1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * Kernel Memory Controller
14 * Copyright (C) 2012 Parallels Inc. and Google Inc.
15 * Authors: Glauber Costa and Suleiman Souhlal
18 * Charge lifetime sanitation
19 * Lockless page tracking & accounting
20 * Unified hierarchy configuration model
21 * Copyright (C) 2015 Red Hat, Inc., Johannes Weiner
23 * This program is free software; you can redistribute it and/or modify
24 * it under the terms of the GNU General Public License as published by
25 * the Free Software Foundation; either version 2 of the License, or
26 * (at your option) any later version.
28 * This program is distributed in the hope that it will be useful,
29 * but WITHOUT ANY WARRANTY; without even the implied warranty of
30 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
31 * GNU General Public License for more details.
34 #include <linux/page_counter.h>
35 #include <linux/memcontrol.h>
36 #include <linux/cgroup.h>
38 #include <linux/sched/mm.h>
39 #include <linux/shmem_fs.h>
40 #include <linux/hugetlb.h>
41 #include <linux/pagemap.h>
42 #include <linux/vm_event_item.h>
43 #include <linux/smp.h>
44 #include <linux/page-flags.h>
45 #include <linux/backing-dev.h>
46 #include <linux/bit_spinlock.h>
47 #include <linux/rcupdate.h>
48 #include <linux/limits.h>
49 #include <linux/export.h>
50 #include <linux/mutex.h>
51 #include <linux/rbtree.h>
52 #include <linux/slab.h>
53 #include <linux/swap.h>
54 #include <linux/swapops.h>
55 #include <linux/spinlock.h>
56 #include <linux/eventfd.h>
57 #include <linux/poll.h>
58 #include <linux/sort.h>
60 #include <linux/seq_file.h>
61 #include <linux/vmpressure.h>
62 #include <linux/mm_inline.h>
63 #include <linux/swap_cgroup.h>
64 #include <linux/cpu.h>
65 #include <linux/oom.h>
66 #include <linux/lockdep.h>
67 #include <linux/file.h>
68 #include <linux/tracehook.h>
74 #include <linux/uaccess.h>
76 #include <trace/events/vmscan.h>
78 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
79 EXPORT_SYMBOL(memory_cgrp_subsys);
81 struct mem_cgroup *root_mem_cgroup __read_mostly;
83 #define MEM_CGROUP_RECLAIM_RETRIES 5
85 /* Socket memory accounting disabled? */
86 static bool cgroup_memory_nosocket;
88 /* Kernel memory accounting disabled? */
89 static bool cgroup_memory_nokmem;
91 /* Whether the swap controller is active */
92 #ifdef CONFIG_MEMCG_SWAP
93 int do_swap_account __read_mostly;
95 #define do_swap_account 0
98 /* Whether legacy memory+swap accounting is active */
99 static bool do_memsw_account(void)
101 return !cgroup_subsys_on_dfl(memory_cgrp_subsys) && do_swap_account;
104 static const char *const mem_cgroup_lru_names[] = {
112 #define THRESHOLDS_EVENTS_TARGET 128
113 #define SOFTLIMIT_EVENTS_TARGET 1024
114 #define NUMAINFO_EVENTS_TARGET 1024
117 * Cgroups above their limits are maintained in a RB-Tree, independent of
118 * their hierarchy representation
121 struct mem_cgroup_tree_per_node {
122 struct rb_root rb_root;
123 struct rb_node *rb_rightmost;
127 struct mem_cgroup_tree {
128 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
131 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
134 struct mem_cgroup_eventfd_list {
135 struct list_head list;
136 struct eventfd_ctx *eventfd;
140 * cgroup_event represents events which userspace want to receive.
142 struct mem_cgroup_event {
144 * memcg which the event belongs to.
146 struct mem_cgroup *memcg;
148 * eventfd to signal userspace about the event.
150 struct eventfd_ctx *eventfd;
152 * Each of these stored in a list by the cgroup.
154 struct list_head list;
156 * register_event() callback will be used to add new userspace
157 * waiter for changes related to this event. Use eventfd_signal()
158 * on eventfd to send notification to userspace.
160 int (*register_event)(struct mem_cgroup *memcg,
161 struct eventfd_ctx *eventfd, const char *args);
163 * unregister_event() callback will be called when userspace closes
164 * the eventfd or on cgroup removing. This callback must be set,
165 * if you want provide notification functionality.
167 void (*unregister_event)(struct mem_cgroup *memcg,
168 struct eventfd_ctx *eventfd);
170 * All fields below needed to unregister event when
171 * userspace closes eventfd.
174 wait_queue_head_t *wqh;
175 wait_queue_entry_t wait;
176 struct work_struct remove;
179 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
180 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
182 /* Stuffs for move charges at task migration. */
184 * Types of charges to be moved.
186 #define MOVE_ANON 0x1U
187 #define MOVE_FILE 0x2U
188 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
190 /* "mc" and its members are protected by cgroup_mutex */
191 static struct move_charge_struct {
192 spinlock_t lock; /* for from, to */
193 struct mm_struct *mm;
194 struct mem_cgroup *from;
195 struct mem_cgroup *to;
197 unsigned long precharge;
198 unsigned long moved_charge;
199 unsigned long moved_swap;
200 struct task_struct *moving_task; /* a task moving charges */
201 wait_queue_head_t waitq; /* a waitq for other context */
203 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
204 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
208 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
209 * limit reclaim to prevent infinite loops, if they ever occur.
211 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
212 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
215 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
216 MEM_CGROUP_CHARGE_TYPE_ANON,
217 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
218 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
222 /* for encoding cft->private value on file */
231 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
232 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
233 #define MEMFILE_ATTR(val) ((val) & 0xffff)
234 /* Used for OOM nofiier */
235 #define OOM_CONTROL (0)
238 * Iteration constructs for visiting all cgroups (under a tree). If
239 * loops are exited prematurely (break), mem_cgroup_iter_break() must
240 * be used for reference counting.
242 #define for_each_mem_cgroup_tree(iter, root) \
243 for (iter = mem_cgroup_iter(root, NULL, NULL); \
245 iter = mem_cgroup_iter(root, iter, NULL))
247 #define for_each_mem_cgroup(iter) \
248 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
250 iter = mem_cgroup_iter(NULL, iter, NULL))
252 static inline bool should_force_charge(void)
254 return tsk_is_oom_victim(current) || fatal_signal_pending(current) ||
255 (current->flags & PF_EXITING);
258 /* Some nice accessors for the vmpressure. */
259 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
262 memcg = root_mem_cgroup;
263 return &memcg->vmpressure;
266 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
268 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
271 #ifdef CONFIG_MEMCG_KMEM
273 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
274 * The main reason for not using cgroup id for this:
275 * this works better in sparse environments, where we have a lot of memcgs,
276 * but only a few kmem-limited. Or also, if we have, for instance, 200
277 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
278 * 200 entry array for that.
280 * The current size of the caches array is stored in memcg_nr_cache_ids. It
281 * will double each time we have to increase it.
283 static DEFINE_IDA(memcg_cache_ida);
284 int memcg_nr_cache_ids;
286 /* Protects memcg_nr_cache_ids */
287 static DECLARE_RWSEM(memcg_cache_ids_sem);
289 void memcg_get_cache_ids(void)
291 down_read(&memcg_cache_ids_sem);
294 void memcg_put_cache_ids(void)
296 up_read(&memcg_cache_ids_sem);
300 * MIN_SIZE is different than 1, because we would like to avoid going through
301 * the alloc/free process all the time. In a small machine, 4 kmem-limited
302 * cgroups is a reasonable guess. In the future, it could be a parameter or
303 * tunable, but that is strictly not necessary.
305 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
306 * this constant directly from cgroup, but it is understandable that this is
307 * better kept as an internal representation in cgroup.c. In any case, the
308 * cgrp_id space is not getting any smaller, and we don't have to necessarily
309 * increase ours as well if it increases.
311 #define MEMCG_CACHES_MIN_SIZE 4
312 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
315 * A lot of the calls to the cache allocation functions are expected to be
316 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
317 * conditional to this static branch, we'll have to allow modules that does
318 * kmem_cache_alloc and the such to see this symbol as well
320 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
321 EXPORT_SYMBOL(memcg_kmem_enabled_key);
323 struct workqueue_struct *memcg_kmem_cache_wq;
325 static int memcg_shrinker_map_size;
326 static DEFINE_MUTEX(memcg_shrinker_map_mutex);
328 static void memcg_free_shrinker_map_rcu(struct rcu_head *head)
330 kvfree(container_of(head, struct memcg_shrinker_map, rcu));
333 static int memcg_expand_one_shrinker_map(struct mem_cgroup *memcg,
334 int size, int old_size)
336 struct memcg_shrinker_map *new, *old;
339 lockdep_assert_held(&memcg_shrinker_map_mutex);
342 old = rcu_dereference_protected(
343 mem_cgroup_nodeinfo(memcg, nid)->shrinker_map, true);
344 /* Not yet online memcg */
348 new = kvmalloc(sizeof(*new) + size, GFP_KERNEL);
352 /* Set all old bits, clear all new bits */
353 memset(new->map, (int)0xff, old_size);
354 memset((void *)new->map + old_size, 0, size - old_size);
356 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, new);
357 call_rcu(&old->rcu, memcg_free_shrinker_map_rcu);
363 static void memcg_free_shrinker_maps(struct mem_cgroup *memcg)
365 struct mem_cgroup_per_node *pn;
366 struct memcg_shrinker_map *map;
369 if (mem_cgroup_is_root(memcg))
373 pn = mem_cgroup_nodeinfo(memcg, nid);
374 map = rcu_dereference_protected(pn->shrinker_map, true);
377 rcu_assign_pointer(pn->shrinker_map, NULL);
381 static int memcg_alloc_shrinker_maps(struct mem_cgroup *memcg)
383 struct memcg_shrinker_map *map;
384 int nid, size, ret = 0;
386 if (mem_cgroup_is_root(memcg))
389 mutex_lock(&memcg_shrinker_map_mutex);
390 size = memcg_shrinker_map_size;
392 map = kvzalloc(sizeof(*map) + size, GFP_KERNEL);
394 memcg_free_shrinker_maps(memcg);
398 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, map);
400 mutex_unlock(&memcg_shrinker_map_mutex);
405 int memcg_expand_shrinker_maps(int new_id)
407 int size, old_size, ret = 0;
408 struct mem_cgroup *memcg;
410 size = DIV_ROUND_UP(new_id + 1, BITS_PER_LONG) * sizeof(unsigned long);
411 old_size = memcg_shrinker_map_size;
412 if (size <= old_size)
415 mutex_lock(&memcg_shrinker_map_mutex);
416 if (!root_mem_cgroup)
419 for_each_mem_cgroup(memcg) {
420 if (mem_cgroup_is_root(memcg))
422 ret = memcg_expand_one_shrinker_map(memcg, size, old_size);
428 memcg_shrinker_map_size = size;
429 mutex_unlock(&memcg_shrinker_map_mutex);
433 void memcg_set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id)
435 if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) {
436 struct memcg_shrinker_map *map;
439 map = rcu_dereference(memcg->nodeinfo[nid]->shrinker_map);
440 /* Pairs with smp mb in shrink_slab() */
441 smp_mb__before_atomic();
442 set_bit(shrinker_id, map->map);
447 #else /* CONFIG_MEMCG_KMEM */
448 static int memcg_alloc_shrinker_maps(struct mem_cgroup *memcg)
452 static void memcg_free_shrinker_maps(struct mem_cgroup *memcg) { }
453 #endif /* CONFIG_MEMCG_KMEM */
456 * mem_cgroup_css_from_page - css of the memcg associated with a page
457 * @page: page of interest
459 * If memcg is bound to the default hierarchy, css of the memcg associated
460 * with @page is returned. The returned css remains associated with @page
461 * until it is released.
463 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
466 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
468 struct mem_cgroup *memcg;
470 memcg = page->mem_cgroup;
472 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
473 memcg = root_mem_cgroup;
479 * page_cgroup_ino - return inode number of the memcg a page is charged to
482 * Look up the closest online ancestor of the memory cgroup @page is charged to
483 * and return its inode number or 0 if @page is not charged to any cgroup. It
484 * is safe to call this function without holding a reference to @page.
486 * Note, this function is inherently racy, because there is nothing to prevent
487 * the cgroup inode from getting torn down and potentially reallocated a moment
488 * after page_cgroup_ino() returns, so it only should be used by callers that
489 * do not care (such as procfs interfaces).
491 ino_t page_cgroup_ino(struct page *page)
493 struct mem_cgroup *memcg;
494 unsigned long ino = 0;
497 memcg = READ_ONCE(page->mem_cgroup);
498 while (memcg && !(memcg->css.flags & CSS_ONLINE))
499 memcg = parent_mem_cgroup(memcg);
501 ino = cgroup_ino(memcg->css.cgroup);
506 static struct mem_cgroup_per_node *
507 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
509 int nid = page_to_nid(page);
511 return memcg->nodeinfo[nid];
514 static struct mem_cgroup_tree_per_node *
515 soft_limit_tree_node(int nid)
517 return soft_limit_tree.rb_tree_per_node[nid];
520 static struct mem_cgroup_tree_per_node *
521 soft_limit_tree_from_page(struct page *page)
523 int nid = page_to_nid(page);
525 return soft_limit_tree.rb_tree_per_node[nid];
528 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
529 struct mem_cgroup_tree_per_node *mctz,
530 unsigned long new_usage_in_excess)
532 struct rb_node **p = &mctz->rb_root.rb_node;
533 struct rb_node *parent = NULL;
534 struct mem_cgroup_per_node *mz_node;
535 bool rightmost = true;
540 mz->usage_in_excess = new_usage_in_excess;
541 if (!mz->usage_in_excess)
545 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
547 if (mz->usage_in_excess < mz_node->usage_in_excess) {
553 * We can't avoid mem cgroups that are over their soft
554 * limit by the same amount
556 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
561 mctz->rb_rightmost = &mz->tree_node;
563 rb_link_node(&mz->tree_node, parent, p);
564 rb_insert_color(&mz->tree_node, &mctz->rb_root);
568 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
569 struct mem_cgroup_tree_per_node *mctz)
574 if (&mz->tree_node == mctz->rb_rightmost)
575 mctz->rb_rightmost = rb_prev(&mz->tree_node);
577 rb_erase(&mz->tree_node, &mctz->rb_root);
581 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
582 struct mem_cgroup_tree_per_node *mctz)
586 spin_lock_irqsave(&mctz->lock, flags);
587 __mem_cgroup_remove_exceeded(mz, mctz);
588 spin_unlock_irqrestore(&mctz->lock, flags);
591 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
593 unsigned long nr_pages = page_counter_read(&memcg->memory);
594 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
595 unsigned long excess = 0;
597 if (nr_pages > soft_limit)
598 excess = nr_pages - soft_limit;
603 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
605 unsigned long excess;
606 struct mem_cgroup_per_node *mz;
607 struct mem_cgroup_tree_per_node *mctz;
609 mctz = soft_limit_tree_from_page(page);
613 * Necessary to update all ancestors when hierarchy is used.
614 * because their event counter is not touched.
616 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
617 mz = mem_cgroup_page_nodeinfo(memcg, page);
618 excess = soft_limit_excess(memcg);
620 * We have to update the tree if mz is on RB-tree or
621 * mem is over its softlimit.
623 if (excess || mz->on_tree) {
626 spin_lock_irqsave(&mctz->lock, flags);
627 /* if on-tree, remove it */
629 __mem_cgroup_remove_exceeded(mz, mctz);
631 * Insert again. mz->usage_in_excess will be updated.
632 * If excess is 0, no tree ops.
634 __mem_cgroup_insert_exceeded(mz, mctz, excess);
635 spin_unlock_irqrestore(&mctz->lock, flags);
640 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
642 struct mem_cgroup_tree_per_node *mctz;
643 struct mem_cgroup_per_node *mz;
647 mz = mem_cgroup_nodeinfo(memcg, nid);
648 mctz = soft_limit_tree_node(nid);
650 mem_cgroup_remove_exceeded(mz, mctz);
654 static struct mem_cgroup_per_node *
655 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
657 struct mem_cgroup_per_node *mz;
661 if (!mctz->rb_rightmost)
662 goto done; /* Nothing to reclaim from */
664 mz = rb_entry(mctz->rb_rightmost,
665 struct mem_cgroup_per_node, tree_node);
667 * Remove the node now but someone else can add it back,
668 * we will to add it back at the end of reclaim to its correct
669 * position in the tree.
671 __mem_cgroup_remove_exceeded(mz, mctz);
672 if (!soft_limit_excess(mz->memcg) ||
673 !css_tryget_online(&mz->memcg->css))
679 static struct mem_cgroup_per_node *
680 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
682 struct mem_cgroup_per_node *mz;
684 spin_lock_irq(&mctz->lock);
685 mz = __mem_cgroup_largest_soft_limit_node(mctz);
686 spin_unlock_irq(&mctz->lock);
690 static unsigned long memcg_sum_events(struct mem_cgroup *memcg,
693 return atomic_long_read(&memcg->events[event]);
696 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
698 bool compound, int nr_pages)
701 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
702 * counted as CACHE even if it's on ANON LRU.
705 __mod_memcg_state(memcg, MEMCG_RSS, nr_pages);
707 __mod_memcg_state(memcg, MEMCG_CACHE, nr_pages);
708 if (PageSwapBacked(page))
709 __mod_memcg_state(memcg, NR_SHMEM, nr_pages);
713 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
714 __mod_memcg_state(memcg, MEMCG_RSS_HUGE, nr_pages);
717 /* pagein of a big page is an event. So, ignore page size */
719 __count_memcg_events(memcg, PGPGIN, 1);
721 __count_memcg_events(memcg, PGPGOUT, 1);
722 nr_pages = -nr_pages; /* for event */
725 __this_cpu_add(memcg->stat_cpu->nr_page_events, nr_pages);
728 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
729 int nid, unsigned int lru_mask)
731 struct lruvec *lruvec = mem_cgroup_lruvec(NODE_DATA(nid), memcg);
732 unsigned long nr = 0;
735 VM_BUG_ON((unsigned)nid >= nr_node_ids);
738 if (!(BIT(lru) & lru_mask))
740 nr += lruvec_page_state(lruvec, NR_LRU_BASE + lru);
745 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
746 unsigned int lru_mask)
748 unsigned long nr = 0;
752 if (!(BIT(lru) & lru_mask))
754 nr += memcg_page_state(memcg, NR_LRU_BASE + lru);
759 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
760 enum mem_cgroup_events_target target)
762 unsigned long val, next;
764 val = __this_cpu_read(memcg->stat_cpu->nr_page_events);
765 next = __this_cpu_read(memcg->stat_cpu->targets[target]);
766 /* from time_after() in jiffies.h */
767 if ((long)(next - val) < 0) {
769 case MEM_CGROUP_TARGET_THRESH:
770 next = val + THRESHOLDS_EVENTS_TARGET;
772 case MEM_CGROUP_TARGET_SOFTLIMIT:
773 next = val + SOFTLIMIT_EVENTS_TARGET;
775 case MEM_CGROUP_TARGET_NUMAINFO:
776 next = val + NUMAINFO_EVENTS_TARGET;
781 __this_cpu_write(memcg->stat_cpu->targets[target], next);
788 * Check events in order.
791 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
793 /* threshold event is triggered in finer grain than soft limit */
794 if (unlikely(mem_cgroup_event_ratelimit(memcg,
795 MEM_CGROUP_TARGET_THRESH))) {
797 bool do_numainfo __maybe_unused;
799 do_softlimit = mem_cgroup_event_ratelimit(memcg,
800 MEM_CGROUP_TARGET_SOFTLIMIT);
802 do_numainfo = mem_cgroup_event_ratelimit(memcg,
803 MEM_CGROUP_TARGET_NUMAINFO);
805 mem_cgroup_threshold(memcg);
806 if (unlikely(do_softlimit))
807 mem_cgroup_update_tree(memcg, page);
809 if (unlikely(do_numainfo))
810 atomic_inc(&memcg->numainfo_events);
815 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
818 * mm_update_next_owner() may clear mm->owner to NULL
819 * if it races with swapoff, page migration, etc.
820 * So this can be called with p == NULL.
825 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
827 EXPORT_SYMBOL(mem_cgroup_from_task);
830 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
831 * @mm: mm from which memcg should be extracted. It can be NULL.
833 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
834 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
837 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
839 struct mem_cgroup *memcg;
841 if (mem_cgroup_disabled())
847 * Page cache insertions can happen withou an
848 * actual mm context, e.g. during disk probing
849 * on boot, loopback IO, acct() writes etc.
852 memcg = root_mem_cgroup;
854 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
855 if (unlikely(!memcg))
856 memcg = root_mem_cgroup;
858 } while (!css_tryget_online(&memcg->css));
862 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
865 * get_mem_cgroup_from_page: Obtain a reference on given page's memcg.
866 * @page: page from which memcg should be extracted.
868 * Obtain a reference on page->memcg and returns it if successful. Otherwise
869 * root_mem_cgroup is returned.
871 struct mem_cgroup *get_mem_cgroup_from_page(struct page *page)
873 struct mem_cgroup *memcg = page->mem_cgroup;
875 if (mem_cgroup_disabled())
879 if (!memcg || !css_tryget_online(&memcg->css))
880 memcg = root_mem_cgroup;
884 EXPORT_SYMBOL(get_mem_cgroup_from_page);
887 * If current->active_memcg is non-NULL, do not fallback to current->mm->memcg.
889 static __always_inline struct mem_cgroup *get_mem_cgroup_from_current(void)
891 if (unlikely(current->active_memcg)) {
892 struct mem_cgroup *memcg = root_mem_cgroup;
895 if (css_tryget_online(¤t->active_memcg->css))
896 memcg = current->active_memcg;
900 return get_mem_cgroup_from_mm(current->mm);
904 * mem_cgroup_iter - iterate over memory cgroup hierarchy
905 * @root: hierarchy root
906 * @prev: previously returned memcg, NULL on first invocation
907 * @reclaim: cookie for shared reclaim walks, NULL for full walks
909 * Returns references to children of the hierarchy below @root, or
910 * @root itself, or %NULL after a full round-trip.
912 * Caller must pass the return value in @prev on subsequent
913 * invocations for reference counting, or use mem_cgroup_iter_break()
914 * to cancel a hierarchy walk before the round-trip is complete.
916 * Reclaimers can specify a node and a priority level in @reclaim to
917 * divide up the memcgs in the hierarchy among all concurrent
918 * reclaimers operating on the same node and priority.
920 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
921 struct mem_cgroup *prev,
922 struct mem_cgroup_reclaim_cookie *reclaim)
924 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
925 struct cgroup_subsys_state *css = NULL;
926 struct mem_cgroup *memcg = NULL;
927 struct mem_cgroup *pos = NULL;
929 if (mem_cgroup_disabled())
933 root = root_mem_cgroup;
935 if (prev && !reclaim)
938 if (!root->use_hierarchy && root != root_mem_cgroup) {
947 struct mem_cgroup_per_node *mz;
949 mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
950 iter = &mz->iter[reclaim->priority];
952 if (prev && reclaim->generation != iter->generation)
956 pos = READ_ONCE(iter->position);
957 if (!pos || css_tryget(&pos->css))
960 * css reference reached zero, so iter->position will
961 * be cleared by ->css_released. However, we should not
962 * rely on this happening soon, because ->css_released
963 * is called from a work queue, and by busy-waiting we
964 * might block it. So we clear iter->position right
967 (void)cmpxchg(&iter->position, pos, NULL);
975 css = css_next_descendant_pre(css, &root->css);
978 * Reclaimers share the hierarchy walk, and a
979 * new one might jump in right at the end of
980 * the hierarchy - make sure they see at least
981 * one group and restart from the beginning.
989 * Verify the css and acquire a reference. The root
990 * is provided by the caller, so we know it's alive
991 * and kicking, and don't take an extra reference.
993 memcg = mem_cgroup_from_css(css);
995 if (css == &root->css)
1006 * The position could have already been updated by a competing
1007 * thread, so check that the value hasn't changed since we read
1008 * it to avoid reclaiming from the same cgroup twice.
1010 (void)cmpxchg(&iter->position, pos, memcg);
1018 reclaim->generation = iter->generation;
1024 if (prev && prev != root)
1025 css_put(&prev->css);
1031 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1032 * @root: hierarchy root
1033 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1035 void mem_cgroup_iter_break(struct mem_cgroup *root,
1036 struct mem_cgroup *prev)
1039 root = root_mem_cgroup;
1040 if (prev && prev != root)
1041 css_put(&prev->css);
1044 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1046 struct mem_cgroup *memcg = dead_memcg;
1047 struct mem_cgroup_reclaim_iter *iter;
1048 struct mem_cgroup_per_node *mz;
1052 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
1053 for_each_node(nid) {
1054 mz = mem_cgroup_nodeinfo(memcg, nid);
1055 for (i = 0; i <= DEF_PRIORITY; i++) {
1056 iter = &mz->iter[i];
1057 cmpxchg(&iter->position,
1065 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1066 * @memcg: hierarchy root
1067 * @fn: function to call for each task
1068 * @arg: argument passed to @fn
1070 * This function iterates over tasks attached to @memcg or to any of its
1071 * descendants and calls @fn for each task. If @fn returns a non-zero
1072 * value, the function breaks the iteration loop and returns the value.
1073 * Otherwise, it will iterate over all tasks and return 0.
1075 * This function must not be called for the root memory cgroup.
1077 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1078 int (*fn)(struct task_struct *, void *), void *arg)
1080 struct mem_cgroup *iter;
1083 BUG_ON(memcg == root_mem_cgroup);
1085 for_each_mem_cgroup_tree(iter, memcg) {
1086 struct css_task_iter it;
1087 struct task_struct *task;
1089 css_task_iter_start(&iter->css, 0, &it);
1090 while (!ret && (task = css_task_iter_next(&it)))
1091 ret = fn(task, arg);
1092 css_task_iter_end(&it);
1094 mem_cgroup_iter_break(memcg, iter);
1102 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1104 * @pgdat: pgdat of the page
1106 * This function is only safe when following the LRU page isolation
1107 * and putback protocol: the LRU lock must be held, and the page must
1108 * either be PageLRU() or the caller must have isolated/allocated it.
1110 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct pglist_data *pgdat)
1112 struct mem_cgroup_per_node *mz;
1113 struct mem_cgroup *memcg;
1114 struct lruvec *lruvec;
1116 if (mem_cgroup_disabled()) {
1117 lruvec = &pgdat->lruvec;
1121 memcg = page->mem_cgroup;
1123 * Swapcache readahead pages are added to the LRU - and
1124 * possibly migrated - before they are charged.
1127 memcg = root_mem_cgroup;
1129 mz = mem_cgroup_page_nodeinfo(memcg, page);
1130 lruvec = &mz->lruvec;
1133 * Since a node can be onlined after the mem_cgroup was created,
1134 * we have to be prepared to initialize lruvec->zone here;
1135 * and if offlined then reonlined, we need to reinitialize it.
1137 if (unlikely(lruvec->pgdat != pgdat))
1138 lruvec->pgdat = pgdat;
1143 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1144 * @lruvec: mem_cgroup per zone lru vector
1145 * @lru: index of lru list the page is sitting on
1146 * @zid: zone id of the accounted pages
1147 * @nr_pages: positive when adding or negative when removing
1149 * This function must be called under lru_lock, just before a page is added
1150 * to or just after a page is removed from an lru list (that ordering being
1151 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1153 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1154 int zid, int nr_pages)
1156 struct mem_cgroup_per_node *mz;
1157 unsigned long *lru_size;
1160 if (mem_cgroup_disabled())
1163 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1164 lru_size = &mz->lru_zone_size[zid][lru];
1167 *lru_size += nr_pages;
1170 if (WARN_ONCE(size < 0,
1171 "%s(%p, %d, %d): lru_size %ld\n",
1172 __func__, lruvec, lru, nr_pages, size)) {
1178 *lru_size += nr_pages;
1181 bool task_in_mem_cgroup(struct task_struct *task, struct mem_cgroup *memcg)
1183 struct mem_cgroup *task_memcg;
1184 struct task_struct *p;
1187 p = find_lock_task_mm(task);
1189 task_memcg = get_mem_cgroup_from_mm(p->mm);
1193 * All threads may have already detached their mm's, but the oom
1194 * killer still needs to detect if they have already been oom
1195 * killed to prevent needlessly killing additional tasks.
1198 task_memcg = mem_cgroup_from_task(task);
1199 css_get(&task_memcg->css);
1202 ret = mem_cgroup_is_descendant(task_memcg, memcg);
1203 css_put(&task_memcg->css);
1208 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1209 * @memcg: the memory cgroup
1211 * Returns the maximum amount of memory @mem can be charged with, in
1214 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1216 unsigned long margin = 0;
1217 unsigned long count;
1218 unsigned long limit;
1220 count = page_counter_read(&memcg->memory);
1221 limit = READ_ONCE(memcg->memory.max);
1223 margin = limit - count;
1225 if (do_memsw_account()) {
1226 count = page_counter_read(&memcg->memsw);
1227 limit = READ_ONCE(memcg->memsw.max);
1229 margin = min(margin, limit - count);
1238 * A routine for checking "mem" is under move_account() or not.
1240 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1241 * moving cgroups. This is for waiting at high-memory pressure
1244 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1246 struct mem_cgroup *from;
1247 struct mem_cgroup *to;
1250 * Unlike task_move routines, we access mc.to, mc.from not under
1251 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1253 spin_lock(&mc.lock);
1259 ret = mem_cgroup_is_descendant(from, memcg) ||
1260 mem_cgroup_is_descendant(to, memcg);
1262 spin_unlock(&mc.lock);
1266 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1268 if (mc.moving_task && current != mc.moving_task) {
1269 if (mem_cgroup_under_move(memcg)) {
1271 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1272 /* moving charge context might have finished. */
1275 finish_wait(&mc.waitq, &wait);
1282 static const unsigned int memcg1_stats[] = {
1293 static const char *const memcg1_stat_names[] = {
1304 #define K(x) ((x) << (PAGE_SHIFT-10))
1306 * mem_cgroup_print_oom_context: Print OOM information relevant to
1307 * memory controller.
1308 * @memcg: The memory cgroup that went over limit
1309 * @p: Task that is going to be killed
1311 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1314 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1319 pr_cont(",oom_memcg=");
1320 pr_cont_cgroup_path(memcg->css.cgroup);
1322 pr_cont(",global_oom");
1324 pr_cont(",task_memcg=");
1325 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1331 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1332 * memory controller.
1333 * @memcg: The memory cgroup that went over limit
1335 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1337 struct mem_cgroup *iter;
1340 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1341 K((u64)page_counter_read(&memcg->memory)),
1342 K((u64)memcg->memory.max), memcg->memory.failcnt);
1343 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1344 K((u64)page_counter_read(&memcg->memsw)),
1345 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1346 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1347 K((u64)page_counter_read(&memcg->kmem)),
1348 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1350 for_each_mem_cgroup_tree(iter, memcg) {
1351 pr_info("Memory cgroup stats for ");
1352 pr_cont_cgroup_path(iter->css.cgroup);
1355 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
1356 if (memcg1_stats[i] == MEMCG_SWAP && !do_swap_account)
1358 pr_cont(" %s:%luKB", memcg1_stat_names[i],
1359 K(memcg_page_state(iter, memcg1_stats[i])));
1362 for (i = 0; i < NR_LRU_LISTS; i++)
1363 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1364 K(memcg_page_state(iter, NR_LRU_BASE + i)));
1371 * Return the memory (and swap, if configured) limit for a memcg.
1373 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1377 max = memcg->memory.max;
1378 if (mem_cgroup_swappiness(memcg)) {
1379 unsigned long memsw_max;
1380 unsigned long swap_max;
1382 memsw_max = memcg->memsw.max;
1383 swap_max = memcg->swap.max;
1384 swap_max = min(swap_max, (unsigned long)total_swap_pages);
1385 max = min(max + swap_max, memsw_max);
1390 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1393 struct oom_control oc = {
1397 .gfp_mask = gfp_mask,
1402 if (mutex_lock_killable(&oom_lock))
1405 * A few threads which were not waiting at mutex_lock_killable() can
1406 * fail to bail out. Therefore, check again after holding oom_lock.
1408 ret = should_force_charge() || out_of_memory(&oc);
1409 mutex_unlock(&oom_lock);
1413 #if MAX_NUMNODES > 1
1416 * test_mem_cgroup_node_reclaimable
1417 * @memcg: the target memcg
1418 * @nid: the node ID to be checked.
1419 * @noswap : specify true here if the user wants flle only information.
1421 * This function returns whether the specified memcg contains any
1422 * reclaimable pages on a node. Returns true if there are any reclaimable
1423 * pages in the node.
1425 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1426 int nid, bool noswap)
1428 struct lruvec *lruvec = mem_cgroup_lruvec(NODE_DATA(nid), memcg);
1430 if (lruvec_page_state(lruvec, NR_INACTIVE_FILE) ||
1431 lruvec_page_state(lruvec, NR_ACTIVE_FILE))
1433 if (noswap || !total_swap_pages)
1435 if (lruvec_page_state(lruvec, NR_INACTIVE_ANON) ||
1436 lruvec_page_state(lruvec, NR_ACTIVE_ANON))
1443 * Always updating the nodemask is not very good - even if we have an empty
1444 * list or the wrong list here, we can start from some node and traverse all
1445 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1448 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1452 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1453 * pagein/pageout changes since the last update.
1455 if (!atomic_read(&memcg->numainfo_events))
1457 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1460 /* make a nodemask where this memcg uses memory from */
1461 memcg->scan_nodes = node_states[N_MEMORY];
1463 for_each_node_mask(nid, node_states[N_MEMORY]) {
1465 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1466 node_clear(nid, memcg->scan_nodes);
1469 atomic_set(&memcg->numainfo_events, 0);
1470 atomic_set(&memcg->numainfo_updating, 0);
1474 * Selecting a node where we start reclaim from. Because what we need is just
1475 * reducing usage counter, start from anywhere is O,K. Considering
1476 * memory reclaim from current node, there are pros. and cons.
1478 * Freeing memory from current node means freeing memory from a node which
1479 * we'll use or we've used. So, it may make LRU bad. And if several threads
1480 * hit limits, it will see a contention on a node. But freeing from remote
1481 * node means more costs for memory reclaim because of memory latency.
1483 * Now, we use round-robin. Better algorithm is welcomed.
1485 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1489 mem_cgroup_may_update_nodemask(memcg);
1490 node = memcg->last_scanned_node;
1492 node = next_node_in(node, memcg->scan_nodes);
1494 * mem_cgroup_may_update_nodemask might have seen no reclaimmable pages
1495 * last time it really checked all the LRUs due to rate limiting.
1496 * Fallback to the current node in that case for simplicity.
1498 if (unlikely(node == MAX_NUMNODES))
1499 node = numa_node_id();
1501 memcg->last_scanned_node = node;
1505 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1511 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1514 unsigned long *total_scanned)
1516 struct mem_cgroup *victim = NULL;
1519 unsigned long excess;
1520 unsigned long nr_scanned;
1521 struct mem_cgroup_reclaim_cookie reclaim = {
1526 excess = soft_limit_excess(root_memcg);
1529 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1534 * If we have not been able to reclaim
1535 * anything, it might because there are
1536 * no reclaimable pages under this hierarchy
1541 * We want to do more targeted reclaim.
1542 * excess >> 2 is not to excessive so as to
1543 * reclaim too much, nor too less that we keep
1544 * coming back to reclaim from this cgroup
1546 if (total >= (excess >> 2) ||
1547 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1552 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1553 pgdat, &nr_scanned);
1554 *total_scanned += nr_scanned;
1555 if (!soft_limit_excess(root_memcg))
1558 mem_cgroup_iter_break(root_memcg, victim);
1562 #ifdef CONFIG_LOCKDEP
1563 static struct lockdep_map memcg_oom_lock_dep_map = {
1564 .name = "memcg_oom_lock",
1568 static DEFINE_SPINLOCK(memcg_oom_lock);
1571 * Check OOM-Killer is already running under our hierarchy.
1572 * If someone is running, return false.
1574 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1576 struct mem_cgroup *iter, *failed = NULL;
1578 spin_lock(&memcg_oom_lock);
1580 for_each_mem_cgroup_tree(iter, memcg) {
1581 if (iter->oom_lock) {
1583 * this subtree of our hierarchy is already locked
1584 * so we cannot give a lock.
1587 mem_cgroup_iter_break(memcg, iter);
1590 iter->oom_lock = true;
1595 * OK, we failed to lock the whole subtree so we have
1596 * to clean up what we set up to the failing subtree
1598 for_each_mem_cgroup_tree(iter, memcg) {
1599 if (iter == failed) {
1600 mem_cgroup_iter_break(memcg, iter);
1603 iter->oom_lock = false;
1606 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1608 spin_unlock(&memcg_oom_lock);
1613 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1615 struct mem_cgroup *iter;
1617 spin_lock(&memcg_oom_lock);
1618 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1619 for_each_mem_cgroup_tree(iter, memcg)
1620 iter->oom_lock = false;
1621 spin_unlock(&memcg_oom_lock);
1624 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1626 struct mem_cgroup *iter;
1628 spin_lock(&memcg_oom_lock);
1629 for_each_mem_cgroup_tree(iter, memcg)
1631 spin_unlock(&memcg_oom_lock);
1634 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1636 struct mem_cgroup *iter;
1639 * When a new child is created while the hierarchy is under oom,
1640 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1642 spin_lock(&memcg_oom_lock);
1643 for_each_mem_cgroup_tree(iter, memcg)
1644 if (iter->under_oom > 0)
1646 spin_unlock(&memcg_oom_lock);
1649 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1651 struct oom_wait_info {
1652 struct mem_cgroup *memcg;
1653 wait_queue_entry_t wait;
1656 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1657 unsigned mode, int sync, void *arg)
1659 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1660 struct mem_cgroup *oom_wait_memcg;
1661 struct oom_wait_info *oom_wait_info;
1663 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1664 oom_wait_memcg = oom_wait_info->memcg;
1666 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1667 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1669 return autoremove_wake_function(wait, mode, sync, arg);
1672 static void memcg_oom_recover(struct mem_cgroup *memcg)
1675 * For the following lockless ->under_oom test, the only required
1676 * guarantee is that it must see the state asserted by an OOM when
1677 * this function is called as a result of userland actions
1678 * triggered by the notification of the OOM. This is trivially
1679 * achieved by invoking mem_cgroup_mark_under_oom() before
1680 * triggering notification.
1682 if (memcg && memcg->under_oom)
1683 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1693 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1695 enum oom_status ret;
1698 if (order > PAGE_ALLOC_COSTLY_ORDER)
1701 memcg_memory_event(memcg, MEMCG_OOM);
1704 * We are in the middle of the charge context here, so we
1705 * don't want to block when potentially sitting on a callstack
1706 * that holds all kinds of filesystem and mm locks.
1708 * cgroup1 allows disabling the OOM killer and waiting for outside
1709 * handling until the charge can succeed; remember the context and put
1710 * the task to sleep at the end of the page fault when all locks are
1713 * On the other hand, in-kernel OOM killer allows for an async victim
1714 * memory reclaim (oom_reaper) and that means that we are not solely
1715 * relying on the oom victim to make a forward progress and we can
1716 * invoke the oom killer here.
1718 * Please note that mem_cgroup_out_of_memory might fail to find a
1719 * victim and then we have to bail out from the charge path.
1721 if (memcg->oom_kill_disable) {
1722 if (!current->in_user_fault)
1724 css_get(&memcg->css);
1725 current->memcg_in_oom = memcg;
1726 current->memcg_oom_gfp_mask = mask;
1727 current->memcg_oom_order = order;
1732 mem_cgroup_mark_under_oom(memcg);
1734 locked = mem_cgroup_oom_trylock(memcg);
1737 mem_cgroup_oom_notify(memcg);
1739 mem_cgroup_unmark_under_oom(memcg);
1740 if (mem_cgroup_out_of_memory(memcg, mask, order))
1746 mem_cgroup_oom_unlock(memcg);
1752 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1753 * @handle: actually kill/wait or just clean up the OOM state
1755 * This has to be called at the end of a page fault if the memcg OOM
1756 * handler was enabled.
1758 * Memcg supports userspace OOM handling where failed allocations must
1759 * sleep on a waitqueue until the userspace task resolves the
1760 * situation. Sleeping directly in the charge context with all kinds
1761 * of locks held is not a good idea, instead we remember an OOM state
1762 * in the task and mem_cgroup_oom_synchronize() has to be called at
1763 * the end of the page fault to complete the OOM handling.
1765 * Returns %true if an ongoing memcg OOM situation was detected and
1766 * completed, %false otherwise.
1768 bool mem_cgroup_oom_synchronize(bool handle)
1770 struct mem_cgroup *memcg = current->memcg_in_oom;
1771 struct oom_wait_info owait;
1774 /* OOM is global, do not handle */
1781 owait.memcg = memcg;
1782 owait.wait.flags = 0;
1783 owait.wait.func = memcg_oom_wake_function;
1784 owait.wait.private = current;
1785 INIT_LIST_HEAD(&owait.wait.entry);
1787 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1788 mem_cgroup_mark_under_oom(memcg);
1790 locked = mem_cgroup_oom_trylock(memcg);
1793 mem_cgroup_oom_notify(memcg);
1795 if (locked && !memcg->oom_kill_disable) {
1796 mem_cgroup_unmark_under_oom(memcg);
1797 finish_wait(&memcg_oom_waitq, &owait.wait);
1798 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1799 current->memcg_oom_order);
1802 mem_cgroup_unmark_under_oom(memcg);
1803 finish_wait(&memcg_oom_waitq, &owait.wait);
1807 mem_cgroup_oom_unlock(memcg);
1809 * There is no guarantee that an OOM-lock contender
1810 * sees the wakeups triggered by the OOM kill
1811 * uncharges. Wake any sleepers explicitely.
1813 memcg_oom_recover(memcg);
1816 current->memcg_in_oom = NULL;
1817 css_put(&memcg->css);
1822 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
1823 * @victim: task to be killed by the OOM killer
1824 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
1826 * Returns a pointer to a memory cgroup, which has to be cleaned up
1827 * by killing all belonging OOM-killable tasks.
1829 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
1831 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
1832 struct mem_cgroup *oom_domain)
1834 struct mem_cgroup *oom_group = NULL;
1835 struct mem_cgroup *memcg;
1837 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
1841 oom_domain = root_mem_cgroup;
1845 memcg = mem_cgroup_from_task(victim);
1846 if (memcg == root_mem_cgroup)
1850 * Traverse the memory cgroup hierarchy from the victim task's
1851 * cgroup up to the OOMing cgroup (or root) to find the
1852 * highest-level memory cgroup with oom.group set.
1854 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
1855 if (memcg->oom_group)
1858 if (memcg == oom_domain)
1863 css_get(&oom_group->css);
1870 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
1872 pr_info("Tasks in ");
1873 pr_cont_cgroup_path(memcg->css.cgroup);
1874 pr_cont(" are going to be killed due to memory.oom.group set\n");
1878 * lock_page_memcg - lock a page->mem_cgroup binding
1881 * This function protects unlocked LRU pages from being moved to
1884 * It ensures lifetime of the returned memcg. Caller is responsible
1885 * for the lifetime of the page; __unlock_page_memcg() is available
1886 * when @page might get freed inside the locked section.
1888 struct mem_cgroup *lock_page_memcg(struct page *page)
1890 struct mem_cgroup *memcg;
1891 unsigned long flags;
1894 * The RCU lock is held throughout the transaction. The fast
1895 * path can get away without acquiring the memcg->move_lock
1896 * because page moving starts with an RCU grace period.
1898 * The RCU lock also protects the memcg from being freed when
1899 * the page state that is going to change is the only thing
1900 * preventing the page itself from being freed. E.g. writeback
1901 * doesn't hold a page reference and relies on PG_writeback to
1902 * keep off truncation, migration and so forth.
1906 if (mem_cgroup_disabled())
1909 memcg = page->mem_cgroup;
1910 if (unlikely(!memcg))
1913 if (atomic_read(&memcg->moving_account) <= 0)
1916 spin_lock_irqsave(&memcg->move_lock, flags);
1917 if (memcg != page->mem_cgroup) {
1918 spin_unlock_irqrestore(&memcg->move_lock, flags);
1923 * When charge migration first begins, we can have locked and
1924 * unlocked page stat updates happening concurrently. Track
1925 * the task who has the lock for unlock_page_memcg().
1927 memcg->move_lock_task = current;
1928 memcg->move_lock_flags = flags;
1932 EXPORT_SYMBOL(lock_page_memcg);
1935 * __unlock_page_memcg - unlock and unpin a memcg
1938 * Unlock and unpin a memcg returned by lock_page_memcg().
1940 void __unlock_page_memcg(struct mem_cgroup *memcg)
1942 if (memcg && memcg->move_lock_task == current) {
1943 unsigned long flags = memcg->move_lock_flags;
1945 memcg->move_lock_task = NULL;
1946 memcg->move_lock_flags = 0;
1948 spin_unlock_irqrestore(&memcg->move_lock, flags);
1955 * unlock_page_memcg - unlock a page->mem_cgroup binding
1958 void unlock_page_memcg(struct page *page)
1960 __unlock_page_memcg(page->mem_cgroup);
1962 EXPORT_SYMBOL(unlock_page_memcg);
1964 struct memcg_stock_pcp {
1965 struct mem_cgroup *cached; /* this never be root cgroup */
1966 unsigned int nr_pages;
1967 struct work_struct work;
1968 unsigned long flags;
1969 #define FLUSHING_CACHED_CHARGE 0
1971 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1972 static DEFINE_MUTEX(percpu_charge_mutex);
1975 * consume_stock: Try to consume stocked charge on this cpu.
1976 * @memcg: memcg to consume from.
1977 * @nr_pages: how many pages to charge.
1979 * The charges will only happen if @memcg matches the current cpu's memcg
1980 * stock, and at least @nr_pages are available in that stock. Failure to
1981 * service an allocation will refill the stock.
1983 * returns true if successful, false otherwise.
1985 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1987 struct memcg_stock_pcp *stock;
1988 unsigned long flags;
1991 if (nr_pages > MEMCG_CHARGE_BATCH)
1994 local_irq_save(flags);
1996 stock = this_cpu_ptr(&memcg_stock);
1997 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
1998 stock->nr_pages -= nr_pages;
2002 local_irq_restore(flags);
2008 * Returns stocks cached in percpu and reset cached information.
2010 static void drain_stock(struct memcg_stock_pcp *stock)
2012 struct mem_cgroup *old = stock->cached;
2014 if (stock->nr_pages) {
2015 page_counter_uncharge(&old->memory, stock->nr_pages);
2016 if (do_memsw_account())
2017 page_counter_uncharge(&old->memsw, stock->nr_pages);
2018 css_put_many(&old->css, stock->nr_pages);
2019 stock->nr_pages = 0;
2021 stock->cached = NULL;
2024 static void drain_local_stock(struct work_struct *dummy)
2026 struct memcg_stock_pcp *stock;
2027 unsigned long flags;
2030 * The only protection from memory hotplug vs. drain_stock races is
2031 * that we always operate on local CPU stock here with IRQ disabled
2033 local_irq_save(flags);
2035 stock = this_cpu_ptr(&memcg_stock);
2037 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2039 local_irq_restore(flags);
2043 * Cache charges(val) to local per_cpu area.
2044 * This will be consumed by consume_stock() function, later.
2046 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2048 struct memcg_stock_pcp *stock;
2049 unsigned long flags;
2051 local_irq_save(flags);
2053 stock = this_cpu_ptr(&memcg_stock);
2054 if (stock->cached != memcg) { /* reset if necessary */
2056 stock->cached = memcg;
2058 stock->nr_pages += nr_pages;
2060 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2063 local_irq_restore(flags);
2067 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2068 * of the hierarchy under it.
2070 static void drain_all_stock(struct mem_cgroup *root_memcg)
2074 /* If someone's already draining, avoid adding running more workers. */
2075 if (!mutex_trylock(&percpu_charge_mutex))
2078 * Notify other cpus that system-wide "drain" is running
2079 * We do not care about races with the cpu hotplug because cpu down
2080 * as well as workers from this path always operate on the local
2081 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2084 for_each_online_cpu(cpu) {
2085 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2086 struct mem_cgroup *memcg;
2088 memcg = stock->cached;
2089 if (!memcg || !stock->nr_pages || !css_tryget(&memcg->css))
2091 if (!mem_cgroup_is_descendant(memcg, root_memcg)) {
2092 css_put(&memcg->css);
2095 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2097 drain_local_stock(&stock->work);
2099 schedule_work_on(cpu, &stock->work);
2101 css_put(&memcg->css);
2104 mutex_unlock(&percpu_charge_mutex);
2107 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2109 struct memcg_stock_pcp *stock;
2110 struct mem_cgroup *memcg;
2112 stock = &per_cpu(memcg_stock, cpu);
2115 for_each_mem_cgroup(memcg) {
2118 for (i = 0; i < MEMCG_NR_STAT; i++) {
2122 x = this_cpu_xchg(memcg->stat_cpu->count[i], 0);
2124 atomic_long_add(x, &memcg->stat[i]);
2126 if (i >= NR_VM_NODE_STAT_ITEMS)
2129 for_each_node(nid) {
2130 struct mem_cgroup_per_node *pn;
2132 pn = mem_cgroup_nodeinfo(memcg, nid);
2133 x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0);
2135 atomic_long_add(x, &pn->lruvec_stat[i]);
2139 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
2142 x = this_cpu_xchg(memcg->stat_cpu->events[i], 0);
2144 atomic_long_add(x, &memcg->events[i]);
2151 static void reclaim_high(struct mem_cgroup *memcg,
2152 unsigned int nr_pages,
2156 if (page_counter_read(&memcg->memory) <= memcg->high)
2158 memcg_memory_event(memcg, MEMCG_HIGH);
2159 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
2160 } while ((memcg = parent_mem_cgroup(memcg)));
2163 static void high_work_func(struct work_struct *work)
2165 struct mem_cgroup *memcg;
2167 memcg = container_of(work, struct mem_cgroup, high_work);
2168 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2172 * Scheduled by try_charge() to be executed from the userland return path
2173 * and reclaims memory over the high limit.
2175 void mem_cgroup_handle_over_high(void)
2177 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2178 struct mem_cgroup *memcg;
2180 if (likely(!nr_pages))
2183 memcg = get_mem_cgroup_from_mm(current->mm);
2184 reclaim_high(memcg, nr_pages, GFP_KERNEL);
2185 css_put(&memcg->css);
2186 current->memcg_nr_pages_over_high = 0;
2189 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2190 unsigned int nr_pages)
2192 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2193 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2194 struct mem_cgroup *mem_over_limit;
2195 struct page_counter *counter;
2196 unsigned long nr_reclaimed;
2197 bool may_swap = true;
2198 bool drained = false;
2200 enum oom_status oom_status;
2202 if (mem_cgroup_is_root(memcg))
2205 if (consume_stock(memcg, nr_pages))
2208 if (!do_memsw_account() ||
2209 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2210 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2212 if (do_memsw_account())
2213 page_counter_uncharge(&memcg->memsw, batch);
2214 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2216 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2220 if (batch > nr_pages) {
2226 * Unlike in global OOM situations, memcg is not in a physical
2227 * memory shortage. Allow dying and OOM-killed tasks to
2228 * bypass the last charges so that they can exit quickly and
2229 * free their memory.
2231 if (unlikely(should_force_charge()))
2235 * Prevent unbounded recursion when reclaim operations need to
2236 * allocate memory. This might exceed the limits temporarily,
2237 * but we prefer facilitating memory reclaim and getting back
2238 * under the limit over triggering OOM kills in these cases.
2240 if (unlikely(current->flags & PF_MEMALLOC))
2243 if (unlikely(task_in_memcg_oom(current)))
2246 if (!gfpflags_allow_blocking(gfp_mask))
2249 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2251 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2252 gfp_mask, may_swap);
2254 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2258 drain_all_stock(mem_over_limit);
2263 if (gfp_mask & __GFP_NORETRY)
2266 * Even though the limit is exceeded at this point, reclaim
2267 * may have been able to free some pages. Retry the charge
2268 * before killing the task.
2270 * Only for regular pages, though: huge pages are rather
2271 * unlikely to succeed so close to the limit, and we fall back
2272 * to regular pages anyway in case of failure.
2274 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2277 * At task move, charge accounts can be doubly counted. So, it's
2278 * better to wait until the end of task_move if something is going on.
2280 if (mem_cgroup_wait_acct_move(mem_over_limit))
2286 if (gfp_mask & __GFP_RETRY_MAYFAIL && oomed)
2289 if (gfp_mask & __GFP_NOFAIL)
2292 if (fatal_signal_pending(current))
2296 * keep retrying as long as the memcg oom killer is able to make
2297 * a forward progress or bypass the charge if the oom killer
2298 * couldn't make any progress.
2300 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2301 get_order(nr_pages * PAGE_SIZE));
2302 switch (oom_status) {
2304 nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2313 if (!(gfp_mask & __GFP_NOFAIL))
2317 * The allocation either can't fail or will lead to more memory
2318 * being freed very soon. Allow memory usage go over the limit
2319 * temporarily by force charging it.
2321 page_counter_charge(&memcg->memory, nr_pages);
2322 if (do_memsw_account())
2323 page_counter_charge(&memcg->memsw, nr_pages);
2324 css_get_many(&memcg->css, nr_pages);
2329 css_get_many(&memcg->css, batch);
2330 if (batch > nr_pages)
2331 refill_stock(memcg, batch - nr_pages);
2334 * If the hierarchy is above the normal consumption range, schedule
2335 * reclaim on returning to userland. We can perform reclaim here
2336 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2337 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2338 * not recorded as it most likely matches current's and won't
2339 * change in the meantime. As high limit is checked again before
2340 * reclaim, the cost of mismatch is negligible.
2343 if (page_counter_read(&memcg->memory) > memcg->high) {
2344 /* Don't bother a random interrupted task */
2345 if (in_interrupt()) {
2346 schedule_work(&memcg->high_work);
2349 current->memcg_nr_pages_over_high += batch;
2350 set_notify_resume(current);
2353 } while ((memcg = parent_mem_cgroup(memcg)));
2358 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2360 if (mem_cgroup_is_root(memcg))
2363 page_counter_uncharge(&memcg->memory, nr_pages);
2364 if (do_memsw_account())
2365 page_counter_uncharge(&memcg->memsw, nr_pages);
2367 css_put_many(&memcg->css, nr_pages);
2370 static void lock_page_lru(struct page *page, int *isolated)
2372 pg_data_t *pgdat = page_pgdat(page);
2374 spin_lock_irq(&pgdat->lru_lock);
2375 if (PageLRU(page)) {
2376 struct lruvec *lruvec;
2378 lruvec = mem_cgroup_page_lruvec(page, pgdat);
2380 del_page_from_lru_list(page, lruvec, page_lru(page));
2386 static void unlock_page_lru(struct page *page, int isolated)
2388 pg_data_t *pgdat = page_pgdat(page);
2391 struct lruvec *lruvec;
2393 lruvec = mem_cgroup_page_lruvec(page, pgdat);
2394 VM_BUG_ON_PAGE(PageLRU(page), page);
2396 add_page_to_lru_list(page, lruvec, page_lru(page));
2398 spin_unlock_irq(&pgdat->lru_lock);
2401 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2406 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2409 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2410 * may already be on some other mem_cgroup's LRU. Take care of it.
2413 lock_page_lru(page, &isolated);
2416 * Nobody should be changing or seriously looking at
2417 * page->mem_cgroup at this point:
2419 * - the page is uncharged
2421 * - the page is off-LRU
2423 * - an anonymous fault has exclusive page access, except for
2424 * a locked page table
2426 * - a page cache insertion, a swapin fault, or a migration
2427 * have the page locked
2429 page->mem_cgroup = memcg;
2432 unlock_page_lru(page, isolated);
2435 #ifdef CONFIG_MEMCG_KMEM
2436 static int memcg_alloc_cache_id(void)
2441 id = ida_simple_get(&memcg_cache_ida,
2442 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2446 if (id < memcg_nr_cache_ids)
2450 * There's no space for the new id in memcg_caches arrays,
2451 * so we have to grow them.
2453 down_write(&memcg_cache_ids_sem);
2455 size = 2 * (id + 1);
2456 if (size < MEMCG_CACHES_MIN_SIZE)
2457 size = MEMCG_CACHES_MIN_SIZE;
2458 else if (size > MEMCG_CACHES_MAX_SIZE)
2459 size = MEMCG_CACHES_MAX_SIZE;
2461 err = memcg_update_all_caches(size);
2463 err = memcg_update_all_list_lrus(size);
2465 memcg_nr_cache_ids = size;
2467 up_write(&memcg_cache_ids_sem);
2470 ida_simple_remove(&memcg_cache_ida, id);
2476 static void memcg_free_cache_id(int id)
2478 ida_simple_remove(&memcg_cache_ida, id);
2481 struct memcg_kmem_cache_create_work {
2482 struct mem_cgroup *memcg;
2483 struct kmem_cache *cachep;
2484 struct work_struct work;
2487 static void memcg_kmem_cache_create_func(struct work_struct *w)
2489 struct memcg_kmem_cache_create_work *cw =
2490 container_of(w, struct memcg_kmem_cache_create_work, work);
2491 struct mem_cgroup *memcg = cw->memcg;
2492 struct kmem_cache *cachep = cw->cachep;
2494 memcg_create_kmem_cache(memcg, cachep);
2496 css_put(&memcg->css);
2501 * Enqueue the creation of a per-memcg kmem_cache.
2503 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2504 struct kmem_cache *cachep)
2506 struct memcg_kmem_cache_create_work *cw;
2508 cw = kmalloc(sizeof(*cw), GFP_NOWAIT | __GFP_NOWARN);
2512 css_get(&memcg->css);
2515 cw->cachep = cachep;
2516 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2518 queue_work(memcg_kmem_cache_wq, &cw->work);
2521 static inline bool memcg_kmem_bypass(void)
2523 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
2529 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2530 * @cachep: the original global kmem cache
2532 * Return the kmem_cache we're supposed to use for a slab allocation.
2533 * We try to use the current memcg's version of the cache.
2535 * If the cache does not exist yet, if we are the first user of it, we
2536 * create it asynchronously in a workqueue and let the current allocation
2537 * go through with the original cache.
2539 * This function takes a reference to the cache it returns to assure it
2540 * won't get destroyed while we are working with it. Once the caller is
2541 * done with it, memcg_kmem_put_cache() must be called to release the
2544 struct kmem_cache *memcg_kmem_get_cache(struct kmem_cache *cachep)
2546 struct mem_cgroup *memcg;
2547 struct kmem_cache *memcg_cachep;
2550 VM_BUG_ON(!is_root_cache(cachep));
2552 if (memcg_kmem_bypass())
2555 memcg = get_mem_cgroup_from_current();
2556 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2560 memcg_cachep = cache_from_memcg_idx(cachep, kmemcg_id);
2561 if (likely(memcg_cachep))
2562 return memcg_cachep;
2565 * If we are in a safe context (can wait, and not in interrupt
2566 * context), we could be be predictable and return right away.
2567 * This would guarantee that the allocation being performed
2568 * already belongs in the new cache.
2570 * However, there are some clashes that can arrive from locking.
2571 * For instance, because we acquire the slab_mutex while doing
2572 * memcg_create_kmem_cache, this means no further allocation
2573 * could happen with the slab_mutex held. So it's better to
2576 memcg_schedule_kmem_cache_create(memcg, cachep);
2578 css_put(&memcg->css);
2583 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
2584 * @cachep: the cache returned by memcg_kmem_get_cache
2586 void memcg_kmem_put_cache(struct kmem_cache *cachep)
2588 if (!is_root_cache(cachep))
2589 css_put(&cachep->memcg_params.memcg->css);
2593 * __memcg_kmem_charge_memcg: charge a kmem page
2594 * @page: page to charge
2595 * @gfp: reclaim mode
2596 * @order: allocation order
2597 * @memcg: memory cgroup to charge
2599 * Returns 0 on success, an error code on failure.
2601 int __memcg_kmem_charge_memcg(struct page *page, gfp_t gfp, int order,
2602 struct mem_cgroup *memcg)
2604 unsigned int nr_pages = 1 << order;
2605 struct page_counter *counter;
2608 ret = try_charge(memcg, gfp, nr_pages);
2612 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
2613 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
2614 cancel_charge(memcg, nr_pages);
2618 page->mem_cgroup = memcg;
2624 * __memcg_kmem_charge: charge a kmem page to the current memory cgroup
2625 * @page: page to charge
2626 * @gfp: reclaim mode
2627 * @order: allocation order
2629 * Returns 0 on success, an error code on failure.
2631 int __memcg_kmem_charge(struct page *page, gfp_t gfp, int order)
2633 struct mem_cgroup *memcg;
2636 if (memcg_kmem_bypass())
2639 memcg = get_mem_cgroup_from_current();
2640 if (!mem_cgroup_is_root(memcg)) {
2641 ret = __memcg_kmem_charge_memcg(page, gfp, order, memcg);
2643 __SetPageKmemcg(page);
2645 css_put(&memcg->css);
2649 * __memcg_kmem_uncharge: uncharge a kmem page
2650 * @page: page to uncharge
2651 * @order: allocation order
2653 void __memcg_kmem_uncharge(struct page *page, int order)
2655 struct mem_cgroup *memcg = page->mem_cgroup;
2656 unsigned int nr_pages = 1 << order;
2661 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2663 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2664 page_counter_uncharge(&memcg->kmem, nr_pages);
2666 page_counter_uncharge(&memcg->memory, nr_pages);
2667 if (do_memsw_account())
2668 page_counter_uncharge(&memcg->memsw, nr_pages);
2670 page->mem_cgroup = NULL;
2672 /* slab pages do not have PageKmemcg flag set */
2673 if (PageKmemcg(page))
2674 __ClearPageKmemcg(page);
2676 css_put_many(&memcg->css, nr_pages);
2678 #endif /* CONFIG_MEMCG_KMEM */
2680 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2683 * Because tail pages are not marked as "used", set it. We're under
2684 * pgdat->lru_lock and migration entries setup in all page mappings.
2686 void mem_cgroup_split_huge_fixup(struct page *head)
2690 if (mem_cgroup_disabled())
2693 for (i = 1; i < HPAGE_PMD_NR; i++)
2694 head[i].mem_cgroup = head->mem_cgroup;
2696 __mod_memcg_state(head->mem_cgroup, MEMCG_RSS_HUGE, -HPAGE_PMD_NR);
2698 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2700 #ifdef CONFIG_MEMCG_SWAP
2702 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2703 * @entry: swap entry to be moved
2704 * @from: mem_cgroup which the entry is moved from
2705 * @to: mem_cgroup which the entry is moved to
2707 * It succeeds only when the swap_cgroup's record for this entry is the same
2708 * as the mem_cgroup's id of @from.
2710 * Returns 0 on success, -EINVAL on failure.
2712 * The caller must have charged to @to, IOW, called page_counter_charge() about
2713 * both res and memsw, and called css_get().
2715 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2716 struct mem_cgroup *from, struct mem_cgroup *to)
2718 unsigned short old_id, new_id;
2720 old_id = mem_cgroup_id(from);
2721 new_id = mem_cgroup_id(to);
2723 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2724 mod_memcg_state(from, MEMCG_SWAP, -1);
2725 mod_memcg_state(to, MEMCG_SWAP, 1);
2731 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2732 struct mem_cgroup *from, struct mem_cgroup *to)
2738 static DEFINE_MUTEX(memcg_max_mutex);
2740 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
2741 unsigned long max, bool memsw)
2743 bool enlarge = false;
2744 bool drained = false;
2746 bool limits_invariant;
2747 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
2750 if (signal_pending(current)) {
2755 mutex_lock(&memcg_max_mutex);
2757 * Make sure that the new limit (memsw or memory limit) doesn't
2758 * break our basic invariant rule memory.max <= memsw.max.
2760 limits_invariant = memsw ? max >= memcg->memory.max :
2761 max <= memcg->memsw.max;
2762 if (!limits_invariant) {
2763 mutex_unlock(&memcg_max_mutex);
2767 if (max > counter->max)
2769 ret = page_counter_set_max(counter, max);
2770 mutex_unlock(&memcg_max_mutex);
2776 drain_all_stock(memcg);
2781 if (!try_to_free_mem_cgroup_pages(memcg, 1,
2782 GFP_KERNEL, !memsw)) {
2788 if (!ret && enlarge)
2789 memcg_oom_recover(memcg);
2794 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
2796 unsigned long *total_scanned)
2798 unsigned long nr_reclaimed = 0;
2799 struct mem_cgroup_per_node *mz, *next_mz = NULL;
2800 unsigned long reclaimed;
2802 struct mem_cgroup_tree_per_node *mctz;
2803 unsigned long excess;
2804 unsigned long nr_scanned;
2809 mctz = soft_limit_tree_node(pgdat->node_id);
2812 * Do not even bother to check the largest node if the root
2813 * is empty. Do it lockless to prevent lock bouncing. Races
2814 * are acceptable as soft limit is best effort anyway.
2816 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
2820 * This loop can run a while, specially if mem_cgroup's continuously
2821 * keep exceeding their soft limit and putting the system under
2828 mz = mem_cgroup_largest_soft_limit_node(mctz);
2833 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
2834 gfp_mask, &nr_scanned);
2835 nr_reclaimed += reclaimed;
2836 *total_scanned += nr_scanned;
2837 spin_lock_irq(&mctz->lock);
2838 __mem_cgroup_remove_exceeded(mz, mctz);
2841 * If we failed to reclaim anything from this memory cgroup
2842 * it is time to move on to the next cgroup
2846 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
2848 excess = soft_limit_excess(mz->memcg);
2850 * One school of thought says that we should not add
2851 * back the node to the tree if reclaim returns 0.
2852 * But our reclaim could return 0, simply because due
2853 * to priority we are exposing a smaller subset of
2854 * memory to reclaim from. Consider this as a longer
2857 /* If excess == 0, no tree ops */
2858 __mem_cgroup_insert_exceeded(mz, mctz, excess);
2859 spin_unlock_irq(&mctz->lock);
2860 css_put(&mz->memcg->css);
2863 * Could not reclaim anything and there are no more
2864 * mem cgroups to try or we seem to be looping without
2865 * reclaiming anything.
2867 if (!nr_reclaimed &&
2869 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
2871 } while (!nr_reclaimed);
2873 css_put(&next_mz->memcg->css);
2874 return nr_reclaimed;
2878 * Test whether @memcg has children, dead or alive. Note that this
2879 * function doesn't care whether @memcg has use_hierarchy enabled and
2880 * returns %true if there are child csses according to the cgroup
2881 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2883 static inline bool memcg_has_children(struct mem_cgroup *memcg)
2888 ret = css_next_child(NULL, &memcg->css);
2894 * Reclaims as many pages from the given memcg as possible.
2896 * Caller is responsible for holding css reference for memcg.
2898 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
2900 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2902 /* we call try-to-free pages for make this cgroup empty */
2903 lru_add_drain_all();
2905 drain_all_stock(memcg);
2907 /* try to free all pages in this cgroup */
2908 while (nr_retries && page_counter_read(&memcg->memory)) {
2911 if (signal_pending(current))
2914 progress = try_to_free_mem_cgroup_pages(memcg, 1,
2918 /* maybe some writeback is necessary */
2919 congestion_wait(BLK_RW_ASYNC, HZ/10);
2927 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
2928 char *buf, size_t nbytes,
2931 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
2933 if (mem_cgroup_is_root(memcg))
2935 return mem_cgroup_force_empty(memcg) ?: nbytes;
2938 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
2941 return mem_cgroup_from_css(css)->use_hierarchy;
2944 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
2945 struct cftype *cft, u64 val)
2948 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2949 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
2951 if (memcg->use_hierarchy == val)
2955 * If parent's use_hierarchy is set, we can't make any modifications
2956 * in the child subtrees. If it is unset, then the change can
2957 * occur, provided the current cgroup has no children.
2959 * For the root cgroup, parent_mem is NULL, we allow value to be
2960 * set if there are no children.
2962 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
2963 (val == 1 || val == 0)) {
2964 if (!memcg_has_children(memcg))
2965 memcg->use_hierarchy = val;
2974 struct accumulated_stats {
2975 unsigned long stat[MEMCG_NR_STAT];
2976 unsigned long events[NR_VM_EVENT_ITEMS];
2977 unsigned long lru_pages[NR_LRU_LISTS];
2978 const unsigned int *stats_array;
2979 const unsigned int *events_array;
2984 static void accumulate_memcg_tree(struct mem_cgroup *memcg,
2985 struct accumulated_stats *acc)
2987 struct mem_cgroup *mi;
2990 for_each_mem_cgroup_tree(mi, memcg) {
2991 for (i = 0; i < acc->stats_size; i++)
2992 acc->stat[i] += memcg_page_state(mi,
2993 acc->stats_array ? acc->stats_array[i] : i);
2995 for (i = 0; i < acc->events_size; i++)
2996 acc->events[i] += memcg_sum_events(mi,
2997 acc->events_array ? acc->events_array[i] : i);
2999 for (i = 0; i < NR_LRU_LISTS; i++)
3000 acc->lru_pages[i] += memcg_page_state(mi,
3005 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3007 unsigned long val = 0;
3009 if (mem_cgroup_is_root(memcg)) {
3010 struct mem_cgroup *iter;
3012 for_each_mem_cgroup_tree(iter, memcg) {
3013 val += memcg_page_state(iter, MEMCG_CACHE);
3014 val += memcg_page_state(iter, MEMCG_RSS);
3016 val += memcg_page_state(iter, MEMCG_SWAP);
3020 val = page_counter_read(&memcg->memory);
3022 val = page_counter_read(&memcg->memsw);
3035 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3038 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3039 struct page_counter *counter;
3041 switch (MEMFILE_TYPE(cft->private)) {
3043 counter = &memcg->memory;
3046 counter = &memcg->memsw;
3049 counter = &memcg->kmem;
3052 counter = &memcg->tcpmem;
3058 switch (MEMFILE_ATTR(cft->private)) {
3060 if (counter == &memcg->memory)
3061 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3062 if (counter == &memcg->memsw)
3063 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3064 return (u64)page_counter_read(counter) * PAGE_SIZE;
3066 return (u64)counter->max * PAGE_SIZE;
3068 return (u64)counter->watermark * PAGE_SIZE;
3070 return counter->failcnt;
3071 case RES_SOFT_LIMIT:
3072 return (u64)memcg->soft_limit * PAGE_SIZE;
3078 #ifdef CONFIG_MEMCG_KMEM
3079 static int memcg_online_kmem(struct mem_cgroup *memcg)
3083 if (cgroup_memory_nokmem)
3086 BUG_ON(memcg->kmemcg_id >= 0);
3087 BUG_ON(memcg->kmem_state);
3089 memcg_id = memcg_alloc_cache_id();
3093 static_branch_inc(&memcg_kmem_enabled_key);
3095 * A memory cgroup is considered kmem-online as soon as it gets
3096 * kmemcg_id. Setting the id after enabling static branching will
3097 * guarantee no one starts accounting before all call sites are
3100 memcg->kmemcg_id = memcg_id;
3101 memcg->kmem_state = KMEM_ONLINE;
3102 INIT_LIST_HEAD(&memcg->kmem_caches);
3107 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3109 struct cgroup_subsys_state *css;
3110 struct mem_cgroup *parent, *child;
3113 if (memcg->kmem_state != KMEM_ONLINE)
3116 * Clear the online state before clearing memcg_caches array
3117 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
3118 * guarantees that no cache will be created for this cgroup
3119 * after we are done (see memcg_create_kmem_cache()).
3121 memcg->kmem_state = KMEM_ALLOCATED;
3123 memcg_deactivate_kmem_caches(memcg);
3125 kmemcg_id = memcg->kmemcg_id;
3126 BUG_ON(kmemcg_id < 0);
3128 parent = parent_mem_cgroup(memcg);
3130 parent = root_mem_cgroup;
3133 * Change kmemcg_id of this cgroup and all its descendants to the
3134 * parent's id, and then move all entries from this cgroup's list_lrus
3135 * to ones of the parent. After we have finished, all list_lrus
3136 * corresponding to this cgroup are guaranteed to remain empty. The
3137 * ordering is imposed by list_lru_node->lock taken by
3138 * memcg_drain_all_list_lrus().
3140 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3141 css_for_each_descendant_pre(css, &memcg->css) {
3142 child = mem_cgroup_from_css(css);
3143 BUG_ON(child->kmemcg_id != kmemcg_id);
3144 child->kmemcg_id = parent->kmemcg_id;
3145 if (!memcg->use_hierarchy)
3150 memcg_drain_all_list_lrus(kmemcg_id, parent);
3152 memcg_free_cache_id(kmemcg_id);
3155 static void memcg_free_kmem(struct mem_cgroup *memcg)
3157 /* css_alloc() failed, offlining didn't happen */
3158 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3159 memcg_offline_kmem(memcg);
3161 if (memcg->kmem_state == KMEM_ALLOCATED) {
3162 memcg_destroy_kmem_caches(memcg);
3163 static_branch_dec(&memcg_kmem_enabled_key);
3164 WARN_ON(page_counter_read(&memcg->kmem));
3168 static int memcg_online_kmem(struct mem_cgroup *memcg)
3172 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3175 static void memcg_free_kmem(struct mem_cgroup *memcg)
3178 #endif /* CONFIG_MEMCG_KMEM */
3180 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3185 mutex_lock(&memcg_max_mutex);
3186 ret = page_counter_set_max(&memcg->kmem, max);
3187 mutex_unlock(&memcg_max_mutex);
3191 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3195 mutex_lock(&memcg_max_mutex);
3197 ret = page_counter_set_max(&memcg->tcpmem, max);
3201 if (!memcg->tcpmem_active) {
3203 * The active flag needs to be written after the static_key
3204 * update. This is what guarantees that the socket activation
3205 * function is the last one to run. See mem_cgroup_sk_alloc()
3206 * for details, and note that we don't mark any socket as
3207 * belonging to this memcg until that flag is up.
3209 * We need to do this, because static_keys will span multiple
3210 * sites, but we can't control their order. If we mark a socket
3211 * as accounted, but the accounting functions are not patched in
3212 * yet, we'll lose accounting.
3214 * We never race with the readers in mem_cgroup_sk_alloc(),
3215 * because when this value change, the code to process it is not
3218 static_branch_inc(&memcg_sockets_enabled_key);
3219 memcg->tcpmem_active = true;
3222 mutex_unlock(&memcg_max_mutex);
3227 * The user of this function is...
3230 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3231 char *buf, size_t nbytes, loff_t off)
3233 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3234 unsigned long nr_pages;
3237 buf = strstrip(buf);
3238 ret = page_counter_memparse(buf, "-1", &nr_pages);
3242 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3244 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3248 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3250 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3253 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3256 ret = memcg_update_kmem_max(memcg, nr_pages);
3259 ret = memcg_update_tcp_max(memcg, nr_pages);
3263 case RES_SOFT_LIMIT:
3264 memcg->soft_limit = nr_pages;
3268 return ret ?: nbytes;
3271 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3272 size_t nbytes, loff_t off)
3274 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3275 struct page_counter *counter;
3277 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3279 counter = &memcg->memory;
3282 counter = &memcg->memsw;
3285 counter = &memcg->kmem;
3288 counter = &memcg->tcpmem;
3294 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3296 page_counter_reset_watermark(counter);
3299 counter->failcnt = 0;
3308 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3311 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3315 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3316 struct cftype *cft, u64 val)
3318 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3320 if (val & ~MOVE_MASK)
3324 * No kind of locking is needed in here, because ->can_attach() will
3325 * check this value once in the beginning of the process, and then carry
3326 * on with stale data. This means that changes to this value will only
3327 * affect task migrations starting after the change.
3329 memcg->move_charge_at_immigrate = val;
3333 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3334 struct cftype *cft, u64 val)
3341 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3345 unsigned int lru_mask;
3348 static const struct numa_stat stats[] = {
3349 { "total", LRU_ALL },
3350 { "file", LRU_ALL_FILE },
3351 { "anon", LRU_ALL_ANON },
3352 { "unevictable", BIT(LRU_UNEVICTABLE) },
3354 const struct numa_stat *stat;
3357 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3359 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3360 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3361 seq_printf(m, "%s=%lu", stat->name, nr);
3362 for_each_node_state(nid, N_MEMORY) {
3363 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3365 seq_printf(m, " N%d=%lu", nid, nr);
3370 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3371 struct mem_cgroup *iter;
3374 for_each_mem_cgroup_tree(iter, memcg)
3375 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3376 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3377 for_each_node_state(nid, N_MEMORY) {
3379 for_each_mem_cgroup_tree(iter, memcg)
3380 nr += mem_cgroup_node_nr_lru_pages(
3381 iter, nid, stat->lru_mask);
3382 seq_printf(m, " N%d=%lu", nid, nr);
3389 #endif /* CONFIG_NUMA */
3391 /* Universal VM events cgroup1 shows, original sort order */
3392 static const unsigned int memcg1_events[] = {
3399 static const char *const memcg1_event_names[] = {
3406 static int memcg_stat_show(struct seq_file *m, void *v)
3408 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3409 unsigned long memory, memsw;
3410 struct mem_cgroup *mi;
3412 struct accumulated_stats acc;
3414 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
3415 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3417 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3418 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3420 seq_printf(m, "%s %lu\n", memcg1_stat_names[i],
3421 memcg_page_state(memcg, memcg1_stats[i]) *
3425 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3426 seq_printf(m, "%s %lu\n", memcg1_event_names[i],
3427 memcg_sum_events(memcg, memcg1_events[i]));
3429 for (i = 0; i < NR_LRU_LISTS; i++)
3430 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3431 memcg_page_state(memcg, NR_LRU_BASE + i) *
3434 /* Hierarchical information */
3435 memory = memsw = PAGE_COUNTER_MAX;
3436 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3437 memory = min(memory, mi->memory.max);
3438 memsw = min(memsw, mi->memsw.max);
3440 seq_printf(m, "hierarchical_memory_limit %llu\n",
3441 (u64)memory * PAGE_SIZE);
3442 if (do_memsw_account())
3443 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3444 (u64)memsw * PAGE_SIZE);
3446 memset(&acc, 0, sizeof(acc));
3447 acc.stats_size = ARRAY_SIZE(memcg1_stats);
3448 acc.stats_array = memcg1_stats;
3449 acc.events_size = ARRAY_SIZE(memcg1_events);
3450 acc.events_array = memcg1_events;
3451 accumulate_memcg_tree(memcg, &acc);
3453 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3454 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3456 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
3457 (u64)acc.stat[i] * PAGE_SIZE);
3460 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3461 seq_printf(m, "total_%s %llu\n", memcg1_event_names[i],
3462 (u64)acc.events[i]);
3464 for (i = 0; i < NR_LRU_LISTS; i++)
3465 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i],
3466 (u64)acc.lru_pages[i] * PAGE_SIZE);
3468 #ifdef CONFIG_DEBUG_VM
3471 struct mem_cgroup_per_node *mz;
3472 struct zone_reclaim_stat *rstat;
3473 unsigned long recent_rotated[2] = {0, 0};
3474 unsigned long recent_scanned[2] = {0, 0};
3476 for_each_online_pgdat(pgdat) {
3477 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
3478 rstat = &mz->lruvec.reclaim_stat;
3480 recent_rotated[0] += rstat->recent_rotated[0];
3481 recent_rotated[1] += rstat->recent_rotated[1];
3482 recent_scanned[0] += rstat->recent_scanned[0];
3483 recent_scanned[1] += rstat->recent_scanned[1];
3485 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3486 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3487 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3488 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3495 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3498 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3500 return mem_cgroup_swappiness(memcg);
3503 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3504 struct cftype *cft, u64 val)
3506 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3512 memcg->swappiness = val;
3514 vm_swappiness = val;
3519 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3521 struct mem_cgroup_threshold_ary *t;
3522 unsigned long usage;
3527 t = rcu_dereference(memcg->thresholds.primary);
3529 t = rcu_dereference(memcg->memsw_thresholds.primary);
3534 usage = mem_cgroup_usage(memcg, swap);
3537 * current_threshold points to threshold just below or equal to usage.
3538 * If it's not true, a threshold was crossed after last
3539 * call of __mem_cgroup_threshold().
3541 i = t->current_threshold;
3544 * Iterate backward over array of thresholds starting from
3545 * current_threshold and check if a threshold is crossed.
3546 * If none of thresholds below usage is crossed, we read
3547 * only one element of the array here.
3549 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3550 eventfd_signal(t->entries[i].eventfd, 1);
3552 /* i = current_threshold + 1 */
3556 * Iterate forward over array of thresholds starting from
3557 * current_threshold+1 and check if a threshold is crossed.
3558 * If none of thresholds above usage is crossed, we read
3559 * only one element of the array here.
3561 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3562 eventfd_signal(t->entries[i].eventfd, 1);
3564 /* Update current_threshold */
3565 t->current_threshold = i - 1;
3570 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3573 __mem_cgroup_threshold(memcg, false);
3574 if (do_memsw_account())
3575 __mem_cgroup_threshold(memcg, true);
3577 memcg = parent_mem_cgroup(memcg);
3581 static int compare_thresholds(const void *a, const void *b)
3583 const struct mem_cgroup_threshold *_a = a;
3584 const struct mem_cgroup_threshold *_b = b;
3586 if (_a->threshold > _b->threshold)
3589 if (_a->threshold < _b->threshold)
3595 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3597 struct mem_cgroup_eventfd_list *ev;
3599 spin_lock(&memcg_oom_lock);
3601 list_for_each_entry(ev, &memcg->oom_notify, list)
3602 eventfd_signal(ev->eventfd, 1);
3604 spin_unlock(&memcg_oom_lock);
3608 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3610 struct mem_cgroup *iter;
3612 for_each_mem_cgroup_tree(iter, memcg)
3613 mem_cgroup_oom_notify_cb(iter);
3616 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3617 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3619 struct mem_cgroup_thresholds *thresholds;
3620 struct mem_cgroup_threshold_ary *new;
3621 unsigned long threshold;
3622 unsigned long usage;
3625 ret = page_counter_memparse(args, "-1", &threshold);
3629 mutex_lock(&memcg->thresholds_lock);
3632 thresholds = &memcg->thresholds;
3633 usage = mem_cgroup_usage(memcg, false);
3634 } else if (type == _MEMSWAP) {
3635 thresholds = &memcg->memsw_thresholds;
3636 usage = mem_cgroup_usage(memcg, true);
3640 /* Check if a threshold crossed before adding a new one */
3641 if (thresholds->primary)
3642 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3644 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3646 /* Allocate memory for new array of thresholds */
3647 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
3654 /* Copy thresholds (if any) to new array */
3655 if (thresholds->primary) {
3656 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3657 sizeof(struct mem_cgroup_threshold));
3660 /* Add new threshold */
3661 new->entries[size - 1].eventfd = eventfd;
3662 new->entries[size - 1].threshold = threshold;
3664 /* Sort thresholds. Registering of new threshold isn't time-critical */
3665 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3666 compare_thresholds, NULL);
3668 /* Find current threshold */
3669 new->current_threshold = -1;
3670 for (i = 0; i < size; i++) {
3671 if (new->entries[i].threshold <= usage) {
3673 * new->current_threshold will not be used until
3674 * rcu_assign_pointer(), so it's safe to increment
3677 ++new->current_threshold;
3682 /* Free old spare buffer and save old primary buffer as spare */
3683 kfree(thresholds->spare);
3684 thresholds->spare = thresholds->primary;
3686 rcu_assign_pointer(thresholds->primary, new);
3688 /* To be sure that nobody uses thresholds */
3692 mutex_unlock(&memcg->thresholds_lock);
3697 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3698 struct eventfd_ctx *eventfd, const char *args)
3700 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
3703 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
3704 struct eventfd_ctx *eventfd, const char *args)
3706 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
3709 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3710 struct eventfd_ctx *eventfd, enum res_type type)
3712 struct mem_cgroup_thresholds *thresholds;
3713 struct mem_cgroup_threshold_ary *new;
3714 unsigned long usage;
3717 mutex_lock(&memcg->thresholds_lock);
3720 thresholds = &memcg->thresholds;
3721 usage = mem_cgroup_usage(memcg, false);
3722 } else if (type == _MEMSWAP) {
3723 thresholds = &memcg->memsw_thresholds;
3724 usage = mem_cgroup_usage(memcg, true);
3728 if (!thresholds->primary)
3731 /* Check if a threshold crossed before removing */
3732 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3734 /* Calculate new number of threshold */
3736 for (i = 0; i < thresholds->primary->size; i++) {
3737 if (thresholds->primary->entries[i].eventfd != eventfd)
3741 new = thresholds->spare;
3743 /* Set thresholds array to NULL if we don't have thresholds */
3752 /* Copy thresholds and find current threshold */
3753 new->current_threshold = -1;
3754 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3755 if (thresholds->primary->entries[i].eventfd == eventfd)
3758 new->entries[j] = thresholds->primary->entries[i];
3759 if (new->entries[j].threshold <= usage) {
3761 * new->current_threshold will not be used
3762 * until rcu_assign_pointer(), so it's safe to increment
3765 ++new->current_threshold;
3771 /* Swap primary and spare array */
3772 thresholds->spare = thresholds->primary;
3774 rcu_assign_pointer(thresholds->primary, new);
3776 /* To be sure that nobody uses thresholds */
3779 /* If all events are unregistered, free the spare array */
3781 kfree(thresholds->spare);
3782 thresholds->spare = NULL;
3785 mutex_unlock(&memcg->thresholds_lock);
3788 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3789 struct eventfd_ctx *eventfd)
3791 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
3794 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3795 struct eventfd_ctx *eventfd)
3797 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
3800 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
3801 struct eventfd_ctx *eventfd, const char *args)
3803 struct mem_cgroup_eventfd_list *event;
3805 event = kmalloc(sizeof(*event), GFP_KERNEL);
3809 spin_lock(&memcg_oom_lock);
3811 event->eventfd = eventfd;
3812 list_add(&event->list, &memcg->oom_notify);
3814 /* already in OOM ? */
3815 if (memcg->under_oom)
3816 eventfd_signal(eventfd, 1);
3817 spin_unlock(&memcg_oom_lock);
3822 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
3823 struct eventfd_ctx *eventfd)
3825 struct mem_cgroup_eventfd_list *ev, *tmp;
3827 spin_lock(&memcg_oom_lock);
3829 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
3830 if (ev->eventfd == eventfd) {
3831 list_del(&ev->list);
3836 spin_unlock(&memcg_oom_lock);
3839 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
3841 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
3843 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
3844 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
3845 seq_printf(sf, "oom_kill %lu\n",
3846 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
3850 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
3851 struct cftype *cft, u64 val)
3853 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3855 /* cannot set to root cgroup and only 0 and 1 are allowed */
3856 if (!css->parent || !((val == 0) || (val == 1)))
3859 memcg->oom_kill_disable = val;
3861 memcg_oom_recover(memcg);
3866 #ifdef CONFIG_CGROUP_WRITEBACK
3868 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3870 return wb_domain_init(&memcg->cgwb_domain, gfp);
3873 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3875 wb_domain_exit(&memcg->cgwb_domain);
3878 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3880 wb_domain_size_changed(&memcg->cgwb_domain);
3883 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
3885 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3887 if (!memcg->css.parent)
3890 return &memcg->cgwb_domain;
3894 * idx can be of type enum memcg_stat_item or node_stat_item.
3895 * Keep in sync with memcg_exact_page().
3897 static unsigned long memcg_exact_page_state(struct mem_cgroup *memcg, int idx)
3899 long x = atomic_long_read(&memcg->stat[idx]);
3902 for_each_online_cpu(cpu)
3903 x += per_cpu_ptr(memcg->stat_cpu, cpu)->count[idx];
3910 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3911 * @wb: bdi_writeback in question
3912 * @pfilepages: out parameter for number of file pages
3913 * @pheadroom: out parameter for number of allocatable pages according to memcg
3914 * @pdirty: out parameter for number of dirty pages
3915 * @pwriteback: out parameter for number of pages under writeback
3917 * Determine the numbers of file, headroom, dirty, and writeback pages in
3918 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
3919 * is a bit more involved.
3921 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
3922 * headroom is calculated as the lowest headroom of itself and the
3923 * ancestors. Note that this doesn't consider the actual amount of
3924 * available memory in the system. The caller should further cap
3925 * *@pheadroom accordingly.
3927 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
3928 unsigned long *pheadroom, unsigned long *pdirty,
3929 unsigned long *pwriteback)
3931 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3932 struct mem_cgroup *parent;
3934 *pdirty = memcg_exact_page_state(memcg, NR_FILE_DIRTY);
3936 /* this should eventually include NR_UNSTABLE_NFS */
3937 *pwriteback = memcg_exact_page_state(memcg, NR_WRITEBACK);
3938 *pfilepages = memcg_exact_page_state(memcg, NR_INACTIVE_FILE) +
3939 memcg_exact_page_state(memcg, NR_ACTIVE_FILE);
3940 *pheadroom = PAGE_COUNTER_MAX;
3942 while ((parent = parent_mem_cgroup(memcg))) {
3943 unsigned long ceiling = min(memcg->memory.max, memcg->high);
3944 unsigned long used = page_counter_read(&memcg->memory);
3946 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
3951 #else /* CONFIG_CGROUP_WRITEBACK */
3953 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3958 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3962 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3966 #endif /* CONFIG_CGROUP_WRITEBACK */
3969 * DO NOT USE IN NEW FILES.
3971 * "cgroup.event_control" implementation.
3973 * This is way over-engineered. It tries to support fully configurable
3974 * events for each user. Such level of flexibility is completely
3975 * unnecessary especially in the light of the planned unified hierarchy.
3977 * Please deprecate this and replace with something simpler if at all
3982 * Unregister event and free resources.
3984 * Gets called from workqueue.
3986 static void memcg_event_remove(struct work_struct *work)
3988 struct mem_cgroup_event *event =
3989 container_of(work, struct mem_cgroup_event, remove);
3990 struct mem_cgroup *memcg = event->memcg;
3992 remove_wait_queue(event->wqh, &event->wait);
3994 event->unregister_event(memcg, event->eventfd);
3996 /* Notify userspace the event is going away. */
3997 eventfd_signal(event->eventfd, 1);
3999 eventfd_ctx_put(event->eventfd);
4001 css_put(&memcg->css);
4005 * Gets called on EPOLLHUP on eventfd when user closes it.
4007 * Called with wqh->lock held and interrupts disabled.
4009 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4010 int sync, void *key)
4012 struct mem_cgroup_event *event =
4013 container_of(wait, struct mem_cgroup_event, wait);
4014 struct mem_cgroup *memcg = event->memcg;
4015 __poll_t flags = key_to_poll(key);
4017 if (flags & EPOLLHUP) {
4019 * If the event has been detached at cgroup removal, we
4020 * can simply return knowing the other side will cleanup
4023 * We can't race against event freeing since the other
4024 * side will require wqh->lock via remove_wait_queue(),
4027 spin_lock(&memcg->event_list_lock);
4028 if (!list_empty(&event->list)) {
4029 list_del_init(&event->list);
4031 * We are in atomic context, but cgroup_event_remove()
4032 * may sleep, so we have to call it in workqueue.
4034 schedule_work(&event->remove);
4036 spin_unlock(&memcg->event_list_lock);
4042 static void memcg_event_ptable_queue_proc(struct file *file,
4043 wait_queue_head_t *wqh, poll_table *pt)
4045 struct mem_cgroup_event *event =
4046 container_of(pt, struct mem_cgroup_event, pt);
4049 add_wait_queue(wqh, &event->wait);
4053 * DO NOT USE IN NEW FILES.
4055 * Parse input and register new cgroup event handler.
4057 * Input must be in format '<event_fd> <control_fd> <args>'.
4058 * Interpretation of args is defined by control file implementation.
4060 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4061 char *buf, size_t nbytes, loff_t off)
4063 struct cgroup_subsys_state *css = of_css(of);
4064 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4065 struct mem_cgroup_event *event;
4066 struct cgroup_subsys_state *cfile_css;
4067 unsigned int efd, cfd;
4074 buf = strstrip(buf);
4076 efd = simple_strtoul(buf, &endp, 10);
4081 cfd = simple_strtoul(buf, &endp, 10);
4082 if ((*endp != ' ') && (*endp != '\0'))
4086 event = kzalloc(sizeof(*event), GFP_KERNEL);
4090 event->memcg = memcg;
4091 INIT_LIST_HEAD(&event->list);
4092 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4093 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4094 INIT_WORK(&event->remove, memcg_event_remove);
4102 event->eventfd = eventfd_ctx_fileget(efile.file);
4103 if (IS_ERR(event->eventfd)) {
4104 ret = PTR_ERR(event->eventfd);
4111 goto out_put_eventfd;
4114 /* the process need read permission on control file */
4115 /* AV: shouldn't we check that it's been opened for read instead? */
4116 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4121 * Determine the event callbacks and set them in @event. This used
4122 * to be done via struct cftype but cgroup core no longer knows
4123 * about these events. The following is crude but the whole thing
4124 * is for compatibility anyway.
4126 * DO NOT ADD NEW FILES.
4128 name = cfile.file->f_path.dentry->d_name.name;
4130 if (!strcmp(name, "memory.usage_in_bytes")) {
4131 event->register_event = mem_cgroup_usage_register_event;
4132 event->unregister_event = mem_cgroup_usage_unregister_event;
4133 } else if (!strcmp(name, "memory.oom_control")) {
4134 event->register_event = mem_cgroup_oom_register_event;
4135 event->unregister_event = mem_cgroup_oom_unregister_event;
4136 } else if (!strcmp(name, "memory.pressure_level")) {
4137 event->register_event = vmpressure_register_event;
4138 event->unregister_event = vmpressure_unregister_event;
4139 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4140 event->register_event = memsw_cgroup_usage_register_event;
4141 event->unregister_event = memsw_cgroup_usage_unregister_event;
4148 * Verify @cfile should belong to @css. Also, remaining events are
4149 * automatically removed on cgroup destruction but the removal is
4150 * asynchronous, so take an extra ref on @css.
4152 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4153 &memory_cgrp_subsys);
4155 if (IS_ERR(cfile_css))
4157 if (cfile_css != css) {
4162 ret = event->register_event(memcg, event->eventfd, buf);
4166 vfs_poll(efile.file, &event->pt);
4168 spin_lock(&memcg->event_list_lock);
4169 list_add(&event->list, &memcg->event_list);
4170 spin_unlock(&memcg->event_list_lock);
4182 eventfd_ctx_put(event->eventfd);
4191 static struct cftype mem_cgroup_legacy_files[] = {
4193 .name = "usage_in_bytes",
4194 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4195 .read_u64 = mem_cgroup_read_u64,
4198 .name = "max_usage_in_bytes",
4199 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4200 .write = mem_cgroup_reset,
4201 .read_u64 = mem_cgroup_read_u64,
4204 .name = "limit_in_bytes",
4205 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4206 .write = mem_cgroup_write,
4207 .read_u64 = mem_cgroup_read_u64,
4210 .name = "soft_limit_in_bytes",
4211 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4212 .write = mem_cgroup_write,
4213 .read_u64 = mem_cgroup_read_u64,
4217 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4218 .write = mem_cgroup_reset,
4219 .read_u64 = mem_cgroup_read_u64,
4223 .seq_show = memcg_stat_show,
4226 .name = "force_empty",
4227 .write = mem_cgroup_force_empty_write,
4230 .name = "use_hierarchy",
4231 .write_u64 = mem_cgroup_hierarchy_write,
4232 .read_u64 = mem_cgroup_hierarchy_read,
4235 .name = "cgroup.event_control", /* XXX: for compat */
4236 .write = memcg_write_event_control,
4237 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4240 .name = "swappiness",
4241 .read_u64 = mem_cgroup_swappiness_read,
4242 .write_u64 = mem_cgroup_swappiness_write,
4245 .name = "move_charge_at_immigrate",
4246 .read_u64 = mem_cgroup_move_charge_read,
4247 .write_u64 = mem_cgroup_move_charge_write,
4250 .name = "oom_control",
4251 .seq_show = mem_cgroup_oom_control_read,
4252 .write_u64 = mem_cgroup_oom_control_write,
4253 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4256 .name = "pressure_level",
4260 .name = "numa_stat",
4261 .seq_show = memcg_numa_stat_show,
4265 .name = "kmem.limit_in_bytes",
4266 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4267 .write = mem_cgroup_write,
4268 .read_u64 = mem_cgroup_read_u64,
4271 .name = "kmem.usage_in_bytes",
4272 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4273 .read_u64 = mem_cgroup_read_u64,
4276 .name = "kmem.failcnt",
4277 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4278 .write = mem_cgroup_reset,
4279 .read_u64 = mem_cgroup_read_u64,
4282 .name = "kmem.max_usage_in_bytes",
4283 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4284 .write = mem_cgroup_reset,
4285 .read_u64 = mem_cgroup_read_u64,
4287 #if defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG)
4289 .name = "kmem.slabinfo",
4290 .seq_start = memcg_slab_start,
4291 .seq_next = memcg_slab_next,
4292 .seq_stop = memcg_slab_stop,
4293 .seq_show = memcg_slab_show,
4297 .name = "kmem.tcp.limit_in_bytes",
4298 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
4299 .write = mem_cgroup_write,
4300 .read_u64 = mem_cgroup_read_u64,
4303 .name = "kmem.tcp.usage_in_bytes",
4304 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
4305 .read_u64 = mem_cgroup_read_u64,
4308 .name = "kmem.tcp.failcnt",
4309 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
4310 .write = mem_cgroup_reset,
4311 .read_u64 = mem_cgroup_read_u64,
4314 .name = "kmem.tcp.max_usage_in_bytes",
4315 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
4316 .write = mem_cgroup_reset,
4317 .read_u64 = mem_cgroup_read_u64,
4319 { }, /* terminate */
4323 * Private memory cgroup IDR
4325 * Swap-out records and page cache shadow entries need to store memcg
4326 * references in constrained space, so we maintain an ID space that is
4327 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4328 * memory-controlled cgroups to 64k.
4330 * However, there usually are many references to the oflline CSS after
4331 * the cgroup has been destroyed, such as page cache or reclaimable
4332 * slab objects, that don't need to hang on to the ID. We want to keep
4333 * those dead CSS from occupying IDs, or we might quickly exhaust the
4334 * relatively small ID space and prevent the creation of new cgroups
4335 * even when there are much fewer than 64k cgroups - possibly none.
4337 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4338 * be freed and recycled when it's no longer needed, which is usually
4339 * when the CSS is offlined.
4341 * The only exception to that are records of swapped out tmpfs/shmem
4342 * pages that need to be attributed to live ancestors on swapin. But
4343 * those references are manageable from userspace.
4346 static DEFINE_IDR(mem_cgroup_idr);
4348 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
4350 if (memcg->id.id > 0) {
4351 idr_remove(&mem_cgroup_idr, memcg->id.id);
4356 static void mem_cgroup_id_get_many(struct mem_cgroup *memcg, unsigned int n)
4358 refcount_add(n, &memcg->id.ref);
4361 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
4363 if (refcount_sub_and_test(n, &memcg->id.ref)) {
4364 mem_cgroup_id_remove(memcg);
4366 /* Memcg ID pins CSS */
4367 css_put(&memcg->css);
4371 static inline void mem_cgroup_id_get(struct mem_cgroup *memcg)
4373 mem_cgroup_id_get_many(memcg, 1);
4376 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
4378 mem_cgroup_id_put_many(memcg, 1);
4382 * mem_cgroup_from_id - look up a memcg from a memcg id
4383 * @id: the memcg id to look up
4385 * Caller must hold rcu_read_lock().
4387 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
4389 WARN_ON_ONCE(!rcu_read_lock_held());
4390 return idr_find(&mem_cgroup_idr, id);
4393 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4395 struct mem_cgroup_per_node *pn;
4398 * This routine is called against possible nodes.
4399 * But it's BUG to call kmalloc() against offline node.
4401 * TODO: this routine can waste much memory for nodes which will
4402 * never be onlined. It's better to use memory hotplug callback
4405 if (!node_state(node, N_NORMAL_MEMORY))
4407 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4411 pn->lruvec_stat_cpu = alloc_percpu(struct lruvec_stat);
4412 if (!pn->lruvec_stat_cpu) {
4417 lruvec_init(&pn->lruvec);
4418 pn->usage_in_excess = 0;
4419 pn->on_tree = false;
4422 memcg->nodeinfo[node] = pn;
4426 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4428 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
4433 free_percpu(pn->lruvec_stat_cpu);
4437 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4442 free_mem_cgroup_per_node_info(memcg, node);
4443 free_percpu(memcg->stat_cpu);
4447 static void mem_cgroup_free(struct mem_cgroup *memcg)
4449 memcg_wb_domain_exit(memcg);
4450 __mem_cgroup_free(memcg);
4453 static struct mem_cgroup *mem_cgroup_alloc(void)
4455 struct mem_cgroup *memcg;
4459 size = sizeof(struct mem_cgroup);
4460 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4462 memcg = kzalloc(size, GFP_KERNEL);
4466 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
4467 1, MEM_CGROUP_ID_MAX,
4469 if (memcg->id.id < 0)
4472 memcg->stat_cpu = alloc_percpu(struct mem_cgroup_stat_cpu);
4473 if (!memcg->stat_cpu)
4477 if (alloc_mem_cgroup_per_node_info(memcg, node))
4480 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
4483 INIT_WORK(&memcg->high_work, high_work_func);
4484 memcg->last_scanned_node = MAX_NUMNODES;
4485 INIT_LIST_HEAD(&memcg->oom_notify);
4486 mutex_init(&memcg->thresholds_lock);
4487 spin_lock_init(&memcg->move_lock);
4488 vmpressure_init(&memcg->vmpressure);
4489 INIT_LIST_HEAD(&memcg->event_list);
4490 spin_lock_init(&memcg->event_list_lock);
4491 memcg->socket_pressure = jiffies;
4492 #ifdef CONFIG_MEMCG_KMEM
4493 memcg->kmemcg_id = -1;
4495 #ifdef CONFIG_CGROUP_WRITEBACK
4496 INIT_LIST_HEAD(&memcg->cgwb_list);
4498 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
4501 mem_cgroup_id_remove(memcg);
4502 __mem_cgroup_free(memcg);
4506 static struct cgroup_subsys_state * __ref
4507 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4509 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
4510 struct mem_cgroup *memcg;
4511 long error = -ENOMEM;
4513 memcg = mem_cgroup_alloc();
4515 return ERR_PTR(error);
4517 memcg->high = PAGE_COUNTER_MAX;
4518 memcg->soft_limit = PAGE_COUNTER_MAX;
4520 memcg->swappiness = mem_cgroup_swappiness(parent);
4521 memcg->oom_kill_disable = parent->oom_kill_disable;
4523 if (parent && parent->use_hierarchy) {
4524 memcg->use_hierarchy = true;
4525 page_counter_init(&memcg->memory, &parent->memory);
4526 page_counter_init(&memcg->swap, &parent->swap);
4527 page_counter_init(&memcg->memsw, &parent->memsw);
4528 page_counter_init(&memcg->kmem, &parent->kmem);
4529 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
4531 page_counter_init(&memcg->memory, NULL);
4532 page_counter_init(&memcg->swap, NULL);
4533 page_counter_init(&memcg->memsw, NULL);
4534 page_counter_init(&memcg->kmem, NULL);
4535 page_counter_init(&memcg->tcpmem, NULL);
4537 * Deeper hierachy with use_hierarchy == false doesn't make
4538 * much sense so let cgroup subsystem know about this
4539 * unfortunate state in our controller.
4541 if (parent != root_mem_cgroup)
4542 memory_cgrp_subsys.broken_hierarchy = true;
4545 /* The following stuff does not apply to the root */
4547 root_mem_cgroup = memcg;
4551 error = memcg_online_kmem(memcg);
4555 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4556 static_branch_inc(&memcg_sockets_enabled_key);
4560 mem_cgroup_id_remove(memcg);
4561 mem_cgroup_free(memcg);
4562 return ERR_PTR(-ENOMEM);
4565 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
4567 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4570 * A memcg must be visible for memcg_expand_shrinker_maps()
4571 * by the time the maps are allocated. So, we allocate maps
4572 * here, when for_each_mem_cgroup() can't skip it.
4574 if (memcg_alloc_shrinker_maps(memcg)) {
4575 mem_cgroup_id_remove(memcg);
4579 /* Online state pins memcg ID, memcg ID pins CSS */
4580 refcount_set(&memcg->id.ref, 1);
4585 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4587 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4588 struct mem_cgroup_event *event, *tmp;
4591 * Unregister events and notify userspace.
4592 * Notify userspace about cgroup removing only after rmdir of cgroup
4593 * directory to avoid race between userspace and kernelspace.
4595 spin_lock(&memcg->event_list_lock);
4596 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4597 list_del_init(&event->list);
4598 schedule_work(&event->remove);
4600 spin_unlock(&memcg->event_list_lock);
4602 page_counter_set_min(&memcg->memory, 0);
4603 page_counter_set_low(&memcg->memory, 0);
4605 memcg_offline_kmem(memcg);
4606 wb_memcg_offline(memcg);
4608 drain_all_stock(memcg);
4610 mem_cgroup_id_put(memcg);
4613 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
4615 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4617 invalidate_reclaim_iterators(memcg);
4620 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4622 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4624 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4625 static_branch_dec(&memcg_sockets_enabled_key);
4627 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
4628 static_branch_dec(&memcg_sockets_enabled_key);
4630 vmpressure_cleanup(&memcg->vmpressure);
4631 cancel_work_sync(&memcg->high_work);
4632 mem_cgroup_remove_from_trees(memcg);
4633 memcg_free_shrinker_maps(memcg);
4634 memcg_free_kmem(memcg);
4635 mem_cgroup_free(memcg);
4639 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4640 * @css: the target css
4642 * Reset the states of the mem_cgroup associated with @css. This is
4643 * invoked when the userland requests disabling on the default hierarchy
4644 * but the memcg is pinned through dependency. The memcg should stop
4645 * applying policies and should revert to the vanilla state as it may be
4646 * made visible again.
4648 * The current implementation only resets the essential configurations.
4649 * This needs to be expanded to cover all the visible parts.
4651 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4653 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4655 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
4656 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
4657 page_counter_set_max(&memcg->memsw, PAGE_COUNTER_MAX);
4658 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
4659 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
4660 page_counter_set_min(&memcg->memory, 0);
4661 page_counter_set_low(&memcg->memory, 0);
4662 memcg->high = PAGE_COUNTER_MAX;
4663 memcg->soft_limit = PAGE_COUNTER_MAX;
4664 memcg_wb_domain_size_changed(memcg);
4668 /* Handlers for move charge at task migration. */
4669 static int mem_cgroup_do_precharge(unsigned long count)
4673 /* Try a single bulk charge without reclaim first, kswapd may wake */
4674 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
4676 mc.precharge += count;
4680 /* Try charges one by one with reclaim, but do not retry */
4682 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
4696 enum mc_target_type {
4703 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4704 unsigned long addr, pte_t ptent)
4706 struct page *page = _vm_normal_page(vma, addr, ptent, true);
4708 if (!page || !page_mapped(page))
4710 if (PageAnon(page)) {
4711 if (!(mc.flags & MOVE_ANON))
4714 if (!(mc.flags & MOVE_FILE))
4717 if (!get_page_unless_zero(page))
4723 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
4724 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4725 pte_t ptent, swp_entry_t *entry)
4727 struct page *page = NULL;
4728 swp_entry_t ent = pte_to_swp_entry(ptent);
4730 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
4734 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
4735 * a device and because they are not accessible by CPU they are store
4736 * as special swap entry in the CPU page table.
4738 if (is_device_private_entry(ent)) {
4739 page = device_private_entry_to_page(ent);
4741 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
4742 * a refcount of 1 when free (unlike normal page)
4744 if (!page_ref_add_unless(page, 1, 1))
4750 * Because lookup_swap_cache() updates some statistics counter,
4751 * we call find_get_page() with swapper_space directly.
4753 page = find_get_page(swap_address_space(ent), swp_offset(ent));
4754 if (do_memsw_account())
4755 entry->val = ent.val;
4760 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4761 pte_t ptent, swp_entry_t *entry)
4767 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4768 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4770 struct page *page = NULL;
4771 struct address_space *mapping;
4774 if (!vma->vm_file) /* anonymous vma */
4776 if (!(mc.flags & MOVE_FILE))
4779 mapping = vma->vm_file->f_mapping;
4780 pgoff = linear_page_index(vma, addr);
4782 /* page is moved even if it's not RSS of this task(page-faulted). */
4784 /* shmem/tmpfs may report page out on swap: account for that too. */
4785 if (shmem_mapping(mapping)) {
4786 page = find_get_entry(mapping, pgoff);
4787 if (xa_is_value(page)) {
4788 swp_entry_t swp = radix_to_swp_entry(page);
4789 if (do_memsw_account())
4791 page = find_get_page(swap_address_space(swp),
4795 page = find_get_page(mapping, pgoff);
4797 page = find_get_page(mapping, pgoff);
4803 * mem_cgroup_move_account - move account of the page
4805 * @compound: charge the page as compound or small page
4806 * @from: mem_cgroup which the page is moved from.
4807 * @to: mem_cgroup which the page is moved to. @from != @to.
4809 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
4811 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4814 static int mem_cgroup_move_account(struct page *page,
4816 struct mem_cgroup *from,
4817 struct mem_cgroup *to)
4819 unsigned long flags;
4820 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
4824 VM_BUG_ON(from == to);
4825 VM_BUG_ON_PAGE(PageLRU(page), page);
4826 VM_BUG_ON(compound && !PageTransHuge(page));
4829 * Prevent mem_cgroup_migrate() from looking at
4830 * page->mem_cgroup of its source page while we change it.
4833 if (!trylock_page(page))
4837 if (page->mem_cgroup != from)
4840 anon = PageAnon(page);
4842 spin_lock_irqsave(&from->move_lock, flags);
4844 if (!anon && page_mapped(page)) {
4845 __mod_memcg_state(from, NR_FILE_MAPPED, -nr_pages);
4846 __mod_memcg_state(to, NR_FILE_MAPPED, nr_pages);
4850 * move_lock grabbed above and caller set from->moving_account, so
4851 * mod_memcg_page_state will serialize updates to PageDirty.
4852 * So mapping should be stable for dirty pages.
4854 if (!anon && PageDirty(page)) {
4855 struct address_space *mapping = page_mapping(page);
4857 if (mapping_cap_account_dirty(mapping)) {
4858 __mod_memcg_state(from, NR_FILE_DIRTY, -nr_pages);
4859 __mod_memcg_state(to, NR_FILE_DIRTY, nr_pages);
4863 if (PageWriteback(page)) {
4864 __mod_memcg_state(from, NR_WRITEBACK, -nr_pages);
4865 __mod_memcg_state(to, NR_WRITEBACK, nr_pages);
4869 * It is safe to change page->mem_cgroup here because the page
4870 * is referenced, charged, and isolated - we can't race with
4871 * uncharging, charging, migration, or LRU putback.
4874 /* caller should have done css_get */
4875 page->mem_cgroup = to;
4876 spin_unlock_irqrestore(&from->move_lock, flags);
4880 local_irq_disable();
4881 mem_cgroup_charge_statistics(to, page, compound, nr_pages);
4882 memcg_check_events(to, page);
4883 mem_cgroup_charge_statistics(from, page, compound, -nr_pages);
4884 memcg_check_events(from, page);
4893 * get_mctgt_type - get target type of moving charge
4894 * @vma: the vma the pte to be checked belongs
4895 * @addr: the address corresponding to the pte to be checked
4896 * @ptent: the pte to be checked
4897 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4900 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4901 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4902 * move charge. if @target is not NULL, the page is stored in target->page
4903 * with extra refcnt got(Callers should handle it).
4904 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4905 * target for charge migration. if @target is not NULL, the entry is stored
4907 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PUBLIC
4908 * or MEMORY_DEVICE_PRIVATE (so ZONE_DEVICE page and thus not on the lru).
4909 * For now we such page is charge like a regular page would be as for all
4910 * intent and purposes it is just special memory taking the place of a
4913 * See Documentations/vm/hmm.txt and include/linux/hmm.h
4915 * Called with pte lock held.
4918 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
4919 unsigned long addr, pte_t ptent, union mc_target *target)
4921 struct page *page = NULL;
4922 enum mc_target_type ret = MC_TARGET_NONE;
4923 swp_entry_t ent = { .val = 0 };
4925 if (pte_present(ptent))
4926 page = mc_handle_present_pte(vma, addr, ptent);
4927 else if (is_swap_pte(ptent))
4928 page = mc_handle_swap_pte(vma, ptent, &ent);
4929 else if (pte_none(ptent))
4930 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4932 if (!page && !ent.val)
4936 * Do only loose check w/o serialization.
4937 * mem_cgroup_move_account() checks the page is valid or
4938 * not under LRU exclusion.
4940 if (page->mem_cgroup == mc.from) {
4941 ret = MC_TARGET_PAGE;
4942 if (is_device_private_page(page) ||
4943 is_device_public_page(page))
4944 ret = MC_TARGET_DEVICE;
4946 target->page = page;
4948 if (!ret || !target)
4952 * There is a swap entry and a page doesn't exist or isn't charged.
4953 * But we cannot move a tail-page in a THP.
4955 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
4956 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
4957 ret = MC_TARGET_SWAP;
4964 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4966 * We don't consider PMD mapped swapping or file mapped pages because THP does
4967 * not support them for now.
4968 * Caller should make sure that pmd_trans_huge(pmd) is true.
4970 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4971 unsigned long addr, pmd_t pmd, union mc_target *target)
4973 struct page *page = NULL;
4974 enum mc_target_type ret = MC_TARGET_NONE;
4976 if (unlikely(is_swap_pmd(pmd))) {
4977 VM_BUG_ON(thp_migration_supported() &&
4978 !is_pmd_migration_entry(pmd));
4981 page = pmd_page(pmd);
4982 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
4983 if (!(mc.flags & MOVE_ANON))
4985 if (page->mem_cgroup == mc.from) {
4986 ret = MC_TARGET_PAGE;
4989 target->page = page;
4995 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4996 unsigned long addr, pmd_t pmd, union mc_target *target)
4998 return MC_TARGET_NONE;
5002 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5003 unsigned long addr, unsigned long end,
5004 struct mm_walk *walk)
5006 struct vm_area_struct *vma = walk->vma;
5010 ptl = pmd_trans_huge_lock(pmd, vma);
5013 * Note their can not be MC_TARGET_DEVICE for now as we do not
5014 * support transparent huge page with MEMORY_DEVICE_PUBLIC or
5015 * MEMORY_DEVICE_PRIVATE but this might change.
5017 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5018 mc.precharge += HPAGE_PMD_NR;
5023 if (pmd_trans_unstable(pmd))
5025 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5026 for (; addr != end; pte++, addr += PAGE_SIZE)
5027 if (get_mctgt_type(vma, addr, *pte, NULL))
5028 mc.precharge++; /* increment precharge temporarily */
5029 pte_unmap_unlock(pte - 1, ptl);
5035 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5037 unsigned long precharge;
5039 struct mm_walk mem_cgroup_count_precharge_walk = {
5040 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5043 down_read(&mm->mmap_sem);
5044 walk_page_range(0, mm->highest_vm_end,
5045 &mem_cgroup_count_precharge_walk);
5046 up_read(&mm->mmap_sem);
5048 precharge = mc.precharge;
5054 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5056 unsigned long precharge = mem_cgroup_count_precharge(mm);
5058 VM_BUG_ON(mc.moving_task);
5059 mc.moving_task = current;
5060 return mem_cgroup_do_precharge(precharge);
5063 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5064 static void __mem_cgroup_clear_mc(void)
5066 struct mem_cgroup *from = mc.from;
5067 struct mem_cgroup *to = mc.to;
5069 /* we must uncharge all the leftover precharges from mc.to */
5071 cancel_charge(mc.to, mc.precharge);
5075 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5076 * we must uncharge here.
5078 if (mc.moved_charge) {
5079 cancel_charge(mc.from, mc.moved_charge);
5080 mc.moved_charge = 0;
5082 /* we must fixup refcnts and charges */
5083 if (mc.moved_swap) {
5084 /* uncharge swap account from the old cgroup */
5085 if (!mem_cgroup_is_root(mc.from))
5086 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5088 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5091 * we charged both to->memory and to->memsw, so we
5092 * should uncharge to->memory.
5094 if (!mem_cgroup_is_root(mc.to))
5095 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5097 mem_cgroup_id_get_many(mc.to, mc.moved_swap);
5098 css_put_many(&mc.to->css, mc.moved_swap);
5102 memcg_oom_recover(from);
5103 memcg_oom_recover(to);
5104 wake_up_all(&mc.waitq);
5107 static void mem_cgroup_clear_mc(void)
5109 struct mm_struct *mm = mc.mm;
5112 * we must clear moving_task before waking up waiters at the end of
5115 mc.moving_task = NULL;
5116 __mem_cgroup_clear_mc();
5117 spin_lock(&mc.lock);
5121 spin_unlock(&mc.lock);
5126 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5128 struct cgroup_subsys_state *css;
5129 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5130 struct mem_cgroup *from;
5131 struct task_struct *leader, *p;
5132 struct mm_struct *mm;
5133 unsigned long move_flags;
5136 /* charge immigration isn't supported on the default hierarchy */
5137 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5141 * Multi-process migrations only happen on the default hierarchy
5142 * where charge immigration is not used. Perform charge
5143 * immigration if @tset contains a leader and whine if there are
5147 cgroup_taskset_for_each_leader(leader, css, tset) {
5150 memcg = mem_cgroup_from_css(css);
5156 * We are now commited to this value whatever it is. Changes in this
5157 * tunable will only affect upcoming migrations, not the current one.
5158 * So we need to save it, and keep it going.
5160 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5164 from = mem_cgroup_from_task(p);
5166 VM_BUG_ON(from == memcg);
5168 mm = get_task_mm(p);
5171 /* We move charges only when we move a owner of the mm */
5172 if (mm->owner == p) {
5175 VM_BUG_ON(mc.precharge);
5176 VM_BUG_ON(mc.moved_charge);
5177 VM_BUG_ON(mc.moved_swap);
5179 spin_lock(&mc.lock);
5183 mc.flags = move_flags;
5184 spin_unlock(&mc.lock);
5185 /* We set mc.moving_task later */
5187 ret = mem_cgroup_precharge_mc(mm);
5189 mem_cgroup_clear_mc();
5196 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5199 mem_cgroup_clear_mc();
5202 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5203 unsigned long addr, unsigned long end,
5204 struct mm_walk *walk)
5207 struct vm_area_struct *vma = walk->vma;
5210 enum mc_target_type target_type;
5211 union mc_target target;
5214 ptl = pmd_trans_huge_lock(pmd, vma);
5216 if (mc.precharge < HPAGE_PMD_NR) {
5220 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5221 if (target_type == MC_TARGET_PAGE) {
5223 if (!isolate_lru_page(page)) {
5224 if (!mem_cgroup_move_account(page, true,
5226 mc.precharge -= HPAGE_PMD_NR;
5227 mc.moved_charge += HPAGE_PMD_NR;
5229 putback_lru_page(page);
5232 } else if (target_type == MC_TARGET_DEVICE) {
5234 if (!mem_cgroup_move_account(page, true,
5236 mc.precharge -= HPAGE_PMD_NR;
5237 mc.moved_charge += HPAGE_PMD_NR;
5245 if (pmd_trans_unstable(pmd))
5248 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5249 for (; addr != end; addr += PAGE_SIZE) {
5250 pte_t ptent = *(pte++);
5251 bool device = false;
5257 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5258 case MC_TARGET_DEVICE:
5261 case MC_TARGET_PAGE:
5264 * We can have a part of the split pmd here. Moving it
5265 * can be done but it would be too convoluted so simply
5266 * ignore such a partial THP and keep it in original
5267 * memcg. There should be somebody mapping the head.
5269 if (PageTransCompound(page))
5271 if (!device && isolate_lru_page(page))
5273 if (!mem_cgroup_move_account(page, false,
5276 /* we uncharge from mc.from later. */
5280 putback_lru_page(page);
5281 put: /* get_mctgt_type() gets the page */
5284 case MC_TARGET_SWAP:
5286 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5288 /* we fixup refcnts and charges later. */
5296 pte_unmap_unlock(pte - 1, ptl);
5301 * We have consumed all precharges we got in can_attach().
5302 * We try charge one by one, but don't do any additional
5303 * charges to mc.to if we have failed in charge once in attach()
5306 ret = mem_cgroup_do_precharge(1);
5314 static void mem_cgroup_move_charge(void)
5316 struct mm_walk mem_cgroup_move_charge_walk = {
5317 .pmd_entry = mem_cgroup_move_charge_pte_range,
5321 lru_add_drain_all();
5323 * Signal lock_page_memcg() to take the memcg's move_lock
5324 * while we're moving its pages to another memcg. Then wait
5325 * for already started RCU-only updates to finish.
5327 atomic_inc(&mc.from->moving_account);
5330 if (unlikely(!down_read_trylock(&mc.mm->mmap_sem))) {
5332 * Someone who are holding the mmap_sem might be waiting in
5333 * waitq. So we cancel all extra charges, wake up all waiters,
5334 * and retry. Because we cancel precharges, we might not be able
5335 * to move enough charges, but moving charge is a best-effort
5336 * feature anyway, so it wouldn't be a big problem.
5338 __mem_cgroup_clear_mc();
5343 * When we have consumed all precharges and failed in doing
5344 * additional charge, the page walk just aborts.
5346 walk_page_range(0, mc.mm->highest_vm_end, &mem_cgroup_move_charge_walk);
5348 up_read(&mc.mm->mmap_sem);
5349 atomic_dec(&mc.from->moving_account);
5352 static void mem_cgroup_move_task(void)
5355 mem_cgroup_move_charge();
5356 mem_cgroup_clear_mc();
5359 #else /* !CONFIG_MMU */
5360 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5364 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5367 static void mem_cgroup_move_task(void)
5373 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5374 * to verify whether we're attached to the default hierarchy on each mount
5377 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5380 * use_hierarchy is forced on the default hierarchy. cgroup core
5381 * guarantees that @root doesn't have any children, so turning it
5382 * on for the root memcg is enough.
5384 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5385 root_mem_cgroup->use_hierarchy = true;
5387 root_mem_cgroup->use_hierarchy = false;
5390 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
5392 if (value == PAGE_COUNTER_MAX)
5393 seq_puts(m, "max\n");
5395 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
5400 static u64 memory_current_read(struct cgroup_subsys_state *css,
5403 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5405 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
5408 static int memory_min_show(struct seq_file *m, void *v)
5410 return seq_puts_memcg_tunable(m,
5411 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
5414 static ssize_t memory_min_write(struct kernfs_open_file *of,
5415 char *buf, size_t nbytes, loff_t off)
5417 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5421 buf = strstrip(buf);
5422 err = page_counter_memparse(buf, "max", &min);
5426 page_counter_set_min(&memcg->memory, min);
5431 static int memory_low_show(struct seq_file *m, void *v)
5433 return seq_puts_memcg_tunable(m,
5434 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
5437 static ssize_t memory_low_write(struct kernfs_open_file *of,
5438 char *buf, size_t nbytes, loff_t off)
5440 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5444 buf = strstrip(buf);
5445 err = page_counter_memparse(buf, "max", &low);
5449 page_counter_set_low(&memcg->memory, low);
5454 static int memory_high_show(struct seq_file *m, void *v)
5456 return seq_puts_memcg_tunable(m, READ_ONCE(mem_cgroup_from_seq(m)->high));
5459 static ssize_t memory_high_write(struct kernfs_open_file *of,
5460 char *buf, size_t nbytes, loff_t off)
5462 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5463 unsigned long nr_pages;
5467 buf = strstrip(buf);
5468 err = page_counter_memparse(buf, "max", &high);
5474 nr_pages = page_counter_read(&memcg->memory);
5475 if (nr_pages > high)
5476 try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
5479 memcg_wb_domain_size_changed(memcg);
5483 static int memory_max_show(struct seq_file *m, void *v)
5485 return seq_puts_memcg_tunable(m,
5486 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
5489 static ssize_t memory_max_write(struct kernfs_open_file *of,
5490 char *buf, size_t nbytes, loff_t off)
5492 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5493 unsigned int nr_reclaims = MEM_CGROUP_RECLAIM_RETRIES;
5494 bool drained = false;
5498 buf = strstrip(buf);
5499 err = page_counter_memparse(buf, "max", &max);
5503 xchg(&memcg->memory.max, max);
5506 unsigned long nr_pages = page_counter_read(&memcg->memory);
5508 if (nr_pages <= max)
5511 if (signal_pending(current)) {
5517 drain_all_stock(memcg);
5523 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
5529 memcg_memory_event(memcg, MEMCG_OOM);
5530 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
5534 memcg_wb_domain_size_changed(memcg);
5538 static int memory_events_show(struct seq_file *m, void *v)
5540 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
5542 seq_printf(m, "low %lu\n",
5543 atomic_long_read(&memcg->memory_events[MEMCG_LOW]));
5544 seq_printf(m, "high %lu\n",
5545 atomic_long_read(&memcg->memory_events[MEMCG_HIGH]));
5546 seq_printf(m, "max %lu\n",
5547 atomic_long_read(&memcg->memory_events[MEMCG_MAX]));
5548 seq_printf(m, "oom %lu\n",
5549 atomic_long_read(&memcg->memory_events[MEMCG_OOM]));
5550 seq_printf(m, "oom_kill %lu\n",
5551 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
5556 static int memory_stat_show(struct seq_file *m, void *v)
5558 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
5559 struct accumulated_stats acc;
5563 * Provide statistics on the state of the memory subsystem as
5564 * well as cumulative event counters that show past behavior.
5566 * This list is ordered following a combination of these gradients:
5567 * 1) generic big picture -> specifics and details
5568 * 2) reflecting userspace activity -> reflecting kernel heuristics
5570 * Current memory state:
5573 memset(&acc, 0, sizeof(acc));
5574 acc.stats_size = MEMCG_NR_STAT;
5575 acc.events_size = NR_VM_EVENT_ITEMS;
5576 accumulate_memcg_tree(memcg, &acc);
5578 seq_printf(m, "anon %llu\n",
5579 (u64)acc.stat[MEMCG_RSS] * PAGE_SIZE);
5580 seq_printf(m, "file %llu\n",
5581 (u64)acc.stat[MEMCG_CACHE] * PAGE_SIZE);
5582 seq_printf(m, "kernel_stack %llu\n",
5583 (u64)acc.stat[MEMCG_KERNEL_STACK_KB] * 1024);
5584 seq_printf(m, "slab %llu\n",
5585 (u64)(acc.stat[NR_SLAB_RECLAIMABLE] +
5586 acc.stat[NR_SLAB_UNRECLAIMABLE]) * PAGE_SIZE);
5587 seq_printf(m, "sock %llu\n",
5588 (u64)acc.stat[MEMCG_SOCK] * PAGE_SIZE);
5590 seq_printf(m, "shmem %llu\n",
5591 (u64)acc.stat[NR_SHMEM] * PAGE_SIZE);
5592 seq_printf(m, "file_mapped %llu\n",
5593 (u64)acc.stat[NR_FILE_MAPPED] * PAGE_SIZE);
5594 seq_printf(m, "file_dirty %llu\n",
5595 (u64)acc.stat[NR_FILE_DIRTY] * PAGE_SIZE);
5596 seq_printf(m, "file_writeback %llu\n",
5597 (u64)acc.stat[NR_WRITEBACK] * PAGE_SIZE);
5600 * TODO: We should eventually replace our own MEMCG_RSS_HUGE counter
5601 * with the NR_ANON_THP vm counter, but right now it's a pain in the
5602 * arse because it requires migrating the work out of rmap to a place
5603 * where the page->mem_cgroup is set up and stable.
5605 seq_printf(m, "anon_thp %llu\n",
5606 (u64)acc.stat[MEMCG_RSS_HUGE] * PAGE_SIZE);
5608 for (i = 0; i < NR_LRU_LISTS; i++)
5609 seq_printf(m, "%s %llu\n", mem_cgroup_lru_names[i],
5610 (u64)acc.lru_pages[i] * PAGE_SIZE);
5612 seq_printf(m, "slab_reclaimable %llu\n",
5613 (u64)acc.stat[NR_SLAB_RECLAIMABLE] * PAGE_SIZE);
5614 seq_printf(m, "slab_unreclaimable %llu\n",
5615 (u64)acc.stat[NR_SLAB_UNRECLAIMABLE] * PAGE_SIZE);
5617 /* Accumulated memory events */
5619 seq_printf(m, "pgfault %lu\n", acc.events[PGFAULT]);
5620 seq_printf(m, "pgmajfault %lu\n", acc.events[PGMAJFAULT]);
5622 seq_printf(m, "workingset_refault %lu\n",
5623 acc.stat[WORKINGSET_REFAULT]);
5624 seq_printf(m, "workingset_activate %lu\n",
5625 acc.stat[WORKINGSET_ACTIVATE]);
5626 seq_printf(m, "workingset_nodereclaim %lu\n",
5627 acc.stat[WORKINGSET_NODERECLAIM]);
5629 seq_printf(m, "pgrefill %lu\n", acc.events[PGREFILL]);
5630 seq_printf(m, "pgscan %lu\n", acc.events[PGSCAN_KSWAPD] +
5631 acc.events[PGSCAN_DIRECT]);
5632 seq_printf(m, "pgsteal %lu\n", acc.events[PGSTEAL_KSWAPD] +
5633 acc.events[PGSTEAL_DIRECT]);
5634 seq_printf(m, "pgactivate %lu\n", acc.events[PGACTIVATE]);
5635 seq_printf(m, "pgdeactivate %lu\n", acc.events[PGDEACTIVATE]);
5636 seq_printf(m, "pglazyfree %lu\n", acc.events[PGLAZYFREE]);
5637 seq_printf(m, "pglazyfreed %lu\n", acc.events[PGLAZYFREED]);
5639 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5640 seq_printf(m, "thp_fault_alloc %lu\n", acc.events[THP_FAULT_ALLOC]);
5641 seq_printf(m, "thp_collapse_alloc %lu\n",
5642 acc.events[THP_COLLAPSE_ALLOC]);
5643 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
5648 static int memory_oom_group_show(struct seq_file *m, void *v)
5650 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
5652 seq_printf(m, "%d\n", memcg->oom_group);
5657 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
5658 char *buf, size_t nbytes, loff_t off)
5660 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5663 buf = strstrip(buf);
5667 ret = kstrtoint(buf, 0, &oom_group);
5671 if (oom_group != 0 && oom_group != 1)
5674 memcg->oom_group = oom_group;
5679 static struct cftype memory_files[] = {
5682 .flags = CFTYPE_NOT_ON_ROOT,
5683 .read_u64 = memory_current_read,
5687 .flags = CFTYPE_NOT_ON_ROOT,
5688 .seq_show = memory_min_show,
5689 .write = memory_min_write,
5693 .flags = CFTYPE_NOT_ON_ROOT,
5694 .seq_show = memory_low_show,
5695 .write = memory_low_write,
5699 .flags = CFTYPE_NOT_ON_ROOT,
5700 .seq_show = memory_high_show,
5701 .write = memory_high_write,
5705 .flags = CFTYPE_NOT_ON_ROOT,
5706 .seq_show = memory_max_show,
5707 .write = memory_max_write,
5711 .flags = CFTYPE_NOT_ON_ROOT,
5712 .file_offset = offsetof(struct mem_cgroup, events_file),
5713 .seq_show = memory_events_show,
5717 .flags = CFTYPE_NOT_ON_ROOT,
5718 .seq_show = memory_stat_show,
5721 .name = "oom.group",
5722 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
5723 .seq_show = memory_oom_group_show,
5724 .write = memory_oom_group_write,
5729 struct cgroup_subsys memory_cgrp_subsys = {
5730 .css_alloc = mem_cgroup_css_alloc,
5731 .css_online = mem_cgroup_css_online,
5732 .css_offline = mem_cgroup_css_offline,
5733 .css_released = mem_cgroup_css_released,
5734 .css_free = mem_cgroup_css_free,
5735 .css_reset = mem_cgroup_css_reset,
5736 .can_attach = mem_cgroup_can_attach,
5737 .cancel_attach = mem_cgroup_cancel_attach,
5738 .post_attach = mem_cgroup_move_task,
5739 .bind = mem_cgroup_bind,
5740 .dfl_cftypes = memory_files,
5741 .legacy_cftypes = mem_cgroup_legacy_files,
5746 * mem_cgroup_protected - check if memory consumption is in the normal range
5747 * @root: the top ancestor of the sub-tree being checked
5748 * @memcg: the memory cgroup to check
5750 * WARNING: This function is not stateless! It can only be used as part
5751 * of a top-down tree iteration, not for isolated queries.
5753 * Returns one of the following:
5754 * MEMCG_PROT_NONE: cgroup memory is not protected
5755 * MEMCG_PROT_LOW: cgroup memory is protected as long there is
5756 * an unprotected supply of reclaimable memory from other cgroups.
5757 * MEMCG_PROT_MIN: cgroup memory is protected
5759 * @root is exclusive; it is never protected when looked at directly
5761 * To provide a proper hierarchical behavior, effective memory.min/low values
5762 * are used. Below is the description of how effective memory.low is calculated.
5763 * Effective memory.min values is calculated in the same way.
5765 * Effective memory.low is always equal or less than the original memory.low.
5766 * If there is no memory.low overcommittment (which is always true for
5767 * top-level memory cgroups), these two values are equal.
5768 * Otherwise, it's a part of parent's effective memory.low,
5769 * calculated as a cgroup's memory.low usage divided by sum of sibling's
5770 * memory.low usages, where memory.low usage is the size of actually
5774 * elow = min( memory.low, parent->elow * ------------------ ),
5775 * siblings_low_usage
5777 * | memory.current, if memory.current < memory.low
5782 * Such definition of the effective memory.low provides the expected
5783 * hierarchical behavior: parent's memory.low value is limiting
5784 * children, unprotected memory is reclaimed first and cgroups,
5785 * which are not using their guarantee do not affect actual memory
5788 * For example, if there are memcgs A, A/B, A/C, A/D and A/E:
5790 * A A/memory.low = 2G, A/memory.current = 6G
5792 * BC DE B/memory.low = 3G B/memory.current = 2G
5793 * C/memory.low = 1G C/memory.current = 2G
5794 * D/memory.low = 0 D/memory.current = 2G
5795 * E/memory.low = 10G E/memory.current = 0
5797 * and the memory pressure is applied, the following memory distribution
5798 * is expected (approximately):
5800 * A/memory.current = 2G
5802 * B/memory.current = 1.3G
5803 * C/memory.current = 0.6G
5804 * D/memory.current = 0
5805 * E/memory.current = 0
5807 * These calculations require constant tracking of the actual low usages
5808 * (see propagate_protected_usage()), as well as recursive calculation of
5809 * effective memory.low values. But as we do call mem_cgroup_protected()
5810 * path for each memory cgroup top-down from the reclaim,
5811 * it's possible to optimize this part, and save calculated elow
5812 * for next usage. This part is intentionally racy, but it's ok,
5813 * as memory.low is a best-effort mechanism.
5815 enum mem_cgroup_protection mem_cgroup_protected(struct mem_cgroup *root,
5816 struct mem_cgroup *memcg)
5818 struct mem_cgroup *parent;
5819 unsigned long emin, parent_emin;
5820 unsigned long elow, parent_elow;
5821 unsigned long usage;
5823 if (mem_cgroup_disabled())
5824 return MEMCG_PROT_NONE;
5827 root = root_mem_cgroup;
5829 return MEMCG_PROT_NONE;
5831 usage = page_counter_read(&memcg->memory);
5833 return MEMCG_PROT_NONE;
5835 emin = memcg->memory.min;
5836 elow = memcg->memory.low;
5838 parent = parent_mem_cgroup(memcg);
5839 /* No parent means a non-hierarchical mode on v1 memcg */
5841 return MEMCG_PROT_NONE;
5846 parent_emin = READ_ONCE(parent->memory.emin);
5847 emin = min(emin, parent_emin);
5848 if (emin && parent_emin) {
5849 unsigned long min_usage, siblings_min_usage;
5851 min_usage = min(usage, memcg->memory.min);
5852 siblings_min_usage = atomic_long_read(
5853 &parent->memory.children_min_usage);
5855 if (min_usage && siblings_min_usage)
5856 emin = min(emin, parent_emin * min_usage /
5857 siblings_min_usage);
5860 parent_elow = READ_ONCE(parent->memory.elow);
5861 elow = min(elow, parent_elow);
5862 if (elow && parent_elow) {
5863 unsigned long low_usage, siblings_low_usage;
5865 low_usage = min(usage, memcg->memory.low);
5866 siblings_low_usage = atomic_long_read(
5867 &parent->memory.children_low_usage);
5869 if (low_usage && siblings_low_usage)
5870 elow = min(elow, parent_elow * low_usage /
5871 siblings_low_usage);
5875 memcg->memory.emin = emin;
5876 memcg->memory.elow = elow;
5879 return MEMCG_PROT_MIN;
5880 else if (usage <= elow)
5881 return MEMCG_PROT_LOW;
5883 return MEMCG_PROT_NONE;
5887 * mem_cgroup_try_charge - try charging a page
5888 * @page: page to charge
5889 * @mm: mm context of the victim
5890 * @gfp_mask: reclaim mode
5891 * @memcgp: charged memcg return
5892 * @compound: charge the page as compound or small page
5894 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5895 * pages according to @gfp_mask if necessary.
5897 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5898 * Otherwise, an error code is returned.
5900 * After page->mapping has been set up, the caller must finalize the
5901 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5902 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5904 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
5905 gfp_t gfp_mask, struct mem_cgroup **memcgp,
5908 struct mem_cgroup *memcg = NULL;
5909 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5912 if (mem_cgroup_disabled())
5915 if (PageSwapCache(page)) {
5917 * Every swap fault against a single page tries to charge the
5918 * page, bail as early as possible. shmem_unuse() encounters
5919 * already charged pages, too. The USED bit is protected by
5920 * the page lock, which serializes swap cache removal, which
5921 * in turn serializes uncharging.
5923 VM_BUG_ON_PAGE(!PageLocked(page), page);
5924 if (compound_head(page)->mem_cgroup)
5927 if (do_swap_account) {
5928 swp_entry_t ent = { .val = page_private(page), };
5929 unsigned short id = lookup_swap_cgroup_id(ent);
5932 memcg = mem_cgroup_from_id(id);
5933 if (memcg && !css_tryget_online(&memcg->css))
5940 memcg = get_mem_cgroup_from_mm(mm);
5942 ret = try_charge(memcg, gfp_mask, nr_pages);
5944 css_put(&memcg->css);
5950 int mem_cgroup_try_charge_delay(struct page *page, struct mm_struct *mm,
5951 gfp_t gfp_mask, struct mem_cgroup **memcgp,
5954 struct mem_cgroup *memcg;
5957 ret = mem_cgroup_try_charge(page, mm, gfp_mask, memcgp, compound);
5959 mem_cgroup_throttle_swaprate(memcg, page_to_nid(page), gfp_mask);
5964 * mem_cgroup_commit_charge - commit a page charge
5965 * @page: page to charge
5966 * @memcg: memcg to charge the page to
5967 * @lrucare: page might be on LRU already
5968 * @compound: charge the page as compound or small page
5970 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5971 * after page->mapping has been set up. This must happen atomically
5972 * as part of the page instantiation, i.e. under the page table lock
5973 * for anonymous pages, under the page lock for page and swap cache.
5975 * In addition, the page must not be on the LRU during the commit, to
5976 * prevent racing with task migration. If it might be, use @lrucare.
5978 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5980 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
5981 bool lrucare, bool compound)
5983 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5985 VM_BUG_ON_PAGE(!page->mapping, page);
5986 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
5988 if (mem_cgroup_disabled())
5991 * Swap faults will attempt to charge the same page multiple
5992 * times. But reuse_swap_page() might have removed the page
5993 * from swapcache already, so we can't check PageSwapCache().
5998 commit_charge(page, memcg, lrucare);
6000 local_irq_disable();
6001 mem_cgroup_charge_statistics(memcg, page, compound, nr_pages);
6002 memcg_check_events(memcg, page);
6005 if (do_memsw_account() && PageSwapCache(page)) {
6006 swp_entry_t entry = { .val = page_private(page) };
6008 * The swap entry might not get freed for a long time,
6009 * let's not wait for it. The page already received a
6010 * memory+swap charge, drop the swap entry duplicate.
6012 mem_cgroup_uncharge_swap(entry, nr_pages);
6017 * mem_cgroup_cancel_charge - cancel a page charge
6018 * @page: page to charge
6019 * @memcg: memcg to charge the page to
6020 * @compound: charge the page as compound or small page
6022 * Cancel a charge transaction started by mem_cgroup_try_charge().
6024 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg,
6027 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6029 if (mem_cgroup_disabled())
6032 * Swap faults will attempt to charge the same page multiple
6033 * times. But reuse_swap_page() might have removed the page
6034 * from swapcache already, so we can't check PageSwapCache().
6039 cancel_charge(memcg, nr_pages);
6042 struct uncharge_gather {
6043 struct mem_cgroup *memcg;
6044 unsigned long pgpgout;
6045 unsigned long nr_anon;
6046 unsigned long nr_file;
6047 unsigned long nr_kmem;
6048 unsigned long nr_huge;
6049 unsigned long nr_shmem;
6050 struct page *dummy_page;
6053 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6055 memset(ug, 0, sizeof(*ug));
6058 static void uncharge_batch(const struct uncharge_gather *ug)
6060 unsigned long nr_pages = ug->nr_anon + ug->nr_file + ug->nr_kmem;
6061 unsigned long flags;
6063 if (!mem_cgroup_is_root(ug->memcg)) {
6064 page_counter_uncharge(&ug->memcg->memory, nr_pages);
6065 if (do_memsw_account())
6066 page_counter_uncharge(&ug->memcg->memsw, nr_pages);
6067 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6068 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6069 memcg_oom_recover(ug->memcg);
6072 local_irq_save(flags);
6073 __mod_memcg_state(ug->memcg, MEMCG_RSS, -ug->nr_anon);
6074 __mod_memcg_state(ug->memcg, MEMCG_CACHE, -ug->nr_file);
6075 __mod_memcg_state(ug->memcg, MEMCG_RSS_HUGE, -ug->nr_huge);
6076 __mod_memcg_state(ug->memcg, NR_SHMEM, -ug->nr_shmem);
6077 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6078 __this_cpu_add(ug->memcg->stat_cpu->nr_page_events, nr_pages);
6079 memcg_check_events(ug->memcg, ug->dummy_page);
6080 local_irq_restore(flags);
6082 if (!mem_cgroup_is_root(ug->memcg))
6083 css_put_many(&ug->memcg->css, nr_pages);
6086 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6088 VM_BUG_ON_PAGE(PageLRU(page), page);
6089 VM_BUG_ON_PAGE(page_count(page) && !is_zone_device_page(page) &&
6090 !PageHWPoison(page) , page);
6092 if (!page->mem_cgroup)
6096 * Nobody should be changing or seriously looking at
6097 * page->mem_cgroup at this point, we have fully
6098 * exclusive access to the page.
6101 if (ug->memcg != page->mem_cgroup) {
6104 uncharge_gather_clear(ug);
6106 ug->memcg = page->mem_cgroup;
6109 if (!PageKmemcg(page)) {
6110 unsigned int nr_pages = 1;
6112 if (PageTransHuge(page)) {
6113 nr_pages <<= compound_order(page);
6114 ug->nr_huge += nr_pages;
6117 ug->nr_anon += nr_pages;
6119 ug->nr_file += nr_pages;
6120 if (PageSwapBacked(page))
6121 ug->nr_shmem += nr_pages;
6125 ug->nr_kmem += 1 << compound_order(page);
6126 __ClearPageKmemcg(page);
6129 ug->dummy_page = page;
6130 page->mem_cgroup = NULL;
6133 static void uncharge_list(struct list_head *page_list)
6135 struct uncharge_gather ug;
6136 struct list_head *next;
6138 uncharge_gather_clear(&ug);
6141 * Note that the list can be a single page->lru; hence the
6142 * do-while loop instead of a simple list_for_each_entry().
6144 next = page_list->next;
6148 page = list_entry(next, struct page, lru);
6149 next = page->lru.next;
6151 uncharge_page(page, &ug);
6152 } while (next != page_list);
6155 uncharge_batch(&ug);
6159 * mem_cgroup_uncharge - uncharge a page
6160 * @page: page to uncharge
6162 * Uncharge a page previously charged with mem_cgroup_try_charge() and
6163 * mem_cgroup_commit_charge().
6165 void mem_cgroup_uncharge(struct page *page)
6167 struct uncharge_gather ug;
6169 if (mem_cgroup_disabled())
6172 /* Don't touch page->lru of any random page, pre-check: */
6173 if (!page->mem_cgroup)
6176 uncharge_gather_clear(&ug);
6177 uncharge_page(page, &ug);
6178 uncharge_batch(&ug);
6182 * mem_cgroup_uncharge_list - uncharge a list of page
6183 * @page_list: list of pages to uncharge
6185 * Uncharge a list of pages previously charged with
6186 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
6188 void mem_cgroup_uncharge_list(struct list_head *page_list)
6190 if (mem_cgroup_disabled())
6193 if (!list_empty(page_list))
6194 uncharge_list(page_list);
6198 * mem_cgroup_migrate - charge a page's replacement
6199 * @oldpage: currently circulating page
6200 * @newpage: replacement page
6202 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6203 * be uncharged upon free.
6205 * Both pages must be locked, @newpage->mapping must be set up.
6207 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6209 struct mem_cgroup *memcg;
6210 unsigned int nr_pages;
6212 unsigned long flags;
6214 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6215 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6216 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6217 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6220 if (mem_cgroup_disabled())
6223 /* Page cache replacement: new page already charged? */
6224 if (newpage->mem_cgroup)
6227 /* Swapcache readahead pages can get replaced before being charged */
6228 memcg = oldpage->mem_cgroup;
6232 /* Force-charge the new page. The old one will be freed soon */
6233 compound = PageTransHuge(newpage);
6234 nr_pages = compound ? hpage_nr_pages(newpage) : 1;
6236 page_counter_charge(&memcg->memory, nr_pages);
6237 if (do_memsw_account())
6238 page_counter_charge(&memcg->memsw, nr_pages);
6239 css_get_many(&memcg->css, nr_pages);
6241 commit_charge(newpage, memcg, false);
6243 local_irq_save(flags);
6244 mem_cgroup_charge_statistics(memcg, newpage, compound, nr_pages);
6245 memcg_check_events(memcg, newpage);
6246 local_irq_restore(flags);
6249 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
6250 EXPORT_SYMBOL(memcg_sockets_enabled_key);
6252 void mem_cgroup_sk_alloc(struct sock *sk)
6254 struct mem_cgroup *memcg;
6256 if (!mem_cgroup_sockets_enabled)
6260 * Socket cloning can throw us here with sk_memcg already
6261 * filled. It won't however, necessarily happen from
6262 * process context. So the test for root memcg given
6263 * the current task's memcg won't help us in this case.
6265 * Respecting the original socket's memcg is a better
6266 * decision in this case.
6269 css_get(&sk->sk_memcg->css);
6274 memcg = mem_cgroup_from_task(current);
6275 if (memcg == root_mem_cgroup)
6277 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
6279 if (css_tryget_online(&memcg->css))
6280 sk->sk_memcg = memcg;
6285 void mem_cgroup_sk_free(struct sock *sk)
6288 css_put(&sk->sk_memcg->css);
6292 * mem_cgroup_charge_skmem - charge socket memory
6293 * @memcg: memcg to charge
6294 * @nr_pages: number of pages to charge
6296 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
6297 * @memcg's configured limit, %false if the charge had to be forced.
6299 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6301 gfp_t gfp_mask = GFP_KERNEL;
6303 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6304 struct page_counter *fail;
6306 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
6307 memcg->tcpmem_pressure = 0;
6310 page_counter_charge(&memcg->tcpmem, nr_pages);
6311 memcg->tcpmem_pressure = 1;
6315 /* Don't block in the packet receive path */
6317 gfp_mask = GFP_NOWAIT;
6319 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
6321 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
6324 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
6329 * mem_cgroup_uncharge_skmem - uncharge socket memory
6330 * @memcg: memcg to uncharge
6331 * @nr_pages: number of pages to uncharge
6333 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6335 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6336 page_counter_uncharge(&memcg->tcpmem, nr_pages);
6340 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
6342 refill_stock(memcg, nr_pages);
6345 static int __init cgroup_memory(char *s)
6349 while ((token = strsep(&s, ",")) != NULL) {
6352 if (!strcmp(token, "nosocket"))
6353 cgroup_memory_nosocket = true;
6354 if (!strcmp(token, "nokmem"))
6355 cgroup_memory_nokmem = true;
6359 __setup("cgroup.memory=", cgroup_memory);
6362 * subsys_initcall() for memory controller.
6364 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
6365 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
6366 * basically everything that doesn't depend on a specific mem_cgroup structure
6367 * should be initialized from here.
6369 static int __init mem_cgroup_init(void)
6373 #ifdef CONFIG_MEMCG_KMEM
6375 * Kmem cache creation is mostly done with the slab_mutex held,
6376 * so use a workqueue with limited concurrency to avoid stalling
6377 * all worker threads in case lots of cgroups are created and
6378 * destroyed simultaneously.
6380 memcg_kmem_cache_wq = alloc_workqueue("memcg_kmem_cache", 0, 1);
6381 BUG_ON(!memcg_kmem_cache_wq);
6384 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
6385 memcg_hotplug_cpu_dead);
6387 for_each_possible_cpu(cpu)
6388 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
6391 for_each_node(node) {
6392 struct mem_cgroup_tree_per_node *rtpn;
6394 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
6395 node_online(node) ? node : NUMA_NO_NODE);
6397 rtpn->rb_root = RB_ROOT;
6398 rtpn->rb_rightmost = NULL;
6399 spin_lock_init(&rtpn->lock);
6400 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6405 subsys_initcall(mem_cgroup_init);
6407 #ifdef CONFIG_MEMCG_SWAP
6408 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
6410 while (!refcount_inc_not_zero(&memcg->id.ref)) {
6412 * The root cgroup cannot be destroyed, so it's refcount must
6415 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
6419 memcg = parent_mem_cgroup(memcg);
6421 memcg = root_mem_cgroup;
6427 * mem_cgroup_swapout - transfer a memsw charge to swap
6428 * @page: page whose memsw charge to transfer
6429 * @entry: swap entry to move the charge to
6431 * Transfer the memsw charge of @page to @entry.
6433 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
6435 struct mem_cgroup *memcg, *swap_memcg;
6436 unsigned int nr_entries;
6437 unsigned short oldid;
6439 VM_BUG_ON_PAGE(PageLRU(page), page);
6440 VM_BUG_ON_PAGE(page_count(page), page);
6442 if (!do_memsw_account())
6445 memcg = page->mem_cgroup;
6447 /* Readahead page, never charged */
6452 * In case the memcg owning these pages has been offlined and doesn't
6453 * have an ID allocated to it anymore, charge the closest online
6454 * ancestor for the swap instead and transfer the memory+swap charge.
6456 swap_memcg = mem_cgroup_id_get_online(memcg);
6457 nr_entries = hpage_nr_pages(page);
6458 /* Get references for the tail pages, too */
6460 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
6461 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
6463 VM_BUG_ON_PAGE(oldid, page);
6464 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
6466 page->mem_cgroup = NULL;
6468 if (!mem_cgroup_is_root(memcg))
6469 page_counter_uncharge(&memcg->memory, nr_entries);
6471 if (memcg != swap_memcg) {
6472 if (!mem_cgroup_is_root(swap_memcg))
6473 page_counter_charge(&swap_memcg->memsw, nr_entries);
6474 page_counter_uncharge(&memcg->memsw, nr_entries);
6478 * Interrupts should be disabled here because the caller holds the
6479 * i_pages lock which is taken with interrupts-off. It is
6480 * important here to have the interrupts disabled because it is the
6481 * only synchronisation we have for updating the per-CPU variables.
6483 VM_BUG_ON(!irqs_disabled());
6484 mem_cgroup_charge_statistics(memcg, page, PageTransHuge(page),
6486 memcg_check_events(memcg, page);
6488 if (!mem_cgroup_is_root(memcg))
6489 css_put_many(&memcg->css, nr_entries);
6493 * mem_cgroup_try_charge_swap - try charging swap space for a page
6494 * @page: page being added to swap
6495 * @entry: swap entry to charge
6497 * Try to charge @page's memcg for the swap space at @entry.
6499 * Returns 0 on success, -ENOMEM on failure.
6501 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
6503 unsigned int nr_pages = hpage_nr_pages(page);
6504 struct page_counter *counter;
6505 struct mem_cgroup *memcg;
6506 unsigned short oldid;
6508 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) || !do_swap_account)
6511 memcg = page->mem_cgroup;
6513 /* Readahead page, never charged */
6518 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
6522 memcg = mem_cgroup_id_get_online(memcg);
6524 if (!mem_cgroup_is_root(memcg) &&
6525 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
6526 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
6527 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
6528 mem_cgroup_id_put(memcg);
6532 /* Get references for the tail pages, too */
6534 mem_cgroup_id_get_many(memcg, nr_pages - 1);
6535 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
6536 VM_BUG_ON_PAGE(oldid, page);
6537 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
6543 * mem_cgroup_uncharge_swap - uncharge swap space
6544 * @entry: swap entry to uncharge
6545 * @nr_pages: the amount of swap space to uncharge
6547 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
6549 struct mem_cgroup *memcg;
6552 if (!do_swap_account)
6555 id = swap_cgroup_record(entry, 0, nr_pages);
6557 memcg = mem_cgroup_from_id(id);
6559 if (!mem_cgroup_is_root(memcg)) {
6560 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
6561 page_counter_uncharge(&memcg->swap, nr_pages);
6563 page_counter_uncharge(&memcg->memsw, nr_pages);
6565 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
6566 mem_cgroup_id_put_many(memcg, nr_pages);
6571 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
6573 long nr_swap_pages = get_nr_swap_pages();
6575 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
6576 return nr_swap_pages;
6577 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
6578 nr_swap_pages = min_t(long, nr_swap_pages,
6579 READ_ONCE(memcg->swap.max) -
6580 page_counter_read(&memcg->swap));
6581 return nr_swap_pages;
6584 bool mem_cgroup_swap_full(struct page *page)
6586 struct mem_cgroup *memcg;
6588 VM_BUG_ON_PAGE(!PageLocked(page), page);
6592 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
6595 memcg = page->mem_cgroup;
6599 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
6600 if (page_counter_read(&memcg->swap) * 2 >= memcg->swap.max)
6606 /* for remember boot option*/
6607 #ifdef CONFIG_MEMCG_SWAP_ENABLED
6608 static int really_do_swap_account __initdata = 1;
6610 static int really_do_swap_account __initdata;
6613 static int __init enable_swap_account(char *s)
6615 if (!strcmp(s, "1"))
6616 really_do_swap_account = 1;
6617 else if (!strcmp(s, "0"))
6618 really_do_swap_account = 0;
6621 __setup("swapaccount=", enable_swap_account);
6623 static u64 swap_current_read(struct cgroup_subsys_state *css,
6626 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6628 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
6631 static int swap_max_show(struct seq_file *m, void *v)
6633 return seq_puts_memcg_tunable(m,
6634 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
6637 static ssize_t swap_max_write(struct kernfs_open_file *of,
6638 char *buf, size_t nbytes, loff_t off)
6640 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6644 buf = strstrip(buf);
6645 err = page_counter_memparse(buf, "max", &max);
6649 xchg(&memcg->swap.max, max);
6654 static int swap_events_show(struct seq_file *m, void *v)
6656 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6658 seq_printf(m, "max %lu\n",
6659 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
6660 seq_printf(m, "fail %lu\n",
6661 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
6666 static struct cftype swap_files[] = {
6668 .name = "swap.current",
6669 .flags = CFTYPE_NOT_ON_ROOT,
6670 .read_u64 = swap_current_read,
6674 .flags = CFTYPE_NOT_ON_ROOT,
6675 .seq_show = swap_max_show,
6676 .write = swap_max_write,
6679 .name = "swap.events",
6680 .flags = CFTYPE_NOT_ON_ROOT,
6681 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
6682 .seq_show = swap_events_show,
6687 static struct cftype memsw_cgroup_files[] = {
6689 .name = "memsw.usage_in_bytes",
6690 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6691 .read_u64 = mem_cgroup_read_u64,
6694 .name = "memsw.max_usage_in_bytes",
6695 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6696 .write = mem_cgroup_reset,
6697 .read_u64 = mem_cgroup_read_u64,
6700 .name = "memsw.limit_in_bytes",
6701 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6702 .write = mem_cgroup_write,
6703 .read_u64 = mem_cgroup_read_u64,
6706 .name = "memsw.failcnt",
6707 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6708 .write = mem_cgroup_reset,
6709 .read_u64 = mem_cgroup_read_u64,
6711 { }, /* terminate */
6714 static int __init mem_cgroup_swap_init(void)
6716 if (!mem_cgroup_disabled() && really_do_swap_account) {
6717 do_swap_account = 1;
6718 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys,
6720 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
6721 memsw_cgroup_files));
6725 subsys_initcall(mem_cgroup_swap_init);
6727 #endif /* CONFIG_MEMCG_SWAP */