On Thu 19-12-24 09:27:52, Chen Ridong wrote:
>
>
> On 2024/12/18 18:22, Michal Hocko wrote:
> > On Wed 18-12-24 17:00:38, Chen Ridong wrote:
> >>
> >>
> >> On 2024/12/18 15:56, Michal Hocko wrote:
> >>> On Wed 18-12-24 15:44:34, Chen Ridong wrote:
> >>>>
> >>>>
> >>>> On 2024/12/17 20:54, Michal Hocko wrote:
> >>>>> On Tue 17-12-24 12:18:28, Chen Ridong wrote:
> >>>>> [...]
> >>>>>> diff --git a/mm/oom_kill.c b/mm/oom_kill.c
> >>>>>> index 1c485beb0b93..14260381cccc 100644
> >>>>>> --- a/mm/oom_kill.c
> >>>>>> +++ b/mm/oom_kill.c
> >>>>>> @@ -390,6 +390,7 @@ static int dump_task(struct task_struct *p, void *arg)
> >>>>>> if (!is_memcg_oom(oc) && !oom_cpuset_eligible(p, oc))
> >>>>>> return 0;
> >>>>>>
> >>>>>> + cond_resched();
> >>>>>> task = find_lock_task_mm(p);
> >>>>>> if (!task) {
> >>>>>> /*
> >>>>>
> >>>>> This is called from RCU read lock for the global OOM killer path and I
> >>>>> do not think you can schedule there. I do not remember specifics of task
> >>>>> traversal for crgoup path but I guess that you might need to silence the
> >>>>> soft lockup detector instead or come up with a different iteration
> >>>>> scheme.
> >>>>
> >>>> Thank you, Michal.
> >>>>
> >>>> I made a mistake. I added cond_resched in the mem_cgroup_scan_tasks
> >>>> function below the fn, but after reconsideration, it may cause
> >>>> unnecessary scheduling for other callers of mem_cgroup_scan_tasks.
> >>>> Therefore, I moved it into the dump_task function. However, I missed the
> >>>> RCU lock from the global OOM.
> >>>>
> >>>> I think we can use touch_nmi_watchdog in place of cond_resched, which
> >>>> can silence the soft lockup detector. Do you think that is acceptable?
> >>>
> >>> It is certainly a way to go. Not the best one at that though. Maybe we
> >>> need different solution for the global and for the memcg OOMs. During
> >>> the global OOM we rarely care about latency as the whole system is
> >>> likely to struggle. Memcg ooms are much more likely. Having that many
> >>> tasks in a memcg certainly requires a further partitioning so if
> >>> configured properly the OOM latency shouldn't be visible much. But I am
> >>> wondering whether the cgroup task iteration could use cond_resched while
> >>> the global one would touch_nmi_watchdog for every N iterations. I might
> >>> be missing something but I do not see any locking required outside of
> >>> css_task_iter_*.
> >>
> >> Do you mean like that:
> >
> > I've had something like this (untested) in mind
> > diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> > index 7b3503d12aaf..37abc94abd2e 100644
> > --- a/mm/memcontrol.c
> > +++ b/mm/memcontrol.c
> > @@ -1167,10 +1167,14 @@ void mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
> > for_each_mem_cgroup_tree(iter, memcg) {
> > struct css_task_iter it;
> > struct task_struct *task;
> > + unsigned int i = 0
> >
> > 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 % 1000)
> > + cond_resched();
> > + }
> > css_task_iter_end(&it);
> > if (ret) {
> > mem_cgroup_iter_break(memcg, iter);
>
> Thank you for your patience.
>
> I had this idea in mind as well.
> However, there are two considerations that led me to reconsider it:
>
> 1. I wasn't convinced about how we should call cond_resched every N
> iterations. Should it be 1000 or 10000?
Sure, there will likely not be any _right_ value. This is mostly to
mitigate the overhead of cond_resched which is not completely free.
Having a system with 1000 tasks is not completely uncommon and we do not
really need cond_resched now.
More importantly we can expect cond_resched will eventually go away with
the PREEMPT_LAZY (or what is the current name of that) so I wouldn't
overthink this.
> 2. I don't think all callers of mem_cgroup_scan_tasks need cond_resched.
> Only fn is expensive (e.g., dump_tasks), and it needs cond_resched. At
> least, I have not encountered any other issue except except when fn is
> dump_tasks.
See above. I wouldn't really overthink this.
--
Michal Hocko
SUSE Labs
