On Thu, Jul 11, 2024 at 4:49 AM Alex Shi <seakeel@xxxxxxxxx> wrote:
>
>
>
> On 7/11/24 4:11 PM, Vlastimil Babka wrote:
> > On 7/10/24 7:43 AM, alexs@xxxxxxxxxx wrote:
> >> From: "Alex Shi (Tencent)" <alexs@xxxxxxxxxx>
> >>
> >> commit 21c690a349ba ("mm: introduce slabobj_ext to support slab object
> >> extensions") selected SLAB_OBJ_EXT on MEMCG just for SLAB_MATCH
> >> memcg_data, that included SLAB_OBJ_EXT for MEMCG. In fact, I didn't see
> >> the necessary to enable SLAB_OBJ_EXT for MEMCG.
> >>
> >> Let's decouple the SLAB_OBJ_EXT from MEMCG and move out
> >> alloc_slab_obj_exts() definition from SLAB_OBJ_EXT only. To alignment
> >> the alloc_slab_obj_exts() return 0 for good. change its return value to
> >> '-1' for always failed with !SLAB_OBJ_EXT. Now we could save unnecessary
> >> code from MEMCG but !SLAB_OBJ_EXT.
> >>
> >> Signed-off-by: Alex Shi (Tencent) <alexs@xxxxxxxxxx>
> >
> > This seems just wrong to me. The memcg hooks for slab do use obj_ext. You
> > made alloc_slab_obj_exts() return -1 and that will just fail all memcg
> > charging (unless alloc profiling selects obj_ext). The kernel will appear to
> > work, but memcg charging for slab won't happen at all.
> >
> > So no, it can't be decoupled for slab, only for pages/folios (patch 1).
>
> Hi Vlastimil,
>
> Thanks a lot for clarification! Yes, the patch isn't correct.
>
> Just forgive my stupidity, why the memcg needs SLAB_OBJ_EXT?
Because when CONFIG_MEMCG_KMEM=y, slabobj_ext contains obj_cgroup
(see: https://elixir.bootlin.com/linux/v6.10-rc7/source/include/linux/memcontrol.h#L1593)
and that's used for memcg accounting. Look into this call chain:
kfree
slab_free
memcg_slab_free_hook
__memcg_slab_free_hook
obj_cgroup_uncharge
>
> And why we need to alloc_slab_obj_exts() at line 3019 with !slab_obj_exts?
> 3015 for (i = 0; i < size; i++) {
> 3016 slab = virt_to_slab(p[i]);
> 3017
> 3018 if (!slab_obj_exts(slab) &&
> 3019 alloc_slab_obj_exts(slab, s, flags, false)) {
> 3020 obj_cgroup_uncharge(objcg, obj_full_size(s));
> 3021 continue;
> 3022 }
>
> Thanks!
> Alex
>
> >
> >
> >> Cc: Randy Dunlap <rdunlap@xxxxxxxxxxxxx>
> >> Cc: Yoann Congal <yoann.congal@xxxxxxxx>
> >> Cc: Masahiro Yamada <masahiroy@xxxxxxxxxx>
> >> Cc: Petr Mladek <pmladek@xxxxxxxx>
> >> ---
> >> init/Kconfig | 1 -
> >> mm/slab.h | 6 +++---
> >> mm/slub.c | 6 +++---
> >> 3 files changed, 6 insertions(+), 7 deletions(-)
> >>
> >> diff --git a/init/Kconfig b/init/Kconfig
> >> index 26bf8bb0a7ce..61e43ac9fe75 100644
> >> --- a/init/Kconfig
> >> +++ b/init/Kconfig
> >> @@ -965,7 +965,6 @@ config MEMCG
> >> bool "Memory controller"
> >> select PAGE_COUNTER
> >> select EVENTFD
> >> - select SLAB_OBJ_EXT
> >> help
> >> Provides control over the memory footprint of tasks in a cgroup.
> >>
> >> diff --git a/mm/slab.h b/mm/slab.h
> >> index 8ffdd4f315f8..6c727ecc1068 100644
> >> --- a/mm/slab.h
> >> +++ b/mm/slab.h
> >> @@ -559,9 +559,6 @@ static inline struct slabobj_ext *slab_obj_exts(struct slab *slab)
> >> return (struct slabobj_ext *)(obj_exts & ~OBJEXTS_FLAGS_MASK);
> >> }
> >>
> >> -int alloc_slab_obj_exts(struct slab *slab, struct kmem_cache *s,
> >> - gfp_t gfp, bool new_slab);
> >> -
> >> #else /* CONFIG_SLAB_OBJ_EXT */
> >>
> >> static inline struct slabobj_ext *slab_obj_exts(struct slab *slab)
> >> @@ -571,6 +568,9 @@ static inline struct slabobj_ext *slab_obj_exts(struct slab *slab)
> >>
> >> #endif /* CONFIG_SLAB_OBJ_EXT */
> >>
> >> +int alloc_slab_obj_exts(struct slab *slab, struct kmem_cache *s,
> >> + gfp_t gfp, bool new_slab);
> >> +
> >> static inline enum node_stat_item cache_vmstat_idx(struct kmem_cache *s)
> >> {
> >> return (s->flags & SLAB_RECLAIM_ACCOUNT) ?
> >> diff --git a/mm/slub.c b/mm/slub.c
> >> index cc11f3869cc6..f531c2d67238 100644
> >> --- a/mm/slub.c
> >> +++ b/mm/slub.c
> >> @@ -2075,10 +2075,10 @@ alloc_tagging_slab_free_hook(struct kmem_cache *s, struct slab *slab, void **p,
> >>
> >> #else /* CONFIG_SLAB_OBJ_EXT */
> >>
> >> -static int alloc_slab_obj_exts(struct slab *slab, struct kmem_cache *s,
> >> - gfp_t gfp, bool new_slab)
> >> +int alloc_slab_obj_exts(struct slab *slab, struct kmem_cache *s,
> >> + gfp_t gfp, bool new_slab)
> >> {
> >> - return 0;
> >> + return -1;
> >> }
> >>
> >> static inline void free_slab_obj_exts(struct slab *slab)
> >
