Re: [PATCH bpf-next v2] bpf: reparent bpf maps on memcg offlining

[Roman Gushchin <roman.gushchin@linux.dev>]

On Tue, Jul 12, 2022 at 03:15:27PM -0700, Alexei Starovoitov wrote:
> On Tue, Jul 12, 2022 at 3:11 PM Shakeel Butt <shakeelb@google.com> wrote:
> >
> > On Tue, Jul 12, 2022 at 2:49 PM Alexei Starovoitov
> > <alexei.starovoitov@gmail.com> wrote:
> > >
> > > On Mon, Jul 11, 2022 at 9:28 AM Roman Gushchin <roman.gushchin@linux.dev> wrote:
> > > >
> > > > The memory consumed by a mpf map is always accounted to the memory
> > > > cgroup of the process which created the map. The map can outlive
> > > > the memory cgroup if it's used by processes in other cgroups or
> > > > is pinned on bpffs. In this case the map pins the original cgroup
> > > > in the dying state.
> > > >
> > > > For other types of objects (slab objects, non-slab kernel allocations,
> > > > percpu objects and recently LRU pages) there is a reparenting process
> > > > implemented: on cgroup offlining charged objects are getting
> > > > reassigned to the parent cgroup. Because all charges and statistics
> > > > are fully recursive it's a fairly cheap operation.
> > > >
> > > > For efficiency and consistency with other types of objects, let's do
> > > > the same for bpf maps. Fortunately thanks to the objcg API, the
> > > > required changes are minimal.
> > > >
> > > > Please, note that individual allocations (slabs, percpu and large
> > > > kmallocs) already have the reparenting mechanism. This commit adds
> > > > it to the saved map->memcg pointer by replacing it to map->objcg.
> > > > Because dying cgroups are not visible for a user and all charges are
> > > > recursive, this commit doesn't bring any behavior changes for a user.
> > > >
> > > > v2:
> > > >   added a missing const qualifier
> > > >
> > > > Signed-off-by: Roman Gushchin <roman.gushchin@linux.dev>
> > > > Reviewed-by: Shakeel Butt <shakeelb@google.com>
> > > > ---
> > > >  include/linux/bpf.h  |  2 +-
> > > >  kernel/bpf/syscall.c | 35 +++++++++++++++++++++++++++--------
> > > >  2 files changed, 28 insertions(+), 9 deletions(-)
> > > >
> > > > diff --git a/include/linux/bpf.h b/include/linux/bpf.h
> > > > index 2b21f2a3452f..85a4db3e0536 100644
> > > > --- a/include/linux/bpf.h
> > > > +++ b/include/linux/bpf.h
> > > > @@ -221,7 +221,7 @@ struct bpf_map {
> > > >         u32 btf_vmlinux_value_type_id;
> > > >         struct btf *btf;
> > > >  #ifdef CONFIG_MEMCG_KMEM
> > > > -       struct mem_cgroup *memcg;
> > > > +       struct obj_cgroup *objcg;
> > > >  #endif
> > > >         char name[BPF_OBJ_NAME_LEN];
> > > >         struct bpf_map_off_arr *off_arr;
> > > > diff --git a/kernel/bpf/syscall.c b/kernel/bpf/syscall.c
> > > > index ab688d85b2c6..ef60dbc21b17 100644
> > > > --- a/kernel/bpf/syscall.c
> > > > +++ b/kernel/bpf/syscall.c
> > > > @@ -419,35 +419,52 @@ void bpf_map_free_id(struct bpf_map *map, bool do_idr_lock)
> > > >  #ifdef CONFIG_MEMCG_KMEM
> > > >  static void bpf_map_save_memcg(struct bpf_map *map)
> > > >  {
> > > > -       map->memcg = get_mem_cgroup_from_mm(current->mm);
> > > > +       /* Currently if a map is created by a process belonging to the root
> > > > +        * memory cgroup, get_obj_cgroup_from_current() will return NULL.
> > > > +        * So we have to check map->objcg for being NULL each time it's
> > > > +        * being used.
> > > > +        */
> > > > +       map->objcg = get_obj_cgroup_from_current();
> > > >  }
> > > >
> > > >  static void bpf_map_release_memcg(struct bpf_map *map)
> > > >  {
> > > > -       mem_cgroup_put(map->memcg);
> > > > +       if (map->objcg)
> > > > +               obj_cgroup_put(map->objcg);
> > > > +}
> > > > +
> > > > +static struct mem_cgroup *bpf_map_get_memcg(const struct bpf_map *map) {
> > > > +       if (map->objcg)
> > > > +               return get_mem_cgroup_from_objcg(map->objcg);
> > > > +
> > > > +       return root_mem_cgroup;
> > > >  }
> > > >
> > > >  void *bpf_map_kmalloc_node(const struct bpf_map *map, size_t size, gfp_t flags,
> > > >                            int node)
> > > >  {
> > > > -       struct mem_cgroup *old_memcg;
> > > > +       struct mem_cgroup *memcg, *old_memcg;
> > > >         void *ptr;
> > > >
> > > > -       old_memcg = set_active_memcg(map->memcg);
> > > > +       memcg = bpf_map_get_memcg(map);
> > > > +       old_memcg = set_active_memcg(memcg);
> > > >         ptr = kmalloc_node(size, flags | __GFP_ACCOUNT, node);
> > > >         set_active_memcg(old_memcg);
> > > > +       mem_cgroup_put(memcg);
> > >
> > > Here we might css_put root_mem_cgroup.
> > > Should we css_get it when returning or
> > > it's marked as CSS_NO_REF ?
> > > But mem_cgroup_alloc() doesn't seem to be doing that marking.
> > > I'm lost at that code.
> >
> > CSS_NO_REF is set for root_mem_cgroup in cgroup_init_subsys().

Yeah, the root cgroups can't be deleted so we save on the refcounting.
> 
> Ahh. I see that
> css = ss->css_alloc(NULL); css->flags |= CSS_NO_REF; now.
> Thanks.

Thanks for applying the patch!