Re: [PATCH 4/5] oom: don't kill random process

[KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>]

(2011/05/24 10:58), David Rientjes wrote:
> On Tue, 24 May 2011, KOSAKI Motohiro wrote:
>
>>>>> This is unnecessary and just makes the oom killer egregiously long.  We
>>>>> are already diagnosing problems here at Google where the oom killer
>>>>> holds
>>>>> tasklist_lock on the readside for far too long, causing other cpus
>>>>> waiting
>>>>> for a write_lock_irq(&tasklist_lock) to encounter issues when irqs are
>>>>> disabled and it is spinning.  A second tasklist scan is simply a
>>>>> non-starter.
>>>>>
>>>>>     [ This is also one of the reasons why we needed to introduce
>>>>>       mm->oom_disable_count to prevent a second, expensive tasklist scan.
>>>>> ]
>>>>
>>>> You misunderstand the code. Both select_bad_process() and
>>>> oom_kill_process()
>>>> are under tasklist_lock(). IOW, no change lock holding time.
>>>>
>>>
>>> A second iteration through the tasklist in select_bad_process() will
>>> extend the time that tasklist_lock is held, which is what your patch does.
>>
>> It never happen usual case. Plz think when happen all process score = 1.
>>
>
> I don't care if it happens in the usual case or extremely rare case.  It
> significantly increases the amount of time that tasklist_lock is held
> which causes writelock starvation on other cpus and causes issues,
> especially if the cpu being starved is updating the timer because it has
> irqs disabled, i.e. write_lock_irq(&tasklist_lock) usually in the clone or
> exit path.  We can do better than that, and that's why I proposed my patch
> to CAI that increases the resolution of the scoring and makes the root
> process bonus proportional to the amount of used memory.

Do I need to say the same word? Please read the code at first.