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/swap_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 /* Whether the swap controller is active */
76 #ifdef CONFIG_MEMCG_SWAP
77 int do_swap_account __read_mostly;
79 #define do_swap_account 0
82 static const char * const mem_cgroup_stat_names[] = {
91 static const char * const mem_cgroup_events_names[] = {
98 static const char * const mem_cgroup_lru_names[] = {
107 * Per memcg event counter is incremented at every pagein/pageout. With THP,
108 * it will be incremated by the number of pages. This counter is used for
109 * for trigger some periodic events. This is straightforward and better
110 * than using jiffies etc. to handle periodic memcg event.
112 enum mem_cgroup_events_target {
113 MEM_CGROUP_TARGET_THRESH,
114 MEM_CGROUP_TARGET_SOFTLIMIT,
115 MEM_CGROUP_TARGET_NUMAINFO,
118 #define THRESHOLDS_EVENTS_TARGET 128
119 #define SOFTLIMIT_EVENTS_TARGET 1024
120 #define NUMAINFO_EVENTS_TARGET 1024
122 struct mem_cgroup_stat_cpu {
123 long count[MEM_CGROUP_STAT_NSTATS];
124 unsigned long events[MEMCG_NR_EVENTS];
125 unsigned long nr_page_events;
126 unsigned long targets[MEM_CGROUP_NTARGETS];
129 struct reclaim_iter {
130 struct mem_cgroup *position;
131 /* scan generation, increased every round-trip */
132 unsigned int generation;
136 * per-zone information in memory controller.
138 struct mem_cgroup_per_zone {
139 struct lruvec lruvec;
140 unsigned long lru_size[NR_LRU_LISTS];
142 struct reclaim_iter iter[DEF_PRIORITY + 1];
144 struct rb_node tree_node; /* RB tree node */
145 unsigned long usage_in_excess;/* Set to the value by which */
146 /* the soft limit is exceeded*/
148 struct mem_cgroup *memcg; /* Back pointer, we cannot */
149 /* use container_of */
152 struct mem_cgroup_per_node {
153 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
157 * Cgroups above their limits are maintained in a RB-Tree, independent of
158 * their hierarchy representation
161 struct mem_cgroup_tree_per_zone {
162 struct rb_root rb_root;
166 struct mem_cgroup_tree_per_node {
167 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
170 struct mem_cgroup_tree {
171 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
174 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
176 struct mem_cgroup_threshold {
177 struct eventfd_ctx *eventfd;
178 unsigned long threshold;
182 struct mem_cgroup_threshold_ary {
183 /* An array index points to threshold just below or equal to usage. */
184 int current_threshold;
185 /* Size of entries[] */
187 /* Array of thresholds */
188 struct mem_cgroup_threshold entries[0];
191 struct mem_cgroup_thresholds {
192 /* Primary thresholds array */
193 struct mem_cgroup_threshold_ary *primary;
195 * Spare threshold array.
196 * This is needed to make mem_cgroup_unregister_event() "never fail".
197 * It must be able to store at least primary->size - 1 entries.
199 struct mem_cgroup_threshold_ary *spare;
203 struct mem_cgroup_eventfd_list {
204 struct list_head list;
205 struct eventfd_ctx *eventfd;
209 * cgroup_event represents events which userspace want to receive.
211 struct mem_cgroup_event {
213 * memcg which the event belongs to.
215 struct mem_cgroup *memcg;
217 * eventfd to signal userspace about the event.
219 struct eventfd_ctx *eventfd;
221 * Each of these stored in a list by the cgroup.
223 struct list_head list;
225 * register_event() callback will be used to add new userspace
226 * waiter for changes related to this event. Use eventfd_signal()
227 * on eventfd to send notification to userspace.
229 int (*register_event)(struct mem_cgroup *memcg,
230 struct eventfd_ctx *eventfd, const char *args);
232 * unregister_event() callback will be called when userspace closes
233 * the eventfd or on cgroup removing. This callback must be set,
234 * if you want provide notification functionality.
236 void (*unregister_event)(struct mem_cgroup *memcg,
237 struct eventfd_ctx *eventfd);
239 * All fields below needed to unregister event when
240 * userspace closes eventfd.
243 wait_queue_head_t *wqh;
245 struct work_struct remove;
248 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
249 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
252 * The memory controller data structure. The memory controller controls both
253 * page cache and RSS per cgroup. We would eventually like to provide
254 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
255 * to help the administrator determine what knobs to tune.
257 * TODO: Add a water mark for the memory controller. Reclaim will begin when
258 * we hit the water mark. May be even add a low water mark, such that
259 * no reclaim occurs from a cgroup at it's low water mark, this is
260 * a feature that will be implemented much later in the future.
263 struct cgroup_subsys_state css;
265 /* Accounted resources */
266 struct page_counter memory;
267 struct page_counter memsw;
268 struct page_counter kmem;
270 /* Normal memory consumption range */
274 unsigned long soft_limit;
276 /* vmpressure notifications */
277 struct vmpressure vmpressure;
279 /* css_online() has been completed */
283 * Should the accounting and control be hierarchical, per subtree?
289 atomic_t oom_wakeups;
292 /* OOM-Killer disable */
293 int oom_kill_disable;
295 /* protect arrays of thresholds */
296 struct mutex thresholds_lock;
298 /* thresholds for memory usage. RCU-protected */
299 struct mem_cgroup_thresholds thresholds;
301 /* thresholds for mem+swap usage. RCU-protected */
302 struct mem_cgroup_thresholds memsw_thresholds;
304 /* For oom notifier event fd */
305 struct list_head oom_notify;
308 * Should we move charges of a task when a task is moved into this
309 * mem_cgroup ? And what type of charges should we move ?
311 unsigned long move_charge_at_immigrate;
313 * set > 0 if pages under this cgroup are moving to other cgroup.
315 atomic_t moving_account;
316 /* taken only while moving_account > 0 */
317 spinlock_t move_lock;
318 struct task_struct *move_lock_task;
319 unsigned long move_lock_flags;
323 struct mem_cgroup_stat_cpu __percpu *stat;
325 * used when a cpu is offlined or other synchronizations
326 * See mem_cgroup_read_stat().
328 struct mem_cgroup_stat_cpu nocpu_base;
329 spinlock_t pcp_counter_lock;
331 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
332 struct cg_proto tcp_mem;
334 #if defined(CONFIG_MEMCG_KMEM)
335 /* Index in the kmem_cache->memcg_params->memcg_caches array */
339 int last_scanned_node;
341 nodemask_t scan_nodes;
342 atomic_t numainfo_events;
343 atomic_t numainfo_updating;
346 /* List of events which userspace want to receive */
347 struct list_head event_list;
348 spinlock_t event_list_lock;
350 struct mem_cgroup_per_node *nodeinfo[0];
351 /* WARNING: nodeinfo must be the last member here */
354 #ifdef CONFIG_MEMCG_KMEM
355 bool memcg_kmem_is_active(struct mem_cgroup *memcg)
357 return memcg->kmemcg_id >= 0;
361 /* Stuffs for move charges at task migration. */
363 * Types of charges to be moved.
365 #define MOVE_ANON 0x1U
366 #define MOVE_FILE 0x2U
367 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
369 /* "mc" and its members are protected by cgroup_mutex */
370 static struct move_charge_struct {
371 spinlock_t lock; /* for from, to */
372 struct mem_cgroup *from;
373 struct mem_cgroup *to;
375 unsigned long precharge;
376 unsigned long moved_charge;
377 unsigned long moved_swap;
378 struct task_struct *moving_task; /* a task moving charges */
379 wait_queue_head_t waitq; /* a waitq for other context */
381 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
382 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
386 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
387 * limit reclaim to prevent infinite loops, if they ever occur.
389 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
390 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
393 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
394 MEM_CGROUP_CHARGE_TYPE_ANON,
395 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
396 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
400 /* for encoding cft->private value on file */
408 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
409 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
410 #define MEMFILE_ATTR(val) ((val) & 0xffff)
411 /* Used for OOM nofiier */
412 #define OOM_CONTROL (0)
415 * The memcg_create_mutex will be held whenever a new cgroup is created.
416 * As a consequence, any change that needs to protect against new child cgroups
417 * appearing has to hold it as well.
419 static DEFINE_MUTEX(memcg_create_mutex);
421 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
423 return s ? container_of(s, struct mem_cgroup, css) : NULL;
426 /* Some nice accessors for the vmpressure. */
427 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
430 memcg = root_mem_cgroup;
431 return &memcg->vmpressure;
434 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
436 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
439 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
441 return (memcg == root_mem_cgroup);
445 * We restrict the id in the range of [1, 65535], so it can fit into
448 #define MEM_CGROUP_ID_MAX USHRT_MAX
450 static inline unsigned short mem_cgroup_id(struct mem_cgroup *memcg)
452 return memcg->css.id;
455 static inline struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
457 struct cgroup_subsys_state *css;
459 css = css_from_id(id, &memory_cgrp_subsys);
460 return mem_cgroup_from_css(css);
463 /* Writing them here to avoid exposing memcg's inner layout */
464 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
466 void sock_update_memcg(struct sock *sk)
468 if (mem_cgroup_sockets_enabled) {
469 struct mem_cgroup *memcg;
470 struct cg_proto *cg_proto;
472 BUG_ON(!sk->sk_prot->proto_cgroup);
474 /* Socket cloning can throw us here with sk_cgrp already
475 * filled. It won't however, necessarily happen from
476 * process context. So the test for root memcg given
477 * the current task's memcg won't help us in this case.
479 * Respecting the original socket's memcg is a better
480 * decision in this case.
483 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
484 css_get(&sk->sk_cgrp->memcg->css);
489 memcg = mem_cgroup_from_task(current);
490 cg_proto = sk->sk_prot->proto_cgroup(memcg);
491 if (!mem_cgroup_is_root(memcg) &&
492 memcg_proto_active(cg_proto) &&
493 css_tryget_online(&memcg->css)) {
494 sk->sk_cgrp = cg_proto;
499 EXPORT_SYMBOL(sock_update_memcg);
501 void sock_release_memcg(struct sock *sk)
503 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
504 struct mem_cgroup *memcg;
505 WARN_ON(!sk->sk_cgrp->memcg);
506 memcg = sk->sk_cgrp->memcg;
507 css_put(&sk->sk_cgrp->memcg->css);
511 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
513 if (!memcg || mem_cgroup_is_root(memcg))
516 return &memcg->tcp_mem;
518 EXPORT_SYMBOL(tcp_proto_cgroup);
520 static void disarm_sock_keys(struct mem_cgroup *memcg)
522 if (!memcg_proto_activated(&memcg->tcp_mem))
524 static_key_slow_dec(&memcg_socket_limit_enabled);
527 static void disarm_sock_keys(struct mem_cgroup *memcg)
532 #ifdef CONFIG_MEMCG_KMEM
534 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
535 * The main reason for not using cgroup id for this:
536 * this works better in sparse environments, where we have a lot of memcgs,
537 * but only a few kmem-limited. Or also, if we have, for instance, 200
538 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
539 * 200 entry array for that.
541 * The current size of the caches array is stored in memcg_nr_cache_ids. It
542 * will double each time we have to increase it.
544 static DEFINE_IDA(memcg_cache_ida);
545 int memcg_nr_cache_ids;
548 * MIN_SIZE is different than 1, because we would like to avoid going through
549 * the alloc/free process all the time. In a small machine, 4 kmem-limited
550 * cgroups is a reasonable guess. In the future, it could be a parameter or
551 * tunable, but that is strictly not necessary.
553 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
554 * this constant directly from cgroup, but it is understandable that this is
555 * better kept as an internal representation in cgroup.c. In any case, the
556 * cgrp_id space is not getting any smaller, and we don't have to necessarily
557 * increase ours as well if it increases.
559 #define MEMCG_CACHES_MIN_SIZE 4
560 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
563 * A lot of the calls to the cache allocation functions are expected to be
564 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
565 * conditional to this static branch, we'll have to allow modules that does
566 * kmem_cache_alloc and the such to see this symbol as well
568 struct static_key memcg_kmem_enabled_key;
569 EXPORT_SYMBOL(memcg_kmem_enabled_key);
571 static void memcg_free_cache_id(int id);
573 static void disarm_kmem_keys(struct mem_cgroup *memcg)
575 if (memcg_kmem_is_active(memcg)) {
576 static_key_slow_dec(&memcg_kmem_enabled_key);
577 memcg_free_cache_id(memcg->kmemcg_id);
580 * This check can't live in kmem destruction function,
581 * since the charges will outlive the cgroup
583 WARN_ON(page_counter_read(&memcg->kmem));
586 static void disarm_kmem_keys(struct mem_cgroup *memcg)
589 #endif /* CONFIG_MEMCG_KMEM */
591 static void disarm_static_keys(struct mem_cgroup *memcg)
593 disarm_sock_keys(memcg);
594 disarm_kmem_keys(memcg);
597 static struct mem_cgroup_per_zone *
598 mem_cgroup_zone_zoneinfo(struct mem_cgroup *memcg, struct zone *zone)
600 int nid = zone_to_nid(zone);
601 int zid = zone_idx(zone);
603 return &memcg->nodeinfo[nid]->zoneinfo[zid];
606 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
611 static struct mem_cgroup_per_zone *
612 mem_cgroup_page_zoneinfo(struct mem_cgroup *memcg, struct page *page)
614 int nid = page_to_nid(page);
615 int zid = page_zonenum(page);
617 return &memcg->nodeinfo[nid]->zoneinfo[zid];
620 static struct mem_cgroup_tree_per_zone *
621 soft_limit_tree_node_zone(int nid, int zid)
623 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
626 static struct mem_cgroup_tree_per_zone *
627 soft_limit_tree_from_page(struct page *page)
629 int nid = page_to_nid(page);
630 int zid = page_zonenum(page);
632 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
635 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone *mz,
636 struct mem_cgroup_tree_per_zone *mctz,
637 unsigned long new_usage_in_excess)
639 struct rb_node **p = &mctz->rb_root.rb_node;
640 struct rb_node *parent = NULL;
641 struct mem_cgroup_per_zone *mz_node;
646 mz->usage_in_excess = new_usage_in_excess;
647 if (!mz->usage_in_excess)
651 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
653 if (mz->usage_in_excess < mz_node->usage_in_excess)
656 * We can't avoid mem cgroups that are over their soft
657 * limit by the same amount
659 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
662 rb_link_node(&mz->tree_node, parent, p);
663 rb_insert_color(&mz->tree_node, &mctz->rb_root);
667 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
668 struct mem_cgroup_tree_per_zone *mctz)
672 rb_erase(&mz->tree_node, &mctz->rb_root);
676 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
677 struct mem_cgroup_tree_per_zone *mctz)
681 spin_lock_irqsave(&mctz->lock, flags);
682 __mem_cgroup_remove_exceeded(mz, mctz);
683 spin_unlock_irqrestore(&mctz->lock, flags);
686 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
688 unsigned long nr_pages = page_counter_read(&memcg->memory);
689 unsigned long soft_limit = ACCESS_ONCE(memcg->soft_limit);
690 unsigned long excess = 0;
692 if (nr_pages > soft_limit)
693 excess = nr_pages - soft_limit;
698 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
700 unsigned long excess;
701 struct mem_cgroup_per_zone *mz;
702 struct mem_cgroup_tree_per_zone *mctz;
704 mctz = soft_limit_tree_from_page(page);
706 * Necessary to update all ancestors when hierarchy is used.
707 * because their event counter is not touched.
709 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
710 mz = mem_cgroup_page_zoneinfo(memcg, page);
711 excess = soft_limit_excess(memcg);
713 * We have to update the tree if mz is on RB-tree or
714 * mem is over its softlimit.
716 if (excess || mz->on_tree) {
719 spin_lock_irqsave(&mctz->lock, flags);
720 /* if on-tree, remove it */
722 __mem_cgroup_remove_exceeded(mz, mctz);
724 * Insert again. mz->usage_in_excess will be updated.
725 * If excess is 0, no tree ops.
727 __mem_cgroup_insert_exceeded(mz, mctz, excess);
728 spin_unlock_irqrestore(&mctz->lock, flags);
733 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
735 struct mem_cgroup_tree_per_zone *mctz;
736 struct mem_cgroup_per_zone *mz;
740 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
741 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
742 mctz = soft_limit_tree_node_zone(nid, zid);
743 mem_cgroup_remove_exceeded(mz, mctz);
748 static struct mem_cgroup_per_zone *
749 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
751 struct rb_node *rightmost = NULL;
752 struct mem_cgroup_per_zone *mz;
756 rightmost = rb_last(&mctz->rb_root);
758 goto done; /* Nothing to reclaim from */
760 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
762 * Remove the node now but someone else can add it back,
763 * we will to add it back at the end of reclaim to its correct
764 * position in the tree.
766 __mem_cgroup_remove_exceeded(mz, mctz);
767 if (!soft_limit_excess(mz->memcg) ||
768 !css_tryget_online(&mz->memcg->css))
774 static struct mem_cgroup_per_zone *
775 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
777 struct mem_cgroup_per_zone *mz;
779 spin_lock_irq(&mctz->lock);
780 mz = __mem_cgroup_largest_soft_limit_node(mctz);
781 spin_unlock_irq(&mctz->lock);
786 * Implementation Note: reading percpu statistics for memcg.
788 * Both of vmstat[] and percpu_counter has threshold and do periodic
789 * synchronization to implement "quick" read. There are trade-off between
790 * reading cost and precision of value. Then, we may have a chance to implement
791 * a periodic synchronizion of counter in memcg's counter.
793 * But this _read() function is used for user interface now. The user accounts
794 * memory usage by memory cgroup and he _always_ requires exact value because
795 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
796 * have to visit all online cpus and make sum. So, for now, unnecessary
797 * synchronization is not implemented. (just implemented for cpu hotplug)
799 * If there are kernel internal actions which can make use of some not-exact
800 * value, and reading all cpu value can be performance bottleneck in some
801 * common workload, threashold and synchonization as vmstat[] should be
804 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
805 enum mem_cgroup_stat_index idx)
811 for_each_online_cpu(cpu)
812 val += per_cpu(memcg->stat->count[idx], cpu);
813 #ifdef CONFIG_HOTPLUG_CPU
814 spin_lock(&memcg->pcp_counter_lock);
815 val += memcg->nocpu_base.count[idx];
816 spin_unlock(&memcg->pcp_counter_lock);
822 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
823 enum mem_cgroup_events_index idx)
825 unsigned long val = 0;
829 for_each_online_cpu(cpu)
830 val += per_cpu(memcg->stat->events[idx], cpu);
831 #ifdef CONFIG_HOTPLUG_CPU
832 spin_lock(&memcg->pcp_counter_lock);
833 val += memcg->nocpu_base.events[idx];
834 spin_unlock(&memcg->pcp_counter_lock);
840 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
845 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
846 * counted as CACHE even if it's on ANON LRU.
849 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
852 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
855 if (PageTransHuge(page))
856 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
859 /* pagein of a big page is an event. So, ignore page size */
861 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
863 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
864 nr_pages = -nr_pages; /* for event */
867 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
870 unsigned long mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
872 struct mem_cgroup_per_zone *mz;
874 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
875 return mz->lru_size[lru];
878 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
880 unsigned int lru_mask)
882 unsigned long nr = 0;
885 VM_BUG_ON((unsigned)nid >= nr_node_ids);
887 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
888 struct mem_cgroup_per_zone *mz;
892 if (!(BIT(lru) & lru_mask))
894 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
895 nr += mz->lru_size[lru];
901 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
902 unsigned int lru_mask)
904 unsigned long nr = 0;
907 for_each_node_state(nid, N_MEMORY)
908 nr += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
912 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
913 enum mem_cgroup_events_target target)
915 unsigned long val, next;
917 val = __this_cpu_read(memcg->stat->nr_page_events);
918 next = __this_cpu_read(memcg->stat->targets[target]);
919 /* from time_after() in jiffies.h */
920 if ((long)next - (long)val < 0) {
922 case MEM_CGROUP_TARGET_THRESH:
923 next = val + THRESHOLDS_EVENTS_TARGET;
925 case MEM_CGROUP_TARGET_SOFTLIMIT:
926 next = val + SOFTLIMIT_EVENTS_TARGET;
928 case MEM_CGROUP_TARGET_NUMAINFO:
929 next = val + NUMAINFO_EVENTS_TARGET;
934 __this_cpu_write(memcg->stat->targets[target], next);
941 * Check events in order.
944 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
946 /* threshold event is triggered in finer grain than soft limit */
947 if (unlikely(mem_cgroup_event_ratelimit(memcg,
948 MEM_CGROUP_TARGET_THRESH))) {
950 bool do_numainfo __maybe_unused;
952 do_softlimit = mem_cgroup_event_ratelimit(memcg,
953 MEM_CGROUP_TARGET_SOFTLIMIT);
955 do_numainfo = mem_cgroup_event_ratelimit(memcg,
956 MEM_CGROUP_TARGET_NUMAINFO);
958 mem_cgroup_threshold(memcg);
959 if (unlikely(do_softlimit))
960 mem_cgroup_update_tree(memcg, page);
962 if (unlikely(do_numainfo))
963 atomic_inc(&memcg->numainfo_events);
968 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
971 * mm_update_next_owner() may clear mm->owner to NULL
972 * if it races with swapoff, page migration, etc.
973 * So this can be called with p == NULL.
978 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
981 static struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
983 struct mem_cgroup *memcg = NULL;
988 * Page cache insertions can happen withou an
989 * actual mm context, e.g. during disk probing
990 * on boot, loopback IO, acct() writes etc.
993 memcg = root_mem_cgroup;
995 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
996 if (unlikely(!memcg))
997 memcg = root_mem_cgroup;
999 } while (!css_tryget_online(&memcg->css));
1005 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1006 * @root: hierarchy root
1007 * @prev: previously returned memcg, NULL on first invocation
1008 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1010 * Returns references to children of the hierarchy below @root, or
1011 * @root itself, or %NULL after a full round-trip.
1013 * Caller must pass the return value in @prev on subsequent
1014 * invocations for reference counting, or use mem_cgroup_iter_break()
1015 * to cancel a hierarchy walk before the round-trip is complete.
1017 * Reclaimers can specify a zone and a priority level in @reclaim to
1018 * divide up the memcgs in the hierarchy among all concurrent
1019 * reclaimers operating on the same zone and priority.
1021 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1022 struct mem_cgroup *prev,
1023 struct mem_cgroup_reclaim_cookie *reclaim)
1025 struct reclaim_iter *uninitialized_var(iter);
1026 struct cgroup_subsys_state *css = NULL;
1027 struct mem_cgroup *memcg = NULL;
1028 struct mem_cgroup *pos = NULL;
1030 if (mem_cgroup_disabled())
1034 root = root_mem_cgroup;
1036 if (prev && !reclaim)
1039 if (!root->use_hierarchy && root != root_mem_cgroup) {
1048 struct mem_cgroup_per_zone *mz;
1050 mz = mem_cgroup_zone_zoneinfo(root, reclaim->zone);
1051 iter = &mz->iter[reclaim->priority];
1053 if (prev && reclaim->generation != iter->generation)
1057 pos = ACCESS_ONCE(iter->position);
1059 * A racing update may change the position and
1060 * put the last reference, hence css_tryget(),
1061 * or retry to see the updated position.
1063 } while (pos && !css_tryget(&pos->css));
1070 css = css_next_descendant_pre(css, &root->css);
1073 * Reclaimers share the hierarchy walk, and a
1074 * new one might jump in right at the end of
1075 * the hierarchy - make sure they see at least
1076 * one group and restart from the beginning.
1084 * Verify the css and acquire a reference. The root
1085 * is provided by the caller, so we know it's alive
1086 * and kicking, and don't take an extra reference.
1088 memcg = mem_cgroup_from_css(css);
1090 if (css == &root->css)
1093 if (css_tryget(css)) {
1095 * Make sure the memcg is initialized:
1096 * mem_cgroup_css_online() orders the the
1097 * initialization against setting the flag.
1099 if (smp_load_acquire(&memcg->initialized))
1109 if (cmpxchg(&iter->position, pos, memcg) == pos) {
1111 css_get(&memcg->css);
1117 * pairs with css_tryget when dereferencing iter->position
1126 reclaim->generation = iter->generation;
1132 if (prev && prev != root)
1133 css_put(&prev->css);
1139 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1140 * @root: hierarchy root
1141 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1143 void mem_cgroup_iter_break(struct mem_cgroup *root,
1144 struct mem_cgroup *prev)
1147 root = root_mem_cgroup;
1148 if (prev && prev != root)
1149 css_put(&prev->css);
1153 * Iteration constructs for visiting all cgroups (under a tree). If
1154 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1155 * be used for reference counting.
1157 #define for_each_mem_cgroup_tree(iter, root) \
1158 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1160 iter = mem_cgroup_iter(root, iter, NULL))
1162 #define for_each_mem_cgroup(iter) \
1163 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1165 iter = mem_cgroup_iter(NULL, iter, NULL))
1167 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1169 struct mem_cgroup *memcg;
1172 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1173 if (unlikely(!memcg))
1178 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1181 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1189 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1192 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1193 * @zone: zone of the wanted lruvec
1194 * @memcg: memcg of the wanted lruvec
1196 * Returns the lru list vector holding pages for the given @zone and
1197 * @mem. This can be the global zone lruvec, if the memory controller
1200 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1201 struct mem_cgroup *memcg)
1203 struct mem_cgroup_per_zone *mz;
1204 struct lruvec *lruvec;
1206 if (mem_cgroup_disabled()) {
1207 lruvec = &zone->lruvec;
1211 mz = mem_cgroup_zone_zoneinfo(memcg, zone);
1212 lruvec = &mz->lruvec;
1215 * Since a node can be onlined after the mem_cgroup was created,
1216 * we have to be prepared to initialize lruvec->zone here;
1217 * and if offlined then reonlined, we need to reinitialize it.
1219 if (unlikely(lruvec->zone != zone))
1220 lruvec->zone = zone;
1225 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1227 * @zone: zone of the page
1229 * This function is only safe when following the LRU page isolation
1230 * and putback protocol: the LRU lock must be held, and the page must
1231 * either be PageLRU() or the caller must have isolated/allocated it.
1233 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1235 struct mem_cgroup_per_zone *mz;
1236 struct mem_cgroup *memcg;
1237 struct lruvec *lruvec;
1239 if (mem_cgroup_disabled()) {
1240 lruvec = &zone->lruvec;
1244 memcg = page->mem_cgroup;
1246 * Swapcache readahead pages are added to the LRU - and
1247 * possibly migrated - before they are charged.
1250 memcg = root_mem_cgroup;
1252 mz = mem_cgroup_page_zoneinfo(memcg, page);
1253 lruvec = &mz->lruvec;
1256 * Since a node can be onlined after the mem_cgroup was created,
1257 * we have to be prepared to initialize lruvec->zone here;
1258 * and if offlined then reonlined, we need to reinitialize it.
1260 if (unlikely(lruvec->zone != zone))
1261 lruvec->zone = zone;
1266 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1267 * @lruvec: mem_cgroup per zone lru vector
1268 * @lru: index of lru list the page is sitting on
1269 * @nr_pages: positive when adding or negative when removing
1271 * This function must be called when a page is added to or removed from an
1274 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1277 struct mem_cgroup_per_zone *mz;
1278 unsigned long *lru_size;
1280 if (mem_cgroup_disabled())
1283 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1284 lru_size = mz->lru_size + lru;
1285 *lru_size += nr_pages;
1286 VM_BUG_ON((long)(*lru_size) < 0);
1289 bool mem_cgroup_is_descendant(struct mem_cgroup *memcg, struct mem_cgroup *root)
1293 if (!root->use_hierarchy)
1295 return cgroup_is_descendant(memcg->css.cgroup, root->css.cgroup);
1298 bool task_in_mem_cgroup(struct task_struct *task, struct mem_cgroup *memcg)
1300 struct mem_cgroup *task_memcg;
1301 struct task_struct *p;
1304 p = find_lock_task_mm(task);
1306 task_memcg = get_mem_cgroup_from_mm(p->mm);
1310 * All threads may have already detached their mm's, but the oom
1311 * killer still needs to detect if they have already been oom
1312 * killed to prevent needlessly killing additional tasks.
1315 task_memcg = mem_cgroup_from_task(task);
1316 css_get(&task_memcg->css);
1319 ret = mem_cgroup_is_descendant(task_memcg, memcg);
1320 css_put(&task_memcg->css);
1324 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1326 unsigned long inactive_ratio;
1327 unsigned long inactive;
1328 unsigned long active;
1331 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1332 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1334 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1336 inactive_ratio = int_sqrt(10 * gb);
1340 return inactive * inactive_ratio < active;
1343 bool mem_cgroup_lruvec_online(struct lruvec *lruvec)
1345 struct mem_cgroup_per_zone *mz;
1346 struct mem_cgroup *memcg;
1348 if (mem_cgroup_disabled())
1351 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1354 return !!(memcg->css.flags & CSS_ONLINE);
1357 #define mem_cgroup_from_counter(counter, member) \
1358 container_of(counter, struct mem_cgroup, member)
1361 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1362 * @memcg: the memory cgroup
1364 * Returns the maximum amount of memory @mem can be charged with, in
1367 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1369 unsigned long margin = 0;
1370 unsigned long count;
1371 unsigned long limit;
1373 count = page_counter_read(&memcg->memory);
1374 limit = ACCESS_ONCE(memcg->memory.limit);
1376 margin = limit - count;
1378 if (do_swap_account) {
1379 count = page_counter_read(&memcg->memsw);
1380 limit = ACCESS_ONCE(memcg->memsw.limit);
1382 margin = min(margin, limit - count);
1388 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1391 if (mem_cgroup_disabled() || !memcg->css.parent)
1392 return vm_swappiness;
1394 return memcg->swappiness;
1398 * A routine for checking "mem" is under move_account() or not.
1400 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1401 * moving cgroups. This is for waiting at high-memory pressure
1404 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1406 struct mem_cgroup *from;
1407 struct mem_cgroup *to;
1410 * Unlike task_move routines, we access mc.to, mc.from not under
1411 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1413 spin_lock(&mc.lock);
1419 ret = mem_cgroup_is_descendant(from, memcg) ||
1420 mem_cgroup_is_descendant(to, memcg);
1422 spin_unlock(&mc.lock);
1426 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1428 if (mc.moving_task && current != mc.moving_task) {
1429 if (mem_cgroup_under_move(memcg)) {
1431 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1432 /* moving charge context might have finished. */
1435 finish_wait(&mc.waitq, &wait);
1442 #define K(x) ((x) << (PAGE_SHIFT-10))
1444 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1445 * @memcg: The memory cgroup that went over limit
1446 * @p: Task that is going to be killed
1448 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1451 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1453 /* oom_info_lock ensures that parallel ooms do not interleave */
1454 static DEFINE_MUTEX(oom_info_lock);
1455 struct mem_cgroup *iter;
1461 mutex_lock(&oom_info_lock);
1464 pr_info("Task in ");
1465 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1466 pr_cont(" killed as a result of limit of ");
1467 pr_cont_cgroup_path(memcg->css.cgroup);
1472 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1473 K((u64)page_counter_read(&memcg->memory)),
1474 K((u64)memcg->memory.limit), memcg->memory.failcnt);
1475 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1476 K((u64)page_counter_read(&memcg->memsw)),
1477 K((u64)memcg->memsw.limit), memcg->memsw.failcnt);
1478 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1479 K((u64)page_counter_read(&memcg->kmem)),
1480 K((u64)memcg->kmem.limit), memcg->kmem.failcnt);
1482 for_each_mem_cgroup_tree(iter, memcg) {
1483 pr_info("Memory cgroup stats for ");
1484 pr_cont_cgroup_path(iter->css.cgroup);
1487 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1488 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1490 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1491 K(mem_cgroup_read_stat(iter, i)));
1494 for (i = 0; i < NR_LRU_LISTS; i++)
1495 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1496 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1500 mutex_unlock(&oom_info_lock);
1504 * This function returns the number of memcg under hierarchy tree. Returns
1505 * 1(self count) if no children.
1507 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1510 struct mem_cgroup *iter;
1512 for_each_mem_cgroup_tree(iter, memcg)
1518 * Return the memory (and swap, if configured) limit for a memcg.
1520 static unsigned long mem_cgroup_get_limit(struct mem_cgroup *memcg)
1522 unsigned long limit;
1524 limit = memcg->memory.limit;
1525 if (mem_cgroup_swappiness(memcg)) {
1526 unsigned long memsw_limit;
1528 memsw_limit = memcg->memsw.limit;
1529 limit = min(limit + total_swap_pages, memsw_limit);
1534 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1537 struct mem_cgroup *iter;
1538 unsigned long chosen_points = 0;
1539 unsigned long totalpages;
1540 unsigned int points = 0;
1541 struct task_struct *chosen = NULL;
1544 * If current has a pending SIGKILL or is exiting, then automatically
1545 * select it. The goal is to allow it to allocate so that it may
1546 * quickly exit and free its memory.
1548 if (fatal_signal_pending(current) || task_will_free_mem(current)) {
1549 mark_tsk_oom_victim(current);
1553 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1554 totalpages = mem_cgroup_get_limit(memcg) ? : 1;
1555 for_each_mem_cgroup_tree(iter, memcg) {
1556 struct css_task_iter it;
1557 struct task_struct *task;
1559 css_task_iter_start(&iter->css, &it);
1560 while ((task = css_task_iter_next(&it))) {
1561 switch (oom_scan_process_thread(task, totalpages, NULL,
1563 case OOM_SCAN_SELECT:
1565 put_task_struct(chosen);
1567 chosen_points = ULONG_MAX;
1568 get_task_struct(chosen);
1570 case OOM_SCAN_CONTINUE:
1572 case OOM_SCAN_ABORT:
1573 css_task_iter_end(&it);
1574 mem_cgroup_iter_break(memcg, iter);
1576 put_task_struct(chosen);
1581 points = oom_badness(task, memcg, NULL, totalpages);
1582 if (!points || points < chosen_points)
1584 /* Prefer thread group leaders for display purposes */
1585 if (points == chosen_points &&
1586 thread_group_leader(chosen))
1590 put_task_struct(chosen);
1592 chosen_points = points;
1593 get_task_struct(chosen);
1595 css_task_iter_end(&it);
1600 points = chosen_points * 1000 / totalpages;
1601 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1602 NULL, "Memory cgroup out of memory");
1605 #if MAX_NUMNODES > 1
1608 * test_mem_cgroup_node_reclaimable
1609 * @memcg: the target memcg
1610 * @nid: the node ID to be checked.
1611 * @noswap : specify true here if the user wants flle only information.
1613 * This function returns whether the specified memcg contains any
1614 * reclaimable pages on a node. Returns true if there are any reclaimable
1615 * pages in the node.
1617 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1618 int nid, bool noswap)
1620 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1622 if (noswap || !total_swap_pages)
1624 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1631 * Always updating the nodemask is not very good - even if we have an empty
1632 * list or the wrong list here, we can start from some node and traverse all
1633 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1636 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1640 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1641 * pagein/pageout changes since the last update.
1643 if (!atomic_read(&memcg->numainfo_events))
1645 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1648 /* make a nodemask where this memcg uses memory from */
1649 memcg->scan_nodes = node_states[N_MEMORY];
1651 for_each_node_mask(nid, node_states[N_MEMORY]) {
1653 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1654 node_clear(nid, memcg->scan_nodes);
1657 atomic_set(&memcg->numainfo_events, 0);
1658 atomic_set(&memcg->numainfo_updating, 0);
1662 * Selecting a node where we start reclaim from. Because what we need is just
1663 * reducing usage counter, start from anywhere is O,K. Considering
1664 * memory reclaim from current node, there are pros. and cons.
1666 * Freeing memory from current node means freeing memory from a node which
1667 * we'll use or we've used. So, it may make LRU bad. And if several threads
1668 * hit limits, it will see a contention on a node. But freeing from remote
1669 * node means more costs for memory reclaim because of memory latency.
1671 * Now, we use round-robin. Better algorithm is welcomed.
1673 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1677 mem_cgroup_may_update_nodemask(memcg);
1678 node = memcg->last_scanned_node;
1680 node = next_node(node, memcg->scan_nodes);
1681 if (node == MAX_NUMNODES)
1682 node = first_node(memcg->scan_nodes);
1684 * We call this when we hit limit, not when pages are added to LRU.
1685 * No LRU may hold pages because all pages are UNEVICTABLE or
1686 * memcg is too small and all pages are not on LRU. In that case,
1687 * we use curret node.
1689 if (unlikely(node == MAX_NUMNODES))
1690 node = numa_node_id();
1692 memcg->last_scanned_node = node;
1696 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1702 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1705 unsigned long *total_scanned)
1707 struct mem_cgroup *victim = NULL;
1710 unsigned long excess;
1711 unsigned long nr_scanned;
1712 struct mem_cgroup_reclaim_cookie reclaim = {
1717 excess = soft_limit_excess(root_memcg);
1720 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1725 * If we have not been able to reclaim
1726 * anything, it might because there are
1727 * no reclaimable pages under this hierarchy
1732 * We want to do more targeted reclaim.
1733 * excess >> 2 is not to excessive so as to
1734 * reclaim too much, nor too less that we keep
1735 * coming back to reclaim from this cgroup
1737 if (total >= (excess >> 2) ||
1738 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1743 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1745 *total_scanned += nr_scanned;
1746 if (!soft_limit_excess(root_memcg))
1749 mem_cgroup_iter_break(root_memcg, victim);
1753 #ifdef CONFIG_LOCKDEP
1754 static struct lockdep_map memcg_oom_lock_dep_map = {
1755 .name = "memcg_oom_lock",
1759 static DEFINE_SPINLOCK(memcg_oom_lock);
1762 * Check OOM-Killer is already running under our hierarchy.
1763 * If someone is running, return false.
1765 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1767 struct mem_cgroup *iter, *failed = NULL;
1769 spin_lock(&memcg_oom_lock);
1771 for_each_mem_cgroup_tree(iter, memcg) {
1772 if (iter->oom_lock) {
1774 * this subtree of our hierarchy is already locked
1775 * so we cannot give a lock.
1778 mem_cgroup_iter_break(memcg, iter);
1781 iter->oom_lock = true;
1786 * OK, we failed to lock the whole subtree so we have
1787 * to clean up what we set up to the failing subtree
1789 for_each_mem_cgroup_tree(iter, memcg) {
1790 if (iter == failed) {
1791 mem_cgroup_iter_break(memcg, iter);
1794 iter->oom_lock = false;
1797 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1799 spin_unlock(&memcg_oom_lock);
1804 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1806 struct mem_cgroup *iter;
1808 spin_lock(&memcg_oom_lock);
1809 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1810 for_each_mem_cgroup_tree(iter, memcg)
1811 iter->oom_lock = false;
1812 spin_unlock(&memcg_oom_lock);
1815 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1817 struct mem_cgroup *iter;
1819 for_each_mem_cgroup_tree(iter, memcg)
1820 atomic_inc(&iter->under_oom);
1823 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1825 struct mem_cgroup *iter;
1828 * When a new child is created while the hierarchy is under oom,
1829 * mem_cgroup_oom_lock() may not be called. We have to use
1830 * atomic_add_unless() here.
1832 for_each_mem_cgroup_tree(iter, memcg)
1833 atomic_add_unless(&iter->under_oom, -1, 0);
1836 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1838 struct oom_wait_info {
1839 struct mem_cgroup *memcg;
1843 static int memcg_oom_wake_function(wait_queue_t *wait,
1844 unsigned mode, int sync, void *arg)
1846 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1847 struct mem_cgroup *oom_wait_memcg;
1848 struct oom_wait_info *oom_wait_info;
1850 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1851 oom_wait_memcg = oom_wait_info->memcg;
1853 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1854 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1856 return autoremove_wake_function(wait, mode, sync, arg);
1859 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
1861 atomic_inc(&memcg->oom_wakeups);
1862 /* for filtering, pass "memcg" as argument. */
1863 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1866 static void memcg_oom_recover(struct mem_cgroup *memcg)
1868 if (memcg && atomic_read(&memcg->under_oom))
1869 memcg_wakeup_oom(memcg);
1872 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1874 if (!current->memcg_oom.may_oom)
1877 * We are in the middle of the charge context here, so we
1878 * don't want to block when potentially sitting on a callstack
1879 * that holds all kinds of filesystem and mm locks.
1881 * Also, the caller may handle a failed allocation gracefully
1882 * (like optional page cache readahead) and so an OOM killer
1883 * invocation might not even be necessary.
1885 * That's why we don't do anything here except remember the
1886 * OOM context and then deal with it at the end of the page
1887 * fault when the stack is unwound, the locks are released,
1888 * and when we know whether the fault was overall successful.
1890 css_get(&memcg->css);
1891 current->memcg_oom.memcg = memcg;
1892 current->memcg_oom.gfp_mask = mask;
1893 current->memcg_oom.order = order;
1897 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1898 * @handle: actually kill/wait or just clean up the OOM state
1900 * This has to be called at the end of a page fault if the memcg OOM
1901 * handler was enabled.
1903 * Memcg supports userspace OOM handling where failed allocations must
1904 * sleep on a waitqueue until the userspace task resolves the
1905 * situation. Sleeping directly in the charge context with all kinds
1906 * of locks held is not a good idea, instead we remember an OOM state
1907 * in the task and mem_cgroup_oom_synchronize() has to be called at
1908 * the end of the page fault to complete the OOM handling.
1910 * Returns %true if an ongoing memcg OOM situation was detected and
1911 * completed, %false otherwise.
1913 bool mem_cgroup_oom_synchronize(bool handle)
1915 struct mem_cgroup *memcg = current->memcg_oom.memcg;
1916 struct oom_wait_info owait;
1919 /* OOM is global, do not handle */
1923 if (!handle || oom_killer_disabled)
1926 owait.memcg = memcg;
1927 owait.wait.flags = 0;
1928 owait.wait.func = memcg_oom_wake_function;
1929 owait.wait.private = current;
1930 INIT_LIST_HEAD(&owait.wait.task_list);
1932 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1933 mem_cgroup_mark_under_oom(memcg);
1935 locked = mem_cgroup_oom_trylock(memcg);
1938 mem_cgroup_oom_notify(memcg);
1940 if (locked && !memcg->oom_kill_disable) {
1941 mem_cgroup_unmark_under_oom(memcg);
1942 finish_wait(&memcg_oom_waitq, &owait.wait);
1943 mem_cgroup_out_of_memory(memcg, current->memcg_oom.gfp_mask,
1944 current->memcg_oom.order);
1947 mem_cgroup_unmark_under_oom(memcg);
1948 finish_wait(&memcg_oom_waitq, &owait.wait);
1952 mem_cgroup_oom_unlock(memcg);
1954 * There is no guarantee that an OOM-lock contender
1955 * sees the wakeups triggered by the OOM kill
1956 * uncharges. Wake any sleepers explicitely.
1958 memcg_oom_recover(memcg);
1961 current->memcg_oom.memcg = NULL;
1962 css_put(&memcg->css);
1967 * mem_cgroup_begin_page_stat - begin a page state statistics transaction
1968 * @page: page that is going to change accounted state
1970 * This function must mark the beginning of an accounted page state
1971 * change to prevent double accounting when the page is concurrently
1972 * being moved to another memcg:
1974 * memcg = mem_cgroup_begin_page_stat(page);
1975 * if (TestClearPageState(page))
1976 * mem_cgroup_update_page_stat(memcg, state, -1);
1977 * mem_cgroup_end_page_stat(memcg);
1979 struct mem_cgroup *mem_cgroup_begin_page_stat(struct page *page)
1981 struct mem_cgroup *memcg;
1982 unsigned long flags;
1985 * The RCU lock is held throughout the transaction. The fast
1986 * path can get away without acquiring the memcg->move_lock
1987 * because page moving starts with an RCU grace period.
1989 * The RCU lock also protects the memcg from being freed when
1990 * the page state that is going to change is the only thing
1991 * preventing the page from being uncharged.
1992 * E.g. end-writeback clearing PageWriteback(), which allows
1993 * migration to go ahead and uncharge the page before the
1994 * account transaction might be complete.
1998 if (mem_cgroup_disabled())
2001 memcg = page->mem_cgroup;
2002 if (unlikely(!memcg))
2005 if (atomic_read(&memcg->moving_account) <= 0)
2008 spin_lock_irqsave(&memcg->move_lock, flags);
2009 if (memcg != page->mem_cgroup) {
2010 spin_unlock_irqrestore(&memcg->move_lock, flags);
2015 * When charge migration first begins, we can have locked and
2016 * unlocked page stat updates happening concurrently. Track
2017 * the task who has the lock for mem_cgroup_end_page_stat().
2019 memcg->move_lock_task = current;
2020 memcg->move_lock_flags = flags;
2026 * mem_cgroup_end_page_stat - finish a page state statistics transaction
2027 * @memcg: the memcg that was accounted against
2029 void mem_cgroup_end_page_stat(struct mem_cgroup *memcg)
2031 if (memcg && memcg->move_lock_task == current) {
2032 unsigned long flags = memcg->move_lock_flags;
2034 memcg->move_lock_task = NULL;
2035 memcg->move_lock_flags = 0;
2037 spin_unlock_irqrestore(&memcg->move_lock, flags);
2044 * mem_cgroup_update_page_stat - update page state statistics
2045 * @memcg: memcg to account against
2046 * @idx: page state item to account
2047 * @val: number of pages (positive or negative)
2049 * See mem_cgroup_begin_page_stat() for locking requirements.
2051 void mem_cgroup_update_page_stat(struct mem_cgroup *memcg,
2052 enum mem_cgroup_stat_index idx, int val)
2054 VM_BUG_ON(!rcu_read_lock_held());
2057 this_cpu_add(memcg->stat->count[idx], val);
2061 * size of first charge trial. "32" comes from vmscan.c's magic value.
2062 * TODO: maybe necessary to use big numbers in big irons.
2064 #define CHARGE_BATCH 32U
2065 struct memcg_stock_pcp {
2066 struct mem_cgroup *cached; /* this never be root cgroup */
2067 unsigned int nr_pages;
2068 struct work_struct work;
2069 unsigned long flags;
2070 #define FLUSHING_CACHED_CHARGE 0
2072 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2073 static DEFINE_MUTEX(percpu_charge_mutex);
2076 * consume_stock: Try to consume stocked charge on this cpu.
2077 * @memcg: memcg to consume from.
2078 * @nr_pages: how many pages to charge.
2080 * The charges will only happen if @memcg matches the current cpu's memcg
2081 * stock, and at least @nr_pages are available in that stock. Failure to
2082 * service an allocation will refill the stock.
2084 * returns true if successful, false otherwise.
2086 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2088 struct memcg_stock_pcp *stock;
2091 if (nr_pages > CHARGE_BATCH)
2094 stock = &get_cpu_var(memcg_stock);
2095 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2096 stock->nr_pages -= nr_pages;
2099 put_cpu_var(memcg_stock);
2104 * Returns stocks cached in percpu and reset cached information.
2106 static void drain_stock(struct memcg_stock_pcp *stock)
2108 struct mem_cgroup *old = stock->cached;
2110 if (stock->nr_pages) {
2111 page_counter_uncharge(&old->memory, stock->nr_pages);
2112 if (do_swap_account)
2113 page_counter_uncharge(&old->memsw, stock->nr_pages);
2114 css_put_many(&old->css, stock->nr_pages);
2115 stock->nr_pages = 0;
2117 stock->cached = NULL;
2121 * This must be called under preempt disabled or must be called by
2122 * a thread which is pinned to local cpu.
2124 static void drain_local_stock(struct work_struct *dummy)
2126 struct memcg_stock_pcp *stock = this_cpu_ptr(&memcg_stock);
2128 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2132 * Cache charges(val) to local per_cpu area.
2133 * This will be consumed by consume_stock() function, later.
2135 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2137 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2139 if (stock->cached != memcg) { /* reset if necessary */
2141 stock->cached = memcg;
2143 stock->nr_pages += nr_pages;
2144 put_cpu_var(memcg_stock);
2148 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2149 * of the hierarchy under it.
2151 static void drain_all_stock(struct mem_cgroup *root_memcg)
2155 /* If someone's already draining, avoid adding running more workers. */
2156 if (!mutex_trylock(&percpu_charge_mutex))
2158 /* Notify other cpus that system-wide "drain" is running */
2161 for_each_online_cpu(cpu) {
2162 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2163 struct mem_cgroup *memcg;
2165 memcg = stock->cached;
2166 if (!memcg || !stock->nr_pages)
2168 if (!mem_cgroup_is_descendant(memcg, root_memcg))
2170 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2172 drain_local_stock(&stock->work);
2174 schedule_work_on(cpu, &stock->work);
2179 mutex_unlock(&percpu_charge_mutex);
2183 * This function drains percpu counter value from DEAD cpu and
2184 * move it to local cpu. Note that this function can be preempted.
2186 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2190 spin_lock(&memcg->pcp_counter_lock);
2191 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2192 long x = per_cpu(memcg->stat->count[i], cpu);
2194 per_cpu(memcg->stat->count[i], cpu) = 0;
2195 memcg->nocpu_base.count[i] += x;
2197 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2198 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2200 per_cpu(memcg->stat->events[i], cpu) = 0;
2201 memcg->nocpu_base.events[i] += x;
2203 spin_unlock(&memcg->pcp_counter_lock);
2206 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
2207 unsigned long action,
2210 int cpu = (unsigned long)hcpu;
2211 struct memcg_stock_pcp *stock;
2212 struct mem_cgroup *iter;
2214 if (action == CPU_ONLINE)
2217 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2220 for_each_mem_cgroup(iter)
2221 mem_cgroup_drain_pcp_counter(iter, cpu);
2223 stock = &per_cpu(memcg_stock, cpu);
2228 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2229 unsigned int nr_pages)
2231 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2232 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2233 struct mem_cgroup *mem_over_limit;
2234 struct page_counter *counter;
2235 unsigned long nr_reclaimed;
2236 bool may_swap = true;
2237 bool drained = false;
2240 if (mem_cgroup_is_root(memcg))
2243 if (consume_stock(memcg, nr_pages))
2246 if (!do_swap_account ||
2247 !page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2248 if (!page_counter_try_charge(&memcg->memory, batch, &counter))
2250 if (do_swap_account)
2251 page_counter_uncharge(&memcg->memsw, batch);
2252 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2254 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2258 if (batch > nr_pages) {
2264 * Unlike in global OOM situations, memcg is not in a physical
2265 * memory shortage. Allow dying and OOM-killed tasks to
2266 * bypass the last charges so that they can exit quickly and
2267 * free their memory.
2269 if (unlikely(test_thread_flag(TIF_MEMDIE) ||
2270 fatal_signal_pending(current) ||
2271 current->flags & PF_EXITING))
2274 if (unlikely(task_in_memcg_oom(current)))
2277 if (!(gfp_mask & __GFP_WAIT))
2280 mem_cgroup_events(mem_over_limit, MEMCG_MAX, 1);
2282 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2283 gfp_mask, may_swap);
2285 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2289 drain_all_stock(mem_over_limit);
2294 if (gfp_mask & __GFP_NORETRY)
2297 * Even though the limit is exceeded at this point, reclaim
2298 * may have been able to free some pages. Retry the charge
2299 * before killing the task.
2301 * Only for regular pages, though: huge pages are rather
2302 * unlikely to succeed so close to the limit, and we fall back
2303 * to regular pages anyway in case of failure.
2305 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2308 * At task move, charge accounts can be doubly counted. So, it's
2309 * better to wait until the end of task_move if something is going on.
2311 if (mem_cgroup_wait_acct_move(mem_over_limit))
2317 if (gfp_mask & __GFP_NOFAIL)
2320 if (fatal_signal_pending(current))
2323 mem_cgroup_events(mem_over_limit, MEMCG_OOM, 1);
2325 mem_cgroup_oom(mem_over_limit, gfp_mask, get_order(nr_pages));
2327 if (!(gfp_mask & __GFP_NOFAIL))
2333 css_get_many(&memcg->css, batch);
2334 if (batch > nr_pages)
2335 refill_stock(memcg, batch - nr_pages);
2337 * If the hierarchy is above the normal consumption range,
2338 * make the charging task trim their excess contribution.
2341 if (page_counter_read(&memcg->memory) <= memcg->high)
2343 mem_cgroup_events(memcg, MEMCG_HIGH, 1);
2344 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
2345 } while ((memcg = parent_mem_cgroup(memcg)));
2350 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2352 if (mem_cgroup_is_root(memcg))
2355 page_counter_uncharge(&memcg->memory, nr_pages);
2356 if (do_swap_account)
2357 page_counter_uncharge(&memcg->memsw, nr_pages);
2359 css_put_many(&memcg->css, nr_pages);
2363 * A helper function to get mem_cgroup from ID. must be called under
2364 * rcu_read_lock(). The caller is responsible for calling
2365 * css_tryget_online() if the mem_cgroup is used for charging. (dropping
2366 * refcnt from swap can be called against removed memcg.)
2368 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2370 /* ID 0 is unused ID */
2373 return mem_cgroup_from_id(id);
2377 * try_get_mem_cgroup_from_page - look up page's memcg association
2380 * Look up, get a css reference, and return the memcg that owns @page.
2382 * The page must be locked to prevent racing with swap-in and page
2383 * cache charges. If coming from an unlocked page table, the caller
2384 * must ensure the page is on the LRU or this can race with charging.
2386 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2388 struct mem_cgroup *memcg;
2392 VM_BUG_ON_PAGE(!PageLocked(page), page);
2394 memcg = page->mem_cgroup;
2396 if (!css_tryget_online(&memcg->css))
2398 } else if (PageSwapCache(page)) {
2399 ent.val = page_private(page);
2400 id = lookup_swap_cgroup_id(ent);
2402 memcg = mem_cgroup_lookup(id);
2403 if (memcg && !css_tryget_online(&memcg->css))
2410 static void lock_page_lru(struct page *page, int *isolated)
2412 struct zone *zone = page_zone(page);
2414 spin_lock_irq(&zone->lru_lock);
2415 if (PageLRU(page)) {
2416 struct lruvec *lruvec;
2418 lruvec = mem_cgroup_page_lruvec(page, zone);
2420 del_page_from_lru_list(page, lruvec, page_lru(page));
2426 static void unlock_page_lru(struct page *page, int isolated)
2428 struct zone *zone = page_zone(page);
2431 struct lruvec *lruvec;
2433 lruvec = mem_cgroup_page_lruvec(page, zone);
2434 VM_BUG_ON_PAGE(PageLRU(page), page);
2436 add_page_to_lru_list(page, lruvec, page_lru(page));
2438 spin_unlock_irq(&zone->lru_lock);
2441 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2446 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2449 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2450 * may already be on some other mem_cgroup's LRU. Take care of it.
2453 lock_page_lru(page, &isolated);
2456 * Nobody should be changing or seriously looking at
2457 * page->mem_cgroup at this point:
2459 * - the page is uncharged
2461 * - the page is off-LRU
2463 * - an anonymous fault has exclusive page access, except for
2464 * a locked page table
2466 * - a page cache insertion, a swapin fault, or a migration
2467 * have the page locked
2469 page->mem_cgroup = memcg;
2472 unlock_page_lru(page, isolated);
2475 #ifdef CONFIG_MEMCG_KMEM
2476 int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp,
2477 unsigned long nr_pages)
2479 struct page_counter *counter;
2482 ret = page_counter_try_charge(&memcg->kmem, nr_pages, &counter);
2486 ret = try_charge(memcg, gfp, nr_pages);
2487 if (ret == -EINTR) {
2489 * try_charge() chose to bypass to root due to OOM kill or
2490 * fatal signal. Since our only options are to either fail
2491 * the allocation or charge it to this cgroup, do it as a
2492 * temporary condition. But we can't fail. From a kmem/slab
2493 * perspective, the cache has already been selected, by
2494 * mem_cgroup_kmem_get_cache(), so it is too late to change
2497 * This condition will only trigger if the task entered
2498 * memcg_charge_kmem in a sane state, but was OOM-killed
2499 * during try_charge() above. Tasks that were already dying
2500 * when the allocation triggers should have been already
2501 * directed to the root cgroup in memcontrol.h
2503 page_counter_charge(&memcg->memory, nr_pages);
2504 if (do_swap_account)
2505 page_counter_charge(&memcg->memsw, nr_pages);
2506 css_get_many(&memcg->css, nr_pages);
2509 page_counter_uncharge(&memcg->kmem, nr_pages);
2514 void memcg_uncharge_kmem(struct mem_cgroup *memcg, unsigned long nr_pages)
2516 page_counter_uncharge(&memcg->memory, nr_pages);
2517 if (do_swap_account)
2518 page_counter_uncharge(&memcg->memsw, nr_pages);
2520 page_counter_uncharge(&memcg->kmem, nr_pages);
2522 css_put_many(&memcg->css, nr_pages);
2526 * helper for acessing a memcg's index. It will be used as an index in the
2527 * child cache array in kmem_cache, and also to derive its name. This function
2528 * will return -1 when this is not a kmem-limited memcg.
2530 int memcg_cache_id(struct mem_cgroup *memcg)
2532 return memcg ? memcg->kmemcg_id : -1;
2535 static int memcg_alloc_cache_id(void)
2540 id = ida_simple_get(&memcg_cache_ida,
2541 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2545 if (id < memcg_nr_cache_ids)
2549 * There's no space for the new id in memcg_caches arrays,
2550 * so we have to grow them.
2553 size = 2 * (id + 1);
2554 if (size < MEMCG_CACHES_MIN_SIZE)
2555 size = MEMCG_CACHES_MIN_SIZE;
2556 else if (size > MEMCG_CACHES_MAX_SIZE)
2557 size = MEMCG_CACHES_MAX_SIZE;
2559 err = memcg_update_all_caches(size);
2561 ida_simple_remove(&memcg_cache_ida, id);
2567 static void memcg_free_cache_id(int id)
2569 ida_simple_remove(&memcg_cache_ida, id);
2573 * We should update the current array size iff all caches updates succeed. This
2574 * can only be done from the slab side. The slab mutex needs to be held when
2577 void memcg_update_array_size(int num)
2579 memcg_nr_cache_ids = num;
2582 struct memcg_kmem_cache_create_work {
2583 struct mem_cgroup *memcg;
2584 struct kmem_cache *cachep;
2585 struct work_struct work;
2588 static void memcg_kmem_cache_create_func(struct work_struct *w)
2590 struct memcg_kmem_cache_create_work *cw =
2591 container_of(w, struct memcg_kmem_cache_create_work, work);
2592 struct mem_cgroup *memcg = cw->memcg;
2593 struct kmem_cache *cachep = cw->cachep;
2595 memcg_create_kmem_cache(memcg, cachep);
2597 css_put(&memcg->css);
2602 * Enqueue the creation of a per-memcg kmem_cache.
2604 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2605 struct kmem_cache *cachep)
2607 struct memcg_kmem_cache_create_work *cw;
2609 cw = kmalloc(sizeof(*cw), GFP_NOWAIT);
2613 css_get(&memcg->css);
2616 cw->cachep = cachep;
2617 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2619 schedule_work(&cw->work);
2622 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2623 struct kmem_cache *cachep)
2626 * We need to stop accounting when we kmalloc, because if the
2627 * corresponding kmalloc cache is not yet created, the first allocation
2628 * in __memcg_schedule_kmem_cache_create will recurse.
2630 * However, it is better to enclose the whole function. Depending on
2631 * the debugging options enabled, INIT_WORK(), for instance, can
2632 * trigger an allocation. This too, will make us recurse. Because at
2633 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2634 * the safest choice is to do it like this, wrapping the whole function.
2636 current->memcg_kmem_skip_account = 1;
2637 __memcg_schedule_kmem_cache_create(memcg, cachep);
2638 current->memcg_kmem_skip_account = 0;
2642 * Return the kmem_cache we're supposed to use for a slab allocation.
2643 * We try to use the current memcg's version of the cache.
2645 * If the cache does not exist yet, if we are the first user of it,
2646 * we either create it immediately, if possible, or create it asynchronously
2648 * In the latter case, we will let the current allocation go through with
2649 * the original cache.
2651 * Can't be called in interrupt context or from kernel threads.
2652 * This function needs to be called with rcu_read_lock() held.
2654 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep)
2656 struct mem_cgroup *memcg;
2657 struct kmem_cache *memcg_cachep;
2659 VM_BUG_ON(!cachep->memcg_params);
2660 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
2662 if (current->memcg_kmem_skip_account)
2665 memcg = get_mem_cgroup_from_mm(current->mm);
2666 if (!memcg_kmem_is_active(memcg))
2669 memcg_cachep = cache_from_memcg_idx(cachep, memcg_cache_id(memcg));
2670 if (likely(memcg_cachep))
2671 return memcg_cachep;
2674 * If we are in a safe context (can wait, and not in interrupt
2675 * context), we could be be predictable and return right away.
2676 * This would guarantee that the allocation being performed
2677 * already belongs in the new cache.
2679 * However, there are some clashes that can arrive from locking.
2680 * For instance, because we acquire the slab_mutex while doing
2681 * memcg_create_kmem_cache, this means no further allocation
2682 * could happen with the slab_mutex held. So it's better to
2685 memcg_schedule_kmem_cache_create(memcg, cachep);
2687 css_put(&memcg->css);
2691 void __memcg_kmem_put_cache(struct kmem_cache *cachep)
2693 if (!is_root_cache(cachep))
2694 css_put(&cachep->memcg_params->memcg->css);
2698 * We need to verify if the allocation against current->mm->owner's memcg is
2699 * possible for the given order. But the page is not allocated yet, so we'll
2700 * need a further commit step to do the final arrangements.
2702 * It is possible for the task to switch cgroups in this mean time, so at
2703 * commit time, we can't rely on task conversion any longer. We'll then use
2704 * the handle argument to return to the caller which cgroup we should commit
2705 * against. We could also return the memcg directly and avoid the pointer
2706 * passing, but a boolean return value gives better semantics considering
2707 * the compiled-out case as well.
2709 * Returning true means the allocation is possible.
2712 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
2714 struct mem_cgroup *memcg;
2719 memcg = get_mem_cgroup_from_mm(current->mm);
2721 if (!memcg_kmem_is_active(memcg)) {
2722 css_put(&memcg->css);
2726 ret = memcg_charge_kmem(memcg, gfp, 1 << order);
2730 css_put(&memcg->css);
2734 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
2737 VM_BUG_ON(mem_cgroup_is_root(memcg));
2739 /* The page allocation failed. Revert */
2741 memcg_uncharge_kmem(memcg, 1 << order);
2744 page->mem_cgroup = memcg;
2747 void __memcg_kmem_uncharge_pages(struct page *page, int order)
2749 struct mem_cgroup *memcg = page->mem_cgroup;
2754 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2756 memcg_uncharge_kmem(memcg, 1 << order);
2757 page->mem_cgroup = NULL;
2759 #endif /* CONFIG_MEMCG_KMEM */
2761 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2764 * Because tail pages are not marked as "used", set it. We're under
2765 * zone->lru_lock, 'splitting on pmd' and compound_lock.
2766 * charge/uncharge will be never happen and move_account() is done under
2767 * compound_lock(), so we don't have to take care of races.
2769 void mem_cgroup_split_huge_fixup(struct page *head)
2773 if (mem_cgroup_disabled())
2776 for (i = 1; i < HPAGE_PMD_NR; i++)
2777 head[i].mem_cgroup = head->mem_cgroup;
2779 __this_cpu_sub(head->mem_cgroup->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
2782 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2785 * mem_cgroup_move_account - move account of the page
2787 * @nr_pages: number of regular pages (>1 for huge pages)
2788 * @from: mem_cgroup which the page is moved from.
2789 * @to: mem_cgroup which the page is moved to. @from != @to.
2791 * The caller must confirm following.
2792 * - page is not on LRU (isolate_page() is useful.)
2793 * - compound_lock is held when nr_pages > 1
2795 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
2798 static int mem_cgroup_move_account(struct page *page,
2799 unsigned int nr_pages,
2800 struct mem_cgroup *from,
2801 struct mem_cgroup *to)
2803 unsigned long flags;
2806 VM_BUG_ON(from == to);
2807 VM_BUG_ON_PAGE(PageLRU(page), page);
2809 * The page is isolated from LRU. So, collapse function
2810 * will not handle this page. But page splitting can happen.
2811 * Do this check under compound_page_lock(). The caller should
2815 if (nr_pages > 1 && !PageTransHuge(page))
2819 * Prevent mem_cgroup_migrate() from looking at page->mem_cgroup
2820 * of its source page while we change it: page migration takes
2821 * both pages off the LRU, but page cache replacement doesn't.
2823 if (!trylock_page(page))
2827 if (page->mem_cgroup != from)
2830 spin_lock_irqsave(&from->move_lock, flags);
2832 if (!PageAnon(page) && page_mapped(page)) {
2833 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
2835 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
2839 if (PageWriteback(page)) {
2840 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK],
2842 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK],
2847 * It is safe to change page->mem_cgroup here because the page
2848 * is referenced, charged, and isolated - we can't race with
2849 * uncharging, charging, migration, or LRU putback.
2852 /* caller should have done css_get */
2853 page->mem_cgroup = to;
2854 spin_unlock_irqrestore(&from->move_lock, flags);
2858 local_irq_disable();
2859 mem_cgroup_charge_statistics(to, page, nr_pages);
2860 memcg_check_events(to, page);
2861 mem_cgroup_charge_statistics(from, page, -nr_pages);
2862 memcg_check_events(from, page);
2870 #ifdef CONFIG_MEMCG_SWAP
2871 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
2874 int val = (charge) ? 1 : -1;
2875 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
2879 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2880 * @entry: swap entry to be moved
2881 * @from: mem_cgroup which the entry is moved from
2882 * @to: mem_cgroup which the entry is moved to
2884 * It succeeds only when the swap_cgroup's record for this entry is the same
2885 * as the mem_cgroup's id of @from.
2887 * Returns 0 on success, -EINVAL on failure.
2889 * The caller must have charged to @to, IOW, called page_counter_charge() about
2890 * both res and memsw, and called css_get().
2892 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2893 struct mem_cgroup *from, struct mem_cgroup *to)
2895 unsigned short old_id, new_id;
2897 old_id = mem_cgroup_id(from);
2898 new_id = mem_cgroup_id(to);
2900 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2901 mem_cgroup_swap_statistics(from, false);
2902 mem_cgroup_swap_statistics(to, true);
2908 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2909 struct mem_cgroup *from, struct mem_cgroup *to)
2915 static DEFINE_MUTEX(memcg_limit_mutex);
2917 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
2918 unsigned long limit)
2920 unsigned long curusage;
2921 unsigned long oldusage;
2922 bool enlarge = false;
2927 * For keeping hierarchical_reclaim simple, how long we should retry
2928 * is depends on callers. We set our retry-count to be function
2929 * of # of children which we should visit in this loop.
2931 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2932 mem_cgroup_count_children(memcg);
2934 oldusage = page_counter_read(&memcg->memory);
2937 if (signal_pending(current)) {
2942 mutex_lock(&memcg_limit_mutex);
2943 if (limit > memcg->memsw.limit) {
2944 mutex_unlock(&memcg_limit_mutex);
2948 if (limit > memcg->memory.limit)
2950 ret = page_counter_limit(&memcg->memory, limit);
2951 mutex_unlock(&memcg_limit_mutex);
2956 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, true);
2958 curusage = page_counter_read(&memcg->memory);
2959 /* Usage is reduced ? */
2960 if (curusage >= oldusage)
2963 oldusage = curusage;
2964 } while (retry_count);
2966 if (!ret && enlarge)
2967 memcg_oom_recover(memcg);
2972 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
2973 unsigned long limit)
2975 unsigned long curusage;
2976 unsigned long oldusage;
2977 bool enlarge = false;
2981 /* see mem_cgroup_resize_res_limit */
2982 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2983 mem_cgroup_count_children(memcg);
2985 oldusage = page_counter_read(&memcg->memsw);
2988 if (signal_pending(current)) {
2993 mutex_lock(&memcg_limit_mutex);
2994 if (limit < memcg->memory.limit) {
2995 mutex_unlock(&memcg_limit_mutex);
2999 if (limit > memcg->memsw.limit)
3001 ret = page_counter_limit(&memcg->memsw, limit);
3002 mutex_unlock(&memcg_limit_mutex);
3007 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, false);
3009 curusage = page_counter_read(&memcg->memsw);
3010 /* Usage is reduced ? */
3011 if (curusage >= oldusage)
3014 oldusage = curusage;
3015 } while (retry_count);
3017 if (!ret && enlarge)
3018 memcg_oom_recover(memcg);
3023 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3025 unsigned long *total_scanned)
3027 unsigned long nr_reclaimed = 0;
3028 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3029 unsigned long reclaimed;
3031 struct mem_cgroup_tree_per_zone *mctz;
3032 unsigned long excess;
3033 unsigned long nr_scanned;
3038 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3040 * This loop can run a while, specially if mem_cgroup's continuously
3041 * keep exceeding their soft limit and putting the system under
3048 mz = mem_cgroup_largest_soft_limit_node(mctz);
3053 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
3054 gfp_mask, &nr_scanned);
3055 nr_reclaimed += reclaimed;
3056 *total_scanned += nr_scanned;
3057 spin_lock_irq(&mctz->lock);
3058 __mem_cgroup_remove_exceeded(mz, mctz);
3061 * If we failed to reclaim anything from this memory cgroup
3062 * it is time to move on to the next cgroup
3066 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3068 excess = soft_limit_excess(mz->memcg);
3070 * One school of thought says that we should not add
3071 * back the node to the tree if reclaim returns 0.
3072 * But our reclaim could return 0, simply because due
3073 * to priority we are exposing a smaller subset of
3074 * memory to reclaim from. Consider this as a longer
3077 /* If excess == 0, no tree ops */
3078 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3079 spin_unlock_irq(&mctz->lock);
3080 css_put(&mz->memcg->css);
3083 * Could not reclaim anything and there are no more
3084 * mem cgroups to try or we seem to be looping without
3085 * reclaiming anything.
3087 if (!nr_reclaimed &&
3089 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3091 } while (!nr_reclaimed);
3093 css_put(&next_mz->memcg->css);
3094 return nr_reclaimed;
3098 * Test whether @memcg has children, dead or alive. Note that this
3099 * function doesn't care whether @memcg has use_hierarchy enabled and
3100 * returns %true if there are child csses according to the cgroup
3101 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
3103 static inline bool memcg_has_children(struct mem_cgroup *memcg)
3108 * The lock does not prevent addition or deletion of children, but
3109 * it prevents a new child from being initialized based on this
3110 * parent in css_online(), so it's enough to decide whether
3111 * hierarchically inherited attributes can still be changed or not.
3113 lockdep_assert_held(&memcg_create_mutex);
3116 ret = css_next_child(NULL, &memcg->css);
3122 * Reclaims as many pages from the given memcg as possible and moves
3123 * the rest to the parent.
3125 * Caller is responsible for holding css reference for memcg.
3127 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3129 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3131 /* we call try-to-free pages for make this cgroup empty */
3132 lru_add_drain_all();
3133 /* try to free all pages in this cgroup */
3134 while (nr_retries && page_counter_read(&memcg->memory)) {
3137 if (signal_pending(current))
3140 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3144 /* maybe some writeback is necessary */
3145 congestion_wait(BLK_RW_ASYNC, HZ/10);
3153 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3154 char *buf, size_t nbytes,
3157 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3159 if (mem_cgroup_is_root(memcg))
3161 return mem_cgroup_force_empty(memcg) ?: nbytes;
3164 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3167 return mem_cgroup_from_css(css)->use_hierarchy;
3170 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3171 struct cftype *cft, u64 val)
3174 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3175 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
3177 mutex_lock(&memcg_create_mutex);
3179 if (memcg->use_hierarchy == val)
3183 * If parent's use_hierarchy is set, we can't make any modifications
3184 * in the child subtrees. If it is unset, then the change can
3185 * occur, provided the current cgroup has no children.
3187 * For the root cgroup, parent_mem is NULL, we allow value to be
3188 * set if there are no children.
3190 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3191 (val == 1 || val == 0)) {
3192 if (!memcg_has_children(memcg))
3193 memcg->use_hierarchy = val;
3200 mutex_unlock(&memcg_create_mutex);
3205 static unsigned long tree_stat(struct mem_cgroup *memcg,
3206 enum mem_cgroup_stat_index idx)
3208 struct mem_cgroup *iter;
3211 /* Per-cpu values can be negative, use a signed accumulator */
3212 for_each_mem_cgroup_tree(iter, memcg)
3213 val += mem_cgroup_read_stat(iter, idx);
3215 if (val < 0) /* race ? */
3220 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3224 if (mem_cgroup_is_root(memcg)) {
3225 val = tree_stat(memcg, MEM_CGROUP_STAT_CACHE);
3226 val += tree_stat(memcg, MEM_CGROUP_STAT_RSS);
3228 val += tree_stat(memcg, MEM_CGROUP_STAT_SWAP);
3231 val = page_counter_read(&memcg->memory);
3233 val = page_counter_read(&memcg->memsw);
3235 return val << PAGE_SHIFT;
3246 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3249 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3250 struct page_counter *counter;
3252 switch (MEMFILE_TYPE(cft->private)) {
3254 counter = &memcg->memory;
3257 counter = &memcg->memsw;
3260 counter = &memcg->kmem;
3266 switch (MEMFILE_ATTR(cft->private)) {
3268 if (counter == &memcg->memory)
3269 return mem_cgroup_usage(memcg, false);
3270 if (counter == &memcg->memsw)
3271 return mem_cgroup_usage(memcg, true);
3272 return (u64)page_counter_read(counter) * PAGE_SIZE;
3274 return (u64)counter->limit * PAGE_SIZE;
3276 return (u64)counter->watermark * PAGE_SIZE;
3278 return counter->failcnt;
3279 case RES_SOFT_LIMIT:
3280 return (u64)memcg->soft_limit * PAGE_SIZE;
3286 #ifdef CONFIG_MEMCG_KMEM
3287 static int memcg_activate_kmem(struct mem_cgroup *memcg,
3288 unsigned long nr_pages)
3293 if (memcg_kmem_is_active(memcg))
3297 * For simplicity, we won't allow this to be disabled. It also can't
3298 * be changed if the cgroup has children already, or if tasks had
3301 * If tasks join before we set the limit, a person looking at
3302 * kmem.usage_in_bytes will have no way to determine when it took
3303 * place, which makes the value quite meaningless.
3305 * After it first became limited, changes in the value of the limit are
3306 * of course permitted.
3308 mutex_lock(&memcg_create_mutex);
3309 if (cgroup_has_tasks(memcg->css.cgroup) ||
3310 (memcg->use_hierarchy && memcg_has_children(memcg)))
3312 mutex_unlock(&memcg_create_mutex);
3316 memcg_id = memcg_alloc_cache_id();
3323 * We couldn't have accounted to this cgroup, because it hasn't got
3324 * activated yet, so this should succeed.
3326 err = page_counter_limit(&memcg->kmem, nr_pages);
3329 static_key_slow_inc(&memcg_kmem_enabled_key);
3331 * A memory cgroup is considered kmem-active as soon as it gets
3332 * kmemcg_id. Setting the id after enabling static branching will
3333 * guarantee no one starts accounting before all call sites are
3336 memcg->kmemcg_id = memcg_id;
3341 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
3342 unsigned long limit)
3346 mutex_lock(&memcg_limit_mutex);
3347 if (!memcg_kmem_is_active(memcg))
3348 ret = memcg_activate_kmem(memcg, limit);
3350 ret = page_counter_limit(&memcg->kmem, limit);
3351 mutex_unlock(&memcg_limit_mutex);
3355 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
3358 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
3363 mutex_lock(&memcg_limit_mutex);
3365 * If the parent cgroup is not kmem-active now, it cannot be activated
3366 * after this point, because it has at least one child already.
3368 if (memcg_kmem_is_active(parent))
3369 ret = memcg_activate_kmem(memcg, PAGE_COUNTER_MAX);
3370 mutex_unlock(&memcg_limit_mutex);
3374 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
3375 unsigned long limit)
3379 #endif /* CONFIG_MEMCG_KMEM */
3382 * The user of this function is...
3385 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3386 char *buf, size_t nbytes, loff_t off)
3388 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3389 unsigned long nr_pages;
3392 buf = strstrip(buf);
3393 ret = page_counter_memparse(buf, "-1", &nr_pages);
3397 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3399 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3403 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3405 ret = mem_cgroup_resize_limit(memcg, nr_pages);
3408 ret = mem_cgroup_resize_memsw_limit(memcg, nr_pages);
3411 ret = memcg_update_kmem_limit(memcg, nr_pages);
3415 case RES_SOFT_LIMIT:
3416 memcg->soft_limit = nr_pages;
3420 return ret ?: nbytes;
3423 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3424 size_t nbytes, loff_t off)
3426 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3427 struct page_counter *counter;
3429 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3431 counter = &memcg->memory;
3434 counter = &memcg->memsw;
3437 counter = &memcg->kmem;
3443 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3445 page_counter_reset_watermark(counter);
3448 counter->failcnt = 0;
3457 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3460 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3464 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3465 struct cftype *cft, u64 val)
3467 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3469 if (val & ~MOVE_MASK)
3473 * No kind of locking is needed in here, because ->can_attach() will
3474 * check this value once in the beginning of the process, and then carry
3475 * on with stale data. This means that changes to this value will only
3476 * affect task migrations starting after the change.
3478 memcg->move_charge_at_immigrate = val;
3482 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3483 struct cftype *cft, u64 val)
3490 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3494 unsigned int lru_mask;
3497 static const struct numa_stat stats[] = {
3498 { "total", LRU_ALL },
3499 { "file", LRU_ALL_FILE },
3500 { "anon", LRU_ALL_ANON },
3501 { "unevictable", BIT(LRU_UNEVICTABLE) },
3503 const struct numa_stat *stat;
3506 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3508 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3509 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3510 seq_printf(m, "%s=%lu", stat->name, nr);
3511 for_each_node_state(nid, N_MEMORY) {
3512 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3514 seq_printf(m, " N%d=%lu", nid, nr);
3519 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3520 struct mem_cgroup *iter;
3523 for_each_mem_cgroup_tree(iter, memcg)
3524 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3525 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3526 for_each_node_state(nid, N_MEMORY) {
3528 for_each_mem_cgroup_tree(iter, memcg)
3529 nr += mem_cgroup_node_nr_lru_pages(
3530 iter, nid, stat->lru_mask);
3531 seq_printf(m, " N%d=%lu", nid, nr);
3538 #endif /* CONFIG_NUMA */
3540 static int memcg_stat_show(struct seq_file *m, void *v)
3542 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3543 unsigned long memory, memsw;
3544 struct mem_cgroup *mi;
3547 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_stat_names) !=
3548 MEM_CGROUP_STAT_NSTATS);
3549 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_events_names) !=
3550 MEM_CGROUP_EVENTS_NSTATS);
3551 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3553 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3554 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
3556 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
3557 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
3560 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
3561 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
3562 mem_cgroup_read_events(memcg, i));
3564 for (i = 0; i < NR_LRU_LISTS; i++)
3565 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3566 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
3568 /* Hierarchical information */
3569 memory = memsw = PAGE_COUNTER_MAX;
3570 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3571 memory = min(memory, mi->memory.limit);
3572 memsw = min(memsw, mi->memsw.limit);
3574 seq_printf(m, "hierarchical_memory_limit %llu\n",
3575 (u64)memory * PAGE_SIZE);
3576 if (do_swap_account)
3577 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3578 (u64)memsw * PAGE_SIZE);
3580 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3583 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
3585 for_each_mem_cgroup_tree(mi, memcg)
3586 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
3587 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
3590 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
3591 unsigned long long val = 0;
3593 for_each_mem_cgroup_tree(mi, memcg)
3594 val += mem_cgroup_read_events(mi, i);
3595 seq_printf(m, "total_%s %llu\n",
3596 mem_cgroup_events_names[i], val);
3599 for (i = 0; i < NR_LRU_LISTS; i++) {
3600 unsigned long long val = 0;
3602 for_each_mem_cgroup_tree(mi, memcg)
3603 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
3604 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
3607 #ifdef CONFIG_DEBUG_VM
3610 struct mem_cgroup_per_zone *mz;
3611 struct zone_reclaim_stat *rstat;
3612 unsigned long recent_rotated[2] = {0, 0};
3613 unsigned long recent_scanned[2] = {0, 0};
3615 for_each_online_node(nid)
3616 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3617 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
3618 rstat = &mz->lruvec.reclaim_stat;
3620 recent_rotated[0] += rstat->recent_rotated[0];
3621 recent_rotated[1] += rstat->recent_rotated[1];
3622 recent_scanned[0] += rstat->recent_scanned[0];
3623 recent_scanned[1] += rstat->recent_scanned[1];
3625 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3626 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3627 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3628 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3635 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3638 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3640 return mem_cgroup_swappiness(memcg);
3643 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3644 struct cftype *cft, u64 val)
3646 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3652 memcg->swappiness = val;
3654 vm_swappiness = val;
3659 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3661 struct mem_cgroup_threshold_ary *t;
3662 unsigned long usage;
3667 t = rcu_dereference(memcg->thresholds.primary);
3669 t = rcu_dereference(memcg->memsw_thresholds.primary);
3674 usage = mem_cgroup_usage(memcg, swap);
3677 * current_threshold points to threshold just below or equal to usage.
3678 * If it's not true, a threshold was crossed after last
3679 * call of __mem_cgroup_threshold().
3681 i = t->current_threshold;
3684 * Iterate backward over array of thresholds starting from
3685 * current_threshold and check if a threshold is crossed.
3686 * If none of thresholds below usage is crossed, we read
3687 * only one element of the array here.
3689 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3690 eventfd_signal(t->entries[i].eventfd, 1);
3692 /* i = current_threshold + 1 */
3696 * Iterate forward over array of thresholds starting from
3697 * current_threshold+1 and check if a threshold is crossed.
3698 * If none of thresholds above usage is crossed, we read
3699 * only one element of the array here.
3701 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3702 eventfd_signal(t->entries[i].eventfd, 1);
3704 /* Update current_threshold */
3705 t->current_threshold = i - 1;
3710 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3713 __mem_cgroup_threshold(memcg, false);
3714 if (do_swap_account)
3715 __mem_cgroup_threshold(memcg, true);
3717 memcg = parent_mem_cgroup(memcg);
3721 static int compare_thresholds(const void *a, const void *b)
3723 const struct mem_cgroup_threshold *_a = a;
3724 const struct mem_cgroup_threshold *_b = b;
3726 if (_a->threshold > _b->threshold)
3729 if (_a->threshold < _b->threshold)
3735 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3737 struct mem_cgroup_eventfd_list *ev;
3739 spin_lock(&memcg_oom_lock);
3741 list_for_each_entry(ev, &memcg->oom_notify, list)
3742 eventfd_signal(ev->eventfd, 1);
3744 spin_unlock(&memcg_oom_lock);
3748 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3750 struct mem_cgroup *iter;
3752 for_each_mem_cgroup_tree(iter, memcg)
3753 mem_cgroup_oom_notify_cb(iter);
3756 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3757 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3759 struct mem_cgroup_thresholds *thresholds;
3760 struct mem_cgroup_threshold_ary *new;
3761 unsigned long threshold;
3762 unsigned long usage;
3765 ret = page_counter_memparse(args, "-1", &threshold);
3769 mutex_lock(&memcg->thresholds_lock);
3772 thresholds = &memcg->thresholds;
3773 usage = mem_cgroup_usage(memcg, false);
3774 } else if (type == _MEMSWAP) {
3775 thresholds = &memcg->memsw_thresholds;
3776 usage = mem_cgroup_usage(memcg, true);
3780 /* Check if a threshold crossed before adding a new one */
3781 if (thresholds->primary)
3782 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3784 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3786 /* Allocate memory for new array of thresholds */
3787 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3795 /* Copy thresholds (if any) to new array */
3796 if (thresholds->primary) {
3797 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3798 sizeof(struct mem_cgroup_threshold));
3801 /* Add new threshold */
3802 new->entries[size - 1].eventfd = eventfd;
3803 new->entries[size - 1].threshold = threshold;
3805 /* Sort thresholds. Registering of new threshold isn't time-critical */
3806 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3807 compare_thresholds, NULL);
3809 /* Find current threshold */
3810 new->current_threshold = -1;
3811 for (i = 0; i < size; i++) {
3812 if (new->entries[i].threshold <= usage) {
3814 * new->current_threshold will not be used until
3815 * rcu_assign_pointer(), so it's safe to increment
3818 ++new->current_threshold;
3823 /* Free old spare buffer and save old primary buffer as spare */
3824 kfree(thresholds->spare);
3825 thresholds->spare = thresholds->primary;
3827 rcu_assign_pointer(thresholds->primary, new);
3829 /* To be sure that nobody uses thresholds */
3833 mutex_unlock(&memcg->thresholds_lock);
3838 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3839 struct eventfd_ctx *eventfd, const char *args)
3841 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
3844 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
3845 struct eventfd_ctx *eventfd, const char *args)
3847 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
3850 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3851 struct eventfd_ctx *eventfd, enum res_type type)
3853 struct mem_cgroup_thresholds *thresholds;
3854 struct mem_cgroup_threshold_ary *new;
3855 unsigned long usage;
3858 mutex_lock(&memcg->thresholds_lock);
3861 thresholds = &memcg->thresholds;
3862 usage = mem_cgroup_usage(memcg, false);
3863 } else if (type == _MEMSWAP) {
3864 thresholds = &memcg->memsw_thresholds;
3865 usage = mem_cgroup_usage(memcg, true);
3869 if (!thresholds->primary)
3872 /* Check if a threshold crossed before removing */
3873 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3875 /* Calculate new number of threshold */
3877 for (i = 0; i < thresholds->primary->size; i++) {
3878 if (thresholds->primary->entries[i].eventfd != eventfd)
3882 new = thresholds->spare;
3884 /* Set thresholds array to NULL if we don't have thresholds */
3893 /* Copy thresholds and find current threshold */
3894 new->current_threshold = -1;
3895 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3896 if (thresholds->primary->entries[i].eventfd == eventfd)
3899 new->entries[j] = thresholds->primary->entries[i];
3900 if (new->entries[j].threshold <= usage) {
3902 * new->current_threshold will not be used
3903 * until rcu_assign_pointer(), so it's safe to increment
3906 ++new->current_threshold;
3912 /* Swap primary and spare array */
3913 thresholds->spare = thresholds->primary;
3914 /* If all events are unregistered, free the spare array */
3916 kfree(thresholds->spare);
3917 thresholds->spare = NULL;
3920 rcu_assign_pointer(thresholds->primary, new);
3922 /* To be sure that nobody uses thresholds */
3925 mutex_unlock(&memcg->thresholds_lock);
3928 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3929 struct eventfd_ctx *eventfd)
3931 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
3934 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3935 struct eventfd_ctx *eventfd)
3937 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
3940 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
3941 struct eventfd_ctx *eventfd, const char *args)
3943 struct mem_cgroup_eventfd_list *event;
3945 event = kmalloc(sizeof(*event), GFP_KERNEL);
3949 spin_lock(&memcg_oom_lock);
3951 event->eventfd = eventfd;
3952 list_add(&event->list, &memcg->oom_notify);
3954 /* already in OOM ? */
3955 if (atomic_read(&memcg->under_oom))
3956 eventfd_signal(eventfd, 1);
3957 spin_unlock(&memcg_oom_lock);
3962 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
3963 struct eventfd_ctx *eventfd)
3965 struct mem_cgroup_eventfd_list *ev, *tmp;
3967 spin_lock(&memcg_oom_lock);
3969 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
3970 if (ev->eventfd == eventfd) {
3971 list_del(&ev->list);
3976 spin_unlock(&memcg_oom_lock);
3979 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
3981 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
3983 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
3984 seq_printf(sf, "under_oom %d\n", (bool)atomic_read(&memcg->under_oom));
3988 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
3989 struct cftype *cft, u64 val)
3991 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3993 /* cannot set to root cgroup and only 0 and 1 are allowed */
3994 if (!css->parent || !((val == 0) || (val == 1)))
3997 memcg->oom_kill_disable = val;
3999 memcg_oom_recover(memcg);
4004 #ifdef CONFIG_MEMCG_KMEM
4005 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
4009 ret = memcg_propagate_kmem(memcg);
4013 return mem_cgroup_sockets_init(memcg, ss);
4016 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
4018 memcg_destroy_kmem_caches(memcg);
4019 mem_cgroup_sockets_destroy(memcg);
4022 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
4027 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
4033 * DO NOT USE IN NEW FILES.
4035 * "cgroup.event_control" implementation.
4037 * This is way over-engineered. It tries to support fully configurable
4038 * events for each user. Such level of flexibility is completely
4039 * unnecessary especially in the light of the planned unified hierarchy.
4041 * Please deprecate this and replace with something simpler if at all
4046 * Unregister event and free resources.
4048 * Gets called from workqueue.
4050 static void memcg_event_remove(struct work_struct *work)
4052 struct mem_cgroup_event *event =
4053 container_of(work, struct mem_cgroup_event, remove);
4054 struct mem_cgroup *memcg = event->memcg;
4056 remove_wait_queue(event->wqh, &event->wait);
4058 event->unregister_event(memcg, event->eventfd);
4060 /* Notify userspace the event is going away. */
4061 eventfd_signal(event->eventfd, 1);
4063 eventfd_ctx_put(event->eventfd);
4065 css_put(&memcg->css);
4069 * Gets called on POLLHUP on eventfd when user closes it.
4071 * Called with wqh->lock held and interrupts disabled.
4073 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
4074 int sync, void *key)
4076 struct mem_cgroup_event *event =
4077 container_of(wait, struct mem_cgroup_event, wait);
4078 struct mem_cgroup *memcg = event->memcg;
4079 unsigned long flags = (unsigned long)key;
4081 if (flags & POLLHUP) {
4083 * If the event has been detached at cgroup removal, we
4084 * can simply return knowing the other side will cleanup
4087 * We can't race against event freeing since the other
4088 * side will require wqh->lock via remove_wait_queue(),
4091 spin_lock(&memcg->event_list_lock);
4092 if (!list_empty(&event->list)) {
4093 list_del_init(&event->list);
4095 * We are in atomic context, but cgroup_event_remove()
4096 * may sleep, so we have to call it in workqueue.
4098 schedule_work(&event->remove);
4100 spin_unlock(&memcg->event_list_lock);
4106 static void memcg_event_ptable_queue_proc(struct file *file,
4107 wait_queue_head_t *wqh, poll_table *pt)
4109 struct mem_cgroup_event *event =
4110 container_of(pt, struct mem_cgroup_event, pt);
4113 add_wait_queue(wqh, &event->wait);
4117 * DO NOT USE IN NEW FILES.
4119 * Parse input and register new cgroup event handler.
4121 * Input must be in format '<event_fd> <control_fd> <args>'.
4122 * Interpretation of args is defined by control file implementation.
4124 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4125 char *buf, size_t nbytes, loff_t off)
4127 struct cgroup_subsys_state *css = of_css(of);
4128 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4129 struct mem_cgroup_event *event;
4130 struct cgroup_subsys_state *cfile_css;
4131 unsigned int efd, cfd;
4138 buf = strstrip(buf);
4140 efd = simple_strtoul(buf, &endp, 10);
4145 cfd = simple_strtoul(buf, &endp, 10);
4146 if ((*endp != ' ') && (*endp != '\0'))
4150 event = kzalloc(sizeof(*event), GFP_KERNEL);
4154 event->memcg = memcg;
4155 INIT_LIST_HEAD(&event->list);
4156 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4157 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4158 INIT_WORK(&event->remove, memcg_event_remove);
4166 event->eventfd = eventfd_ctx_fileget(efile.file);
4167 if (IS_ERR(event->eventfd)) {
4168 ret = PTR_ERR(event->eventfd);
4175 goto out_put_eventfd;
4178 /* the process need read permission on control file */
4179 /* AV: shouldn't we check that it's been opened for read instead? */
4180 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4185 * Determine the event callbacks and set them in @event. This used
4186 * to be done via struct cftype but cgroup core no longer knows
4187 * about these events. The following is crude but the whole thing
4188 * is for compatibility anyway.
4190 * DO NOT ADD NEW FILES.
4192 name = cfile.file->f_path.dentry->d_name.name;
4194 if (!strcmp(name, "memory.usage_in_bytes")) {
4195 event->register_event = mem_cgroup_usage_register_event;
4196 event->unregister_event = mem_cgroup_usage_unregister_event;
4197 } else if (!strcmp(name, "memory.oom_control")) {
4198 event->register_event = mem_cgroup_oom_register_event;
4199 event->unregister_event = mem_cgroup_oom_unregister_event;
4200 } else if (!strcmp(name, "memory.pressure_level")) {
4201 event->register_event = vmpressure_register_event;
4202 event->unregister_event = vmpressure_unregister_event;
4203 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4204 event->register_event = memsw_cgroup_usage_register_event;
4205 event->unregister_event = memsw_cgroup_usage_unregister_event;
4212 * Verify @cfile should belong to @css. Also, remaining events are
4213 * automatically removed on cgroup destruction but the removal is
4214 * asynchronous, so take an extra ref on @css.
4216 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4217 &memory_cgrp_subsys);
4219 if (IS_ERR(cfile_css))
4221 if (cfile_css != css) {
4226 ret = event->register_event(memcg, event->eventfd, buf);
4230 efile.file->f_op->poll(efile.file, &event->pt);
4232 spin_lock(&memcg->event_list_lock);
4233 list_add(&event->list, &memcg->event_list);
4234 spin_unlock(&memcg->event_list_lock);
4246 eventfd_ctx_put(event->eventfd);
4255 static struct cftype mem_cgroup_legacy_files[] = {
4257 .name = "usage_in_bytes",
4258 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4259 .read_u64 = mem_cgroup_read_u64,
4262 .name = "max_usage_in_bytes",
4263 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4264 .write = mem_cgroup_reset,
4265 .read_u64 = mem_cgroup_read_u64,
4268 .name = "limit_in_bytes",
4269 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4270 .write = mem_cgroup_write,
4271 .read_u64 = mem_cgroup_read_u64,
4274 .name = "soft_limit_in_bytes",
4275 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4276 .write = mem_cgroup_write,
4277 .read_u64 = mem_cgroup_read_u64,
4281 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4282 .write = mem_cgroup_reset,
4283 .read_u64 = mem_cgroup_read_u64,
4287 .seq_show = memcg_stat_show,
4290 .name = "force_empty",
4291 .write = mem_cgroup_force_empty_write,
4294 .name = "use_hierarchy",
4295 .write_u64 = mem_cgroup_hierarchy_write,
4296 .read_u64 = mem_cgroup_hierarchy_read,
4299 .name = "cgroup.event_control", /* XXX: for compat */
4300 .write = memcg_write_event_control,
4301 .flags = CFTYPE_NO_PREFIX,
4305 .name = "swappiness",
4306 .read_u64 = mem_cgroup_swappiness_read,
4307 .write_u64 = mem_cgroup_swappiness_write,
4310 .name = "move_charge_at_immigrate",
4311 .read_u64 = mem_cgroup_move_charge_read,
4312 .write_u64 = mem_cgroup_move_charge_write,
4315 .name = "oom_control",
4316 .seq_show = mem_cgroup_oom_control_read,
4317 .write_u64 = mem_cgroup_oom_control_write,
4318 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4321 .name = "pressure_level",
4325 .name = "numa_stat",
4326 .seq_show = memcg_numa_stat_show,
4329 #ifdef CONFIG_MEMCG_KMEM
4331 .name = "kmem.limit_in_bytes",
4332 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4333 .write = mem_cgroup_write,
4334 .read_u64 = mem_cgroup_read_u64,
4337 .name = "kmem.usage_in_bytes",
4338 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4339 .read_u64 = mem_cgroup_read_u64,
4342 .name = "kmem.failcnt",
4343 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4344 .write = mem_cgroup_reset,
4345 .read_u64 = mem_cgroup_read_u64,
4348 .name = "kmem.max_usage_in_bytes",
4349 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4350 .write = mem_cgroup_reset,
4351 .read_u64 = mem_cgroup_read_u64,
4353 #ifdef CONFIG_SLABINFO
4355 .name = "kmem.slabinfo",
4356 .seq_start = slab_start,
4357 .seq_next = slab_next,
4358 .seq_stop = slab_stop,
4359 .seq_show = memcg_slab_show,
4363 { }, /* terminate */
4366 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4368 struct mem_cgroup_per_node *pn;
4369 struct mem_cgroup_per_zone *mz;
4370 int zone, tmp = node;
4372 * This routine is called against possible nodes.
4373 * But it's BUG to call kmalloc() against offline node.
4375 * TODO: this routine can waste much memory for nodes which will
4376 * never be onlined. It's better to use memory hotplug callback
4379 if (!node_state(node, N_NORMAL_MEMORY))
4381 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4385 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4386 mz = &pn->zoneinfo[zone];
4387 lruvec_init(&mz->lruvec);
4388 mz->usage_in_excess = 0;
4389 mz->on_tree = false;
4392 memcg->nodeinfo[node] = pn;
4396 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4398 kfree(memcg->nodeinfo[node]);
4401 static struct mem_cgroup *mem_cgroup_alloc(void)
4403 struct mem_cgroup *memcg;
4406 size = sizeof(struct mem_cgroup);
4407 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4409 memcg = kzalloc(size, GFP_KERNEL);
4413 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4416 spin_lock_init(&memcg->pcp_counter_lock);
4425 * At destroying mem_cgroup, references from swap_cgroup can remain.
4426 * (scanning all at force_empty is too costly...)
4428 * Instead of clearing all references at force_empty, we remember
4429 * the number of reference from swap_cgroup and free mem_cgroup when
4430 * it goes down to 0.
4432 * Removal of cgroup itself succeeds regardless of refs from swap.
4435 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4439 mem_cgroup_remove_from_trees(memcg);
4442 free_mem_cgroup_per_zone_info(memcg, node);
4444 free_percpu(memcg->stat);
4446 disarm_static_keys(memcg);
4451 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4453 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4455 if (!memcg->memory.parent)
4457 return mem_cgroup_from_counter(memcg->memory.parent, memory);
4459 EXPORT_SYMBOL(parent_mem_cgroup);
4461 static struct cgroup_subsys_state * __ref
4462 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4464 struct mem_cgroup *memcg;
4465 long error = -ENOMEM;
4468 memcg = mem_cgroup_alloc();
4470 return ERR_PTR(error);
4473 if (alloc_mem_cgroup_per_zone_info(memcg, node))
4477 if (parent_css == NULL) {
4478 root_mem_cgroup = memcg;
4479 page_counter_init(&memcg->memory, NULL);
4480 memcg->high = PAGE_COUNTER_MAX;
4481 memcg->soft_limit = PAGE_COUNTER_MAX;
4482 page_counter_init(&memcg->memsw, NULL);
4483 page_counter_init(&memcg->kmem, NULL);
4486 memcg->last_scanned_node = MAX_NUMNODES;
4487 INIT_LIST_HEAD(&memcg->oom_notify);
4488 memcg->move_charge_at_immigrate = 0;
4489 mutex_init(&memcg->thresholds_lock);
4490 spin_lock_init(&memcg->move_lock);
4491 vmpressure_init(&memcg->vmpressure);
4492 INIT_LIST_HEAD(&memcg->event_list);
4493 spin_lock_init(&memcg->event_list_lock);
4494 #ifdef CONFIG_MEMCG_KMEM
4495 memcg->kmemcg_id = -1;
4501 __mem_cgroup_free(memcg);
4502 return ERR_PTR(error);
4506 mem_cgroup_css_online(struct cgroup_subsys_state *css)
4508 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4509 struct mem_cgroup *parent = mem_cgroup_from_css(css->parent);
4512 if (css->id > MEM_CGROUP_ID_MAX)
4518 mutex_lock(&memcg_create_mutex);
4520 memcg->use_hierarchy = parent->use_hierarchy;
4521 memcg->oom_kill_disable = parent->oom_kill_disable;
4522 memcg->swappiness = mem_cgroup_swappiness(parent);
4524 if (parent->use_hierarchy) {
4525 page_counter_init(&memcg->memory, &parent->memory);
4526 memcg->high = PAGE_COUNTER_MAX;
4527 memcg->soft_limit = PAGE_COUNTER_MAX;
4528 page_counter_init(&memcg->memsw, &parent->memsw);
4529 page_counter_init(&memcg->kmem, &parent->kmem);
4532 * No need to take a reference to the parent because cgroup
4533 * core guarantees its existence.
4536 page_counter_init(&memcg->memory, NULL);
4537 memcg->high = PAGE_COUNTER_MAX;
4538 memcg->soft_limit = PAGE_COUNTER_MAX;
4539 page_counter_init(&memcg->memsw, NULL);
4540 page_counter_init(&memcg->kmem, NULL);
4542 * Deeper hierachy with use_hierarchy == false doesn't make
4543 * much sense so let cgroup subsystem know about this
4544 * unfortunate state in our controller.
4546 if (parent != root_mem_cgroup)
4547 memory_cgrp_subsys.broken_hierarchy = true;
4549 mutex_unlock(&memcg_create_mutex);
4551 ret = memcg_init_kmem(memcg, &memory_cgrp_subsys);
4556 * Make sure the memcg is initialized: mem_cgroup_iter()
4557 * orders reading memcg->initialized against its callers
4558 * reading the memcg members.
4560 smp_store_release(&memcg->initialized, 1);
4565 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4567 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4568 struct mem_cgroup_event *event, *tmp;
4571 * Unregister events and notify userspace.
4572 * Notify userspace about cgroup removing only after rmdir of cgroup
4573 * directory to avoid race between userspace and kernelspace.
4575 spin_lock(&memcg->event_list_lock);
4576 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4577 list_del_init(&event->list);
4578 schedule_work(&event->remove);
4580 spin_unlock(&memcg->event_list_lock);
4582 vmpressure_cleanup(&memcg->vmpressure);
4585 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4587 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4589 memcg_destroy_kmem(memcg);
4590 __mem_cgroup_free(memcg);
4594 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4595 * @css: the target css
4597 * Reset the states of the mem_cgroup associated with @css. This is
4598 * invoked when the userland requests disabling on the default hierarchy
4599 * but the memcg is pinned through dependency. The memcg should stop
4600 * applying policies and should revert to the vanilla state as it may be
4601 * made visible again.
4603 * The current implementation only resets the essential configurations.
4604 * This needs to be expanded to cover all the visible parts.
4606 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4608 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4610 mem_cgroup_resize_limit(memcg, PAGE_COUNTER_MAX);
4611 mem_cgroup_resize_memsw_limit(memcg, PAGE_COUNTER_MAX);
4612 memcg_update_kmem_limit(memcg, PAGE_COUNTER_MAX);
4614 memcg->high = PAGE_COUNTER_MAX;
4615 memcg->soft_limit = PAGE_COUNTER_MAX;
4619 /* Handlers for move charge at task migration. */
4620 static int mem_cgroup_do_precharge(unsigned long count)
4624 /* Try a single bulk charge without reclaim first */
4625 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_WAIT, count);
4627 mc.precharge += count;
4630 if (ret == -EINTR) {
4631 cancel_charge(root_mem_cgroup, count);
4635 /* Try charges one by one with reclaim */
4637 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_NORETRY, 1);
4639 * In case of failure, any residual charges against
4640 * mc.to will be dropped by mem_cgroup_clear_mc()
4641 * later on. However, cancel any charges that are
4642 * bypassed to root right away or they'll be lost.
4645 cancel_charge(root_mem_cgroup, 1);
4655 * get_mctgt_type - get target type of moving charge
4656 * @vma: the vma the pte to be checked belongs
4657 * @addr: the address corresponding to the pte to be checked
4658 * @ptent: the pte to be checked
4659 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4662 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4663 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4664 * move charge. if @target is not NULL, the page is stored in target->page
4665 * with extra refcnt got(Callers should handle it).
4666 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4667 * target for charge migration. if @target is not NULL, the entry is stored
4670 * Called with pte lock held.
4677 enum mc_target_type {
4683 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4684 unsigned long addr, pte_t ptent)
4686 struct page *page = vm_normal_page(vma, addr, ptent);
4688 if (!page || !page_mapped(page))
4690 if (PageAnon(page)) {
4691 if (!(mc.flags & MOVE_ANON))
4694 if (!(mc.flags & MOVE_FILE))
4697 if (!get_page_unless_zero(page))
4704 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4705 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4707 struct page *page = NULL;
4708 swp_entry_t ent = pte_to_swp_entry(ptent);
4710 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
4713 * Because lookup_swap_cache() updates some statistics counter,
4714 * we call find_get_page() with swapper_space directly.
4716 page = find_get_page(swap_address_space(ent), ent.val);
4717 if (do_swap_account)
4718 entry->val = ent.val;
4723 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4724 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4730 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4731 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4733 struct page *page = NULL;
4734 struct address_space *mapping;
4737 if (!vma->vm_file) /* anonymous vma */
4739 if (!(mc.flags & MOVE_FILE))
4742 mapping = vma->vm_file->f_mapping;
4743 pgoff = linear_page_index(vma, addr);
4745 /* page is moved even if it's not RSS of this task(page-faulted). */
4747 /* shmem/tmpfs may report page out on swap: account for that too. */
4748 if (shmem_mapping(mapping)) {
4749 page = find_get_entry(mapping, pgoff);
4750 if (radix_tree_exceptional_entry(page)) {
4751 swp_entry_t swp = radix_to_swp_entry(page);
4752 if (do_swap_account)
4754 page = find_get_page(swap_address_space(swp), swp.val);
4757 page = find_get_page(mapping, pgoff);
4759 page = find_get_page(mapping, pgoff);
4764 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
4765 unsigned long addr, pte_t ptent, union mc_target *target)
4767 struct page *page = NULL;
4768 enum mc_target_type ret = MC_TARGET_NONE;
4769 swp_entry_t ent = { .val = 0 };
4771 if (pte_present(ptent))
4772 page = mc_handle_present_pte(vma, addr, ptent);
4773 else if (is_swap_pte(ptent))
4774 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
4775 else if (pte_none(ptent))
4776 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4778 if (!page && !ent.val)
4782 * Do only loose check w/o serialization.
4783 * mem_cgroup_move_account() checks the page is valid or
4784 * not under LRU exclusion.
4786 if (page->mem_cgroup == mc.from) {
4787 ret = MC_TARGET_PAGE;
4789 target->page = page;
4791 if (!ret || !target)
4794 /* There is a swap entry and a page doesn't exist or isn't charged */
4795 if (ent.val && !ret &&
4796 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
4797 ret = MC_TARGET_SWAP;
4804 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4806 * We don't consider swapping or file mapped pages because THP does not
4807 * support them for now.
4808 * Caller should make sure that pmd_trans_huge(pmd) is true.
4810 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4811 unsigned long addr, pmd_t pmd, union mc_target *target)
4813 struct page *page = NULL;
4814 enum mc_target_type ret = MC_TARGET_NONE;
4816 page = pmd_page(pmd);
4817 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
4818 if (!(mc.flags & MOVE_ANON))
4820 if (page->mem_cgroup == mc.from) {
4821 ret = MC_TARGET_PAGE;
4824 target->page = page;
4830 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4831 unsigned long addr, pmd_t pmd, union mc_target *target)
4833 return MC_TARGET_NONE;
4837 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4838 unsigned long addr, unsigned long end,
4839 struct mm_walk *walk)
4841 struct vm_area_struct *vma = walk->vma;
4845 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
4846 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
4847 mc.precharge += HPAGE_PMD_NR;
4852 if (pmd_trans_unstable(pmd))
4854 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4855 for (; addr != end; pte++, addr += PAGE_SIZE)
4856 if (get_mctgt_type(vma, addr, *pte, NULL))
4857 mc.precharge++; /* increment precharge temporarily */
4858 pte_unmap_unlock(pte - 1, ptl);
4864 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4866 unsigned long precharge;
4868 struct mm_walk mem_cgroup_count_precharge_walk = {
4869 .pmd_entry = mem_cgroup_count_precharge_pte_range,
4872 down_read(&mm->mmap_sem);
4873 walk_page_range(0, ~0UL, &mem_cgroup_count_precharge_walk);
4874 up_read(&mm->mmap_sem);
4876 precharge = mc.precharge;
4882 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4884 unsigned long precharge = mem_cgroup_count_precharge(mm);
4886 VM_BUG_ON(mc.moving_task);
4887 mc.moving_task = current;
4888 return mem_cgroup_do_precharge(precharge);
4891 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4892 static void __mem_cgroup_clear_mc(void)
4894 struct mem_cgroup *from = mc.from;
4895 struct mem_cgroup *to = mc.to;
4897 /* we must uncharge all the leftover precharges from mc.to */
4899 cancel_charge(mc.to, mc.precharge);
4903 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4904 * we must uncharge here.
4906 if (mc.moved_charge) {
4907 cancel_charge(mc.from, mc.moved_charge);
4908 mc.moved_charge = 0;
4910 /* we must fixup refcnts and charges */
4911 if (mc.moved_swap) {
4912 /* uncharge swap account from the old cgroup */
4913 if (!mem_cgroup_is_root(mc.from))
4914 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
4917 * we charged both to->memory and to->memsw, so we
4918 * should uncharge to->memory.
4920 if (!mem_cgroup_is_root(mc.to))
4921 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
4923 css_put_many(&mc.from->css, mc.moved_swap);
4925 /* we've already done css_get(mc.to) */
4928 memcg_oom_recover(from);
4929 memcg_oom_recover(to);
4930 wake_up_all(&mc.waitq);
4933 static void mem_cgroup_clear_mc(void)
4936 * we must clear moving_task before waking up waiters at the end of
4939 mc.moving_task = NULL;
4940 __mem_cgroup_clear_mc();
4941 spin_lock(&mc.lock);
4944 spin_unlock(&mc.lock);
4947 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
4948 struct cgroup_taskset *tset)
4950 struct task_struct *p = cgroup_taskset_first(tset);
4952 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4953 unsigned long move_flags;
4956 * We are now commited to this value whatever it is. Changes in this
4957 * tunable will only affect upcoming migrations, not the current one.
4958 * So we need to save it, and keep it going.
4960 move_flags = ACCESS_ONCE(memcg->move_charge_at_immigrate);
4962 struct mm_struct *mm;
4963 struct mem_cgroup *from = mem_cgroup_from_task(p);
4965 VM_BUG_ON(from == memcg);
4967 mm = get_task_mm(p);
4970 /* We move charges only when we move a owner of the mm */
4971 if (mm->owner == p) {
4974 VM_BUG_ON(mc.precharge);
4975 VM_BUG_ON(mc.moved_charge);
4976 VM_BUG_ON(mc.moved_swap);
4978 spin_lock(&mc.lock);
4981 mc.flags = move_flags;
4982 spin_unlock(&mc.lock);
4983 /* We set mc.moving_task later */
4985 ret = mem_cgroup_precharge_mc(mm);
4987 mem_cgroup_clear_mc();
4994 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
4995 struct cgroup_taskset *tset)
4998 mem_cgroup_clear_mc();
5001 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5002 unsigned long addr, unsigned long end,
5003 struct mm_walk *walk)
5006 struct vm_area_struct *vma = walk->vma;
5009 enum mc_target_type target_type;
5010 union mc_target target;
5014 * We don't take compound_lock() here but no race with splitting thp
5016 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
5017 * under splitting, which means there's no concurrent thp split,
5018 * - if another thread runs into split_huge_page() just after we
5019 * entered this if-block, the thread must wait for page table lock
5020 * to be unlocked in __split_huge_page_splitting(), where the main
5021 * part of thp split is not executed yet.
5023 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
5024 if (mc.precharge < HPAGE_PMD_NR) {
5028 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5029 if (target_type == MC_TARGET_PAGE) {
5031 if (!isolate_lru_page(page)) {
5032 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
5034 mc.precharge -= HPAGE_PMD_NR;
5035 mc.moved_charge += HPAGE_PMD_NR;
5037 putback_lru_page(page);
5045 if (pmd_trans_unstable(pmd))
5048 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5049 for (; addr != end; addr += PAGE_SIZE) {
5050 pte_t ptent = *(pte++);
5056 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5057 case MC_TARGET_PAGE:
5059 if (isolate_lru_page(page))
5061 if (!mem_cgroup_move_account(page, 1, mc.from, mc.to)) {
5063 /* we uncharge from mc.from later. */
5066 putback_lru_page(page);
5067 put: /* get_mctgt_type() gets the page */
5070 case MC_TARGET_SWAP:
5072 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5074 /* we fixup refcnts and charges later. */
5082 pte_unmap_unlock(pte - 1, ptl);
5087 * We have consumed all precharges we got in can_attach().
5088 * We try charge one by one, but don't do any additional
5089 * charges to mc.to if we have failed in charge once in attach()
5092 ret = mem_cgroup_do_precharge(1);
5100 static void mem_cgroup_move_charge(struct mm_struct *mm)
5102 struct mm_walk mem_cgroup_move_charge_walk = {
5103 .pmd_entry = mem_cgroup_move_charge_pte_range,
5107 lru_add_drain_all();
5109 * Signal mem_cgroup_begin_page_stat() to take the memcg's
5110 * move_lock while we're moving its pages to another memcg.
5111 * Then wait for already started RCU-only updates to finish.
5113 atomic_inc(&mc.from->moving_account);
5116 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5118 * Someone who are holding the mmap_sem might be waiting in
5119 * waitq. So we cancel all extra charges, wake up all waiters,
5120 * and retry. Because we cancel precharges, we might not be able
5121 * to move enough charges, but moving charge is a best-effort
5122 * feature anyway, so it wouldn't be a big problem.
5124 __mem_cgroup_clear_mc();
5129 * When we have consumed all precharges and failed in doing
5130 * additional charge, the page walk just aborts.
5132 walk_page_range(0, ~0UL, &mem_cgroup_move_charge_walk);
5133 up_read(&mm->mmap_sem);
5134 atomic_dec(&mc.from->moving_account);
5137 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
5138 struct cgroup_taskset *tset)
5140 struct task_struct *p = cgroup_taskset_first(tset);
5141 struct mm_struct *mm = get_task_mm(p);
5145 mem_cgroup_move_charge(mm);
5149 mem_cgroup_clear_mc();
5151 #else /* !CONFIG_MMU */
5152 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
5153 struct cgroup_taskset *tset)
5157 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
5158 struct cgroup_taskset *tset)
5161 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
5162 struct cgroup_taskset *tset)
5168 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5169 * to verify whether we're attached to the default hierarchy on each mount
5172 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5175 * use_hierarchy is forced on the default hierarchy. cgroup core
5176 * guarantees that @root doesn't have any children, so turning it
5177 * on for the root memcg is enough.
5179 if (cgroup_on_dfl(root_css->cgroup))
5180 mem_cgroup_from_css(root_css)->use_hierarchy = true;
5183 static u64 memory_current_read(struct cgroup_subsys_state *css,
5186 return mem_cgroup_usage(mem_cgroup_from_css(css), false);
5189 static int memory_low_show(struct seq_file *m, void *v)
5191 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5192 unsigned long low = ACCESS_ONCE(memcg->low);
5194 if (low == PAGE_COUNTER_MAX)
5195 seq_puts(m, "infinity\n");
5197 seq_printf(m, "%llu\n", (u64)low * PAGE_SIZE);
5202 static ssize_t memory_low_write(struct kernfs_open_file *of,
5203 char *buf, size_t nbytes, loff_t off)
5205 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5209 buf = strstrip(buf);
5210 err = page_counter_memparse(buf, "infinity", &low);
5219 static int memory_high_show(struct seq_file *m, void *v)
5221 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5222 unsigned long high = ACCESS_ONCE(memcg->high);
5224 if (high == PAGE_COUNTER_MAX)
5225 seq_puts(m, "infinity\n");
5227 seq_printf(m, "%llu\n", (u64)high * PAGE_SIZE);
5232 static ssize_t memory_high_write(struct kernfs_open_file *of,
5233 char *buf, size_t nbytes, loff_t off)
5235 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5239 buf = strstrip(buf);
5240 err = page_counter_memparse(buf, "infinity", &high);
5249 static int memory_max_show(struct seq_file *m, void *v)
5251 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5252 unsigned long max = ACCESS_ONCE(memcg->memory.limit);
5254 if (max == PAGE_COUNTER_MAX)
5255 seq_puts(m, "infinity\n");
5257 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
5262 static ssize_t memory_max_write(struct kernfs_open_file *of,
5263 char *buf, size_t nbytes, loff_t off)
5265 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5269 buf = strstrip(buf);
5270 err = page_counter_memparse(buf, "infinity", &max);
5274 err = mem_cgroup_resize_limit(memcg, max);
5281 static int memory_events_show(struct seq_file *m, void *v)
5283 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5285 seq_printf(m, "low %lu\n", mem_cgroup_read_events(memcg, MEMCG_LOW));
5286 seq_printf(m, "high %lu\n", mem_cgroup_read_events(memcg, MEMCG_HIGH));
5287 seq_printf(m, "max %lu\n", mem_cgroup_read_events(memcg, MEMCG_MAX));
5288 seq_printf(m, "oom %lu\n", mem_cgroup_read_events(memcg, MEMCG_OOM));
5293 static struct cftype memory_files[] = {
5296 .read_u64 = memory_current_read,
5300 .flags = CFTYPE_NOT_ON_ROOT,
5301 .seq_show = memory_low_show,
5302 .write = memory_low_write,
5306 .flags = CFTYPE_NOT_ON_ROOT,
5307 .seq_show = memory_high_show,
5308 .write = memory_high_write,
5312 .flags = CFTYPE_NOT_ON_ROOT,
5313 .seq_show = memory_max_show,
5314 .write = memory_max_write,
5318 .flags = CFTYPE_NOT_ON_ROOT,
5319 .seq_show = memory_events_show,
5324 struct cgroup_subsys memory_cgrp_subsys = {
5325 .css_alloc = mem_cgroup_css_alloc,
5326 .css_online = mem_cgroup_css_online,
5327 .css_offline = mem_cgroup_css_offline,
5328 .css_free = mem_cgroup_css_free,
5329 .css_reset = mem_cgroup_css_reset,
5330 .can_attach = mem_cgroup_can_attach,
5331 .cancel_attach = mem_cgroup_cancel_attach,
5332 .attach = mem_cgroup_move_task,
5333 .bind = mem_cgroup_bind,
5334 .dfl_cftypes = memory_files,
5335 .legacy_cftypes = mem_cgroup_legacy_files,
5340 * mem_cgroup_events - count memory events against a cgroup
5341 * @memcg: the memory cgroup
5342 * @idx: the event index
5343 * @nr: the number of events to account for
5345 void mem_cgroup_events(struct mem_cgroup *memcg,
5346 enum mem_cgroup_events_index idx,
5349 this_cpu_add(memcg->stat->events[idx], nr);
5353 * mem_cgroup_low - check if memory consumption is below the normal range
5354 * @root: the highest ancestor to consider
5355 * @memcg: the memory cgroup to check
5357 * Returns %true if memory consumption of @memcg, and that of all
5358 * configurable ancestors up to @root, is below the normal range.
5360 bool mem_cgroup_low(struct mem_cgroup *root, struct mem_cgroup *memcg)
5362 if (mem_cgroup_disabled())
5366 * The toplevel group doesn't have a configurable range, so
5367 * it's never low when looked at directly, and it is not
5368 * considered an ancestor when assessing the hierarchy.
5371 if (memcg == root_mem_cgroup)
5374 if (page_counter_read(&memcg->memory) > memcg->low)
5377 while (memcg != root) {
5378 memcg = parent_mem_cgroup(memcg);
5380 if (memcg == root_mem_cgroup)
5383 if (page_counter_read(&memcg->memory) > memcg->low)
5390 * mem_cgroup_try_charge - try charging a page
5391 * @page: page to charge
5392 * @mm: mm context of the victim
5393 * @gfp_mask: reclaim mode
5394 * @memcgp: charged memcg return
5396 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5397 * pages according to @gfp_mask if necessary.
5399 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5400 * Otherwise, an error code is returned.
5402 * After page->mapping has been set up, the caller must finalize the
5403 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5404 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5406 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
5407 gfp_t gfp_mask, struct mem_cgroup **memcgp)
5409 struct mem_cgroup *memcg = NULL;
5410 unsigned int nr_pages = 1;
5413 if (mem_cgroup_disabled())
5416 if (PageSwapCache(page)) {
5418 * Every swap fault against a single page tries to charge the
5419 * page, bail as early as possible. shmem_unuse() encounters
5420 * already charged pages, too. The USED bit is protected by
5421 * the page lock, which serializes swap cache removal, which
5422 * in turn serializes uncharging.
5424 if (page->mem_cgroup)
5428 if (PageTransHuge(page)) {
5429 nr_pages <<= compound_order(page);
5430 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5433 if (do_swap_account && PageSwapCache(page))
5434 memcg = try_get_mem_cgroup_from_page(page);
5436 memcg = get_mem_cgroup_from_mm(mm);
5438 ret = try_charge(memcg, gfp_mask, nr_pages);
5440 css_put(&memcg->css);
5442 if (ret == -EINTR) {
5443 memcg = root_mem_cgroup;
5452 * mem_cgroup_commit_charge - commit a page charge
5453 * @page: page to charge
5454 * @memcg: memcg to charge the page to
5455 * @lrucare: page might be on LRU already
5457 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5458 * after page->mapping has been set up. This must happen atomically
5459 * as part of the page instantiation, i.e. under the page table lock
5460 * for anonymous pages, under the page lock for page and swap cache.
5462 * In addition, the page must not be on the LRU during the commit, to
5463 * prevent racing with task migration. If it might be, use @lrucare.
5465 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5467 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
5470 unsigned int nr_pages = 1;
5472 VM_BUG_ON_PAGE(!page->mapping, page);
5473 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
5475 if (mem_cgroup_disabled())
5478 * Swap faults will attempt to charge the same page multiple
5479 * times. But reuse_swap_page() might have removed the page
5480 * from swapcache already, so we can't check PageSwapCache().
5485 commit_charge(page, memcg, lrucare);
5487 if (PageTransHuge(page)) {
5488 nr_pages <<= compound_order(page);
5489 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5492 local_irq_disable();
5493 mem_cgroup_charge_statistics(memcg, page, nr_pages);
5494 memcg_check_events(memcg, page);
5497 if (do_swap_account && PageSwapCache(page)) {
5498 swp_entry_t entry = { .val = page_private(page) };
5500 * The swap entry might not get freed for a long time,
5501 * let's not wait for it. The page already received a
5502 * memory+swap charge, drop the swap entry duplicate.
5504 mem_cgroup_uncharge_swap(entry);
5509 * mem_cgroup_cancel_charge - cancel a page charge
5510 * @page: page to charge
5511 * @memcg: memcg to charge the page to
5513 * Cancel a charge transaction started by mem_cgroup_try_charge().
5515 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg)
5517 unsigned int nr_pages = 1;
5519 if (mem_cgroup_disabled())
5522 * Swap faults will attempt to charge the same page multiple
5523 * times. But reuse_swap_page() might have removed the page
5524 * from swapcache already, so we can't check PageSwapCache().
5529 if (PageTransHuge(page)) {
5530 nr_pages <<= compound_order(page);
5531 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5534 cancel_charge(memcg, nr_pages);
5537 static void uncharge_batch(struct mem_cgroup *memcg, unsigned long pgpgout,
5538 unsigned long nr_anon, unsigned long nr_file,
5539 unsigned long nr_huge, struct page *dummy_page)
5541 unsigned long nr_pages = nr_anon + nr_file;
5542 unsigned long flags;
5544 if (!mem_cgroup_is_root(memcg)) {
5545 page_counter_uncharge(&memcg->memory, nr_pages);
5546 if (do_swap_account)
5547 page_counter_uncharge(&memcg->memsw, nr_pages);
5548 memcg_oom_recover(memcg);
5551 local_irq_save(flags);
5552 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS], nr_anon);
5553 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_CACHE], nr_file);
5554 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE], nr_huge);
5555 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT], pgpgout);
5556 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
5557 memcg_check_events(memcg, dummy_page);
5558 local_irq_restore(flags);
5560 if (!mem_cgroup_is_root(memcg))
5561 css_put_many(&memcg->css, nr_pages);
5564 static void uncharge_list(struct list_head *page_list)
5566 struct mem_cgroup *memcg = NULL;
5567 unsigned long nr_anon = 0;
5568 unsigned long nr_file = 0;
5569 unsigned long nr_huge = 0;
5570 unsigned long pgpgout = 0;
5571 struct list_head *next;
5574 next = page_list->next;
5576 unsigned int nr_pages = 1;
5578 page = list_entry(next, struct page, lru);
5579 next = page->lru.next;
5581 VM_BUG_ON_PAGE(PageLRU(page), page);
5582 VM_BUG_ON_PAGE(page_count(page), page);
5584 if (!page->mem_cgroup)
5588 * Nobody should be changing or seriously looking at
5589 * page->mem_cgroup at this point, we have fully
5590 * exclusive access to the page.
5593 if (memcg != page->mem_cgroup) {
5595 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5597 pgpgout = nr_anon = nr_file = nr_huge = 0;
5599 memcg = page->mem_cgroup;
5602 if (PageTransHuge(page)) {
5603 nr_pages <<= compound_order(page);
5604 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5605 nr_huge += nr_pages;
5609 nr_anon += nr_pages;
5611 nr_file += nr_pages;
5613 page->mem_cgroup = NULL;
5616 } while (next != page_list);
5619 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5624 * mem_cgroup_uncharge - uncharge a page
5625 * @page: page to uncharge
5627 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5628 * mem_cgroup_commit_charge().
5630 void mem_cgroup_uncharge(struct page *page)
5632 if (mem_cgroup_disabled())
5635 /* Don't touch page->lru of any random page, pre-check: */
5636 if (!page->mem_cgroup)
5639 INIT_LIST_HEAD(&page->lru);
5640 uncharge_list(&page->lru);
5644 * mem_cgroup_uncharge_list - uncharge a list of page
5645 * @page_list: list of pages to uncharge
5647 * Uncharge a list of pages previously charged with
5648 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5650 void mem_cgroup_uncharge_list(struct list_head *page_list)
5652 if (mem_cgroup_disabled())
5655 if (!list_empty(page_list))
5656 uncharge_list(page_list);
5660 * mem_cgroup_migrate - migrate a charge to another page
5661 * @oldpage: currently charged page
5662 * @newpage: page to transfer the charge to
5663 * @lrucare: either or both pages might be on the LRU already
5665 * Migrate the charge from @oldpage to @newpage.
5667 * Both pages must be locked, @newpage->mapping must be set up.
5669 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage,
5672 struct mem_cgroup *memcg;
5675 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
5676 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
5677 VM_BUG_ON_PAGE(!lrucare && PageLRU(oldpage), oldpage);
5678 VM_BUG_ON_PAGE(!lrucare && PageLRU(newpage), newpage);
5679 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
5680 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
5683 if (mem_cgroup_disabled())
5686 /* Page cache replacement: new page already charged? */
5687 if (newpage->mem_cgroup)
5691 * Swapcache readahead pages can get migrated before being
5692 * charged, and migration from compaction can happen to an
5693 * uncharged page when the PFN walker finds a page that
5694 * reclaim just put back on the LRU but has not released yet.
5696 memcg = oldpage->mem_cgroup;
5701 lock_page_lru(oldpage, &isolated);
5703 oldpage->mem_cgroup = NULL;
5706 unlock_page_lru(oldpage, isolated);
5708 commit_charge(newpage, memcg, lrucare);
5712 * subsys_initcall() for memory controller.
5714 * Some parts like hotcpu_notifier() have to be initialized from this context
5715 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5716 * everything that doesn't depend on a specific mem_cgroup structure should
5717 * be initialized from here.
5719 static int __init mem_cgroup_init(void)
5723 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5725 for_each_possible_cpu(cpu)
5726 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
5729 for_each_node(node) {
5730 struct mem_cgroup_tree_per_node *rtpn;
5733 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
5734 node_online(node) ? node : NUMA_NO_NODE);
5736 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5737 struct mem_cgroup_tree_per_zone *rtpz;
5739 rtpz = &rtpn->rb_tree_per_zone[zone];
5740 rtpz->rb_root = RB_ROOT;
5741 spin_lock_init(&rtpz->lock);
5743 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5748 subsys_initcall(mem_cgroup_init);
5750 #ifdef CONFIG_MEMCG_SWAP
5752 * mem_cgroup_swapout - transfer a memsw charge to swap
5753 * @page: page whose memsw charge to transfer
5754 * @entry: swap entry to move the charge to
5756 * Transfer the memsw charge of @page to @entry.
5758 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
5760 struct mem_cgroup *memcg;
5761 unsigned short oldid;
5763 VM_BUG_ON_PAGE(PageLRU(page), page);
5764 VM_BUG_ON_PAGE(page_count(page), page);
5766 if (!do_swap_account)
5769 memcg = page->mem_cgroup;
5771 /* Readahead page, never charged */
5775 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg));
5776 VM_BUG_ON_PAGE(oldid, page);
5777 mem_cgroup_swap_statistics(memcg, true);
5779 page->mem_cgroup = NULL;
5781 if (!mem_cgroup_is_root(memcg))
5782 page_counter_uncharge(&memcg->memory, 1);
5784 /* XXX: caller holds IRQ-safe mapping->tree_lock */
5785 VM_BUG_ON(!irqs_disabled());
5787 mem_cgroup_charge_statistics(memcg, page, -1);
5788 memcg_check_events(memcg, page);
5792 * mem_cgroup_uncharge_swap - uncharge a swap entry
5793 * @entry: swap entry to uncharge
5795 * Drop the memsw charge associated with @entry.
5797 void mem_cgroup_uncharge_swap(swp_entry_t entry)
5799 struct mem_cgroup *memcg;
5802 if (!do_swap_account)
5805 id = swap_cgroup_record(entry, 0);
5807 memcg = mem_cgroup_lookup(id);
5809 if (!mem_cgroup_is_root(memcg))
5810 page_counter_uncharge(&memcg->memsw, 1);
5811 mem_cgroup_swap_statistics(memcg, false);
5812 css_put(&memcg->css);
5817 /* for remember boot option*/
5818 #ifdef CONFIG_MEMCG_SWAP_ENABLED
5819 static int really_do_swap_account __initdata = 1;
5821 static int really_do_swap_account __initdata;
5824 static int __init enable_swap_account(char *s)
5826 if (!strcmp(s, "1"))
5827 really_do_swap_account = 1;
5828 else if (!strcmp(s, "0"))
5829 really_do_swap_account = 0;
5832 __setup("swapaccount=", enable_swap_account);
5834 static struct cftype memsw_cgroup_files[] = {
5836 .name = "memsw.usage_in_bytes",
5837 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
5838 .read_u64 = mem_cgroup_read_u64,
5841 .name = "memsw.max_usage_in_bytes",
5842 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
5843 .write = mem_cgroup_reset,
5844 .read_u64 = mem_cgroup_read_u64,
5847 .name = "memsw.limit_in_bytes",
5848 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
5849 .write = mem_cgroup_write,
5850 .read_u64 = mem_cgroup_read_u64,
5853 .name = "memsw.failcnt",
5854 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
5855 .write = mem_cgroup_reset,
5856 .read_u64 = mem_cgroup_read_u64,
5858 { }, /* terminate */
5861 static int __init mem_cgroup_swap_init(void)
5863 if (!mem_cgroup_disabled() && really_do_swap_account) {
5864 do_swap_account = 1;
5865 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
5866 memsw_cgroup_files));
5870 subsys_initcall(mem_cgroup_swap_init);
5872 #endif /* CONFIG_MEMCG_SWAP */