Re: [RFC PATCH v2 1/5] mm, oom: Introduce bpf_oom_evaluate_task

[Chuyi Zhou]

Hello,
在 2023/8/17 11:22, Alexei Starovoitov 写道:
> On Wed, Aug 16, 2023 at 7:51 PM Chuyi Zhou <zhouchuyi@xxxxxxxxxxxxx> wrote:
> > Hello,
> >
> > 在 2023/8/17 10:07, Alexei Starovoitov 写道:
> > > On Thu, Aug 10, 2023 at 1:13 AM Chuyi Zhou <zhouchuyi@xxxxxxxxxxxxx> wrote:
> > > >    static int oom_evaluate_task(struct task_struct *task, void *arg)
> > > >    {
> > > >           struct oom_control *oc = arg;
> > > > @@ -317,6 +339,26 @@ static int oom_evaluate_task(struct task_struct *task, void *arg)
> > > >           if (!is_memcg_oom(oc) && !oom_cpuset_eligible(task, oc))
> > > >                   goto next;
> > > >
> > > > +       /*
> > > > +        * If task is allocating a lot of memory and has been marked to be
> > > > +        * killed first if it triggers an oom, then select it.
> > > > +        */
> > > > +       if (oom_task_origin(task)) {
> > > > +               points = LONG_MAX;
> > > > +               goto select;
> > > > +       }
> > > > +
> > > > +       switch (bpf_oom_evaluate_task(task, oc)) {
> > > > +       case BPF_EVAL_ABORT:
> > > > +               goto abort; /* abort search process */
> > > > +       case BPF_EVAL_NEXT:
> > > > +               goto next; /* ignore the task */
> > > > +       case BPF_EVAL_SELECT:
> > > > +               goto select; /* select the task */
> > > > +       default:
> > > > +               break; /* No BPF policy */
> > > > +       }
> > > > +
> > >
> > > I think forcing bpf prog to look at every task is going to be limiting
> > > long term.
> > > It's more flexible to invoke bpf prog from out_of_memory()
> > > and if it doesn't choose a task then fallback to select_bad_process().
> > > I believe that's what Roman was proposing.
> > > bpf can choose to iterate memcg or it might have some side knowledge
> > > that there are processes that can be set as oc->chosen right away,
> > > so it can skip the iteration.
> >
> > IIUC, We may need some new bpf features if we want to iterating
> > tasks/memcg in BPF, sush as:
> > bpf_for_each_task
> > bpf_for_each_memcg
> > bpf_for_each_task_in_memcg
> > ...
> >
> > It seems we have some work to do first in the BPF side.
> > Will these iterating features be useful in other BPF scenario except OOM
> > Policy?
>
> Yes.
> Use open coded iterators though.
> Like example in
> https://lore.kernel.org/all/20230810183513.684836-4-davemarchevsky@xxxxxx/
>
> bpf_for_each(task_vma, vma, task, 0) { ... }
> will safely iterate vma-s of the task.
> Similarly struct css_task_iter can be hidden inside bpf open coded iterator.
OK. I think the following APIs whould be useful and I am willing to 
start with these in another bpf-next RFC patchset:

1. bpf_for_each(task). Just like for_each_process(p) in kernel to 
itearing all tasks in the system with rcu_read_lock().

2. bpf_for_each(css_task, task, css). It works like 
css_task_iter_{start, next, end} and would be used to iterating 
tasks/threads under a css.

3. bpf_for_each(descendant_css, css, root_css, {PRE, POST}). It works 
like css_next_descendant_{pre, post} to iterating all descendant.

If you have better ideas or any advice, please let me know.
Thanks.