Re: [RFC PATCH 1/2] mm, oom: Introduce bpf_select_task

[Chuyi Zhou]

在 2023/8/4 21:34, Michal Hocko 写道:
> On Fri 04-08-23 21:15:57, Chuyi Zhou wrote:
> [...]
> > > +	switch (bpf_oom_evaluate_task(task, oc, &points)) {
> > > +		case -EOPNOTSUPP: break; /* No BPF policy */
> > > +		case -EBUSY: goto abort; /* abort search process */
> > > +		case 0: goto next; /* ignore process */
> > > +		default: goto select; /* note the task */
> > > +	}
> >
> > Why we need to change the *points* value if we do not care about oom_badness
> > ? Is it used to record some state? If so, we could record it through bpf
> > map.
>
> Strictly speaking we do not need to. That would require BPF to keep the
> state internally. Many will do I suppose but we have to keep track of
> the victim so that the oom killer knows what to kill so I thought that
> it doesn't hurt to keep track of an abstract concept of points as well.
> If you think this is not needed then oc->points could be always 0 for
> bpf selected victims. The value is not used anyway in the proposed
> scheme.
>
> Btw. we will need another hook or metadata for the reporting side of
> things. Generally dump_header() to know what has been the selection
> policy.
OK. Maybe a integer like policy_type is enough to distinguish different 
policies and the default method is zero. Or we can let BPF return a 
string like policy_name.

Which one should I start implementing in next version? Do you have a 
better idea?

Thanks.