On Fri, Oct 25, 2019 at 03:41:18PM -0400, Johannes Weiner wrote:
> On Thu, Oct 17, 2019 at 05:28:13PM -0700, Roman Gushchin wrote:
> > +static inline struct kmem_cache *memcg_slab_pre_alloc_hook(struct kmem_cache *s,
> > + struct mem_cgroup **memcgp,
> > + size_t size, gfp_t flags)
> > +{
> > + struct kmem_cache *cachep;
> > +
> > + cachep = memcg_kmem_get_cache(s, memcgp);
> > + if (is_root_cache(cachep))
> > + return s;
> > +
> > + if (__memcg_kmem_charge_subpage(*memcgp, size * s->size, flags)) {
> > + mem_cgroup_put(*memcgp);
> > + memcg_kmem_put_cache(cachep);
> > + cachep = NULL;
> > + }
> > +
> > + return cachep;
> > +}
> > +
> > static inline void memcg_slab_post_alloc_hook(struct kmem_cache *s,
> > struct mem_cgroup *memcg,
> > size_t size, void **p)
> > {
> > struct mem_cgroup_ptr *memcg_ptr;
> > + struct lruvec *lruvec;
> > struct page *page;
> > unsigned long off;
> > size_t i;
> > @@ -439,6 +393,11 @@ static inline void memcg_slab_post_alloc_hook(struct kmem_cache *s,
> > off = obj_to_index(s, page, p[i]);
> > mem_cgroup_ptr_get(memcg_ptr);
> > page->mem_cgroup_vec[off] = memcg_ptr;
> > + lruvec = mem_cgroup_lruvec(page_pgdat(page), memcg);
> > + mod_lruvec_memcg_state(lruvec, cache_vmstat_idx(s),
> > + s->size);
> > + } else {
> > + __memcg_kmem_uncharge_subpage(memcg, s->size);
> > }
> > }
> > mem_cgroup_ptr_put(memcg_ptr);
>
> The memcg_ptr as a collection vessel for object references makes a lot
> of sense. But this code showcases that it should be a first-class
> memory tracking API that the allocator interacts with, rather than
> having to deal with a combination of memcg_ptr and memcg.
>
> In the two hunks here, on one hand we charge bytes to the memcg
> object, and then handle all the refcounting through a different
> bucketing object. To support that in the first place, we have to
> overload the memcg API all the way down to try_charge() to support
> bytes and pages. This is difficult to follow throughout all layers.
>
> What would be better is for this to be an abstraction layer for a
> subpage object tracker that sits on top of the memcg page tracker -
> not unlike the page allocator and the slab allocators themselves.
>
> And then the slab allocator would only interact with the subpage
> object tracker, and the object tracker would deal with the memcg page
> tracker under the hood.
Yes, the idea makes total sense to me. I'm not sure I like the new naming
(I have to spend some time with it first), but the idea of moving
stocks and leftovers to the memcg_ptr/obj_cgroup level is really good.
I'll include something based on your proposal into the next version
of the patchset.
Thank you!
