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
17 * This program is free software; you can redistribute it and/or modify
18 * it under the terms of the GNU General Public License as published by
19 * the Free Software Foundation; either version 2 of the License, or
20 * (at your option) any later version.
22 * This program is distributed in the hope that it will be useful,
23 * but WITHOUT ANY WARRANTY; without even the implied warranty of
24 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
25 * GNU General Public License for more details.
28 #include <linux/page_counter.h>
29 #include <linux/memcontrol.h>
30 #include <linux/cgroup.h>
32 #include <linux/hugetlb.h>
33 #include <linux/pagemap.h>
34 #include <linux/smp.h>
35 #include <linux/page-flags.h>
36 #include <linux/backing-dev.h>
37 #include <linux/bit_spinlock.h>
38 #include <linux/rcupdate.h>
39 #include <linux/limits.h>
40 #include <linux/export.h>
41 #include <linux/mutex.h>
42 #include <linux/rbtree.h>
43 #include <linux/slab.h>
44 #include <linux/swap.h>
45 #include <linux/swapops.h>
46 #include <linux/spinlock.h>
47 #include <linux/eventfd.h>
48 #include <linux/poll.h>
49 #include <linux/sort.h>
51 #include <linux/seq_file.h>
52 #include <linux/vmpressure.h>
53 #include <linux/mm_inline.h>
54 #include <linux/page_cgroup.h>
55 #include <linux/cpu.h>
56 #include <linux/oom.h>
57 #include <linux/lockdep.h>
58 #include <linux/file.h>
62 #include <net/tcp_memcontrol.h>
65 #include <asm/uaccess.h>
67 #include <trace/events/vmscan.h>
69 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
70 EXPORT_SYMBOL(memory_cgrp_subsys);
72 #define MEM_CGROUP_RECLAIM_RETRIES 5
73 static struct mem_cgroup *root_mem_cgroup __read_mostly;
75 #ifdef CONFIG_MEMCG_SWAP
76 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
77 int do_swap_account __read_mostly;
79 /* for remember boot option*/
80 #ifdef CONFIG_MEMCG_SWAP_ENABLED
81 static int really_do_swap_account __initdata = 1;
83 static int really_do_swap_account __initdata;
87 #define do_swap_account 0
91 static const char * const mem_cgroup_stat_names[] = {
100 enum mem_cgroup_events_index {
101 MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */
102 MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */
103 MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */
104 MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */
105 MEM_CGROUP_EVENTS_NSTATS,
108 static const char * const mem_cgroup_events_names[] = {
115 static const char * const mem_cgroup_lru_names[] = {
124 * Per memcg event counter is incremented at every pagein/pageout. With THP,
125 * it will be incremated by the number of pages. This counter is used for
126 * for trigger some periodic events. This is straightforward and better
127 * than using jiffies etc. to handle periodic memcg event.
129 enum mem_cgroup_events_target {
130 MEM_CGROUP_TARGET_THRESH,
131 MEM_CGROUP_TARGET_SOFTLIMIT,
132 MEM_CGROUP_TARGET_NUMAINFO,
135 #define THRESHOLDS_EVENTS_TARGET 128
136 #define SOFTLIMIT_EVENTS_TARGET 1024
137 #define NUMAINFO_EVENTS_TARGET 1024
139 struct mem_cgroup_stat_cpu {
140 long count[MEM_CGROUP_STAT_NSTATS];
141 unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
142 unsigned long nr_page_events;
143 unsigned long targets[MEM_CGROUP_NTARGETS];
146 struct mem_cgroup_reclaim_iter {
148 * last scanned hierarchy member. Valid only if last_dead_count
149 * matches memcg->dead_count of the hierarchy root group.
151 struct mem_cgroup *last_visited;
154 /* scan generation, increased every round-trip */
155 unsigned int generation;
159 * per-zone information in memory controller.
161 struct mem_cgroup_per_zone {
162 struct lruvec lruvec;
163 unsigned long lru_size[NR_LRU_LISTS];
165 struct mem_cgroup_reclaim_iter reclaim_iter[DEF_PRIORITY + 1];
167 struct rb_node tree_node; /* RB tree node */
168 unsigned long usage_in_excess;/* Set to the value by which */
169 /* the soft limit is exceeded*/
171 struct mem_cgroup *memcg; /* Back pointer, we cannot */
172 /* use container_of */
175 struct mem_cgroup_per_node {
176 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
180 * Cgroups above their limits are maintained in a RB-Tree, independent of
181 * their hierarchy representation
184 struct mem_cgroup_tree_per_zone {
185 struct rb_root rb_root;
189 struct mem_cgroup_tree_per_node {
190 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
193 struct mem_cgroup_tree {
194 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
197 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
199 struct mem_cgroup_threshold {
200 struct eventfd_ctx *eventfd;
201 unsigned long threshold;
205 struct mem_cgroup_threshold_ary {
206 /* An array index points to threshold just below or equal to usage. */
207 int current_threshold;
208 /* Size of entries[] */
210 /* Array of thresholds */
211 struct mem_cgroup_threshold entries[0];
214 struct mem_cgroup_thresholds {
215 /* Primary thresholds array */
216 struct mem_cgroup_threshold_ary *primary;
218 * Spare threshold array.
219 * This is needed to make mem_cgroup_unregister_event() "never fail".
220 * It must be able to store at least primary->size - 1 entries.
222 struct mem_cgroup_threshold_ary *spare;
226 struct mem_cgroup_eventfd_list {
227 struct list_head list;
228 struct eventfd_ctx *eventfd;
232 * cgroup_event represents events which userspace want to receive.
234 struct mem_cgroup_event {
236 * memcg which the event belongs to.
238 struct mem_cgroup *memcg;
240 * eventfd to signal userspace about the event.
242 struct eventfd_ctx *eventfd;
244 * Each of these stored in a list by the cgroup.
246 struct list_head list;
248 * register_event() callback will be used to add new userspace
249 * waiter for changes related to this event. Use eventfd_signal()
250 * on eventfd to send notification to userspace.
252 int (*register_event)(struct mem_cgroup *memcg,
253 struct eventfd_ctx *eventfd, const char *args);
255 * unregister_event() callback will be called when userspace closes
256 * the eventfd or on cgroup removing. This callback must be set,
257 * if you want provide notification functionality.
259 void (*unregister_event)(struct mem_cgroup *memcg,
260 struct eventfd_ctx *eventfd);
262 * All fields below needed to unregister event when
263 * userspace closes eventfd.
266 wait_queue_head_t *wqh;
268 struct work_struct remove;
271 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
272 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
275 * The memory controller data structure. The memory controller controls both
276 * page cache and RSS per cgroup. We would eventually like to provide
277 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
278 * to help the administrator determine what knobs to tune.
280 * TODO: Add a water mark for the memory controller. Reclaim will begin when
281 * we hit the water mark. May be even add a low water mark, such that
282 * no reclaim occurs from a cgroup at it's low water mark, this is
283 * a feature that will be implemented much later in the future.
286 struct cgroup_subsys_state css;
288 /* Accounted resources */
289 struct page_counter memory;
290 struct page_counter memsw;
291 struct page_counter kmem;
293 unsigned long soft_limit;
295 /* vmpressure notifications */
296 struct vmpressure vmpressure;
298 /* css_online() has been completed */
302 * Should the accounting and control be hierarchical, per subtree?
305 unsigned long kmem_account_flags; /* See KMEM_ACCOUNTED_*, below */
309 atomic_t oom_wakeups;
312 /* OOM-Killer disable */
313 int oom_kill_disable;
315 /* protect arrays of thresholds */
316 struct mutex thresholds_lock;
318 /* thresholds for memory usage. RCU-protected */
319 struct mem_cgroup_thresholds thresholds;
321 /* thresholds for mem+swap usage. RCU-protected */
322 struct mem_cgroup_thresholds memsw_thresholds;
324 /* For oom notifier event fd */
325 struct list_head oom_notify;
328 * Should we move charges of a task when a task is moved into this
329 * mem_cgroup ? And what type of charges should we move ?
331 unsigned long move_charge_at_immigrate;
333 * set > 0 if pages under this cgroup are moving to other cgroup.
335 atomic_t moving_account;
336 /* taken only while moving_account > 0 */
337 spinlock_t move_lock;
341 struct mem_cgroup_stat_cpu __percpu *stat;
343 * used when a cpu is offlined or other synchronizations
344 * See mem_cgroup_read_stat().
346 struct mem_cgroup_stat_cpu nocpu_base;
347 spinlock_t pcp_counter_lock;
350 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
351 struct cg_proto tcp_mem;
353 #if defined(CONFIG_MEMCG_KMEM)
354 /* analogous to slab_common's slab_caches list, but per-memcg;
355 * protected by memcg_slab_mutex */
356 struct list_head memcg_slab_caches;
357 /* Index in the kmem_cache->memcg_params->memcg_caches array */
361 int last_scanned_node;
363 nodemask_t scan_nodes;
364 atomic_t numainfo_events;
365 atomic_t numainfo_updating;
368 /* List of events which userspace want to receive */
369 struct list_head event_list;
370 spinlock_t event_list_lock;
372 struct mem_cgroup_per_node *nodeinfo[0];
373 /* WARNING: nodeinfo must be the last member here */
376 /* internal only representation about the status of kmem accounting. */
378 KMEM_ACCOUNTED_ACTIVE, /* accounted by this cgroup itself */
379 KMEM_ACCOUNTED_DEAD, /* dead memcg with pending kmem charges */
382 #ifdef CONFIG_MEMCG_KMEM
383 static inline void memcg_kmem_set_active(struct mem_cgroup *memcg)
385 set_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
388 static bool memcg_kmem_is_active(struct mem_cgroup *memcg)
390 return test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
393 static void memcg_kmem_mark_dead(struct mem_cgroup *memcg)
396 * Our caller must use css_get() first, because memcg_uncharge_kmem()
397 * will call css_put() if it sees the memcg is dead.
400 if (test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags))
401 set_bit(KMEM_ACCOUNTED_DEAD, &memcg->kmem_account_flags);
404 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup *memcg)
406 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD,
407 &memcg->kmem_account_flags);
411 /* Stuffs for move charges at task migration. */
413 * Types of charges to be moved. "move_charge_at_immitgrate" and
414 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
417 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
418 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
422 /* "mc" and its members are protected by cgroup_mutex */
423 static struct move_charge_struct {
424 spinlock_t lock; /* for from, to */
425 struct mem_cgroup *from;
426 struct mem_cgroup *to;
427 unsigned long immigrate_flags;
428 unsigned long precharge;
429 unsigned long moved_charge;
430 unsigned long moved_swap;
431 struct task_struct *moving_task; /* a task moving charges */
432 wait_queue_head_t waitq; /* a waitq for other context */
434 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
435 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
438 static bool move_anon(void)
440 return test_bit(MOVE_CHARGE_TYPE_ANON, &mc.immigrate_flags);
443 static bool move_file(void)
445 return test_bit(MOVE_CHARGE_TYPE_FILE, &mc.immigrate_flags);
449 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
450 * limit reclaim to prevent infinite loops, if they ever occur.
452 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
453 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
456 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
457 MEM_CGROUP_CHARGE_TYPE_ANON,
458 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
459 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
463 /* for encoding cft->private value on file */
471 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
472 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
473 #define MEMFILE_ATTR(val) ((val) & 0xffff)
474 /* Used for OOM nofiier */
475 #define OOM_CONTROL (0)
478 * The memcg_create_mutex will be held whenever a new cgroup is created.
479 * As a consequence, any change that needs to protect against new child cgroups
480 * appearing has to hold it as well.
482 static DEFINE_MUTEX(memcg_create_mutex);
484 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
486 return s ? container_of(s, struct mem_cgroup, css) : NULL;
489 /* Some nice accessors for the vmpressure. */
490 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
493 memcg = root_mem_cgroup;
494 return &memcg->vmpressure;
497 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
499 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
502 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
504 return (memcg == root_mem_cgroup);
508 * We restrict the id in the range of [1, 65535], so it can fit into
511 #define MEM_CGROUP_ID_MAX USHRT_MAX
513 static inline unsigned short mem_cgroup_id(struct mem_cgroup *memcg)
515 return memcg->css.id;
518 static inline struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
520 struct cgroup_subsys_state *css;
522 css = css_from_id(id, &memory_cgrp_subsys);
523 return mem_cgroup_from_css(css);
526 /* Writing them here to avoid exposing memcg's inner layout */
527 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
529 void sock_update_memcg(struct sock *sk)
531 if (mem_cgroup_sockets_enabled) {
532 struct mem_cgroup *memcg;
533 struct cg_proto *cg_proto;
535 BUG_ON(!sk->sk_prot->proto_cgroup);
537 /* Socket cloning can throw us here with sk_cgrp already
538 * filled. It won't however, necessarily happen from
539 * process context. So the test for root memcg given
540 * the current task's memcg won't help us in this case.
542 * Respecting the original socket's memcg is a better
543 * decision in this case.
546 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
547 css_get(&sk->sk_cgrp->memcg->css);
552 memcg = mem_cgroup_from_task(current);
553 cg_proto = sk->sk_prot->proto_cgroup(memcg);
554 if (!mem_cgroup_is_root(memcg) &&
555 memcg_proto_active(cg_proto) &&
556 css_tryget_online(&memcg->css)) {
557 sk->sk_cgrp = cg_proto;
562 EXPORT_SYMBOL(sock_update_memcg);
564 void sock_release_memcg(struct sock *sk)
566 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
567 struct mem_cgroup *memcg;
568 WARN_ON(!sk->sk_cgrp->memcg);
569 memcg = sk->sk_cgrp->memcg;
570 css_put(&sk->sk_cgrp->memcg->css);
574 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
576 if (!memcg || mem_cgroup_is_root(memcg))
579 return &memcg->tcp_mem;
581 EXPORT_SYMBOL(tcp_proto_cgroup);
583 static void disarm_sock_keys(struct mem_cgroup *memcg)
585 if (!memcg_proto_activated(&memcg->tcp_mem))
587 static_key_slow_dec(&memcg_socket_limit_enabled);
590 static void disarm_sock_keys(struct mem_cgroup *memcg)
595 #ifdef CONFIG_MEMCG_KMEM
597 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
598 * The main reason for not using cgroup id for this:
599 * this works better in sparse environments, where we have a lot of memcgs,
600 * but only a few kmem-limited. Or also, if we have, for instance, 200
601 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
602 * 200 entry array for that.
604 * The current size of the caches array is stored in
605 * memcg_limited_groups_array_size. It will double each time we have to
608 static DEFINE_IDA(kmem_limited_groups);
609 int memcg_limited_groups_array_size;
612 * MIN_SIZE is different than 1, because we would like to avoid going through
613 * the alloc/free process all the time. In a small machine, 4 kmem-limited
614 * cgroups is a reasonable guess. In the future, it could be a parameter or
615 * tunable, but that is strictly not necessary.
617 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
618 * this constant directly from cgroup, but it is understandable that this is
619 * better kept as an internal representation in cgroup.c. In any case, the
620 * cgrp_id space is not getting any smaller, and we don't have to necessarily
621 * increase ours as well if it increases.
623 #define MEMCG_CACHES_MIN_SIZE 4
624 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
627 * A lot of the calls to the cache allocation functions are expected to be
628 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
629 * conditional to this static branch, we'll have to allow modules that does
630 * kmem_cache_alloc and the such to see this symbol as well
632 struct static_key memcg_kmem_enabled_key;
633 EXPORT_SYMBOL(memcg_kmem_enabled_key);
635 static void memcg_free_cache_id(int id);
637 static void disarm_kmem_keys(struct mem_cgroup *memcg)
639 if (memcg_kmem_is_active(memcg)) {
640 static_key_slow_dec(&memcg_kmem_enabled_key);
641 memcg_free_cache_id(memcg->kmemcg_id);
644 * This check can't live in kmem destruction function,
645 * since the charges will outlive the cgroup
647 WARN_ON(page_counter_read(&memcg->kmem));
650 static void disarm_kmem_keys(struct mem_cgroup *memcg)
653 #endif /* CONFIG_MEMCG_KMEM */
655 static void disarm_static_keys(struct mem_cgroup *memcg)
657 disarm_sock_keys(memcg);
658 disarm_kmem_keys(memcg);
661 static void drain_all_stock_async(struct mem_cgroup *memcg);
663 static struct mem_cgroup_per_zone *
664 mem_cgroup_zone_zoneinfo(struct mem_cgroup *memcg, struct zone *zone)
666 int nid = zone_to_nid(zone);
667 int zid = zone_idx(zone);
669 return &memcg->nodeinfo[nid]->zoneinfo[zid];
672 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
677 static struct mem_cgroup_per_zone *
678 mem_cgroup_page_zoneinfo(struct mem_cgroup *memcg, struct page *page)
680 int nid = page_to_nid(page);
681 int zid = page_zonenum(page);
683 return &memcg->nodeinfo[nid]->zoneinfo[zid];
686 static struct mem_cgroup_tree_per_zone *
687 soft_limit_tree_node_zone(int nid, int zid)
689 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
692 static struct mem_cgroup_tree_per_zone *
693 soft_limit_tree_from_page(struct page *page)
695 int nid = page_to_nid(page);
696 int zid = page_zonenum(page);
698 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
701 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone *mz,
702 struct mem_cgroup_tree_per_zone *mctz,
703 unsigned long new_usage_in_excess)
705 struct rb_node **p = &mctz->rb_root.rb_node;
706 struct rb_node *parent = NULL;
707 struct mem_cgroup_per_zone *mz_node;
712 mz->usage_in_excess = new_usage_in_excess;
713 if (!mz->usage_in_excess)
717 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
719 if (mz->usage_in_excess < mz_node->usage_in_excess)
722 * We can't avoid mem cgroups that are over their soft
723 * limit by the same amount
725 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
728 rb_link_node(&mz->tree_node, parent, p);
729 rb_insert_color(&mz->tree_node, &mctz->rb_root);
733 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
734 struct mem_cgroup_tree_per_zone *mctz)
738 rb_erase(&mz->tree_node, &mctz->rb_root);
742 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
743 struct mem_cgroup_tree_per_zone *mctz)
747 spin_lock_irqsave(&mctz->lock, flags);
748 __mem_cgroup_remove_exceeded(mz, mctz);
749 spin_unlock_irqrestore(&mctz->lock, flags);
752 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
754 unsigned long nr_pages = page_counter_read(&memcg->memory);
755 unsigned long soft_limit = ACCESS_ONCE(memcg->soft_limit);
756 unsigned long excess = 0;
758 if (nr_pages > soft_limit)
759 excess = nr_pages - soft_limit;
764 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
766 unsigned long excess;
767 struct mem_cgroup_per_zone *mz;
768 struct mem_cgroup_tree_per_zone *mctz;
770 mctz = soft_limit_tree_from_page(page);
772 * Necessary to update all ancestors when hierarchy is used.
773 * because their event counter is not touched.
775 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
776 mz = mem_cgroup_page_zoneinfo(memcg, page);
777 excess = soft_limit_excess(memcg);
779 * We have to update the tree if mz is on RB-tree or
780 * mem is over its softlimit.
782 if (excess || mz->on_tree) {
785 spin_lock_irqsave(&mctz->lock, flags);
786 /* if on-tree, remove it */
788 __mem_cgroup_remove_exceeded(mz, mctz);
790 * Insert again. mz->usage_in_excess will be updated.
791 * If excess is 0, no tree ops.
793 __mem_cgroup_insert_exceeded(mz, mctz, excess);
794 spin_unlock_irqrestore(&mctz->lock, flags);
799 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
801 struct mem_cgroup_tree_per_zone *mctz;
802 struct mem_cgroup_per_zone *mz;
806 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
807 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
808 mctz = soft_limit_tree_node_zone(nid, zid);
809 mem_cgroup_remove_exceeded(mz, mctz);
814 static struct mem_cgroup_per_zone *
815 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
817 struct rb_node *rightmost = NULL;
818 struct mem_cgroup_per_zone *mz;
822 rightmost = rb_last(&mctz->rb_root);
824 goto done; /* Nothing to reclaim from */
826 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
828 * Remove the node now but someone else can add it back,
829 * we will to add it back at the end of reclaim to its correct
830 * position in the tree.
832 __mem_cgroup_remove_exceeded(mz, mctz);
833 if (!soft_limit_excess(mz->memcg) ||
834 !css_tryget_online(&mz->memcg->css))
840 static struct mem_cgroup_per_zone *
841 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
843 struct mem_cgroup_per_zone *mz;
845 spin_lock_irq(&mctz->lock);
846 mz = __mem_cgroup_largest_soft_limit_node(mctz);
847 spin_unlock_irq(&mctz->lock);
852 * Implementation Note: reading percpu statistics for memcg.
854 * Both of vmstat[] and percpu_counter has threshold and do periodic
855 * synchronization to implement "quick" read. There are trade-off between
856 * reading cost and precision of value. Then, we may have a chance to implement
857 * a periodic synchronizion of counter in memcg's counter.
859 * But this _read() function is used for user interface now. The user accounts
860 * memory usage by memory cgroup and he _always_ requires exact value because
861 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
862 * have to visit all online cpus and make sum. So, for now, unnecessary
863 * synchronization is not implemented. (just implemented for cpu hotplug)
865 * If there are kernel internal actions which can make use of some not-exact
866 * value, and reading all cpu value can be performance bottleneck in some
867 * common workload, threashold and synchonization as vmstat[] should be
870 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
871 enum mem_cgroup_stat_index idx)
877 for_each_online_cpu(cpu)
878 val += per_cpu(memcg->stat->count[idx], cpu);
879 #ifdef CONFIG_HOTPLUG_CPU
880 spin_lock(&memcg->pcp_counter_lock);
881 val += memcg->nocpu_base.count[idx];
882 spin_unlock(&memcg->pcp_counter_lock);
888 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
889 enum mem_cgroup_events_index idx)
891 unsigned long val = 0;
895 for_each_online_cpu(cpu)
896 val += per_cpu(memcg->stat->events[idx], cpu);
897 #ifdef CONFIG_HOTPLUG_CPU
898 spin_lock(&memcg->pcp_counter_lock);
899 val += memcg->nocpu_base.events[idx];
900 spin_unlock(&memcg->pcp_counter_lock);
906 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
911 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
912 * counted as CACHE even if it's on ANON LRU.
915 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
918 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
921 if (PageTransHuge(page))
922 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
925 /* pagein of a big page is an event. So, ignore page size */
927 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
929 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
930 nr_pages = -nr_pages; /* for event */
933 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
936 unsigned long mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
938 struct mem_cgroup_per_zone *mz;
940 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
941 return mz->lru_size[lru];
944 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
946 unsigned int lru_mask)
948 unsigned long nr = 0;
951 VM_BUG_ON((unsigned)nid >= nr_node_ids);
953 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
954 struct mem_cgroup_per_zone *mz;
958 if (!(BIT(lru) & lru_mask))
960 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
961 nr += mz->lru_size[lru];
967 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
968 unsigned int lru_mask)
970 unsigned long nr = 0;
973 for_each_node_state(nid, N_MEMORY)
974 nr += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
978 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
979 enum mem_cgroup_events_target target)
981 unsigned long val, next;
983 val = __this_cpu_read(memcg->stat->nr_page_events);
984 next = __this_cpu_read(memcg->stat->targets[target]);
985 /* from time_after() in jiffies.h */
986 if ((long)next - (long)val < 0) {
988 case MEM_CGROUP_TARGET_THRESH:
989 next = val + THRESHOLDS_EVENTS_TARGET;
991 case MEM_CGROUP_TARGET_SOFTLIMIT:
992 next = val + SOFTLIMIT_EVENTS_TARGET;
994 case MEM_CGROUP_TARGET_NUMAINFO:
995 next = val + NUMAINFO_EVENTS_TARGET;
1000 __this_cpu_write(memcg->stat->targets[target], next);
1007 * Check events in order.
1010 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
1012 /* threshold event is triggered in finer grain than soft limit */
1013 if (unlikely(mem_cgroup_event_ratelimit(memcg,
1014 MEM_CGROUP_TARGET_THRESH))) {
1016 bool do_numainfo __maybe_unused;
1018 do_softlimit = mem_cgroup_event_ratelimit(memcg,
1019 MEM_CGROUP_TARGET_SOFTLIMIT);
1020 #if MAX_NUMNODES > 1
1021 do_numainfo = mem_cgroup_event_ratelimit(memcg,
1022 MEM_CGROUP_TARGET_NUMAINFO);
1024 mem_cgroup_threshold(memcg);
1025 if (unlikely(do_softlimit))
1026 mem_cgroup_update_tree(memcg, page);
1027 #if MAX_NUMNODES > 1
1028 if (unlikely(do_numainfo))
1029 atomic_inc(&memcg->numainfo_events);
1034 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1037 * mm_update_next_owner() may clear mm->owner to NULL
1038 * if it races with swapoff, page migration, etc.
1039 * So this can be called with p == NULL.
1044 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
1047 static struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
1049 struct mem_cgroup *memcg = NULL;
1054 * Page cache insertions can happen withou an
1055 * actual mm context, e.g. during disk probing
1056 * on boot, loopback IO, acct() writes etc.
1059 memcg = root_mem_cgroup;
1061 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1062 if (unlikely(!memcg))
1063 memcg = root_mem_cgroup;
1065 } while (!css_tryget_online(&memcg->css));
1071 * Returns a next (in a pre-order walk) alive memcg (with elevated css
1072 * ref. count) or NULL if the whole root's subtree has been visited.
1074 * helper function to be used by mem_cgroup_iter
1076 static struct mem_cgroup *__mem_cgroup_iter_next(struct mem_cgroup *root,
1077 struct mem_cgroup *last_visited)
1079 struct cgroup_subsys_state *prev_css, *next_css;
1081 prev_css = last_visited ? &last_visited->css : NULL;
1083 next_css = css_next_descendant_pre(prev_css, &root->css);
1086 * Even if we found a group we have to make sure it is
1087 * alive. css && !memcg means that the groups should be
1088 * skipped and we should continue the tree walk.
1089 * last_visited css is safe to use because it is
1090 * protected by css_get and the tree walk is rcu safe.
1092 * We do not take a reference on the root of the tree walk
1093 * because we might race with the root removal when it would
1094 * be the only node in the iterated hierarchy and mem_cgroup_iter
1095 * would end up in an endless loop because it expects that at
1096 * least one valid node will be returned. Root cannot disappear
1097 * because caller of the iterator should hold it already so
1098 * skipping css reference should be safe.
1101 struct mem_cgroup *memcg = mem_cgroup_from_css(next_css);
1103 if (next_css == &root->css)
1106 if (css_tryget_online(next_css)) {
1108 * Make sure the memcg is initialized:
1109 * mem_cgroup_css_online() orders the the
1110 * initialization against setting the flag.
1112 if (smp_load_acquire(&memcg->initialized))
1117 prev_css = next_css;
1124 static void mem_cgroup_iter_invalidate(struct mem_cgroup *root)
1127 * When a group in the hierarchy below root is destroyed, the
1128 * hierarchy iterator can no longer be trusted since it might
1129 * have pointed to the destroyed group. Invalidate it.
1131 atomic_inc(&root->dead_count);
1134 static struct mem_cgroup *
1135 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter *iter,
1136 struct mem_cgroup *root,
1139 struct mem_cgroup *position = NULL;
1141 * A cgroup destruction happens in two stages: offlining and
1142 * release. They are separated by a RCU grace period.
1144 * If the iterator is valid, we may still race with an
1145 * offlining. The RCU lock ensures the object won't be
1146 * released, tryget will fail if we lost the race.
1148 *sequence = atomic_read(&root->dead_count);
1149 if (iter->last_dead_count == *sequence) {
1151 position = iter->last_visited;
1154 * We cannot take a reference to root because we might race
1155 * with root removal and returning NULL would end up in
1156 * an endless loop on the iterator user level when root
1157 * would be returned all the time.
1159 if (position && position != root &&
1160 !css_tryget_online(&position->css))
1166 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter *iter,
1167 struct mem_cgroup *last_visited,
1168 struct mem_cgroup *new_position,
1169 struct mem_cgroup *root,
1172 /* root reference counting symmetric to mem_cgroup_iter_load */
1173 if (last_visited && last_visited != root)
1174 css_put(&last_visited->css);
1176 * We store the sequence count from the time @last_visited was
1177 * loaded successfully instead of rereading it here so that we
1178 * don't lose destruction events in between. We could have
1179 * raced with the destruction of @new_position after all.
1181 iter->last_visited = new_position;
1183 iter->last_dead_count = sequence;
1187 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1188 * @root: hierarchy root
1189 * @prev: previously returned memcg, NULL on first invocation
1190 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1192 * Returns references to children of the hierarchy below @root, or
1193 * @root itself, or %NULL after a full round-trip.
1195 * Caller must pass the return value in @prev on subsequent
1196 * invocations for reference counting, or use mem_cgroup_iter_break()
1197 * to cancel a hierarchy walk before the round-trip is complete.
1199 * Reclaimers can specify a zone and a priority level in @reclaim to
1200 * divide up the memcgs in the hierarchy among all concurrent
1201 * reclaimers operating on the same zone and priority.
1203 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1204 struct mem_cgroup *prev,
1205 struct mem_cgroup_reclaim_cookie *reclaim)
1207 struct mem_cgroup *memcg = NULL;
1208 struct mem_cgroup *last_visited = NULL;
1210 if (mem_cgroup_disabled())
1214 root = root_mem_cgroup;
1216 if (prev && !reclaim)
1217 last_visited = prev;
1219 if (!root->use_hierarchy && root != root_mem_cgroup) {
1227 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1228 int uninitialized_var(seq);
1231 struct mem_cgroup_per_zone *mz;
1233 mz = mem_cgroup_zone_zoneinfo(root, reclaim->zone);
1234 iter = &mz->reclaim_iter[reclaim->priority];
1235 if (prev && reclaim->generation != iter->generation) {
1236 iter->last_visited = NULL;
1240 last_visited = mem_cgroup_iter_load(iter, root, &seq);
1243 memcg = __mem_cgroup_iter_next(root, last_visited);
1246 mem_cgroup_iter_update(iter, last_visited, memcg, root,
1251 else if (!prev && memcg)
1252 reclaim->generation = iter->generation;
1261 if (prev && prev != root)
1262 css_put(&prev->css);
1268 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1269 * @root: hierarchy root
1270 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1272 void mem_cgroup_iter_break(struct mem_cgroup *root,
1273 struct mem_cgroup *prev)
1276 root = root_mem_cgroup;
1277 if (prev && prev != root)
1278 css_put(&prev->css);
1282 * Iteration constructs for visiting all cgroups (under a tree). If
1283 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1284 * be used for reference counting.
1286 #define for_each_mem_cgroup_tree(iter, root) \
1287 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1289 iter = mem_cgroup_iter(root, iter, NULL))
1291 #define for_each_mem_cgroup(iter) \
1292 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1294 iter = mem_cgroup_iter(NULL, iter, NULL))
1296 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1298 struct mem_cgroup *memcg;
1301 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1302 if (unlikely(!memcg))
1307 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1310 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1318 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1321 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1322 * @zone: zone of the wanted lruvec
1323 * @memcg: memcg of the wanted lruvec
1325 * Returns the lru list vector holding pages for the given @zone and
1326 * @mem. This can be the global zone lruvec, if the memory controller
1329 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1330 struct mem_cgroup *memcg)
1332 struct mem_cgroup_per_zone *mz;
1333 struct lruvec *lruvec;
1335 if (mem_cgroup_disabled()) {
1336 lruvec = &zone->lruvec;
1340 mz = mem_cgroup_zone_zoneinfo(memcg, zone);
1341 lruvec = &mz->lruvec;
1344 * Since a node can be onlined after the mem_cgroup was created,
1345 * we have to be prepared to initialize lruvec->zone here;
1346 * and if offlined then reonlined, we need to reinitialize it.
1348 if (unlikely(lruvec->zone != zone))
1349 lruvec->zone = zone;
1354 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1356 * @zone: zone of the page
1358 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1360 struct mem_cgroup_per_zone *mz;
1361 struct mem_cgroup *memcg;
1362 struct page_cgroup *pc;
1363 struct lruvec *lruvec;
1365 if (mem_cgroup_disabled()) {
1366 lruvec = &zone->lruvec;
1370 pc = lookup_page_cgroup(page);
1371 memcg = pc->mem_cgroup;
1374 * Surreptitiously switch any uncharged offlist page to root:
1375 * an uncharged page off lru does nothing to secure
1376 * its former mem_cgroup from sudden removal.
1378 * Our caller holds lru_lock, and PageCgroupUsed is updated
1379 * under page_cgroup lock: between them, they make all uses
1380 * of pc->mem_cgroup safe.
1382 if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1383 pc->mem_cgroup = memcg = root_mem_cgroup;
1385 mz = mem_cgroup_page_zoneinfo(memcg, page);
1386 lruvec = &mz->lruvec;
1389 * Since a node can be onlined after the mem_cgroup was created,
1390 * we have to be prepared to initialize lruvec->zone here;
1391 * and if offlined then reonlined, we need to reinitialize it.
1393 if (unlikely(lruvec->zone != zone))
1394 lruvec->zone = zone;
1399 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1400 * @lruvec: mem_cgroup per zone lru vector
1401 * @lru: index of lru list the page is sitting on
1402 * @nr_pages: positive when adding or negative when removing
1404 * This function must be called when a page is added to or removed from an
1407 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1410 struct mem_cgroup_per_zone *mz;
1411 unsigned long *lru_size;
1413 if (mem_cgroup_disabled())
1416 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1417 lru_size = mz->lru_size + lru;
1418 *lru_size += nr_pages;
1419 VM_BUG_ON((long)(*lru_size) < 0);
1423 * Checks whether given mem is same or in the root_mem_cgroup's
1426 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1427 struct mem_cgroup *memcg)
1429 if (root_memcg == memcg)
1431 if (!root_memcg->use_hierarchy || !memcg)
1433 return cgroup_is_descendant(memcg->css.cgroup, root_memcg->css.cgroup);
1436 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1437 struct mem_cgroup *memcg)
1442 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1447 bool task_in_mem_cgroup(struct task_struct *task,
1448 const struct mem_cgroup *memcg)
1450 struct mem_cgroup *curr = NULL;
1451 struct task_struct *p;
1454 p = find_lock_task_mm(task);
1456 curr = get_mem_cgroup_from_mm(p->mm);
1460 * All threads may have already detached their mm's, but the oom
1461 * killer still needs to detect if they have already been oom
1462 * killed to prevent needlessly killing additional tasks.
1465 curr = mem_cgroup_from_task(task);
1467 css_get(&curr->css);
1471 * We should check use_hierarchy of "memcg" not "curr". Because checking
1472 * use_hierarchy of "curr" here make this function true if hierarchy is
1473 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1474 * hierarchy(even if use_hierarchy is disabled in "memcg").
1476 ret = mem_cgroup_same_or_subtree(memcg, curr);
1477 css_put(&curr->css);
1481 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1483 unsigned long inactive_ratio;
1484 unsigned long inactive;
1485 unsigned long active;
1488 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1489 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1491 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1493 inactive_ratio = int_sqrt(10 * gb);
1497 return inactive * inactive_ratio < active;
1500 #define mem_cgroup_from_counter(counter, member) \
1501 container_of(counter, struct mem_cgroup, member)
1504 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1505 * @memcg: the memory cgroup
1507 * Returns the maximum amount of memory @mem can be charged with, in
1510 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1512 unsigned long margin = 0;
1513 unsigned long count;
1514 unsigned long limit;
1516 count = page_counter_read(&memcg->memory);
1517 limit = ACCESS_ONCE(memcg->memory.limit);
1519 margin = limit - count;
1521 if (do_swap_account) {
1522 count = page_counter_read(&memcg->memsw);
1523 limit = ACCESS_ONCE(memcg->memsw.limit);
1525 margin = min(margin, limit - count);
1531 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1534 if (mem_cgroup_disabled() || !memcg->css.parent)
1535 return vm_swappiness;
1537 return memcg->swappiness;
1541 * memcg->moving_account is used for checking possibility that some thread is
1542 * calling move_account(). When a thread on CPU-A starts moving pages under
1543 * a memcg, other threads should check memcg->moving_account under
1544 * rcu_read_lock(), like this:
1548 * memcg->moving_account+1 if (memcg->mocing_account)
1550 * synchronize_rcu() update something.
1555 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1557 atomic_inc(&memcg->moving_account);
1561 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1564 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1565 * We check NULL in callee rather than caller.
1568 atomic_dec(&memcg->moving_account);
1572 * A routine for checking "mem" is under move_account() or not.
1574 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1575 * moving cgroups. This is for waiting at high-memory pressure
1578 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1580 struct mem_cgroup *from;
1581 struct mem_cgroup *to;
1584 * Unlike task_move routines, we access mc.to, mc.from not under
1585 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1587 spin_lock(&mc.lock);
1593 ret = mem_cgroup_same_or_subtree(memcg, from)
1594 || mem_cgroup_same_or_subtree(memcg, to);
1596 spin_unlock(&mc.lock);
1600 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1602 if (mc.moving_task && current != mc.moving_task) {
1603 if (mem_cgroup_under_move(memcg)) {
1605 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1606 /* moving charge context might have finished. */
1609 finish_wait(&mc.waitq, &wait);
1617 * Take this lock when
1618 * - a code tries to modify page's memcg while it's USED.
1619 * - a code tries to modify page state accounting in a memcg.
1621 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1622 unsigned long *flags)
1624 spin_lock_irqsave(&memcg->move_lock, *flags);
1627 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1628 unsigned long *flags)
1630 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1633 #define K(x) ((x) << (PAGE_SHIFT-10))
1635 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1636 * @memcg: The memory cgroup that went over limit
1637 * @p: Task that is going to be killed
1639 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1642 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1644 /* oom_info_lock ensures that parallel ooms do not interleave */
1645 static DEFINE_MUTEX(oom_info_lock);
1646 struct mem_cgroup *iter;
1652 mutex_lock(&oom_info_lock);
1655 pr_info("Task in ");
1656 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1657 pr_info(" killed as a result of limit of ");
1658 pr_cont_cgroup_path(memcg->css.cgroup);
1663 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1664 K((u64)page_counter_read(&memcg->memory)),
1665 K((u64)memcg->memory.limit), memcg->memory.failcnt);
1666 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1667 K((u64)page_counter_read(&memcg->memsw)),
1668 K((u64)memcg->memsw.limit), memcg->memsw.failcnt);
1669 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1670 K((u64)page_counter_read(&memcg->kmem)),
1671 K((u64)memcg->kmem.limit), memcg->kmem.failcnt);
1673 for_each_mem_cgroup_tree(iter, memcg) {
1674 pr_info("Memory cgroup stats for ");
1675 pr_cont_cgroup_path(iter->css.cgroup);
1678 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1679 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1681 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1682 K(mem_cgroup_read_stat(iter, i)));
1685 for (i = 0; i < NR_LRU_LISTS; i++)
1686 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1687 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1691 mutex_unlock(&oom_info_lock);
1695 * This function returns the number of memcg under hierarchy tree. Returns
1696 * 1(self count) if no children.
1698 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1701 struct mem_cgroup *iter;
1703 for_each_mem_cgroup_tree(iter, memcg)
1709 * Return the memory (and swap, if configured) limit for a memcg.
1711 static unsigned long mem_cgroup_get_limit(struct mem_cgroup *memcg)
1713 unsigned long limit;
1715 limit = memcg->memory.limit;
1716 if (mem_cgroup_swappiness(memcg)) {
1717 unsigned long memsw_limit;
1719 memsw_limit = memcg->memsw.limit;
1720 limit = min(limit + total_swap_pages, memsw_limit);
1725 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1728 struct mem_cgroup *iter;
1729 unsigned long chosen_points = 0;
1730 unsigned long totalpages;
1731 unsigned int points = 0;
1732 struct task_struct *chosen = NULL;
1735 * If current has a pending SIGKILL or is exiting, then automatically
1736 * select it. The goal is to allow it to allocate so that it may
1737 * quickly exit and free its memory.
1739 if (fatal_signal_pending(current) || current->flags & PF_EXITING) {
1740 set_thread_flag(TIF_MEMDIE);
1744 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1745 totalpages = mem_cgroup_get_limit(memcg) ? : 1;
1746 for_each_mem_cgroup_tree(iter, memcg) {
1747 struct css_task_iter it;
1748 struct task_struct *task;
1750 css_task_iter_start(&iter->css, &it);
1751 while ((task = css_task_iter_next(&it))) {
1752 switch (oom_scan_process_thread(task, totalpages, NULL,
1754 case OOM_SCAN_SELECT:
1756 put_task_struct(chosen);
1758 chosen_points = ULONG_MAX;
1759 get_task_struct(chosen);
1761 case OOM_SCAN_CONTINUE:
1763 case OOM_SCAN_ABORT:
1764 css_task_iter_end(&it);
1765 mem_cgroup_iter_break(memcg, iter);
1767 put_task_struct(chosen);
1772 points = oom_badness(task, memcg, NULL, totalpages);
1773 if (!points || points < chosen_points)
1775 /* Prefer thread group leaders for display purposes */
1776 if (points == chosen_points &&
1777 thread_group_leader(chosen))
1781 put_task_struct(chosen);
1783 chosen_points = points;
1784 get_task_struct(chosen);
1786 css_task_iter_end(&it);
1791 points = chosen_points * 1000 / totalpages;
1792 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1793 NULL, "Memory cgroup out of memory");
1797 * test_mem_cgroup_node_reclaimable
1798 * @memcg: the target memcg
1799 * @nid: the node ID to be checked.
1800 * @noswap : specify true here if the user wants flle only information.
1802 * This function returns whether the specified memcg contains any
1803 * reclaimable pages on a node. Returns true if there are any reclaimable
1804 * pages in the node.
1806 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1807 int nid, bool noswap)
1809 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1811 if (noswap || !total_swap_pages)
1813 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1818 #if MAX_NUMNODES > 1
1821 * Always updating the nodemask is not very good - even if we have an empty
1822 * list or the wrong list here, we can start from some node and traverse all
1823 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1826 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1830 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1831 * pagein/pageout changes since the last update.
1833 if (!atomic_read(&memcg->numainfo_events))
1835 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1838 /* make a nodemask where this memcg uses memory from */
1839 memcg->scan_nodes = node_states[N_MEMORY];
1841 for_each_node_mask(nid, node_states[N_MEMORY]) {
1843 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1844 node_clear(nid, memcg->scan_nodes);
1847 atomic_set(&memcg->numainfo_events, 0);
1848 atomic_set(&memcg->numainfo_updating, 0);
1852 * Selecting a node where we start reclaim from. Because what we need is just
1853 * reducing usage counter, start from anywhere is O,K. Considering
1854 * memory reclaim from current node, there are pros. and cons.
1856 * Freeing memory from current node means freeing memory from a node which
1857 * we'll use or we've used. So, it may make LRU bad. And if several threads
1858 * hit limits, it will see a contention on a node. But freeing from remote
1859 * node means more costs for memory reclaim because of memory latency.
1861 * Now, we use round-robin. Better algorithm is welcomed.
1863 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1867 mem_cgroup_may_update_nodemask(memcg);
1868 node = memcg->last_scanned_node;
1870 node = next_node(node, memcg->scan_nodes);
1871 if (node == MAX_NUMNODES)
1872 node = first_node(memcg->scan_nodes);
1874 * We call this when we hit limit, not when pages are added to LRU.
1875 * No LRU may hold pages because all pages are UNEVICTABLE or
1876 * memcg is too small and all pages are not on LRU. In that case,
1877 * we use curret node.
1879 if (unlikely(node == MAX_NUMNODES))
1880 node = numa_node_id();
1882 memcg->last_scanned_node = node;
1887 * Check all nodes whether it contains reclaimable pages or not.
1888 * For quick scan, we make use of scan_nodes. This will allow us to skip
1889 * unused nodes. But scan_nodes is lazily updated and may not cotain
1890 * enough new information. We need to do double check.
1892 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1897 * quick check...making use of scan_node.
1898 * We can skip unused nodes.
1900 if (!nodes_empty(memcg->scan_nodes)) {
1901 for (nid = first_node(memcg->scan_nodes);
1903 nid = next_node(nid, memcg->scan_nodes)) {
1905 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1910 * Check rest of nodes.
1912 for_each_node_state(nid, N_MEMORY) {
1913 if (node_isset(nid, memcg->scan_nodes))
1915 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1922 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1927 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1929 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
1933 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1936 unsigned long *total_scanned)
1938 struct mem_cgroup *victim = NULL;
1941 unsigned long excess;
1942 unsigned long nr_scanned;
1943 struct mem_cgroup_reclaim_cookie reclaim = {
1948 excess = soft_limit_excess(root_memcg);
1951 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1956 * If we have not been able to reclaim
1957 * anything, it might because there are
1958 * no reclaimable pages under this hierarchy
1963 * We want to do more targeted reclaim.
1964 * excess >> 2 is not to excessive so as to
1965 * reclaim too much, nor too less that we keep
1966 * coming back to reclaim from this cgroup
1968 if (total >= (excess >> 2) ||
1969 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1974 if (!mem_cgroup_reclaimable(victim, false))
1976 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1978 *total_scanned += nr_scanned;
1979 if (!soft_limit_excess(root_memcg))
1982 mem_cgroup_iter_break(root_memcg, victim);
1986 #ifdef CONFIG_LOCKDEP
1987 static struct lockdep_map memcg_oom_lock_dep_map = {
1988 .name = "memcg_oom_lock",
1992 static DEFINE_SPINLOCK(memcg_oom_lock);
1995 * Check OOM-Killer is already running under our hierarchy.
1996 * If someone is running, return false.
1998 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
2000 struct mem_cgroup *iter, *failed = NULL;
2002 spin_lock(&memcg_oom_lock);
2004 for_each_mem_cgroup_tree(iter, memcg) {
2005 if (iter->oom_lock) {
2007 * this subtree of our hierarchy is already locked
2008 * so we cannot give a lock.
2011 mem_cgroup_iter_break(memcg, iter);
2014 iter->oom_lock = true;
2019 * OK, we failed to lock the whole subtree so we have
2020 * to clean up what we set up to the failing subtree
2022 for_each_mem_cgroup_tree(iter, memcg) {
2023 if (iter == failed) {
2024 mem_cgroup_iter_break(memcg, iter);
2027 iter->oom_lock = false;
2030 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
2032 spin_unlock(&memcg_oom_lock);
2037 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
2039 struct mem_cgroup *iter;
2041 spin_lock(&memcg_oom_lock);
2042 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
2043 for_each_mem_cgroup_tree(iter, memcg)
2044 iter->oom_lock = false;
2045 spin_unlock(&memcg_oom_lock);
2048 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
2050 struct mem_cgroup *iter;
2052 for_each_mem_cgroup_tree(iter, memcg)
2053 atomic_inc(&iter->under_oom);
2056 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
2058 struct mem_cgroup *iter;
2061 * When a new child is created while the hierarchy is under oom,
2062 * mem_cgroup_oom_lock() may not be called. We have to use
2063 * atomic_add_unless() here.
2065 for_each_mem_cgroup_tree(iter, memcg)
2066 atomic_add_unless(&iter->under_oom, -1, 0);
2069 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
2071 struct oom_wait_info {
2072 struct mem_cgroup *memcg;
2076 static int memcg_oom_wake_function(wait_queue_t *wait,
2077 unsigned mode, int sync, void *arg)
2079 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
2080 struct mem_cgroup *oom_wait_memcg;
2081 struct oom_wait_info *oom_wait_info;
2083 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
2084 oom_wait_memcg = oom_wait_info->memcg;
2087 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2088 * Then we can use css_is_ancestor without taking care of RCU.
2090 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
2091 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
2093 return autoremove_wake_function(wait, mode, sync, arg);
2096 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
2098 atomic_inc(&memcg->oom_wakeups);
2099 /* for filtering, pass "memcg" as argument. */
2100 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
2103 static void memcg_oom_recover(struct mem_cgroup *memcg)
2105 if (memcg && atomic_read(&memcg->under_oom))
2106 memcg_wakeup_oom(memcg);
2109 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
2111 if (!current->memcg_oom.may_oom)
2114 * We are in the middle of the charge context here, so we
2115 * don't want to block when potentially sitting on a callstack
2116 * that holds all kinds of filesystem and mm locks.
2118 * Also, the caller may handle a failed allocation gracefully
2119 * (like optional page cache readahead) and so an OOM killer
2120 * invocation might not even be necessary.
2122 * That's why we don't do anything here except remember the
2123 * OOM context and then deal with it at the end of the page
2124 * fault when the stack is unwound, the locks are released,
2125 * and when we know whether the fault was overall successful.
2127 css_get(&memcg->css);
2128 current->memcg_oom.memcg = memcg;
2129 current->memcg_oom.gfp_mask = mask;
2130 current->memcg_oom.order = order;
2134 * mem_cgroup_oom_synchronize - complete memcg OOM handling
2135 * @handle: actually kill/wait or just clean up the OOM state
2137 * This has to be called at the end of a page fault if the memcg OOM
2138 * handler was enabled.
2140 * Memcg supports userspace OOM handling where failed allocations must
2141 * sleep on a waitqueue until the userspace task resolves the
2142 * situation. Sleeping directly in the charge context with all kinds
2143 * of locks held is not a good idea, instead we remember an OOM state
2144 * in the task and mem_cgroup_oom_synchronize() has to be called at
2145 * the end of the page fault to complete the OOM handling.
2147 * Returns %true if an ongoing memcg OOM situation was detected and
2148 * completed, %false otherwise.
2150 bool mem_cgroup_oom_synchronize(bool handle)
2152 struct mem_cgroup *memcg = current->memcg_oom.memcg;
2153 struct oom_wait_info owait;
2156 /* OOM is global, do not handle */
2163 owait.memcg = memcg;
2164 owait.wait.flags = 0;
2165 owait.wait.func = memcg_oom_wake_function;
2166 owait.wait.private = current;
2167 INIT_LIST_HEAD(&owait.wait.task_list);
2169 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2170 mem_cgroup_mark_under_oom(memcg);
2172 locked = mem_cgroup_oom_trylock(memcg);
2175 mem_cgroup_oom_notify(memcg);
2177 if (locked && !memcg->oom_kill_disable) {
2178 mem_cgroup_unmark_under_oom(memcg);
2179 finish_wait(&memcg_oom_waitq, &owait.wait);
2180 mem_cgroup_out_of_memory(memcg, current->memcg_oom.gfp_mask,
2181 current->memcg_oom.order);
2184 mem_cgroup_unmark_under_oom(memcg);
2185 finish_wait(&memcg_oom_waitq, &owait.wait);
2189 mem_cgroup_oom_unlock(memcg);
2191 * There is no guarantee that an OOM-lock contender
2192 * sees the wakeups triggered by the OOM kill
2193 * uncharges. Wake any sleepers explicitely.
2195 memcg_oom_recover(memcg);
2198 current->memcg_oom.memcg = NULL;
2199 css_put(&memcg->css);
2204 * mem_cgroup_begin_page_stat - begin a page state statistics transaction
2205 * @page: page that is going to change accounted state
2206 * @locked: &memcg->move_lock slowpath was taken
2207 * @flags: IRQ-state flags for &memcg->move_lock
2209 * This function must mark the beginning of an accounted page state
2210 * change to prevent double accounting when the page is concurrently
2211 * being moved to another memcg:
2213 * memcg = mem_cgroup_begin_page_stat(page, &locked, &flags);
2214 * if (TestClearPageState(page))
2215 * mem_cgroup_update_page_stat(memcg, state, -1);
2216 * mem_cgroup_end_page_stat(memcg, locked, flags);
2218 * The RCU lock is held throughout the transaction. The fast path can
2219 * get away without acquiring the memcg->move_lock (@locked is false)
2220 * because page moving starts with an RCU grace period.
2222 * The RCU lock also protects the memcg from being freed when the page
2223 * state that is going to change is the only thing preventing the page
2224 * from being uncharged. E.g. end-writeback clearing PageWriteback(),
2225 * which allows migration to go ahead and uncharge the page before the
2226 * account transaction might be complete.
2228 struct mem_cgroup *mem_cgroup_begin_page_stat(struct page *page,
2230 unsigned long *flags)
2232 struct mem_cgroup *memcg;
2233 struct page_cgroup *pc;
2237 if (mem_cgroup_disabled())
2240 pc = lookup_page_cgroup(page);
2242 memcg = pc->mem_cgroup;
2243 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2247 if (atomic_read(&memcg->moving_account) <= 0)
2250 move_lock_mem_cgroup(memcg, flags);
2251 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2252 move_unlock_mem_cgroup(memcg, flags);
2261 * mem_cgroup_end_page_stat - finish a page state statistics transaction
2262 * @memcg: the memcg that was accounted against
2263 * @locked: value received from mem_cgroup_begin_page_stat()
2264 * @flags: value received from mem_cgroup_begin_page_stat()
2266 void mem_cgroup_end_page_stat(struct mem_cgroup *memcg, bool locked,
2267 unsigned long flags)
2269 if (memcg && locked)
2270 move_unlock_mem_cgroup(memcg, &flags);
2276 * mem_cgroup_update_page_stat - update page state statistics
2277 * @memcg: memcg to account against
2278 * @idx: page state item to account
2279 * @val: number of pages (positive or negative)
2281 * See mem_cgroup_begin_page_stat() for locking requirements.
2283 void mem_cgroup_update_page_stat(struct mem_cgroup *memcg,
2284 enum mem_cgroup_stat_index idx, int val)
2286 VM_BUG_ON(!rcu_read_lock_held());
2289 this_cpu_add(memcg->stat->count[idx], val);
2293 * size of first charge trial. "32" comes from vmscan.c's magic value.
2294 * TODO: maybe necessary to use big numbers in big irons.
2296 #define CHARGE_BATCH 32U
2297 struct memcg_stock_pcp {
2298 struct mem_cgroup *cached; /* this never be root cgroup */
2299 unsigned int nr_pages;
2300 struct work_struct work;
2301 unsigned long flags;
2302 #define FLUSHING_CACHED_CHARGE 0
2304 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2305 static DEFINE_MUTEX(percpu_charge_mutex);
2308 * consume_stock: Try to consume stocked charge on this cpu.
2309 * @memcg: memcg to consume from.
2310 * @nr_pages: how many pages to charge.
2312 * The charges will only happen if @memcg matches the current cpu's memcg
2313 * stock, and at least @nr_pages are available in that stock. Failure to
2314 * service an allocation will refill the stock.
2316 * returns true if successful, false otherwise.
2318 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2320 struct memcg_stock_pcp *stock;
2323 if (nr_pages > CHARGE_BATCH)
2326 stock = &get_cpu_var(memcg_stock);
2327 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2328 stock->nr_pages -= nr_pages;
2331 put_cpu_var(memcg_stock);
2336 * Returns stocks cached in percpu and reset cached information.
2338 static void drain_stock(struct memcg_stock_pcp *stock)
2340 struct mem_cgroup *old = stock->cached;
2342 if (stock->nr_pages) {
2343 page_counter_uncharge(&old->memory, stock->nr_pages);
2344 if (do_swap_account)
2345 page_counter_uncharge(&old->memsw, stock->nr_pages);
2346 stock->nr_pages = 0;
2348 stock->cached = NULL;
2352 * This must be called under preempt disabled or must be called by
2353 * a thread which is pinned to local cpu.
2355 static void drain_local_stock(struct work_struct *dummy)
2357 struct memcg_stock_pcp *stock = this_cpu_ptr(&memcg_stock);
2359 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2362 static void __init memcg_stock_init(void)
2366 for_each_possible_cpu(cpu) {
2367 struct memcg_stock_pcp *stock =
2368 &per_cpu(memcg_stock, cpu);
2369 INIT_WORK(&stock->work, drain_local_stock);
2374 * Cache charges(val) to local per_cpu area.
2375 * This will be consumed by consume_stock() function, later.
2377 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2379 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2381 if (stock->cached != memcg) { /* reset if necessary */
2383 stock->cached = memcg;
2385 stock->nr_pages += nr_pages;
2386 put_cpu_var(memcg_stock);
2390 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2391 * of the hierarchy under it. sync flag says whether we should block
2392 * until the work is done.
2394 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2398 /* Notify other cpus that system-wide "drain" is running */
2401 for_each_online_cpu(cpu) {
2402 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2403 struct mem_cgroup *memcg;
2405 memcg = stock->cached;
2406 if (!memcg || !stock->nr_pages)
2408 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2410 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2412 drain_local_stock(&stock->work);
2414 schedule_work_on(cpu, &stock->work);
2422 for_each_online_cpu(cpu) {
2423 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2424 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2425 flush_work(&stock->work);
2432 * Tries to drain stocked charges in other cpus. This function is asynchronous
2433 * and just put a work per cpu for draining localy on each cpu. Caller can
2434 * expects some charges will be back later but cannot wait for it.
2436 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2439 * If someone calls draining, avoid adding more kworker runs.
2441 if (!mutex_trylock(&percpu_charge_mutex))
2443 drain_all_stock(root_memcg, false);
2444 mutex_unlock(&percpu_charge_mutex);
2447 /* This is a synchronous drain interface. */
2448 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2450 /* called when force_empty is called */
2451 mutex_lock(&percpu_charge_mutex);
2452 drain_all_stock(root_memcg, true);
2453 mutex_unlock(&percpu_charge_mutex);
2457 * This function drains percpu counter value from DEAD cpu and
2458 * move it to local cpu. Note that this function can be preempted.
2460 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2464 spin_lock(&memcg->pcp_counter_lock);
2465 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2466 long x = per_cpu(memcg->stat->count[i], cpu);
2468 per_cpu(memcg->stat->count[i], cpu) = 0;
2469 memcg->nocpu_base.count[i] += x;
2471 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2472 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2474 per_cpu(memcg->stat->events[i], cpu) = 0;
2475 memcg->nocpu_base.events[i] += x;
2477 spin_unlock(&memcg->pcp_counter_lock);
2480 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
2481 unsigned long action,
2484 int cpu = (unsigned long)hcpu;
2485 struct memcg_stock_pcp *stock;
2486 struct mem_cgroup *iter;
2488 if (action == CPU_ONLINE)
2491 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2494 for_each_mem_cgroup(iter)
2495 mem_cgroup_drain_pcp_counter(iter, cpu);
2497 stock = &per_cpu(memcg_stock, cpu);
2502 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2503 unsigned int nr_pages)
2505 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2506 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2507 struct mem_cgroup *mem_over_limit;
2508 struct page_counter *counter;
2509 unsigned long nr_reclaimed;
2510 bool may_swap = true;
2511 bool drained = false;
2514 if (mem_cgroup_is_root(memcg))
2517 if (consume_stock(memcg, nr_pages))
2520 if (!do_swap_account ||
2521 !page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2522 if (!page_counter_try_charge(&memcg->memory, batch, &counter))
2524 if (do_swap_account)
2525 page_counter_uncharge(&memcg->memsw, batch);
2526 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2528 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2532 if (batch > nr_pages) {
2538 * Unlike in global OOM situations, memcg is not in a physical
2539 * memory shortage. Allow dying and OOM-killed tasks to
2540 * bypass the last charges so that they can exit quickly and
2541 * free their memory.
2543 if (unlikely(test_thread_flag(TIF_MEMDIE) ||
2544 fatal_signal_pending(current) ||
2545 current->flags & PF_EXITING))
2548 if (unlikely(task_in_memcg_oom(current)))
2551 if (!(gfp_mask & __GFP_WAIT))
2554 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2555 gfp_mask, may_swap);
2557 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2561 drain_all_stock_async(mem_over_limit);
2566 if (gfp_mask & __GFP_NORETRY)
2569 * Even though the limit is exceeded at this point, reclaim
2570 * may have been able to free some pages. Retry the charge
2571 * before killing the task.
2573 * Only for regular pages, though: huge pages are rather
2574 * unlikely to succeed so close to the limit, and we fall back
2575 * to regular pages anyway in case of failure.
2577 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2580 * At task move, charge accounts can be doubly counted. So, it's
2581 * better to wait until the end of task_move if something is going on.
2583 if (mem_cgroup_wait_acct_move(mem_over_limit))
2589 if (gfp_mask & __GFP_NOFAIL)
2592 if (fatal_signal_pending(current))
2595 mem_cgroup_oom(mem_over_limit, gfp_mask, get_order(nr_pages));
2597 if (!(gfp_mask & __GFP_NOFAIL))
2603 if (batch > nr_pages)
2604 refill_stock(memcg, batch - nr_pages);
2609 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2611 if (mem_cgroup_is_root(memcg))
2614 page_counter_uncharge(&memcg->memory, nr_pages);
2615 if (do_swap_account)
2616 page_counter_uncharge(&memcg->memsw, nr_pages);
2620 * A helper function to get mem_cgroup from ID. must be called under
2621 * rcu_read_lock(). The caller is responsible for calling
2622 * css_tryget_online() if the mem_cgroup is used for charging. (dropping
2623 * refcnt from swap can be called against removed memcg.)
2625 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2627 /* ID 0 is unused ID */
2630 return mem_cgroup_from_id(id);
2634 * try_get_mem_cgroup_from_page - look up page's memcg association
2637 * Look up, get a css reference, and return the memcg that owns @page.
2639 * The page must be locked to prevent racing with swap-in and page
2640 * cache charges. If coming from an unlocked page table, the caller
2641 * must ensure the page is on the LRU or this can race with charging.
2643 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2645 struct mem_cgroup *memcg = NULL;
2646 struct page_cgroup *pc;
2650 VM_BUG_ON_PAGE(!PageLocked(page), page);
2652 pc = lookup_page_cgroup(page);
2653 if (PageCgroupUsed(pc)) {
2654 memcg = pc->mem_cgroup;
2655 if (memcg && !css_tryget_online(&memcg->css))
2657 } else if (PageSwapCache(page)) {
2658 ent.val = page_private(page);
2659 id = lookup_swap_cgroup_id(ent);
2661 memcg = mem_cgroup_lookup(id);
2662 if (memcg && !css_tryget_online(&memcg->css))
2669 static void lock_page_lru(struct page *page, int *isolated)
2671 struct zone *zone = page_zone(page);
2673 spin_lock_irq(&zone->lru_lock);
2674 if (PageLRU(page)) {
2675 struct lruvec *lruvec;
2677 lruvec = mem_cgroup_page_lruvec(page, zone);
2679 del_page_from_lru_list(page, lruvec, page_lru(page));
2685 static void unlock_page_lru(struct page *page, int isolated)
2687 struct zone *zone = page_zone(page);
2690 struct lruvec *lruvec;
2692 lruvec = mem_cgroup_page_lruvec(page, zone);
2693 VM_BUG_ON_PAGE(PageLRU(page), page);
2695 add_page_to_lru_list(page, lruvec, page_lru(page));
2697 spin_unlock_irq(&zone->lru_lock);
2700 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2703 struct page_cgroup *pc = lookup_page_cgroup(page);
2706 VM_BUG_ON_PAGE(PageCgroupUsed(pc), page);
2708 * we don't need page_cgroup_lock about tail pages, becase they are not
2709 * accessed by any other context at this point.
2713 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2714 * may already be on some other mem_cgroup's LRU. Take care of it.
2717 lock_page_lru(page, &isolated);
2720 * Nobody should be changing or seriously looking at
2721 * pc->mem_cgroup and pc->flags at this point:
2723 * - the page is uncharged
2725 * - the page is off-LRU
2727 * - an anonymous fault has exclusive page access, except for
2728 * a locked page table
2730 * - a page cache insertion, a swapin fault, or a migration
2731 * have the page locked
2733 pc->mem_cgroup = memcg;
2734 pc->flags = PCG_USED | PCG_MEM | (do_swap_account ? PCG_MEMSW : 0);
2737 unlock_page_lru(page, isolated);
2740 #ifdef CONFIG_MEMCG_KMEM
2742 * The memcg_slab_mutex is held whenever a per memcg kmem cache is created or
2743 * destroyed. It protects memcg_caches arrays and memcg_slab_caches lists.
2745 static DEFINE_MUTEX(memcg_slab_mutex);
2747 static DEFINE_MUTEX(activate_kmem_mutex);
2750 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2751 * in the memcg_cache_params struct.
2753 static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
2755 struct kmem_cache *cachep;
2757 VM_BUG_ON(p->is_root_cache);
2758 cachep = p->root_cache;
2759 return cache_from_memcg_idx(cachep, memcg_cache_id(p->memcg));
2762 #ifdef CONFIG_SLABINFO
2763 static int mem_cgroup_slabinfo_read(struct seq_file *m, void *v)
2765 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
2766 struct memcg_cache_params *params;
2768 if (!memcg_kmem_is_active(memcg))
2771 print_slabinfo_header(m);
2773 mutex_lock(&memcg_slab_mutex);
2774 list_for_each_entry(params, &memcg->memcg_slab_caches, list)
2775 cache_show(memcg_params_to_cache(params), m);
2776 mutex_unlock(&memcg_slab_mutex);
2782 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp,
2783 unsigned long nr_pages)
2785 struct page_counter *counter;
2788 ret = page_counter_try_charge(&memcg->kmem, nr_pages, &counter);
2792 ret = try_charge(memcg, gfp, nr_pages);
2793 if (ret == -EINTR) {
2795 * try_charge() chose to bypass to root due to OOM kill or
2796 * fatal signal. Since our only options are to either fail
2797 * the allocation or charge it to this cgroup, do it as a
2798 * temporary condition. But we can't fail. From a kmem/slab
2799 * perspective, the cache has already been selected, by
2800 * mem_cgroup_kmem_get_cache(), so it is too late to change
2803 * This condition will only trigger if the task entered
2804 * memcg_charge_kmem in a sane state, but was OOM-killed
2805 * during try_charge() above. Tasks that were already dying
2806 * when the allocation triggers should have been already
2807 * directed to the root cgroup in memcontrol.h
2809 page_counter_charge(&memcg->memory, nr_pages);
2810 if (do_swap_account)
2811 page_counter_charge(&memcg->memsw, nr_pages);
2814 page_counter_uncharge(&memcg->kmem, nr_pages);
2819 static void memcg_uncharge_kmem(struct mem_cgroup *memcg,
2820 unsigned long nr_pages)
2822 page_counter_uncharge(&memcg->memory, nr_pages);
2823 if (do_swap_account)
2824 page_counter_uncharge(&memcg->memsw, nr_pages);
2827 if (page_counter_uncharge(&memcg->kmem, nr_pages))
2831 * Releases a reference taken in kmem_cgroup_css_offline in case
2832 * this last uncharge is racing with the offlining code or it is
2833 * outliving the memcg existence.
2835 * The memory barrier imposed by test&clear is paired with the
2836 * explicit one in memcg_kmem_mark_dead().
2838 if (memcg_kmem_test_and_clear_dead(memcg))
2839 css_put(&memcg->css);
2843 * helper for acessing a memcg's index. It will be used as an index in the
2844 * child cache array in kmem_cache, and also to derive its name. This function
2845 * will return -1 when this is not a kmem-limited memcg.
2847 int memcg_cache_id(struct mem_cgroup *memcg)
2849 return memcg ? memcg->kmemcg_id : -1;
2852 static int memcg_alloc_cache_id(void)
2857 id = ida_simple_get(&kmem_limited_groups,
2858 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2862 if (id < memcg_limited_groups_array_size)
2866 * There's no space for the new id in memcg_caches arrays,
2867 * so we have to grow them.
2870 size = 2 * (id + 1);
2871 if (size < MEMCG_CACHES_MIN_SIZE)
2872 size = MEMCG_CACHES_MIN_SIZE;
2873 else if (size > MEMCG_CACHES_MAX_SIZE)
2874 size = MEMCG_CACHES_MAX_SIZE;
2876 mutex_lock(&memcg_slab_mutex);
2877 err = memcg_update_all_caches(size);
2878 mutex_unlock(&memcg_slab_mutex);
2881 ida_simple_remove(&kmem_limited_groups, id);
2887 static void memcg_free_cache_id(int id)
2889 ida_simple_remove(&kmem_limited_groups, id);
2893 * We should update the current array size iff all caches updates succeed. This
2894 * can only be done from the slab side. The slab mutex needs to be held when
2897 void memcg_update_array_size(int num)
2899 memcg_limited_groups_array_size = num;
2902 static void memcg_register_cache(struct mem_cgroup *memcg,
2903 struct kmem_cache *root_cache)
2905 static char memcg_name_buf[NAME_MAX + 1]; /* protected by
2907 struct kmem_cache *cachep;
2910 lockdep_assert_held(&memcg_slab_mutex);
2912 id = memcg_cache_id(memcg);
2915 * Since per-memcg caches are created asynchronously on first
2916 * allocation (see memcg_kmem_get_cache()), several threads can try to
2917 * create the same cache, but only one of them may succeed.
2919 if (cache_from_memcg_idx(root_cache, id))
2922 cgroup_name(memcg->css.cgroup, memcg_name_buf, NAME_MAX + 1);
2923 cachep = memcg_create_kmem_cache(memcg, root_cache, memcg_name_buf);
2925 * If we could not create a memcg cache, do not complain, because
2926 * that's not critical at all as we can always proceed with the root
2932 css_get(&memcg->css);
2933 list_add(&cachep->memcg_params->list, &memcg->memcg_slab_caches);
2936 * Since readers won't lock (see cache_from_memcg_idx()), we need a
2937 * barrier here to ensure nobody will see the kmem_cache partially
2942 BUG_ON(root_cache->memcg_params->memcg_caches[id]);
2943 root_cache->memcg_params->memcg_caches[id] = cachep;
2946 static void memcg_unregister_cache(struct kmem_cache *cachep)
2948 struct kmem_cache *root_cache;
2949 struct mem_cgroup *memcg;
2952 lockdep_assert_held(&memcg_slab_mutex);
2954 BUG_ON(is_root_cache(cachep));
2956 root_cache = cachep->memcg_params->root_cache;
2957 memcg = cachep->memcg_params->memcg;
2958 id = memcg_cache_id(memcg);
2960 BUG_ON(root_cache->memcg_params->memcg_caches[id] != cachep);
2961 root_cache->memcg_params->memcg_caches[id] = NULL;
2963 list_del(&cachep->memcg_params->list);
2965 kmem_cache_destroy(cachep);
2967 /* drop the reference taken in memcg_register_cache */
2968 css_put(&memcg->css);
2972 * During the creation a new cache, we need to disable our accounting mechanism
2973 * altogether. This is true even if we are not creating, but rather just
2974 * enqueing new caches to be created.
2976 * This is because that process will trigger allocations; some visible, like
2977 * explicit kmallocs to auxiliary data structures, name strings and internal
2978 * cache structures; some well concealed, like INIT_WORK() that can allocate
2979 * objects during debug.
2981 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
2982 * to it. This may not be a bounded recursion: since the first cache creation
2983 * failed to complete (waiting on the allocation), we'll just try to create the
2984 * cache again, failing at the same point.
2986 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
2987 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
2988 * inside the following two functions.
2990 static inline void memcg_stop_kmem_account(void)
2992 VM_BUG_ON(!current->mm);
2993 current->memcg_kmem_skip_account++;
2996 static inline void memcg_resume_kmem_account(void)
2998 VM_BUG_ON(!current->mm);
2999 current->memcg_kmem_skip_account--;
3002 int __memcg_cleanup_cache_params(struct kmem_cache *s)
3004 struct kmem_cache *c;
3007 mutex_lock(&memcg_slab_mutex);
3008 for_each_memcg_cache_index(i) {
3009 c = cache_from_memcg_idx(s, i);
3013 memcg_unregister_cache(c);
3015 if (cache_from_memcg_idx(s, i))
3018 mutex_unlock(&memcg_slab_mutex);
3022 static void memcg_unregister_all_caches(struct mem_cgroup *memcg)
3024 struct kmem_cache *cachep;
3025 struct memcg_cache_params *params, *tmp;
3027 if (!memcg_kmem_is_active(memcg))
3030 mutex_lock(&memcg_slab_mutex);
3031 list_for_each_entry_safe(params, tmp, &memcg->memcg_slab_caches, list) {
3032 cachep = memcg_params_to_cache(params);
3033 kmem_cache_shrink(cachep);
3034 if (atomic_read(&cachep->memcg_params->nr_pages) == 0)
3035 memcg_unregister_cache(cachep);
3037 mutex_unlock(&memcg_slab_mutex);
3040 struct memcg_register_cache_work {
3041 struct mem_cgroup *memcg;
3042 struct kmem_cache *cachep;
3043 struct work_struct work;
3046 static void memcg_register_cache_func(struct work_struct *w)
3048 struct memcg_register_cache_work *cw =
3049 container_of(w, struct memcg_register_cache_work, work);
3050 struct mem_cgroup *memcg = cw->memcg;
3051 struct kmem_cache *cachep = cw->cachep;
3053 mutex_lock(&memcg_slab_mutex);
3054 memcg_register_cache(memcg, cachep);
3055 mutex_unlock(&memcg_slab_mutex);
3057 css_put(&memcg->css);
3062 * Enqueue the creation of a per-memcg kmem_cache.
3064 static void __memcg_schedule_register_cache(struct mem_cgroup *memcg,
3065 struct kmem_cache *cachep)
3067 struct memcg_register_cache_work *cw;
3069 cw = kmalloc(sizeof(*cw), GFP_NOWAIT);
3071 css_put(&memcg->css);
3076 cw->cachep = cachep;
3078 INIT_WORK(&cw->work, memcg_register_cache_func);
3079 schedule_work(&cw->work);
3082 static void memcg_schedule_register_cache(struct mem_cgroup *memcg,
3083 struct kmem_cache *cachep)
3086 * We need to stop accounting when we kmalloc, because if the
3087 * corresponding kmalloc cache is not yet created, the first allocation
3088 * in __memcg_schedule_register_cache will recurse.
3090 * However, it is better to enclose the whole function. Depending on
3091 * the debugging options enabled, INIT_WORK(), for instance, can
3092 * trigger an allocation. This too, will make us recurse. Because at
3093 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3094 * the safest choice is to do it like this, wrapping the whole function.
3096 memcg_stop_kmem_account();
3097 __memcg_schedule_register_cache(memcg, cachep);
3098 memcg_resume_kmem_account();
3101 int __memcg_charge_slab(struct kmem_cache *cachep, gfp_t gfp, int order)
3103 unsigned int nr_pages = 1 << order;
3106 res = memcg_charge_kmem(cachep->memcg_params->memcg, gfp, nr_pages);
3108 atomic_add(nr_pages, &cachep->memcg_params->nr_pages);
3112 void __memcg_uncharge_slab(struct kmem_cache *cachep, int order)
3114 unsigned int nr_pages = 1 << order;
3116 memcg_uncharge_kmem(cachep->memcg_params->memcg, nr_pages);
3117 atomic_sub(nr_pages, &cachep->memcg_params->nr_pages);
3121 * Return the kmem_cache we're supposed to use for a slab allocation.
3122 * We try to use the current memcg's version of the cache.
3124 * If the cache does not exist yet, if we are the first user of it,
3125 * we either create it immediately, if possible, or create it asynchronously
3127 * In the latter case, we will let the current allocation go through with
3128 * the original cache.
3130 * Can't be called in interrupt context or from kernel threads.
3131 * This function needs to be called with rcu_read_lock() held.
3133 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
3136 struct mem_cgroup *memcg;
3137 struct kmem_cache *memcg_cachep;
3139 VM_BUG_ON(!cachep->memcg_params);
3140 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
3142 if (!current->mm || current->memcg_kmem_skip_account)
3146 memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
3148 if (!memcg_kmem_is_active(memcg))
3151 memcg_cachep = cache_from_memcg_idx(cachep, memcg_cache_id(memcg));
3152 if (likely(memcg_cachep)) {
3153 cachep = memcg_cachep;
3157 /* The corresponding put will be done in the workqueue. */
3158 if (!css_tryget_online(&memcg->css))
3163 * If we are in a safe context (can wait, and not in interrupt
3164 * context), we could be be predictable and return right away.
3165 * This would guarantee that the allocation being performed
3166 * already belongs in the new cache.
3168 * However, there are some clashes that can arrive from locking.
3169 * For instance, because we acquire the slab_mutex while doing
3170 * memcg_create_kmem_cache, this means no further allocation
3171 * could happen with the slab_mutex held. So it's better to
3174 memcg_schedule_register_cache(memcg, cachep);
3182 * We need to verify if the allocation against current->mm->owner's memcg is
3183 * possible for the given order. But the page is not allocated yet, so we'll
3184 * need a further commit step to do the final arrangements.
3186 * It is possible for the task to switch cgroups in this mean time, so at
3187 * commit time, we can't rely on task conversion any longer. We'll then use
3188 * the handle argument to return to the caller which cgroup we should commit
3189 * against. We could also return the memcg directly and avoid the pointer
3190 * passing, but a boolean return value gives better semantics considering
3191 * the compiled-out case as well.
3193 * Returning true means the allocation is possible.
3196 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3198 struct mem_cgroup *memcg;
3204 * Disabling accounting is only relevant for some specific memcg
3205 * internal allocations. Therefore we would initially not have such
3206 * check here, since direct calls to the page allocator that are
3207 * accounted to kmemcg (alloc_kmem_pages and friends) only happen
3208 * outside memcg core. We are mostly concerned with cache allocations,
3209 * and by having this test at memcg_kmem_get_cache, we are already able
3210 * to relay the allocation to the root cache and bypass the memcg cache
3213 * There is one exception, though: the SLUB allocator does not create
3214 * large order caches, but rather service large kmallocs directly from
3215 * the page allocator. Therefore, the following sequence when backed by
3216 * the SLUB allocator:
3218 * memcg_stop_kmem_account();
3219 * kmalloc(<large_number>)
3220 * memcg_resume_kmem_account();
3222 * would effectively ignore the fact that we should skip accounting,
3223 * since it will drive us directly to this function without passing
3224 * through the cache selector memcg_kmem_get_cache. Such large
3225 * allocations are extremely rare but can happen, for instance, for the
3226 * cache arrays. We bring this test here.
3228 if (!current->mm || current->memcg_kmem_skip_account)
3231 memcg = get_mem_cgroup_from_mm(current->mm);
3233 if (!memcg_kmem_is_active(memcg)) {
3234 css_put(&memcg->css);
3238 ret = memcg_charge_kmem(memcg, gfp, 1 << order);
3242 css_put(&memcg->css);
3246 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3249 struct page_cgroup *pc;
3251 VM_BUG_ON(mem_cgroup_is_root(memcg));
3253 /* The page allocation failed. Revert */
3255 memcg_uncharge_kmem(memcg, 1 << order);
3259 * The page is freshly allocated and not visible to any
3260 * outside callers yet. Set up pc non-atomically.
3262 pc = lookup_page_cgroup(page);
3263 pc->mem_cgroup = memcg;
3264 pc->flags = PCG_USED;
3267 void __memcg_kmem_uncharge_pages(struct page *page, int order)
3269 struct mem_cgroup *memcg = NULL;
3270 struct page_cgroup *pc;
3273 pc = lookup_page_cgroup(page);
3274 if (!PageCgroupUsed(pc))
3277 memcg = pc->mem_cgroup;
3281 * We trust that only if there is a memcg associated with the page, it
3282 * is a valid allocation
3287 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
3288 memcg_uncharge_kmem(memcg, 1 << order);
3291 static inline void memcg_unregister_all_caches(struct mem_cgroup *memcg)
3294 #endif /* CONFIG_MEMCG_KMEM */
3296 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3299 * Because tail pages are not marked as "used", set it. We're under
3300 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3301 * charge/uncharge will be never happen and move_account() is done under
3302 * compound_lock(), so we don't have to take care of races.
3304 void mem_cgroup_split_huge_fixup(struct page *head)
3306 struct page_cgroup *head_pc = lookup_page_cgroup(head);
3307 struct page_cgroup *pc;
3308 struct mem_cgroup *memcg;
3311 if (mem_cgroup_disabled())
3314 memcg = head_pc->mem_cgroup;
3315 for (i = 1; i < HPAGE_PMD_NR; i++) {
3317 pc->mem_cgroup = memcg;
3318 pc->flags = head_pc->flags;
3320 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
3323 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3326 * mem_cgroup_move_account - move account of the page
3328 * @nr_pages: number of regular pages (>1 for huge pages)
3329 * @pc: page_cgroup of the page.
3330 * @from: mem_cgroup which the page is moved from.
3331 * @to: mem_cgroup which the page is moved to. @from != @to.
3333 * The caller must confirm following.
3334 * - page is not on LRU (isolate_page() is useful.)
3335 * - compound_lock is held when nr_pages > 1
3337 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3340 static int mem_cgroup_move_account(struct page *page,
3341 unsigned int nr_pages,
3342 struct page_cgroup *pc,
3343 struct mem_cgroup *from,
3344 struct mem_cgroup *to)
3346 unsigned long flags;
3349 VM_BUG_ON(from == to);
3350 VM_BUG_ON_PAGE(PageLRU(page), page);
3352 * The page is isolated from LRU. So, collapse function
3353 * will not handle this page. But page splitting can happen.
3354 * Do this check under compound_page_lock(). The caller should
3358 if (nr_pages > 1 && !PageTransHuge(page))
3362 * Prevent mem_cgroup_migrate() from looking at pc->mem_cgroup
3363 * of its source page while we change it: page migration takes
3364 * both pages off the LRU, but page cache replacement doesn't.
3366 if (!trylock_page(page))
3370 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3373 move_lock_mem_cgroup(from, &flags);
3375 if (!PageAnon(page) && page_mapped(page)) {
3376 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
3378 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
3382 if (PageWriteback(page)) {
3383 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK],
3385 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK],
3390 * It is safe to change pc->mem_cgroup here because the page
3391 * is referenced, charged, and isolated - we can't race with
3392 * uncharging, charging, migration, or LRU putback.
3395 /* caller should have done css_get */
3396 pc->mem_cgroup = to;
3397 move_unlock_mem_cgroup(from, &flags);
3400 local_irq_disable();
3401 mem_cgroup_charge_statistics(to, page, nr_pages);
3402 memcg_check_events(to, page);
3403 mem_cgroup_charge_statistics(from, page, -nr_pages);
3404 memcg_check_events(from, page);
3413 * mem_cgroup_move_parent - moves page to the parent group
3414 * @page: the page to move
3415 * @pc: page_cgroup of the page
3416 * @child: page's cgroup
3418 * move charges to its parent or the root cgroup if the group has no
3419 * parent (aka use_hierarchy==0).
3420 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3421 * mem_cgroup_move_account fails) the failure is always temporary and
3422 * it signals a race with a page removal/uncharge or migration. In the
3423 * first case the page is on the way out and it will vanish from the LRU
3424 * on the next attempt and the call should be retried later.
3425 * Isolation from the LRU fails only if page has been isolated from
3426 * the LRU since we looked at it and that usually means either global
3427 * reclaim or migration going on. The page will either get back to the
3429 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3430 * (!PageCgroupUsed) or moved to a different group. The page will
3431 * disappear in the next attempt.
3433 static int mem_cgroup_move_parent(struct page *page,
3434 struct page_cgroup *pc,
3435 struct mem_cgroup *child)
3437 struct mem_cgroup *parent;
3438 unsigned int nr_pages;
3439 unsigned long uninitialized_var(flags);
3442 VM_BUG_ON(mem_cgroup_is_root(child));
3445 if (!get_page_unless_zero(page))
3447 if (isolate_lru_page(page))
3450 nr_pages = hpage_nr_pages(page);
3452 parent = parent_mem_cgroup(child);
3454 * If no parent, move charges to root cgroup.
3457 parent = root_mem_cgroup;
3460 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
3461 flags = compound_lock_irqsave(page);
3464 ret = mem_cgroup_move_account(page, nr_pages,
3467 /* Take charge off the local counters */
3468 page_counter_cancel(&child->memory, nr_pages);
3469 if (do_swap_account)
3470 page_counter_cancel(&child->memsw, nr_pages);
3474 compound_unlock_irqrestore(page, flags);
3475 putback_lru_page(page);
3482 #ifdef CONFIG_MEMCG_SWAP
3483 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
3486 int val = (charge) ? 1 : -1;
3487 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
3491 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3492 * @entry: swap entry to be moved
3493 * @from: mem_cgroup which the entry is moved from
3494 * @to: mem_cgroup which the entry is moved to
3496 * It succeeds only when the swap_cgroup's record for this entry is the same
3497 * as the mem_cgroup's id of @from.
3499 * Returns 0 on success, -EINVAL on failure.
3501 * The caller must have charged to @to, IOW, called page_counter_charge() about
3502 * both res and memsw, and called css_get().
3504 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3505 struct mem_cgroup *from, struct mem_cgroup *to)
3507 unsigned short old_id, new_id;
3509 old_id = mem_cgroup_id(from);
3510 new_id = mem_cgroup_id(to);
3512 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3513 mem_cgroup_swap_statistics(from, false);
3514 mem_cgroup_swap_statistics(to, true);
3516 * This function is only called from task migration context now.
3517 * It postpones page_counter and refcount handling till the end
3518 * of task migration(mem_cgroup_clear_mc()) for performance
3519 * improvement. But we cannot postpone css_get(to) because if
3520 * the process that has been moved to @to does swap-in, the
3521 * refcount of @to might be decreased to 0.
3523 * We are in attach() phase, so the cgroup is guaranteed to be
3524 * alive, so we can just call css_get().
3532 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3533 struct mem_cgroup *from, struct mem_cgroup *to)
3539 #ifdef CONFIG_DEBUG_VM
3540 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3542 struct page_cgroup *pc;
3544 pc = lookup_page_cgroup(page);
3546 * Can be NULL while feeding pages into the page allocator for
3547 * the first time, i.e. during boot or memory hotplug;
3548 * or when mem_cgroup_disabled().
3550 if (likely(pc) && PageCgroupUsed(pc))
3555 bool mem_cgroup_bad_page_check(struct page *page)
3557 if (mem_cgroup_disabled())
3560 return lookup_page_cgroup_used(page) != NULL;
3563 void mem_cgroup_print_bad_page(struct page *page)
3565 struct page_cgroup *pc;
3567 pc = lookup_page_cgroup_used(page);
3569 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
3570 pc, pc->flags, pc->mem_cgroup);
3575 static DEFINE_MUTEX(memcg_limit_mutex);
3577 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3578 unsigned long limit)
3580 unsigned long curusage;
3581 unsigned long oldusage;
3582 bool enlarge = false;
3587 * For keeping hierarchical_reclaim simple, how long we should retry
3588 * is depends on callers. We set our retry-count to be function
3589 * of # of children which we should visit in this loop.
3591 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
3592 mem_cgroup_count_children(memcg);
3594 oldusage = page_counter_read(&memcg->memory);
3597 if (signal_pending(current)) {
3602 mutex_lock(&memcg_limit_mutex);
3603 if (limit > memcg->memsw.limit) {
3604 mutex_unlock(&memcg_limit_mutex);
3608 if (limit > memcg->memory.limit)
3610 ret = page_counter_limit(&memcg->memory, limit);
3611 mutex_unlock(&memcg_limit_mutex);
3616 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, true);
3618 curusage = page_counter_read(&memcg->memory);
3619 /* Usage is reduced ? */
3620 if (curusage >= oldusage)
3623 oldusage = curusage;
3624 } while (retry_count);
3626 if (!ret && enlarge)
3627 memcg_oom_recover(memcg);
3632 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3633 unsigned long limit)
3635 unsigned long curusage;
3636 unsigned long oldusage;
3637 bool enlarge = false;
3641 /* see mem_cgroup_resize_res_limit */
3642 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
3643 mem_cgroup_count_children(memcg);
3645 oldusage = page_counter_read(&memcg->memsw);
3648 if (signal_pending(current)) {
3653 mutex_lock(&memcg_limit_mutex);
3654 if (limit < memcg->memory.limit) {
3655 mutex_unlock(&memcg_limit_mutex);
3659 if (limit > memcg->memsw.limit)
3661 ret = page_counter_limit(&memcg->memsw, limit);
3662 mutex_unlock(&memcg_limit_mutex);
3667 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, false);
3669 curusage = page_counter_read(&memcg->memsw);
3670 /* Usage is reduced ? */
3671 if (curusage >= oldusage)
3674 oldusage = curusage;
3675 } while (retry_count);
3677 if (!ret && enlarge)
3678 memcg_oom_recover(memcg);
3683 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3685 unsigned long *total_scanned)
3687 unsigned long nr_reclaimed = 0;
3688 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3689 unsigned long reclaimed;
3691 struct mem_cgroup_tree_per_zone *mctz;
3692 unsigned long excess;
3693 unsigned long nr_scanned;
3698 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3700 * This loop can run a while, specially if mem_cgroup's continuously
3701 * keep exceeding their soft limit and putting the system under
3708 mz = mem_cgroup_largest_soft_limit_node(mctz);
3713 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
3714 gfp_mask, &nr_scanned);
3715 nr_reclaimed += reclaimed;
3716 *total_scanned += nr_scanned;
3717 spin_lock_irq(&mctz->lock);
3720 * If we failed to reclaim anything from this memory cgroup
3721 * it is time to move on to the next cgroup
3727 * Loop until we find yet another one.
3729 * By the time we get the soft_limit lock
3730 * again, someone might have aded the
3731 * group back on the RB tree. Iterate to
3732 * make sure we get a different mem.
3733 * mem_cgroup_largest_soft_limit_node returns
3734 * NULL if no other cgroup is present on
3738 __mem_cgroup_largest_soft_limit_node(mctz);
3740 css_put(&next_mz->memcg->css);
3741 else /* next_mz == NULL or other memcg */
3745 __mem_cgroup_remove_exceeded(mz, mctz);
3746 excess = soft_limit_excess(mz->memcg);
3748 * One school of thought says that we should not add
3749 * back the node to the tree if reclaim returns 0.
3750 * But our reclaim could return 0, simply because due
3751 * to priority we are exposing a smaller subset of
3752 * memory to reclaim from. Consider this as a longer
3755 /* If excess == 0, no tree ops */
3756 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3757 spin_unlock_irq(&mctz->lock);
3758 css_put(&mz->memcg->css);
3761 * Could not reclaim anything and there are no more
3762 * mem cgroups to try or we seem to be looping without
3763 * reclaiming anything.
3765 if (!nr_reclaimed &&
3767 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3769 } while (!nr_reclaimed);
3771 css_put(&next_mz->memcg->css);
3772 return nr_reclaimed;
3776 * mem_cgroup_force_empty_list - clears LRU of a group
3777 * @memcg: group to clear
3780 * @lru: lru to to clear
3782 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
3783 * reclaim the pages page themselves - pages are moved to the parent (or root)
3786 static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
3787 int node, int zid, enum lru_list lru)
3789 struct lruvec *lruvec;
3790 unsigned long flags;
3791 struct list_head *list;
3795 zone = &NODE_DATA(node)->node_zones[zid];
3796 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
3797 list = &lruvec->lists[lru];
3801 struct page_cgroup *pc;
3804 spin_lock_irqsave(&zone->lru_lock, flags);
3805 if (list_empty(list)) {
3806 spin_unlock_irqrestore(&zone->lru_lock, flags);
3809 page = list_entry(list->prev, struct page, lru);
3811 list_move(&page->lru, list);
3813 spin_unlock_irqrestore(&zone->lru_lock, flags);
3816 spin_unlock_irqrestore(&zone->lru_lock, flags);
3818 pc = lookup_page_cgroup(page);
3820 if (mem_cgroup_move_parent(page, pc, memcg)) {
3821 /* found lock contention or "pc" is obsolete. */
3826 } while (!list_empty(list));
3830 * make mem_cgroup's charge to be 0 if there is no task by moving
3831 * all the charges and pages to the parent.
3832 * This enables deleting this mem_cgroup.
3834 * Caller is responsible for holding css reference on the memcg.
3836 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
3841 /* This is for making all *used* pages to be on LRU. */
3842 lru_add_drain_all();
3843 drain_all_stock_sync(memcg);
3844 mem_cgroup_start_move(memcg);
3845 for_each_node_state(node, N_MEMORY) {
3846 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3849 mem_cgroup_force_empty_list(memcg,
3854 mem_cgroup_end_move(memcg);
3855 memcg_oom_recover(memcg);
3859 * Kernel memory may not necessarily be trackable to a specific
3860 * process. So they are not migrated, and therefore we can't
3861 * expect their value to drop to 0 here.
3862 * Having res filled up with kmem only is enough.
3864 * This is a safety check because mem_cgroup_force_empty_list
3865 * could have raced with mem_cgroup_replace_page_cache callers
3866 * so the lru seemed empty but the page could have been added
3867 * right after the check. RES_USAGE should be safe as we always
3868 * charge before adding to the LRU.
3870 } while (page_counter_read(&memcg->memory) -
3871 page_counter_read(&memcg->kmem) > 0);
3875 * Test whether @memcg has children, dead or alive. Note that this
3876 * function doesn't care whether @memcg has use_hierarchy enabled and
3877 * returns %true if there are child csses according to the cgroup
3878 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
3880 static inline bool memcg_has_children(struct mem_cgroup *memcg)
3885 * The lock does not prevent addition or deletion of children, but
3886 * it prevents a new child from being initialized based on this
3887 * parent in css_online(), so it's enough to decide whether
3888 * hierarchically inherited attributes can still be changed or not.
3890 lockdep_assert_held(&memcg_create_mutex);
3893 ret = css_next_child(NULL, &memcg->css);
3899 * Reclaims as many pages from the given memcg as possible and moves
3900 * the rest to the parent.
3902 * Caller is responsible for holding css reference for memcg.
3904 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3906 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3908 /* we call try-to-free pages for make this cgroup empty */
3909 lru_add_drain_all();
3910 /* try to free all pages in this cgroup */
3911 while (nr_retries && page_counter_read(&memcg->memory)) {
3914 if (signal_pending(current))
3917 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3921 /* maybe some writeback is necessary */
3922 congestion_wait(BLK_RW_ASYNC, HZ/10);
3930 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3931 char *buf, size_t nbytes,
3934 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3936 if (mem_cgroup_is_root(memcg))
3938 return mem_cgroup_force_empty(memcg) ?: nbytes;
3941 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3944 return mem_cgroup_from_css(css)->use_hierarchy;
3947 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3948 struct cftype *cft, u64 val)
3951 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3952 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
3954 mutex_lock(&memcg_create_mutex);
3956 if (memcg->use_hierarchy == val)
3960 * If parent's use_hierarchy is set, we can't make any modifications
3961 * in the child subtrees. If it is unset, then the change can
3962 * occur, provided the current cgroup has no children.
3964 * For the root cgroup, parent_mem is NULL, we allow value to be
3965 * set if there are no children.
3967 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3968 (val == 1 || val == 0)) {
3969 if (!memcg_has_children(memcg))
3970 memcg->use_hierarchy = val;
3977 mutex_unlock(&memcg_create_mutex);
3982 static unsigned long tree_stat(struct mem_cgroup *memcg,
3983 enum mem_cgroup_stat_index idx)
3985 struct mem_cgroup *iter;
3988 /* Per-cpu values can be negative, use a signed accumulator */
3989 for_each_mem_cgroup_tree(iter, memcg)
3990 val += mem_cgroup_read_stat(iter, idx);
3992 if (val < 0) /* race ? */
3997 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
4001 if (mem_cgroup_is_root(memcg)) {
4002 val = tree_stat(memcg, MEM_CGROUP_STAT_CACHE);
4003 val += tree_stat(memcg, MEM_CGROUP_STAT_RSS);
4005 val += tree_stat(memcg, MEM_CGROUP_STAT_SWAP);
4008 val = page_counter_read(&memcg->memory);
4010 val = page_counter_read(&memcg->memsw);
4012 return val << PAGE_SHIFT;
4023 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
4026 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4027 struct page_counter *counter;
4029 switch (MEMFILE_TYPE(cft->private)) {
4031 counter = &memcg->memory;
4034 counter = &memcg->memsw;
4037 counter = &memcg->kmem;
4043 switch (MEMFILE_ATTR(cft->private)) {
4045 if (counter == &memcg->memory)
4046 return mem_cgroup_usage(memcg, false);
4047 if (counter == &memcg->memsw)
4048 return mem_cgroup_usage(memcg, true);
4049 return (u64)page_counter_read(counter) * PAGE_SIZE;
4051 return (u64)counter->limit * PAGE_SIZE;
4053 return (u64)counter->watermark * PAGE_SIZE;
4055 return counter->failcnt;
4056 case RES_SOFT_LIMIT:
4057 return (u64)memcg->soft_limit * PAGE_SIZE;
4063 #ifdef CONFIG_MEMCG_KMEM
4064 /* should be called with activate_kmem_mutex held */
4065 static int __memcg_activate_kmem(struct mem_cgroup *memcg,
4066 unsigned long nr_pages)
4071 if (memcg_kmem_is_active(memcg))
4075 * We are going to allocate memory for data shared by all memory
4076 * cgroups so let's stop accounting here.
4078 memcg_stop_kmem_account();
4081 * For simplicity, we won't allow this to be disabled. It also can't
4082 * be changed if the cgroup has children already, or if tasks had
4085 * If tasks join before we set the limit, a person looking at
4086 * kmem.usage_in_bytes will have no way to determine when it took
4087 * place, which makes the value quite meaningless.
4089 * After it first became limited, changes in the value of the limit are
4090 * of course permitted.
4092 mutex_lock(&memcg_create_mutex);
4093 if (cgroup_has_tasks(memcg->css.cgroup) ||
4094 (memcg->use_hierarchy && memcg_has_children(memcg)))
4096 mutex_unlock(&memcg_create_mutex);
4100 memcg_id = memcg_alloc_cache_id();
4106 memcg->kmemcg_id = memcg_id;
4107 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
4110 * We couldn't have accounted to this cgroup, because it hasn't got the
4111 * active bit set yet, so this should succeed.
4113 err = page_counter_limit(&memcg->kmem, nr_pages);
4116 static_key_slow_inc(&memcg_kmem_enabled_key);
4118 * Setting the active bit after enabling static branching will
4119 * guarantee no one starts accounting before all call sites are
4122 memcg_kmem_set_active(memcg);
4124 memcg_resume_kmem_account();
4128 static int memcg_activate_kmem(struct mem_cgroup *memcg,
4129 unsigned long nr_pages)
4133 mutex_lock(&activate_kmem_mutex);
4134 ret = __memcg_activate_kmem(memcg, nr_pages);
4135 mutex_unlock(&activate_kmem_mutex);
4139 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
4140 unsigned long limit)
4144 mutex_lock(&memcg_limit_mutex);
4145 if (!memcg_kmem_is_active(memcg))
4146 ret = memcg_activate_kmem(memcg, limit);
4148 ret = page_counter_limit(&memcg->kmem, limit);
4149 mutex_unlock(&memcg_limit_mutex);
4153 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
4156 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4161 mutex_lock(&activate_kmem_mutex);
4163 * If the parent cgroup is not kmem-active now, it cannot be activated
4164 * after this point, because it has at least one child already.
4166 if (memcg_kmem_is_active(parent))
4167 ret = __memcg_activate_kmem(memcg, PAGE_COUNTER_MAX);
4168 mutex_unlock(&activate_kmem_mutex);
4172 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
4173 unsigned long limit)
4177 #endif /* CONFIG_MEMCG_KMEM */
4180 * The user of this function is...
4183 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
4184 char *buf, size_t nbytes, loff_t off)
4186 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
4187 unsigned long nr_pages;
4190 buf = strstrip(buf);
4191 ret = page_counter_memparse(buf, &nr_pages);
4195 switch (MEMFILE_ATTR(of_cft(of)->private)) {
4197 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
4201 switch (MEMFILE_TYPE(of_cft(of)->private)) {
4203 ret = mem_cgroup_resize_limit(memcg, nr_pages);
4206 ret = mem_cgroup_resize_memsw_limit(memcg, nr_pages);
4209 ret = memcg_update_kmem_limit(memcg, nr_pages);
4213 case RES_SOFT_LIMIT:
4214 memcg->soft_limit = nr_pages;
4218 return ret ?: nbytes;
4221 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
4222 size_t nbytes, loff_t off)
4224 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
4225 struct page_counter *counter;
4227 switch (MEMFILE_TYPE(of_cft(of)->private)) {
4229 counter = &memcg->memory;
4232 counter = &memcg->memsw;
4235 counter = &memcg->kmem;
4241 switch (MEMFILE_ATTR(of_cft(of)->private)) {
4243 page_counter_reset_watermark(counter);
4246 counter->failcnt = 0;
4255 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
4258 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
4262 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
4263 struct cftype *cft, u64 val)
4265 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4267 if (val >= (1 << NR_MOVE_TYPE))
4271 * No kind of locking is needed in here, because ->can_attach() will
4272 * check this value once in the beginning of the process, and then carry
4273 * on with stale data. This means that changes to this value will only
4274 * affect task migrations starting after the change.
4276 memcg->move_charge_at_immigrate = val;
4280 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
4281 struct cftype *cft, u64 val)
4288 static int memcg_numa_stat_show(struct seq_file *m, void *v)
4292 unsigned int lru_mask;
4295 static const struct numa_stat stats[] = {
4296 { "total", LRU_ALL },
4297 { "file", LRU_ALL_FILE },
4298 { "anon", LRU_ALL_ANON },
4299 { "unevictable", BIT(LRU_UNEVICTABLE) },
4301 const struct numa_stat *stat;
4304 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
4306 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
4307 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
4308 seq_printf(m, "%s=%lu", stat->name, nr);
4309 for_each_node_state(nid, N_MEMORY) {
4310 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4312 seq_printf(m, " N%d=%lu", nid, nr);
4317 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
4318 struct mem_cgroup *iter;
4321 for_each_mem_cgroup_tree(iter, memcg)
4322 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
4323 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
4324 for_each_node_state(nid, N_MEMORY) {
4326 for_each_mem_cgroup_tree(iter, memcg)
4327 nr += mem_cgroup_node_nr_lru_pages(
4328 iter, nid, stat->lru_mask);
4329 seq_printf(m, " N%d=%lu", nid, nr);
4336 #endif /* CONFIG_NUMA */
4338 static inline void mem_cgroup_lru_names_not_uptodate(void)
4340 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
4343 static int memcg_stat_show(struct seq_file *m, void *v)
4345 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
4346 unsigned long memory, memsw;
4347 struct mem_cgroup *mi;
4350 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
4351 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
4353 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
4354 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
4357 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
4358 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
4359 mem_cgroup_read_events(memcg, i));
4361 for (i = 0; i < NR_LRU_LISTS; i++)
4362 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
4363 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
4365 /* Hierarchical information */
4366 memory = memsw = PAGE_COUNTER_MAX;
4367 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
4368 memory = min(memory, mi->memory.limit);
4369 memsw = min(memsw, mi->memsw.limit);
4371 seq_printf(m, "hierarchical_memory_limit %llu\n",
4372 (u64)memory * PAGE_SIZE);
4373 if (do_swap_account)
4374 seq_printf(m, "hierarchical_memsw_limit %llu\n",
4375 (u64)memsw * PAGE_SIZE);
4377 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
4380 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
4382 for_each_mem_cgroup_tree(mi, memcg)
4383 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
4384 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
4387 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
4388 unsigned long long val = 0;
4390 for_each_mem_cgroup_tree(mi, memcg)
4391 val += mem_cgroup_read_events(mi, i);
4392 seq_printf(m, "total_%s %llu\n",
4393 mem_cgroup_events_names[i], val);
4396 for (i = 0; i < NR_LRU_LISTS; i++) {
4397 unsigned long long val = 0;
4399 for_each_mem_cgroup_tree(mi, memcg)
4400 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
4401 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
4404 #ifdef CONFIG_DEBUG_VM
4407 struct mem_cgroup_per_zone *mz;
4408 struct zone_reclaim_stat *rstat;
4409 unsigned long recent_rotated[2] = {0, 0};
4410 unsigned long recent_scanned[2] = {0, 0};
4412 for_each_online_node(nid)
4413 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4414 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
4415 rstat = &mz->lruvec.reclaim_stat;
4417 recent_rotated[0] += rstat->recent_rotated[0];
4418 recent_rotated[1] += rstat->recent_rotated[1];
4419 recent_scanned[0] += rstat->recent_scanned[0];
4420 recent_scanned[1] += rstat->recent_scanned[1];
4422 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
4423 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
4424 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
4425 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
4432 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
4435 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4437 return mem_cgroup_swappiness(memcg);
4440 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
4441 struct cftype *cft, u64 val)
4443 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4449 memcg->swappiness = val;
4451 vm_swappiness = val;
4456 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4458 struct mem_cgroup_threshold_ary *t;
4459 unsigned long usage;
4464 t = rcu_dereference(memcg->thresholds.primary);
4466 t = rcu_dereference(memcg->memsw_thresholds.primary);
4471 usage = mem_cgroup_usage(memcg, swap);
4474 * current_threshold points to threshold just below or equal to usage.
4475 * If it's not true, a threshold was crossed after last
4476 * call of __mem_cgroup_threshold().
4478 i = t->current_threshold;
4481 * Iterate backward over array of thresholds starting from
4482 * current_threshold and check if a threshold is crossed.
4483 * If none of thresholds below usage is crossed, we read
4484 * only one element of the array here.
4486 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4487 eventfd_signal(t->entries[i].eventfd, 1);
4489 /* i = current_threshold + 1 */
4493 * Iterate forward over array of thresholds starting from
4494 * current_threshold+1 and check if a threshold is crossed.
4495 * If none of thresholds above usage is crossed, we read
4496 * only one element of the array here.
4498 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4499 eventfd_signal(t->entries[i].eventfd, 1);
4501 /* Update current_threshold */
4502 t->current_threshold = i - 1;
4507 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4510 __mem_cgroup_threshold(memcg, false);
4511 if (do_swap_account)
4512 __mem_cgroup_threshold(memcg, true);
4514 memcg = parent_mem_cgroup(memcg);
4518 static int compare_thresholds(const void *a, const void *b)
4520 const struct mem_cgroup_threshold *_a = a;
4521 const struct mem_cgroup_threshold *_b = b;
4523 if (_a->threshold > _b->threshold)
4526 if (_a->threshold < _b->threshold)
4532 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4534 struct mem_cgroup_eventfd_list *ev;
4536 spin_lock(&memcg_oom_lock);
4538 list_for_each_entry(ev, &memcg->oom_notify, list)
4539 eventfd_signal(ev->eventfd, 1);
4541 spin_unlock(&memcg_oom_lock);
4545 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4547 struct mem_cgroup *iter;
4549 for_each_mem_cgroup_tree(iter, memcg)
4550 mem_cgroup_oom_notify_cb(iter);
4553 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4554 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4556 struct mem_cgroup_thresholds *thresholds;
4557 struct mem_cgroup_threshold_ary *new;
4558 unsigned long threshold;
4559 unsigned long usage;
4562 ret = page_counter_memparse(args, &threshold);
4566 mutex_lock(&memcg->thresholds_lock);
4569 thresholds = &memcg->thresholds;
4570 usage = mem_cgroup_usage(memcg, false);
4571 } else if (type == _MEMSWAP) {
4572 thresholds = &memcg->memsw_thresholds;
4573 usage = mem_cgroup_usage(memcg, true);
4577 /* Check if a threshold crossed before adding a new one */
4578 if (thresholds->primary)
4579 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4581 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4583 /* Allocate memory for new array of thresholds */
4584 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4592 /* Copy thresholds (if any) to new array */
4593 if (thresholds->primary) {
4594 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4595 sizeof(struct mem_cgroup_threshold));
4598 /* Add new threshold */
4599 new->entries[size - 1].eventfd = eventfd;
4600 new->entries[size - 1].threshold = threshold;
4602 /* Sort thresholds. Registering of new threshold isn't time-critical */
4603 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4604 compare_thresholds, NULL);
4606 /* Find current threshold */
4607 new->current_threshold = -1;
4608 for (i = 0; i < size; i++) {
4609 if (new->entries[i].threshold <= usage) {
4611 * new->current_threshold will not be used until
4612 * rcu_assign_pointer(), so it's safe to increment
4615 ++new->current_threshold;
4620 /* Free old spare buffer and save old primary buffer as spare */
4621 kfree(thresholds->spare);
4622 thresholds->spare = thresholds->primary;
4624 rcu_assign_pointer(thresholds->primary, new);
4626 /* To be sure that nobody uses thresholds */
4630 mutex_unlock(&memcg->thresholds_lock);
4635 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4636 struct eventfd_ctx *eventfd, const char *args)
4638 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4641 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4642 struct eventfd_ctx *eventfd, const char *args)
4644 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4647 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4648 struct eventfd_ctx *eventfd, enum res_type type)
4650 struct mem_cgroup_thresholds *thresholds;
4651 struct mem_cgroup_threshold_ary *new;
4652 unsigned long usage;
4655 mutex_lock(&memcg->thresholds_lock);
4658 thresholds = &memcg->thresholds;
4659 usage = mem_cgroup_usage(memcg, false);
4660 } else if (type == _MEMSWAP) {
4661 thresholds = &memcg->memsw_thresholds;
4662 usage = mem_cgroup_usage(memcg, true);
4666 if (!thresholds->primary)
4669 /* Check if a threshold crossed before removing */
4670 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4672 /* Calculate new number of threshold */
4674 for (i = 0; i < thresholds->primary->size; i++) {
4675 if (thresholds->primary->entries[i].eventfd != eventfd)
4679 new = thresholds->spare;
4681 /* Set thresholds array to NULL if we don't have thresholds */
4690 /* Copy thresholds and find current threshold */
4691 new->current_threshold = -1;
4692 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4693 if (thresholds->primary->entries[i].eventfd == eventfd)
4696 new->entries[j] = thresholds->primary->entries[i];
4697 if (new->entries[j].threshold <= usage) {
4699 * new->current_threshold will not be used
4700 * until rcu_assign_pointer(), so it's safe to increment
4703 ++new->current_threshold;
4709 /* Swap primary and spare array */
4710 thresholds->spare = thresholds->primary;
4711 /* If all events are unregistered, free the spare array */
4713 kfree(thresholds->spare);
4714 thresholds->spare = NULL;
4717 rcu_assign_pointer(thresholds->primary, new);
4719 /* To be sure that nobody uses thresholds */
4722 mutex_unlock(&memcg->thresholds_lock);
4725 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4726 struct eventfd_ctx *eventfd)
4728 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4731 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4732 struct eventfd_ctx *eventfd)
4734 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4737 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4738 struct eventfd_ctx *eventfd, const char *args)
4740 struct mem_cgroup_eventfd_list *event;
4742 event = kmalloc(sizeof(*event), GFP_KERNEL);
4746 spin_lock(&memcg_oom_lock);
4748 event->eventfd = eventfd;
4749 list_add(&event->list, &memcg->oom_notify);
4751 /* already in OOM ? */
4752 if (atomic_read(&memcg->under_oom))
4753 eventfd_signal(eventfd, 1);
4754 spin_unlock(&memcg_oom_lock);
4759 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4760 struct eventfd_ctx *eventfd)
4762 struct mem_cgroup_eventfd_list *ev, *tmp;
4764 spin_lock(&memcg_oom_lock);
4766 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4767 if (ev->eventfd == eventfd) {
4768 list_del(&ev->list);
4773 spin_unlock(&memcg_oom_lock);
4776 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4778 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
4780 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4781 seq_printf(sf, "under_oom %d\n", (bool)atomic_read(&memcg->under_oom));
4785 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4786 struct cftype *cft, u64 val)
4788 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4790 /* cannot set to root cgroup and only 0 and 1 are allowed */
4791 if (!css->parent || !((val == 0) || (val == 1)))
4794 memcg->oom_kill_disable = val;
4796 memcg_oom_recover(memcg);
4801 #ifdef CONFIG_MEMCG_KMEM
4802 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
4806 memcg->kmemcg_id = -1;
4807 ret = memcg_propagate_kmem(memcg);
4811 return mem_cgroup_sockets_init(memcg, ss);
4814 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
4816 mem_cgroup_sockets_destroy(memcg);
4819 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
4821 if (!memcg_kmem_is_active(memcg))
4825 * kmem charges can outlive the cgroup. In the case of slab
4826 * pages, for instance, a page contain objects from various
4827 * processes. As we prevent from taking a reference for every
4828 * such allocation we have to be careful when doing uncharge
4829 * (see memcg_uncharge_kmem) and here during offlining.
4831 * The idea is that that only the _last_ uncharge which sees
4832 * the dead memcg will drop the last reference. An additional
4833 * reference is taken here before the group is marked dead
4834 * which is then paired with css_put during uncharge resp. here.
4836 * Although this might sound strange as this path is called from
4837 * css_offline() when the referencemight have dropped down to 0 and
4838 * shouldn't be incremented anymore (css_tryget_online() would
4839 * fail) we do not have other options because of the kmem
4840 * allocations lifetime.
4842 css_get(&memcg->css);
4844 memcg_kmem_mark_dead(memcg);
4846 if (page_counter_read(&memcg->kmem))
4849 if (memcg_kmem_test_and_clear_dead(memcg))
4850 css_put(&memcg->css);
4853 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
4858 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
4862 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
4868 * DO NOT USE IN NEW FILES.
4870 * "cgroup.event_control" implementation.
4872 * This is way over-engineered. It tries to support fully configurable
4873 * events for each user. Such level of flexibility is completely
4874 * unnecessary especially in the light of the planned unified hierarchy.
4876 * Please deprecate this and replace with something simpler if at all
4881 * Unregister event and free resources.
4883 * Gets called from workqueue.
4885 static void memcg_event_remove(struct work_struct *work)
4887 struct mem_cgroup_event *event =
4888 container_of(work, struct mem_cgroup_event, remove);
4889 struct mem_cgroup *memcg = event->memcg;
4891 remove_wait_queue(event->wqh, &event->wait);
4893 event->unregister_event(memcg, event->eventfd);
4895 /* Notify userspace the event is going away. */
4896 eventfd_signal(event->eventfd, 1);
4898 eventfd_ctx_put(event->eventfd);
4900 css_put(&memcg->css);
4904 * Gets called on POLLHUP on eventfd when user closes it.
4906 * Called with wqh->lock held and interrupts disabled.
4908 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
4909 int sync, void *key)
4911 struct mem_cgroup_event *event =
4912 container_of(wait, struct mem_cgroup_event, wait);
4913 struct mem_cgroup *memcg = event->memcg;
4914 unsigned long flags = (unsigned long)key;
4916 if (flags & POLLHUP) {
4918 * If the event has been detached at cgroup removal, we
4919 * can simply return knowing the other side will cleanup
4922 * We can't race against event freeing since the other
4923 * side will require wqh->lock via remove_wait_queue(),
4926 spin_lock(&memcg->event_list_lock);
4927 if (!list_empty(&event->list)) {
4928 list_del_init(&event->list);
4930 * We are in atomic context, but cgroup_event_remove()
4931 * may sleep, so we have to call it in workqueue.
4933 schedule_work(&event->remove);
4935 spin_unlock(&memcg->event_list_lock);
4941 static void memcg_event_ptable_queue_proc(struct file *file,
4942 wait_queue_head_t *wqh, poll_table *pt)
4944 struct mem_cgroup_event *event =
4945 container_of(pt, struct mem_cgroup_event, pt);
4948 add_wait_queue(wqh, &event->wait);
4952 * DO NOT USE IN NEW FILES.
4954 * Parse input and register new cgroup event handler.
4956 * Input must be in format '<event_fd> <control_fd> <args>'.
4957 * Interpretation of args is defined by control file implementation.
4959 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4960 char *buf, size_t nbytes, loff_t off)
4962 struct cgroup_subsys_state *css = of_css(of);
4963 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4964 struct mem_cgroup_event *event;
4965 struct cgroup_subsys_state *cfile_css;
4966 unsigned int efd, cfd;
4973 buf = strstrip(buf);
4975 efd = simple_strtoul(buf, &endp, 10);
4980 cfd = simple_strtoul(buf, &endp, 10);
4981 if ((*endp != ' ') && (*endp != '\0'))
4985 event = kzalloc(sizeof(*event), GFP_KERNEL);
4989 event->memcg = memcg;
4990 INIT_LIST_HEAD(&event->list);
4991 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4992 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4993 INIT_WORK(&event->remove, memcg_event_remove);
5001 event->eventfd = eventfd_ctx_fileget(efile.file);
5002 if (IS_ERR(event->eventfd)) {
5003 ret = PTR_ERR(event->eventfd);
5010 goto out_put_eventfd;
5013 /* the process need read permission on control file */
5014 /* AV: shouldn't we check that it's been opened for read instead? */
5015 ret = inode_permission(file_inode(cfile.file), MAY_READ);
5020 * Determine the event callbacks and set them in @event. This used
5021 * to be done via struct cftype but cgroup core no longer knows
5022 * about these events. The following is crude but the whole thing
5023 * is for compatibility anyway.
5025 * DO NOT ADD NEW FILES.
5027 name = cfile.file->f_dentry->d_name.name;
5029 if (!strcmp(name, "memory.usage_in_bytes")) {
5030 event->register_event = mem_cgroup_usage_register_event;
5031 event->unregister_event = mem_cgroup_usage_unregister_event;
5032 } else if (!strcmp(name, "memory.oom_control")) {
5033 event->register_event = mem_cgroup_oom_register_event;
5034 event->unregister_event = mem_cgroup_oom_unregister_event;
5035 } else if (!strcmp(name, "memory.pressure_level")) {
5036 event->register_event = vmpressure_register_event;
5037 event->unregister_event = vmpressure_unregister_event;
5038 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
5039 event->register_event = memsw_cgroup_usage_register_event;
5040 event->unregister_event = memsw_cgroup_usage_unregister_event;
5047 * Verify @cfile should belong to @css. Also, remaining events are
5048 * automatically removed on cgroup destruction but the removal is
5049 * asynchronous, so take an extra ref on @css.
5051 cfile_css = css_tryget_online_from_dir(cfile.file->f_dentry->d_parent,
5052 &memory_cgrp_subsys);
5054 if (IS_ERR(cfile_css))
5056 if (cfile_css != css) {
5061 ret = event->register_event(memcg, event->eventfd, buf);
5065 efile.file->f_op->poll(efile.file, &event->pt);
5067 spin_lock(&memcg->event_list_lock);
5068 list_add(&event->list, &memcg->event_list);
5069 spin_unlock(&memcg->event_list_lock);
5081 eventfd_ctx_put(event->eventfd);
5090 static struct cftype mem_cgroup_files[] = {
5092 .name = "usage_in_bytes",
5093 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
5094 .read_u64 = mem_cgroup_read_u64,
5097 .name = "max_usage_in_bytes",
5098 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
5099 .write = mem_cgroup_reset,
5100 .read_u64 = mem_cgroup_read_u64,
5103 .name = "limit_in_bytes",
5104 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
5105 .write = mem_cgroup_write,
5106 .read_u64 = mem_cgroup_read_u64,
5109 .name = "soft_limit_in_bytes",
5110 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
5111 .write = mem_cgroup_write,
5112 .read_u64 = mem_cgroup_read_u64,
5116 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
5117 .write = mem_cgroup_reset,
5118 .read_u64 = mem_cgroup_read_u64,
5122 .seq_show = memcg_stat_show,
5125 .name = "force_empty",
5126 .write = mem_cgroup_force_empty_write,
5129 .name = "use_hierarchy",
5130 .write_u64 = mem_cgroup_hierarchy_write,
5131 .read_u64 = mem_cgroup_hierarchy_read,
5134 .name = "cgroup.event_control", /* XXX: for compat */
5135 .write = memcg_write_event_control,
5136 .flags = CFTYPE_NO_PREFIX,
5140 .name = "swappiness",
5141 .read_u64 = mem_cgroup_swappiness_read,
5142 .write_u64 = mem_cgroup_swappiness_write,
5145 .name = "move_charge_at_immigrate",
5146 .read_u64 = mem_cgroup_move_charge_read,
5147 .write_u64 = mem_cgroup_move_charge_write,
5150 .name = "oom_control",
5151 .seq_show = mem_cgroup_oom_control_read,
5152 .write_u64 = mem_cgroup_oom_control_write,
5153 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
5156 .name = "pressure_level",
5160 .name = "numa_stat",
5161 .seq_show = memcg_numa_stat_show,
5164 #ifdef CONFIG_MEMCG_KMEM
5166 .name = "kmem.limit_in_bytes",
5167 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
5168 .write = mem_cgroup_write,
5169 .read_u64 = mem_cgroup_read_u64,
5172 .name = "kmem.usage_in_bytes",
5173 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5174 .read_u64 = mem_cgroup_read_u64,
5177 .name = "kmem.failcnt",
5178 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5179 .write = mem_cgroup_reset,
5180 .read_u64 = mem_cgroup_read_u64,
5183 .name = "kmem.max_usage_in_bytes",
5184 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5185 .write = mem_cgroup_reset,
5186 .read_u64 = mem_cgroup_read_u64,
5188 #ifdef CONFIG_SLABINFO
5190 .name = "kmem.slabinfo",
5191 .seq_show = mem_cgroup_slabinfo_read,
5195 { }, /* terminate */
5198 #ifdef CONFIG_MEMCG_SWAP
5199 static struct cftype memsw_cgroup_files[] = {
5201 .name = "memsw.usage_in_bytes",
5202 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
5203 .read_u64 = mem_cgroup_read_u64,
5206 .name = "memsw.max_usage_in_bytes",
5207 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
5208 .write = mem_cgroup_reset,
5209 .read_u64 = mem_cgroup_read_u64,
5212 .name = "memsw.limit_in_bytes",
5213 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
5214 .write = mem_cgroup_write,
5215 .read_u64 = mem_cgroup_read_u64,
5218 .name = "memsw.failcnt",
5219 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
5220 .write = mem_cgroup_reset,
5221 .read_u64 = mem_cgroup_read_u64,
5223 { }, /* terminate */
5226 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
5228 struct mem_cgroup_per_node *pn;
5229 struct mem_cgroup_per_zone *mz;
5230 int zone, tmp = node;
5232 * This routine is called against possible nodes.
5233 * But it's BUG to call kmalloc() against offline node.
5235 * TODO: this routine can waste much memory for nodes which will
5236 * never be onlined. It's better to use memory hotplug callback
5239 if (!node_state(node, N_NORMAL_MEMORY))
5241 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5245 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5246 mz = &pn->zoneinfo[zone];
5247 lruvec_init(&mz->lruvec);
5248 mz->usage_in_excess = 0;
5249 mz->on_tree = false;
5252 memcg->nodeinfo[node] = pn;
5256 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
5258 kfree(memcg->nodeinfo[node]);
5261 static struct mem_cgroup *mem_cgroup_alloc(void)
5263 struct mem_cgroup *memcg;
5266 size = sizeof(struct mem_cgroup);
5267 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
5269 memcg = kzalloc(size, GFP_KERNEL);
5273 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
5276 spin_lock_init(&memcg->pcp_counter_lock);
5285 * At destroying mem_cgroup, references from swap_cgroup can remain.
5286 * (scanning all at force_empty is too costly...)
5288 * Instead of clearing all references at force_empty, we remember
5289 * the number of reference from swap_cgroup and free mem_cgroup when
5290 * it goes down to 0.
5292 * Removal of cgroup itself succeeds regardless of refs from swap.
5295 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5299 mem_cgroup_remove_from_trees(memcg);
5302 free_mem_cgroup_per_zone_info(memcg, node);
5304 free_percpu(memcg->stat);
5307 * We need to make sure that (at least for now), the jump label
5308 * destruction code runs outside of the cgroup lock. This is because
5309 * get_online_cpus(), which is called from the static_branch update,
5310 * can't be called inside the cgroup_lock. cpusets are the ones
5311 * enforcing this dependency, so if they ever change, we might as well.
5313 * schedule_work() will guarantee this happens. Be careful if you need
5314 * to move this code around, and make sure it is outside
5317 disarm_static_keys(memcg);
5322 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
5324 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
5326 if (!memcg->memory.parent)
5328 return mem_cgroup_from_counter(memcg->memory.parent, memory);
5330 EXPORT_SYMBOL(parent_mem_cgroup);
5332 static void __init mem_cgroup_soft_limit_tree_init(void)
5334 struct mem_cgroup_tree_per_node *rtpn;
5335 struct mem_cgroup_tree_per_zone *rtpz;
5336 int tmp, node, zone;
5338 for_each_node(node) {
5340 if (!node_state(node, N_NORMAL_MEMORY))
5342 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
5345 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5347 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5348 rtpz = &rtpn->rb_tree_per_zone[zone];
5349 rtpz->rb_root = RB_ROOT;
5350 spin_lock_init(&rtpz->lock);
5355 static struct cgroup_subsys_state * __ref
5356 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5358 struct mem_cgroup *memcg;
5359 long error = -ENOMEM;
5362 memcg = mem_cgroup_alloc();
5364 return ERR_PTR(error);
5367 if (alloc_mem_cgroup_per_zone_info(memcg, node))
5371 if (parent_css == NULL) {
5372 root_mem_cgroup = memcg;
5373 page_counter_init(&memcg->memory, NULL);
5374 page_counter_init(&memcg->memsw, NULL);
5375 page_counter_init(&memcg->kmem, NULL);
5378 memcg->last_scanned_node = MAX_NUMNODES;
5379 INIT_LIST_HEAD(&memcg->oom_notify);
5380 memcg->move_charge_at_immigrate = 0;
5381 mutex_init(&memcg->thresholds_lock);
5382 spin_lock_init(&memcg->move_lock);
5383 vmpressure_init(&memcg->vmpressure);
5384 INIT_LIST_HEAD(&memcg->event_list);
5385 spin_lock_init(&memcg->event_list_lock);
5390 __mem_cgroup_free(memcg);
5391 return ERR_PTR(error);
5395 mem_cgroup_css_online(struct cgroup_subsys_state *css)
5397 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5398 struct mem_cgroup *parent = mem_cgroup_from_css(css->parent);
5401 if (css->id > MEM_CGROUP_ID_MAX)
5407 mutex_lock(&memcg_create_mutex);
5409 memcg->use_hierarchy = parent->use_hierarchy;
5410 memcg->oom_kill_disable = parent->oom_kill_disable;
5411 memcg->swappiness = mem_cgroup_swappiness(parent);
5413 if (parent->use_hierarchy) {
5414 page_counter_init(&memcg->memory, &parent->memory);
5415 page_counter_init(&memcg->memsw, &parent->memsw);
5416 page_counter_init(&memcg->kmem, &parent->kmem);
5419 * No need to take a reference to the parent because cgroup
5420 * core guarantees its existence.
5423 page_counter_init(&memcg->memory, NULL);
5424 page_counter_init(&memcg->memsw, NULL);
5425 page_counter_init(&memcg->kmem, NULL);
5427 * Deeper hierachy with use_hierarchy == false doesn't make
5428 * much sense so let cgroup subsystem know about this
5429 * unfortunate state in our controller.
5431 if (parent != root_mem_cgroup)
5432 memory_cgrp_subsys.broken_hierarchy = true;
5434 mutex_unlock(&memcg_create_mutex);
5436 ret = memcg_init_kmem(memcg, &memory_cgrp_subsys);
5441 * Make sure the memcg is initialized: mem_cgroup_iter()
5442 * orders reading memcg->initialized against its callers
5443 * reading the memcg members.
5445 smp_store_release(&memcg->initialized, 1);
5451 * Announce all parents that a group from their hierarchy is gone.
5453 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup *memcg)
5455 struct mem_cgroup *parent = memcg;
5457 while ((parent = parent_mem_cgroup(parent)))
5458 mem_cgroup_iter_invalidate(parent);
5461 * if the root memcg is not hierarchical we have to check it
5464 if (!root_mem_cgroup->use_hierarchy)
5465 mem_cgroup_iter_invalidate(root_mem_cgroup);
5468 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5470 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5471 struct mem_cgroup_event *event, *tmp;
5472 struct cgroup_subsys_state *iter;
5475 * Unregister events and notify userspace.
5476 * Notify userspace about cgroup removing only after rmdir of cgroup
5477 * directory to avoid race between userspace and kernelspace.
5479 spin_lock(&memcg->event_list_lock);
5480 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5481 list_del_init(&event->list);
5482 schedule_work(&event->remove);
5484 spin_unlock(&memcg->event_list_lock);
5486 kmem_cgroup_css_offline(memcg);
5488 mem_cgroup_invalidate_reclaim_iterators(memcg);
5491 * This requires that offlining is serialized. Right now that is
5492 * guaranteed because css_killed_work_fn() holds the cgroup_mutex.
5494 css_for_each_descendant_post(iter, css)
5495 mem_cgroup_reparent_charges(mem_cgroup_from_css(iter));
5497 memcg_unregister_all_caches(memcg);
5498 vmpressure_cleanup(&memcg->vmpressure);
5501 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5503 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5505 * XXX: css_offline() would be where we should reparent all
5506 * memory to prepare the cgroup for destruction. However,
5507 * memcg does not do css_tryget_online() and page_counter charging
5508 * under the same RCU lock region, which means that charging
5509 * could race with offlining. Offlining only happens to
5510 * cgroups with no tasks in them but charges can show up
5511 * without any tasks from the swapin path when the target
5512 * memcg is looked up from the swapout record and not from the
5513 * current task as it usually is. A race like this can leak
5514 * charges and put pages with stale cgroup pointers into
5518 * lookup_swap_cgroup_id()
5520 * mem_cgroup_lookup()
5521 * css_tryget_online()
5523 * disable css_tryget_online()
5526 * reparent_charges()
5527 * page_counter_try_charge()
5530 * pc->mem_cgroup = dead memcg
5533 * The bulk of the charges are still moved in offline_css() to
5534 * avoid pinning a lot of pages in case a long-term reference
5535 * like a swapout record is deferring the css_free() to long
5536 * after offlining. But this makes sure we catch any charges
5537 * made after offlining:
5539 mem_cgroup_reparent_charges(memcg);
5541 memcg_destroy_kmem(memcg);
5542 __mem_cgroup_free(memcg);
5546 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5547 * @css: the target css
5549 * Reset the states of the mem_cgroup associated with @css. This is
5550 * invoked when the userland requests disabling on the default hierarchy
5551 * but the memcg is pinned through dependency. The memcg should stop
5552 * applying policies and should revert to the vanilla state as it may be
5553 * made visible again.
5555 * The current implementation only resets the essential configurations.
5556 * This needs to be expanded to cover all the visible parts.
5558 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5560 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5562 mem_cgroup_resize_limit(memcg, PAGE_COUNTER_MAX);
5563 mem_cgroup_resize_memsw_limit(memcg, PAGE_COUNTER_MAX);
5564 memcg_update_kmem_limit(memcg, PAGE_COUNTER_MAX);
5565 memcg->soft_limit = 0;
5569 /* Handlers for move charge at task migration. */
5570 static int mem_cgroup_do_precharge(unsigned long count)
5574 /* Try a single bulk charge without reclaim first */
5575 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_WAIT, count);
5577 mc.precharge += count;
5580 if (ret == -EINTR) {
5581 cancel_charge(root_mem_cgroup, count);
5585 /* Try charges one by one with reclaim */
5587 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_NORETRY, 1);
5589 * In case of failure, any residual charges against
5590 * mc.to will be dropped by mem_cgroup_clear_mc()
5591 * later on. However, cancel any charges that are
5592 * bypassed to root right away or they'll be lost.
5595 cancel_charge(root_mem_cgroup, 1);
5605 * get_mctgt_type - get target type of moving charge
5606 * @vma: the vma the pte to be checked belongs
5607 * @addr: the address corresponding to the pte to be checked
5608 * @ptent: the pte to be checked
5609 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5612 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5613 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5614 * move charge. if @target is not NULL, the page is stored in target->page
5615 * with extra refcnt got(Callers should handle it).
5616 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5617 * target for charge migration. if @target is not NULL, the entry is stored
5620 * Called with pte lock held.
5627 enum mc_target_type {
5633 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5634 unsigned long addr, pte_t ptent)
5636 struct page *page = vm_normal_page(vma, addr, ptent);
5638 if (!page || !page_mapped(page))
5640 if (PageAnon(page)) {
5641 /* we don't move shared anon */
5644 } else if (!move_file())
5645 /* we ignore mapcount for file pages */
5647 if (!get_page_unless_zero(page))
5654 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5655 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5657 struct page *page = NULL;
5658 swp_entry_t ent = pte_to_swp_entry(ptent);
5660 if (!move_anon() || non_swap_entry(ent))
5663 * Because lookup_swap_cache() updates some statistics counter,
5664 * we call find_get_page() with swapper_space directly.
5666 page = find_get_page(swap_address_space(ent), ent.val);
5667 if (do_swap_account)
5668 entry->val = ent.val;
5673 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5674 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5680 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5681 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5683 struct page *page = NULL;
5684 struct address_space *mapping;
5687 if (!vma->vm_file) /* anonymous vma */
5692 mapping = vma->vm_file->f_mapping;
5693 if (pte_none(ptent))
5694 pgoff = linear_page_index(vma, addr);
5695 else /* pte_file(ptent) is true */
5696 pgoff = pte_to_pgoff(ptent);
5698 /* page is moved even if it's not RSS of this task(page-faulted). */
5700 /* shmem/tmpfs may report page out on swap: account for that too. */
5701 if (shmem_mapping(mapping)) {
5702 page = find_get_entry(mapping, pgoff);
5703 if (radix_tree_exceptional_entry(page)) {
5704 swp_entry_t swp = radix_to_swp_entry(page);
5705 if (do_swap_account)
5707 page = find_get_page(swap_address_space(swp), swp.val);
5710 page = find_get_page(mapping, pgoff);
5712 page = find_get_page(mapping, pgoff);
5717 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5718 unsigned long addr, pte_t ptent, union mc_target *target)
5720 struct page *page = NULL;
5721 struct page_cgroup *pc;
5722 enum mc_target_type ret = MC_TARGET_NONE;
5723 swp_entry_t ent = { .val = 0 };
5725 if (pte_present(ptent))
5726 page = mc_handle_present_pte(vma, addr, ptent);
5727 else if (is_swap_pte(ptent))
5728 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5729 else if (pte_none(ptent) || pte_file(ptent))
5730 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5732 if (!page && !ent.val)
5735 pc = lookup_page_cgroup(page);
5737 * Do only loose check w/o serialization.
5738 * mem_cgroup_move_account() checks the pc is valid or
5739 * not under LRU exclusion.
5741 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5742 ret = MC_TARGET_PAGE;
5744 target->page = page;
5746 if (!ret || !target)
5749 /* There is a swap entry and a page doesn't exist or isn't charged */
5750 if (ent.val && !ret &&
5751 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5752 ret = MC_TARGET_SWAP;
5759 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5761 * We don't consider swapping or file mapped pages because THP does not
5762 * support them for now.
5763 * Caller should make sure that pmd_trans_huge(pmd) is true.
5765 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5766 unsigned long addr, pmd_t pmd, union mc_target *target)
5768 struct page *page = NULL;
5769 struct page_cgroup *pc;
5770 enum mc_target_type ret = MC_TARGET_NONE;
5772 page = pmd_page(pmd);
5773 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5776 pc = lookup_page_cgroup(page);
5777 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5778 ret = MC_TARGET_PAGE;
5781 target->page = page;
5787 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5788 unsigned long addr, pmd_t pmd, union mc_target *target)
5790 return MC_TARGET_NONE;
5794 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5795 unsigned long addr, unsigned long end,
5796 struct mm_walk *walk)
5798 struct vm_area_struct *vma = walk->private;
5802 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
5803 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5804 mc.precharge += HPAGE_PMD_NR;
5809 if (pmd_trans_unstable(pmd))
5811 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5812 for (; addr != end; pte++, addr += PAGE_SIZE)
5813 if (get_mctgt_type(vma, addr, *pte, NULL))
5814 mc.precharge++; /* increment precharge temporarily */
5815 pte_unmap_unlock(pte - 1, ptl);
5821 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5823 unsigned long precharge;
5824 struct vm_area_struct *vma;
5826 down_read(&mm->mmap_sem);
5827 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5828 struct mm_walk mem_cgroup_count_precharge_walk = {
5829 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5833 if (is_vm_hugetlb_page(vma))
5835 walk_page_range(vma->vm_start, vma->vm_end,
5836 &mem_cgroup_count_precharge_walk);
5838 up_read(&mm->mmap_sem);
5840 precharge = mc.precharge;
5846 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5848 unsigned long precharge = mem_cgroup_count_precharge(mm);
5850 VM_BUG_ON(mc.moving_task);
5851 mc.moving_task = current;
5852 return mem_cgroup_do_precharge(precharge);
5855 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5856 static void __mem_cgroup_clear_mc(void)
5858 struct mem_cgroup *from = mc.from;
5859 struct mem_cgroup *to = mc.to;
5862 /* we must uncharge all the leftover precharges from mc.to */
5864 cancel_charge(mc.to, mc.precharge);
5868 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5869 * we must uncharge here.
5871 if (mc.moved_charge) {
5872 cancel_charge(mc.from, mc.moved_charge);
5873 mc.moved_charge = 0;
5875 /* we must fixup refcnts and charges */
5876 if (mc.moved_swap) {
5877 /* uncharge swap account from the old cgroup */
5878 if (!mem_cgroup_is_root(mc.from))
5879 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5882 * we charged both to->memory and to->memsw, so we
5883 * should uncharge to->memory.
5885 if (!mem_cgroup_is_root(mc.to))
5886 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5888 for (i = 0; i < mc.moved_swap; i++)
5889 css_put(&mc.from->css);
5891 /* we've already done css_get(mc.to) */
5894 memcg_oom_recover(from);
5895 memcg_oom_recover(to);
5896 wake_up_all(&mc.waitq);
5899 static void mem_cgroup_clear_mc(void)
5901 struct mem_cgroup *from = mc.from;
5904 * we must clear moving_task before waking up waiters at the end of
5907 mc.moving_task = NULL;
5908 __mem_cgroup_clear_mc();
5909 spin_lock(&mc.lock);
5912 spin_unlock(&mc.lock);
5913 mem_cgroup_end_move(from);
5916 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
5917 struct cgroup_taskset *tset)
5919 struct task_struct *p = cgroup_taskset_first(tset);
5921 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5922 unsigned long move_charge_at_immigrate;
5925 * We are now commited to this value whatever it is. Changes in this
5926 * tunable will only affect upcoming migrations, not the current one.
5927 * So we need to save it, and keep it going.
5929 move_charge_at_immigrate = memcg->move_charge_at_immigrate;
5930 if (move_charge_at_immigrate) {
5931 struct mm_struct *mm;
5932 struct mem_cgroup *from = mem_cgroup_from_task(p);
5934 VM_BUG_ON(from == memcg);
5936 mm = get_task_mm(p);
5939 /* We move charges only when we move a owner of the mm */
5940 if (mm->owner == p) {
5943 VM_BUG_ON(mc.precharge);
5944 VM_BUG_ON(mc.moved_charge);
5945 VM_BUG_ON(mc.moved_swap);
5946 mem_cgroup_start_move(from);
5947 spin_lock(&mc.lock);
5950 mc.immigrate_flags = move_charge_at_immigrate;
5951 spin_unlock(&mc.lock);
5952 /* We set mc.moving_task later */
5954 ret = mem_cgroup_precharge_mc(mm);
5956 mem_cgroup_clear_mc();
5963 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
5964 struct cgroup_taskset *tset)
5966 mem_cgroup_clear_mc();
5969 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5970 unsigned long addr, unsigned long end,
5971 struct mm_walk *walk)
5974 struct vm_area_struct *vma = walk->private;
5977 enum mc_target_type target_type;
5978 union mc_target target;
5980 struct page_cgroup *pc;
5983 * We don't take compound_lock() here but no race with splitting thp
5985 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
5986 * under splitting, which means there's no concurrent thp split,
5987 * - if another thread runs into split_huge_page() just after we
5988 * entered this if-block, the thread must wait for page table lock
5989 * to be unlocked in __split_huge_page_splitting(), where the main
5990 * part of thp split is not executed yet.
5992 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
5993 if (mc.precharge < HPAGE_PMD_NR) {
5997 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5998 if (target_type == MC_TARGET_PAGE) {
6000 if (!isolate_lru_page(page)) {
6001 pc = lookup_page_cgroup(page);
6002 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
6003 pc, mc.from, mc.to)) {
6004 mc.precharge -= HPAGE_PMD_NR;
6005 mc.moved_charge += HPAGE_PMD_NR;
6007 putback_lru_page(page);
6015 if (pmd_trans_unstable(pmd))
6018 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6019 for (; addr != end; addr += PAGE_SIZE) {
6020 pte_t ptent = *(pte++);
6026 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6027 case MC_TARGET_PAGE:
6029 if (isolate_lru_page(page))
6031 pc = lookup_page_cgroup(page);
6032 if (!mem_cgroup_move_account(page, 1, pc,
6035 /* we uncharge from mc.from later. */
6038 putback_lru_page(page);
6039 put: /* get_mctgt_type() gets the page */
6042 case MC_TARGET_SWAP:
6044 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6046 /* we fixup refcnts and charges later. */
6054 pte_unmap_unlock(pte - 1, ptl);
6059 * We have consumed all precharges we got in can_attach().
6060 * We try charge one by one, but don't do any additional
6061 * charges to mc.to if we have failed in charge once in attach()
6064 ret = mem_cgroup_do_precharge(1);
6072 static void mem_cgroup_move_charge(struct mm_struct *mm)
6074 struct vm_area_struct *vma;
6076 lru_add_drain_all();
6078 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
6080 * Someone who are holding the mmap_sem might be waiting in
6081 * waitq. So we cancel all extra charges, wake up all waiters,
6082 * and retry. Because we cancel precharges, we might not be able
6083 * to move enough charges, but moving charge is a best-effort
6084 * feature anyway, so it wouldn't be a big problem.
6086 __mem_cgroup_clear_mc();
6090 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6092 struct mm_walk mem_cgroup_move_charge_walk = {
6093 .pmd_entry = mem_cgroup_move_charge_pte_range,
6097 if (is_vm_hugetlb_page(vma))
6099 ret = walk_page_range(vma->vm_start, vma->vm_end,
6100 &mem_cgroup_move_charge_walk);
6103 * means we have consumed all precharges and failed in
6104 * doing additional charge. Just abandon here.
6108 up_read(&mm->mmap_sem);
6111 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
6112 struct cgroup_taskset *tset)
6114 struct task_struct *p = cgroup_taskset_first(tset);
6115 struct mm_struct *mm = get_task_mm(p);
6119 mem_cgroup_move_charge(mm);
6123 mem_cgroup_clear_mc();
6125 #else /* !CONFIG_MMU */
6126 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
6127 struct cgroup_taskset *tset)
6131 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
6132 struct cgroup_taskset *tset)
6135 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
6136 struct cgroup_taskset *tset)
6142 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6143 * to verify whether we're attached to the default hierarchy on each mount
6146 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
6149 * use_hierarchy is forced on the default hierarchy. cgroup core
6150 * guarantees that @root doesn't have any children, so turning it
6151 * on for the root memcg is enough.
6153 if (cgroup_on_dfl(root_css->cgroup))
6154 mem_cgroup_from_css(root_css)->use_hierarchy = true;
6157 struct cgroup_subsys memory_cgrp_subsys = {
6158 .css_alloc = mem_cgroup_css_alloc,
6159 .css_online = mem_cgroup_css_online,
6160 .css_offline = mem_cgroup_css_offline,
6161 .css_free = mem_cgroup_css_free,
6162 .css_reset = mem_cgroup_css_reset,
6163 .can_attach = mem_cgroup_can_attach,
6164 .cancel_attach = mem_cgroup_cancel_attach,
6165 .attach = mem_cgroup_move_task,
6166 .bind = mem_cgroup_bind,
6167 .legacy_cftypes = mem_cgroup_files,
6171 #ifdef CONFIG_MEMCG_SWAP
6172 static int __init enable_swap_account(char *s)
6174 if (!strcmp(s, "1"))
6175 really_do_swap_account = 1;
6176 else if (!strcmp(s, "0"))
6177 really_do_swap_account = 0;
6180 __setup("swapaccount=", enable_swap_account);
6182 static void __init memsw_file_init(void)
6184 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
6185 memsw_cgroup_files));
6188 static void __init enable_swap_cgroup(void)
6190 if (!mem_cgroup_disabled() && really_do_swap_account) {
6191 do_swap_account = 1;
6197 static void __init enable_swap_cgroup(void)
6202 #ifdef CONFIG_MEMCG_SWAP
6204 * mem_cgroup_swapout - transfer a memsw charge to swap
6205 * @page: page whose memsw charge to transfer
6206 * @entry: swap entry to move the charge to
6208 * Transfer the memsw charge of @page to @entry.
6210 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
6212 struct page_cgroup *pc;
6213 unsigned short oldid;
6215 VM_BUG_ON_PAGE(PageLRU(page), page);
6216 VM_BUG_ON_PAGE(page_count(page), page);
6218 if (!do_swap_account)
6221 pc = lookup_page_cgroup(page);
6223 /* Readahead page, never charged */
6224 if (!PageCgroupUsed(pc))
6227 VM_BUG_ON_PAGE(!(pc->flags & PCG_MEMSW), page);
6229 oldid = swap_cgroup_record(entry, mem_cgroup_id(pc->mem_cgroup));
6230 VM_BUG_ON_PAGE(oldid, page);
6232 pc->flags &= ~PCG_MEMSW;
6233 css_get(&pc->mem_cgroup->css);
6234 mem_cgroup_swap_statistics(pc->mem_cgroup, true);
6238 * mem_cgroup_uncharge_swap - uncharge a swap entry
6239 * @entry: swap entry to uncharge
6241 * Drop the memsw charge associated with @entry.
6243 void mem_cgroup_uncharge_swap(swp_entry_t entry)
6245 struct mem_cgroup *memcg;
6248 if (!do_swap_account)
6251 id = swap_cgroup_record(entry, 0);
6253 memcg = mem_cgroup_lookup(id);
6255 if (!mem_cgroup_is_root(memcg))
6256 page_counter_uncharge(&memcg->memsw, 1);
6257 mem_cgroup_swap_statistics(memcg, false);
6258 css_put(&memcg->css);
6265 * mem_cgroup_try_charge - try charging a page
6266 * @page: page to charge
6267 * @mm: mm context of the victim
6268 * @gfp_mask: reclaim mode
6269 * @memcgp: charged memcg return
6271 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6272 * pages according to @gfp_mask if necessary.
6274 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
6275 * Otherwise, an error code is returned.
6277 * After page->mapping has been set up, the caller must finalize the
6278 * charge with mem_cgroup_commit_charge(). Or abort the transaction
6279 * with mem_cgroup_cancel_charge() in case page instantiation fails.
6281 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
6282 gfp_t gfp_mask, struct mem_cgroup **memcgp)
6284 struct mem_cgroup *memcg = NULL;
6285 unsigned int nr_pages = 1;
6288 if (mem_cgroup_disabled())
6291 if (PageSwapCache(page)) {
6292 struct page_cgroup *pc = lookup_page_cgroup(page);
6294 * Every swap fault against a single page tries to charge the
6295 * page, bail as early as possible. shmem_unuse() encounters
6296 * already charged pages, too. The USED bit is protected by
6297 * the page lock, which serializes swap cache removal, which
6298 * in turn serializes uncharging.
6300 if (PageCgroupUsed(pc))
6304 if (PageTransHuge(page)) {
6305 nr_pages <<= compound_order(page);
6306 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
6309 if (do_swap_account && PageSwapCache(page))
6310 memcg = try_get_mem_cgroup_from_page(page);
6312 memcg = get_mem_cgroup_from_mm(mm);
6314 ret = try_charge(memcg, gfp_mask, nr_pages);
6316 css_put(&memcg->css);
6318 if (ret == -EINTR) {
6319 memcg = root_mem_cgroup;
6328 * mem_cgroup_commit_charge - commit a page charge
6329 * @page: page to charge
6330 * @memcg: memcg to charge the page to
6331 * @lrucare: page might be on LRU already
6333 * Finalize a charge transaction started by mem_cgroup_try_charge(),
6334 * after page->mapping has been set up. This must happen atomically
6335 * as part of the page instantiation, i.e. under the page table lock
6336 * for anonymous pages, under the page lock for page and swap cache.
6338 * In addition, the page must not be on the LRU during the commit, to
6339 * prevent racing with task migration. If it might be, use @lrucare.
6341 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
6343 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
6346 unsigned int nr_pages = 1;
6348 VM_BUG_ON_PAGE(!page->mapping, page);
6349 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
6351 if (mem_cgroup_disabled())
6354 * Swap faults will attempt to charge the same page multiple
6355 * times. But reuse_swap_page() might have removed the page
6356 * from swapcache already, so we can't check PageSwapCache().
6361 commit_charge(page, memcg, lrucare);
6363 if (PageTransHuge(page)) {
6364 nr_pages <<= compound_order(page);
6365 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
6368 local_irq_disable();
6369 mem_cgroup_charge_statistics(memcg, page, nr_pages);
6370 memcg_check_events(memcg, page);
6373 if (do_swap_account && PageSwapCache(page)) {
6374 swp_entry_t entry = { .val = page_private(page) };
6376 * The swap entry might not get freed for a long time,
6377 * let's not wait for it. The page already received a
6378 * memory+swap charge, drop the swap entry duplicate.
6380 mem_cgroup_uncharge_swap(entry);
6385 * mem_cgroup_cancel_charge - cancel a page charge
6386 * @page: page to charge
6387 * @memcg: memcg to charge the page to
6389 * Cancel a charge transaction started by mem_cgroup_try_charge().
6391 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg)
6393 unsigned int nr_pages = 1;
6395 if (mem_cgroup_disabled())
6398 * Swap faults will attempt to charge the same page multiple
6399 * times. But reuse_swap_page() might have removed the page
6400 * from swapcache already, so we can't check PageSwapCache().
6405 if (PageTransHuge(page)) {
6406 nr_pages <<= compound_order(page);
6407 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
6410 cancel_charge(memcg, nr_pages);
6413 static void uncharge_batch(struct mem_cgroup *memcg, unsigned long pgpgout,
6414 unsigned long nr_mem, unsigned long nr_memsw,
6415 unsigned long nr_anon, unsigned long nr_file,
6416 unsigned long nr_huge, struct page *dummy_page)
6418 unsigned long flags;
6420 if (!mem_cgroup_is_root(memcg)) {
6422 page_counter_uncharge(&memcg->memory, nr_mem);
6424 page_counter_uncharge(&memcg->memsw, nr_memsw);
6425 memcg_oom_recover(memcg);
6428 local_irq_save(flags);
6429 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS], nr_anon);
6430 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_CACHE], nr_file);
6431 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE], nr_huge);
6432 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT], pgpgout);
6433 __this_cpu_add(memcg->stat->nr_page_events, nr_anon + nr_file);
6434 memcg_check_events(memcg, dummy_page);
6435 local_irq_restore(flags);
6438 static void uncharge_list(struct list_head *page_list)
6440 struct mem_cgroup *memcg = NULL;
6441 unsigned long nr_memsw = 0;
6442 unsigned long nr_anon = 0;
6443 unsigned long nr_file = 0;
6444 unsigned long nr_huge = 0;
6445 unsigned long pgpgout = 0;
6446 unsigned long nr_mem = 0;
6447 struct list_head *next;
6450 next = page_list->next;
6452 unsigned int nr_pages = 1;
6453 struct page_cgroup *pc;
6455 page = list_entry(next, struct page, lru);
6456 next = page->lru.next;
6458 VM_BUG_ON_PAGE(PageLRU(page), page);
6459 VM_BUG_ON_PAGE(page_count(page), page);
6461 pc = lookup_page_cgroup(page);
6462 if (!PageCgroupUsed(pc))
6466 * Nobody should be changing or seriously looking at
6467 * pc->mem_cgroup and pc->flags at this point, we have
6468 * fully exclusive access to the page.
6471 if (memcg != pc->mem_cgroup) {
6473 uncharge_batch(memcg, pgpgout, nr_mem, nr_memsw,
6474 nr_anon, nr_file, nr_huge, page);
6475 pgpgout = nr_mem = nr_memsw = 0;
6476 nr_anon = nr_file = nr_huge = 0;
6478 memcg = pc->mem_cgroup;
6481 if (PageTransHuge(page)) {
6482 nr_pages <<= compound_order(page);
6483 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
6484 nr_huge += nr_pages;
6488 nr_anon += nr_pages;
6490 nr_file += nr_pages;
6492 if (pc->flags & PCG_MEM)
6494 if (pc->flags & PCG_MEMSW)
6495 nr_memsw += nr_pages;
6499 } while (next != page_list);
6502 uncharge_batch(memcg, pgpgout, nr_mem, nr_memsw,
6503 nr_anon, nr_file, nr_huge, page);
6507 * mem_cgroup_uncharge - uncharge a page
6508 * @page: page to uncharge
6510 * Uncharge a page previously charged with mem_cgroup_try_charge() and
6511 * mem_cgroup_commit_charge().
6513 void mem_cgroup_uncharge(struct page *page)
6515 struct page_cgroup *pc;
6517 if (mem_cgroup_disabled())
6520 /* Don't touch page->lru of any random page, pre-check: */
6521 pc = lookup_page_cgroup(page);
6522 if (!PageCgroupUsed(pc))
6525 INIT_LIST_HEAD(&page->lru);
6526 uncharge_list(&page->lru);
6530 * mem_cgroup_uncharge_list - uncharge a list of page
6531 * @page_list: list of pages to uncharge
6533 * Uncharge a list of pages previously charged with
6534 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
6536 void mem_cgroup_uncharge_list(struct list_head *page_list)
6538 if (mem_cgroup_disabled())
6541 if (!list_empty(page_list))
6542 uncharge_list(page_list);
6546 * mem_cgroup_migrate - migrate a charge to another page
6547 * @oldpage: currently charged page
6548 * @newpage: page to transfer the charge to
6549 * @lrucare: both pages might be on the LRU already
6551 * Migrate the charge from @oldpage to @newpage.
6553 * Both pages must be locked, @newpage->mapping must be set up.
6555 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage,
6558 struct page_cgroup *pc;
6561 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6562 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6563 VM_BUG_ON_PAGE(!lrucare && PageLRU(oldpage), oldpage);
6564 VM_BUG_ON_PAGE(!lrucare && PageLRU(newpage), newpage);
6565 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6566 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6569 if (mem_cgroup_disabled())
6572 /* Page cache replacement: new page already charged? */
6573 pc = lookup_page_cgroup(newpage);
6574 if (PageCgroupUsed(pc))
6577 /* Re-entrant migration: old page already uncharged? */
6578 pc = lookup_page_cgroup(oldpage);
6579 if (!PageCgroupUsed(pc))
6582 VM_BUG_ON_PAGE(!(pc->flags & PCG_MEM), oldpage);
6583 VM_BUG_ON_PAGE(do_swap_account && !(pc->flags & PCG_MEMSW), oldpage);
6586 lock_page_lru(oldpage, &isolated);
6591 unlock_page_lru(oldpage, isolated);
6593 commit_charge(newpage, pc->mem_cgroup, lrucare);
6597 * subsys_initcall() for memory controller.
6599 * Some parts like hotcpu_notifier() have to be initialized from this context
6600 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
6601 * everything that doesn't depend on a specific mem_cgroup structure should
6602 * be initialized from here.
6604 static int __init mem_cgroup_init(void)
6606 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
6607 enable_swap_cgroup();
6608 mem_cgroup_soft_limit_tree_init();
6612 subsys_initcall(mem_cgroup_init);