Re: [PATCH] mm/vmscan: make do_shrink_slab more robust.

[Michal Hocko]

On Mon 27-11-17 15:26:54, jiang.biao2@xxxxxxxxxx wrote:
> On Mon, Nov 27, 2017 at 02:27:20PM +0800, jiang.biao2@xxxxxxxxxx wrote:> > I agree with your concern.  How about we take another way by
> > > adding some warning in such case? such as,
> > >         freeable = shrinker->count_objects(shrinker, shrinkctl);
> > > +       if (unlikely(freeable < 0)) {
> > > +               pr_err("shrink_slab: %pF negative objects returned. freeable=%ld\n",
> > > +                       shrinker->scan_objects, freeable);
> > > +               freeable = 0;  //maybe not needed?
> > > +       }
> > >         if (freeable == 0)
> > >                 return 0;
> > > In this way, we would not break the API, but could alert user exception
> > > with message, and make it more robust in such case.
> >
> > True but it would be a problem robust vs. effectivess tradeoff.
> > Think about that everyone want to make thier code robust.
> > It means they start to dump lots of defensive code so code start
> > to look like complicated as well as binary bloating.
> > So, whenever we add some more, we should think how effective
> > the code I am putting?
> > 
> > In this case, I'm skeptical, Sorry. But others might have different
> > opinions. :)
> 
> With all due respect. I still think the robustness is more important than 
> effectiveness in this case. :)

This is a slow path so I wouldn't worry about the performance much. On
the other hand I agree that the API is well documented so adding a
warning is too defensive. We simply assume that the kernel running in
the kernel is reasonable. So I would say, fix your code.

-- 
Michal Hocko
SUSE Labs