Re: [PATCH] mm: delay oom_reaper only for the process using robust-futex

[zhongjinji <zhongjinji@honor.com>]

Yeah, it is enough time for any process on Android to exit within 2 seconds,
so the oom reaper basically does not work.
Currently, we have not found any users of robust futex on Android.
As we know, adding killed processes to the oom_reaper queue can help
free their memory faster.
Delaying the oom reaper for all processes might lead to less efficient
memory reclamation, which could be undesirable.
>
>> 
>> But I noticed that many processes do not use robust-futex, so they do not
>> access user-space memory during do_exit and do not run into the problem
>> mentioned in that patch.
>> 
>> So, this change delays the oom_reaper only when the process uses
>> robust-futex, letting the oom_reaper work properly in more cases.
>
>The direction seems reasonable.
>
>>  
>> +static inline bool exit_may_access_user(struct task_struct *task)
>> +{
>> +	return task->robust_list || task->compat_robust_list;
>
>Here I am not sure. This robust_list seems like a per-task list and we
>are making a process level decision based on a given task's usage of
>robust list. Can we have a scenario where some tasks/threads of a
>process does not have robust list and others have? If yes this can cause
>similar similar which the original patch tried to solve, right?

Yes, we definitely need to check whether all threads have a robust list.
Thanks for the reminder.
I will include this fix in the next version.