mm, slab: move memcg charging to post-alloc hook
authorVlastimil Babka <vbabka@suse.cz>
Tue, 26 Mar 2024 10:37:38 +0000 (11:37 +0100)
committerAndrew Morton <akpm@linux-foundation.org>
Fri, 26 Apr 2024 03:56:16 +0000 (20:56 -0700)
Patch series "memcg_kmem hooks refactoring", v3.

This patch (of 2):

The MEMCG_KMEM integration with slab currently relies on two hooks during
allocation.  memcg_slab_pre_alloc_hook() determines the objcg and charges
it, and memcg_slab_post_alloc_hook() assigns the objcg pointer to the
allocated object(s).

As Linus pointed out, this is unnecessarily complex.  Failing to charge
due to memcg limits should be rare, so we can optimistically allocate the
object(s) and do the charging together with assigning the objcg pointer in
a single post_alloc hook.  In the rare case the charging fails, we can
free the object(s) back.

This simplifies the code (no need to pass around the objcg pointer) and
potentially allows to separate charging from allocation in cases where
it's common that the allocation would be immediately freed, and the memcg
handling overhead could be saved.

[vbabka@suse.cz: fix call to memcg_alloc_abort_single()]
Link: https://lkml.kernel.org/r/4af50be2-4109-45e5-8a36-2136252a635e@suse.cz
[roman.gushchin@linux.dev: comment fixup]
Link: https://lkml.kernel.org/r/Zg2LsNm6twOmG69l@P9FQF9L96D.corp.robot.car
Link: https://lkml.kernel.org/r/20240326-slab-memcg-v3-0-d85d2563287a@suse.cz
Link: https://lkml.kernel.org/r/20240326-slab-memcg-v3-1-d85d2563287a@suse.cz
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Roman Gushchin <roman.gushchin@linux.dev>
Suggested-by: Linus Torvalds <torvalds@linux-foundation.org>
Link: https://lore.kernel.org/all/CAHk-=whYOOdM7jWy5jdrAm8LxcgCMFyk2bt8fYYvZzM4U-zAQA@mail.gmail.com/
Reviewed-by: Roman Gushchin <roman.gushchin@linux.dev>
Reviewed-by: Chengming Zhou <chengming.zhou@linux.dev>
Cc: Al Viro <viro@ZenIV.linux.org.uk>
Cc: Christian Brauner <brauner@kernel.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Chuck Lever <chuck.lever@oracle.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Hyeonggon Yoo <42.hyeyoo@gmail.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jeff Layton <jlayton@kernel.org>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Josh Poimboeuf <jpoimboe@kernel.org>
Cc: Kees Cook <kees@kernel.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: Shakeel Butt <shakeel.butt@linux.dev>
Cc: Aishwarya TCV <aishwarya.tcv@arm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
mm/memcontrol.c
mm/slub.c

index 45dd209012827b3100d9f078f7dd7297611779d5..a2aa2ff6cf7788b4f87d3cda3a2faf77c85ab694 100644 (file)
@@ -350,7 +350,7 @@ static void memcg_reparent_objcgs(struct mem_cgroup *memcg,
 
 /*
  * A lot of the calls to the cache allocation functions are expected to be
- * inlined by the compiler. Since the calls to memcg_slab_pre_alloc_hook() are
+ * inlined by the compiler. Since the calls to memcg_slab_post_alloc_hook() are
  * conditional to this static branch, we'll have to allow modules that does
  * kmem_cache_alloc and the such to see this symbol as well
  */
index e09d8260adff5dbe23b8fff726eb6c5a195702a2..f83efdbe15b7e60e44cd1d5149220eff6dd7587c 100644 (file)
--- a/mm/slub.c
+++ b/mm/slub.c
@@ -2092,23 +2092,36 @@ static inline size_t obj_full_size(struct kmem_cache *s)
        return s->size + sizeof(struct obj_cgroup *);
 }
 
-/*
- * Returns false if the allocation should fail.
- */
-static bool __memcg_slab_pre_alloc_hook(struct kmem_cache *s,
-                                       struct list_lru *lru,
-                                       struct obj_cgroup **objcgp,
-                                       size_t objects, gfp_t flags)
+static bool __memcg_slab_post_alloc_hook(struct kmem_cache *s,
+                                        struct list_lru *lru,
+                                        gfp_t flags, size_t size,
+                                        void **p)
 {
+       struct obj_cgroup *objcg;
+       struct slab *slab;
+       unsigned long off;
+       size_t i;
+
        /*
         * The obtained objcg pointer is safe to use within the current scope,
         * defined by current task or set_active_memcg() pair.
         * obj_cgroup_get() is used to get a permanent reference.
         */
-       struct obj_cgroup *objcg = current_obj_cgroup();
+       objcg = current_obj_cgroup();
        if (!objcg)
                return true;
 
+       /*
+        * slab_alloc_node() avoids the NULL check, so we might be called with a
+        * single NULL object. kmem_cache_alloc_bulk() aborts if it can't fill
+        * the whole requested size.
+        * return success as there's nothing to free back
+        */
+       if (unlikely(*p == NULL))
+               return true;
+
+       flags &= gfp_allowed_mask;
+
        if (lru) {
                int ret;
                struct mem_cgroup *memcg;
@@ -2121,71 +2134,51 @@ static bool __memcg_slab_pre_alloc_hook(struct kmem_cache *s,
                        return false;
        }
 
-       if (obj_cgroup_charge(objcg, flags, objects * obj_full_size(s)))
+       if (obj_cgroup_charge(objcg, flags, size * obj_full_size(s)))
                return false;
 
-       *objcgp = objcg;
+       for (i = 0; i < size; i++) {
+               slab = virt_to_slab(p[i]);
+
+               if (!slab_obj_exts(slab) &&
+                   alloc_slab_obj_exts(slab, s, flags, false)) {
+                       obj_cgroup_uncharge(objcg, obj_full_size(s));
+                       continue;
+               }
+
+               off = obj_to_index(s, slab, p[i]);
+               obj_cgroup_get(objcg);
+               slab_obj_exts(slab)[off].objcg = objcg;
+               mod_objcg_state(objcg, slab_pgdat(slab),
+                               cache_vmstat_idx(s), obj_full_size(s));
+       }
+
        return true;
 }
 
-/*
- * Returns false if the allocation should fail.
- */
+static void memcg_alloc_abort_single(struct kmem_cache *s, void *object);
+
 static __fastpath_inline
-bool memcg_slab_pre_alloc_hook(struct kmem_cache *s, struct list_lru *lru,
-                              struct obj_cgroup **objcgp, size_t objects,
-                              gfp_t flags)
+bool memcg_slab_post_alloc_hook(struct kmem_cache *s, struct list_lru *lru,
+                               gfp_t flags, size_t size, void **p)
 {
-       if (!memcg_kmem_online())
+       if (likely(!memcg_kmem_online()))
                return true;
 
        if (likely(!(flags & __GFP_ACCOUNT) && !(s->flags & SLAB_ACCOUNT)))
                return true;
 
-       return likely(__memcg_slab_pre_alloc_hook(s, lru, objcgp, objects,
-                                                 flags));
-}
-
-static void __memcg_slab_post_alloc_hook(struct kmem_cache *s,
-                                        struct obj_cgroup *objcg,
-                                        gfp_t flags, size_t size,
-                                        void **p)
-{
-       struct slab *slab;
-       unsigned long off;
-       size_t i;
-
-       flags &= gfp_allowed_mask;
-
-       for (i = 0; i < size; i++) {
-               if (likely(p[i])) {
-                       slab = virt_to_slab(p[i]);
-
-                       if (!slab_obj_exts(slab) &&
-                           alloc_slab_obj_exts(slab, s, flags, false)) {
-                               obj_cgroup_uncharge(objcg, obj_full_size(s));
-                               continue;
-                       }
+       if (likely(__memcg_slab_post_alloc_hook(s, lru, flags, size, p)))
+               return true;
 
-                       off = obj_to_index(s, slab, p[i]);
-                       obj_cgroup_get(objcg);
-                       slab_obj_exts(slab)[off].objcg = objcg;
-                       mod_objcg_state(objcg, slab_pgdat(slab),
-                                       cache_vmstat_idx(s), obj_full_size(s));
-               } else {
-                       obj_cgroup_uncharge(objcg, obj_full_size(s));
-               }
+       if (likely(size == 1)) {
+               memcg_alloc_abort_single(s, *p);
+               *p = NULL;
+       } else {
+               kmem_cache_free_bulk(s, size, p);
        }
-}
-
-static __fastpath_inline
-void memcg_slab_post_alloc_hook(struct kmem_cache *s, struct obj_cgroup *objcg,
-                               gfp_t flags, size_t size, void **p)
-{
-       if (likely(!memcg_kmem_online() || !objcg))
-               return;
 
-       return __memcg_slab_post_alloc_hook(s, objcg, flags, size, p);
+       return false;
 }
 
 static void __memcg_slab_free_hook(struct kmem_cache *s, struct slab *slab,
@@ -2224,40 +2217,19 @@ void memcg_slab_free_hook(struct kmem_cache *s, struct slab *slab, void **p,
 
        __memcg_slab_free_hook(s, slab, p, objects, obj_exts);
 }
-
-static inline
-void memcg_slab_alloc_error_hook(struct kmem_cache *s, int objects,
-                          struct obj_cgroup *objcg)
-{
-       if (objcg)
-               obj_cgroup_uncharge(objcg, objects * obj_full_size(s));
-}
 #else /* CONFIG_MEMCG_KMEM */
-static inline bool memcg_slab_pre_alloc_hook(struct kmem_cache *s,
-                                            struct list_lru *lru,
-                                            struct obj_cgroup **objcgp,
-                                            size_t objects, gfp_t flags)
-{
-       return true;
-}
-
-static inline void memcg_slab_post_alloc_hook(struct kmem_cache *s,
-                                             struct obj_cgroup *objcg,
+static inline bool memcg_slab_post_alloc_hook(struct kmem_cache *s,
+                                             struct list_lru *lru,
                                              gfp_t flags, size_t size,
                                              void **p)
 {
+       return true;
 }
 
 static inline void memcg_slab_free_hook(struct kmem_cache *s, struct slab *slab,
                                        void **p, int objects)
 {
 }
-
-static inline
-void memcg_slab_alloc_error_hook(struct kmem_cache *s, int objects,
-                                struct obj_cgroup *objcg)
-{
-}
 #endif /* CONFIG_MEMCG_KMEM */
 
 /*
@@ -3937,10 +3909,7 @@ noinline int should_failslab(struct kmem_cache *s, gfp_t gfpflags)
 ALLOW_ERROR_INJECTION(should_failslab, ERRNO);
 
 static __fastpath_inline
-struct kmem_cache *slab_pre_alloc_hook(struct kmem_cache *s,
-                                      struct list_lru *lru,
-                                      struct obj_cgroup **objcgp,
-                                      size_t size, gfp_t flags)
+struct kmem_cache *slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
 {
        flags &= gfp_allowed_mask;
 
@@ -3949,14 +3918,11 @@ struct kmem_cache *slab_pre_alloc_hook(struct kmem_cache *s,
        if (unlikely(should_failslab(s, flags)))
                return NULL;
 
-       if (unlikely(!memcg_slab_pre_alloc_hook(s, lru, objcgp, size, flags)))
-               return NULL;
-
        return s;
 }
 
 static __fastpath_inline
-void slab_post_alloc_hook(struct kmem_cache *s,        struct obj_cgroup *objcg,
+bool slab_post_alloc_hook(struct kmem_cache *s, struct list_lru *lru,
                          gfp_t flags, size_t size, void **p, bool init,
                          unsigned int orig_size)
 {
@@ -4018,7 +3984,7 @@ void slab_post_alloc_hook(struct kmem_cache *s,   struct obj_cgroup *objcg,
                }
        }
 
-       memcg_slab_post_alloc_hook(s, objcg, flags, size, p);
+       return memcg_slab_post_alloc_hook(s, lru, flags, size, p);
 }
 
 /*
@@ -4035,10 +4001,9 @@ static __fastpath_inline void *slab_alloc_node(struct kmem_cache *s, struct list
                gfp_t gfpflags, int node, unsigned long addr, size_t orig_size)
 {
        void *object;
-       struct obj_cgroup *objcg = NULL;
        bool init = false;
 
-       s = slab_pre_alloc_hook(s, lru, &objcg, 1, gfpflags);
+       s = slab_pre_alloc_hook(s, gfpflags);
        if (unlikely(!s))
                return NULL;
 
@@ -4055,8 +4020,10 @@ out:
        /*
         * When init equals 'true', like for kzalloc() family, only
         * @orig_size bytes might be zeroed instead of s->object_size
+        * In case this fails due to memcg_slab_post_alloc_hook(),
+        * object is set to NULL
         */
-       slab_post_alloc_hook(s, objcg, gfpflags, 1, &object, init, orig_size);
+       slab_post_alloc_hook(s, lru, gfpflags, 1, &object, init, orig_size);
 
        return object;
 }
@@ -4496,6 +4463,16 @@ void slab_free(struct kmem_cache *s, struct slab *slab, void *object,
                do_slab_free(s, slab, object, object, 1, addr);
 }
 
+#ifdef CONFIG_MEMCG_KMEM
+/* Do not inline the rare memcg charging failed path into the allocation path */
+static noinline
+void memcg_alloc_abort_single(struct kmem_cache *s, void *object)
+{
+       if (likely(slab_free_hook(s, object, slab_want_init_on_free(s))))
+               do_slab_free(s, virt_to_slab(object), object, object, 1, _RET_IP_);
+}
+#endif
+
 static __fastpath_inline
 void slab_free_bulk(struct kmem_cache *s, struct slab *slab, void *head,
                    void *tail, void **p, int cnt, unsigned long addr)
@@ -4832,29 +4809,26 @@ int kmem_cache_alloc_bulk_noprof(struct kmem_cache *s, gfp_t flags, size_t size,
                                 void **p)
 {
        int i;
-       struct obj_cgroup *objcg = NULL;
 
        if (!size)
                return 0;
 
-       /* memcg and kmem_cache debug support */
-       s = slab_pre_alloc_hook(s, NULL, &objcg, size, flags);
+       s = slab_pre_alloc_hook(s, flags);
        if (unlikely(!s))
                return 0;
 
        i = __kmem_cache_alloc_bulk(s, flags, size, p);
+       if (unlikely(i == 0))
+               return 0;
 
        /*
         * memcg and kmem_cache debug support and memory initialization.
         * Done outside of the IRQ disabled fastpath loop.
         */
-       if (likely(i != 0)) {
-               slab_post_alloc_hook(s, objcg, flags, size, p,
-                       slab_want_init_on_alloc(flags, s), s->object_size);
-       } else {
-               memcg_slab_alloc_error_hook(s, size, objcg);
+       if (unlikely(!slab_post_alloc_hook(s, NULL, flags, size, p,
+                   slab_want_init_on_alloc(flags, s), s->object_size))) {
+               return 0;
        }
-
        return i;
 }
 EXPORT_SYMBOL(kmem_cache_alloc_bulk_noprof);