Re: WARN @lib/refcount.c:128 during hot unplug of I/O adapter.

[Tyrel Datwyler]

On 04/11/2017 02:00 AM, Michael Ellerman wrote:
> Tyrel Datwyler <tyreld@xxxxxxxxxxxxxxxxxx> writes:
> 
>> On 04/06/2017 09:04 PM, Michael Ellerman wrote:
>>> Tyrel Datwyler <tyreld@xxxxxxxxxxxxxxxxxx> writes:
>>>
>>>> On 04/06/2017 03:27 AM, Sachin Sant wrote:
>>>>> On a POWER8 LPAR running 4.11.0-rc5, a hot unplug operation on
>>>>> any I/O adapter results in the following warning
>>>>>
>>>>> This problem has been in the code for some time now. I had first seen this in
>>>>> -next tree.
>>>>>
>>
>> <snip>
>>
>>>>> Have attached the dmesg log from the system. Let me know if any additional
>>>>> information is required to help debug this problem.
>>>>
>>>> I remember you mentioning this when the issue was brought up for CPUs. I
>>>> assume the case is the same here where the issue is only seen with
>>>> adapters that were hot-added after boot (ie. hot-remove of adapter
>>>> present at boot doesn't trip the warning)?
>>>
>>> So who's fixing this?
>>
>> I started looking at it when Bharata submitted a patch trying to fix the
>> issue for CPUs, but got side tracked by other things. I suspect that
>> this underflow has actually been an issue for quite some time, and we
>> are just now becoming aware of it thanks to the recount_t patchset being
>> merged.
> 
> Yes I agree. Which means it might be broken in existing distros.

Definitely. I did some profiling last night, and I understand the
hotplug case. It turns out to be as I suggested in the original thread
about CPUs. When the devicetree code was worked to move the tree out of
proc and into sysfs the sysfs detach code added a of_node_put to remove
the original of_init reference. pSeries Being the sole original
*dynamic* device tree user we had always issued a of_node_put in our
dlpar specific detach function to achieve that end. So, this should be a
pretty straight forward trivial fix.

However, for the case where devices are present at boot it appears we a
leaking a lot of references resulting in the device nodes never actually
being released/freed after a dlpar remove. In the CPU case after boot I
count 8 more references taken than the hotplug case, and corresponding
of_node_put's are not called at dlpar remove time either. That will take
some time to track them down, review and clean up.

-Tyrel

> 
>> I'll look into it again this week.
> 
> Thanks.
> 
> cheers
>