Re: [PATCH] mm: don't be stuck to rmap lock on reclaim path

[Matthew Wilcox]

On Wed, May 04, 2022 at 08:52:13AM -0700, Minchan Kim wrote:
> On Wed, May 04, 2022 at 07:09:37AM +0100, Matthew Wilcox wrote:
> > On Tue, May 03, 2022 at 09:30:38PM -0700, Minchan Kim wrote:
> > > On Wed, May 04, 2022 at 04:32:13AM +0100, Matthew Wilcox wrote:
> > > > On Tue, May 03, 2022 at 10:03:41AM -0700, Minchan Kim wrote:
> > > > > -void rmap_walk(struct folio *folio, const struct rmap_walk_control *rwc);
> > > > > -void rmap_walk_locked(struct folio *folio, const struct rmap_walk_control *rwc);
> > > > > +void rmap_walk(struct folio *folio, struct rmap_walk_control *rwc);
> > > > > +void rmap_walk_locked(struct folio *folio, struct rmap_walk_control *rwc);
> > > > 
> > > > I see the build bot already beat me to pointing out why this is wrong,
> > > > but do you not look at git log to figure out why code was changed to be
> > > > the way it is now, before you change it back?
> > > 
> > > This patch added a new field as out param like compact_control so
> > > the rmap_walk_control is not immutable.
> > 
> > ... but we have a user which treats it as if it is.
> 
> True. I don't think it will show sizable benefit on runtime overhead
> since rmap_walk is already one of the most expensive operation in MM.
> 
> I could reintroduce the typecast for page_idle_clear_pte_refs to remove
> the const as we had several years.
> 
> If your concern was to make rmap_walk_control mutable back, I could
> change rmap_walk function having return value or adding a addtional
> new out param. However, I thought rmap_walk_control is more readable/
> easier than them.

I haven't thought deeply about it, but I suspect the right approach is
to remove the rather dubious optimisation in page_idle_clear_pte_refs().