On Fri, Feb 2, 2024 at 8:13 PM Shakeel Butt <shakeelb@xxxxxxxxxx> wrote:
>
> On Fri, Feb 2, 2024 at 4:34 PM Yosry Ahmed <yosryahmed@xxxxxxxxxx> wrote:
> >
> > Commit da10d7e140196 ("mm: memcg: optimize parent iteration in
> > memcg_rstat_updated()") added two additional pointers to the end of
> > struct memcg_vmstats_percpu with CACHELINE_PADDING to put them in a
> > separate cacheline. This caused the struct size to increase from 1200 to
> > 1280 on my config (80 extra bytes instead of 16).
> >
> > Upon revisiting, the relevant struct members do not need to be on a
> > separate cacheline, they just need to fit in a single one. This is a
> > percpu struct, so there shouldn't be any contention on that cacheline
> > anyway. Move the members to the beginning of the struct and cachealign
> > the first member. Add a comment about the members that need to fit
> > together in a cacheline.
> >
> > The struct size is now 1216 on my config with this change.
> >
> > Fixes: da10d7e140196 ("mm: memcg: optimize parent iteration in memcg_rstat_updated()")
> > Reported-by: Greg Thelen <gthelen@xxxxxxxxxx>
> > Signed-off-by: Yosry Ahmed <yosryahmed@xxxxxxxxxx>
> > ---
> > mm/memcontrol.c | 19 +++++++++----------
> > 1 file changed, 9 insertions(+), 10 deletions(-)
> >
> > diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> > index d9ca0fdbe4ab0..09f09f37e397e 100644
> > --- a/mm/memcontrol.c
> > +++ b/mm/memcontrol.c
> > @@ -621,6 +621,15 @@ static inline int memcg_events_index(enum vm_event_item idx)
> > }
> >
> > struct memcg_vmstats_percpu {
> > + /* Stats updates since the last flush */
> > + unsigned int stats_updates ____cacheline_aligned;
>
> Why do you need ____cacheline_aligned here? From what I understand for
> the previous patch you want stats_updates, parent and vmstats on the
> same cacheline, right?
Yes. I am trying to ensure that stats_updates sits at the beginning of
a cacheline to ensure they all fit in one cacheline. Is this
implicitly guaranteed somehow?
>
> I would say just remove the CACHELINE_PADDING() from the previous
> patch and we are good.
IIUC, without CACHELINE_PADDING(), they may end up on different cache
lines, depending on the size of the arrays before them in the struct
(which depends on several configs). Am I misunderstanding?
>
> In the followup I plan to add usage of __cacheline_group_begin() and
> __cacheline_group_end() usage in memcg code. If you want, take a stab
> at it.
For now, I am just looking for something simple to fix the struct size
proliferation for v6.8, but this would be interesting to see. I wonder
how __cacheline_group_end() works since the end is decided already by
__cacheline_group_begin() and the cacheline size.
