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 bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
729 enum mem_cgroup_events_target target)
731 unsigned long val, next;
733 val = __this_cpu_read(memcg->stat_cpu->nr_page_events);
734 next = __this_cpu_read(memcg->stat_cpu->targets[target]);
735 /* from time_after() in jiffies.h */
736 if ((long)(next - val) < 0) {
738 case MEM_CGROUP_TARGET_THRESH:
739 next = val + THRESHOLDS_EVENTS_TARGET;
741 case MEM_CGROUP_TARGET_SOFTLIMIT:
742 next = val + SOFTLIMIT_EVENTS_TARGET;
744 case MEM_CGROUP_TARGET_NUMAINFO:
745 next = val + NUMAINFO_EVENTS_TARGET;
750 __this_cpu_write(memcg->stat_cpu->targets[target], next);
757 * Check events in order.
760 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
762 /* threshold event is triggered in finer grain than soft limit */
763 if (unlikely(mem_cgroup_event_ratelimit(memcg,
764 MEM_CGROUP_TARGET_THRESH))) {
766 bool do_numainfo __maybe_unused;
768 do_softlimit = mem_cgroup_event_ratelimit(memcg,
769 MEM_CGROUP_TARGET_SOFTLIMIT);
771 do_numainfo = mem_cgroup_event_ratelimit(memcg,
772 MEM_CGROUP_TARGET_NUMAINFO);
774 mem_cgroup_threshold(memcg);
775 if (unlikely(do_softlimit))
776 mem_cgroup_update_tree(memcg, page);
778 if (unlikely(do_numainfo))
779 atomic_inc(&memcg->numainfo_events);
784 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
787 * mm_update_next_owner() may clear mm->owner to NULL
788 * if it races with swapoff, page migration, etc.
789 * So this can be called with p == NULL.
794 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
796 EXPORT_SYMBOL(mem_cgroup_from_task);
799 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
800 * @mm: mm from which memcg should be extracted. It can be NULL.
802 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
803 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
806 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
808 struct mem_cgroup *memcg;
810 if (mem_cgroup_disabled())
816 * Page cache insertions can happen withou an
817 * actual mm context, e.g. during disk probing
818 * on boot, loopback IO, acct() writes etc.
821 memcg = root_mem_cgroup;
823 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
824 if (unlikely(!memcg))
825 memcg = root_mem_cgroup;
827 } while (!css_tryget_online(&memcg->css));
831 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
834 * get_mem_cgroup_from_page: Obtain a reference on given page's memcg.
835 * @page: page from which memcg should be extracted.
837 * Obtain a reference on page->memcg and returns it if successful. Otherwise
838 * root_mem_cgroup is returned.
840 struct mem_cgroup *get_mem_cgroup_from_page(struct page *page)
842 struct mem_cgroup *memcg = page->mem_cgroup;
844 if (mem_cgroup_disabled())
848 if (!memcg || !css_tryget_online(&memcg->css))
849 memcg = root_mem_cgroup;
853 EXPORT_SYMBOL(get_mem_cgroup_from_page);
856 * If current->active_memcg is non-NULL, do not fallback to current->mm->memcg.
858 static __always_inline struct mem_cgroup *get_mem_cgroup_from_current(void)
860 if (unlikely(current->active_memcg)) {
861 struct mem_cgroup *memcg = root_mem_cgroup;
864 if (css_tryget_online(¤t->active_memcg->css))
865 memcg = current->active_memcg;
869 return get_mem_cgroup_from_mm(current->mm);
873 * mem_cgroup_iter - iterate over memory cgroup hierarchy
874 * @root: hierarchy root
875 * @prev: previously returned memcg, NULL on first invocation
876 * @reclaim: cookie for shared reclaim walks, NULL for full walks
878 * Returns references to children of the hierarchy below @root, or
879 * @root itself, or %NULL after a full round-trip.
881 * Caller must pass the return value in @prev on subsequent
882 * invocations for reference counting, or use mem_cgroup_iter_break()
883 * to cancel a hierarchy walk before the round-trip is complete.
885 * Reclaimers can specify a node and a priority level in @reclaim to
886 * divide up the memcgs in the hierarchy among all concurrent
887 * reclaimers operating on the same node and priority.
889 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
890 struct mem_cgroup *prev,
891 struct mem_cgroup_reclaim_cookie *reclaim)
893 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
894 struct cgroup_subsys_state *css = NULL;
895 struct mem_cgroup *memcg = NULL;
896 struct mem_cgroup *pos = NULL;
898 if (mem_cgroup_disabled())
902 root = root_mem_cgroup;
904 if (prev && !reclaim)
907 if (!root->use_hierarchy && root != root_mem_cgroup) {
916 struct mem_cgroup_per_node *mz;
918 mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
919 iter = &mz->iter[reclaim->priority];
921 if (prev && reclaim->generation != iter->generation)
925 pos = READ_ONCE(iter->position);
926 if (!pos || css_tryget(&pos->css))
929 * css reference reached zero, so iter->position will
930 * be cleared by ->css_released. However, we should not
931 * rely on this happening soon, because ->css_released
932 * is called from a work queue, and by busy-waiting we
933 * might block it. So we clear iter->position right
936 (void)cmpxchg(&iter->position, pos, NULL);
944 css = css_next_descendant_pre(css, &root->css);
947 * Reclaimers share the hierarchy walk, and a
948 * new one might jump in right at the end of
949 * the hierarchy - make sure they see at least
950 * one group and restart from the beginning.
958 * Verify the css and acquire a reference. The root
959 * is provided by the caller, so we know it's alive
960 * and kicking, and don't take an extra reference.
962 memcg = mem_cgroup_from_css(css);
964 if (css == &root->css)
975 * The position could have already been updated by a competing
976 * thread, so check that the value hasn't changed since we read
977 * it to avoid reclaiming from the same cgroup twice.
979 (void)cmpxchg(&iter->position, pos, memcg);
987 reclaim->generation = iter->generation;
993 if (prev && prev != root)
1000 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1001 * @root: hierarchy root
1002 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1004 void mem_cgroup_iter_break(struct mem_cgroup *root,
1005 struct mem_cgroup *prev)
1008 root = root_mem_cgroup;
1009 if (prev && prev != root)
1010 css_put(&prev->css);
1013 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1015 struct mem_cgroup *memcg = dead_memcg;
1016 struct mem_cgroup_reclaim_iter *iter;
1017 struct mem_cgroup_per_node *mz;
1021 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
1022 for_each_node(nid) {
1023 mz = mem_cgroup_nodeinfo(memcg, nid);
1024 for (i = 0; i <= DEF_PRIORITY; i++) {
1025 iter = &mz->iter[i];
1026 cmpxchg(&iter->position,
1034 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1035 * @memcg: hierarchy root
1036 * @fn: function to call for each task
1037 * @arg: argument passed to @fn
1039 * This function iterates over tasks attached to @memcg or to any of its
1040 * descendants and calls @fn for each task. If @fn returns a non-zero
1041 * value, the function breaks the iteration loop and returns the value.
1042 * Otherwise, it will iterate over all tasks and return 0.
1044 * This function must not be called for the root memory cgroup.
1046 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1047 int (*fn)(struct task_struct *, void *), void *arg)
1049 struct mem_cgroup *iter;
1052 BUG_ON(memcg == root_mem_cgroup);
1054 for_each_mem_cgroup_tree(iter, memcg) {
1055 struct css_task_iter it;
1056 struct task_struct *task;
1058 css_task_iter_start(&iter->css, 0, &it);
1059 while (!ret && (task = css_task_iter_next(&it)))
1060 ret = fn(task, arg);
1061 css_task_iter_end(&it);
1063 mem_cgroup_iter_break(memcg, iter);
1071 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1073 * @pgdat: pgdat of the page
1075 * This function is only safe when following the LRU page isolation
1076 * and putback protocol: the LRU lock must be held, and the page must
1077 * either be PageLRU() or the caller must have isolated/allocated it.
1079 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct pglist_data *pgdat)
1081 struct mem_cgroup_per_node *mz;
1082 struct mem_cgroup *memcg;
1083 struct lruvec *lruvec;
1085 if (mem_cgroup_disabled()) {
1086 lruvec = &pgdat->lruvec;
1090 memcg = page->mem_cgroup;
1092 * Swapcache readahead pages are added to the LRU - and
1093 * possibly migrated - before they are charged.
1096 memcg = root_mem_cgroup;
1098 mz = mem_cgroup_page_nodeinfo(memcg, page);
1099 lruvec = &mz->lruvec;
1102 * Since a node can be onlined after the mem_cgroup was created,
1103 * we have to be prepared to initialize lruvec->zone here;
1104 * and if offlined then reonlined, we need to reinitialize it.
1106 if (unlikely(lruvec->pgdat != pgdat))
1107 lruvec->pgdat = pgdat;
1112 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1113 * @lruvec: mem_cgroup per zone lru vector
1114 * @lru: index of lru list the page is sitting on
1115 * @zid: zone id of the accounted pages
1116 * @nr_pages: positive when adding or negative when removing
1118 * This function must be called under lru_lock, just before a page is added
1119 * to or just after a page is removed from an lru list (that ordering being
1120 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1122 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1123 int zid, int nr_pages)
1125 struct mem_cgroup_per_node *mz;
1126 unsigned long *lru_size;
1129 if (mem_cgroup_disabled())
1132 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1133 lru_size = &mz->lru_zone_size[zid][lru];
1136 *lru_size += nr_pages;
1139 if (WARN_ONCE(size < 0,
1140 "%s(%p, %d, %d): lru_size %ld\n",
1141 __func__, lruvec, lru, nr_pages, size)) {
1147 *lru_size += nr_pages;
1150 bool task_in_mem_cgroup(struct task_struct *task, struct mem_cgroup *memcg)
1152 struct mem_cgroup *task_memcg;
1153 struct task_struct *p;
1156 p = find_lock_task_mm(task);
1158 task_memcg = get_mem_cgroup_from_mm(p->mm);
1162 * All threads may have already detached their mm's, but the oom
1163 * killer still needs to detect if they have already been oom
1164 * killed to prevent needlessly killing additional tasks.
1167 task_memcg = mem_cgroup_from_task(task);
1168 css_get(&task_memcg->css);
1171 ret = mem_cgroup_is_descendant(task_memcg, memcg);
1172 css_put(&task_memcg->css);
1177 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1178 * @memcg: the memory cgroup
1180 * Returns the maximum amount of memory @mem can be charged with, in
1183 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1185 unsigned long margin = 0;
1186 unsigned long count;
1187 unsigned long limit;
1189 count = page_counter_read(&memcg->memory);
1190 limit = READ_ONCE(memcg->memory.max);
1192 margin = limit - count;
1194 if (do_memsw_account()) {
1195 count = page_counter_read(&memcg->memsw);
1196 limit = READ_ONCE(memcg->memsw.max);
1198 margin = min(margin, limit - count);
1207 * A routine for checking "mem" is under move_account() or not.
1209 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1210 * moving cgroups. This is for waiting at high-memory pressure
1213 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1215 struct mem_cgroup *from;
1216 struct mem_cgroup *to;
1219 * Unlike task_move routines, we access mc.to, mc.from not under
1220 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1222 spin_lock(&mc.lock);
1228 ret = mem_cgroup_is_descendant(from, memcg) ||
1229 mem_cgroup_is_descendant(to, memcg);
1231 spin_unlock(&mc.lock);
1235 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1237 if (mc.moving_task && current != mc.moving_task) {
1238 if (mem_cgroup_under_move(memcg)) {
1240 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1241 /* moving charge context might have finished. */
1244 finish_wait(&mc.waitq, &wait);
1251 static const unsigned int memcg1_stats[] = {
1262 static const char *const memcg1_stat_names[] = {
1273 #define K(x) ((x) << (PAGE_SHIFT-10))
1275 * mem_cgroup_print_oom_context: Print OOM information relevant to
1276 * memory controller.
1277 * @memcg: The memory cgroup that went over limit
1278 * @p: Task that is going to be killed
1280 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1283 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1288 pr_cont(",oom_memcg=");
1289 pr_cont_cgroup_path(memcg->css.cgroup);
1291 pr_cont(",global_oom");
1293 pr_cont(",task_memcg=");
1294 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1300 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1301 * memory controller.
1302 * @memcg: The memory cgroup that went over limit
1304 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1306 struct mem_cgroup *iter;
1309 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1310 K((u64)page_counter_read(&memcg->memory)),
1311 K((u64)memcg->memory.max), memcg->memory.failcnt);
1312 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1313 K((u64)page_counter_read(&memcg->memsw)),
1314 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1315 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1316 K((u64)page_counter_read(&memcg->kmem)),
1317 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1319 for_each_mem_cgroup_tree(iter, memcg) {
1320 pr_info("Memory cgroup stats for ");
1321 pr_cont_cgroup_path(iter->css.cgroup);
1324 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
1325 if (memcg1_stats[i] == MEMCG_SWAP && !do_swap_account)
1327 pr_cont(" %s:%luKB", memcg1_stat_names[i],
1328 K(memcg_page_state(iter, memcg1_stats[i])));
1331 for (i = 0; i < NR_LRU_LISTS; i++)
1332 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1333 K(memcg_page_state(iter, NR_LRU_BASE + i)));
1340 * Return the memory (and swap, if configured) limit for a memcg.
1342 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1346 max = memcg->memory.max;
1347 if (mem_cgroup_swappiness(memcg)) {
1348 unsigned long memsw_max;
1349 unsigned long swap_max;
1351 memsw_max = memcg->memsw.max;
1352 swap_max = memcg->swap.max;
1353 swap_max = min(swap_max, (unsigned long)total_swap_pages);
1354 max = min(max + swap_max, memsw_max);
1359 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1362 struct oom_control oc = {
1366 .gfp_mask = gfp_mask,
1371 if (mutex_lock_killable(&oom_lock))
1374 * A few threads which were not waiting at mutex_lock_killable() can
1375 * fail to bail out. Therefore, check again after holding oom_lock.
1377 ret = should_force_charge() || out_of_memory(&oc);
1378 mutex_unlock(&oom_lock);
1382 #if MAX_NUMNODES > 1
1385 * test_mem_cgroup_node_reclaimable
1386 * @memcg: the target memcg
1387 * @nid: the node ID to be checked.
1388 * @noswap : specify true here if the user wants flle only information.
1390 * This function returns whether the specified memcg contains any
1391 * reclaimable pages on a node. Returns true if there are any reclaimable
1392 * pages in the node.
1394 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1395 int nid, bool noswap)
1397 struct lruvec *lruvec = mem_cgroup_lruvec(NODE_DATA(nid), memcg);
1399 if (lruvec_page_state(lruvec, NR_INACTIVE_FILE) ||
1400 lruvec_page_state(lruvec, NR_ACTIVE_FILE))
1402 if (noswap || !total_swap_pages)
1404 if (lruvec_page_state(lruvec, NR_INACTIVE_ANON) ||
1405 lruvec_page_state(lruvec, NR_ACTIVE_ANON))
1412 * Always updating the nodemask is not very good - even if we have an empty
1413 * list or the wrong list here, we can start from some node and traverse all
1414 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1417 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1421 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1422 * pagein/pageout changes since the last update.
1424 if (!atomic_read(&memcg->numainfo_events))
1426 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1429 /* make a nodemask where this memcg uses memory from */
1430 memcg->scan_nodes = node_states[N_MEMORY];
1432 for_each_node_mask(nid, node_states[N_MEMORY]) {
1434 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1435 node_clear(nid, memcg->scan_nodes);
1438 atomic_set(&memcg->numainfo_events, 0);
1439 atomic_set(&memcg->numainfo_updating, 0);
1443 * Selecting a node where we start reclaim from. Because what we need is just
1444 * reducing usage counter, start from anywhere is O,K. Considering
1445 * memory reclaim from current node, there are pros. and cons.
1447 * Freeing memory from current node means freeing memory from a node which
1448 * we'll use or we've used. So, it may make LRU bad. And if several threads
1449 * hit limits, it will see a contention on a node. But freeing from remote
1450 * node means more costs for memory reclaim because of memory latency.
1452 * Now, we use round-robin. Better algorithm is welcomed.
1454 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1458 mem_cgroup_may_update_nodemask(memcg);
1459 node = memcg->last_scanned_node;
1461 node = next_node_in(node, memcg->scan_nodes);
1463 * mem_cgroup_may_update_nodemask might have seen no reclaimmable pages
1464 * last time it really checked all the LRUs due to rate limiting.
1465 * Fallback to the current node in that case for simplicity.
1467 if (unlikely(node == MAX_NUMNODES))
1468 node = numa_node_id();
1470 memcg->last_scanned_node = node;
1474 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1480 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1483 unsigned long *total_scanned)
1485 struct mem_cgroup *victim = NULL;
1488 unsigned long excess;
1489 unsigned long nr_scanned;
1490 struct mem_cgroup_reclaim_cookie reclaim = {
1495 excess = soft_limit_excess(root_memcg);
1498 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1503 * If we have not been able to reclaim
1504 * anything, it might because there are
1505 * no reclaimable pages under this hierarchy
1510 * We want to do more targeted reclaim.
1511 * excess >> 2 is not to excessive so as to
1512 * reclaim too much, nor too less that we keep
1513 * coming back to reclaim from this cgroup
1515 if (total >= (excess >> 2) ||
1516 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1521 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1522 pgdat, &nr_scanned);
1523 *total_scanned += nr_scanned;
1524 if (!soft_limit_excess(root_memcg))
1527 mem_cgroup_iter_break(root_memcg, victim);
1531 #ifdef CONFIG_LOCKDEP
1532 static struct lockdep_map memcg_oom_lock_dep_map = {
1533 .name = "memcg_oom_lock",
1537 static DEFINE_SPINLOCK(memcg_oom_lock);
1540 * Check OOM-Killer is already running under our hierarchy.
1541 * If someone is running, return false.
1543 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1545 struct mem_cgroup *iter, *failed = NULL;
1547 spin_lock(&memcg_oom_lock);
1549 for_each_mem_cgroup_tree(iter, memcg) {
1550 if (iter->oom_lock) {
1552 * this subtree of our hierarchy is already locked
1553 * so we cannot give a lock.
1556 mem_cgroup_iter_break(memcg, iter);
1559 iter->oom_lock = true;
1564 * OK, we failed to lock the whole subtree so we have
1565 * to clean up what we set up to the failing subtree
1567 for_each_mem_cgroup_tree(iter, memcg) {
1568 if (iter == failed) {
1569 mem_cgroup_iter_break(memcg, iter);
1572 iter->oom_lock = false;
1575 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1577 spin_unlock(&memcg_oom_lock);
1582 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1584 struct mem_cgroup *iter;
1586 spin_lock(&memcg_oom_lock);
1587 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1588 for_each_mem_cgroup_tree(iter, memcg)
1589 iter->oom_lock = false;
1590 spin_unlock(&memcg_oom_lock);
1593 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1595 struct mem_cgroup *iter;
1597 spin_lock(&memcg_oom_lock);
1598 for_each_mem_cgroup_tree(iter, memcg)
1600 spin_unlock(&memcg_oom_lock);
1603 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1605 struct mem_cgroup *iter;
1608 * When a new child is created while the hierarchy is under oom,
1609 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1611 spin_lock(&memcg_oom_lock);
1612 for_each_mem_cgroup_tree(iter, memcg)
1613 if (iter->under_oom > 0)
1615 spin_unlock(&memcg_oom_lock);
1618 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1620 struct oom_wait_info {
1621 struct mem_cgroup *memcg;
1622 wait_queue_entry_t wait;
1625 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1626 unsigned mode, int sync, void *arg)
1628 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1629 struct mem_cgroup *oom_wait_memcg;
1630 struct oom_wait_info *oom_wait_info;
1632 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1633 oom_wait_memcg = oom_wait_info->memcg;
1635 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1636 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1638 return autoremove_wake_function(wait, mode, sync, arg);
1641 static void memcg_oom_recover(struct mem_cgroup *memcg)
1644 * For the following lockless ->under_oom test, the only required
1645 * guarantee is that it must see the state asserted by an OOM when
1646 * this function is called as a result of userland actions
1647 * triggered by the notification of the OOM. This is trivially
1648 * achieved by invoking mem_cgroup_mark_under_oom() before
1649 * triggering notification.
1651 if (memcg && memcg->under_oom)
1652 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1662 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1664 enum oom_status ret;
1667 if (order > PAGE_ALLOC_COSTLY_ORDER)
1670 memcg_memory_event(memcg, MEMCG_OOM);
1673 * We are in the middle of the charge context here, so we
1674 * don't want to block when potentially sitting on a callstack
1675 * that holds all kinds of filesystem and mm locks.
1677 * cgroup1 allows disabling the OOM killer and waiting for outside
1678 * handling until the charge can succeed; remember the context and put
1679 * the task to sleep at the end of the page fault when all locks are
1682 * On the other hand, in-kernel OOM killer allows for an async victim
1683 * memory reclaim (oom_reaper) and that means that we are not solely
1684 * relying on the oom victim to make a forward progress and we can
1685 * invoke the oom killer here.
1687 * Please note that mem_cgroup_out_of_memory might fail to find a
1688 * victim and then we have to bail out from the charge path.
1690 if (memcg->oom_kill_disable) {
1691 if (!current->in_user_fault)
1693 css_get(&memcg->css);
1694 current->memcg_in_oom = memcg;
1695 current->memcg_oom_gfp_mask = mask;
1696 current->memcg_oom_order = order;
1701 mem_cgroup_mark_under_oom(memcg);
1703 locked = mem_cgroup_oom_trylock(memcg);
1706 mem_cgroup_oom_notify(memcg);
1708 mem_cgroup_unmark_under_oom(memcg);
1709 if (mem_cgroup_out_of_memory(memcg, mask, order))
1715 mem_cgroup_oom_unlock(memcg);
1721 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1722 * @handle: actually kill/wait or just clean up the OOM state
1724 * This has to be called at the end of a page fault if the memcg OOM
1725 * handler was enabled.
1727 * Memcg supports userspace OOM handling where failed allocations must
1728 * sleep on a waitqueue until the userspace task resolves the
1729 * situation. Sleeping directly in the charge context with all kinds
1730 * of locks held is not a good idea, instead we remember an OOM state
1731 * in the task and mem_cgroup_oom_synchronize() has to be called at
1732 * the end of the page fault to complete the OOM handling.
1734 * Returns %true if an ongoing memcg OOM situation was detected and
1735 * completed, %false otherwise.
1737 bool mem_cgroup_oom_synchronize(bool handle)
1739 struct mem_cgroup *memcg = current->memcg_in_oom;
1740 struct oom_wait_info owait;
1743 /* OOM is global, do not handle */
1750 owait.memcg = memcg;
1751 owait.wait.flags = 0;
1752 owait.wait.func = memcg_oom_wake_function;
1753 owait.wait.private = current;
1754 INIT_LIST_HEAD(&owait.wait.entry);
1756 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1757 mem_cgroup_mark_under_oom(memcg);
1759 locked = mem_cgroup_oom_trylock(memcg);
1762 mem_cgroup_oom_notify(memcg);
1764 if (locked && !memcg->oom_kill_disable) {
1765 mem_cgroup_unmark_under_oom(memcg);
1766 finish_wait(&memcg_oom_waitq, &owait.wait);
1767 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1768 current->memcg_oom_order);
1771 mem_cgroup_unmark_under_oom(memcg);
1772 finish_wait(&memcg_oom_waitq, &owait.wait);
1776 mem_cgroup_oom_unlock(memcg);
1778 * There is no guarantee that an OOM-lock contender
1779 * sees the wakeups triggered by the OOM kill
1780 * uncharges. Wake any sleepers explicitely.
1782 memcg_oom_recover(memcg);
1785 current->memcg_in_oom = NULL;
1786 css_put(&memcg->css);
1791 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
1792 * @victim: task to be killed by the OOM killer
1793 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
1795 * Returns a pointer to a memory cgroup, which has to be cleaned up
1796 * by killing all belonging OOM-killable tasks.
1798 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
1800 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
1801 struct mem_cgroup *oom_domain)
1803 struct mem_cgroup *oom_group = NULL;
1804 struct mem_cgroup *memcg;
1806 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
1810 oom_domain = root_mem_cgroup;
1814 memcg = mem_cgroup_from_task(victim);
1815 if (memcg == root_mem_cgroup)
1819 * Traverse the memory cgroup hierarchy from the victim task's
1820 * cgroup up to the OOMing cgroup (or root) to find the
1821 * highest-level memory cgroup with oom.group set.
1823 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
1824 if (memcg->oom_group)
1827 if (memcg == oom_domain)
1832 css_get(&oom_group->css);
1839 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
1841 pr_info("Tasks in ");
1842 pr_cont_cgroup_path(memcg->css.cgroup);
1843 pr_cont(" are going to be killed due to memory.oom.group set\n");
1847 * lock_page_memcg - lock a page->mem_cgroup binding
1850 * This function protects unlocked LRU pages from being moved to
1853 * It ensures lifetime of the returned memcg. Caller is responsible
1854 * for the lifetime of the page; __unlock_page_memcg() is available
1855 * when @page might get freed inside the locked section.
1857 struct mem_cgroup *lock_page_memcg(struct page *page)
1859 struct mem_cgroup *memcg;
1860 unsigned long flags;
1863 * The RCU lock is held throughout the transaction. The fast
1864 * path can get away without acquiring the memcg->move_lock
1865 * because page moving starts with an RCU grace period.
1867 * The RCU lock also protects the memcg from being freed when
1868 * the page state that is going to change is the only thing
1869 * preventing the page itself from being freed. E.g. writeback
1870 * doesn't hold a page reference and relies on PG_writeback to
1871 * keep off truncation, migration and so forth.
1875 if (mem_cgroup_disabled())
1878 memcg = page->mem_cgroup;
1879 if (unlikely(!memcg))
1882 if (atomic_read(&memcg->moving_account) <= 0)
1885 spin_lock_irqsave(&memcg->move_lock, flags);
1886 if (memcg != page->mem_cgroup) {
1887 spin_unlock_irqrestore(&memcg->move_lock, flags);
1892 * When charge migration first begins, we can have locked and
1893 * unlocked page stat updates happening concurrently. Track
1894 * the task who has the lock for unlock_page_memcg().
1896 memcg->move_lock_task = current;
1897 memcg->move_lock_flags = flags;
1901 EXPORT_SYMBOL(lock_page_memcg);
1904 * __unlock_page_memcg - unlock and unpin a memcg
1907 * Unlock and unpin a memcg returned by lock_page_memcg().
1909 void __unlock_page_memcg(struct mem_cgroup *memcg)
1911 if (memcg && memcg->move_lock_task == current) {
1912 unsigned long flags = memcg->move_lock_flags;
1914 memcg->move_lock_task = NULL;
1915 memcg->move_lock_flags = 0;
1917 spin_unlock_irqrestore(&memcg->move_lock, flags);
1924 * unlock_page_memcg - unlock a page->mem_cgroup binding
1927 void unlock_page_memcg(struct page *page)
1929 __unlock_page_memcg(page->mem_cgroup);
1931 EXPORT_SYMBOL(unlock_page_memcg);
1933 struct memcg_stock_pcp {
1934 struct mem_cgroup *cached; /* this never be root cgroup */
1935 unsigned int nr_pages;
1936 struct work_struct work;
1937 unsigned long flags;
1938 #define FLUSHING_CACHED_CHARGE 0
1940 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1941 static DEFINE_MUTEX(percpu_charge_mutex);
1944 * consume_stock: Try to consume stocked charge on this cpu.
1945 * @memcg: memcg to consume from.
1946 * @nr_pages: how many pages to charge.
1948 * The charges will only happen if @memcg matches the current cpu's memcg
1949 * stock, and at least @nr_pages are available in that stock. Failure to
1950 * service an allocation will refill the stock.
1952 * returns true if successful, false otherwise.
1954 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1956 struct memcg_stock_pcp *stock;
1957 unsigned long flags;
1960 if (nr_pages > MEMCG_CHARGE_BATCH)
1963 local_irq_save(flags);
1965 stock = this_cpu_ptr(&memcg_stock);
1966 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
1967 stock->nr_pages -= nr_pages;
1971 local_irq_restore(flags);
1977 * Returns stocks cached in percpu and reset cached information.
1979 static void drain_stock(struct memcg_stock_pcp *stock)
1981 struct mem_cgroup *old = stock->cached;
1983 if (stock->nr_pages) {
1984 page_counter_uncharge(&old->memory, stock->nr_pages);
1985 if (do_memsw_account())
1986 page_counter_uncharge(&old->memsw, stock->nr_pages);
1987 css_put_many(&old->css, stock->nr_pages);
1988 stock->nr_pages = 0;
1990 stock->cached = NULL;
1993 static void drain_local_stock(struct work_struct *dummy)
1995 struct memcg_stock_pcp *stock;
1996 unsigned long flags;
1999 * The only protection from memory hotplug vs. drain_stock races is
2000 * that we always operate on local CPU stock here with IRQ disabled
2002 local_irq_save(flags);
2004 stock = this_cpu_ptr(&memcg_stock);
2006 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2008 local_irq_restore(flags);
2012 * Cache charges(val) to local per_cpu area.
2013 * This will be consumed by consume_stock() function, later.
2015 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2017 struct memcg_stock_pcp *stock;
2018 unsigned long flags;
2020 local_irq_save(flags);
2022 stock = this_cpu_ptr(&memcg_stock);
2023 if (stock->cached != memcg) { /* reset if necessary */
2025 stock->cached = memcg;
2027 stock->nr_pages += nr_pages;
2029 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2032 local_irq_restore(flags);
2036 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2037 * of the hierarchy under it.
2039 static void drain_all_stock(struct mem_cgroup *root_memcg)
2043 /* If someone's already draining, avoid adding running more workers. */
2044 if (!mutex_trylock(&percpu_charge_mutex))
2047 * Notify other cpus that system-wide "drain" is running
2048 * We do not care about races with the cpu hotplug because cpu down
2049 * as well as workers from this path always operate on the local
2050 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2053 for_each_online_cpu(cpu) {
2054 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2055 struct mem_cgroup *memcg;
2057 memcg = stock->cached;
2058 if (!memcg || !stock->nr_pages || !css_tryget(&memcg->css))
2060 if (!mem_cgroup_is_descendant(memcg, root_memcg)) {
2061 css_put(&memcg->css);
2064 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2066 drain_local_stock(&stock->work);
2068 schedule_work_on(cpu, &stock->work);
2070 css_put(&memcg->css);
2073 mutex_unlock(&percpu_charge_mutex);
2076 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2078 struct memcg_stock_pcp *stock;
2079 struct mem_cgroup *memcg;
2081 stock = &per_cpu(memcg_stock, cpu);
2084 for_each_mem_cgroup(memcg) {
2087 for (i = 0; i < MEMCG_NR_STAT; i++) {
2091 x = this_cpu_xchg(memcg->stat_cpu->count[i], 0);
2093 atomic_long_add(x, &memcg->stat[i]);
2095 if (i >= NR_VM_NODE_STAT_ITEMS)
2098 for_each_node(nid) {
2099 struct mem_cgroup_per_node *pn;
2101 pn = mem_cgroup_nodeinfo(memcg, nid);
2102 x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0);
2104 atomic_long_add(x, &pn->lruvec_stat[i]);
2108 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
2111 x = this_cpu_xchg(memcg->stat_cpu->events[i], 0);
2113 atomic_long_add(x, &memcg->events[i]);
2120 static void reclaim_high(struct mem_cgroup *memcg,
2121 unsigned int nr_pages,
2125 if (page_counter_read(&memcg->memory) <= memcg->high)
2127 memcg_memory_event(memcg, MEMCG_HIGH);
2128 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
2129 } while ((memcg = parent_mem_cgroup(memcg)));
2132 static void high_work_func(struct work_struct *work)
2134 struct mem_cgroup *memcg;
2136 memcg = container_of(work, struct mem_cgroup, high_work);
2137 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2141 * Scheduled by try_charge() to be executed from the userland return path
2142 * and reclaims memory over the high limit.
2144 void mem_cgroup_handle_over_high(void)
2146 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2147 struct mem_cgroup *memcg;
2149 if (likely(!nr_pages))
2152 memcg = get_mem_cgroup_from_mm(current->mm);
2153 reclaim_high(memcg, nr_pages, GFP_KERNEL);
2154 css_put(&memcg->css);
2155 current->memcg_nr_pages_over_high = 0;
2158 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2159 unsigned int nr_pages)
2161 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2162 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2163 struct mem_cgroup *mem_over_limit;
2164 struct page_counter *counter;
2165 unsigned long nr_reclaimed;
2166 bool may_swap = true;
2167 bool drained = false;
2169 enum oom_status oom_status;
2171 if (mem_cgroup_is_root(memcg))
2174 if (consume_stock(memcg, nr_pages))
2177 if (!do_memsw_account() ||
2178 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2179 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2181 if (do_memsw_account())
2182 page_counter_uncharge(&memcg->memsw, batch);
2183 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2185 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2189 if (batch > nr_pages) {
2195 * Unlike in global OOM situations, memcg is not in a physical
2196 * memory shortage. Allow dying and OOM-killed tasks to
2197 * bypass the last charges so that they can exit quickly and
2198 * free their memory.
2200 if (unlikely(should_force_charge()))
2204 * Prevent unbounded recursion when reclaim operations need to
2205 * allocate memory. This might exceed the limits temporarily,
2206 * but we prefer facilitating memory reclaim and getting back
2207 * under the limit over triggering OOM kills in these cases.
2209 if (unlikely(current->flags & PF_MEMALLOC))
2212 if (unlikely(task_in_memcg_oom(current)))
2215 if (!gfpflags_allow_blocking(gfp_mask))
2218 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2220 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2221 gfp_mask, may_swap);
2223 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2227 drain_all_stock(mem_over_limit);
2232 if (gfp_mask & __GFP_NORETRY)
2235 * Even though the limit is exceeded at this point, reclaim
2236 * may have been able to free some pages. Retry the charge
2237 * before killing the task.
2239 * Only for regular pages, though: huge pages are rather
2240 * unlikely to succeed so close to the limit, and we fall back
2241 * to regular pages anyway in case of failure.
2243 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2246 * At task move, charge accounts can be doubly counted. So, it's
2247 * better to wait until the end of task_move if something is going on.
2249 if (mem_cgroup_wait_acct_move(mem_over_limit))
2255 if (gfp_mask & __GFP_RETRY_MAYFAIL && oomed)
2258 if (gfp_mask & __GFP_NOFAIL)
2261 if (fatal_signal_pending(current))
2265 * keep retrying as long as the memcg oom killer is able to make
2266 * a forward progress or bypass the charge if the oom killer
2267 * couldn't make any progress.
2269 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2270 get_order(nr_pages * PAGE_SIZE));
2271 switch (oom_status) {
2273 nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2282 if (!(gfp_mask & __GFP_NOFAIL))
2286 * The allocation either can't fail or will lead to more memory
2287 * being freed very soon. Allow memory usage go over the limit
2288 * temporarily by force charging it.
2290 page_counter_charge(&memcg->memory, nr_pages);
2291 if (do_memsw_account())
2292 page_counter_charge(&memcg->memsw, nr_pages);
2293 css_get_many(&memcg->css, nr_pages);
2298 css_get_many(&memcg->css, batch);
2299 if (batch > nr_pages)
2300 refill_stock(memcg, batch - nr_pages);
2303 * If the hierarchy is above the normal consumption range, schedule
2304 * reclaim on returning to userland. We can perform reclaim here
2305 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2306 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2307 * not recorded as it most likely matches current's and won't
2308 * change in the meantime. As high limit is checked again before
2309 * reclaim, the cost of mismatch is negligible.
2312 if (page_counter_read(&memcg->memory) > memcg->high) {
2313 /* Don't bother a random interrupted task */
2314 if (in_interrupt()) {
2315 schedule_work(&memcg->high_work);
2318 current->memcg_nr_pages_over_high += batch;
2319 set_notify_resume(current);
2322 } while ((memcg = parent_mem_cgroup(memcg)));
2327 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2329 if (mem_cgroup_is_root(memcg))
2332 page_counter_uncharge(&memcg->memory, nr_pages);
2333 if (do_memsw_account())
2334 page_counter_uncharge(&memcg->memsw, nr_pages);
2336 css_put_many(&memcg->css, nr_pages);
2339 static void lock_page_lru(struct page *page, int *isolated)
2341 pg_data_t *pgdat = page_pgdat(page);
2343 spin_lock_irq(&pgdat->lru_lock);
2344 if (PageLRU(page)) {
2345 struct lruvec *lruvec;
2347 lruvec = mem_cgroup_page_lruvec(page, pgdat);
2349 del_page_from_lru_list(page, lruvec, page_lru(page));
2355 static void unlock_page_lru(struct page *page, int isolated)
2357 pg_data_t *pgdat = page_pgdat(page);
2360 struct lruvec *lruvec;
2362 lruvec = mem_cgroup_page_lruvec(page, pgdat);
2363 VM_BUG_ON_PAGE(PageLRU(page), page);
2365 add_page_to_lru_list(page, lruvec, page_lru(page));
2367 spin_unlock_irq(&pgdat->lru_lock);
2370 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2375 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2378 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2379 * may already be on some other mem_cgroup's LRU. Take care of it.
2382 lock_page_lru(page, &isolated);
2385 * Nobody should be changing or seriously looking at
2386 * page->mem_cgroup at this point:
2388 * - the page is uncharged
2390 * - the page is off-LRU
2392 * - an anonymous fault has exclusive page access, except for
2393 * a locked page table
2395 * - a page cache insertion, a swapin fault, or a migration
2396 * have the page locked
2398 page->mem_cgroup = memcg;
2401 unlock_page_lru(page, isolated);
2404 #ifdef CONFIG_MEMCG_KMEM
2405 static int memcg_alloc_cache_id(void)
2410 id = ida_simple_get(&memcg_cache_ida,
2411 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2415 if (id < memcg_nr_cache_ids)
2419 * There's no space for the new id in memcg_caches arrays,
2420 * so we have to grow them.
2422 down_write(&memcg_cache_ids_sem);
2424 size = 2 * (id + 1);
2425 if (size < MEMCG_CACHES_MIN_SIZE)
2426 size = MEMCG_CACHES_MIN_SIZE;
2427 else if (size > MEMCG_CACHES_MAX_SIZE)
2428 size = MEMCG_CACHES_MAX_SIZE;
2430 err = memcg_update_all_caches(size);
2432 err = memcg_update_all_list_lrus(size);
2434 memcg_nr_cache_ids = size;
2436 up_write(&memcg_cache_ids_sem);
2439 ida_simple_remove(&memcg_cache_ida, id);
2445 static void memcg_free_cache_id(int id)
2447 ida_simple_remove(&memcg_cache_ida, id);
2450 struct memcg_kmem_cache_create_work {
2451 struct mem_cgroup *memcg;
2452 struct kmem_cache *cachep;
2453 struct work_struct work;
2456 static void memcg_kmem_cache_create_func(struct work_struct *w)
2458 struct memcg_kmem_cache_create_work *cw =
2459 container_of(w, struct memcg_kmem_cache_create_work, work);
2460 struct mem_cgroup *memcg = cw->memcg;
2461 struct kmem_cache *cachep = cw->cachep;
2463 memcg_create_kmem_cache(memcg, cachep);
2465 css_put(&memcg->css);
2470 * Enqueue the creation of a per-memcg kmem_cache.
2472 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2473 struct kmem_cache *cachep)
2475 struct memcg_kmem_cache_create_work *cw;
2477 cw = kmalloc(sizeof(*cw), GFP_NOWAIT | __GFP_NOWARN);
2481 css_get(&memcg->css);
2484 cw->cachep = cachep;
2485 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2487 queue_work(memcg_kmem_cache_wq, &cw->work);
2490 static inline bool memcg_kmem_bypass(void)
2492 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
2498 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2499 * @cachep: the original global kmem cache
2501 * Return the kmem_cache we're supposed to use for a slab allocation.
2502 * We try to use the current memcg's version of the cache.
2504 * If the cache does not exist yet, if we are the first user of it, we
2505 * create it asynchronously in a workqueue and let the current allocation
2506 * go through with the original cache.
2508 * This function takes a reference to the cache it returns to assure it
2509 * won't get destroyed while we are working with it. Once the caller is
2510 * done with it, memcg_kmem_put_cache() must be called to release the
2513 struct kmem_cache *memcg_kmem_get_cache(struct kmem_cache *cachep)
2515 struct mem_cgroup *memcg;
2516 struct kmem_cache *memcg_cachep;
2519 VM_BUG_ON(!is_root_cache(cachep));
2521 if (memcg_kmem_bypass())
2524 memcg = get_mem_cgroup_from_current();
2525 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2529 memcg_cachep = cache_from_memcg_idx(cachep, kmemcg_id);
2530 if (likely(memcg_cachep))
2531 return memcg_cachep;
2534 * If we are in a safe context (can wait, and not in interrupt
2535 * context), we could be be predictable and return right away.
2536 * This would guarantee that the allocation being performed
2537 * already belongs in the new cache.
2539 * However, there are some clashes that can arrive from locking.
2540 * For instance, because we acquire the slab_mutex while doing
2541 * memcg_create_kmem_cache, this means no further allocation
2542 * could happen with the slab_mutex held. So it's better to
2545 memcg_schedule_kmem_cache_create(memcg, cachep);
2547 css_put(&memcg->css);
2552 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
2553 * @cachep: the cache returned by memcg_kmem_get_cache
2555 void memcg_kmem_put_cache(struct kmem_cache *cachep)
2557 if (!is_root_cache(cachep))
2558 css_put(&cachep->memcg_params.memcg->css);
2562 * __memcg_kmem_charge_memcg: charge a kmem page
2563 * @page: page to charge
2564 * @gfp: reclaim mode
2565 * @order: allocation order
2566 * @memcg: memory cgroup to charge
2568 * Returns 0 on success, an error code on failure.
2570 int __memcg_kmem_charge_memcg(struct page *page, gfp_t gfp, int order,
2571 struct mem_cgroup *memcg)
2573 unsigned int nr_pages = 1 << order;
2574 struct page_counter *counter;
2577 ret = try_charge(memcg, gfp, nr_pages);
2581 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
2582 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
2583 cancel_charge(memcg, nr_pages);
2587 page->mem_cgroup = memcg;
2593 * __memcg_kmem_charge: charge a kmem page to the current memory cgroup
2594 * @page: page to charge
2595 * @gfp: reclaim mode
2596 * @order: allocation order
2598 * Returns 0 on success, an error code on failure.
2600 int __memcg_kmem_charge(struct page *page, gfp_t gfp, int order)
2602 struct mem_cgroup *memcg;
2605 if (memcg_kmem_bypass())
2608 memcg = get_mem_cgroup_from_current();
2609 if (!mem_cgroup_is_root(memcg)) {
2610 ret = __memcg_kmem_charge_memcg(page, gfp, order, memcg);
2612 __SetPageKmemcg(page);
2614 css_put(&memcg->css);
2618 * __memcg_kmem_uncharge: uncharge a kmem page
2619 * @page: page to uncharge
2620 * @order: allocation order
2622 void __memcg_kmem_uncharge(struct page *page, int order)
2624 struct mem_cgroup *memcg = page->mem_cgroup;
2625 unsigned int nr_pages = 1 << order;
2630 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2632 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2633 page_counter_uncharge(&memcg->kmem, nr_pages);
2635 page_counter_uncharge(&memcg->memory, nr_pages);
2636 if (do_memsw_account())
2637 page_counter_uncharge(&memcg->memsw, nr_pages);
2639 page->mem_cgroup = NULL;
2641 /* slab pages do not have PageKmemcg flag set */
2642 if (PageKmemcg(page))
2643 __ClearPageKmemcg(page);
2645 css_put_many(&memcg->css, nr_pages);
2647 #endif /* CONFIG_MEMCG_KMEM */
2649 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2652 * Because tail pages are not marked as "used", set it. We're under
2653 * pgdat->lru_lock and migration entries setup in all page mappings.
2655 void mem_cgroup_split_huge_fixup(struct page *head)
2659 if (mem_cgroup_disabled())
2662 for (i = 1; i < HPAGE_PMD_NR; i++)
2663 head[i].mem_cgroup = head->mem_cgroup;
2665 __mod_memcg_state(head->mem_cgroup, MEMCG_RSS_HUGE, -HPAGE_PMD_NR);
2667 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2669 #ifdef CONFIG_MEMCG_SWAP
2671 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2672 * @entry: swap entry to be moved
2673 * @from: mem_cgroup which the entry is moved from
2674 * @to: mem_cgroup which the entry is moved to
2676 * It succeeds only when the swap_cgroup's record for this entry is the same
2677 * as the mem_cgroup's id of @from.
2679 * Returns 0 on success, -EINVAL on failure.
2681 * The caller must have charged to @to, IOW, called page_counter_charge() about
2682 * both res and memsw, and called css_get().
2684 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2685 struct mem_cgroup *from, struct mem_cgroup *to)
2687 unsigned short old_id, new_id;
2689 old_id = mem_cgroup_id(from);
2690 new_id = mem_cgroup_id(to);
2692 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2693 mod_memcg_state(from, MEMCG_SWAP, -1);
2694 mod_memcg_state(to, MEMCG_SWAP, 1);
2700 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2701 struct mem_cgroup *from, struct mem_cgroup *to)
2707 static DEFINE_MUTEX(memcg_max_mutex);
2709 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
2710 unsigned long max, bool memsw)
2712 bool enlarge = false;
2713 bool drained = false;
2715 bool limits_invariant;
2716 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
2719 if (signal_pending(current)) {
2724 mutex_lock(&memcg_max_mutex);
2726 * Make sure that the new limit (memsw or memory limit) doesn't
2727 * break our basic invariant rule memory.max <= memsw.max.
2729 limits_invariant = memsw ? max >= memcg->memory.max :
2730 max <= memcg->memsw.max;
2731 if (!limits_invariant) {
2732 mutex_unlock(&memcg_max_mutex);
2736 if (max > counter->max)
2738 ret = page_counter_set_max(counter, max);
2739 mutex_unlock(&memcg_max_mutex);
2745 drain_all_stock(memcg);
2750 if (!try_to_free_mem_cgroup_pages(memcg, 1,
2751 GFP_KERNEL, !memsw)) {
2757 if (!ret && enlarge)
2758 memcg_oom_recover(memcg);
2763 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
2765 unsigned long *total_scanned)
2767 unsigned long nr_reclaimed = 0;
2768 struct mem_cgroup_per_node *mz, *next_mz = NULL;
2769 unsigned long reclaimed;
2771 struct mem_cgroup_tree_per_node *mctz;
2772 unsigned long excess;
2773 unsigned long nr_scanned;
2778 mctz = soft_limit_tree_node(pgdat->node_id);
2781 * Do not even bother to check the largest node if the root
2782 * is empty. Do it lockless to prevent lock bouncing. Races
2783 * are acceptable as soft limit is best effort anyway.
2785 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
2789 * This loop can run a while, specially if mem_cgroup's continuously
2790 * keep exceeding their soft limit and putting the system under
2797 mz = mem_cgroup_largest_soft_limit_node(mctz);
2802 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
2803 gfp_mask, &nr_scanned);
2804 nr_reclaimed += reclaimed;
2805 *total_scanned += nr_scanned;
2806 spin_lock_irq(&mctz->lock);
2807 __mem_cgroup_remove_exceeded(mz, mctz);
2810 * If we failed to reclaim anything from this memory cgroup
2811 * it is time to move on to the next cgroup
2815 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
2817 excess = soft_limit_excess(mz->memcg);
2819 * One school of thought says that we should not add
2820 * back the node to the tree if reclaim returns 0.
2821 * But our reclaim could return 0, simply because due
2822 * to priority we are exposing a smaller subset of
2823 * memory to reclaim from. Consider this as a longer
2826 /* If excess == 0, no tree ops */
2827 __mem_cgroup_insert_exceeded(mz, mctz, excess);
2828 spin_unlock_irq(&mctz->lock);
2829 css_put(&mz->memcg->css);
2832 * Could not reclaim anything and there are no more
2833 * mem cgroups to try or we seem to be looping without
2834 * reclaiming anything.
2836 if (!nr_reclaimed &&
2838 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
2840 } while (!nr_reclaimed);
2842 css_put(&next_mz->memcg->css);
2843 return nr_reclaimed;
2847 * Test whether @memcg has children, dead or alive. Note that this
2848 * function doesn't care whether @memcg has use_hierarchy enabled and
2849 * returns %true if there are child csses according to the cgroup
2850 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2852 static inline bool memcg_has_children(struct mem_cgroup *memcg)
2857 ret = css_next_child(NULL, &memcg->css);
2863 * Reclaims as many pages from the given memcg as possible.
2865 * Caller is responsible for holding css reference for memcg.
2867 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
2869 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2871 /* we call try-to-free pages for make this cgroup empty */
2872 lru_add_drain_all();
2874 drain_all_stock(memcg);
2876 /* try to free all pages in this cgroup */
2877 while (nr_retries && page_counter_read(&memcg->memory)) {
2880 if (signal_pending(current))
2883 progress = try_to_free_mem_cgroup_pages(memcg, 1,
2887 /* maybe some writeback is necessary */
2888 congestion_wait(BLK_RW_ASYNC, HZ/10);
2896 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
2897 char *buf, size_t nbytes,
2900 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
2902 if (mem_cgroup_is_root(memcg))
2904 return mem_cgroup_force_empty(memcg) ?: nbytes;
2907 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
2910 return mem_cgroup_from_css(css)->use_hierarchy;
2913 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
2914 struct cftype *cft, u64 val)
2917 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2918 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
2920 if (memcg->use_hierarchy == val)
2924 * If parent's use_hierarchy is set, we can't make any modifications
2925 * in the child subtrees. If it is unset, then the change can
2926 * occur, provided the current cgroup has no children.
2928 * For the root cgroup, parent_mem is NULL, we allow value to be
2929 * set if there are no children.
2931 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
2932 (val == 1 || val == 0)) {
2933 if (!memcg_has_children(memcg))
2934 memcg->use_hierarchy = val;
2943 struct accumulated_stats {
2944 unsigned long stat[MEMCG_NR_STAT];
2945 unsigned long events[NR_VM_EVENT_ITEMS];
2946 unsigned long lru_pages[NR_LRU_LISTS];
2947 const unsigned int *stats_array;
2948 const unsigned int *events_array;
2953 static void accumulate_memcg_tree(struct mem_cgroup *memcg,
2954 struct accumulated_stats *acc)
2956 struct mem_cgroup *mi;
2959 for_each_mem_cgroup_tree(mi, memcg) {
2960 for (i = 0; i < acc->stats_size; i++)
2961 acc->stat[i] += memcg_page_state(mi,
2962 acc->stats_array ? acc->stats_array[i] : i);
2964 for (i = 0; i < acc->events_size; i++)
2965 acc->events[i] += memcg_sum_events(mi,
2966 acc->events_array ? acc->events_array[i] : i);
2968 for (i = 0; i < NR_LRU_LISTS; i++)
2969 acc->lru_pages[i] += memcg_page_state(mi,
2974 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
2976 unsigned long val = 0;
2978 if (mem_cgroup_is_root(memcg)) {
2979 struct mem_cgroup *iter;
2981 for_each_mem_cgroup_tree(iter, memcg) {
2982 val += memcg_page_state(iter, MEMCG_CACHE);
2983 val += memcg_page_state(iter, MEMCG_RSS);
2985 val += memcg_page_state(iter, MEMCG_SWAP);
2989 val = page_counter_read(&memcg->memory);
2991 val = page_counter_read(&memcg->memsw);
3004 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3007 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3008 struct page_counter *counter;
3010 switch (MEMFILE_TYPE(cft->private)) {
3012 counter = &memcg->memory;
3015 counter = &memcg->memsw;
3018 counter = &memcg->kmem;
3021 counter = &memcg->tcpmem;
3027 switch (MEMFILE_ATTR(cft->private)) {
3029 if (counter == &memcg->memory)
3030 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3031 if (counter == &memcg->memsw)
3032 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3033 return (u64)page_counter_read(counter) * PAGE_SIZE;
3035 return (u64)counter->max * PAGE_SIZE;
3037 return (u64)counter->watermark * PAGE_SIZE;
3039 return counter->failcnt;
3040 case RES_SOFT_LIMIT:
3041 return (u64)memcg->soft_limit * PAGE_SIZE;
3047 #ifdef CONFIG_MEMCG_KMEM
3048 static int memcg_online_kmem(struct mem_cgroup *memcg)
3052 if (cgroup_memory_nokmem)
3055 BUG_ON(memcg->kmemcg_id >= 0);
3056 BUG_ON(memcg->kmem_state);
3058 memcg_id = memcg_alloc_cache_id();
3062 static_branch_inc(&memcg_kmem_enabled_key);
3064 * A memory cgroup is considered kmem-online as soon as it gets
3065 * kmemcg_id. Setting the id after enabling static branching will
3066 * guarantee no one starts accounting before all call sites are
3069 memcg->kmemcg_id = memcg_id;
3070 memcg->kmem_state = KMEM_ONLINE;
3071 INIT_LIST_HEAD(&memcg->kmem_caches);
3076 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3078 struct cgroup_subsys_state *css;
3079 struct mem_cgroup *parent, *child;
3082 if (memcg->kmem_state != KMEM_ONLINE)
3085 * Clear the online state before clearing memcg_caches array
3086 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
3087 * guarantees that no cache will be created for this cgroup
3088 * after we are done (see memcg_create_kmem_cache()).
3090 memcg->kmem_state = KMEM_ALLOCATED;
3092 memcg_deactivate_kmem_caches(memcg);
3094 kmemcg_id = memcg->kmemcg_id;
3095 BUG_ON(kmemcg_id < 0);
3097 parent = parent_mem_cgroup(memcg);
3099 parent = root_mem_cgroup;
3102 * Change kmemcg_id of this cgroup and all its descendants to the
3103 * parent's id, and then move all entries from this cgroup's list_lrus
3104 * to ones of the parent. After we have finished, all list_lrus
3105 * corresponding to this cgroup are guaranteed to remain empty. The
3106 * ordering is imposed by list_lru_node->lock taken by
3107 * memcg_drain_all_list_lrus().
3109 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3110 css_for_each_descendant_pre(css, &memcg->css) {
3111 child = mem_cgroup_from_css(css);
3112 BUG_ON(child->kmemcg_id != kmemcg_id);
3113 child->kmemcg_id = parent->kmemcg_id;
3114 if (!memcg->use_hierarchy)
3119 memcg_drain_all_list_lrus(kmemcg_id, parent);
3121 memcg_free_cache_id(kmemcg_id);
3124 static void memcg_free_kmem(struct mem_cgroup *memcg)
3126 /* css_alloc() failed, offlining didn't happen */
3127 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3128 memcg_offline_kmem(memcg);
3130 if (memcg->kmem_state == KMEM_ALLOCATED) {
3131 memcg_destroy_kmem_caches(memcg);
3132 static_branch_dec(&memcg_kmem_enabled_key);
3133 WARN_ON(page_counter_read(&memcg->kmem));
3137 static int memcg_online_kmem(struct mem_cgroup *memcg)
3141 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3144 static void memcg_free_kmem(struct mem_cgroup *memcg)
3147 #endif /* CONFIG_MEMCG_KMEM */
3149 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3154 mutex_lock(&memcg_max_mutex);
3155 ret = page_counter_set_max(&memcg->kmem, max);
3156 mutex_unlock(&memcg_max_mutex);
3160 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3164 mutex_lock(&memcg_max_mutex);
3166 ret = page_counter_set_max(&memcg->tcpmem, max);
3170 if (!memcg->tcpmem_active) {
3172 * The active flag needs to be written after the static_key
3173 * update. This is what guarantees that the socket activation
3174 * function is the last one to run. See mem_cgroup_sk_alloc()
3175 * for details, and note that we don't mark any socket as
3176 * belonging to this memcg until that flag is up.
3178 * We need to do this, because static_keys will span multiple
3179 * sites, but we can't control their order. If we mark a socket
3180 * as accounted, but the accounting functions are not patched in
3181 * yet, we'll lose accounting.
3183 * We never race with the readers in mem_cgroup_sk_alloc(),
3184 * because when this value change, the code to process it is not
3187 static_branch_inc(&memcg_sockets_enabled_key);
3188 memcg->tcpmem_active = true;
3191 mutex_unlock(&memcg_max_mutex);
3196 * The user of this function is...
3199 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3200 char *buf, size_t nbytes, loff_t off)
3202 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3203 unsigned long nr_pages;
3206 buf = strstrip(buf);
3207 ret = page_counter_memparse(buf, "-1", &nr_pages);
3211 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3213 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3217 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3219 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3222 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3225 ret = memcg_update_kmem_max(memcg, nr_pages);
3228 ret = memcg_update_tcp_max(memcg, nr_pages);
3232 case RES_SOFT_LIMIT:
3233 memcg->soft_limit = nr_pages;
3237 return ret ?: nbytes;
3240 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3241 size_t nbytes, loff_t off)
3243 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3244 struct page_counter *counter;
3246 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3248 counter = &memcg->memory;
3251 counter = &memcg->memsw;
3254 counter = &memcg->kmem;
3257 counter = &memcg->tcpmem;
3263 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3265 page_counter_reset_watermark(counter);
3268 counter->failcnt = 0;
3277 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3280 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3284 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3285 struct cftype *cft, u64 val)
3287 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3289 if (val & ~MOVE_MASK)
3293 * No kind of locking is needed in here, because ->can_attach() will
3294 * check this value once in the beginning of the process, and then carry
3295 * on with stale data. This means that changes to this value will only
3296 * affect task migrations starting after the change.
3298 memcg->move_charge_at_immigrate = val;
3302 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3303 struct cftype *cft, u64 val)
3311 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3312 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3313 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3315 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3316 int nid, unsigned int lru_mask)
3318 struct lruvec *lruvec = mem_cgroup_lruvec(NODE_DATA(nid), memcg);
3319 unsigned long nr = 0;
3322 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3325 if (!(BIT(lru) & lru_mask))
3327 nr += lruvec_page_state(lruvec, NR_LRU_BASE + lru);
3332 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3333 unsigned int lru_mask)
3335 unsigned long nr = 0;
3339 if (!(BIT(lru) & lru_mask))
3341 nr += memcg_page_state(memcg, NR_LRU_BASE + lru);
3346 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3350 unsigned int lru_mask;
3353 static const struct numa_stat stats[] = {
3354 { "total", LRU_ALL },
3355 { "file", LRU_ALL_FILE },
3356 { "anon", LRU_ALL_ANON },
3357 { "unevictable", BIT(LRU_UNEVICTABLE) },
3359 const struct numa_stat *stat;
3362 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3364 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3365 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3366 seq_printf(m, "%s=%lu", stat->name, nr);
3367 for_each_node_state(nid, N_MEMORY) {
3368 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3370 seq_printf(m, " N%d=%lu", nid, nr);
3375 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3376 struct mem_cgroup *iter;
3379 for_each_mem_cgroup_tree(iter, memcg)
3380 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3381 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3382 for_each_node_state(nid, N_MEMORY) {
3384 for_each_mem_cgroup_tree(iter, memcg)
3385 nr += mem_cgroup_node_nr_lru_pages(
3386 iter, nid, stat->lru_mask);
3387 seq_printf(m, " N%d=%lu", nid, nr);
3394 #endif /* CONFIG_NUMA */
3396 /* Universal VM events cgroup1 shows, original sort order */
3397 static const unsigned int memcg1_events[] = {
3404 static const char *const memcg1_event_names[] = {
3411 static int memcg_stat_show(struct seq_file *m, void *v)
3413 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3414 unsigned long memory, memsw;
3415 struct mem_cgroup *mi;
3417 struct accumulated_stats acc;
3419 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
3420 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3422 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3423 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3425 seq_printf(m, "%s %lu\n", memcg1_stat_names[i],
3426 memcg_page_state(memcg, memcg1_stats[i]) *
3430 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3431 seq_printf(m, "%s %lu\n", memcg1_event_names[i],
3432 memcg_sum_events(memcg, memcg1_events[i]));
3434 for (i = 0; i < NR_LRU_LISTS; i++)
3435 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3436 memcg_page_state(memcg, NR_LRU_BASE + i) *
3439 /* Hierarchical information */
3440 memory = memsw = PAGE_COUNTER_MAX;
3441 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3442 memory = min(memory, mi->memory.max);
3443 memsw = min(memsw, mi->memsw.max);
3445 seq_printf(m, "hierarchical_memory_limit %llu\n",
3446 (u64)memory * PAGE_SIZE);
3447 if (do_memsw_account())
3448 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3449 (u64)memsw * PAGE_SIZE);
3451 memset(&acc, 0, sizeof(acc));
3452 acc.stats_size = ARRAY_SIZE(memcg1_stats);
3453 acc.stats_array = memcg1_stats;
3454 acc.events_size = ARRAY_SIZE(memcg1_events);
3455 acc.events_array = memcg1_events;
3456 accumulate_memcg_tree(memcg, &acc);
3458 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3459 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3461 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
3462 (u64)acc.stat[i] * PAGE_SIZE);
3465 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3466 seq_printf(m, "total_%s %llu\n", memcg1_event_names[i],
3467 (u64)acc.events[i]);
3469 for (i = 0; i < NR_LRU_LISTS; i++)
3470 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i],
3471 (u64)acc.lru_pages[i] * PAGE_SIZE);
3473 #ifdef CONFIG_DEBUG_VM
3476 struct mem_cgroup_per_node *mz;
3477 struct zone_reclaim_stat *rstat;
3478 unsigned long recent_rotated[2] = {0, 0};
3479 unsigned long recent_scanned[2] = {0, 0};
3481 for_each_online_pgdat(pgdat) {
3482 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
3483 rstat = &mz->lruvec.reclaim_stat;
3485 recent_rotated[0] += rstat->recent_rotated[0];
3486 recent_rotated[1] += rstat->recent_rotated[1];
3487 recent_scanned[0] += rstat->recent_scanned[0];
3488 recent_scanned[1] += rstat->recent_scanned[1];
3490 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3491 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3492 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3493 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3500 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3503 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3505 return mem_cgroup_swappiness(memcg);
3508 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3509 struct cftype *cft, u64 val)
3511 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3517 memcg->swappiness = val;
3519 vm_swappiness = val;
3524 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3526 struct mem_cgroup_threshold_ary *t;
3527 unsigned long usage;
3532 t = rcu_dereference(memcg->thresholds.primary);
3534 t = rcu_dereference(memcg->memsw_thresholds.primary);
3539 usage = mem_cgroup_usage(memcg, swap);
3542 * current_threshold points to threshold just below or equal to usage.
3543 * If it's not true, a threshold was crossed after last
3544 * call of __mem_cgroup_threshold().
3546 i = t->current_threshold;
3549 * Iterate backward over array of thresholds starting from
3550 * current_threshold and check if a threshold is crossed.
3551 * If none of thresholds below usage is crossed, we read
3552 * only one element of the array here.
3554 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3555 eventfd_signal(t->entries[i].eventfd, 1);
3557 /* i = current_threshold + 1 */
3561 * Iterate forward over array of thresholds starting from
3562 * current_threshold+1 and check if a threshold is crossed.
3563 * If none of thresholds above usage is crossed, we read
3564 * only one element of the array here.
3566 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3567 eventfd_signal(t->entries[i].eventfd, 1);
3569 /* Update current_threshold */
3570 t->current_threshold = i - 1;
3575 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3578 __mem_cgroup_threshold(memcg, false);
3579 if (do_memsw_account())
3580 __mem_cgroup_threshold(memcg, true);
3582 memcg = parent_mem_cgroup(memcg);
3586 static int compare_thresholds(const void *a, const void *b)
3588 const struct mem_cgroup_threshold *_a = a;
3589 const struct mem_cgroup_threshold *_b = b;
3591 if (_a->threshold > _b->threshold)
3594 if (_a->threshold < _b->threshold)
3600 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3602 struct mem_cgroup_eventfd_list *ev;
3604 spin_lock(&memcg_oom_lock);
3606 list_for_each_entry(ev, &memcg->oom_notify, list)
3607 eventfd_signal(ev->eventfd, 1);
3609 spin_unlock(&memcg_oom_lock);
3613 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3615 struct mem_cgroup *iter;
3617 for_each_mem_cgroup_tree(iter, memcg)
3618 mem_cgroup_oom_notify_cb(iter);
3621 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3622 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3624 struct mem_cgroup_thresholds *thresholds;
3625 struct mem_cgroup_threshold_ary *new;
3626 unsigned long threshold;
3627 unsigned long usage;
3630 ret = page_counter_memparse(args, "-1", &threshold);
3634 mutex_lock(&memcg->thresholds_lock);
3637 thresholds = &memcg->thresholds;
3638 usage = mem_cgroup_usage(memcg, false);
3639 } else if (type == _MEMSWAP) {
3640 thresholds = &memcg->memsw_thresholds;
3641 usage = mem_cgroup_usage(memcg, true);
3645 /* Check if a threshold crossed before adding a new one */
3646 if (thresholds->primary)
3647 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3649 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3651 /* Allocate memory for new array of thresholds */
3652 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
3659 /* Copy thresholds (if any) to new array */
3660 if (thresholds->primary) {
3661 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3662 sizeof(struct mem_cgroup_threshold));
3665 /* Add new threshold */
3666 new->entries[size - 1].eventfd = eventfd;
3667 new->entries[size - 1].threshold = threshold;
3669 /* Sort thresholds. Registering of new threshold isn't time-critical */
3670 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3671 compare_thresholds, NULL);
3673 /* Find current threshold */
3674 new->current_threshold = -1;
3675 for (i = 0; i < size; i++) {
3676 if (new->entries[i].threshold <= usage) {
3678 * new->current_threshold will not be used until
3679 * rcu_assign_pointer(), so it's safe to increment
3682 ++new->current_threshold;
3687 /* Free old spare buffer and save old primary buffer as spare */
3688 kfree(thresholds->spare);
3689 thresholds->spare = thresholds->primary;
3691 rcu_assign_pointer(thresholds->primary, new);
3693 /* To be sure that nobody uses thresholds */
3697 mutex_unlock(&memcg->thresholds_lock);
3702 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3703 struct eventfd_ctx *eventfd, const char *args)
3705 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
3708 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
3709 struct eventfd_ctx *eventfd, const char *args)
3711 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
3714 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3715 struct eventfd_ctx *eventfd, enum res_type type)
3717 struct mem_cgroup_thresholds *thresholds;
3718 struct mem_cgroup_threshold_ary *new;
3719 unsigned long usage;
3722 mutex_lock(&memcg->thresholds_lock);
3725 thresholds = &memcg->thresholds;
3726 usage = mem_cgroup_usage(memcg, false);
3727 } else if (type == _MEMSWAP) {
3728 thresholds = &memcg->memsw_thresholds;
3729 usage = mem_cgroup_usage(memcg, true);
3733 if (!thresholds->primary)
3736 /* Check if a threshold crossed before removing */
3737 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3739 /* Calculate new number of threshold */
3741 for (i = 0; i < thresholds->primary->size; i++) {
3742 if (thresholds->primary->entries[i].eventfd != eventfd)
3746 new = thresholds->spare;
3748 /* Set thresholds array to NULL if we don't have thresholds */
3757 /* Copy thresholds and find current threshold */
3758 new->current_threshold = -1;
3759 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3760 if (thresholds->primary->entries[i].eventfd == eventfd)
3763 new->entries[j] = thresholds->primary->entries[i];
3764 if (new->entries[j].threshold <= usage) {
3766 * new->current_threshold will not be used
3767 * until rcu_assign_pointer(), so it's safe to increment
3770 ++new->current_threshold;
3776 /* Swap primary and spare array */
3777 thresholds->spare = thresholds->primary;
3779 rcu_assign_pointer(thresholds->primary, new);
3781 /* To be sure that nobody uses thresholds */
3784 /* If all events are unregistered, free the spare array */
3786 kfree(thresholds->spare);
3787 thresholds->spare = NULL;
3790 mutex_unlock(&memcg->thresholds_lock);
3793 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3794 struct eventfd_ctx *eventfd)
3796 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
3799 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3800 struct eventfd_ctx *eventfd)
3802 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
3805 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
3806 struct eventfd_ctx *eventfd, const char *args)
3808 struct mem_cgroup_eventfd_list *event;
3810 event = kmalloc(sizeof(*event), GFP_KERNEL);
3814 spin_lock(&memcg_oom_lock);
3816 event->eventfd = eventfd;
3817 list_add(&event->list, &memcg->oom_notify);
3819 /* already in OOM ? */
3820 if (memcg->under_oom)
3821 eventfd_signal(eventfd, 1);
3822 spin_unlock(&memcg_oom_lock);
3827 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
3828 struct eventfd_ctx *eventfd)
3830 struct mem_cgroup_eventfd_list *ev, *tmp;
3832 spin_lock(&memcg_oom_lock);
3834 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
3835 if (ev->eventfd == eventfd) {
3836 list_del(&ev->list);
3841 spin_unlock(&memcg_oom_lock);
3844 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
3846 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
3848 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
3849 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
3850 seq_printf(sf, "oom_kill %lu\n",
3851 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
3855 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
3856 struct cftype *cft, u64 val)
3858 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3860 /* cannot set to root cgroup and only 0 and 1 are allowed */
3861 if (!css->parent || !((val == 0) || (val == 1)))
3864 memcg->oom_kill_disable = val;
3866 memcg_oom_recover(memcg);
3871 #ifdef CONFIG_CGROUP_WRITEBACK
3873 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3875 return wb_domain_init(&memcg->cgwb_domain, gfp);
3878 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3880 wb_domain_exit(&memcg->cgwb_domain);
3883 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3885 wb_domain_size_changed(&memcg->cgwb_domain);
3888 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
3890 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3892 if (!memcg->css.parent)
3895 return &memcg->cgwb_domain;
3899 * idx can be of type enum memcg_stat_item or node_stat_item.
3900 * Keep in sync with memcg_exact_page().
3902 static unsigned long memcg_exact_page_state(struct mem_cgroup *memcg, int idx)
3904 long x = atomic_long_read(&memcg->stat[idx]);
3907 for_each_online_cpu(cpu)
3908 x += per_cpu_ptr(memcg->stat_cpu, cpu)->count[idx];
3915 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3916 * @wb: bdi_writeback in question
3917 * @pfilepages: out parameter for number of file pages
3918 * @pheadroom: out parameter for number of allocatable pages according to memcg
3919 * @pdirty: out parameter for number of dirty pages
3920 * @pwriteback: out parameter for number of pages under writeback
3922 * Determine the numbers of file, headroom, dirty, and writeback pages in
3923 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
3924 * is a bit more involved.
3926 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
3927 * headroom is calculated as the lowest headroom of itself and the
3928 * ancestors. Note that this doesn't consider the actual amount of
3929 * available memory in the system. The caller should further cap
3930 * *@pheadroom accordingly.
3932 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
3933 unsigned long *pheadroom, unsigned long *pdirty,
3934 unsigned long *pwriteback)
3936 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3937 struct mem_cgroup *parent;
3939 *pdirty = memcg_exact_page_state(memcg, NR_FILE_DIRTY);
3941 /* this should eventually include NR_UNSTABLE_NFS */
3942 *pwriteback = memcg_exact_page_state(memcg, NR_WRITEBACK);
3943 *pfilepages = memcg_exact_page_state(memcg, NR_INACTIVE_FILE) +
3944 memcg_exact_page_state(memcg, NR_ACTIVE_FILE);
3945 *pheadroom = PAGE_COUNTER_MAX;
3947 while ((parent = parent_mem_cgroup(memcg))) {
3948 unsigned long ceiling = min(memcg->memory.max, memcg->high);
3949 unsigned long used = page_counter_read(&memcg->memory);
3951 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
3956 #else /* CONFIG_CGROUP_WRITEBACK */
3958 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3963 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3967 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3971 #endif /* CONFIG_CGROUP_WRITEBACK */
3974 * DO NOT USE IN NEW FILES.
3976 * "cgroup.event_control" implementation.
3978 * This is way over-engineered. It tries to support fully configurable
3979 * events for each user. Such level of flexibility is completely
3980 * unnecessary especially in the light of the planned unified hierarchy.
3982 * Please deprecate this and replace with something simpler if at all
3987 * Unregister event and free resources.
3989 * Gets called from workqueue.
3991 static void memcg_event_remove(struct work_struct *work)
3993 struct mem_cgroup_event *event =
3994 container_of(work, struct mem_cgroup_event, remove);
3995 struct mem_cgroup *memcg = event->memcg;
3997 remove_wait_queue(event->wqh, &event->wait);
3999 event->unregister_event(memcg, event->eventfd);
4001 /* Notify userspace the event is going away. */
4002 eventfd_signal(event->eventfd, 1);
4004 eventfd_ctx_put(event->eventfd);
4006 css_put(&memcg->css);
4010 * Gets called on EPOLLHUP on eventfd when user closes it.
4012 * Called with wqh->lock held and interrupts disabled.
4014 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4015 int sync, void *key)
4017 struct mem_cgroup_event *event =
4018 container_of(wait, struct mem_cgroup_event, wait);
4019 struct mem_cgroup *memcg = event->memcg;
4020 __poll_t flags = key_to_poll(key);
4022 if (flags & EPOLLHUP) {
4024 * If the event has been detached at cgroup removal, we
4025 * can simply return knowing the other side will cleanup
4028 * We can't race against event freeing since the other
4029 * side will require wqh->lock via remove_wait_queue(),
4032 spin_lock(&memcg->event_list_lock);
4033 if (!list_empty(&event->list)) {
4034 list_del_init(&event->list);
4036 * We are in atomic context, but cgroup_event_remove()
4037 * may sleep, so we have to call it in workqueue.
4039 schedule_work(&event->remove);
4041 spin_unlock(&memcg->event_list_lock);
4047 static void memcg_event_ptable_queue_proc(struct file *file,
4048 wait_queue_head_t *wqh, poll_table *pt)
4050 struct mem_cgroup_event *event =
4051 container_of(pt, struct mem_cgroup_event, pt);
4054 add_wait_queue(wqh, &event->wait);
4058 * DO NOT USE IN NEW FILES.
4060 * Parse input and register new cgroup event handler.
4062 * Input must be in format '<event_fd> <control_fd> <args>'.
4063 * Interpretation of args is defined by control file implementation.
4065 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4066 char *buf, size_t nbytes, loff_t off)
4068 struct cgroup_subsys_state *css = of_css(of);
4069 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4070 struct mem_cgroup_event *event;
4071 struct cgroup_subsys_state *cfile_css;
4072 unsigned int efd, cfd;
4079 buf = strstrip(buf);
4081 efd = simple_strtoul(buf, &endp, 10);
4086 cfd = simple_strtoul(buf, &endp, 10);
4087 if ((*endp != ' ') && (*endp != '\0'))
4091 event = kzalloc(sizeof(*event), GFP_KERNEL);
4095 event->memcg = memcg;
4096 INIT_LIST_HEAD(&event->list);
4097 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4098 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4099 INIT_WORK(&event->remove, memcg_event_remove);
4107 event->eventfd = eventfd_ctx_fileget(efile.file);
4108 if (IS_ERR(event->eventfd)) {
4109 ret = PTR_ERR(event->eventfd);
4116 goto out_put_eventfd;
4119 /* the process need read permission on control file */
4120 /* AV: shouldn't we check that it's been opened for read instead? */
4121 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4126 * Determine the event callbacks and set them in @event. This used
4127 * to be done via struct cftype but cgroup core no longer knows
4128 * about these events. The following is crude but the whole thing
4129 * is for compatibility anyway.
4131 * DO NOT ADD NEW FILES.
4133 name = cfile.file->f_path.dentry->d_name.name;
4135 if (!strcmp(name, "memory.usage_in_bytes")) {
4136 event->register_event = mem_cgroup_usage_register_event;
4137 event->unregister_event = mem_cgroup_usage_unregister_event;
4138 } else if (!strcmp(name, "memory.oom_control")) {
4139 event->register_event = mem_cgroup_oom_register_event;
4140 event->unregister_event = mem_cgroup_oom_unregister_event;
4141 } else if (!strcmp(name, "memory.pressure_level")) {
4142 event->register_event = vmpressure_register_event;
4143 event->unregister_event = vmpressure_unregister_event;
4144 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4145 event->register_event = memsw_cgroup_usage_register_event;
4146 event->unregister_event = memsw_cgroup_usage_unregister_event;
4153 * Verify @cfile should belong to @css. Also, remaining events are
4154 * automatically removed on cgroup destruction but the removal is
4155 * asynchronous, so take an extra ref on @css.
4157 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4158 &memory_cgrp_subsys);
4160 if (IS_ERR(cfile_css))
4162 if (cfile_css != css) {
4167 ret = event->register_event(memcg, event->eventfd, buf);
4171 vfs_poll(efile.file, &event->pt);
4173 spin_lock(&memcg->event_list_lock);
4174 list_add(&event->list, &memcg->event_list);
4175 spin_unlock(&memcg->event_list_lock);
4187 eventfd_ctx_put(event->eventfd);
4196 static struct cftype mem_cgroup_legacy_files[] = {
4198 .name = "usage_in_bytes",
4199 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4200 .read_u64 = mem_cgroup_read_u64,
4203 .name = "max_usage_in_bytes",
4204 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4205 .write = mem_cgroup_reset,
4206 .read_u64 = mem_cgroup_read_u64,
4209 .name = "limit_in_bytes",
4210 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4211 .write = mem_cgroup_write,
4212 .read_u64 = mem_cgroup_read_u64,
4215 .name = "soft_limit_in_bytes",
4216 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4217 .write = mem_cgroup_write,
4218 .read_u64 = mem_cgroup_read_u64,
4222 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4223 .write = mem_cgroup_reset,
4224 .read_u64 = mem_cgroup_read_u64,
4228 .seq_show = memcg_stat_show,
4231 .name = "force_empty",
4232 .write = mem_cgroup_force_empty_write,
4235 .name = "use_hierarchy",
4236 .write_u64 = mem_cgroup_hierarchy_write,
4237 .read_u64 = mem_cgroup_hierarchy_read,
4240 .name = "cgroup.event_control", /* XXX: for compat */
4241 .write = memcg_write_event_control,
4242 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4245 .name = "swappiness",
4246 .read_u64 = mem_cgroup_swappiness_read,
4247 .write_u64 = mem_cgroup_swappiness_write,
4250 .name = "move_charge_at_immigrate",
4251 .read_u64 = mem_cgroup_move_charge_read,
4252 .write_u64 = mem_cgroup_move_charge_write,
4255 .name = "oom_control",
4256 .seq_show = mem_cgroup_oom_control_read,
4257 .write_u64 = mem_cgroup_oom_control_write,
4258 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4261 .name = "pressure_level",
4265 .name = "numa_stat",
4266 .seq_show = memcg_numa_stat_show,
4270 .name = "kmem.limit_in_bytes",
4271 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4272 .write = mem_cgroup_write,
4273 .read_u64 = mem_cgroup_read_u64,
4276 .name = "kmem.usage_in_bytes",
4277 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4278 .read_u64 = mem_cgroup_read_u64,
4281 .name = "kmem.failcnt",
4282 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4283 .write = mem_cgroup_reset,
4284 .read_u64 = mem_cgroup_read_u64,
4287 .name = "kmem.max_usage_in_bytes",
4288 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4289 .write = mem_cgroup_reset,
4290 .read_u64 = mem_cgroup_read_u64,
4292 #if defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG)
4294 .name = "kmem.slabinfo",
4295 .seq_start = memcg_slab_start,
4296 .seq_next = memcg_slab_next,
4297 .seq_stop = memcg_slab_stop,
4298 .seq_show = memcg_slab_show,
4302 .name = "kmem.tcp.limit_in_bytes",
4303 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
4304 .write = mem_cgroup_write,
4305 .read_u64 = mem_cgroup_read_u64,
4308 .name = "kmem.tcp.usage_in_bytes",
4309 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
4310 .read_u64 = mem_cgroup_read_u64,
4313 .name = "kmem.tcp.failcnt",
4314 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
4315 .write = mem_cgroup_reset,
4316 .read_u64 = mem_cgroup_read_u64,
4319 .name = "kmem.tcp.max_usage_in_bytes",
4320 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
4321 .write = mem_cgroup_reset,
4322 .read_u64 = mem_cgroup_read_u64,
4324 { }, /* terminate */
4328 * Private memory cgroup IDR
4330 * Swap-out records and page cache shadow entries need to store memcg
4331 * references in constrained space, so we maintain an ID space that is
4332 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4333 * memory-controlled cgroups to 64k.
4335 * However, there usually are many references to the oflline CSS after
4336 * the cgroup has been destroyed, such as page cache or reclaimable
4337 * slab objects, that don't need to hang on to the ID. We want to keep
4338 * those dead CSS from occupying IDs, or we might quickly exhaust the
4339 * relatively small ID space and prevent the creation of new cgroups
4340 * even when there are much fewer than 64k cgroups - possibly none.
4342 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4343 * be freed and recycled when it's no longer needed, which is usually
4344 * when the CSS is offlined.
4346 * The only exception to that are records of swapped out tmpfs/shmem
4347 * pages that need to be attributed to live ancestors on swapin. But
4348 * those references are manageable from userspace.
4351 static DEFINE_IDR(mem_cgroup_idr);
4353 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
4355 if (memcg->id.id > 0) {
4356 idr_remove(&mem_cgroup_idr, memcg->id.id);
4361 static void mem_cgroup_id_get_many(struct mem_cgroup *memcg, unsigned int n)
4363 refcount_add(n, &memcg->id.ref);
4366 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
4368 if (refcount_sub_and_test(n, &memcg->id.ref)) {
4369 mem_cgroup_id_remove(memcg);
4371 /* Memcg ID pins CSS */
4372 css_put(&memcg->css);
4376 static inline void mem_cgroup_id_get(struct mem_cgroup *memcg)
4378 mem_cgroup_id_get_many(memcg, 1);
4381 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
4383 mem_cgroup_id_put_many(memcg, 1);
4387 * mem_cgroup_from_id - look up a memcg from a memcg id
4388 * @id: the memcg id to look up
4390 * Caller must hold rcu_read_lock().
4392 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
4394 WARN_ON_ONCE(!rcu_read_lock_held());
4395 return idr_find(&mem_cgroup_idr, id);
4398 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4400 struct mem_cgroup_per_node *pn;
4403 * This routine is called against possible nodes.
4404 * But it's BUG to call kmalloc() against offline node.
4406 * TODO: this routine can waste much memory for nodes which will
4407 * never be onlined. It's better to use memory hotplug callback
4410 if (!node_state(node, N_NORMAL_MEMORY))
4412 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4416 pn->lruvec_stat_cpu = alloc_percpu(struct lruvec_stat);
4417 if (!pn->lruvec_stat_cpu) {
4422 lruvec_init(&pn->lruvec);
4423 pn->usage_in_excess = 0;
4424 pn->on_tree = false;
4427 memcg->nodeinfo[node] = pn;
4431 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4433 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
4438 free_percpu(pn->lruvec_stat_cpu);
4442 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4447 free_mem_cgroup_per_node_info(memcg, node);
4448 free_percpu(memcg->stat_cpu);
4452 static void mem_cgroup_free(struct mem_cgroup *memcg)
4454 memcg_wb_domain_exit(memcg);
4455 __mem_cgroup_free(memcg);
4458 static struct mem_cgroup *mem_cgroup_alloc(void)
4460 struct mem_cgroup *memcg;
4464 size = sizeof(struct mem_cgroup);
4465 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4467 memcg = kzalloc(size, GFP_KERNEL);
4471 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
4472 1, MEM_CGROUP_ID_MAX,
4474 if (memcg->id.id < 0)
4477 memcg->stat_cpu = alloc_percpu(struct mem_cgroup_stat_cpu);
4478 if (!memcg->stat_cpu)
4482 if (alloc_mem_cgroup_per_node_info(memcg, node))
4485 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
4488 INIT_WORK(&memcg->high_work, high_work_func);
4489 memcg->last_scanned_node = MAX_NUMNODES;
4490 INIT_LIST_HEAD(&memcg->oom_notify);
4491 mutex_init(&memcg->thresholds_lock);
4492 spin_lock_init(&memcg->move_lock);
4493 vmpressure_init(&memcg->vmpressure);
4494 INIT_LIST_HEAD(&memcg->event_list);
4495 spin_lock_init(&memcg->event_list_lock);
4496 memcg->socket_pressure = jiffies;
4497 #ifdef CONFIG_MEMCG_KMEM
4498 memcg->kmemcg_id = -1;
4500 #ifdef CONFIG_CGROUP_WRITEBACK
4501 INIT_LIST_HEAD(&memcg->cgwb_list);
4503 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
4506 mem_cgroup_id_remove(memcg);
4507 __mem_cgroup_free(memcg);
4511 static struct cgroup_subsys_state * __ref
4512 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4514 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
4515 struct mem_cgroup *memcg;
4516 long error = -ENOMEM;
4518 memcg = mem_cgroup_alloc();
4520 return ERR_PTR(error);
4522 memcg->high = PAGE_COUNTER_MAX;
4523 memcg->soft_limit = PAGE_COUNTER_MAX;
4525 memcg->swappiness = mem_cgroup_swappiness(parent);
4526 memcg->oom_kill_disable = parent->oom_kill_disable;
4528 if (parent && parent->use_hierarchy) {
4529 memcg->use_hierarchy = true;
4530 page_counter_init(&memcg->memory, &parent->memory);
4531 page_counter_init(&memcg->swap, &parent->swap);
4532 page_counter_init(&memcg->memsw, &parent->memsw);
4533 page_counter_init(&memcg->kmem, &parent->kmem);
4534 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
4536 page_counter_init(&memcg->memory, NULL);
4537 page_counter_init(&memcg->swap, NULL);
4538 page_counter_init(&memcg->memsw, NULL);
4539 page_counter_init(&memcg->kmem, NULL);
4540 page_counter_init(&memcg->tcpmem, NULL);
4542 * Deeper hierachy with use_hierarchy == false doesn't make
4543 * much sense so let cgroup subsystem know about this
4544 * unfortunate state in our controller.
4546 if (parent != root_mem_cgroup)
4547 memory_cgrp_subsys.broken_hierarchy = true;
4550 /* The following stuff does not apply to the root */
4552 root_mem_cgroup = memcg;
4556 error = memcg_online_kmem(memcg);
4560 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4561 static_branch_inc(&memcg_sockets_enabled_key);
4565 mem_cgroup_id_remove(memcg);
4566 mem_cgroup_free(memcg);
4567 return ERR_PTR(-ENOMEM);
4570 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
4572 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4575 * A memcg must be visible for memcg_expand_shrinker_maps()
4576 * by the time the maps are allocated. So, we allocate maps
4577 * here, when for_each_mem_cgroup() can't skip it.
4579 if (memcg_alloc_shrinker_maps(memcg)) {
4580 mem_cgroup_id_remove(memcg);
4584 /* Online state pins memcg ID, memcg ID pins CSS */
4585 refcount_set(&memcg->id.ref, 1);
4590 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4592 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4593 struct mem_cgroup_event *event, *tmp;
4596 * Unregister events and notify userspace.
4597 * Notify userspace about cgroup removing only after rmdir of cgroup
4598 * directory to avoid race between userspace and kernelspace.
4600 spin_lock(&memcg->event_list_lock);
4601 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4602 list_del_init(&event->list);
4603 schedule_work(&event->remove);
4605 spin_unlock(&memcg->event_list_lock);
4607 page_counter_set_min(&memcg->memory, 0);
4608 page_counter_set_low(&memcg->memory, 0);
4610 memcg_offline_kmem(memcg);
4611 wb_memcg_offline(memcg);
4613 drain_all_stock(memcg);
4615 mem_cgroup_id_put(memcg);
4618 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
4620 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4622 invalidate_reclaim_iterators(memcg);
4625 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4627 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4629 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4630 static_branch_dec(&memcg_sockets_enabled_key);
4632 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
4633 static_branch_dec(&memcg_sockets_enabled_key);
4635 vmpressure_cleanup(&memcg->vmpressure);
4636 cancel_work_sync(&memcg->high_work);
4637 mem_cgroup_remove_from_trees(memcg);
4638 memcg_free_shrinker_maps(memcg);
4639 memcg_free_kmem(memcg);
4640 mem_cgroup_free(memcg);
4644 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4645 * @css: the target css
4647 * Reset the states of the mem_cgroup associated with @css. This is
4648 * invoked when the userland requests disabling on the default hierarchy
4649 * but the memcg is pinned through dependency. The memcg should stop
4650 * applying policies and should revert to the vanilla state as it may be
4651 * made visible again.
4653 * The current implementation only resets the essential configurations.
4654 * This needs to be expanded to cover all the visible parts.
4656 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4658 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4660 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
4661 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
4662 page_counter_set_max(&memcg->memsw, PAGE_COUNTER_MAX);
4663 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
4664 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
4665 page_counter_set_min(&memcg->memory, 0);
4666 page_counter_set_low(&memcg->memory, 0);
4667 memcg->high = PAGE_COUNTER_MAX;
4668 memcg->soft_limit = PAGE_COUNTER_MAX;
4669 memcg_wb_domain_size_changed(memcg);
4673 /* Handlers for move charge at task migration. */
4674 static int mem_cgroup_do_precharge(unsigned long count)
4678 /* Try a single bulk charge without reclaim first, kswapd may wake */
4679 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
4681 mc.precharge += count;
4685 /* Try charges one by one with reclaim, but do not retry */
4687 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
4701 enum mc_target_type {
4708 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4709 unsigned long addr, pte_t ptent)
4711 struct page *page = _vm_normal_page(vma, addr, ptent, true);
4713 if (!page || !page_mapped(page))
4715 if (PageAnon(page)) {
4716 if (!(mc.flags & MOVE_ANON))
4719 if (!(mc.flags & MOVE_FILE))
4722 if (!get_page_unless_zero(page))
4728 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
4729 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4730 pte_t ptent, swp_entry_t *entry)
4732 struct page *page = NULL;
4733 swp_entry_t ent = pte_to_swp_entry(ptent);
4735 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
4739 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
4740 * a device and because they are not accessible by CPU they are store
4741 * as special swap entry in the CPU page table.
4743 if (is_device_private_entry(ent)) {
4744 page = device_private_entry_to_page(ent);
4746 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
4747 * a refcount of 1 when free (unlike normal page)
4749 if (!page_ref_add_unless(page, 1, 1))
4755 * Because lookup_swap_cache() updates some statistics counter,
4756 * we call find_get_page() with swapper_space directly.
4758 page = find_get_page(swap_address_space(ent), swp_offset(ent));
4759 if (do_memsw_account())
4760 entry->val = ent.val;
4765 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4766 pte_t ptent, swp_entry_t *entry)
4772 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4773 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4775 struct page *page = NULL;
4776 struct address_space *mapping;
4779 if (!vma->vm_file) /* anonymous vma */
4781 if (!(mc.flags & MOVE_FILE))
4784 mapping = vma->vm_file->f_mapping;
4785 pgoff = linear_page_index(vma, addr);
4787 /* page is moved even if it's not RSS of this task(page-faulted). */
4789 /* shmem/tmpfs may report page out on swap: account for that too. */
4790 if (shmem_mapping(mapping)) {
4791 page = find_get_entry(mapping, pgoff);
4792 if (xa_is_value(page)) {
4793 swp_entry_t swp = radix_to_swp_entry(page);
4794 if (do_memsw_account())
4796 page = find_get_page(swap_address_space(swp),
4800 page = find_get_page(mapping, pgoff);
4802 page = find_get_page(mapping, pgoff);
4808 * mem_cgroup_move_account - move account of the page
4810 * @compound: charge the page as compound or small page
4811 * @from: mem_cgroup which the page is moved from.
4812 * @to: mem_cgroup which the page is moved to. @from != @to.
4814 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
4816 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4819 static int mem_cgroup_move_account(struct page *page,
4821 struct mem_cgroup *from,
4822 struct mem_cgroup *to)
4824 unsigned long flags;
4825 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
4829 VM_BUG_ON(from == to);
4830 VM_BUG_ON_PAGE(PageLRU(page), page);
4831 VM_BUG_ON(compound && !PageTransHuge(page));
4834 * Prevent mem_cgroup_migrate() from looking at
4835 * page->mem_cgroup of its source page while we change it.
4838 if (!trylock_page(page))
4842 if (page->mem_cgroup != from)
4845 anon = PageAnon(page);
4847 spin_lock_irqsave(&from->move_lock, flags);
4849 if (!anon && page_mapped(page)) {
4850 __mod_memcg_state(from, NR_FILE_MAPPED, -nr_pages);
4851 __mod_memcg_state(to, NR_FILE_MAPPED, nr_pages);
4855 * move_lock grabbed above and caller set from->moving_account, so
4856 * mod_memcg_page_state will serialize updates to PageDirty.
4857 * So mapping should be stable for dirty pages.
4859 if (!anon && PageDirty(page)) {
4860 struct address_space *mapping = page_mapping(page);
4862 if (mapping_cap_account_dirty(mapping)) {
4863 __mod_memcg_state(from, NR_FILE_DIRTY, -nr_pages);
4864 __mod_memcg_state(to, NR_FILE_DIRTY, nr_pages);
4868 if (PageWriteback(page)) {
4869 __mod_memcg_state(from, NR_WRITEBACK, -nr_pages);
4870 __mod_memcg_state(to, NR_WRITEBACK, nr_pages);
4874 * It is safe to change page->mem_cgroup here because the page
4875 * is referenced, charged, and isolated - we can't race with
4876 * uncharging, charging, migration, or LRU putback.
4879 /* caller should have done css_get */
4880 page->mem_cgroup = to;
4881 spin_unlock_irqrestore(&from->move_lock, flags);
4885 local_irq_disable();
4886 mem_cgroup_charge_statistics(to, page, compound, nr_pages);
4887 memcg_check_events(to, page);
4888 mem_cgroup_charge_statistics(from, page, compound, -nr_pages);
4889 memcg_check_events(from, page);
4898 * get_mctgt_type - get target type of moving charge
4899 * @vma: the vma the pte to be checked belongs
4900 * @addr: the address corresponding to the pte to be checked
4901 * @ptent: the pte to be checked
4902 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4905 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4906 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4907 * move charge. if @target is not NULL, the page is stored in target->page
4908 * with extra refcnt got(Callers should handle it).
4909 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4910 * target for charge migration. if @target is not NULL, the entry is stored
4912 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PUBLIC
4913 * or MEMORY_DEVICE_PRIVATE (so ZONE_DEVICE page and thus not on the lru).
4914 * For now we such page is charge like a regular page would be as for all
4915 * intent and purposes it is just special memory taking the place of a
4918 * See Documentations/vm/hmm.txt and include/linux/hmm.h
4920 * Called with pte lock held.
4923 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
4924 unsigned long addr, pte_t ptent, union mc_target *target)
4926 struct page *page = NULL;
4927 enum mc_target_type ret = MC_TARGET_NONE;
4928 swp_entry_t ent = { .val = 0 };
4930 if (pte_present(ptent))
4931 page = mc_handle_present_pte(vma, addr, ptent);
4932 else if (is_swap_pte(ptent))
4933 page = mc_handle_swap_pte(vma, ptent, &ent);
4934 else if (pte_none(ptent))
4935 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4937 if (!page && !ent.val)
4941 * Do only loose check w/o serialization.
4942 * mem_cgroup_move_account() checks the page is valid or
4943 * not under LRU exclusion.
4945 if (page->mem_cgroup == mc.from) {
4946 ret = MC_TARGET_PAGE;
4947 if (is_device_private_page(page) ||
4948 is_device_public_page(page))
4949 ret = MC_TARGET_DEVICE;
4951 target->page = page;
4953 if (!ret || !target)
4957 * There is a swap entry and a page doesn't exist or isn't charged.
4958 * But we cannot move a tail-page in a THP.
4960 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
4961 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
4962 ret = MC_TARGET_SWAP;
4969 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4971 * We don't consider PMD mapped swapping or file mapped pages because THP does
4972 * not support them for now.
4973 * Caller should make sure that pmd_trans_huge(pmd) is true.
4975 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4976 unsigned long addr, pmd_t pmd, union mc_target *target)
4978 struct page *page = NULL;
4979 enum mc_target_type ret = MC_TARGET_NONE;
4981 if (unlikely(is_swap_pmd(pmd))) {
4982 VM_BUG_ON(thp_migration_supported() &&
4983 !is_pmd_migration_entry(pmd));
4986 page = pmd_page(pmd);
4987 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
4988 if (!(mc.flags & MOVE_ANON))
4990 if (page->mem_cgroup == mc.from) {
4991 ret = MC_TARGET_PAGE;
4994 target->page = page;
5000 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5001 unsigned long addr, pmd_t pmd, union mc_target *target)
5003 return MC_TARGET_NONE;
5007 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5008 unsigned long addr, unsigned long end,
5009 struct mm_walk *walk)
5011 struct vm_area_struct *vma = walk->vma;
5015 ptl = pmd_trans_huge_lock(pmd, vma);
5018 * Note their can not be MC_TARGET_DEVICE for now as we do not
5019 * support transparent huge page with MEMORY_DEVICE_PUBLIC or
5020 * MEMORY_DEVICE_PRIVATE but this might change.
5022 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5023 mc.precharge += HPAGE_PMD_NR;
5028 if (pmd_trans_unstable(pmd))
5030 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5031 for (; addr != end; pte++, addr += PAGE_SIZE)
5032 if (get_mctgt_type(vma, addr, *pte, NULL))
5033 mc.precharge++; /* increment precharge temporarily */
5034 pte_unmap_unlock(pte - 1, ptl);
5040 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5042 unsigned long precharge;
5044 struct mm_walk mem_cgroup_count_precharge_walk = {
5045 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5048 down_read(&mm->mmap_sem);
5049 walk_page_range(0, mm->highest_vm_end,
5050 &mem_cgroup_count_precharge_walk);
5051 up_read(&mm->mmap_sem);
5053 precharge = mc.precharge;
5059 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5061 unsigned long precharge = mem_cgroup_count_precharge(mm);
5063 VM_BUG_ON(mc.moving_task);
5064 mc.moving_task = current;
5065 return mem_cgroup_do_precharge(precharge);
5068 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5069 static void __mem_cgroup_clear_mc(void)
5071 struct mem_cgroup *from = mc.from;
5072 struct mem_cgroup *to = mc.to;
5074 /* we must uncharge all the leftover precharges from mc.to */
5076 cancel_charge(mc.to, mc.precharge);
5080 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5081 * we must uncharge here.
5083 if (mc.moved_charge) {
5084 cancel_charge(mc.from, mc.moved_charge);
5085 mc.moved_charge = 0;
5087 /* we must fixup refcnts and charges */
5088 if (mc.moved_swap) {
5089 /* uncharge swap account from the old cgroup */
5090 if (!mem_cgroup_is_root(mc.from))
5091 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5093 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5096 * we charged both to->memory and to->memsw, so we
5097 * should uncharge to->memory.
5099 if (!mem_cgroup_is_root(mc.to))
5100 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5102 mem_cgroup_id_get_many(mc.to, mc.moved_swap);
5103 css_put_many(&mc.to->css, mc.moved_swap);
5107 memcg_oom_recover(from);
5108 memcg_oom_recover(to);
5109 wake_up_all(&mc.waitq);
5112 static void mem_cgroup_clear_mc(void)
5114 struct mm_struct *mm = mc.mm;
5117 * we must clear moving_task before waking up waiters at the end of
5120 mc.moving_task = NULL;
5121 __mem_cgroup_clear_mc();
5122 spin_lock(&mc.lock);
5126 spin_unlock(&mc.lock);
5131 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5133 struct cgroup_subsys_state *css;
5134 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5135 struct mem_cgroup *from;
5136 struct task_struct *leader, *p;
5137 struct mm_struct *mm;
5138 unsigned long move_flags;
5141 /* charge immigration isn't supported on the default hierarchy */
5142 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5146 * Multi-process migrations only happen on the default hierarchy
5147 * where charge immigration is not used. Perform charge
5148 * immigration if @tset contains a leader and whine if there are
5152 cgroup_taskset_for_each_leader(leader, css, tset) {
5155 memcg = mem_cgroup_from_css(css);
5161 * We are now commited to this value whatever it is. Changes in this
5162 * tunable will only affect upcoming migrations, not the current one.
5163 * So we need to save it, and keep it going.
5165 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5169 from = mem_cgroup_from_task(p);
5171 VM_BUG_ON(from == memcg);
5173 mm = get_task_mm(p);
5176 /* We move charges only when we move a owner of the mm */
5177 if (mm->owner == p) {
5180 VM_BUG_ON(mc.precharge);
5181 VM_BUG_ON(mc.moved_charge);
5182 VM_BUG_ON(mc.moved_swap);
5184 spin_lock(&mc.lock);
5188 mc.flags = move_flags;
5189 spin_unlock(&mc.lock);
5190 /* We set mc.moving_task later */
5192 ret = mem_cgroup_precharge_mc(mm);
5194 mem_cgroup_clear_mc();
5201 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5204 mem_cgroup_clear_mc();
5207 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5208 unsigned long addr, unsigned long end,
5209 struct mm_walk *walk)
5212 struct vm_area_struct *vma = walk->vma;
5215 enum mc_target_type target_type;
5216 union mc_target target;
5219 ptl = pmd_trans_huge_lock(pmd, vma);
5221 if (mc.precharge < HPAGE_PMD_NR) {
5225 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5226 if (target_type == MC_TARGET_PAGE) {
5228 if (!isolate_lru_page(page)) {
5229 if (!mem_cgroup_move_account(page, true,
5231 mc.precharge -= HPAGE_PMD_NR;
5232 mc.moved_charge += HPAGE_PMD_NR;
5234 putback_lru_page(page);
5237 } else if (target_type == MC_TARGET_DEVICE) {
5239 if (!mem_cgroup_move_account(page, true,
5241 mc.precharge -= HPAGE_PMD_NR;
5242 mc.moved_charge += HPAGE_PMD_NR;
5250 if (pmd_trans_unstable(pmd))
5253 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5254 for (; addr != end; addr += PAGE_SIZE) {
5255 pte_t ptent = *(pte++);
5256 bool device = false;
5262 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5263 case MC_TARGET_DEVICE:
5266 case MC_TARGET_PAGE:
5269 * We can have a part of the split pmd here. Moving it
5270 * can be done but it would be too convoluted so simply
5271 * ignore such a partial THP and keep it in original
5272 * memcg. There should be somebody mapping the head.
5274 if (PageTransCompound(page))
5276 if (!device && isolate_lru_page(page))
5278 if (!mem_cgroup_move_account(page, false,
5281 /* we uncharge from mc.from later. */
5285 putback_lru_page(page);
5286 put: /* get_mctgt_type() gets the page */
5289 case MC_TARGET_SWAP:
5291 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5293 /* we fixup refcnts and charges later. */
5301 pte_unmap_unlock(pte - 1, ptl);
5306 * We have consumed all precharges we got in can_attach().
5307 * We try charge one by one, but don't do any additional
5308 * charges to mc.to if we have failed in charge once in attach()
5311 ret = mem_cgroup_do_precharge(1);
5319 static void mem_cgroup_move_charge(void)
5321 struct mm_walk mem_cgroup_move_charge_walk = {
5322 .pmd_entry = mem_cgroup_move_charge_pte_range,
5326 lru_add_drain_all();
5328 * Signal lock_page_memcg() to take the memcg's move_lock
5329 * while we're moving its pages to another memcg. Then wait
5330 * for already started RCU-only updates to finish.
5332 atomic_inc(&mc.from->moving_account);
5335 if (unlikely(!down_read_trylock(&mc.mm->mmap_sem))) {
5337 * Someone who are holding the mmap_sem might be waiting in
5338 * waitq. So we cancel all extra charges, wake up all waiters,
5339 * and retry. Because we cancel precharges, we might not be able
5340 * to move enough charges, but moving charge is a best-effort
5341 * feature anyway, so it wouldn't be a big problem.
5343 __mem_cgroup_clear_mc();
5348 * When we have consumed all precharges and failed in doing
5349 * additional charge, the page walk just aborts.
5351 walk_page_range(0, mc.mm->highest_vm_end, &mem_cgroup_move_charge_walk);
5353 up_read(&mc.mm->mmap_sem);
5354 atomic_dec(&mc.from->moving_account);
5357 static void mem_cgroup_move_task(void)
5360 mem_cgroup_move_charge();
5361 mem_cgroup_clear_mc();
5364 #else /* !CONFIG_MMU */
5365 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5369 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5372 static void mem_cgroup_move_task(void)
5378 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5379 * to verify whether we're attached to the default hierarchy on each mount
5382 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5385 * use_hierarchy is forced on the default hierarchy. cgroup core
5386 * guarantees that @root doesn't have any children, so turning it
5387 * on for the root memcg is enough.
5389 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5390 root_mem_cgroup->use_hierarchy = true;
5392 root_mem_cgroup->use_hierarchy = false;
5395 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
5397 if (value == PAGE_COUNTER_MAX)
5398 seq_puts(m, "max\n");
5400 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
5405 static u64 memory_current_read(struct cgroup_subsys_state *css,
5408 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5410 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
5413 static int memory_min_show(struct seq_file *m, void *v)
5415 return seq_puts_memcg_tunable(m,
5416 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
5419 static ssize_t memory_min_write(struct kernfs_open_file *of,
5420 char *buf, size_t nbytes, loff_t off)
5422 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5426 buf = strstrip(buf);
5427 err = page_counter_memparse(buf, "max", &min);
5431 page_counter_set_min(&memcg->memory, min);
5436 static int memory_low_show(struct seq_file *m, void *v)
5438 return seq_puts_memcg_tunable(m,
5439 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
5442 static ssize_t memory_low_write(struct kernfs_open_file *of,
5443 char *buf, size_t nbytes, loff_t off)
5445 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5449 buf = strstrip(buf);
5450 err = page_counter_memparse(buf, "max", &low);
5454 page_counter_set_low(&memcg->memory, low);
5459 static int memory_high_show(struct seq_file *m, void *v)
5461 return seq_puts_memcg_tunable(m, READ_ONCE(mem_cgroup_from_seq(m)->high));
5464 static ssize_t memory_high_write(struct kernfs_open_file *of,
5465 char *buf, size_t nbytes, loff_t off)
5467 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5468 unsigned long nr_pages;
5472 buf = strstrip(buf);
5473 err = page_counter_memparse(buf, "max", &high);
5479 nr_pages = page_counter_read(&memcg->memory);
5480 if (nr_pages > high)
5481 try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
5484 memcg_wb_domain_size_changed(memcg);
5488 static int memory_max_show(struct seq_file *m, void *v)
5490 return seq_puts_memcg_tunable(m,
5491 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
5494 static ssize_t memory_max_write(struct kernfs_open_file *of,
5495 char *buf, size_t nbytes, loff_t off)
5497 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5498 unsigned int nr_reclaims = MEM_CGROUP_RECLAIM_RETRIES;
5499 bool drained = false;
5503 buf = strstrip(buf);
5504 err = page_counter_memparse(buf, "max", &max);
5508 xchg(&memcg->memory.max, max);
5511 unsigned long nr_pages = page_counter_read(&memcg->memory);
5513 if (nr_pages <= max)
5516 if (signal_pending(current)) {
5522 drain_all_stock(memcg);
5528 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
5534 memcg_memory_event(memcg, MEMCG_OOM);
5535 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
5539 memcg_wb_domain_size_changed(memcg);
5543 static int memory_events_show(struct seq_file *m, void *v)
5545 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
5547 seq_printf(m, "low %lu\n",
5548 atomic_long_read(&memcg->memory_events[MEMCG_LOW]));
5549 seq_printf(m, "high %lu\n",
5550 atomic_long_read(&memcg->memory_events[MEMCG_HIGH]));
5551 seq_printf(m, "max %lu\n",
5552 atomic_long_read(&memcg->memory_events[MEMCG_MAX]));
5553 seq_printf(m, "oom %lu\n",
5554 atomic_long_read(&memcg->memory_events[MEMCG_OOM]));
5555 seq_printf(m, "oom_kill %lu\n",
5556 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
5561 static int memory_stat_show(struct seq_file *m, void *v)
5563 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
5564 struct accumulated_stats acc;
5568 * Provide statistics on the state of the memory subsystem as
5569 * well as cumulative event counters that show past behavior.
5571 * This list is ordered following a combination of these gradients:
5572 * 1) generic big picture -> specifics and details
5573 * 2) reflecting userspace activity -> reflecting kernel heuristics
5575 * Current memory state:
5578 memset(&acc, 0, sizeof(acc));
5579 acc.stats_size = MEMCG_NR_STAT;
5580 acc.events_size = NR_VM_EVENT_ITEMS;
5581 accumulate_memcg_tree(memcg, &acc);
5583 seq_printf(m, "anon %llu\n",
5584 (u64)acc.stat[MEMCG_RSS] * PAGE_SIZE);
5585 seq_printf(m, "file %llu\n",
5586 (u64)acc.stat[MEMCG_CACHE] * PAGE_SIZE);
5587 seq_printf(m, "kernel_stack %llu\n",
5588 (u64)acc.stat[MEMCG_KERNEL_STACK_KB] * 1024);
5589 seq_printf(m, "slab %llu\n",
5590 (u64)(acc.stat[NR_SLAB_RECLAIMABLE] +
5591 acc.stat[NR_SLAB_UNRECLAIMABLE]) * PAGE_SIZE);
5592 seq_printf(m, "sock %llu\n",
5593 (u64)acc.stat[MEMCG_SOCK] * PAGE_SIZE);
5595 seq_printf(m, "shmem %llu\n",
5596 (u64)acc.stat[NR_SHMEM] * PAGE_SIZE);
5597 seq_printf(m, "file_mapped %llu\n",
5598 (u64)acc.stat[NR_FILE_MAPPED] * PAGE_SIZE);
5599 seq_printf(m, "file_dirty %llu\n",
5600 (u64)acc.stat[NR_FILE_DIRTY] * PAGE_SIZE);
5601 seq_printf(m, "file_writeback %llu\n",
5602 (u64)acc.stat[NR_WRITEBACK] * PAGE_SIZE);
5605 * TODO: We should eventually replace our own MEMCG_RSS_HUGE counter
5606 * with the NR_ANON_THP vm counter, but right now it's a pain in the
5607 * arse because it requires migrating the work out of rmap to a place
5608 * where the page->mem_cgroup is set up and stable.
5610 seq_printf(m, "anon_thp %llu\n",
5611 (u64)acc.stat[MEMCG_RSS_HUGE] * PAGE_SIZE);
5613 for (i = 0; i < NR_LRU_LISTS; i++)
5614 seq_printf(m, "%s %llu\n", mem_cgroup_lru_names[i],
5615 (u64)acc.lru_pages[i] * PAGE_SIZE);
5617 seq_printf(m, "slab_reclaimable %llu\n",
5618 (u64)acc.stat[NR_SLAB_RECLAIMABLE] * PAGE_SIZE);
5619 seq_printf(m, "slab_unreclaimable %llu\n",
5620 (u64)acc.stat[NR_SLAB_UNRECLAIMABLE] * PAGE_SIZE);
5622 /* Accumulated memory events */
5624 seq_printf(m, "pgfault %lu\n", acc.events[PGFAULT]);
5625 seq_printf(m, "pgmajfault %lu\n", acc.events[PGMAJFAULT]);
5627 seq_printf(m, "workingset_refault %lu\n",
5628 acc.stat[WORKINGSET_REFAULT]);
5629 seq_printf(m, "workingset_activate %lu\n",
5630 acc.stat[WORKINGSET_ACTIVATE]);
5631 seq_printf(m, "workingset_nodereclaim %lu\n",
5632 acc.stat[WORKINGSET_NODERECLAIM]);
5634 seq_printf(m, "pgrefill %lu\n", acc.events[PGREFILL]);
5635 seq_printf(m, "pgscan %lu\n", acc.events[PGSCAN_KSWAPD] +
5636 acc.events[PGSCAN_DIRECT]);
5637 seq_printf(m, "pgsteal %lu\n", acc.events[PGSTEAL_KSWAPD] +
5638 acc.events[PGSTEAL_DIRECT]);
5639 seq_printf(m, "pgactivate %lu\n", acc.events[PGACTIVATE]);
5640 seq_printf(m, "pgdeactivate %lu\n", acc.events[PGDEACTIVATE]);
5641 seq_printf(m, "pglazyfree %lu\n", acc.events[PGLAZYFREE]);
5642 seq_printf(m, "pglazyfreed %lu\n", acc.events[PGLAZYFREED]);
5644 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5645 seq_printf(m, "thp_fault_alloc %lu\n", acc.events[THP_FAULT_ALLOC]);
5646 seq_printf(m, "thp_collapse_alloc %lu\n",
5647 acc.events[THP_COLLAPSE_ALLOC]);
5648 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
5653 static int memory_oom_group_show(struct seq_file *m, void *v)
5655 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
5657 seq_printf(m, "%d\n", memcg->oom_group);
5662 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
5663 char *buf, size_t nbytes, loff_t off)
5665 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5668 buf = strstrip(buf);
5672 ret = kstrtoint(buf, 0, &oom_group);
5676 if (oom_group != 0 && oom_group != 1)
5679 memcg->oom_group = oom_group;
5684 static struct cftype memory_files[] = {
5687 .flags = CFTYPE_NOT_ON_ROOT,
5688 .read_u64 = memory_current_read,
5692 .flags = CFTYPE_NOT_ON_ROOT,
5693 .seq_show = memory_min_show,
5694 .write = memory_min_write,
5698 .flags = CFTYPE_NOT_ON_ROOT,
5699 .seq_show = memory_low_show,
5700 .write = memory_low_write,
5704 .flags = CFTYPE_NOT_ON_ROOT,
5705 .seq_show = memory_high_show,
5706 .write = memory_high_write,
5710 .flags = CFTYPE_NOT_ON_ROOT,
5711 .seq_show = memory_max_show,
5712 .write = memory_max_write,
5716 .flags = CFTYPE_NOT_ON_ROOT,
5717 .file_offset = offsetof(struct mem_cgroup, events_file),
5718 .seq_show = memory_events_show,
5722 .flags = CFTYPE_NOT_ON_ROOT,
5723 .seq_show = memory_stat_show,
5726 .name = "oom.group",
5727 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
5728 .seq_show = memory_oom_group_show,
5729 .write = memory_oom_group_write,
5734 struct cgroup_subsys memory_cgrp_subsys = {
5735 .css_alloc = mem_cgroup_css_alloc,
5736 .css_online = mem_cgroup_css_online,
5737 .css_offline = mem_cgroup_css_offline,
5738 .css_released = mem_cgroup_css_released,
5739 .css_free = mem_cgroup_css_free,
5740 .css_reset = mem_cgroup_css_reset,
5741 .can_attach = mem_cgroup_can_attach,
5742 .cancel_attach = mem_cgroup_cancel_attach,
5743 .post_attach = mem_cgroup_move_task,
5744 .bind = mem_cgroup_bind,
5745 .dfl_cftypes = memory_files,
5746 .legacy_cftypes = mem_cgroup_legacy_files,
5751 * mem_cgroup_protected - check if memory consumption is in the normal range
5752 * @root: the top ancestor of the sub-tree being checked
5753 * @memcg: the memory cgroup to check
5755 * WARNING: This function is not stateless! It can only be used as part
5756 * of a top-down tree iteration, not for isolated queries.
5758 * Returns one of the following:
5759 * MEMCG_PROT_NONE: cgroup memory is not protected
5760 * MEMCG_PROT_LOW: cgroup memory is protected as long there is
5761 * an unprotected supply of reclaimable memory from other cgroups.
5762 * MEMCG_PROT_MIN: cgroup memory is protected
5764 * @root is exclusive; it is never protected when looked at directly
5766 * To provide a proper hierarchical behavior, effective memory.min/low values
5767 * are used. Below is the description of how effective memory.low is calculated.
5768 * Effective memory.min values is calculated in the same way.
5770 * Effective memory.low is always equal or less than the original memory.low.
5771 * If there is no memory.low overcommittment (which is always true for
5772 * top-level memory cgroups), these two values are equal.
5773 * Otherwise, it's a part of parent's effective memory.low,
5774 * calculated as a cgroup's memory.low usage divided by sum of sibling's
5775 * memory.low usages, where memory.low usage is the size of actually
5779 * elow = min( memory.low, parent->elow * ------------------ ),
5780 * siblings_low_usage
5782 * | memory.current, if memory.current < memory.low
5787 * Such definition of the effective memory.low provides the expected
5788 * hierarchical behavior: parent's memory.low value is limiting
5789 * children, unprotected memory is reclaimed first and cgroups,
5790 * which are not using their guarantee do not affect actual memory
5793 * For example, if there are memcgs A, A/B, A/C, A/D and A/E:
5795 * A A/memory.low = 2G, A/memory.current = 6G
5797 * BC DE B/memory.low = 3G B/memory.current = 2G
5798 * C/memory.low = 1G C/memory.current = 2G
5799 * D/memory.low = 0 D/memory.current = 2G
5800 * E/memory.low = 10G E/memory.current = 0
5802 * and the memory pressure is applied, the following memory distribution
5803 * is expected (approximately):
5805 * A/memory.current = 2G
5807 * B/memory.current = 1.3G
5808 * C/memory.current = 0.6G
5809 * D/memory.current = 0
5810 * E/memory.current = 0
5812 * These calculations require constant tracking of the actual low usages
5813 * (see propagate_protected_usage()), as well as recursive calculation of
5814 * effective memory.low values. But as we do call mem_cgroup_protected()
5815 * path for each memory cgroup top-down from the reclaim,
5816 * it's possible to optimize this part, and save calculated elow
5817 * for next usage. This part is intentionally racy, but it's ok,
5818 * as memory.low is a best-effort mechanism.
5820 enum mem_cgroup_protection mem_cgroup_protected(struct mem_cgroup *root,
5821 struct mem_cgroup *memcg)
5823 struct mem_cgroup *parent;
5824 unsigned long emin, parent_emin;
5825 unsigned long elow, parent_elow;
5826 unsigned long usage;
5828 if (mem_cgroup_disabled())
5829 return MEMCG_PROT_NONE;
5832 root = root_mem_cgroup;
5834 return MEMCG_PROT_NONE;
5836 usage = page_counter_read(&memcg->memory);
5838 return MEMCG_PROT_NONE;
5840 emin = memcg->memory.min;
5841 elow = memcg->memory.low;
5843 parent = parent_mem_cgroup(memcg);
5844 /* No parent means a non-hierarchical mode on v1 memcg */
5846 return MEMCG_PROT_NONE;
5851 parent_emin = READ_ONCE(parent->memory.emin);
5852 emin = min(emin, parent_emin);
5853 if (emin && parent_emin) {
5854 unsigned long min_usage, siblings_min_usage;
5856 min_usage = min(usage, memcg->memory.min);
5857 siblings_min_usage = atomic_long_read(
5858 &parent->memory.children_min_usage);
5860 if (min_usage && siblings_min_usage)
5861 emin = min(emin, parent_emin * min_usage /
5862 siblings_min_usage);
5865 parent_elow = READ_ONCE(parent->memory.elow);
5866 elow = min(elow, parent_elow);
5867 if (elow && parent_elow) {
5868 unsigned long low_usage, siblings_low_usage;
5870 low_usage = min(usage, memcg->memory.low);
5871 siblings_low_usage = atomic_long_read(
5872 &parent->memory.children_low_usage);
5874 if (low_usage && siblings_low_usage)
5875 elow = min(elow, parent_elow * low_usage /
5876 siblings_low_usage);
5880 memcg->memory.emin = emin;
5881 memcg->memory.elow = elow;
5884 return MEMCG_PROT_MIN;
5885 else if (usage <= elow)
5886 return MEMCG_PROT_LOW;
5888 return MEMCG_PROT_NONE;
5892 * mem_cgroup_try_charge - try charging a page
5893 * @page: page to charge
5894 * @mm: mm context of the victim
5895 * @gfp_mask: reclaim mode
5896 * @memcgp: charged memcg return
5897 * @compound: charge the page as compound or small page
5899 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5900 * pages according to @gfp_mask if necessary.
5902 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5903 * Otherwise, an error code is returned.
5905 * After page->mapping has been set up, the caller must finalize the
5906 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5907 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5909 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
5910 gfp_t gfp_mask, struct mem_cgroup **memcgp,
5913 struct mem_cgroup *memcg = NULL;
5914 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5917 if (mem_cgroup_disabled())
5920 if (PageSwapCache(page)) {
5922 * Every swap fault against a single page tries to charge the
5923 * page, bail as early as possible. shmem_unuse() encounters
5924 * already charged pages, too. The USED bit is protected by
5925 * the page lock, which serializes swap cache removal, which
5926 * in turn serializes uncharging.
5928 VM_BUG_ON_PAGE(!PageLocked(page), page);
5929 if (compound_head(page)->mem_cgroup)
5932 if (do_swap_account) {
5933 swp_entry_t ent = { .val = page_private(page), };
5934 unsigned short id = lookup_swap_cgroup_id(ent);
5937 memcg = mem_cgroup_from_id(id);
5938 if (memcg && !css_tryget_online(&memcg->css))
5945 memcg = get_mem_cgroup_from_mm(mm);
5947 ret = try_charge(memcg, gfp_mask, nr_pages);
5949 css_put(&memcg->css);
5955 int mem_cgroup_try_charge_delay(struct page *page, struct mm_struct *mm,
5956 gfp_t gfp_mask, struct mem_cgroup **memcgp,
5959 struct mem_cgroup *memcg;
5962 ret = mem_cgroup_try_charge(page, mm, gfp_mask, memcgp, compound);
5964 mem_cgroup_throttle_swaprate(memcg, page_to_nid(page), gfp_mask);
5969 * mem_cgroup_commit_charge - commit a page charge
5970 * @page: page to charge
5971 * @memcg: memcg to charge the page to
5972 * @lrucare: page might be on LRU already
5973 * @compound: charge the page as compound or small page
5975 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5976 * after page->mapping has been set up. This must happen atomically
5977 * as part of the page instantiation, i.e. under the page table lock
5978 * for anonymous pages, under the page lock for page and swap cache.
5980 * In addition, the page must not be on the LRU during the commit, to
5981 * prevent racing with task migration. If it might be, use @lrucare.
5983 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5985 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
5986 bool lrucare, bool compound)
5988 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5990 VM_BUG_ON_PAGE(!page->mapping, page);
5991 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
5993 if (mem_cgroup_disabled())
5996 * Swap faults will attempt to charge the same page multiple
5997 * times. But reuse_swap_page() might have removed the page
5998 * from swapcache already, so we can't check PageSwapCache().
6003 commit_charge(page, memcg, lrucare);
6005 local_irq_disable();
6006 mem_cgroup_charge_statistics(memcg, page, compound, nr_pages);
6007 memcg_check_events(memcg, page);
6010 if (do_memsw_account() && PageSwapCache(page)) {
6011 swp_entry_t entry = { .val = page_private(page) };
6013 * The swap entry might not get freed for a long time,
6014 * let's not wait for it. The page already received a
6015 * memory+swap charge, drop the swap entry duplicate.
6017 mem_cgroup_uncharge_swap(entry, nr_pages);
6022 * mem_cgroup_cancel_charge - cancel a page charge
6023 * @page: page to charge
6024 * @memcg: memcg to charge the page to
6025 * @compound: charge the page as compound or small page
6027 * Cancel a charge transaction started by mem_cgroup_try_charge().
6029 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg,
6032 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6034 if (mem_cgroup_disabled())
6037 * Swap faults will attempt to charge the same page multiple
6038 * times. But reuse_swap_page() might have removed the page
6039 * from swapcache already, so we can't check PageSwapCache().
6044 cancel_charge(memcg, nr_pages);
6047 struct uncharge_gather {
6048 struct mem_cgroup *memcg;
6049 unsigned long pgpgout;
6050 unsigned long nr_anon;
6051 unsigned long nr_file;
6052 unsigned long nr_kmem;
6053 unsigned long nr_huge;
6054 unsigned long nr_shmem;
6055 struct page *dummy_page;
6058 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6060 memset(ug, 0, sizeof(*ug));
6063 static void uncharge_batch(const struct uncharge_gather *ug)
6065 unsigned long nr_pages = ug->nr_anon + ug->nr_file + ug->nr_kmem;
6066 unsigned long flags;
6068 if (!mem_cgroup_is_root(ug->memcg)) {
6069 page_counter_uncharge(&ug->memcg->memory, nr_pages);
6070 if (do_memsw_account())
6071 page_counter_uncharge(&ug->memcg->memsw, nr_pages);
6072 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6073 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6074 memcg_oom_recover(ug->memcg);
6077 local_irq_save(flags);
6078 __mod_memcg_state(ug->memcg, MEMCG_RSS, -ug->nr_anon);
6079 __mod_memcg_state(ug->memcg, MEMCG_CACHE, -ug->nr_file);
6080 __mod_memcg_state(ug->memcg, MEMCG_RSS_HUGE, -ug->nr_huge);
6081 __mod_memcg_state(ug->memcg, NR_SHMEM, -ug->nr_shmem);
6082 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6083 __this_cpu_add(ug->memcg->stat_cpu->nr_page_events, nr_pages);
6084 memcg_check_events(ug->memcg, ug->dummy_page);
6085 local_irq_restore(flags);
6087 if (!mem_cgroup_is_root(ug->memcg))
6088 css_put_many(&ug->memcg->css, nr_pages);
6091 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6093 VM_BUG_ON_PAGE(PageLRU(page), page);
6094 VM_BUG_ON_PAGE(page_count(page) && !is_zone_device_page(page) &&
6095 !PageHWPoison(page) , page);
6097 if (!page->mem_cgroup)
6101 * Nobody should be changing or seriously looking at
6102 * page->mem_cgroup at this point, we have fully
6103 * exclusive access to the page.
6106 if (ug->memcg != page->mem_cgroup) {
6109 uncharge_gather_clear(ug);
6111 ug->memcg = page->mem_cgroup;
6114 if (!PageKmemcg(page)) {
6115 unsigned int nr_pages = 1;
6117 if (PageTransHuge(page)) {
6118 nr_pages <<= compound_order(page);
6119 ug->nr_huge += nr_pages;
6122 ug->nr_anon += nr_pages;
6124 ug->nr_file += nr_pages;
6125 if (PageSwapBacked(page))
6126 ug->nr_shmem += nr_pages;
6130 ug->nr_kmem += 1 << compound_order(page);
6131 __ClearPageKmemcg(page);
6134 ug->dummy_page = page;
6135 page->mem_cgroup = NULL;
6138 static void uncharge_list(struct list_head *page_list)
6140 struct uncharge_gather ug;
6141 struct list_head *next;
6143 uncharge_gather_clear(&ug);
6146 * Note that the list can be a single page->lru; hence the
6147 * do-while loop instead of a simple list_for_each_entry().
6149 next = page_list->next;
6153 page = list_entry(next, struct page, lru);
6154 next = page->lru.next;
6156 uncharge_page(page, &ug);
6157 } while (next != page_list);
6160 uncharge_batch(&ug);
6164 * mem_cgroup_uncharge - uncharge a page
6165 * @page: page to uncharge
6167 * Uncharge a page previously charged with mem_cgroup_try_charge() and
6168 * mem_cgroup_commit_charge().
6170 void mem_cgroup_uncharge(struct page *page)
6172 struct uncharge_gather ug;
6174 if (mem_cgroup_disabled())
6177 /* Don't touch page->lru of any random page, pre-check: */
6178 if (!page->mem_cgroup)
6181 uncharge_gather_clear(&ug);
6182 uncharge_page(page, &ug);
6183 uncharge_batch(&ug);
6187 * mem_cgroup_uncharge_list - uncharge a list of page
6188 * @page_list: list of pages to uncharge
6190 * Uncharge a list of pages previously charged with
6191 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
6193 void mem_cgroup_uncharge_list(struct list_head *page_list)
6195 if (mem_cgroup_disabled())
6198 if (!list_empty(page_list))
6199 uncharge_list(page_list);
6203 * mem_cgroup_migrate - charge a page's replacement
6204 * @oldpage: currently circulating page
6205 * @newpage: replacement page
6207 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6208 * be uncharged upon free.
6210 * Both pages must be locked, @newpage->mapping must be set up.
6212 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6214 struct mem_cgroup *memcg;
6215 unsigned int nr_pages;
6217 unsigned long flags;
6219 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6220 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6221 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6222 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6225 if (mem_cgroup_disabled())
6228 /* Page cache replacement: new page already charged? */
6229 if (newpage->mem_cgroup)
6232 /* Swapcache readahead pages can get replaced before being charged */
6233 memcg = oldpage->mem_cgroup;
6237 /* Force-charge the new page. The old one will be freed soon */
6238 compound = PageTransHuge(newpage);
6239 nr_pages = compound ? hpage_nr_pages(newpage) : 1;
6241 page_counter_charge(&memcg->memory, nr_pages);
6242 if (do_memsw_account())
6243 page_counter_charge(&memcg->memsw, nr_pages);
6244 css_get_many(&memcg->css, nr_pages);
6246 commit_charge(newpage, memcg, false);
6248 local_irq_save(flags);
6249 mem_cgroup_charge_statistics(memcg, newpage, compound, nr_pages);
6250 memcg_check_events(memcg, newpage);
6251 local_irq_restore(flags);
6254 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
6255 EXPORT_SYMBOL(memcg_sockets_enabled_key);
6257 void mem_cgroup_sk_alloc(struct sock *sk)
6259 struct mem_cgroup *memcg;
6261 if (!mem_cgroup_sockets_enabled)
6265 * Socket cloning can throw us here with sk_memcg already
6266 * filled. It won't however, necessarily happen from
6267 * process context. So the test for root memcg given
6268 * the current task's memcg won't help us in this case.
6270 * Respecting the original socket's memcg is a better
6271 * decision in this case.
6274 css_get(&sk->sk_memcg->css);
6279 memcg = mem_cgroup_from_task(current);
6280 if (memcg == root_mem_cgroup)
6282 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
6284 if (css_tryget_online(&memcg->css))
6285 sk->sk_memcg = memcg;
6290 void mem_cgroup_sk_free(struct sock *sk)
6293 css_put(&sk->sk_memcg->css);
6297 * mem_cgroup_charge_skmem - charge socket memory
6298 * @memcg: memcg to charge
6299 * @nr_pages: number of pages to charge
6301 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
6302 * @memcg's configured limit, %false if the charge had to be forced.
6304 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6306 gfp_t gfp_mask = GFP_KERNEL;
6308 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6309 struct page_counter *fail;
6311 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
6312 memcg->tcpmem_pressure = 0;
6315 page_counter_charge(&memcg->tcpmem, nr_pages);
6316 memcg->tcpmem_pressure = 1;
6320 /* Don't block in the packet receive path */
6322 gfp_mask = GFP_NOWAIT;
6324 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
6326 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
6329 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
6334 * mem_cgroup_uncharge_skmem - uncharge socket memory
6335 * @memcg: memcg to uncharge
6336 * @nr_pages: number of pages to uncharge
6338 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6340 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6341 page_counter_uncharge(&memcg->tcpmem, nr_pages);
6345 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
6347 refill_stock(memcg, nr_pages);
6350 static int __init cgroup_memory(char *s)
6354 while ((token = strsep(&s, ",")) != NULL) {
6357 if (!strcmp(token, "nosocket"))
6358 cgroup_memory_nosocket = true;
6359 if (!strcmp(token, "nokmem"))
6360 cgroup_memory_nokmem = true;
6364 __setup("cgroup.memory=", cgroup_memory);
6367 * subsys_initcall() for memory controller.
6369 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
6370 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
6371 * basically everything that doesn't depend on a specific mem_cgroup structure
6372 * should be initialized from here.
6374 static int __init mem_cgroup_init(void)
6378 #ifdef CONFIG_MEMCG_KMEM
6380 * Kmem cache creation is mostly done with the slab_mutex held,
6381 * so use a workqueue with limited concurrency to avoid stalling
6382 * all worker threads in case lots of cgroups are created and
6383 * destroyed simultaneously.
6385 memcg_kmem_cache_wq = alloc_workqueue("memcg_kmem_cache", 0, 1);
6386 BUG_ON(!memcg_kmem_cache_wq);
6389 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
6390 memcg_hotplug_cpu_dead);
6392 for_each_possible_cpu(cpu)
6393 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
6396 for_each_node(node) {
6397 struct mem_cgroup_tree_per_node *rtpn;
6399 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
6400 node_online(node) ? node : NUMA_NO_NODE);
6402 rtpn->rb_root = RB_ROOT;
6403 rtpn->rb_rightmost = NULL;
6404 spin_lock_init(&rtpn->lock);
6405 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6410 subsys_initcall(mem_cgroup_init);
6412 #ifdef CONFIG_MEMCG_SWAP
6413 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
6415 while (!refcount_inc_not_zero(&memcg->id.ref)) {
6417 * The root cgroup cannot be destroyed, so it's refcount must
6420 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
6424 memcg = parent_mem_cgroup(memcg);
6426 memcg = root_mem_cgroup;
6432 * mem_cgroup_swapout - transfer a memsw charge to swap
6433 * @page: page whose memsw charge to transfer
6434 * @entry: swap entry to move the charge to
6436 * Transfer the memsw charge of @page to @entry.
6438 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
6440 struct mem_cgroup *memcg, *swap_memcg;
6441 unsigned int nr_entries;
6442 unsigned short oldid;
6444 VM_BUG_ON_PAGE(PageLRU(page), page);
6445 VM_BUG_ON_PAGE(page_count(page), page);
6447 if (!do_memsw_account())
6450 memcg = page->mem_cgroup;
6452 /* Readahead page, never charged */
6457 * In case the memcg owning these pages has been offlined and doesn't
6458 * have an ID allocated to it anymore, charge the closest online
6459 * ancestor for the swap instead and transfer the memory+swap charge.
6461 swap_memcg = mem_cgroup_id_get_online(memcg);
6462 nr_entries = hpage_nr_pages(page);
6463 /* Get references for the tail pages, too */
6465 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
6466 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
6468 VM_BUG_ON_PAGE(oldid, page);
6469 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
6471 page->mem_cgroup = NULL;
6473 if (!mem_cgroup_is_root(memcg))
6474 page_counter_uncharge(&memcg->memory, nr_entries);
6476 if (memcg != swap_memcg) {
6477 if (!mem_cgroup_is_root(swap_memcg))
6478 page_counter_charge(&swap_memcg->memsw, nr_entries);
6479 page_counter_uncharge(&memcg->memsw, nr_entries);
6483 * Interrupts should be disabled here because the caller holds the
6484 * i_pages lock which is taken with interrupts-off. It is
6485 * important here to have the interrupts disabled because it is the
6486 * only synchronisation we have for updating the per-CPU variables.
6488 VM_BUG_ON(!irqs_disabled());
6489 mem_cgroup_charge_statistics(memcg, page, PageTransHuge(page),
6491 memcg_check_events(memcg, page);
6493 if (!mem_cgroup_is_root(memcg))
6494 css_put_many(&memcg->css, nr_entries);
6498 * mem_cgroup_try_charge_swap - try charging swap space for a page
6499 * @page: page being added to swap
6500 * @entry: swap entry to charge
6502 * Try to charge @page's memcg for the swap space at @entry.
6504 * Returns 0 on success, -ENOMEM on failure.
6506 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
6508 unsigned int nr_pages = hpage_nr_pages(page);
6509 struct page_counter *counter;
6510 struct mem_cgroup *memcg;
6511 unsigned short oldid;
6513 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) || !do_swap_account)
6516 memcg = page->mem_cgroup;
6518 /* Readahead page, never charged */
6523 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
6527 memcg = mem_cgroup_id_get_online(memcg);
6529 if (!mem_cgroup_is_root(memcg) &&
6530 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
6531 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
6532 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
6533 mem_cgroup_id_put(memcg);
6537 /* Get references for the tail pages, too */
6539 mem_cgroup_id_get_many(memcg, nr_pages - 1);
6540 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
6541 VM_BUG_ON_PAGE(oldid, page);
6542 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
6548 * mem_cgroup_uncharge_swap - uncharge swap space
6549 * @entry: swap entry to uncharge
6550 * @nr_pages: the amount of swap space to uncharge
6552 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
6554 struct mem_cgroup *memcg;
6557 if (!do_swap_account)
6560 id = swap_cgroup_record(entry, 0, nr_pages);
6562 memcg = mem_cgroup_from_id(id);
6564 if (!mem_cgroup_is_root(memcg)) {
6565 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
6566 page_counter_uncharge(&memcg->swap, nr_pages);
6568 page_counter_uncharge(&memcg->memsw, nr_pages);
6570 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
6571 mem_cgroup_id_put_many(memcg, nr_pages);
6576 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
6578 long nr_swap_pages = get_nr_swap_pages();
6580 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
6581 return nr_swap_pages;
6582 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
6583 nr_swap_pages = min_t(long, nr_swap_pages,
6584 READ_ONCE(memcg->swap.max) -
6585 page_counter_read(&memcg->swap));
6586 return nr_swap_pages;
6589 bool mem_cgroup_swap_full(struct page *page)
6591 struct mem_cgroup *memcg;
6593 VM_BUG_ON_PAGE(!PageLocked(page), page);
6597 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
6600 memcg = page->mem_cgroup;
6604 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
6605 if (page_counter_read(&memcg->swap) * 2 >= memcg->swap.max)
6611 /* for remember boot option*/
6612 #ifdef CONFIG_MEMCG_SWAP_ENABLED
6613 static int really_do_swap_account __initdata = 1;
6615 static int really_do_swap_account __initdata;
6618 static int __init enable_swap_account(char *s)
6620 if (!strcmp(s, "1"))
6621 really_do_swap_account = 1;
6622 else if (!strcmp(s, "0"))
6623 really_do_swap_account = 0;
6626 __setup("swapaccount=", enable_swap_account);
6628 static u64 swap_current_read(struct cgroup_subsys_state *css,
6631 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6633 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
6636 static int swap_max_show(struct seq_file *m, void *v)
6638 return seq_puts_memcg_tunable(m,
6639 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
6642 static ssize_t swap_max_write(struct kernfs_open_file *of,
6643 char *buf, size_t nbytes, loff_t off)
6645 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6649 buf = strstrip(buf);
6650 err = page_counter_memparse(buf, "max", &max);
6654 xchg(&memcg->swap.max, max);
6659 static int swap_events_show(struct seq_file *m, void *v)
6661 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6663 seq_printf(m, "max %lu\n",
6664 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
6665 seq_printf(m, "fail %lu\n",
6666 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
6671 static struct cftype swap_files[] = {
6673 .name = "swap.current",
6674 .flags = CFTYPE_NOT_ON_ROOT,
6675 .read_u64 = swap_current_read,
6679 .flags = CFTYPE_NOT_ON_ROOT,
6680 .seq_show = swap_max_show,
6681 .write = swap_max_write,
6684 .name = "swap.events",
6685 .flags = CFTYPE_NOT_ON_ROOT,
6686 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
6687 .seq_show = swap_events_show,
6692 static struct cftype memsw_cgroup_files[] = {
6694 .name = "memsw.usage_in_bytes",
6695 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6696 .read_u64 = mem_cgroup_read_u64,
6699 .name = "memsw.max_usage_in_bytes",
6700 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6701 .write = mem_cgroup_reset,
6702 .read_u64 = mem_cgroup_read_u64,
6705 .name = "memsw.limit_in_bytes",
6706 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6707 .write = mem_cgroup_write,
6708 .read_u64 = mem_cgroup_read_u64,
6711 .name = "memsw.failcnt",
6712 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6713 .write = mem_cgroup_reset,
6714 .read_u64 = mem_cgroup_read_u64,
6716 { }, /* terminate */
6719 static int __init mem_cgroup_swap_init(void)
6721 if (!mem_cgroup_disabled() && really_do_swap_account) {
6722 do_swap_account = 1;
6723 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys,
6725 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
6726 memsw_cgroup_files));
6730 subsys_initcall(mem_cgroup_swap_init);
6732 #endif /* CONFIG_MEMCG_SWAP */