64-bit 5.15.x kernel. The NFS configuration remained
>>>>> unchanged between the two systems.
>>>>>
>>>>> On two separate occasions since the upgrade (5 Jan under 5.15.10, 14 Jan
>>>>> under 5.15.12) the kernel has oopsed at around the time that an NFS client
>>>>> machine is turned on for the day. On both occasions the call trace was
>>>>> essentially identical. The full oops sequence is at the end of this email.
>>>>> The oops was not observed when running the 4.14.128 kernel.
>>>>>
>>>>> Is there anything more I can provide to help track down the cause of the
>>>>> oops?
>>>>
>>>> A possible culprit is 7f024fcd5c97 ("Keep read and write fds with each
>>>> nlm_file"), which was introduced in or around v5.15. You could try a
>>>> simple test and back the server down to v5.14.y to see if the problem
>>>> persists.
>>>
>>> I could do this, but only perhaps on Monday when I'm next on site. It may
>>> take a while to get an answer though, since it seems we hit the fault only
>>> around once every 2 weeks. Since it's a production server we are of course
>>> limited in the things I can do.
>>>
>>> I *may* be able to set up another system as an NFS server and hit that with
>>> repeated mount requests. That could help reduce the time we have to wait
>>> for an answer.
>>
>> Given the callback information you provided, I believe that the problem
>> is due to a client reboot, not a mount request. The callback shows the
>> crash occurs while your server is processing an SM_NOTIFY request from
>> one of your clients.
>>
>>
>>> Is it worth considering a revert of 7f024fcd5c97? I guess it depends on how
>>> many later patches depended on it.
>>
>> You can try reverting 7f024fcd5c97, but as I recall there are some
>> subsequent changes that depend on that one.
>
> NLM locking on reexports would stop working. Which is a new (and
> imperfect) feature, so less important than avoiding this NULL
> dereference, if push came to shove. But, let's see if we can just fix
> it.....
Agreed. I was suggested reverting only as an experiment.
--
Chuck Lever
