On Thu 19-03-20 10:36:06, Matthew Wilcox wrote:
> On Thu, Mar 19, 2020 at 04:20:25PM +0000, Chris Down wrote:
[...]
> > > -static bool cgroup_reclaim(struct scan_control *sc)
> > > +static struct mem_cgroup *cgroup_reclaim(struct scan_control *sc)
> > > {
> > > - return false;
> > > + return NULL;
> > > }
>
> I think this is actually the important bit. For those who build
> their kernels with cgroups disabled, it will save a small number of
> instructions since cgroup_reclaim() will be NULL rather than dereferencing
> sc->target_mem_group. It'd be nice to have that saving quantified as
> part of the changelog.
I gave it a try and you are right that !MEMCG is slightly better. But
MEMCG=y which is a more common configuration I would say is worse
text data bss dec hex filename
40661 24060 12 64733 fcdd mm/vmscan.memcg.after.o
40556 24060 12 64628 fc74 mm/vmscan.memcg.before.o
36240 23076 8 59324 e7bc mm/vmscan.nomemcg.after.o
36283 23076 8 59367 e7e7 mm/vmscan.nomemcg.before.o
This is with gcc version 9.2.1 20200306 (Debian 9.2.1-31).
--
Michal Hocko
SUSE Labs
