On Wed, May 22, 2024 at 09:35:57PM -0700, Yosry Ahmed wrote:
> On Wed, May 22, 2024 at 8:48 PM Shakeel Butt <shakeel.butt@xxxxxxxxx> wrote:
[...]
> >
> > struct mem_cgroup_per_node {
> > - struct lruvec lruvec;
> > + /* Keep the read-only fields at the start */
> > + struct mem_cgroup *memcg; /* Back pointer, we cannot */
> > + /* use container_of */
> >
> > struct lruvec_stats_percpu __percpu *lruvec_stats_percpu;
> > struct lruvec_stats *lruvec_stats;
> > -
> > - unsigned long lru_zone_size[MAX_NR_ZONES][NR_LRU_LISTS];
> > -
> > - struct mem_cgroup_reclaim_iter iter;
> > -
> > struct shrinker_info __rcu *shrinker_info;
> >
> > + /* memcg-v1 only stuff in middle */
> > +
> > struct rb_node tree_node; /* RB tree node */
> > unsigned long usage_in_excess;/* Set to the value by which */
> > /* the soft limit is exceeded*/
> > bool on_tree;
> > - struct mem_cgroup *memcg; /* Back pointer, we cannot */
> > - /* use container_of */
>
> Do we need CACHELINE_PADDING() here (or maybe make struct lruvec
> cache-aligned) to make sure the false cacheline sharing doesn't happen
> again with the fields below, or is the idea that the fields that get
> read in hot paths (memcg, lruvec_stats_percpu, lruvec_stats) are far
> at the top, and the memcg v1 elements in the middle act as a buffer?
>
> IOW, is sharing between the fields below and memcg v1 fields okay
> because they are not read in the hot path? If yes, I believe it's
> worth a comment. It can be easily missed if the memcg v1 soft limit is
> removed later for example.
>
For 6.10, I wanted to keep the change simple and yes, the memcg v1 stuff
as a buffer between the pointers and lruvec/lru_zone_size fields. For
6.11 or later kernels, I am planning to use some asserts to make sure
these fields don't share a cacheline, so later when we remove the
v1-only stuff, the asserts will make sure we keep the separate cacheline
property intact.
