Re: [PATCH v2] memcg: fix soft lockup in the OOM process

[Chen Ridong]

On 2024/12/21 15:28, Michal Hocko wrote:
> On Fri 20-12-24 14:47:34, Andrew Morton wrote:
>> On Fri, 20 Dec 2024 10:31:23 +0000 Chen Ridong <chenridong@xxxxxxxxxxxxxxx> wrote:
>>
>>> From: Chen Ridong <chenridong@xxxxxxxxxx>
>>>
>>> A soft lockup issue was found in the product with about 56,000 tasks were
>>> in the OOM cgroup, it was traversing them when the soft lockup was
>>> triggered.
>>>
>>> ...
>>>
>>> This is because thousands of processes are in the OOM cgroup, it takes a
>>> long time to traverse all of them. As a result, this lead to soft lockup
>>> in the OOM process.
>>>
>>> To fix this issue, call 'cond_resched' in the 'mem_cgroup_scan_tasks'
>>> function per 1000 iterations. For global OOM, call
>>> 'touch_softlockup_watchdog' per 1000 iterations to avoid this issue.
>>>
>>> ...
>>>
>>> --- a/include/linux/oom.h
>>> +++ b/include/linux/oom.h
>>> @@ -14,6 +14,13 @@ struct notifier_block;
>>>  struct mem_cgroup;
>>>  struct task_struct;
>>>  
>>> +/* When it traverses for long time,  to prevent softlockup, call
>>> + * cond_resched/touch_softlockup_watchdog very 1000 iterations.
>>> + * The 1000 value  is not exactly right, it's used to mitigate the overhead
>>> + * of cond_resched/touch_softlockup_watchdog.
>>> + */
>>> +#define SOFTLOCKUP_PREVENTION_LIMIT 1000
>>
>> If this is to have potentially kernel-wide scope, its name should
>> identify which subsystem it belongs to.  Maybe OOM_KILL_RESCHED or
>> something.
>>
>> But I'm not sure that this really needs to exist.  Are the two usage
>> sites particularly related?
> 
> Yes, I do not think this needs to pretend to be a more generic mechanism
> to prevent soft lockups. The number of iterations highly depends on the
> operation itself.
> 

Thanks， I will update.

>>
>>>  enum oom_constraint {
>>>  	CONSTRAINT_NONE,
>>>  	CONSTRAINT_CPUSET,
>>> diff --git a/mm/memcontrol.c b/mm/memcontrol.c
>>> index 5c373d275e7a..f4c12d6e7b37 100644
>>> --- a/mm/memcontrol.c
>>> +++ b/mm/memcontrol.c
>>> @@ -1161,6 +1161,7 @@ void mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
>>>  {
>>>  	struct mem_cgroup *iter;
>>>  	int ret = 0;
>>> +	int i = 0;
>>>  
>>>  	BUG_ON(mem_cgroup_is_root(memcg));
>>>  
>>> @@ -1169,8 +1170,11 @@ void mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
>>>  		struct task_struct *task;
>>>  
>>>  		css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it);
>>> -		while (!ret && (task = css_task_iter_next(&it)))
>>> +		while (!ret && (task = css_task_iter_next(&it))) {
>>>  			ret = fn(task, arg);
>>> +			if (++i % SOFTLOCKUP_PREVENTION_LIMIT)
>>
>> And a modulus operation is somewhat expensive.
> 
> This is a cold path used during OOM. While we can make it more optimal I
> doubt it matters in practice so we should aim at readbility. I do not
> mind either way, I just wanted to note that this is not performance
> sensitive.
> 

I think '(++i & 1023)' is much better, I will update.
Thank you all gays.

Best regards
Ridong

>>
>> Perhaps a simple
>>
>> 		/* Avoid potential softlockup warning */
>> 		if ((++i & 1023) == 0)
>>
>> at both sites will suffice.  Opinions might vary...
>>
>