Re: [RFC PATCH 2/2] lib: debugobjects: touch watchdog to avoid softlockup when !CONFIG_PREEMPT

["Yang Shi" <yang.s@alibaba-inc.com>]

On 11/27/17 10:18 AM, Waiman Long wrote:
> On 11/27/2017 12:54 PM, Yang Shi wrote:
>> Hi Waiman,
>>
>> The second patch of this series.
>>
>> Thanks,
>> Yang
>>
>>
>> On 11/17/17 11:43 AM, Yang Shi wrote:
>>> There are nested loops on debug objects free path, sometimes it may take
>>> over hundred thousands of loops, then cause soft lockup with
>>> !CONFIG_PREEMPT
>>> occasionally, like below:
>>>
>>>   ...
>>>
>>> The code path might be called in either atomic or non-atomic context,
>>> so touching softlockup watchdog instead of calling cond_resched() which
>>> might fall asleep. However, it is unnecessary to touch the watchdog
>>> every loop, so just touch the watchdog at every 10000 (best estimate)
>>> loops.
>>>
>>> Signed-off-by: Yang Shi <yang.s@alibaba-inc.com>
> 
> I do have some concern about suppressing the soft lockup warning
> entirely. If the system feels unresponsive for a certain period of time
> (e.g. 22s), most users would like to know what is going on. It can be a
> custom message with less scary warning. Alternatively, some opt-out

I'm not sure if it is necessary for debug code since the 
unresponsiveness is introduced by debug config and is expected somehow, 
so the user is supposed to know what they are doing, and it sounds 
preferred to disregard the soft lockup message reported by object debug 
for the most time.

We do have some other examples which suppress soft lockup completely in 
kernel, i.e. kdb debug, printing some verbose debug or error message, 
some slow driver code, etc.

> mechanism can be added to explicitly disable soft lookup warning for
> debugobjs is OK as long as it is not the default.
If we really want to some opt-out, we should be able to add a proc knob 
to disable soft lockup as the patch does, but not default. If this is 
too overkilling, we may just add some comment in the Kconfig help text 
to tell users the side effect.

Thanks,
Yang

> 
> Cheers,
> Longman
>