1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * Kernel Memory Controller
14 * Copyright (C) 2012 Parallels Inc. and Google Inc.
15 * Authors: Glauber Costa and Suleiman Souhlal
18 * Charge lifetime sanitation
19 * Lockless page tracking & accounting
20 * Unified hierarchy configuration model
21 * Copyright (C) 2015 Red Hat, Inc., Johannes Weiner
23 * This program is free software; you can redistribute it and/or modify
24 * it under the terms of the GNU General Public License as published by
25 * the Free Software Foundation; either version 2 of the License, or
26 * (at your option) any later version.
28 * This program is distributed in the hope that it will be useful,
29 * but WITHOUT ANY WARRANTY; without even the implied warranty of
30 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
31 * GNU General Public License for more details.
34 #include <linux/page_counter.h>
35 #include <linux/memcontrol.h>
36 #include <linux/cgroup.h>
38 #include <linux/hugetlb.h>
39 #include <linux/pagemap.h>
40 #include <linux/smp.h>
41 #include <linux/page-flags.h>
42 #include <linux/backing-dev.h>
43 #include <linux/bit_spinlock.h>
44 #include <linux/rcupdate.h>
45 #include <linux/limits.h>
46 #include <linux/export.h>
47 #include <linux/mutex.h>
48 #include <linux/rbtree.h>
49 #include <linux/slab.h>
50 #include <linux/swap.h>
51 #include <linux/swapops.h>
52 #include <linux/spinlock.h>
53 #include <linux/eventfd.h>
54 #include <linux/poll.h>
55 #include <linux/sort.h>
57 #include <linux/seq_file.h>
58 #include <linux/vmpressure.h>
59 #include <linux/mm_inline.h>
60 #include <linux/swap_cgroup.h>
61 #include <linux/cpu.h>
62 #include <linux/oom.h>
63 #include <linux/lockdep.h>
64 #include <linux/file.h>
68 #include <net/tcp_memcontrol.h>
69 #include <linux/locallock.h>
73 #include <asm/uaccess.h>
75 #include <trace/events/vmscan.h>
77 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
78 EXPORT_SYMBOL(memory_cgrp_subsys);
80 #define MEM_CGROUP_RECLAIM_RETRIES 5
81 static struct mem_cgroup *root_mem_cgroup __read_mostly;
83 /* Whether the swap controller is active */
84 #ifdef CONFIG_MEMCG_SWAP
85 int do_swap_account __read_mostly;
87 #define do_swap_account 0
90 static DEFINE_LOCAL_IRQ_LOCK(event_lock);
91 static const char * const mem_cgroup_stat_names[] = {
100 static const char * const mem_cgroup_events_names[] = {
107 static const char * const mem_cgroup_lru_names[] = {
116 * Per memcg event counter is incremented at every pagein/pageout. With THP,
117 * it will be incremated by the number of pages. This counter is used for
118 * for trigger some periodic events. This is straightforward and better
119 * than using jiffies etc. to handle periodic memcg event.
121 enum mem_cgroup_events_target {
122 MEM_CGROUP_TARGET_THRESH,
123 MEM_CGROUP_TARGET_SOFTLIMIT,
124 MEM_CGROUP_TARGET_NUMAINFO,
127 #define THRESHOLDS_EVENTS_TARGET 128
128 #define SOFTLIMIT_EVENTS_TARGET 1024
129 #define NUMAINFO_EVENTS_TARGET 1024
131 struct mem_cgroup_stat_cpu {
132 long count[MEM_CGROUP_STAT_NSTATS];
133 unsigned long events[MEMCG_NR_EVENTS];
134 unsigned long nr_page_events;
135 unsigned long targets[MEM_CGROUP_NTARGETS];
138 struct reclaim_iter {
139 struct mem_cgroup *position;
140 /* scan generation, increased every round-trip */
141 unsigned int generation;
145 * per-zone information in memory controller.
147 struct mem_cgroup_per_zone {
148 struct lruvec lruvec;
149 unsigned long lru_size[NR_LRU_LISTS];
151 struct reclaim_iter iter[DEF_PRIORITY + 1];
153 struct rb_node tree_node; /* RB tree node */
154 unsigned long usage_in_excess;/* Set to the value by which */
155 /* the soft limit is exceeded*/
157 struct mem_cgroup *memcg; /* Back pointer, we cannot */
158 /* use container_of */
161 struct mem_cgroup_per_node {
162 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
166 * Cgroups above their limits are maintained in a RB-Tree, independent of
167 * their hierarchy representation
170 struct mem_cgroup_tree_per_zone {
171 struct rb_root rb_root;
175 struct mem_cgroup_tree_per_node {
176 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
179 struct mem_cgroup_tree {
180 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
183 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
185 struct mem_cgroup_threshold {
186 struct eventfd_ctx *eventfd;
187 unsigned long threshold;
191 struct mem_cgroup_threshold_ary {
192 /* An array index points to threshold just below or equal to usage. */
193 int current_threshold;
194 /* Size of entries[] */
196 /* Array of thresholds */
197 struct mem_cgroup_threshold entries[0];
200 struct mem_cgroup_thresholds {
201 /* Primary thresholds array */
202 struct mem_cgroup_threshold_ary *primary;
204 * Spare threshold array.
205 * This is needed to make mem_cgroup_unregister_event() "never fail".
206 * It must be able to store at least primary->size - 1 entries.
208 struct mem_cgroup_threshold_ary *spare;
212 struct mem_cgroup_eventfd_list {
213 struct list_head list;
214 struct eventfd_ctx *eventfd;
218 * cgroup_event represents events which userspace want to receive.
220 struct mem_cgroup_event {
222 * memcg which the event belongs to.
224 struct mem_cgroup *memcg;
226 * eventfd to signal userspace about the event.
228 struct eventfd_ctx *eventfd;
230 * Each of these stored in a list by the cgroup.
232 struct list_head list;
234 * register_event() callback will be used to add new userspace
235 * waiter for changes related to this event. Use eventfd_signal()
236 * on eventfd to send notification to userspace.
238 int (*register_event)(struct mem_cgroup *memcg,
239 struct eventfd_ctx *eventfd, const char *args);
241 * unregister_event() callback will be called when userspace closes
242 * the eventfd or on cgroup removing. This callback must be set,
243 * if you want provide notification functionality.
245 void (*unregister_event)(struct mem_cgroup *memcg,
246 struct eventfd_ctx *eventfd);
248 * All fields below needed to unregister event when
249 * userspace closes eventfd.
252 wait_queue_head_t *wqh;
254 struct work_struct remove;
257 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
258 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
261 * The memory controller data structure. The memory controller controls both
262 * page cache and RSS per cgroup. We would eventually like to provide
263 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
264 * to help the administrator determine what knobs to tune.
267 struct cgroup_subsys_state css;
269 /* Accounted resources */
270 struct page_counter memory;
271 struct page_counter memsw;
272 struct page_counter kmem;
274 /* Normal memory consumption range */
278 unsigned long soft_limit;
280 /* vmpressure notifications */
281 struct vmpressure vmpressure;
283 /* css_online() has been completed */
287 * Should the accounting and control be hierarchical, per subtree?
293 atomic_t oom_wakeups;
296 /* OOM-Killer disable */
297 int oom_kill_disable;
299 /* protect arrays of thresholds */
300 struct mutex thresholds_lock;
302 /* thresholds for memory usage. RCU-protected */
303 struct mem_cgroup_thresholds thresholds;
305 /* thresholds for mem+swap usage. RCU-protected */
306 struct mem_cgroup_thresholds memsw_thresholds;
308 /* For oom notifier event fd */
309 struct list_head oom_notify;
312 * Should we move charges of a task when a task is moved into this
313 * mem_cgroup ? And what type of charges should we move ?
315 unsigned long move_charge_at_immigrate;
317 * set > 0 if pages under this cgroup are moving to other cgroup.
319 atomic_t moving_account;
320 /* taken only while moving_account > 0 */
321 spinlock_t move_lock;
322 struct task_struct *move_lock_task;
323 unsigned long move_lock_flags;
327 struct mem_cgroup_stat_cpu __percpu *stat;
329 * used when a cpu is offlined or other synchronizations
330 * See mem_cgroup_read_stat().
332 struct mem_cgroup_stat_cpu nocpu_base;
333 spinlock_t pcp_counter_lock;
335 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
336 struct cg_proto tcp_mem;
338 #if defined(CONFIG_MEMCG_KMEM)
339 /* Index in the kmem_cache->memcg_params.memcg_caches array */
341 bool kmem_acct_activated;
342 bool kmem_acct_active;
345 int last_scanned_node;
347 nodemask_t scan_nodes;
348 atomic_t numainfo_events;
349 atomic_t numainfo_updating;
352 /* List of events which userspace want to receive */
353 struct list_head event_list;
354 spinlock_t event_list_lock;
356 struct mem_cgroup_per_node *nodeinfo[0];
357 /* WARNING: nodeinfo must be the last member here */
360 #ifdef CONFIG_MEMCG_KMEM
361 bool memcg_kmem_is_active(struct mem_cgroup *memcg)
363 return memcg->kmem_acct_active;
367 /* Stuffs for move charges at task migration. */
369 * Types of charges to be moved.
371 #define MOVE_ANON 0x1U
372 #define MOVE_FILE 0x2U
373 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
375 /* "mc" and its members are protected by cgroup_mutex */
376 static struct move_charge_struct {
377 spinlock_t lock; /* for from, to */
378 struct mem_cgroup *from;
379 struct mem_cgroup *to;
381 unsigned long precharge;
382 unsigned long moved_charge;
383 unsigned long moved_swap;
384 struct task_struct *moving_task; /* a task moving charges */
385 wait_queue_head_t waitq; /* a waitq for other context */
387 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
388 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
392 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
393 * limit reclaim to prevent infinite loops, if they ever occur.
395 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
396 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
399 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
400 MEM_CGROUP_CHARGE_TYPE_ANON,
401 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
402 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
406 /* for encoding cft->private value on file */
414 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
415 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
416 #define MEMFILE_ATTR(val) ((val) & 0xffff)
417 /* Used for OOM nofiier */
418 #define OOM_CONTROL (0)
421 * The memcg_create_mutex will be held whenever a new cgroup is created.
422 * As a consequence, any change that needs to protect against new child cgroups
423 * appearing has to hold it as well.
425 static DEFINE_MUTEX(memcg_create_mutex);
427 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
429 return s ? container_of(s, struct mem_cgroup, css) : NULL;
432 /* Some nice accessors for the vmpressure. */
433 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
436 memcg = root_mem_cgroup;
437 return &memcg->vmpressure;
440 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
442 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
445 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
447 return (memcg == root_mem_cgroup);
451 * We restrict the id in the range of [1, 65535], so it can fit into
454 #define MEM_CGROUP_ID_MAX USHRT_MAX
456 static inline unsigned short mem_cgroup_id(struct mem_cgroup *memcg)
458 return memcg->css.id;
462 * A helper function to get mem_cgroup from ID. must be called under
463 * rcu_read_lock(). The caller is responsible for calling
464 * css_tryget_online() if the mem_cgroup is used for charging. (dropping
465 * refcnt from swap can be called against removed memcg.)
467 static inline struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
469 struct cgroup_subsys_state *css;
471 css = css_from_id(id, &memory_cgrp_subsys);
472 return mem_cgroup_from_css(css);
475 /* Writing them here to avoid exposing memcg's inner layout */
476 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
478 void sock_update_memcg(struct sock *sk)
480 if (mem_cgroup_sockets_enabled) {
481 struct mem_cgroup *memcg;
482 struct cg_proto *cg_proto;
484 BUG_ON(!sk->sk_prot->proto_cgroup);
486 /* Socket cloning can throw us here with sk_cgrp already
487 * filled. It won't however, necessarily happen from
488 * process context. So the test for root memcg given
489 * the current task's memcg won't help us in this case.
491 * Respecting the original socket's memcg is a better
492 * decision in this case.
495 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
496 css_get(&sk->sk_cgrp->memcg->css);
501 memcg = mem_cgroup_from_task(current);
502 cg_proto = sk->sk_prot->proto_cgroup(memcg);
503 if (!mem_cgroup_is_root(memcg) &&
504 memcg_proto_active(cg_proto) &&
505 css_tryget_online(&memcg->css)) {
506 sk->sk_cgrp = cg_proto;
511 EXPORT_SYMBOL(sock_update_memcg);
513 void sock_release_memcg(struct sock *sk)
515 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
516 struct mem_cgroup *memcg;
517 WARN_ON(!sk->sk_cgrp->memcg);
518 memcg = sk->sk_cgrp->memcg;
519 css_put(&sk->sk_cgrp->memcg->css);
523 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
525 if (!memcg || mem_cgroup_is_root(memcg))
528 return &memcg->tcp_mem;
530 EXPORT_SYMBOL(tcp_proto_cgroup);
534 #ifdef CONFIG_MEMCG_KMEM
536 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
537 * The main reason for not using cgroup id for this:
538 * this works better in sparse environments, where we have a lot of memcgs,
539 * but only a few kmem-limited. Or also, if we have, for instance, 200
540 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
541 * 200 entry array for that.
543 * The current size of the caches array is stored in memcg_nr_cache_ids. It
544 * will double each time we have to increase it.
546 static DEFINE_IDA(memcg_cache_ida);
547 int memcg_nr_cache_ids;
549 /* Protects memcg_nr_cache_ids */
550 static DECLARE_RWSEM(memcg_cache_ids_sem);
552 void memcg_get_cache_ids(void)
554 down_read(&memcg_cache_ids_sem);
557 void memcg_put_cache_ids(void)
559 up_read(&memcg_cache_ids_sem);
563 * MIN_SIZE is different than 1, because we would like to avoid going through
564 * the alloc/free process all the time. In a small machine, 4 kmem-limited
565 * cgroups is a reasonable guess. In the future, it could be a parameter or
566 * tunable, but that is strictly not necessary.
568 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
569 * this constant directly from cgroup, but it is understandable that this is
570 * better kept as an internal representation in cgroup.c. In any case, the
571 * cgrp_id space is not getting any smaller, and we don't have to necessarily
572 * increase ours as well if it increases.
574 #define MEMCG_CACHES_MIN_SIZE 4
575 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
578 * A lot of the calls to the cache allocation functions are expected to be
579 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
580 * conditional to this static branch, we'll have to allow modules that does
581 * kmem_cache_alloc and the such to see this symbol as well
583 struct static_key memcg_kmem_enabled_key;
584 EXPORT_SYMBOL(memcg_kmem_enabled_key);
586 #endif /* CONFIG_MEMCG_KMEM */
588 static struct mem_cgroup_per_zone *
589 mem_cgroup_zone_zoneinfo(struct mem_cgroup *memcg, struct zone *zone)
591 int nid = zone_to_nid(zone);
592 int zid = zone_idx(zone);
594 return &memcg->nodeinfo[nid]->zoneinfo[zid];
597 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
602 static struct mem_cgroup_per_zone *
603 mem_cgroup_page_zoneinfo(struct mem_cgroup *memcg, struct page *page)
605 int nid = page_to_nid(page);
606 int zid = page_zonenum(page);
608 return &memcg->nodeinfo[nid]->zoneinfo[zid];
611 static struct mem_cgroup_tree_per_zone *
612 soft_limit_tree_node_zone(int nid, int zid)
614 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
617 static struct mem_cgroup_tree_per_zone *
618 soft_limit_tree_from_page(struct page *page)
620 int nid = page_to_nid(page);
621 int zid = page_zonenum(page);
623 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
626 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone *mz,
627 struct mem_cgroup_tree_per_zone *mctz,
628 unsigned long new_usage_in_excess)
630 struct rb_node **p = &mctz->rb_root.rb_node;
631 struct rb_node *parent = NULL;
632 struct mem_cgroup_per_zone *mz_node;
637 mz->usage_in_excess = new_usage_in_excess;
638 if (!mz->usage_in_excess)
642 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
644 if (mz->usage_in_excess < mz_node->usage_in_excess)
647 * We can't avoid mem cgroups that are over their soft
648 * limit by the same amount
650 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
653 rb_link_node(&mz->tree_node, parent, p);
654 rb_insert_color(&mz->tree_node, &mctz->rb_root);
658 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
659 struct mem_cgroup_tree_per_zone *mctz)
663 rb_erase(&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 spin_lock_irqsave(&mctz->lock, flags);
673 __mem_cgroup_remove_exceeded(mz, mctz);
674 spin_unlock_irqrestore(&mctz->lock, flags);
677 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
679 unsigned long nr_pages = page_counter_read(&memcg->memory);
680 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
681 unsigned long excess = 0;
683 if (nr_pages > soft_limit)
684 excess = nr_pages - soft_limit;
689 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
691 unsigned long excess;
692 struct mem_cgroup_per_zone *mz;
693 struct mem_cgroup_tree_per_zone *mctz;
695 mctz = soft_limit_tree_from_page(page);
697 * Necessary to update all ancestors when hierarchy is used.
698 * because their event counter is not touched.
700 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
701 mz = mem_cgroup_page_zoneinfo(memcg, page);
702 excess = soft_limit_excess(memcg);
704 * We have to update the tree if mz is on RB-tree or
705 * mem is over its softlimit.
707 if (excess || mz->on_tree) {
710 spin_lock_irqsave(&mctz->lock, flags);
711 /* if on-tree, remove it */
713 __mem_cgroup_remove_exceeded(mz, mctz);
715 * Insert again. mz->usage_in_excess will be updated.
716 * If excess is 0, no tree ops.
718 __mem_cgroup_insert_exceeded(mz, mctz, excess);
719 spin_unlock_irqrestore(&mctz->lock, flags);
724 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
726 struct mem_cgroup_tree_per_zone *mctz;
727 struct mem_cgroup_per_zone *mz;
731 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
732 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
733 mctz = soft_limit_tree_node_zone(nid, zid);
734 mem_cgroup_remove_exceeded(mz, mctz);
739 static struct mem_cgroup_per_zone *
740 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
742 struct rb_node *rightmost = NULL;
743 struct mem_cgroup_per_zone *mz;
747 rightmost = rb_last(&mctz->rb_root);
749 goto done; /* Nothing to reclaim from */
751 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
753 * Remove the node now but someone else can add it back,
754 * we will to add it back at the end of reclaim to its correct
755 * position in the tree.
757 __mem_cgroup_remove_exceeded(mz, mctz);
758 if (!soft_limit_excess(mz->memcg) ||
759 !css_tryget_online(&mz->memcg->css))
765 static struct mem_cgroup_per_zone *
766 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
768 struct mem_cgroup_per_zone *mz;
770 spin_lock_irq(&mctz->lock);
771 mz = __mem_cgroup_largest_soft_limit_node(mctz);
772 spin_unlock_irq(&mctz->lock);
777 * Implementation Note: reading percpu statistics for memcg.
779 * Both of vmstat[] and percpu_counter has threshold and do periodic
780 * synchronization to implement "quick" read. There are trade-off between
781 * reading cost and precision of value. Then, we may have a chance to implement
782 * a periodic synchronizion of counter in memcg's counter.
784 * But this _read() function is used for user interface now. The user accounts
785 * memory usage by memory cgroup and he _always_ requires exact value because
786 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
787 * have to visit all online cpus and make sum. So, for now, unnecessary
788 * synchronization is not implemented. (just implemented for cpu hotplug)
790 * If there are kernel internal actions which can make use of some not-exact
791 * value, and reading all cpu value can be performance bottleneck in some
792 * common workload, threashold and synchonization as vmstat[] should be
795 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
796 enum mem_cgroup_stat_index idx)
802 for_each_online_cpu(cpu)
803 val += per_cpu(memcg->stat->count[idx], cpu);
804 #ifdef CONFIG_HOTPLUG_CPU
805 spin_lock(&memcg->pcp_counter_lock);
806 val += memcg->nocpu_base.count[idx];
807 spin_unlock(&memcg->pcp_counter_lock);
813 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
814 enum mem_cgroup_events_index idx)
816 unsigned long val = 0;
820 for_each_online_cpu(cpu)
821 val += per_cpu(memcg->stat->events[idx], cpu);
822 #ifdef CONFIG_HOTPLUG_CPU
823 spin_lock(&memcg->pcp_counter_lock);
824 val += memcg->nocpu_base.events[idx];
825 spin_unlock(&memcg->pcp_counter_lock);
831 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
836 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
837 * counted as CACHE even if it's on ANON LRU.
840 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
843 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
846 if (PageTransHuge(page))
847 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
850 /* pagein of a big page is an event. So, ignore page size */
852 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
854 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
855 nr_pages = -nr_pages; /* for event */
858 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
861 unsigned long mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
863 struct mem_cgroup_per_zone *mz;
865 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
866 return mz->lru_size[lru];
869 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
871 unsigned int lru_mask)
873 unsigned long nr = 0;
876 VM_BUG_ON((unsigned)nid >= nr_node_ids);
878 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
879 struct mem_cgroup_per_zone *mz;
883 if (!(BIT(lru) & lru_mask))
885 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
886 nr += mz->lru_size[lru];
892 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
893 unsigned int lru_mask)
895 unsigned long nr = 0;
898 for_each_node_state(nid, N_MEMORY)
899 nr += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
903 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
904 enum mem_cgroup_events_target target)
906 unsigned long val, next;
908 val = __this_cpu_read(memcg->stat->nr_page_events);
909 next = __this_cpu_read(memcg->stat->targets[target]);
910 /* from time_after() in jiffies.h */
911 if ((long)next - (long)val < 0) {
913 case MEM_CGROUP_TARGET_THRESH:
914 next = val + THRESHOLDS_EVENTS_TARGET;
916 case MEM_CGROUP_TARGET_SOFTLIMIT:
917 next = val + SOFTLIMIT_EVENTS_TARGET;
919 case MEM_CGROUP_TARGET_NUMAINFO:
920 next = val + NUMAINFO_EVENTS_TARGET;
925 __this_cpu_write(memcg->stat->targets[target], next);
932 * Check events in order.
935 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
937 /* threshold event is triggered in finer grain than soft limit */
938 if (unlikely(mem_cgroup_event_ratelimit(memcg,
939 MEM_CGROUP_TARGET_THRESH))) {
941 bool do_numainfo __maybe_unused;
943 do_softlimit = mem_cgroup_event_ratelimit(memcg,
944 MEM_CGROUP_TARGET_SOFTLIMIT);
946 do_numainfo = mem_cgroup_event_ratelimit(memcg,
947 MEM_CGROUP_TARGET_NUMAINFO);
949 mem_cgroup_threshold(memcg);
950 if (unlikely(do_softlimit))
951 mem_cgroup_update_tree(memcg, page);
953 if (unlikely(do_numainfo))
954 atomic_inc(&memcg->numainfo_events);
959 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
962 * mm_update_next_owner() may clear mm->owner to NULL
963 * if it races with swapoff, page migration, etc.
964 * So this can be called with p == NULL.
969 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
972 static struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
974 struct mem_cgroup *memcg = NULL;
979 * Page cache insertions can happen withou an
980 * actual mm context, e.g. during disk probing
981 * on boot, loopback IO, acct() writes etc.
984 memcg = root_mem_cgroup;
986 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
987 if (unlikely(!memcg))
988 memcg = root_mem_cgroup;
990 } while (!css_tryget_online(&memcg->css));
996 * mem_cgroup_iter - iterate over memory cgroup hierarchy
997 * @root: hierarchy root
998 * @prev: previously returned memcg, NULL on first invocation
999 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1001 * Returns references to children of the hierarchy below @root, or
1002 * @root itself, or %NULL after a full round-trip.
1004 * Caller must pass the return value in @prev on subsequent
1005 * invocations for reference counting, or use mem_cgroup_iter_break()
1006 * to cancel a hierarchy walk before the round-trip is complete.
1008 * Reclaimers can specify a zone and a priority level in @reclaim to
1009 * divide up the memcgs in the hierarchy among all concurrent
1010 * reclaimers operating on the same zone and priority.
1012 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1013 struct mem_cgroup *prev,
1014 struct mem_cgroup_reclaim_cookie *reclaim)
1016 struct reclaim_iter *uninitialized_var(iter);
1017 struct cgroup_subsys_state *css = NULL;
1018 struct mem_cgroup *memcg = NULL;
1019 struct mem_cgroup *pos = NULL;
1021 if (mem_cgroup_disabled())
1025 root = root_mem_cgroup;
1027 if (prev && !reclaim)
1030 if (!root->use_hierarchy && root != root_mem_cgroup) {
1039 struct mem_cgroup_per_zone *mz;
1041 mz = mem_cgroup_zone_zoneinfo(root, reclaim->zone);
1042 iter = &mz->iter[reclaim->priority];
1044 if (prev && reclaim->generation != iter->generation)
1048 pos = READ_ONCE(iter->position);
1050 * A racing update may change the position and
1051 * put the last reference, hence css_tryget(),
1052 * or retry to see the updated position.
1054 } while (pos && !css_tryget(&pos->css));
1061 css = css_next_descendant_pre(css, &root->css);
1064 * Reclaimers share the hierarchy walk, and a
1065 * new one might jump in right at the end of
1066 * the hierarchy - make sure they see at least
1067 * one group and restart from the beginning.
1075 * Verify the css and acquire a reference. The root
1076 * is provided by the caller, so we know it's alive
1077 * and kicking, and don't take an extra reference.
1079 memcg = mem_cgroup_from_css(css);
1081 if (css == &root->css)
1084 if (css_tryget(css)) {
1086 * Make sure the memcg is initialized:
1087 * mem_cgroup_css_online() orders the the
1088 * initialization against setting the flag.
1090 if (smp_load_acquire(&memcg->initialized))
1100 if (cmpxchg(&iter->position, pos, memcg) == pos) {
1102 css_get(&memcg->css);
1108 * pairs with css_tryget when dereferencing iter->position
1117 reclaim->generation = iter->generation;
1123 if (prev && prev != root)
1124 css_put(&prev->css);
1130 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1131 * @root: hierarchy root
1132 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1134 void mem_cgroup_iter_break(struct mem_cgroup *root,
1135 struct mem_cgroup *prev)
1138 root = root_mem_cgroup;
1139 if (prev && prev != root)
1140 css_put(&prev->css);
1144 * Iteration constructs for visiting all cgroups (under a tree). If
1145 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1146 * be used for reference counting.
1148 #define for_each_mem_cgroup_tree(iter, root) \
1149 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1151 iter = mem_cgroup_iter(root, iter, NULL))
1153 #define for_each_mem_cgroup(iter) \
1154 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1156 iter = mem_cgroup_iter(NULL, iter, NULL))
1158 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1160 struct mem_cgroup *memcg;
1163 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1164 if (unlikely(!memcg))
1169 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1172 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1180 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1183 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1184 * @zone: zone of the wanted lruvec
1185 * @memcg: memcg of the wanted lruvec
1187 * Returns the lru list vector holding pages for the given @zone and
1188 * @mem. This can be the global zone lruvec, if the memory controller
1191 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1192 struct mem_cgroup *memcg)
1194 struct mem_cgroup_per_zone *mz;
1195 struct lruvec *lruvec;
1197 if (mem_cgroup_disabled()) {
1198 lruvec = &zone->lruvec;
1202 mz = mem_cgroup_zone_zoneinfo(memcg, zone);
1203 lruvec = &mz->lruvec;
1206 * Since a node can be onlined after the mem_cgroup was created,
1207 * we have to be prepared to initialize lruvec->zone here;
1208 * and if offlined then reonlined, we need to reinitialize it.
1210 if (unlikely(lruvec->zone != zone))
1211 lruvec->zone = zone;
1216 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1218 * @zone: zone of the page
1220 * This function is only safe when following the LRU page isolation
1221 * and putback protocol: the LRU lock must be held, and the page must
1222 * either be PageLRU() or the caller must have isolated/allocated it.
1224 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1226 struct mem_cgroup_per_zone *mz;
1227 struct mem_cgroup *memcg;
1228 struct lruvec *lruvec;
1230 if (mem_cgroup_disabled()) {
1231 lruvec = &zone->lruvec;
1235 memcg = page->mem_cgroup;
1237 * Swapcache readahead pages are added to the LRU - and
1238 * possibly migrated - before they are charged.
1241 memcg = root_mem_cgroup;
1243 mz = mem_cgroup_page_zoneinfo(memcg, page);
1244 lruvec = &mz->lruvec;
1247 * Since a node can be onlined after the mem_cgroup was created,
1248 * we have to be prepared to initialize lruvec->zone here;
1249 * and if offlined then reonlined, we need to reinitialize it.
1251 if (unlikely(lruvec->zone != zone))
1252 lruvec->zone = zone;
1257 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1258 * @lruvec: mem_cgroup per zone lru vector
1259 * @lru: index of lru list the page is sitting on
1260 * @nr_pages: positive when adding or negative when removing
1262 * This function must be called when a page is added to or removed from an
1265 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1268 struct mem_cgroup_per_zone *mz;
1269 unsigned long *lru_size;
1271 if (mem_cgroup_disabled())
1274 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1275 lru_size = mz->lru_size + lru;
1276 *lru_size += nr_pages;
1277 VM_BUG_ON((long)(*lru_size) < 0);
1280 bool mem_cgroup_is_descendant(struct mem_cgroup *memcg, struct mem_cgroup *root)
1284 if (!root->use_hierarchy)
1286 return cgroup_is_descendant(memcg->css.cgroup, root->css.cgroup);
1289 bool task_in_mem_cgroup(struct task_struct *task, struct mem_cgroup *memcg)
1291 struct mem_cgroup *task_memcg;
1292 struct task_struct *p;
1295 p = find_lock_task_mm(task);
1297 task_memcg = get_mem_cgroup_from_mm(p->mm);
1301 * All threads may have already detached their mm's, but the oom
1302 * killer still needs to detect if they have already been oom
1303 * killed to prevent needlessly killing additional tasks.
1306 task_memcg = mem_cgroup_from_task(task);
1307 css_get(&task_memcg->css);
1310 ret = mem_cgroup_is_descendant(task_memcg, memcg);
1311 css_put(&task_memcg->css);
1315 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1317 unsigned long inactive_ratio;
1318 unsigned long inactive;
1319 unsigned long active;
1322 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1323 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1325 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1327 inactive_ratio = int_sqrt(10 * gb);
1331 return inactive * inactive_ratio < active;
1334 bool mem_cgroup_lruvec_online(struct lruvec *lruvec)
1336 struct mem_cgroup_per_zone *mz;
1337 struct mem_cgroup *memcg;
1339 if (mem_cgroup_disabled())
1342 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1345 return !!(memcg->css.flags & CSS_ONLINE);
1348 #define mem_cgroup_from_counter(counter, member) \
1349 container_of(counter, struct mem_cgroup, member)
1352 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1353 * @memcg: the memory cgroup
1355 * Returns the maximum amount of memory @mem can be charged with, in
1358 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1360 unsigned long margin = 0;
1361 unsigned long count;
1362 unsigned long limit;
1364 count = page_counter_read(&memcg->memory);
1365 limit = READ_ONCE(memcg->memory.limit);
1367 margin = limit - count;
1369 if (do_swap_account) {
1370 count = page_counter_read(&memcg->memsw);
1371 limit = READ_ONCE(memcg->memsw.limit);
1373 margin = min(margin, limit - count);
1379 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1382 if (mem_cgroup_disabled() || !memcg->css.parent)
1383 return vm_swappiness;
1385 return memcg->swappiness;
1389 * A routine for checking "mem" is under move_account() or not.
1391 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1392 * moving cgroups. This is for waiting at high-memory pressure
1395 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1397 struct mem_cgroup *from;
1398 struct mem_cgroup *to;
1401 * Unlike task_move routines, we access mc.to, mc.from not under
1402 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1404 spin_lock(&mc.lock);
1410 ret = mem_cgroup_is_descendant(from, memcg) ||
1411 mem_cgroup_is_descendant(to, memcg);
1413 spin_unlock(&mc.lock);
1417 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1419 if (mc.moving_task && current != mc.moving_task) {
1420 if (mem_cgroup_under_move(memcg)) {
1422 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1423 /* moving charge context might have finished. */
1426 finish_wait(&mc.waitq, &wait);
1433 #define K(x) ((x) << (PAGE_SHIFT-10))
1435 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1436 * @memcg: The memory cgroup that went over limit
1437 * @p: Task that is going to be killed
1439 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1442 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1444 /* oom_info_lock ensures that parallel ooms do not interleave */
1445 static DEFINE_MUTEX(oom_info_lock);
1446 struct mem_cgroup *iter;
1449 mutex_lock(&oom_info_lock);
1453 pr_info("Task in ");
1454 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1455 pr_cont(" killed as a result of limit of ");
1457 pr_info("Memory limit reached of cgroup ");
1460 pr_cont_cgroup_path(memcg->css.cgroup);
1465 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1466 K((u64)page_counter_read(&memcg->memory)),
1467 K((u64)memcg->memory.limit), memcg->memory.failcnt);
1468 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1469 K((u64)page_counter_read(&memcg->memsw)),
1470 K((u64)memcg->memsw.limit), memcg->memsw.failcnt);
1471 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1472 K((u64)page_counter_read(&memcg->kmem)),
1473 K((u64)memcg->kmem.limit), memcg->kmem.failcnt);
1475 for_each_mem_cgroup_tree(iter, memcg) {
1476 pr_info("Memory cgroup stats for ");
1477 pr_cont_cgroup_path(iter->css.cgroup);
1480 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1481 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1483 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1484 K(mem_cgroup_read_stat(iter, i)));
1487 for (i = 0; i < NR_LRU_LISTS; i++)
1488 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1489 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1493 mutex_unlock(&oom_info_lock);
1497 * This function returns the number of memcg under hierarchy tree. Returns
1498 * 1(self count) if no children.
1500 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1503 struct mem_cgroup *iter;
1505 for_each_mem_cgroup_tree(iter, memcg)
1511 * Return the memory (and swap, if configured) limit for a memcg.
1513 static unsigned long mem_cgroup_get_limit(struct mem_cgroup *memcg)
1515 unsigned long limit;
1517 limit = memcg->memory.limit;
1518 if (mem_cgroup_swappiness(memcg)) {
1519 unsigned long memsw_limit;
1521 memsw_limit = memcg->memsw.limit;
1522 limit = min(limit + total_swap_pages, memsw_limit);
1527 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1530 struct mem_cgroup *iter;
1531 unsigned long chosen_points = 0;
1532 unsigned long totalpages;
1533 unsigned int points = 0;
1534 struct task_struct *chosen = NULL;
1537 * If current has a pending SIGKILL or is exiting, then automatically
1538 * select it. The goal is to allow it to allocate so that it may
1539 * quickly exit and free its memory.
1541 if (fatal_signal_pending(current) || task_will_free_mem(current)) {
1542 mark_tsk_oom_victim(current);
1546 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL, memcg);
1547 totalpages = mem_cgroup_get_limit(memcg) ? : 1;
1548 for_each_mem_cgroup_tree(iter, memcg) {
1549 struct css_task_iter it;
1550 struct task_struct *task;
1552 css_task_iter_start(&iter->css, &it);
1553 while ((task = css_task_iter_next(&it))) {
1554 switch (oom_scan_process_thread(task, totalpages, NULL,
1556 case OOM_SCAN_SELECT:
1558 put_task_struct(chosen);
1560 chosen_points = ULONG_MAX;
1561 get_task_struct(chosen);
1563 case OOM_SCAN_CONTINUE:
1565 case OOM_SCAN_ABORT:
1566 css_task_iter_end(&it);
1567 mem_cgroup_iter_break(memcg, iter);
1569 put_task_struct(chosen);
1574 points = oom_badness(task, memcg, NULL, totalpages);
1575 if (!points || points < chosen_points)
1577 /* Prefer thread group leaders for display purposes */
1578 if (points == chosen_points &&
1579 thread_group_leader(chosen))
1583 put_task_struct(chosen);
1585 chosen_points = points;
1586 get_task_struct(chosen);
1588 css_task_iter_end(&it);
1593 points = chosen_points * 1000 / totalpages;
1594 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1595 NULL, "Memory cgroup out of memory");
1598 #if MAX_NUMNODES > 1
1601 * test_mem_cgroup_node_reclaimable
1602 * @memcg: the target memcg
1603 * @nid: the node ID to be checked.
1604 * @noswap : specify true here if the user wants flle only information.
1606 * This function returns whether the specified memcg contains any
1607 * reclaimable pages on a node. Returns true if there are any reclaimable
1608 * pages in the node.
1610 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1611 int nid, bool noswap)
1613 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1615 if (noswap || !total_swap_pages)
1617 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1624 * Always updating the nodemask is not very good - even if we have an empty
1625 * list or the wrong list here, we can start from some node and traverse all
1626 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1629 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1633 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1634 * pagein/pageout changes since the last update.
1636 if (!atomic_read(&memcg->numainfo_events))
1638 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1641 /* make a nodemask where this memcg uses memory from */
1642 memcg->scan_nodes = node_states[N_MEMORY];
1644 for_each_node_mask(nid, node_states[N_MEMORY]) {
1646 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1647 node_clear(nid, memcg->scan_nodes);
1650 atomic_set(&memcg->numainfo_events, 0);
1651 atomic_set(&memcg->numainfo_updating, 0);
1655 * Selecting a node where we start reclaim from. Because what we need is just
1656 * reducing usage counter, start from anywhere is O,K. Considering
1657 * memory reclaim from current node, there are pros. and cons.
1659 * Freeing memory from current node means freeing memory from a node which
1660 * we'll use or we've used. So, it may make LRU bad. And if several threads
1661 * hit limits, it will see a contention on a node. But freeing from remote
1662 * node means more costs for memory reclaim because of memory latency.
1664 * Now, we use round-robin. Better algorithm is welcomed.
1666 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1670 mem_cgroup_may_update_nodemask(memcg);
1671 node = memcg->last_scanned_node;
1673 node = next_node(node, memcg->scan_nodes);
1674 if (node == MAX_NUMNODES)
1675 node = first_node(memcg->scan_nodes);
1677 * We call this when we hit limit, not when pages are added to LRU.
1678 * No LRU may hold pages because all pages are UNEVICTABLE or
1679 * memcg is too small and all pages are not on LRU. In that case,
1680 * we use curret node.
1682 if (unlikely(node == MAX_NUMNODES))
1683 node = numa_node_id();
1685 memcg->last_scanned_node = node;
1689 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1695 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1698 unsigned long *total_scanned)
1700 struct mem_cgroup *victim = NULL;
1703 unsigned long excess;
1704 unsigned long nr_scanned;
1705 struct mem_cgroup_reclaim_cookie reclaim = {
1710 excess = soft_limit_excess(root_memcg);
1713 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1718 * If we have not been able to reclaim
1719 * anything, it might because there are
1720 * no reclaimable pages under this hierarchy
1725 * We want to do more targeted reclaim.
1726 * excess >> 2 is not to excessive so as to
1727 * reclaim too much, nor too less that we keep
1728 * coming back to reclaim from this cgroup
1730 if (total >= (excess >> 2) ||
1731 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1736 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1738 *total_scanned += nr_scanned;
1739 if (!soft_limit_excess(root_memcg))
1742 mem_cgroup_iter_break(root_memcg, victim);
1746 #ifdef CONFIG_LOCKDEP
1747 static struct lockdep_map memcg_oom_lock_dep_map = {
1748 .name = "memcg_oom_lock",
1752 static DEFINE_SPINLOCK(memcg_oom_lock);
1755 * Check OOM-Killer is already running under our hierarchy.
1756 * If someone is running, return false.
1758 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1760 struct mem_cgroup *iter, *failed = NULL;
1762 spin_lock(&memcg_oom_lock);
1764 for_each_mem_cgroup_tree(iter, memcg) {
1765 if (iter->oom_lock) {
1767 * this subtree of our hierarchy is already locked
1768 * so we cannot give a lock.
1771 mem_cgroup_iter_break(memcg, iter);
1774 iter->oom_lock = true;
1779 * OK, we failed to lock the whole subtree so we have
1780 * to clean up what we set up to the failing subtree
1782 for_each_mem_cgroup_tree(iter, memcg) {
1783 if (iter == failed) {
1784 mem_cgroup_iter_break(memcg, iter);
1787 iter->oom_lock = false;
1790 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1792 spin_unlock(&memcg_oom_lock);
1797 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1799 struct mem_cgroup *iter;
1801 spin_lock(&memcg_oom_lock);
1802 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1803 for_each_mem_cgroup_tree(iter, memcg)
1804 iter->oom_lock = false;
1805 spin_unlock(&memcg_oom_lock);
1808 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1810 struct mem_cgroup *iter;
1812 for_each_mem_cgroup_tree(iter, memcg)
1813 atomic_inc(&iter->under_oom);
1816 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1818 struct mem_cgroup *iter;
1821 * When a new child is created while the hierarchy is under oom,
1822 * mem_cgroup_oom_lock() may not be called. We have to use
1823 * atomic_add_unless() here.
1825 for_each_mem_cgroup_tree(iter, memcg)
1826 atomic_add_unless(&iter->under_oom, -1, 0);
1829 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1831 struct oom_wait_info {
1832 struct mem_cgroup *memcg;
1836 static int memcg_oom_wake_function(wait_queue_t *wait,
1837 unsigned mode, int sync, void *arg)
1839 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1840 struct mem_cgroup *oom_wait_memcg;
1841 struct oom_wait_info *oom_wait_info;
1843 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1844 oom_wait_memcg = oom_wait_info->memcg;
1846 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1847 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1849 return autoremove_wake_function(wait, mode, sync, arg);
1852 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
1854 atomic_inc(&memcg->oom_wakeups);
1855 /* for filtering, pass "memcg" as argument. */
1856 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1859 static void memcg_oom_recover(struct mem_cgroup *memcg)
1861 if (memcg && atomic_read(&memcg->under_oom))
1862 memcg_wakeup_oom(memcg);
1865 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1867 if (!current->memcg_oom.may_oom)
1870 * We are in the middle of the charge context here, so we
1871 * don't want to block when potentially sitting on a callstack
1872 * that holds all kinds of filesystem and mm locks.
1874 * Also, the caller may handle a failed allocation gracefully
1875 * (like optional page cache readahead) and so an OOM killer
1876 * invocation might not even be necessary.
1878 * That's why we don't do anything here except remember the
1879 * OOM context and then deal with it at the end of the page
1880 * fault when the stack is unwound, the locks are released,
1881 * and when we know whether the fault was overall successful.
1883 css_get(&memcg->css);
1884 current->memcg_oom.memcg = memcg;
1885 current->memcg_oom.gfp_mask = mask;
1886 current->memcg_oom.order = order;
1890 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1891 * @handle: actually kill/wait or just clean up the OOM state
1893 * This has to be called at the end of a page fault if the memcg OOM
1894 * handler was enabled.
1896 * Memcg supports userspace OOM handling where failed allocations must
1897 * sleep on a waitqueue until the userspace task resolves the
1898 * situation. Sleeping directly in the charge context with all kinds
1899 * of locks held is not a good idea, instead we remember an OOM state
1900 * in the task and mem_cgroup_oom_synchronize() has to be called at
1901 * the end of the page fault to complete the OOM handling.
1903 * Returns %true if an ongoing memcg OOM situation was detected and
1904 * completed, %false otherwise.
1906 bool mem_cgroup_oom_synchronize(bool handle)
1908 struct mem_cgroup *memcg = current->memcg_oom.memcg;
1909 struct oom_wait_info owait;
1912 /* OOM is global, do not handle */
1916 if (!handle || oom_killer_disabled)
1919 owait.memcg = memcg;
1920 owait.wait.flags = 0;
1921 owait.wait.func = memcg_oom_wake_function;
1922 owait.wait.private = current;
1923 INIT_LIST_HEAD(&owait.wait.task_list);
1925 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1926 mem_cgroup_mark_under_oom(memcg);
1928 locked = mem_cgroup_oom_trylock(memcg);
1931 mem_cgroup_oom_notify(memcg);
1933 if (locked && !memcg->oom_kill_disable) {
1934 mem_cgroup_unmark_under_oom(memcg);
1935 finish_wait(&memcg_oom_waitq, &owait.wait);
1936 mem_cgroup_out_of_memory(memcg, current->memcg_oom.gfp_mask,
1937 current->memcg_oom.order);
1940 mem_cgroup_unmark_under_oom(memcg);
1941 finish_wait(&memcg_oom_waitq, &owait.wait);
1945 mem_cgroup_oom_unlock(memcg);
1947 * There is no guarantee that an OOM-lock contender
1948 * sees the wakeups triggered by the OOM kill
1949 * uncharges. Wake any sleepers explicitely.
1951 memcg_oom_recover(memcg);
1954 current->memcg_oom.memcg = NULL;
1955 css_put(&memcg->css);
1960 * mem_cgroup_begin_page_stat - begin a page state statistics transaction
1961 * @page: page that is going to change accounted state
1963 * This function must mark the beginning of an accounted page state
1964 * change to prevent double accounting when the page is concurrently
1965 * being moved to another memcg:
1967 * memcg = mem_cgroup_begin_page_stat(page);
1968 * if (TestClearPageState(page))
1969 * mem_cgroup_update_page_stat(memcg, state, -1);
1970 * mem_cgroup_end_page_stat(memcg);
1972 struct mem_cgroup *mem_cgroup_begin_page_stat(struct page *page)
1974 struct mem_cgroup *memcg;
1975 unsigned long flags;
1978 * The RCU lock is held throughout the transaction. The fast
1979 * path can get away without acquiring the memcg->move_lock
1980 * because page moving starts with an RCU grace period.
1982 * The RCU lock also protects the memcg from being freed when
1983 * the page state that is going to change is the only thing
1984 * preventing the page from being uncharged.
1985 * E.g. end-writeback clearing PageWriteback(), which allows
1986 * migration to go ahead and uncharge the page before the
1987 * account transaction might be complete.
1991 if (mem_cgroup_disabled())
1994 memcg = page->mem_cgroup;
1995 if (unlikely(!memcg))
1998 if (atomic_read(&memcg->moving_account) <= 0)
2001 spin_lock_irqsave(&memcg->move_lock, flags);
2002 if (memcg != page->mem_cgroup) {
2003 spin_unlock_irqrestore(&memcg->move_lock, flags);
2008 * When charge migration first begins, we can have locked and
2009 * unlocked page stat updates happening concurrently. Track
2010 * the task who has the lock for mem_cgroup_end_page_stat().
2012 memcg->move_lock_task = current;
2013 memcg->move_lock_flags = flags;
2019 * mem_cgroup_end_page_stat - finish a page state statistics transaction
2020 * @memcg: the memcg that was accounted against
2022 void mem_cgroup_end_page_stat(struct mem_cgroup *memcg)
2024 if (memcg && memcg->move_lock_task == current) {
2025 unsigned long flags = memcg->move_lock_flags;
2027 memcg->move_lock_task = NULL;
2028 memcg->move_lock_flags = 0;
2030 spin_unlock_irqrestore(&memcg->move_lock, flags);
2037 * mem_cgroup_update_page_stat - update page state statistics
2038 * @memcg: memcg to account against
2039 * @idx: page state item to account
2040 * @val: number of pages (positive or negative)
2042 * See mem_cgroup_begin_page_stat() for locking requirements.
2044 void mem_cgroup_update_page_stat(struct mem_cgroup *memcg,
2045 enum mem_cgroup_stat_index idx, int val)
2047 VM_BUG_ON(!rcu_read_lock_held());
2050 this_cpu_add(memcg->stat->count[idx], val);
2054 * size of first charge trial. "32" comes from vmscan.c's magic value.
2055 * TODO: maybe necessary to use big numbers in big irons.
2057 #define CHARGE_BATCH 32U
2058 struct memcg_stock_pcp {
2059 struct mem_cgroup *cached; /* this never be root cgroup */
2060 unsigned int nr_pages;
2061 struct work_struct work;
2062 unsigned long flags;
2063 #define FLUSHING_CACHED_CHARGE 0
2065 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2066 static DEFINE_MUTEX(percpu_charge_mutex);
2069 * consume_stock: Try to consume stocked charge on this cpu.
2070 * @memcg: memcg to consume from.
2071 * @nr_pages: how many pages to charge.
2073 * The charges will only happen if @memcg matches the current cpu's memcg
2074 * stock, and at least @nr_pages are available in that stock. Failure to
2075 * service an allocation will refill the stock.
2077 * returns true if successful, false otherwise.
2079 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2081 struct memcg_stock_pcp *stock;
2084 if (nr_pages > CHARGE_BATCH)
2087 stock = &get_cpu_var(memcg_stock);
2088 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2089 stock->nr_pages -= nr_pages;
2092 put_cpu_var(memcg_stock);
2097 * Returns stocks cached in percpu and reset cached information.
2099 static void drain_stock(struct memcg_stock_pcp *stock)
2101 struct mem_cgroup *old = stock->cached;
2103 if (stock->nr_pages) {
2104 page_counter_uncharge(&old->memory, stock->nr_pages);
2105 if (do_swap_account)
2106 page_counter_uncharge(&old->memsw, stock->nr_pages);
2107 css_put_many(&old->css, stock->nr_pages);
2108 stock->nr_pages = 0;
2110 stock->cached = NULL;
2114 * This must be called under preempt disabled or must be called by
2115 * a thread which is pinned to local cpu.
2117 static void drain_local_stock(struct work_struct *dummy)
2119 struct memcg_stock_pcp *stock = this_cpu_ptr(&memcg_stock);
2121 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2125 * Cache charges(val) to local per_cpu area.
2126 * This will be consumed by consume_stock() function, later.
2128 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2130 struct memcg_stock_pcp *stock;
2131 int cpu = get_cpu_light();
2133 stock = &per_cpu(memcg_stock, cpu);
2135 if (stock->cached != memcg) { /* reset if necessary */
2137 stock->cached = memcg;
2139 stock->nr_pages += nr_pages;
2144 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2145 * of the hierarchy under it.
2147 static void drain_all_stock(struct mem_cgroup *root_memcg)
2151 /* If someone's already draining, avoid adding running more workers. */
2152 if (!mutex_trylock(&percpu_charge_mutex))
2154 /* Notify other cpus that system-wide "drain" is running */
2156 curcpu = get_cpu_light();
2157 for_each_online_cpu(cpu) {
2158 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2159 struct mem_cgroup *memcg;
2161 memcg = stock->cached;
2162 if (!memcg || !stock->nr_pages)
2164 if (!mem_cgroup_is_descendant(memcg, root_memcg))
2166 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2168 drain_local_stock(&stock->work);
2170 schedule_work_on(cpu, &stock->work);
2175 mutex_unlock(&percpu_charge_mutex);
2179 * This function drains percpu counter value from DEAD cpu and
2180 * move it to local cpu. Note that this function can be preempted.
2182 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2186 spin_lock(&memcg->pcp_counter_lock);
2187 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2188 long x = per_cpu(memcg->stat->count[i], cpu);
2190 per_cpu(memcg->stat->count[i], cpu) = 0;
2191 memcg->nocpu_base.count[i] += x;
2193 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2194 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2196 per_cpu(memcg->stat->events[i], cpu) = 0;
2197 memcg->nocpu_base.events[i] += x;
2199 spin_unlock(&memcg->pcp_counter_lock);
2202 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
2203 unsigned long action,
2206 int cpu = (unsigned long)hcpu;
2207 struct memcg_stock_pcp *stock;
2208 struct mem_cgroup *iter;
2210 if (action == CPU_ONLINE)
2213 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2216 for_each_mem_cgroup(iter)
2217 mem_cgroup_drain_pcp_counter(iter, cpu);
2219 stock = &per_cpu(memcg_stock, cpu);
2224 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2225 unsigned int nr_pages)
2227 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2228 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2229 struct mem_cgroup *mem_over_limit;
2230 struct page_counter *counter;
2231 unsigned long nr_reclaimed;
2232 bool may_swap = true;
2233 bool drained = false;
2236 if (mem_cgroup_is_root(memcg))
2239 if (consume_stock(memcg, nr_pages))
2242 if (!do_swap_account ||
2243 !page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2244 if (!page_counter_try_charge(&memcg->memory, batch, &counter))
2246 if (do_swap_account)
2247 page_counter_uncharge(&memcg->memsw, batch);
2248 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2250 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2254 if (batch > nr_pages) {
2260 * Unlike in global OOM situations, memcg is not in a physical
2261 * memory shortage. Allow dying and OOM-killed tasks to
2262 * bypass the last charges so that they can exit quickly and
2263 * free their memory.
2265 if (unlikely(test_thread_flag(TIF_MEMDIE) ||
2266 fatal_signal_pending(current) ||
2267 current->flags & PF_EXITING))
2270 if (unlikely(task_in_memcg_oom(current)))
2273 if (!(gfp_mask & __GFP_WAIT))
2276 mem_cgroup_events(mem_over_limit, MEMCG_MAX, 1);
2278 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2279 gfp_mask, may_swap);
2281 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2285 drain_all_stock(mem_over_limit);
2290 if (gfp_mask & __GFP_NORETRY)
2293 * Even though the limit is exceeded at this point, reclaim
2294 * may have been able to free some pages. Retry the charge
2295 * before killing the task.
2297 * Only for regular pages, though: huge pages are rather
2298 * unlikely to succeed so close to the limit, and we fall back
2299 * to regular pages anyway in case of failure.
2301 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2304 * At task move, charge accounts can be doubly counted. So, it's
2305 * better to wait until the end of task_move if something is going on.
2307 if (mem_cgroup_wait_acct_move(mem_over_limit))
2313 if (gfp_mask & __GFP_NOFAIL)
2316 if (fatal_signal_pending(current))
2319 mem_cgroup_events(mem_over_limit, MEMCG_OOM, 1);
2321 mem_cgroup_oom(mem_over_limit, gfp_mask, get_order(nr_pages));
2323 if (!(gfp_mask & __GFP_NOFAIL))
2329 css_get_many(&memcg->css, batch);
2330 if (batch > nr_pages)
2331 refill_stock(memcg, batch - nr_pages);
2332 if (!(gfp_mask & __GFP_WAIT))
2335 * If the hierarchy is above the normal consumption range,
2336 * make the charging task trim their excess contribution.
2339 if (page_counter_read(&memcg->memory) <= memcg->high)
2341 mem_cgroup_events(memcg, MEMCG_HIGH, 1);
2342 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
2343 } while ((memcg = parent_mem_cgroup(memcg)));
2348 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2350 if (mem_cgroup_is_root(memcg))
2353 page_counter_uncharge(&memcg->memory, nr_pages);
2354 if (do_swap_account)
2355 page_counter_uncharge(&memcg->memsw, nr_pages);
2357 css_put_many(&memcg->css, nr_pages);
2361 * try_get_mem_cgroup_from_page - look up page's memcg association
2364 * Look up, get a css reference, and return the memcg that owns @page.
2366 * The page must be locked to prevent racing with swap-in and page
2367 * cache charges. If coming from an unlocked page table, the caller
2368 * must ensure the page is on the LRU or this can race with charging.
2370 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2372 struct mem_cgroup *memcg;
2376 VM_BUG_ON_PAGE(!PageLocked(page), page);
2378 memcg = page->mem_cgroup;
2380 if (!css_tryget_online(&memcg->css))
2382 } else if (PageSwapCache(page)) {
2383 ent.val = page_private(page);
2384 id = lookup_swap_cgroup_id(ent);
2386 memcg = mem_cgroup_from_id(id);
2387 if (memcg && !css_tryget_online(&memcg->css))
2394 static void lock_page_lru(struct page *page, int *isolated)
2396 struct zone *zone = page_zone(page);
2398 spin_lock_irq(&zone->lru_lock);
2399 if (PageLRU(page)) {
2400 struct lruvec *lruvec;
2402 lruvec = mem_cgroup_page_lruvec(page, zone);
2404 del_page_from_lru_list(page, lruvec, page_lru(page));
2410 static void unlock_page_lru(struct page *page, int isolated)
2412 struct zone *zone = page_zone(page);
2415 struct lruvec *lruvec;
2417 lruvec = mem_cgroup_page_lruvec(page, zone);
2418 VM_BUG_ON_PAGE(PageLRU(page), page);
2420 add_page_to_lru_list(page, lruvec, page_lru(page));
2422 spin_unlock_irq(&zone->lru_lock);
2425 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2430 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2433 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2434 * may already be on some other mem_cgroup's LRU. Take care of it.
2437 lock_page_lru(page, &isolated);
2440 * Nobody should be changing or seriously looking at
2441 * page->mem_cgroup at this point:
2443 * - the page is uncharged
2445 * - the page is off-LRU
2447 * - an anonymous fault has exclusive page access, except for
2448 * a locked page table
2450 * - a page cache insertion, a swapin fault, or a migration
2451 * have the page locked
2453 page->mem_cgroup = memcg;
2456 unlock_page_lru(page, isolated);
2459 #ifdef CONFIG_MEMCG_KMEM
2460 int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp,
2461 unsigned long nr_pages)
2463 struct page_counter *counter;
2466 ret = page_counter_try_charge(&memcg->kmem, nr_pages, &counter);
2470 ret = try_charge(memcg, gfp, nr_pages);
2471 if (ret == -EINTR) {
2473 * try_charge() chose to bypass to root due to OOM kill or
2474 * fatal signal. Since our only options are to either fail
2475 * the allocation or charge it to this cgroup, do it as a
2476 * temporary condition. But we can't fail. From a kmem/slab
2477 * perspective, the cache has already been selected, by
2478 * mem_cgroup_kmem_get_cache(), so it is too late to change
2481 * This condition will only trigger if the task entered
2482 * memcg_charge_kmem in a sane state, but was OOM-killed
2483 * during try_charge() above. Tasks that were already dying
2484 * when the allocation triggers should have been already
2485 * directed to the root cgroup in memcontrol.h
2487 page_counter_charge(&memcg->memory, nr_pages);
2488 if (do_swap_account)
2489 page_counter_charge(&memcg->memsw, nr_pages);
2490 css_get_many(&memcg->css, nr_pages);
2493 page_counter_uncharge(&memcg->kmem, nr_pages);
2498 void memcg_uncharge_kmem(struct mem_cgroup *memcg, unsigned long nr_pages)
2500 page_counter_uncharge(&memcg->memory, nr_pages);
2501 if (do_swap_account)
2502 page_counter_uncharge(&memcg->memsw, nr_pages);
2504 page_counter_uncharge(&memcg->kmem, nr_pages);
2506 css_put_many(&memcg->css, nr_pages);
2510 * helper for acessing a memcg's index. It will be used as an index in the
2511 * child cache array in kmem_cache, and also to derive its name. This function
2512 * will return -1 when this is not a kmem-limited memcg.
2514 int memcg_cache_id(struct mem_cgroup *memcg)
2516 return memcg ? memcg->kmemcg_id : -1;
2519 static int memcg_alloc_cache_id(void)
2524 id = ida_simple_get(&memcg_cache_ida,
2525 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2529 if (id < memcg_nr_cache_ids)
2533 * There's no space for the new id in memcg_caches arrays,
2534 * so we have to grow them.
2536 down_write(&memcg_cache_ids_sem);
2538 size = 2 * (id + 1);
2539 if (size < MEMCG_CACHES_MIN_SIZE)
2540 size = MEMCG_CACHES_MIN_SIZE;
2541 else if (size > MEMCG_CACHES_MAX_SIZE)
2542 size = MEMCG_CACHES_MAX_SIZE;
2544 err = memcg_update_all_caches(size);
2546 err = memcg_update_all_list_lrus(size);
2548 memcg_nr_cache_ids = size;
2550 up_write(&memcg_cache_ids_sem);
2553 ida_simple_remove(&memcg_cache_ida, id);
2559 static void memcg_free_cache_id(int id)
2561 ida_simple_remove(&memcg_cache_ida, id);
2564 struct memcg_kmem_cache_create_work {
2565 struct mem_cgroup *memcg;
2566 struct kmem_cache *cachep;
2567 struct work_struct work;
2570 static void memcg_kmem_cache_create_func(struct work_struct *w)
2572 struct memcg_kmem_cache_create_work *cw =
2573 container_of(w, struct memcg_kmem_cache_create_work, work);
2574 struct mem_cgroup *memcg = cw->memcg;
2575 struct kmem_cache *cachep = cw->cachep;
2577 memcg_create_kmem_cache(memcg, cachep);
2579 css_put(&memcg->css);
2584 * Enqueue the creation of a per-memcg kmem_cache.
2586 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2587 struct kmem_cache *cachep)
2589 struct memcg_kmem_cache_create_work *cw;
2591 cw = kmalloc(sizeof(*cw), GFP_NOWAIT);
2595 css_get(&memcg->css);
2598 cw->cachep = cachep;
2599 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2601 schedule_work(&cw->work);
2604 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2605 struct kmem_cache *cachep)
2608 * We need to stop accounting when we kmalloc, because if the
2609 * corresponding kmalloc cache is not yet created, the first allocation
2610 * in __memcg_schedule_kmem_cache_create will recurse.
2612 * However, it is better to enclose the whole function. Depending on
2613 * the debugging options enabled, INIT_WORK(), for instance, can
2614 * trigger an allocation. This too, will make us recurse. Because at
2615 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2616 * the safest choice is to do it like this, wrapping the whole function.
2618 current->memcg_kmem_skip_account = 1;
2619 __memcg_schedule_kmem_cache_create(memcg, cachep);
2620 current->memcg_kmem_skip_account = 0;
2624 * Return the kmem_cache we're supposed to use for a slab allocation.
2625 * We try to use the current memcg's version of the cache.
2627 * If the cache does not exist yet, if we are the first user of it,
2628 * we either create it immediately, if possible, or create it asynchronously
2630 * In the latter case, we will let the current allocation go through with
2631 * the original cache.
2633 * Can't be called in interrupt context or from kernel threads.
2634 * This function needs to be called with rcu_read_lock() held.
2636 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep)
2638 struct mem_cgroup *memcg;
2639 struct kmem_cache *memcg_cachep;
2642 VM_BUG_ON(!is_root_cache(cachep));
2644 if (current->memcg_kmem_skip_account)
2647 memcg = get_mem_cgroup_from_mm(current->mm);
2648 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2652 memcg_cachep = cache_from_memcg_idx(cachep, kmemcg_id);
2653 if (likely(memcg_cachep))
2654 return memcg_cachep;
2657 * If we are in a safe context (can wait, and not in interrupt
2658 * context), we could be be predictable and return right away.
2659 * This would guarantee that the allocation being performed
2660 * already belongs in the new cache.
2662 * However, there are some clashes that can arrive from locking.
2663 * For instance, because we acquire the slab_mutex while doing
2664 * memcg_create_kmem_cache, this means no further allocation
2665 * could happen with the slab_mutex held. So it's better to
2668 memcg_schedule_kmem_cache_create(memcg, cachep);
2670 css_put(&memcg->css);
2674 void __memcg_kmem_put_cache(struct kmem_cache *cachep)
2676 if (!is_root_cache(cachep))
2677 css_put(&cachep->memcg_params.memcg->css);
2681 * We need to verify if the allocation against current->mm->owner's memcg is
2682 * possible for the given order. But the page is not allocated yet, so we'll
2683 * need a further commit step to do the final arrangements.
2685 * It is possible for the task to switch cgroups in this mean time, so at
2686 * commit time, we can't rely on task conversion any longer. We'll then use
2687 * the handle argument to return to the caller which cgroup we should commit
2688 * against. We could also return the memcg directly and avoid the pointer
2689 * passing, but a boolean return value gives better semantics considering
2690 * the compiled-out case as well.
2692 * Returning true means the allocation is possible.
2695 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
2697 struct mem_cgroup *memcg;
2702 memcg = get_mem_cgroup_from_mm(current->mm);
2704 if (!memcg_kmem_is_active(memcg)) {
2705 css_put(&memcg->css);
2709 ret = memcg_charge_kmem(memcg, gfp, 1 << order);
2713 css_put(&memcg->css);
2717 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
2720 VM_BUG_ON(mem_cgroup_is_root(memcg));
2722 /* The page allocation failed. Revert */
2724 memcg_uncharge_kmem(memcg, 1 << order);
2727 page->mem_cgroup = memcg;
2730 void __memcg_kmem_uncharge_pages(struct page *page, int order)
2732 struct mem_cgroup *memcg = page->mem_cgroup;
2737 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2739 memcg_uncharge_kmem(memcg, 1 << order);
2740 page->mem_cgroup = NULL;
2743 struct mem_cgroup *__mem_cgroup_from_kmem(void *ptr)
2745 struct mem_cgroup *memcg = NULL;
2746 struct kmem_cache *cachep;
2749 page = virt_to_head_page(ptr);
2750 if (PageSlab(page)) {
2751 cachep = page->slab_cache;
2752 if (!is_root_cache(cachep))
2753 memcg = cachep->memcg_params.memcg;
2755 /* page allocated by alloc_kmem_pages */
2756 memcg = page->mem_cgroup;
2760 #endif /* CONFIG_MEMCG_KMEM */
2762 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2765 * Because tail pages are not marked as "used", set it. We're under
2766 * zone->lru_lock, 'splitting on pmd' and compound_lock.
2767 * charge/uncharge will be never happen and move_account() is done under
2768 * compound_lock(), so we don't have to take care of races.
2770 void mem_cgroup_split_huge_fixup(struct page *head)
2774 if (mem_cgroup_disabled())
2777 for (i = 1; i < HPAGE_PMD_NR; i++)
2778 head[i].mem_cgroup = head->mem_cgroup;
2780 __this_cpu_sub(head->mem_cgroup->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
2783 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2785 #ifdef CONFIG_MEMCG_SWAP
2786 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
2789 int val = (charge) ? 1 : -1;
2790 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
2794 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2795 * @entry: swap entry to be moved
2796 * @from: mem_cgroup which the entry is moved from
2797 * @to: mem_cgroup which the entry is moved to
2799 * It succeeds only when the swap_cgroup's record for this entry is the same
2800 * as the mem_cgroup's id of @from.
2802 * Returns 0 on success, -EINVAL on failure.
2804 * The caller must have charged to @to, IOW, called page_counter_charge() about
2805 * both res and memsw, and called css_get().
2807 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2808 struct mem_cgroup *from, struct mem_cgroup *to)
2810 unsigned short old_id, new_id;
2812 old_id = mem_cgroup_id(from);
2813 new_id = mem_cgroup_id(to);
2815 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2816 mem_cgroup_swap_statistics(from, false);
2817 mem_cgroup_swap_statistics(to, true);
2823 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2824 struct mem_cgroup *from, struct mem_cgroup *to)
2830 static DEFINE_MUTEX(memcg_limit_mutex);
2832 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
2833 unsigned long limit)
2835 unsigned long curusage;
2836 unsigned long oldusage;
2837 bool enlarge = false;
2842 * For keeping hierarchical_reclaim simple, how long we should retry
2843 * is depends on callers. We set our retry-count to be function
2844 * of # of children which we should visit in this loop.
2846 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2847 mem_cgroup_count_children(memcg);
2849 oldusage = page_counter_read(&memcg->memory);
2852 if (signal_pending(current)) {
2857 mutex_lock(&memcg_limit_mutex);
2858 if (limit > memcg->memsw.limit) {
2859 mutex_unlock(&memcg_limit_mutex);
2863 if (limit > memcg->memory.limit)
2865 ret = page_counter_limit(&memcg->memory, limit);
2866 mutex_unlock(&memcg_limit_mutex);
2871 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, true);
2873 curusage = page_counter_read(&memcg->memory);
2874 /* Usage is reduced ? */
2875 if (curusage >= oldusage)
2878 oldusage = curusage;
2879 } while (retry_count);
2881 if (!ret && enlarge)
2882 memcg_oom_recover(memcg);
2887 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
2888 unsigned long limit)
2890 unsigned long curusage;
2891 unsigned long oldusage;
2892 bool enlarge = false;
2896 /* see mem_cgroup_resize_res_limit */
2897 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2898 mem_cgroup_count_children(memcg);
2900 oldusage = page_counter_read(&memcg->memsw);
2903 if (signal_pending(current)) {
2908 mutex_lock(&memcg_limit_mutex);
2909 if (limit < memcg->memory.limit) {
2910 mutex_unlock(&memcg_limit_mutex);
2914 if (limit > memcg->memsw.limit)
2916 ret = page_counter_limit(&memcg->memsw, limit);
2917 mutex_unlock(&memcg_limit_mutex);
2922 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, false);
2924 curusage = page_counter_read(&memcg->memsw);
2925 /* Usage is reduced ? */
2926 if (curusage >= oldusage)
2929 oldusage = curusage;
2930 } while (retry_count);
2932 if (!ret && enlarge)
2933 memcg_oom_recover(memcg);
2938 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
2940 unsigned long *total_scanned)
2942 unsigned long nr_reclaimed = 0;
2943 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
2944 unsigned long reclaimed;
2946 struct mem_cgroup_tree_per_zone *mctz;
2947 unsigned long excess;
2948 unsigned long nr_scanned;
2953 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
2955 * This loop can run a while, specially if mem_cgroup's continuously
2956 * keep exceeding their soft limit and putting the system under
2963 mz = mem_cgroup_largest_soft_limit_node(mctz);
2968 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
2969 gfp_mask, &nr_scanned);
2970 nr_reclaimed += reclaimed;
2971 *total_scanned += nr_scanned;
2972 spin_lock_irq(&mctz->lock);
2973 __mem_cgroup_remove_exceeded(mz, mctz);
2976 * If we failed to reclaim anything from this memory cgroup
2977 * it is time to move on to the next cgroup
2981 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
2983 excess = soft_limit_excess(mz->memcg);
2985 * One school of thought says that we should not add
2986 * back the node to the tree if reclaim returns 0.
2987 * But our reclaim could return 0, simply because due
2988 * to priority we are exposing a smaller subset of
2989 * memory to reclaim from. Consider this as a longer
2992 /* If excess == 0, no tree ops */
2993 __mem_cgroup_insert_exceeded(mz, mctz, excess);
2994 spin_unlock_irq(&mctz->lock);
2995 css_put(&mz->memcg->css);
2998 * Could not reclaim anything and there are no more
2999 * mem cgroups to try or we seem to be looping without
3000 * reclaiming anything.
3002 if (!nr_reclaimed &&
3004 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3006 } while (!nr_reclaimed);
3008 css_put(&next_mz->memcg->css);
3009 return nr_reclaimed;
3013 * Test whether @memcg has children, dead or alive. Note that this
3014 * function doesn't care whether @memcg has use_hierarchy enabled and
3015 * returns %true if there are child csses according to the cgroup
3016 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
3018 static inline bool memcg_has_children(struct mem_cgroup *memcg)
3023 * The lock does not prevent addition or deletion of children, but
3024 * it prevents a new child from being initialized based on this
3025 * parent in css_online(), so it's enough to decide whether
3026 * hierarchically inherited attributes can still be changed or not.
3028 lockdep_assert_held(&memcg_create_mutex);
3031 ret = css_next_child(NULL, &memcg->css);
3037 * Reclaims as many pages from the given memcg as possible and moves
3038 * the rest to the parent.
3040 * Caller is responsible for holding css reference for memcg.
3042 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3044 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3046 /* we call try-to-free pages for make this cgroup empty */
3047 lru_add_drain_all();
3048 /* try to free all pages in this cgroup */
3049 while (nr_retries && page_counter_read(&memcg->memory)) {
3052 if (signal_pending(current))
3055 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3059 /* maybe some writeback is necessary */
3060 congestion_wait(BLK_RW_ASYNC, HZ/10);
3068 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3069 char *buf, size_t nbytes,
3072 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3074 if (mem_cgroup_is_root(memcg))
3076 return mem_cgroup_force_empty(memcg) ?: nbytes;
3079 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3082 return mem_cgroup_from_css(css)->use_hierarchy;
3085 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3086 struct cftype *cft, u64 val)
3089 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3090 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
3092 mutex_lock(&memcg_create_mutex);
3094 if (memcg->use_hierarchy == val)
3098 * If parent's use_hierarchy is set, we can't make any modifications
3099 * in the child subtrees. If it is unset, then the change can
3100 * occur, provided the current cgroup has no children.
3102 * For the root cgroup, parent_mem is NULL, we allow value to be
3103 * set if there are no children.
3105 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3106 (val == 1 || val == 0)) {
3107 if (!memcg_has_children(memcg))
3108 memcg->use_hierarchy = val;
3115 mutex_unlock(&memcg_create_mutex);
3120 static unsigned long tree_stat(struct mem_cgroup *memcg,
3121 enum mem_cgroup_stat_index idx)
3123 struct mem_cgroup *iter;
3126 /* Per-cpu values can be negative, use a signed accumulator */
3127 for_each_mem_cgroup_tree(iter, memcg)
3128 val += mem_cgroup_read_stat(iter, idx);
3130 if (val < 0) /* race ? */
3135 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3139 if (mem_cgroup_is_root(memcg)) {
3140 val = tree_stat(memcg, MEM_CGROUP_STAT_CACHE);
3141 val += tree_stat(memcg, MEM_CGROUP_STAT_RSS);
3143 val += tree_stat(memcg, MEM_CGROUP_STAT_SWAP);
3146 val = page_counter_read(&memcg->memory);
3148 val = page_counter_read(&memcg->memsw);
3150 return val << PAGE_SHIFT;
3161 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3164 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3165 struct page_counter *counter;
3167 switch (MEMFILE_TYPE(cft->private)) {
3169 counter = &memcg->memory;
3172 counter = &memcg->memsw;
3175 counter = &memcg->kmem;
3181 switch (MEMFILE_ATTR(cft->private)) {
3183 if (counter == &memcg->memory)
3184 return mem_cgroup_usage(memcg, false);
3185 if (counter == &memcg->memsw)
3186 return mem_cgroup_usage(memcg, true);
3187 return (u64)page_counter_read(counter) * PAGE_SIZE;
3189 return (u64)counter->limit * PAGE_SIZE;
3191 return (u64)counter->watermark * PAGE_SIZE;
3193 return counter->failcnt;
3194 case RES_SOFT_LIMIT:
3195 return (u64)memcg->soft_limit * PAGE_SIZE;
3201 #ifdef CONFIG_MEMCG_KMEM
3202 static int memcg_activate_kmem(struct mem_cgroup *memcg,
3203 unsigned long nr_pages)
3208 BUG_ON(memcg->kmemcg_id >= 0);
3209 BUG_ON(memcg->kmem_acct_activated);
3210 BUG_ON(memcg->kmem_acct_active);
3213 * For simplicity, we won't allow this to be disabled. It also can't
3214 * be changed if the cgroup has children already, or if tasks had
3217 * If tasks join before we set the limit, a person looking at
3218 * kmem.usage_in_bytes will have no way to determine when it took
3219 * place, which makes the value quite meaningless.
3221 * After it first became limited, changes in the value of the limit are
3222 * of course permitted.
3224 mutex_lock(&memcg_create_mutex);
3225 if (cgroup_has_tasks(memcg->css.cgroup) ||
3226 (memcg->use_hierarchy && memcg_has_children(memcg)))
3228 mutex_unlock(&memcg_create_mutex);
3232 memcg_id = memcg_alloc_cache_id();
3239 * We couldn't have accounted to this cgroup, because it hasn't got
3240 * activated yet, so this should succeed.
3242 err = page_counter_limit(&memcg->kmem, nr_pages);
3245 static_key_slow_inc(&memcg_kmem_enabled_key);
3247 * A memory cgroup is considered kmem-active as soon as it gets
3248 * kmemcg_id. Setting the id after enabling static branching will
3249 * guarantee no one starts accounting before all call sites are
3252 memcg->kmemcg_id = memcg_id;
3253 memcg->kmem_acct_activated = true;
3254 memcg->kmem_acct_active = true;
3259 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
3260 unsigned long limit)
3264 mutex_lock(&memcg_limit_mutex);
3265 if (!memcg_kmem_is_active(memcg))
3266 ret = memcg_activate_kmem(memcg, limit);
3268 ret = page_counter_limit(&memcg->kmem, limit);
3269 mutex_unlock(&memcg_limit_mutex);
3273 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
3276 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
3281 mutex_lock(&memcg_limit_mutex);
3283 * If the parent cgroup is not kmem-active now, it cannot be activated
3284 * after this point, because it has at least one child already.
3286 if (memcg_kmem_is_active(parent))
3287 ret = memcg_activate_kmem(memcg, PAGE_COUNTER_MAX);
3288 mutex_unlock(&memcg_limit_mutex);
3292 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
3293 unsigned long limit)
3297 #endif /* CONFIG_MEMCG_KMEM */
3300 * The user of this function is...
3303 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3304 char *buf, size_t nbytes, loff_t off)
3306 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3307 unsigned long nr_pages;
3310 buf = strstrip(buf);
3311 ret = page_counter_memparse(buf, "-1", &nr_pages);
3315 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3317 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3321 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3323 ret = mem_cgroup_resize_limit(memcg, nr_pages);
3326 ret = mem_cgroup_resize_memsw_limit(memcg, nr_pages);
3329 ret = memcg_update_kmem_limit(memcg, nr_pages);
3333 case RES_SOFT_LIMIT:
3334 memcg->soft_limit = nr_pages;
3338 return ret ?: nbytes;
3341 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3342 size_t nbytes, loff_t off)
3344 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3345 struct page_counter *counter;
3347 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3349 counter = &memcg->memory;
3352 counter = &memcg->memsw;
3355 counter = &memcg->kmem;
3361 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3363 page_counter_reset_watermark(counter);
3366 counter->failcnt = 0;
3375 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3378 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3382 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3383 struct cftype *cft, u64 val)
3385 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3387 if (val & ~MOVE_MASK)
3391 * No kind of locking is needed in here, because ->can_attach() will
3392 * check this value once in the beginning of the process, and then carry
3393 * on with stale data. This means that changes to this value will only
3394 * affect task migrations starting after the change.
3396 memcg->move_charge_at_immigrate = val;
3400 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3401 struct cftype *cft, u64 val)
3408 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3412 unsigned int lru_mask;
3415 static const struct numa_stat stats[] = {
3416 { "total", LRU_ALL },
3417 { "file", LRU_ALL_FILE },
3418 { "anon", LRU_ALL_ANON },
3419 { "unevictable", BIT(LRU_UNEVICTABLE) },
3421 const struct numa_stat *stat;
3424 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3426 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3427 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3428 seq_printf(m, "%s=%lu", stat->name, nr);
3429 for_each_node_state(nid, N_MEMORY) {
3430 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3432 seq_printf(m, " N%d=%lu", nid, nr);
3437 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3438 struct mem_cgroup *iter;
3441 for_each_mem_cgroup_tree(iter, memcg)
3442 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3443 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3444 for_each_node_state(nid, N_MEMORY) {
3446 for_each_mem_cgroup_tree(iter, memcg)
3447 nr += mem_cgroup_node_nr_lru_pages(
3448 iter, nid, stat->lru_mask);
3449 seq_printf(m, " N%d=%lu", nid, nr);
3456 #endif /* CONFIG_NUMA */
3458 static int memcg_stat_show(struct seq_file *m, void *v)
3460 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3461 unsigned long memory, memsw;
3462 struct mem_cgroup *mi;
3465 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_stat_names) !=
3466 MEM_CGROUP_STAT_NSTATS);
3467 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_events_names) !=
3468 MEM_CGROUP_EVENTS_NSTATS);
3469 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3471 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3472 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
3474 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
3475 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
3478 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
3479 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
3480 mem_cgroup_read_events(memcg, i));
3482 for (i = 0; i < NR_LRU_LISTS; i++)
3483 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3484 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
3486 /* Hierarchical information */
3487 memory = memsw = PAGE_COUNTER_MAX;
3488 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3489 memory = min(memory, mi->memory.limit);
3490 memsw = min(memsw, mi->memsw.limit);
3492 seq_printf(m, "hierarchical_memory_limit %llu\n",
3493 (u64)memory * PAGE_SIZE);
3494 if (do_swap_account)
3495 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3496 (u64)memsw * PAGE_SIZE);
3498 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3501 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
3503 for_each_mem_cgroup_tree(mi, memcg)
3504 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
3505 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
3508 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
3509 unsigned long long val = 0;
3511 for_each_mem_cgroup_tree(mi, memcg)
3512 val += mem_cgroup_read_events(mi, i);
3513 seq_printf(m, "total_%s %llu\n",
3514 mem_cgroup_events_names[i], val);
3517 for (i = 0; i < NR_LRU_LISTS; i++) {
3518 unsigned long long val = 0;
3520 for_each_mem_cgroup_tree(mi, memcg)
3521 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
3522 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
3525 #ifdef CONFIG_DEBUG_VM
3528 struct mem_cgroup_per_zone *mz;
3529 struct zone_reclaim_stat *rstat;
3530 unsigned long recent_rotated[2] = {0, 0};
3531 unsigned long recent_scanned[2] = {0, 0};
3533 for_each_online_node(nid)
3534 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3535 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
3536 rstat = &mz->lruvec.reclaim_stat;
3538 recent_rotated[0] += rstat->recent_rotated[0];
3539 recent_rotated[1] += rstat->recent_rotated[1];
3540 recent_scanned[0] += rstat->recent_scanned[0];
3541 recent_scanned[1] += rstat->recent_scanned[1];
3543 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3544 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3545 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3546 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3553 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3556 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3558 return mem_cgroup_swappiness(memcg);
3561 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3562 struct cftype *cft, u64 val)
3564 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3570 memcg->swappiness = val;
3572 vm_swappiness = val;
3577 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3579 struct mem_cgroup_threshold_ary *t;
3580 unsigned long usage;
3585 t = rcu_dereference(memcg->thresholds.primary);
3587 t = rcu_dereference(memcg->memsw_thresholds.primary);
3592 usage = mem_cgroup_usage(memcg, swap);
3595 * current_threshold points to threshold just below or equal to usage.
3596 * If it's not true, a threshold was crossed after last
3597 * call of __mem_cgroup_threshold().
3599 i = t->current_threshold;
3602 * Iterate backward over array of thresholds starting from
3603 * current_threshold and check if a threshold is crossed.
3604 * If none of thresholds below usage is crossed, we read
3605 * only one element of the array here.
3607 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3608 eventfd_signal(t->entries[i].eventfd, 1);
3610 /* i = current_threshold + 1 */
3614 * Iterate forward over array of thresholds starting from
3615 * current_threshold+1 and check if a threshold is crossed.
3616 * If none of thresholds above usage is crossed, we read
3617 * only one element of the array here.
3619 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3620 eventfd_signal(t->entries[i].eventfd, 1);
3622 /* Update current_threshold */
3623 t->current_threshold = i - 1;
3628 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3631 __mem_cgroup_threshold(memcg, false);
3632 if (do_swap_account)
3633 __mem_cgroup_threshold(memcg, true);
3635 memcg = parent_mem_cgroup(memcg);
3639 static int compare_thresholds(const void *a, const void *b)
3641 const struct mem_cgroup_threshold *_a = a;
3642 const struct mem_cgroup_threshold *_b = b;
3644 if (_a->threshold > _b->threshold)
3647 if (_a->threshold < _b->threshold)
3653 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3655 struct mem_cgroup_eventfd_list *ev;
3657 spin_lock(&memcg_oom_lock);
3659 list_for_each_entry(ev, &memcg->oom_notify, list)
3660 eventfd_signal(ev->eventfd, 1);
3662 spin_unlock(&memcg_oom_lock);
3666 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3668 struct mem_cgroup *iter;
3670 for_each_mem_cgroup_tree(iter, memcg)
3671 mem_cgroup_oom_notify_cb(iter);
3674 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3675 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3677 struct mem_cgroup_thresholds *thresholds;
3678 struct mem_cgroup_threshold_ary *new;
3679 unsigned long threshold;
3680 unsigned long usage;
3683 ret = page_counter_memparse(args, "-1", &threshold);
3687 mutex_lock(&memcg->thresholds_lock);
3690 thresholds = &memcg->thresholds;
3691 usage = mem_cgroup_usage(memcg, false);
3692 } else if (type == _MEMSWAP) {
3693 thresholds = &memcg->memsw_thresholds;
3694 usage = mem_cgroup_usage(memcg, true);
3698 /* Check if a threshold crossed before adding a new one */
3699 if (thresholds->primary)
3700 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3702 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3704 /* Allocate memory for new array of thresholds */
3705 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3713 /* Copy thresholds (if any) to new array */
3714 if (thresholds->primary) {
3715 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3716 sizeof(struct mem_cgroup_threshold));
3719 /* Add new threshold */
3720 new->entries[size - 1].eventfd = eventfd;
3721 new->entries[size - 1].threshold = threshold;
3723 /* Sort thresholds. Registering of new threshold isn't time-critical */
3724 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3725 compare_thresholds, NULL);
3727 /* Find current threshold */
3728 new->current_threshold = -1;
3729 for (i = 0; i < size; i++) {
3730 if (new->entries[i].threshold <= usage) {
3732 * new->current_threshold will not be used until
3733 * rcu_assign_pointer(), so it's safe to increment
3736 ++new->current_threshold;
3741 /* Free old spare buffer and save old primary buffer as spare */
3742 kfree(thresholds->spare);
3743 thresholds->spare = thresholds->primary;
3745 rcu_assign_pointer(thresholds->primary, new);
3747 /* To be sure that nobody uses thresholds */
3751 mutex_unlock(&memcg->thresholds_lock);
3756 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3757 struct eventfd_ctx *eventfd, const char *args)
3759 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
3762 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
3763 struct eventfd_ctx *eventfd, const char *args)
3765 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
3768 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3769 struct eventfd_ctx *eventfd, enum res_type type)
3771 struct mem_cgroup_thresholds *thresholds;
3772 struct mem_cgroup_threshold_ary *new;
3773 unsigned long usage;
3776 mutex_lock(&memcg->thresholds_lock);
3779 thresholds = &memcg->thresholds;
3780 usage = mem_cgroup_usage(memcg, false);
3781 } else if (type == _MEMSWAP) {
3782 thresholds = &memcg->memsw_thresholds;
3783 usage = mem_cgroup_usage(memcg, true);
3787 if (!thresholds->primary)
3790 /* Check if a threshold crossed before removing */
3791 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3793 /* Calculate new number of threshold */
3795 for (i = 0; i < thresholds->primary->size; i++) {
3796 if (thresholds->primary->entries[i].eventfd != eventfd)
3800 new = thresholds->spare;
3802 /* Set thresholds array to NULL if we don't have thresholds */
3811 /* Copy thresholds and find current threshold */
3812 new->current_threshold = -1;
3813 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3814 if (thresholds->primary->entries[i].eventfd == eventfd)
3817 new->entries[j] = thresholds->primary->entries[i];
3818 if (new->entries[j].threshold <= usage) {
3820 * new->current_threshold will not be used
3821 * until rcu_assign_pointer(), so it's safe to increment
3824 ++new->current_threshold;
3830 /* Swap primary and spare array */
3831 thresholds->spare = thresholds->primary;
3832 /* If all events are unregistered, free the spare array */
3834 kfree(thresholds->spare);
3835 thresholds->spare = NULL;
3838 rcu_assign_pointer(thresholds->primary, new);
3840 /* To be sure that nobody uses thresholds */
3843 mutex_unlock(&memcg->thresholds_lock);
3846 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3847 struct eventfd_ctx *eventfd)
3849 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
3852 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3853 struct eventfd_ctx *eventfd)
3855 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
3858 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
3859 struct eventfd_ctx *eventfd, const char *args)
3861 struct mem_cgroup_eventfd_list *event;
3863 event = kmalloc(sizeof(*event), GFP_KERNEL);
3867 spin_lock(&memcg_oom_lock);
3869 event->eventfd = eventfd;
3870 list_add(&event->list, &memcg->oom_notify);
3872 /* already in OOM ? */
3873 if (atomic_read(&memcg->under_oom))
3874 eventfd_signal(eventfd, 1);
3875 spin_unlock(&memcg_oom_lock);
3880 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
3881 struct eventfd_ctx *eventfd)
3883 struct mem_cgroup_eventfd_list *ev, *tmp;
3885 spin_lock(&memcg_oom_lock);
3887 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
3888 if (ev->eventfd == eventfd) {
3889 list_del(&ev->list);
3894 spin_unlock(&memcg_oom_lock);
3897 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
3899 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
3901 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
3902 seq_printf(sf, "under_oom %d\n", (bool)atomic_read(&memcg->under_oom));
3906 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
3907 struct cftype *cft, u64 val)
3909 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3911 /* cannot set to root cgroup and only 0 and 1 are allowed */
3912 if (!css->parent || !((val == 0) || (val == 1)))
3915 memcg->oom_kill_disable = val;
3917 memcg_oom_recover(memcg);
3922 #ifdef CONFIG_MEMCG_KMEM
3923 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
3927 ret = memcg_propagate_kmem(memcg);
3931 return mem_cgroup_sockets_init(memcg, ss);
3934 static void memcg_deactivate_kmem(struct mem_cgroup *memcg)
3936 struct cgroup_subsys_state *css;
3937 struct mem_cgroup *parent, *child;
3940 if (!memcg->kmem_acct_active)
3944 * Clear the 'active' flag before clearing memcg_caches arrays entries.
3945 * Since we take the slab_mutex in memcg_deactivate_kmem_caches(), it
3946 * guarantees no cache will be created for this cgroup after we are
3947 * done (see memcg_create_kmem_cache()).
3949 memcg->kmem_acct_active = false;
3951 memcg_deactivate_kmem_caches(memcg);
3953 kmemcg_id = memcg->kmemcg_id;
3954 BUG_ON(kmemcg_id < 0);
3956 parent = parent_mem_cgroup(memcg);
3958 parent = root_mem_cgroup;
3961 * Change kmemcg_id of this cgroup and all its descendants to the
3962 * parent's id, and then move all entries from this cgroup's list_lrus
3963 * to ones of the parent. After we have finished, all list_lrus
3964 * corresponding to this cgroup are guaranteed to remain empty. The
3965 * ordering is imposed by list_lru_node->lock taken by
3966 * memcg_drain_all_list_lrus().
3968 css_for_each_descendant_pre(css, &memcg->css) {
3969 child = mem_cgroup_from_css(css);
3970 BUG_ON(child->kmemcg_id != kmemcg_id);
3971 child->kmemcg_id = parent->kmemcg_id;
3972 if (!memcg->use_hierarchy)
3975 memcg_drain_all_list_lrus(kmemcg_id, parent->kmemcg_id);
3977 memcg_free_cache_id(kmemcg_id);
3980 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
3982 if (memcg->kmem_acct_activated) {
3983 memcg_destroy_kmem_caches(memcg);
3984 static_key_slow_dec(&memcg_kmem_enabled_key);
3985 WARN_ON(page_counter_read(&memcg->kmem));
3987 mem_cgroup_sockets_destroy(memcg);
3990 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
3995 static void memcg_deactivate_kmem(struct mem_cgroup *memcg)
3999 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
4005 * DO NOT USE IN NEW FILES.
4007 * "cgroup.event_control" implementation.
4009 * This is way over-engineered. It tries to support fully configurable
4010 * events for each user. Such level of flexibility is completely
4011 * unnecessary especially in the light of the planned unified hierarchy.
4013 * Please deprecate this and replace with something simpler if at all
4018 * Unregister event and free resources.
4020 * Gets called from workqueue.
4022 static void memcg_event_remove(struct work_struct *work)
4024 struct mem_cgroup_event *event =
4025 container_of(work, struct mem_cgroup_event, remove);
4026 struct mem_cgroup *memcg = event->memcg;
4028 remove_wait_queue(event->wqh, &event->wait);
4030 event->unregister_event(memcg, event->eventfd);
4032 /* Notify userspace the event is going away. */
4033 eventfd_signal(event->eventfd, 1);
4035 eventfd_ctx_put(event->eventfd);
4037 css_put(&memcg->css);
4041 * Gets called on POLLHUP on eventfd when user closes it.
4043 * Called with wqh->lock held and interrupts disabled.
4045 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
4046 int sync, void *key)
4048 struct mem_cgroup_event *event =
4049 container_of(wait, struct mem_cgroup_event, wait);
4050 struct mem_cgroup *memcg = event->memcg;
4051 unsigned long flags = (unsigned long)key;
4053 if (flags & POLLHUP) {
4055 * If the event has been detached at cgroup removal, we
4056 * can simply return knowing the other side will cleanup
4059 * We can't race against event freeing since the other
4060 * side will require wqh->lock via remove_wait_queue(),
4063 spin_lock(&memcg->event_list_lock);
4064 if (!list_empty(&event->list)) {
4065 list_del_init(&event->list);
4067 * We are in atomic context, but cgroup_event_remove()
4068 * may sleep, so we have to call it in workqueue.
4070 schedule_work(&event->remove);
4072 spin_unlock(&memcg->event_list_lock);
4078 static void memcg_event_ptable_queue_proc(struct file *file,
4079 wait_queue_head_t *wqh, poll_table *pt)
4081 struct mem_cgroup_event *event =
4082 container_of(pt, struct mem_cgroup_event, pt);
4085 add_wait_queue(wqh, &event->wait);
4089 * DO NOT USE IN NEW FILES.
4091 * Parse input and register new cgroup event handler.
4093 * Input must be in format '<event_fd> <control_fd> <args>'.
4094 * Interpretation of args is defined by control file implementation.
4096 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4097 char *buf, size_t nbytes, loff_t off)
4099 struct cgroup_subsys_state *css = of_css(of);
4100 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4101 struct mem_cgroup_event *event;
4102 struct cgroup_subsys_state *cfile_css;
4103 unsigned int efd, cfd;
4110 buf = strstrip(buf);
4112 efd = simple_strtoul(buf, &endp, 10);
4117 cfd = simple_strtoul(buf, &endp, 10);
4118 if ((*endp != ' ') && (*endp != '\0'))
4122 event = kzalloc(sizeof(*event), GFP_KERNEL);
4126 event->memcg = memcg;
4127 INIT_LIST_HEAD(&event->list);
4128 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4129 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4130 INIT_WORK(&event->remove, memcg_event_remove);
4138 event->eventfd = eventfd_ctx_fileget(efile.file);
4139 if (IS_ERR(event->eventfd)) {
4140 ret = PTR_ERR(event->eventfd);
4147 goto out_put_eventfd;
4150 /* the process need read permission on control file */
4151 /* AV: shouldn't we check that it's been opened for read instead? */
4152 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4157 * Determine the event callbacks and set them in @event. This used
4158 * to be done via struct cftype but cgroup core no longer knows
4159 * about these events. The following is crude but the whole thing
4160 * is for compatibility anyway.
4162 * DO NOT ADD NEW FILES.
4164 name = cfile.file->f_path.dentry->d_name.name;
4166 if (!strcmp(name, "memory.usage_in_bytes")) {
4167 event->register_event = mem_cgroup_usage_register_event;
4168 event->unregister_event = mem_cgroup_usage_unregister_event;
4169 } else if (!strcmp(name, "memory.oom_control")) {
4170 event->register_event = mem_cgroup_oom_register_event;
4171 event->unregister_event = mem_cgroup_oom_unregister_event;
4172 } else if (!strcmp(name, "memory.pressure_level")) {
4173 event->register_event = vmpressure_register_event;
4174 event->unregister_event = vmpressure_unregister_event;
4175 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4176 event->register_event = memsw_cgroup_usage_register_event;
4177 event->unregister_event = memsw_cgroup_usage_unregister_event;
4184 * Verify @cfile should belong to @css. Also, remaining events are
4185 * automatically removed on cgroup destruction but the removal is
4186 * asynchronous, so take an extra ref on @css.
4188 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4189 &memory_cgrp_subsys);
4191 if (IS_ERR(cfile_css))
4193 if (cfile_css != css) {
4198 ret = event->register_event(memcg, event->eventfd, buf);
4202 efile.file->f_op->poll(efile.file, &event->pt);
4204 spin_lock(&memcg->event_list_lock);
4205 list_add(&event->list, &memcg->event_list);
4206 spin_unlock(&memcg->event_list_lock);
4218 eventfd_ctx_put(event->eventfd);
4227 static struct cftype mem_cgroup_legacy_files[] = {
4229 .name = "usage_in_bytes",
4230 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4231 .read_u64 = mem_cgroup_read_u64,
4234 .name = "max_usage_in_bytes",
4235 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4236 .write = mem_cgroup_reset,
4237 .read_u64 = mem_cgroup_read_u64,
4240 .name = "limit_in_bytes",
4241 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4242 .write = mem_cgroup_write,
4243 .read_u64 = mem_cgroup_read_u64,
4246 .name = "soft_limit_in_bytes",
4247 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4248 .write = mem_cgroup_write,
4249 .read_u64 = mem_cgroup_read_u64,
4253 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4254 .write = mem_cgroup_reset,
4255 .read_u64 = mem_cgroup_read_u64,
4259 .seq_show = memcg_stat_show,
4262 .name = "force_empty",
4263 .write = mem_cgroup_force_empty_write,
4266 .name = "use_hierarchy",
4267 .write_u64 = mem_cgroup_hierarchy_write,
4268 .read_u64 = mem_cgroup_hierarchy_read,
4271 .name = "cgroup.event_control", /* XXX: for compat */
4272 .write = memcg_write_event_control,
4273 .flags = CFTYPE_NO_PREFIX,
4277 .name = "swappiness",
4278 .read_u64 = mem_cgroup_swappiness_read,
4279 .write_u64 = mem_cgroup_swappiness_write,
4282 .name = "move_charge_at_immigrate",
4283 .read_u64 = mem_cgroup_move_charge_read,
4284 .write_u64 = mem_cgroup_move_charge_write,
4287 .name = "oom_control",
4288 .seq_show = mem_cgroup_oom_control_read,
4289 .write_u64 = mem_cgroup_oom_control_write,
4290 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4293 .name = "pressure_level",
4297 .name = "numa_stat",
4298 .seq_show = memcg_numa_stat_show,
4301 #ifdef CONFIG_MEMCG_KMEM
4303 .name = "kmem.limit_in_bytes",
4304 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4305 .write = mem_cgroup_write,
4306 .read_u64 = mem_cgroup_read_u64,
4309 .name = "kmem.usage_in_bytes",
4310 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4311 .read_u64 = mem_cgroup_read_u64,
4314 .name = "kmem.failcnt",
4315 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4316 .write = mem_cgroup_reset,
4317 .read_u64 = mem_cgroup_read_u64,
4320 .name = "kmem.max_usage_in_bytes",
4321 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4322 .write = mem_cgroup_reset,
4323 .read_u64 = mem_cgroup_read_u64,
4325 #ifdef CONFIG_SLABINFO
4327 .name = "kmem.slabinfo",
4328 .seq_start = slab_start,
4329 .seq_next = slab_next,
4330 .seq_stop = slab_stop,
4331 .seq_show = memcg_slab_show,
4335 { }, /* terminate */
4338 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4340 struct mem_cgroup_per_node *pn;
4341 struct mem_cgroup_per_zone *mz;
4342 int zone, tmp = node;
4344 * This routine is called against possible nodes.
4345 * But it's BUG to call kmalloc() against offline node.
4347 * TODO: this routine can waste much memory for nodes which will
4348 * never be onlined. It's better to use memory hotplug callback
4351 if (!node_state(node, N_NORMAL_MEMORY))
4353 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4357 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4358 mz = &pn->zoneinfo[zone];
4359 lruvec_init(&mz->lruvec);
4360 mz->usage_in_excess = 0;
4361 mz->on_tree = false;
4364 memcg->nodeinfo[node] = pn;
4368 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4370 kfree(memcg->nodeinfo[node]);
4373 static struct mem_cgroup *mem_cgroup_alloc(void)
4375 struct mem_cgroup *memcg;
4378 size = sizeof(struct mem_cgroup);
4379 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4381 memcg = kzalloc(size, GFP_KERNEL);
4385 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4388 spin_lock_init(&memcg->pcp_counter_lock);
4397 * At destroying mem_cgroup, references from swap_cgroup can remain.
4398 * (scanning all at force_empty is too costly...)
4400 * Instead of clearing all references at force_empty, we remember
4401 * the number of reference from swap_cgroup and free mem_cgroup when
4402 * it goes down to 0.
4404 * Removal of cgroup itself succeeds regardless of refs from swap.
4407 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4411 mem_cgroup_remove_from_trees(memcg);
4414 free_mem_cgroup_per_zone_info(memcg, node);
4416 free_percpu(memcg->stat);
4421 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4423 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4425 if (!memcg->memory.parent)
4427 return mem_cgroup_from_counter(memcg->memory.parent, memory);
4429 EXPORT_SYMBOL(parent_mem_cgroup);
4431 static struct cgroup_subsys_state * __ref
4432 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4434 struct mem_cgroup *memcg;
4435 long error = -ENOMEM;
4438 memcg = mem_cgroup_alloc();
4440 return ERR_PTR(error);
4443 if (alloc_mem_cgroup_per_zone_info(memcg, node))
4447 if (parent_css == NULL) {
4448 root_mem_cgroup = memcg;
4449 page_counter_init(&memcg->memory, NULL);
4450 memcg->high = PAGE_COUNTER_MAX;
4451 memcg->soft_limit = PAGE_COUNTER_MAX;
4452 page_counter_init(&memcg->memsw, NULL);
4453 page_counter_init(&memcg->kmem, NULL);
4456 memcg->last_scanned_node = MAX_NUMNODES;
4457 INIT_LIST_HEAD(&memcg->oom_notify);
4458 memcg->move_charge_at_immigrate = 0;
4459 mutex_init(&memcg->thresholds_lock);
4460 spin_lock_init(&memcg->move_lock);
4461 vmpressure_init(&memcg->vmpressure);
4462 INIT_LIST_HEAD(&memcg->event_list);
4463 spin_lock_init(&memcg->event_list_lock);
4464 #ifdef CONFIG_MEMCG_KMEM
4465 memcg->kmemcg_id = -1;
4471 __mem_cgroup_free(memcg);
4472 return ERR_PTR(error);
4476 mem_cgroup_css_online(struct cgroup_subsys_state *css)
4478 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4479 struct mem_cgroup *parent = mem_cgroup_from_css(css->parent);
4482 if (css->id > MEM_CGROUP_ID_MAX)
4488 mutex_lock(&memcg_create_mutex);
4490 memcg->use_hierarchy = parent->use_hierarchy;
4491 memcg->oom_kill_disable = parent->oom_kill_disable;
4492 memcg->swappiness = mem_cgroup_swappiness(parent);
4494 if (parent->use_hierarchy) {
4495 page_counter_init(&memcg->memory, &parent->memory);
4496 memcg->high = PAGE_COUNTER_MAX;
4497 memcg->soft_limit = PAGE_COUNTER_MAX;
4498 page_counter_init(&memcg->memsw, &parent->memsw);
4499 page_counter_init(&memcg->kmem, &parent->kmem);
4502 * No need to take a reference to the parent because cgroup
4503 * core guarantees its existence.
4506 page_counter_init(&memcg->memory, NULL);
4507 memcg->high = PAGE_COUNTER_MAX;
4508 memcg->soft_limit = PAGE_COUNTER_MAX;
4509 page_counter_init(&memcg->memsw, NULL);
4510 page_counter_init(&memcg->kmem, NULL);
4512 * Deeper hierachy with use_hierarchy == false doesn't make
4513 * much sense so let cgroup subsystem know about this
4514 * unfortunate state in our controller.
4516 if (parent != root_mem_cgroup)
4517 memory_cgrp_subsys.broken_hierarchy = true;
4519 mutex_unlock(&memcg_create_mutex);
4521 ret = memcg_init_kmem(memcg, &memory_cgrp_subsys);
4526 * Make sure the memcg is initialized: mem_cgroup_iter()
4527 * orders reading memcg->initialized against its callers
4528 * reading the memcg members.
4530 smp_store_release(&memcg->initialized, 1);
4535 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4537 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4538 struct mem_cgroup_event *event, *tmp;
4541 * Unregister events and notify userspace.
4542 * Notify userspace about cgroup removing only after rmdir of cgroup
4543 * directory to avoid race between userspace and kernelspace.
4545 spin_lock(&memcg->event_list_lock);
4546 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4547 list_del_init(&event->list);
4548 schedule_work(&event->remove);
4550 spin_unlock(&memcg->event_list_lock);
4552 vmpressure_cleanup(&memcg->vmpressure);
4554 memcg_deactivate_kmem(memcg);
4557 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4559 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4561 memcg_destroy_kmem(memcg);
4562 __mem_cgroup_free(memcg);
4566 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4567 * @css: the target css
4569 * Reset the states of the mem_cgroup associated with @css. This is
4570 * invoked when the userland requests disabling on the default hierarchy
4571 * but the memcg is pinned through dependency. The memcg should stop
4572 * applying policies and should revert to the vanilla state as it may be
4573 * made visible again.
4575 * The current implementation only resets the essential configurations.
4576 * This needs to be expanded to cover all the visible parts.
4578 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4580 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4582 mem_cgroup_resize_limit(memcg, PAGE_COUNTER_MAX);
4583 mem_cgroup_resize_memsw_limit(memcg, PAGE_COUNTER_MAX);
4584 memcg_update_kmem_limit(memcg, PAGE_COUNTER_MAX);
4586 memcg->high = PAGE_COUNTER_MAX;
4587 memcg->soft_limit = PAGE_COUNTER_MAX;
4591 /* Handlers for move charge at task migration. */
4592 static int mem_cgroup_do_precharge(unsigned long count)
4596 /* Try a single bulk charge without reclaim first */
4597 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_WAIT, count);
4599 mc.precharge += count;
4602 if (ret == -EINTR) {
4603 cancel_charge(root_mem_cgroup, count);
4607 /* Try charges one by one with reclaim */
4609 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_NORETRY, 1);
4611 * In case of failure, any residual charges against
4612 * mc.to will be dropped by mem_cgroup_clear_mc()
4613 * later on. However, cancel any charges that are
4614 * bypassed to root right away or they'll be lost.
4617 cancel_charge(root_mem_cgroup, 1);
4627 * get_mctgt_type - get target type of moving charge
4628 * @vma: the vma the pte to be checked belongs
4629 * @addr: the address corresponding to the pte to be checked
4630 * @ptent: the pte to be checked
4631 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4634 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4635 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4636 * move charge. if @target is not NULL, the page is stored in target->page
4637 * with extra refcnt got(Callers should handle it).
4638 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4639 * target for charge migration. if @target is not NULL, the entry is stored
4642 * Called with pte lock held.
4649 enum mc_target_type {
4655 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4656 unsigned long addr, pte_t ptent)
4658 struct page *page = vm_normal_page(vma, addr, ptent);
4660 if (!page || !page_mapped(page))
4662 if (PageAnon(page)) {
4663 if (!(mc.flags & MOVE_ANON))
4666 if (!(mc.flags & MOVE_FILE))
4669 if (!get_page_unless_zero(page))
4676 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4677 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4679 struct page *page = NULL;
4680 swp_entry_t ent = pte_to_swp_entry(ptent);
4682 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
4685 * Because lookup_swap_cache() updates some statistics counter,
4686 * we call find_get_page() with swapper_space directly.
4688 page = find_get_page(swap_address_space(ent), ent.val);
4689 if (do_swap_account)
4690 entry->val = ent.val;
4695 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4696 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4702 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4703 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4705 struct page *page = NULL;
4706 struct address_space *mapping;
4709 if (!vma->vm_file) /* anonymous vma */
4711 if (!(mc.flags & MOVE_FILE))
4714 mapping = vma->vm_file->f_mapping;
4715 pgoff = linear_page_index(vma, addr);
4717 /* page is moved even if it's not RSS of this task(page-faulted). */
4719 /* shmem/tmpfs may report page out on swap: account for that too. */
4720 if (shmem_mapping(mapping)) {
4721 page = find_get_entry(mapping, pgoff);
4722 if (radix_tree_exceptional_entry(page)) {
4723 swp_entry_t swp = radix_to_swp_entry(page);
4724 if (do_swap_account)
4726 page = find_get_page(swap_address_space(swp), swp.val);
4729 page = find_get_page(mapping, pgoff);
4731 page = find_get_page(mapping, pgoff);
4737 * mem_cgroup_move_account - move account of the page
4739 * @nr_pages: number of regular pages (>1 for huge pages)
4740 * @from: mem_cgroup which the page is moved from.
4741 * @to: mem_cgroup which the page is moved to. @from != @to.
4743 * The caller must confirm following.
4744 * - page is not on LRU (isolate_page() is useful.)
4745 * - compound_lock is held when nr_pages > 1
4747 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4750 static int mem_cgroup_move_account(struct page *page,
4751 unsigned int nr_pages,
4752 struct mem_cgroup *from,
4753 struct mem_cgroup *to)
4755 unsigned long flags;
4758 VM_BUG_ON(from == to);
4759 VM_BUG_ON_PAGE(PageLRU(page), page);
4761 * The page is isolated from LRU. So, collapse function
4762 * will not handle this page. But page splitting can happen.
4763 * Do this check under compound_page_lock(). The caller should
4767 if (nr_pages > 1 && !PageTransHuge(page))
4771 * Prevent mem_cgroup_migrate() from looking at page->mem_cgroup
4772 * of its source page while we change it: page migration takes
4773 * both pages off the LRU, but page cache replacement doesn't.
4775 if (!trylock_page(page))
4779 if (page->mem_cgroup != from)
4782 spin_lock_irqsave(&from->move_lock, flags);
4784 if (!PageAnon(page) && page_mapped(page)) {
4785 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4787 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4791 if (PageWriteback(page)) {
4792 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4794 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4799 * It is safe to change page->mem_cgroup here because the page
4800 * is referenced, charged, and isolated - we can't race with
4801 * uncharging, charging, migration, or LRU putback.
4804 /* caller should have done css_get */
4805 page->mem_cgroup = to;
4806 spin_unlock_irqrestore(&from->move_lock, flags);
4810 local_lock_irq(event_lock);
4811 mem_cgroup_charge_statistics(to, page, nr_pages);
4812 memcg_check_events(to, page);
4813 mem_cgroup_charge_statistics(from, page, -nr_pages);
4814 memcg_check_events(from, page);
4815 local_unlock_irq(event_lock);
4822 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
4823 unsigned long addr, pte_t ptent, union mc_target *target)
4825 struct page *page = NULL;
4826 enum mc_target_type ret = MC_TARGET_NONE;
4827 swp_entry_t ent = { .val = 0 };
4829 if (pte_present(ptent))
4830 page = mc_handle_present_pte(vma, addr, ptent);
4831 else if (is_swap_pte(ptent))
4832 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
4833 else if (pte_none(ptent))
4834 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4836 if (!page && !ent.val)
4840 * Do only loose check w/o serialization.
4841 * mem_cgroup_move_account() checks the page is valid or
4842 * not under LRU exclusion.
4844 if (page->mem_cgroup == mc.from) {
4845 ret = MC_TARGET_PAGE;
4847 target->page = page;
4849 if (!ret || !target)
4852 /* There is a swap entry and a page doesn't exist or isn't charged */
4853 if (ent.val && !ret &&
4854 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
4855 ret = MC_TARGET_SWAP;
4862 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4864 * We don't consider swapping or file mapped pages because THP does not
4865 * support them for now.
4866 * Caller should make sure that pmd_trans_huge(pmd) is true.
4868 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4869 unsigned long addr, pmd_t pmd, union mc_target *target)
4871 struct page *page = NULL;
4872 enum mc_target_type ret = MC_TARGET_NONE;
4874 page = pmd_page(pmd);
4875 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
4876 if (!(mc.flags & MOVE_ANON))
4878 if (page->mem_cgroup == mc.from) {
4879 ret = MC_TARGET_PAGE;
4882 target->page = page;
4888 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4889 unsigned long addr, pmd_t pmd, union mc_target *target)
4891 return MC_TARGET_NONE;
4895 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4896 unsigned long addr, unsigned long end,
4897 struct mm_walk *walk)
4899 struct vm_area_struct *vma = walk->vma;
4903 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
4904 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
4905 mc.precharge += HPAGE_PMD_NR;
4910 if (pmd_trans_unstable(pmd))
4912 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4913 for (; addr != end; pte++, addr += PAGE_SIZE)
4914 if (get_mctgt_type(vma, addr, *pte, NULL))
4915 mc.precharge++; /* increment precharge temporarily */
4916 pte_unmap_unlock(pte - 1, ptl);
4922 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4924 unsigned long precharge;
4926 struct mm_walk mem_cgroup_count_precharge_walk = {
4927 .pmd_entry = mem_cgroup_count_precharge_pte_range,
4930 down_read(&mm->mmap_sem);
4931 walk_page_range(0, ~0UL, &mem_cgroup_count_precharge_walk);
4932 up_read(&mm->mmap_sem);
4934 precharge = mc.precharge;
4940 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4942 unsigned long precharge = mem_cgroup_count_precharge(mm);
4944 VM_BUG_ON(mc.moving_task);
4945 mc.moving_task = current;
4946 return mem_cgroup_do_precharge(precharge);
4949 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4950 static void __mem_cgroup_clear_mc(void)
4952 struct mem_cgroup *from = mc.from;
4953 struct mem_cgroup *to = mc.to;
4955 /* we must uncharge all the leftover precharges from mc.to */
4957 cancel_charge(mc.to, mc.precharge);
4961 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4962 * we must uncharge here.
4964 if (mc.moved_charge) {
4965 cancel_charge(mc.from, mc.moved_charge);
4966 mc.moved_charge = 0;
4968 /* we must fixup refcnts and charges */
4969 if (mc.moved_swap) {
4970 /* uncharge swap account from the old cgroup */
4971 if (!mem_cgroup_is_root(mc.from))
4972 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
4975 * we charged both to->memory and to->memsw, so we
4976 * should uncharge to->memory.
4978 if (!mem_cgroup_is_root(mc.to))
4979 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
4981 css_put_many(&mc.from->css, mc.moved_swap);
4983 /* we've already done css_get(mc.to) */
4986 memcg_oom_recover(from);
4987 memcg_oom_recover(to);
4988 wake_up_all(&mc.waitq);
4991 static void mem_cgroup_clear_mc(void)
4994 * we must clear moving_task before waking up waiters at the end of
4997 mc.moving_task = NULL;
4998 __mem_cgroup_clear_mc();
4999 spin_lock(&mc.lock);
5002 spin_unlock(&mc.lock);
5005 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
5006 struct cgroup_taskset *tset)
5008 struct task_struct *p = cgroup_taskset_first(tset);
5010 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5011 unsigned long move_flags;
5014 * We are now commited to this value whatever it is. Changes in this
5015 * tunable will only affect upcoming migrations, not the current one.
5016 * So we need to save it, and keep it going.
5018 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5020 struct mm_struct *mm;
5021 struct mem_cgroup *from = mem_cgroup_from_task(p);
5023 VM_BUG_ON(from == memcg);
5025 mm = get_task_mm(p);
5028 /* We move charges only when we move a owner of the mm */
5029 if (mm->owner == p) {
5032 VM_BUG_ON(mc.precharge);
5033 VM_BUG_ON(mc.moved_charge);
5034 VM_BUG_ON(mc.moved_swap);
5036 spin_lock(&mc.lock);
5039 mc.flags = move_flags;
5040 spin_unlock(&mc.lock);
5041 /* We set mc.moving_task later */
5043 ret = mem_cgroup_precharge_mc(mm);
5045 mem_cgroup_clear_mc();
5052 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
5053 struct cgroup_taskset *tset)
5056 mem_cgroup_clear_mc();
5059 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5060 unsigned long addr, unsigned long end,
5061 struct mm_walk *walk)
5064 struct vm_area_struct *vma = walk->vma;
5067 enum mc_target_type target_type;
5068 union mc_target target;
5072 * We don't take compound_lock() here but no race with splitting thp
5074 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
5075 * under splitting, which means there's no concurrent thp split,
5076 * - if another thread runs into split_huge_page() just after we
5077 * entered this if-block, the thread must wait for page table lock
5078 * to be unlocked in __split_huge_page_splitting(), where the main
5079 * part of thp split is not executed yet.
5081 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
5082 if (mc.precharge < HPAGE_PMD_NR) {
5086 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5087 if (target_type == MC_TARGET_PAGE) {
5089 if (!isolate_lru_page(page)) {
5090 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
5092 mc.precharge -= HPAGE_PMD_NR;
5093 mc.moved_charge += HPAGE_PMD_NR;
5095 putback_lru_page(page);
5103 if (pmd_trans_unstable(pmd))
5106 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5107 for (; addr != end; addr += PAGE_SIZE) {
5108 pte_t ptent = *(pte++);
5114 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5115 case MC_TARGET_PAGE:
5117 if (isolate_lru_page(page))
5119 if (!mem_cgroup_move_account(page, 1, mc.from, mc.to)) {
5121 /* we uncharge from mc.from later. */
5124 putback_lru_page(page);
5125 put: /* get_mctgt_type() gets the page */
5128 case MC_TARGET_SWAP:
5130 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5132 /* we fixup refcnts and charges later. */
5140 pte_unmap_unlock(pte - 1, ptl);
5145 * We have consumed all precharges we got in can_attach().
5146 * We try charge one by one, but don't do any additional
5147 * charges to mc.to if we have failed in charge once in attach()
5150 ret = mem_cgroup_do_precharge(1);
5158 static void mem_cgroup_move_charge(struct mm_struct *mm)
5160 struct mm_walk mem_cgroup_move_charge_walk = {
5161 .pmd_entry = mem_cgroup_move_charge_pte_range,
5165 lru_add_drain_all();
5167 * Signal mem_cgroup_begin_page_stat() to take the memcg's
5168 * move_lock while we're moving its pages to another memcg.
5169 * Then wait for already started RCU-only updates to finish.
5171 atomic_inc(&mc.from->moving_account);
5174 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5176 * Someone who are holding the mmap_sem might be waiting in
5177 * waitq. So we cancel all extra charges, wake up all waiters,
5178 * and retry. Because we cancel precharges, we might not be able
5179 * to move enough charges, but moving charge is a best-effort
5180 * feature anyway, so it wouldn't be a big problem.
5182 __mem_cgroup_clear_mc();
5187 * When we have consumed all precharges and failed in doing
5188 * additional charge, the page walk just aborts.
5190 walk_page_range(0, ~0UL, &mem_cgroup_move_charge_walk);
5191 up_read(&mm->mmap_sem);
5192 atomic_dec(&mc.from->moving_account);
5195 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
5196 struct cgroup_taskset *tset)
5198 struct task_struct *p = cgroup_taskset_first(tset);
5199 struct mm_struct *mm = get_task_mm(p);
5203 mem_cgroup_move_charge(mm);
5207 mem_cgroup_clear_mc();
5209 #else /* !CONFIG_MMU */
5210 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
5211 struct cgroup_taskset *tset)
5215 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
5216 struct cgroup_taskset *tset)
5219 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
5220 struct cgroup_taskset *tset)
5226 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5227 * to verify whether we're attached to the default hierarchy on each mount
5230 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5233 * use_hierarchy is forced on the default hierarchy. cgroup core
5234 * guarantees that @root doesn't have any children, so turning it
5235 * on for the root memcg is enough.
5237 if (cgroup_on_dfl(root_css->cgroup))
5238 root_mem_cgroup->use_hierarchy = true;
5240 root_mem_cgroup->use_hierarchy = false;
5243 static u64 memory_current_read(struct cgroup_subsys_state *css,
5246 return mem_cgroup_usage(mem_cgroup_from_css(css), false);
5249 static int memory_low_show(struct seq_file *m, void *v)
5251 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5252 unsigned long low = READ_ONCE(memcg->low);
5254 if (low == PAGE_COUNTER_MAX)
5255 seq_puts(m, "max\n");
5257 seq_printf(m, "%llu\n", (u64)low * PAGE_SIZE);
5262 static ssize_t memory_low_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, "max", &low);
5279 static int memory_high_show(struct seq_file *m, void *v)
5281 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5282 unsigned long high = READ_ONCE(memcg->high);
5284 if (high == PAGE_COUNTER_MAX)
5285 seq_puts(m, "max\n");
5287 seq_printf(m, "%llu\n", (u64)high * PAGE_SIZE);
5292 static ssize_t memory_high_write(struct kernfs_open_file *of,
5293 char *buf, size_t nbytes, loff_t off)
5295 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5299 buf = strstrip(buf);
5300 err = page_counter_memparse(buf, "max", &high);
5309 static int memory_max_show(struct seq_file *m, void *v)
5311 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5312 unsigned long max = READ_ONCE(memcg->memory.limit);
5314 if (max == PAGE_COUNTER_MAX)
5315 seq_puts(m, "max\n");
5317 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
5322 static ssize_t memory_max_write(struct kernfs_open_file *of,
5323 char *buf, size_t nbytes, loff_t off)
5325 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5329 buf = strstrip(buf);
5330 err = page_counter_memparse(buf, "max", &max);
5334 err = mem_cgroup_resize_limit(memcg, max);
5341 static int memory_events_show(struct seq_file *m, void *v)
5343 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5345 seq_printf(m, "low %lu\n", mem_cgroup_read_events(memcg, MEMCG_LOW));
5346 seq_printf(m, "high %lu\n", mem_cgroup_read_events(memcg, MEMCG_HIGH));
5347 seq_printf(m, "max %lu\n", mem_cgroup_read_events(memcg, MEMCG_MAX));
5348 seq_printf(m, "oom %lu\n", mem_cgroup_read_events(memcg, MEMCG_OOM));
5353 static struct cftype memory_files[] = {
5356 .read_u64 = memory_current_read,
5360 .flags = CFTYPE_NOT_ON_ROOT,
5361 .seq_show = memory_low_show,
5362 .write = memory_low_write,
5366 .flags = CFTYPE_NOT_ON_ROOT,
5367 .seq_show = memory_high_show,
5368 .write = memory_high_write,
5372 .flags = CFTYPE_NOT_ON_ROOT,
5373 .seq_show = memory_max_show,
5374 .write = memory_max_write,
5378 .flags = CFTYPE_NOT_ON_ROOT,
5379 .seq_show = memory_events_show,
5384 struct cgroup_subsys memory_cgrp_subsys = {
5385 .css_alloc = mem_cgroup_css_alloc,
5386 .css_online = mem_cgroup_css_online,
5387 .css_offline = mem_cgroup_css_offline,
5388 .css_free = mem_cgroup_css_free,
5389 .css_reset = mem_cgroup_css_reset,
5390 .can_attach = mem_cgroup_can_attach,
5391 .cancel_attach = mem_cgroup_cancel_attach,
5392 .attach = mem_cgroup_move_task,
5393 .bind = mem_cgroup_bind,
5394 .dfl_cftypes = memory_files,
5395 .legacy_cftypes = mem_cgroup_legacy_files,
5400 * mem_cgroup_events - count memory events against a cgroup
5401 * @memcg: the memory cgroup
5402 * @idx: the event index
5403 * @nr: the number of events to account for
5405 void mem_cgroup_events(struct mem_cgroup *memcg,
5406 enum mem_cgroup_events_index idx,
5409 this_cpu_add(memcg->stat->events[idx], nr);
5413 * mem_cgroup_low - check if memory consumption is below the normal range
5414 * @root: the highest ancestor to consider
5415 * @memcg: the memory cgroup to check
5417 * Returns %true if memory consumption of @memcg, and that of all
5418 * configurable ancestors up to @root, is below the normal range.
5420 bool mem_cgroup_low(struct mem_cgroup *root, struct mem_cgroup *memcg)
5422 if (mem_cgroup_disabled())
5426 * The toplevel group doesn't have a configurable range, so
5427 * it's never low when looked at directly, and it is not
5428 * considered an ancestor when assessing the hierarchy.
5431 if (memcg == root_mem_cgroup)
5434 if (page_counter_read(&memcg->memory) >= memcg->low)
5437 while (memcg != root) {
5438 memcg = parent_mem_cgroup(memcg);
5440 if (memcg == root_mem_cgroup)
5443 if (page_counter_read(&memcg->memory) >= memcg->low)
5450 * mem_cgroup_try_charge - try charging a page
5451 * @page: page to charge
5452 * @mm: mm context of the victim
5453 * @gfp_mask: reclaim mode
5454 * @memcgp: charged memcg return
5456 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5457 * pages according to @gfp_mask if necessary.
5459 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5460 * Otherwise, an error code is returned.
5462 * After page->mapping has been set up, the caller must finalize the
5463 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5464 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5466 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
5467 gfp_t gfp_mask, struct mem_cgroup **memcgp)
5469 struct mem_cgroup *memcg = NULL;
5470 unsigned int nr_pages = 1;
5473 if (mem_cgroup_disabled())
5476 if (PageSwapCache(page)) {
5478 * Every swap fault against a single page tries to charge the
5479 * page, bail as early as possible. shmem_unuse() encounters
5480 * already charged pages, too. The USED bit is protected by
5481 * the page lock, which serializes swap cache removal, which
5482 * in turn serializes uncharging.
5484 if (page->mem_cgroup)
5488 if (PageTransHuge(page)) {
5489 nr_pages <<= compound_order(page);
5490 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5493 if (do_swap_account && PageSwapCache(page))
5494 memcg = try_get_mem_cgroup_from_page(page);
5496 memcg = get_mem_cgroup_from_mm(mm);
5498 ret = try_charge(memcg, gfp_mask, nr_pages);
5500 css_put(&memcg->css);
5502 if (ret == -EINTR) {
5503 memcg = root_mem_cgroup;
5512 * mem_cgroup_commit_charge - commit a page charge
5513 * @page: page to charge
5514 * @memcg: memcg to charge the page to
5515 * @lrucare: page might be on LRU already
5517 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5518 * after page->mapping has been set up. This must happen atomically
5519 * as part of the page instantiation, i.e. under the page table lock
5520 * for anonymous pages, under the page lock for page and swap cache.
5522 * In addition, the page must not be on the LRU during the commit, to
5523 * prevent racing with task migration. If it might be, use @lrucare.
5525 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5527 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
5530 unsigned int nr_pages = 1;
5532 VM_BUG_ON_PAGE(!page->mapping, page);
5533 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
5535 if (mem_cgroup_disabled())
5538 * Swap faults will attempt to charge the same page multiple
5539 * times. But reuse_swap_page() might have removed the page
5540 * from swapcache already, so we can't check PageSwapCache().
5545 commit_charge(page, memcg, lrucare);
5547 if (PageTransHuge(page)) {
5548 nr_pages <<= compound_order(page);
5549 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5552 local_lock_irq(event_lock);
5553 mem_cgroup_charge_statistics(memcg, page, nr_pages);
5554 memcg_check_events(memcg, page);
5555 local_unlock_irq(event_lock);
5557 if (do_swap_account && PageSwapCache(page)) {
5558 swp_entry_t entry = { .val = page_private(page) };
5560 * The swap entry might not get freed for a long time,
5561 * let's not wait for it. The page already received a
5562 * memory+swap charge, drop the swap entry duplicate.
5564 mem_cgroup_uncharge_swap(entry);
5569 * mem_cgroup_cancel_charge - cancel a page charge
5570 * @page: page to charge
5571 * @memcg: memcg to charge the page to
5573 * Cancel a charge transaction started by mem_cgroup_try_charge().
5575 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg)
5577 unsigned int nr_pages = 1;
5579 if (mem_cgroup_disabled())
5582 * Swap faults will attempt to charge the same page multiple
5583 * times. But reuse_swap_page() might have removed the page
5584 * from swapcache already, so we can't check PageSwapCache().
5589 if (PageTransHuge(page)) {
5590 nr_pages <<= compound_order(page);
5591 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5594 cancel_charge(memcg, nr_pages);
5597 static void uncharge_batch(struct mem_cgroup *memcg, unsigned long pgpgout,
5598 unsigned long nr_anon, unsigned long nr_file,
5599 unsigned long nr_huge, struct page *dummy_page)
5601 unsigned long nr_pages = nr_anon + nr_file;
5602 unsigned long flags;
5604 if (!mem_cgroup_is_root(memcg)) {
5605 page_counter_uncharge(&memcg->memory, nr_pages);
5606 if (do_swap_account)
5607 page_counter_uncharge(&memcg->memsw, nr_pages);
5608 memcg_oom_recover(memcg);
5611 local_lock_irqsave(event_lock, flags);
5612 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS], nr_anon);
5613 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_CACHE], nr_file);
5614 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE], nr_huge);
5615 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT], pgpgout);
5616 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
5617 memcg_check_events(memcg, dummy_page);
5618 local_unlock_irqrestore(event_lock, flags);
5620 if (!mem_cgroup_is_root(memcg))
5621 css_put_many(&memcg->css, nr_pages);
5624 static void uncharge_list(struct list_head *page_list)
5626 struct mem_cgroup *memcg = NULL;
5627 unsigned long nr_anon = 0;
5628 unsigned long nr_file = 0;
5629 unsigned long nr_huge = 0;
5630 unsigned long pgpgout = 0;
5631 struct list_head *next;
5634 next = page_list->next;
5636 unsigned int nr_pages = 1;
5638 page = list_entry(next, struct page, lru);
5639 next = page->lru.next;
5641 VM_BUG_ON_PAGE(PageLRU(page), page);
5642 VM_BUG_ON_PAGE(page_count(page), page);
5644 if (!page->mem_cgroup)
5648 * Nobody should be changing or seriously looking at
5649 * page->mem_cgroup at this point, we have fully
5650 * exclusive access to the page.
5653 if (memcg != page->mem_cgroup) {
5655 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5657 pgpgout = nr_anon = nr_file = nr_huge = 0;
5659 memcg = page->mem_cgroup;
5662 if (PageTransHuge(page)) {
5663 nr_pages <<= compound_order(page);
5664 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5665 nr_huge += nr_pages;
5669 nr_anon += nr_pages;
5671 nr_file += nr_pages;
5673 page->mem_cgroup = NULL;
5676 } while (next != page_list);
5679 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5684 * mem_cgroup_uncharge - uncharge a page
5685 * @page: page to uncharge
5687 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5688 * mem_cgroup_commit_charge().
5690 void mem_cgroup_uncharge(struct page *page)
5692 if (mem_cgroup_disabled())
5695 /* Don't touch page->lru of any random page, pre-check: */
5696 if (!page->mem_cgroup)
5699 INIT_LIST_HEAD(&page->lru);
5700 uncharge_list(&page->lru);
5704 * mem_cgroup_uncharge_list - uncharge a list of page
5705 * @page_list: list of pages to uncharge
5707 * Uncharge a list of pages previously charged with
5708 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5710 void mem_cgroup_uncharge_list(struct list_head *page_list)
5712 if (mem_cgroup_disabled())
5715 if (!list_empty(page_list))
5716 uncharge_list(page_list);
5720 * mem_cgroup_migrate - migrate a charge to another page
5721 * @oldpage: currently charged page
5722 * @newpage: page to transfer the charge to
5723 * @lrucare: either or both pages might be on the LRU already
5725 * Migrate the charge from @oldpage to @newpage.
5727 * Both pages must be locked, @newpage->mapping must be set up.
5729 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage,
5732 struct mem_cgroup *memcg;
5735 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
5736 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
5737 VM_BUG_ON_PAGE(!lrucare && PageLRU(oldpage), oldpage);
5738 VM_BUG_ON_PAGE(!lrucare && PageLRU(newpage), newpage);
5739 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
5740 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
5743 if (mem_cgroup_disabled())
5746 /* Page cache replacement: new page already charged? */
5747 if (newpage->mem_cgroup)
5751 * Swapcache readahead pages can get migrated before being
5752 * charged, and migration from compaction can happen to an
5753 * uncharged page when the PFN walker finds a page that
5754 * reclaim just put back on the LRU but has not released yet.
5756 memcg = oldpage->mem_cgroup;
5761 lock_page_lru(oldpage, &isolated);
5763 oldpage->mem_cgroup = NULL;
5766 unlock_page_lru(oldpage, isolated);
5768 commit_charge(newpage, memcg, lrucare);
5772 * subsys_initcall() for memory controller.
5774 * Some parts like hotcpu_notifier() have to be initialized from this context
5775 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5776 * everything that doesn't depend on a specific mem_cgroup structure should
5777 * be initialized from here.
5779 static int __init mem_cgroup_init(void)
5783 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5785 for_each_possible_cpu(cpu)
5786 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
5789 for_each_node(node) {
5790 struct mem_cgroup_tree_per_node *rtpn;
5793 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
5794 node_online(node) ? node : NUMA_NO_NODE);
5796 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5797 struct mem_cgroup_tree_per_zone *rtpz;
5799 rtpz = &rtpn->rb_tree_per_zone[zone];
5800 rtpz->rb_root = RB_ROOT;
5801 spin_lock_init(&rtpz->lock);
5803 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5808 subsys_initcall(mem_cgroup_init);
5810 #ifdef CONFIG_MEMCG_SWAP
5812 * mem_cgroup_swapout - transfer a memsw charge to swap
5813 * @page: page whose memsw charge to transfer
5814 * @entry: swap entry to move the charge to
5816 * Transfer the memsw charge of @page to @entry.
5818 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
5820 struct mem_cgroup *memcg;
5821 unsigned short oldid;
5822 unsigned long flags;
5824 VM_BUG_ON_PAGE(PageLRU(page), page);
5825 VM_BUG_ON_PAGE(page_count(page), page);
5827 if (!do_swap_account)
5830 memcg = page->mem_cgroup;
5832 /* Readahead page, never charged */
5836 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg));
5837 VM_BUG_ON_PAGE(oldid, page);
5838 mem_cgroup_swap_statistics(memcg, true);
5840 page->mem_cgroup = NULL;
5842 if (!mem_cgroup_is_root(memcg))
5843 page_counter_uncharge(&memcg->memory, 1);
5845 local_lock_irqsave(event_lock, flags);
5846 /* Caller disabled preemption with mapping->tree_lock */
5847 mem_cgroup_charge_statistics(memcg, page, -1);
5848 memcg_check_events(memcg, page);
5849 local_unlock_irqrestore(event_lock, flags);
5853 * mem_cgroup_uncharge_swap - uncharge a swap entry
5854 * @entry: swap entry to uncharge
5856 * Drop the memsw charge associated with @entry.
5858 void mem_cgroup_uncharge_swap(swp_entry_t entry)
5860 struct mem_cgroup *memcg;
5863 if (!do_swap_account)
5866 id = swap_cgroup_record(entry, 0);
5868 memcg = mem_cgroup_from_id(id);
5870 if (!mem_cgroup_is_root(memcg))
5871 page_counter_uncharge(&memcg->memsw, 1);
5872 mem_cgroup_swap_statistics(memcg, false);
5873 css_put(&memcg->css);
5878 /* for remember boot option*/
5879 #ifdef CONFIG_MEMCG_SWAP_ENABLED
5880 static int really_do_swap_account __initdata = 1;
5882 static int really_do_swap_account __initdata;
5885 static int __init enable_swap_account(char *s)
5887 if (!strcmp(s, "1"))
5888 really_do_swap_account = 1;
5889 else if (!strcmp(s, "0"))
5890 really_do_swap_account = 0;
5893 __setup("swapaccount=", enable_swap_account);
5895 static struct cftype memsw_cgroup_files[] = {
5897 .name = "memsw.usage_in_bytes",
5898 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
5899 .read_u64 = mem_cgroup_read_u64,
5902 .name = "memsw.max_usage_in_bytes",
5903 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
5904 .write = mem_cgroup_reset,
5905 .read_u64 = mem_cgroup_read_u64,
5908 .name = "memsw.limit_in_bytes",
5909 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
5910 .write = mem_cgroup_write,
5911 .read_u64 = mem_cgroup_read_u64,
5914 .name = "memsw.failcnt",
5915 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
5916 .write = mem_cgroup_reset,
5917 .read_u64 = mem_cgroup_read_u64,
5919 { }, /* terminate */
5922 static int __init mem_cgroup_swap_init(void)
5924 if (!mem_cgroup_disabled() && really_do_swap_account) {
5925 do_swap_account = 1;
5926 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
5927 memsw_cgroup_files));
5931 subsys_initcall(mem_cgroup_swap_init);
5933 #endif /* CONFIG_MEMCG_SWAP */