On 1/5/22 03:41, Roman Gushchin wrote: > On Tue, Jan 04, 2022 at 01:10:37AM +0100, Vlastimil Babka wrote: >> page->memcg_data is used with MEMCG_DATA_OBJCGS flag only for slab pages >> so convert all the related infrastructure to struct slab. Also use >> struct folio instead of struct page when resolving object pointers. >> >> This is not just mechanistic changing of types and names. Now in >> mem_cgroup_from_obj() we use folio_test_slab() to decide if we interpret >> the folio as a real slab instead of a large kmalloc, instead of relying >> on MEMCG_DATA_OBJCGS bit that used to be checked in page_objcgs_check(). >> Similarly in memcg_slab_free_hook() where we can encounter >> kmalloc_large() pages (here the folio slab flag check is implied by >> virt_to_slab()). As a result, page_objcgs_check() can be dropped instead >> of converted. >> >> To avoid include cycles, move the inline definition of slab_objcgs() >> from memcontrol.h to mm/slab.h. >> >> Signed-off-by: Vlastimil Babka <vbabka@xxxxxxx> >> Cc: Johannes Weiner <hannes@xxxxxxxxxxx> >> Cc: Michal Hocko <mhocko@xxxxxxxxxx> >> Cc: Vladimir Davydov <vdavydov.dev@xxxxxxxxx> >> Cc: <cgroups@xxxxxxxxxxxxxxx> >> /* >> * Slab objects are accounted individually, not per-page. >> * Memcg membership data for each individual object is saved in >> * the page->obj_cgroups. > ^^^^^^^^^^^^^^^^^ > slab->memcg_data Good catch, fixed. >> */ >> - if (page_objcgs_check(page)) { >> - struct obj_cgroup *objcg; >> + if (folio_test_slab(folio)) { >> + struct obj_cgroup **objcgs; >> + struct slab *slab; >> unsigned int off; >> >> - off = obj_to_index(page->slab_cache, page_slab(page), p); >> - objcg = page_objcgs(page)[off]; >> - if (objcg) >> - return obj_cgroup_memcg(objcg); >> + slab = folio_slab(folio); >> + objcgs = slab_objcgs(slab); >> + if (!objcgs) >> + return NULL; >> + >> + off = obj_to_index(slab->slab_cache, slab, p); >> + if (objcgs[off]) >> + return obj_cgroup_memcg(objcgs[off]); >> >> return NULL; >> } > > There is a comment below, which needs some changes: > /* > * page_memcg_check() is used here, because page_has_obj_cgroups() > * check above could fail because the object cgroups vector wasn't set > * at that moment, but it can be set concurrently. > * page_memcg_check(page) will guarantee that a proper memory > * cgroup pointer or NULL will be returned. > */ > > In reality the folio's slab flag can be cleared before releasing the objcgs \ > vector. It seems that there is no such possibility at setting the flag, > it's always set before allocating and assigning the objcg vector. You're right. I'm changing it to: * page_memcg_check() is used here, because in theory we can encounter * a folio where the slab flag has been cleared already, but * slab->memcg_data has not been freed yet * page_memcg_check(page) will guarantee that a proper memory * cgroup pointer or NULL will be returned. I wrote "in theory" because AFAICS it implies a race as we would have to be freeing a slab and at the same time query an object address. We probably could have used the non-check version, but at this point I don't want to make any functional changes besides these comment fixes. I assume your patch on top would cover it? >> @@ -2896,7 +2901,7 @@ struct mem_cgroup *mem_cgroup_from_obj(void *p) >> * page_memcg_check(page) will guarantee that a proper memory >> * cgroup pointer or NULL will be returned. >> */ >> - return page_memcg_check(page); >> + return page_memcg_check(folio_page(folio, 0)); >> } >> >> __always_inline struct obj_cgroup *get_obj_cgroup_from_current(void) >> diff --git a/mm/slab.h b/mm/slab.h >> index bca9181e96d7..36e0022d8267 100644 >> --- a/mm/slab.h >> +++ b/mm/slab.h >> @@ -412,15 +412,36 @@ static inline bool kmem_cache_debug_flags(struct kmem_cache *s, slab_flags_t fla >> } >> >> #ifdef CONFIG_MEMCG_KMEM >> -int memcg_alloc_page_obj_cgroups(struct page *page, struct kmem_cache *s, >> - gfp_t gfp, bool new_page); >> +/* >> + * slab_objcgs - get the object cgroups vector associated with a slab >> + * @slab: a pointer to the slab struct >> + * >> + * Returns a pointer to the object cgroups vector associated with the slab, >> + * or NULL. This function assumes that the slab is known to have an >> + * associated object cgroups vector. It's not safe to call this function >> + * against slabs with underlying pages, which might have an associated memory >> + * cgroup: e.g. kernel stack pages. > > Hm, is it still true? I don't think so. It must be safe to call it for any > slab now. Right, forgot to update after removing the _check variant. Changing to: * Returns a pointer to the object cgroups vector associated with the slab, * or NULL if no such vector has been associated yet. > The rest looks good to me, please feel free to add > Reviewed-by: Roman Gushchin <guro@xxxxxx> > after fixing these comments. Thanks! > Thanks!