Re: [PATCH v5 2/2] mm: add swapiness= arg to memory.reclaim

[Yu Zhao <yuzhao@google.com>]

On Tue, Jan 2, 2024 at 8:21 AM Dan Schatzberg <schatzberg.dan@gmail.com> wrote:
>
> Hi Yu Zhao,
>
> Thanks for the feedback, sorry for the delayed response.
>
> On Thu, Dec 21, 2023 at 10:31:59PM -0700, Yu Zhao wrote:
> > On Wed, Dec 20, 2023 at 8:27 AM Dan Schatzberg <schatzberg.dan@gmail.com> wrote:
> > >
> > > ...
> >
> > The cover letter says:
> > "Previously, this exact interface addition was proposed by Yosry[3]."
> >
> > So I think it should be acknowledged with a Suggested-by, based on:
> > "A Suggested-by: tag indicates that the patch idea is suggested by the
> > person named and ensures credit to the person for the idea."
> > from
> > https://docs.kernel.org/process/submitting-patches.html#using-reported-by-tested-by-reviewed-by-suggested-by-and-fixes
>
> Sure, will do.
>
> > > diff --git a/mm/vmscan.c b/mm/vmscan.c
> > > index d91963e2d47f..aa5666842c49 100644
> > > --- a/mm/vmscan.c
> > > +++ b/mm/vmscan.c
> > > @@ -92,6 +92,9 @@ struct scan_control {
> > >         unsigned long   anon_cost;
> > >         unsigned long   file_cost;
> > >
> > > +       /* Swappiness value for reclaim. NULL will fall back to per-memcg/global value */
> > > +       int *swappiness;
> >
> > Using a pointer to indicate whether the type it points to is
> > overridden isn't really a good practice.
> >
> > A better alternative was suggested during the v2:
> > "Perhaps the negative to avoid unnecessary dereferences."
> > https://lore.kernel.org/linux-mm/dhhjw4h22q4ngwtxmhuyifv32zjd6z2relrcjgnxsw6zys3mod@o6dh5dy53ae3/
>
> I did have a couple versions with a negative but it creates
> initialization issues where every instantiation of scan_control needs
> to make sure to initialize swappiness or else it will behave as if
> swappiness is 0. That's pretty error prone so using the pointer seemed
> the better approach.
>
> > Since only proactive reclaim can override swappiness, meaning it only
> > happens if sc->proactive is true, I think the best way to make it work
> > without spending much effort is create a helper as Michal suggest but
> > it should look like:
> >
> > sc_swappiness()
> > {
> >   return sc->proactive ? sc->swappiness : mem_cgroup_swappiness(memcg);
> > }
> >
> > In this patchset, sc->swappiness really means
> > sc->proactive_swappiness. So it should be renamed accordingly.
>
> Helper aside, I disagree with this point about coupling with the
> proactive flag.

Sure. But I would like to hear a *concrete* counterexample.

> The fact that the only user currently is proactive
> reclaim

Yes, that's a fact, and we should make the decision based on the
current known facts.

> doesn't imply to me that the interface (in scan_control)
> should be coupled to the use-case.

Future always has its uncertainty which I would not worry so much about.

> It's easier to reason about a
> swappiness field that overrides swappiness for all scans that set it
> regardless of the users.

For example? And how likely would that happen in the next few years?