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>
71 #include <asm/uaccess.h>
73 #include <trace/events/vmscan.h>
75 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
76 EXPORT_SYMBOL(memory_cgrp_subsys);
78 #define MEM_CGROUP_RECLAIM_RETRIES 5
79 static struct mem_cgroup *root_mem_cgroup __read_mostly;
81 /* Whether the swap controller is active */
82 #ifdef CONFIG_MEMCG_SWAP
83 int do_swap_account __read_mostly;
85 #define do_swap_account 0
88 static const char * const mem_cgroup_stat_names[] = {
97 static const char * const mem_cgroup_events_names[] = {
104 static const char * const mem_cgroup_lru_names[] = {
113 * Per memcg event counter is incremented at every pagein/pageout. With THP,
114 * it will be incremated by the number of pages. This counter is used for
115 * for trigger some periodic events. This is straightforward and better
116 * than using jiffies etc. to handle periodic memcg event.
118 enum mem_cgroup_events_target {
119 MEM_CGROUP_TARGET_THRESH,
120 MEM_CGROUP_TARGET_SOFTLIMIT,
121 MEM_CGROUP_TARGET_NUMAINFO,
124 #define THRESHOLDS_EVENTS_TARGET 128
125 #define SOFTLIMIT_EVENTS_TARGET 1024
126 #define NUMAINFO_EVENTS_TARGET 1024
128 struct mem_cgroup_stat_cpu {
129 long count[MEM_CGROUP_STAT_NSTATS];
130 unsigned long events[MEMCG_NR_EVENTS];
131 unsigned long nr_page_events;
132 unsigned long targets[MEM_CGROUP_NTARGETS];
135 struct reclaim_iter {
136 struct mem_cgroup *position;
137 /* scan generation, increased every round-trip */
138 unsigned int generation;
142 * per-zone information in memory controller.
144 struct mem_cgroup_per_zone {
145 struct lruvec lruvec;
146 unsigned long lru_size[NR_LRU_LISTS];
148 struct reclaim_iter iter[DEF_PRIORITY + 1];
150 struct rb_node tree_node; /* RB tree node */
151 unsigned long usage_in_excess;/* Set to the value by which */
152 /* the soft limit is exceeded*/
154 struct mem_cgroup *memcg; /* Back pointer, we cannot */
155 /* use container_of */
158 struct mem_cgroup_per_node {
159 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
163 * Cgroups above their limits are maintained in a RB-Tree, independent of
164 * their hierarchy representation
167 struct mem_cgroup_tree_per_zone {
168 struct rb_root rb_root;
172 struct mem_cgroup_tree_per_node {
173 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
176 struct mem_cgroup_tree {
177 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
180 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
182 struct mem_cgroup_threshold {
183 struct eventfd_ctx *eventfd;
184 unsigned long threshold;
188 struct mem_cgroup_threshold_ary {
189 /* An array index points to threshold just below or equal to usage. */
190 int current_threshold;
191 /* Size of entries[] */
193 /* Array of thresholds */
194 struct mem_cgroup_threshold entries[0];
197 struct mem_cgroup_thresholds {
198 /* Primary thresholds array */
199 struct mem_cgroup_threshold_ary *primary;
201 * Spare threshold array.
202 * This is needed to make mem_cgroup_unregister_event() "never fail".
203 * It must be able to store at least primary->size - 1 entries.
205 struct mem_cgroup_threshold_ary *spare;
209 struct mem_cgroup_eventfd_list {
210 struct list_head list;
211 struct eventfd_ctx *eventfd;
215 * cgroup_event represents events which userspace want to receive.
217 struct mem_cgroup_event {
219 * memcg which the event belongs to.
221 struct mem_cgroup *memcg;
223 * eventfd to signal userspace about the event.
225 struct eventfd_ctx *eventfd;
227 * Each of these stored in a list by the cgroup.
229 struct list_head list;
231 * register_event() callback will be used to add new userspace
232 * waiter for changes related to this event. Use eventfd_signal()
233 * on eventfd to send notification to userspace.
235 int (*register_event)(struct mem_cgroup *memcg,
236 struct eventfd_ctx *eventfd, const char *args);
238 * unregister_event() callback will be called when userspace closes
239 * the eventfd or on cgroup removing. This callback must be set,
240 * if you want provide notification functionality.
242 void (*unregister_event)(struct mem_cgroup *memcg,
243 struct eventfd_ctx *eventfd);
245 * All fields below needed to unregister event when
246 * userspace closes eventfd.
249 wait_queue_head_t *wqh;
251 struct work_struct remove;
254 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
255 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
258 * The memory controller data structure. The memory controller controls both
259 * page cache and RSS per cgroup. We would eventually like to provide
260 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
261 * to help the administrator determine what knobs to tune.
263 * TODO: Add a water mark for the memory controller. Reclaim will begin when
264 * we hit the water mark. May be even add a low water mark, such that
265 * no reclaim occurs from a cgroup at it's low water mark, this is
266 * a feature that will be implemented much later in the future.
269 struct cgroup_subsys_state css;
271 /* Accounted resources */
272 struct page_counter memory;
273 struct page_counter memsw;
274 struct page_counter kmem;
276 /* Normal memory consumption range */
280 unsigned long soft_limit;
282 /* vmpressure notifications */
283 struct vmpressure vmpressure;
285 /* css_online() has been completed */
289 * Should the accounting and control be hierarchical, per subtree?
295 atomic_t oom_wakeups;
298 /* OOM-Killer disable */
299 int oom_kill_disable;
301 /* protect arrays of thresholds */
302 struct mutex thresholds_lock;
304 /* thresholds for memory usage. RCU-protected */
305 struct mem_cgroup_thresholds thresholds;
307 /* thresholds for mem+swap usage. RCU-protected */
308 struct mem_cgroup_thresholds memsw_thresholds;
310 /* For oom notifier event fd */
311 struct list_head oom_notify;
314 * Should we move charges of a task when a task is moved into this
315 * mem_cgroup ? And what type of charges should we move ?
317 unsigned long move_charge_at_immigrate;
319 * set > 0 if pages under this cgroup are moving to other cgroup.
321 atomic_t moving_account;
322 /* taken only while moving_account > 0 */
323 spinlock_t move_lock;
324 struct task_struct *move_lock_task;
325 unsigned long move_lock_flags;
329 struct mem_cgroup_stat_cpu __percpu *stat;
331 * used when a cpu is offlined or other synchronizations
332 * See mem_cgroup_read_stat().
334 struct mem_cgroup_stat_cpu nocpu_base;
335 spinlock_t pcp_counter_lock;
337 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
338 struct cg_proto tcp_mem;
340 #if defined(CONFIG_MEMCG_KMEM)
341 /* Index in the kmem_cache->memcg_params.memcg_caches array */
343 bool kmem_acct_activated;
344 bool kmem_acct_active;
347 int last_scanned_node;
349 nodemask_t scan_nodes;
350 atomic_t numainfo_events;
351 atomic_t numainfo_updating;
354 /* List of events which userspace want to receive */
355 struct list_head event_list;
356 spinlock_t event_list_lock;
358 struct mem_cgroup_per_node *nodeinfo[0];
359 /* WARNING: nodeinfo must be the last member here */
362 #ifdef CONFIG_MEMCG_KMEM
363 bool memcg_kmem_is_active(struct mem_cgroup *memcg)
365 return memcg->kmem_acct_active;
369 /* Stuffs for move charges at task migration. */
371 * Types of charges to be moved.
373 #define MOVE_ANON 0x1U
374 #define MOVE_FILE 0x2U
375 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
377 /* "mc" and its members are protected by cgroup_mutex */
378 static struct move_charge_struct {
379 spinlock_t lock; /* for from, to */
380 struct mem_cgroup *from;
381 struct mem_cgroup *to;
383 unsigned long precharge;
384 unsigned long moved_charge;
385 unsigned long moved_swap;
386 struct task_struct *moving_task; /* a task moving charges */
387 wait_queue_head_t waitq; /* a waitq for other context */
389 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
390 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
394 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
395 * limit reclaim to prevent infinite loops, if they ever occur.
397 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
398 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
401 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
402 MEM_CGROUP_CHARGE_TYPE_ANON,
403 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
404 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
408 /* for encoding cft->private value on file */
416 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
417 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
418 #define MEMFILE_ATTR(val) ((val) & 0xffff)
419 /* Used for OOM nofiier */
420 #define OOM_CONTROL (0)
423 * The memcg_create_mutex will be held whenever a new cgroup is created.
424 * As a consequence, any change that needs to protect against new child cgroups
425 * appearing has to hold it as well.
427 static DEFINE_MUTEX(memcg_create_mutex);
429 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
431 return s ? container_of(s, struct mem_cgroup, css) : NULL;
434 /* Some nice accessors for the vmpressure. */
435 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
438 memcg = root_mem_cgroup;
439 return &memcg->vmpressure;
442 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
444 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
447 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
449 return (memcg == root_mem_cgroup);
453 * We restrict the id in the range of [1, 65535], so it can fit into
456 #define MEM_CGROUP_ID_MAX USHRT_MAX
458 static inline unsigned short mem_cgroup_id(struct mem_cgroup *memcg)
460 return memcg->css.id;
464 * A helper function to get mem_cgroup from ID. must be called under
465 * rcu_read_lock(). The caller is responsible for calling
466 * css_tryget_online() if the mem_cgroup is used for charging. (dropping
467 * refcnt from swap can be called against removed memcg.)
469 static inline struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
471 struct cgroup_subsys_state *css;
473 css = css_from_id(id, &memory_cgrp_subsys);
474 return mem_cgroup_from_css(css);
477 /* Writing them here to avoid exposing memcg's inner layout */
478 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
480 void sock_update_memcg(struct sock *sk)
482 if (mem_cgroup_sockets_enabled) {
483 struct mem_cgroup *memcg;
484 struct cg_proto *cg_proto;
486 BUG_ON(!sk->sk_prot->proto_cgroup);
488 /* Socket cloning can throw us here with sk_cgrp already
489 * filled. It won't however, necessarily happen from
490 * process context. So the test for root memcg given
491 * the current task's memcg won't help us in this case.
493 * Respecting the original socket's memcg is a better
494 * decision in this case.
497 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
498 css_get(&sk->sk_cgrp->memcg->css);
503 memcg = mem_cgroup_from_task(current);
504 cg_proto = sk->sk_prot->proto_cgroup(memcg);
505 if (!mem_cgroup_is_root(memcg) &&
506 memcg_proto_active(cg_proto) &&
507 css_tryget_online(&memcg->css)) {
508 sk->sk_cgrp = cg_proto;
513 EXPORT_SYMBOL(sock_update_memcg);
515 void sock_release_memcg(struct sock *sk)
517 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
518 struct mem_cgroup *memcg;
519 WARN_ON(!sk->sk_cgrp->memcg);
520 memcg = sk->sk_cgrp->memcg;
521 css_put(&sk->sk_cgrp->memcg->css);
525 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
527 if (!memcg || mem_cgroup_is_root(memcg))
530 return &memcg->tcp_mem;
532 EXPORT_SYMBOL(tcp_proto_cgroup);
536 #ifdef CONFIG_MEMCG_KMEM
538 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
539 * The main reason for not using cgroup id for this:
540 * this works better in sparse environments, where we have a lot of memcgs,
541 * but only a few kmem-limited. Or also, if we have, for instance, 200
542 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
543 * 200 entry array for that.
545 * The current size of the caches array is stored in memcg_nr_cache_ids. It
546 * will double each time we have to increase it.
548 static DEFINE_IDA(memcg_cache_ida);
549 int memcg_nr_cache_ids;
551 /* Protects memcg_nr_cache_ids */
552 static DECLARE_RWSEM(memcg_cache_ids_sem);
554 void memcg_get_cache_ids(void)
556 down_read(&memcg_cache_ids_sem);
559 void memcg_put_cache_ids(void)
561 up_read(&memcg_cache_ids_sem);
565 * MIN_SIZE is different than 1, because we would like to avoid going through
566 * the alloc/free process all the time. In a small machine, 4 kmem-limited
567 * cgroups is a reasonable guess. In the future, it could be a parameter or
568 * tunable, but that is strictly not necessary.
570 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
571 * this constant directly from cgroup, but it is understandable that this is
572 * better kept as an internal representation in cgroup.c. In any case, the
573 * cgrp_id space is not getting any smaller, and we don't have to necessarily
574 * increase ours as well if it increases.
576 #define MEMCG_CACHES_MIN_SIZE 4
577 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
580 * A lot of the calls to the cache allocation functions are expected to be
581 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
582 * conditional to this static branch, we'll have to allow modules that does
583 * kmem_cache_alloc and the such to see this symbol as well
585 struct static_key memcg_kmem_enabled_key;
586 EXPORT_SYMBOL(memcg_kmem_enabled_key);
588 #endif /* CONFIG_MEMCG_KMEM */
590 static struct mem_cgroup_per_zone *
591 mem_cgroup_zone_zoneinfo(struct mem_cgroup *memcg, struct zone *zone)
593 int nid = zone_to_nid(zone);
594 int zid = zone_idx(zone);
596 return &memcg->nodeinfo[nid]->zoneinfo[zid];
599 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
604 static struct mem_cgroup_per_zone *
605 mem_cgroup_page_zoneinfo(struct mem_cgroup *memcg, struct page *page)
607 int nid = page_to_nid(page);
608 int zid = page_zonenum(page);
610 return &memcg->nodeinfo[nid]->zoneinfo[zid];
613 static struct mem_cgroup_tree_per_zone *
614 soft_limit_tree_node_zone(int nid, int zid)
616 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
619 static struct mem_cgroup_tree_per_zone *
620 soft_limit_tree_from_page(struct page *page)
622 int nid = page_to_nid(page);
623 int zid = page_zonenum(page);
625 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
628 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone *mz,
629 struct mem_cgroup_tree_per_zone *mctz,
630 unsigned long new_usage_in_excess)
632 struct rb_node **p = &mctz->rb_root.rb_node;
633 struct rb_node *parent = NULL;
634 struct mem_cgroup_per_zone *mz_node;
639 mz->usage_in_excess = new_usage_in_excess;
640 if (!mz->usage_in_excess)
644 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
646 if (mz->usage_in_excess < mz_node->usage_in_excess)
649 * We can't avoid mem cgroups that are over their soft
650 * limit by the same amount
652 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
655 rb_link_node(&mz->tree_node, parent, p);
656 rb_insert_color(&mz->tree_node, &mctz->rb_root);
660 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
661 struct mem_cgroup_tree_per_zone *mctz)
665 rb_erase(&mz->tree_node, &mctz->rb_root);
669 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
670 struct mem_cgroup_tree_per_zone *mctz)
674 spin_lock_irqsave(&mctz->lock, flags);
675 __mem_cgroup_remove_exceeded(mz, mctz);
676 spin_unlock_irqrestore(&mctz->lock, flags);
679 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
681 unsigned long nr_pages = page_counter_read(&memcg->memory);
682 unsigned long soft_limit = ACCESS_ONCE(memcg->soft_limit);
683 unsigned long excess = 0;
685 if (nr_pages > soft_limit)
686 excess = nr_pages - soft_limit;
691 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
693 unsigned long excess;
694 struct mem_cgroup_per_zone *mz;
695 struct mem_cgroup_tree_per_zone *mctz;
697 mctz = soft_limit_tree_from_page(page);
699 * Necessary to update all ancestors when hierarchy is used.
700 * because their event counter is not touched.
702 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
703 mz = mem_cgroup_page_zoneinfo(memcg, page);
704 excess = soft_limit_excess(memcg);
706 * We have to update the tree if mz is on RB-tree or
707 * mem is over its softlimit.
709 if (excess || mz->on_tree) {
712 spin_lock_irqsave(&mctz->lock, flags);
713 /* if on-tree, remove it */
715 __mem_cgroup_remove_exceeded(mz, mctz);
717 * Insert again. mz->usage_in_excess will be updated.
718 * If excess is 0, no tree ops.
720 __mem_cgroup_insert_exceeded(mz, mctz, excess);
721 spin_unlock_irqrestore(&mctz->lock, flags);
726 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
728 struct mem_cgroup_tree_per_zone *mctz;
729 struct mem_cgroup_per_zone *mz;
733 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
734 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
735 mctz = soft_limit_tree_node_zone(nid, zid);
736 mem_cgroup_remove_exceeded(mz, mctz);
741 static struct mem_cgroup_per_zone *
742 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
744 struct rb_node *rightmost = NULL;
745 struct mem_cgroup_per_zone *mz;
749 rightmost = rb_last(&mctz->rb_root);
751 goto done; /* Nothing to reclaim from */
753 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
755 * Remove the node now but someone else can add it back,
756 * we will to add it back at the end of reclaim to its correct
757 * position in the tree.
759 __mem_cgroup_remove_exceeded(mz, mctz);
760 if (!soft_limit_excess(mz->memcg) ||
761 !css_tryget_online(&mz->memcg->css))
767 static struct mem_cgroup_per_zone *
768 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
770 struct mem_cgroup_per_zone *mz;
772 spin_lock_irq(&mctz->lock);
773 mz = __mem_cgroup_largest_soft_limit_node(mctz);
774 spin_unlock_irq(&mctz->lock);
779 * Implementation Note: reading percpu statistics for memcg.
781 * Both of vmstat[] and percpu_counter has threshold and do periodic
782 * synchronization to implement "quick" read. There are trade-off between
783 * reading cost and precision of value. Then, we may have a chance to implement
784 * a periodic synchronizion of counter in memcg's counter.
786 * But this _read() function is used for user interface now. The user accounts
787 * memory usage by memory cgroup and he _always_ requires exact value because
788 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
789 * have to visit all online cpus and make sum. So, for now, unnecessary
790 * synchronization is not implemented. (just implemented for cpu hotplug)
792 * If there are kernel internal actions which can make use of some not-exact
793 * value, and reading all cpu value can be performance bottleneck in some
794 * common workload, threashold and synchonization as vmstat[] should be
797 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
798 enum mem_cgroup_stat_index idx)
804 for_each_online_cpu(cpu)
805 val += per_cpu(memcg->stat->count[idx], cpu);
806 #ifdef CONFIG_HOTPLUG_CPU
807 spin_lock(&memcg->pcp_counter_lock);
808 val += memcg->nocpu_base.count[idx];
809 spin_unlock(&memcg->pcp_counter_lock);
815 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
816 enum mem_cgroup_events_index idx)
818 unsigned long val = 0;
822 for_each_online_cpu(cpu)
823 val += per_cpu(memcg->stat->events[idx], cpu);
824 #ifdef CONFIG_HOTPLUG_CPU
825 spin_lock(&memcg->pcp_counter_lock);
826 val += memcg->nocpu_base.events[idx];
827 spin_unlock(&memcg->pcp_counter_lock);
833 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
838 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
839 * counted as CACHE even if it's on ANON LRU.
842 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
845 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
848 if (PageTransHuge(page))
849 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
852 /* pagein of a big page is an event. So, ignore page size */
854 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
856 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
857 nr_pages = -nr_pages; /* for event */
860 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
863 unsigned long mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
865 struct mem_cgroup_per_zone *mz;
867 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
868 return mz->lru_size[lru];
871 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
873 unsigned int lru_mask)
875 unsigned long nr = 0;
878 VM_BUG_ON((unsigned)nid >= nr_node_ids);
880 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
881 struct mem_cgroup_per_zone *mz;
885 if (!(BIT(lru) & lru_mask))
887 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
888 nr += mz->lru_size[lru];
894 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
895 unsigned int lru_mask)
897 unsigned long nr = 0;
900 for_each_node_state(nid, N_MEMORY)
901 nr += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
905 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
906 enum mem_cgroup_events_target target)
908 unsigned long val, next;
910 val = __this_cpu_read(memcg->stat->nr_page_events);
911 next = __this_cpu_read(memcg->stat->targets[target]);
912 /* from time_after() in jiffies.h */
913 if ((long)next - (long)val < 0) {
915 case MEM_CGROUP_TARGET_THRESH:
916 next = val + THRESHOLDS_EVENTS_TARGET;
918 case MEM_CGROUP_TARGET_SOFTLIMIT:
919 next = val + SOFTLIMIT_EVENTS_TARGET;
921 case MEM_CGROUP_TARGET_NUMAINFO:
922 next = val + NUMAINFO_EVENTS_TARGET;
927 __this_cpu_write(memcg->stat->targets[target], next);
934 * Check events in order.
937 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
939 /* threshold event is triggered in finer grain than soft limit */
940 if (unlikely(mem_cgroup_event_ratelimit(memcg,
941 MEM_CGROUP_TARGET_THRESH))) {
943 bool do_numainfo __maybe_unused;
945 do_softlimit = mem_cgroup_event_ratelimit(memcg,
946 MEM_CGROUP_TARGET_SOFTLIMIT);
948 do_numainfo = mem_cgroup_event_ratelimit(memcg,
949 MEM_CGROUP_TARGET_NUMAINFO);
951 mem_cgroup_threshold(memcg);
952 if (unlikely(do_softlimit))
953 mem_cgroup_update_tree(memcg, page);
955 if (unlikely(do_numainfo))
956 atomic_inc(&memcg->numainfo_events);
961 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
964 * mm_update_next_owner() may clear mm->owner to NULL
965 * if it races with swapoff, page migration, etc.
966 * So this can be called with p == NULL.
971 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
974 static struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
976 struct mem_cgroup *memcg = NULL;
981 * Page cache insertions can happen withou an
982 * actual mm context, e.g. during disk probing
983 * on boot, loopback IO, acct() writes etc.
986 memcg = root_mem_cgroup;
988 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
989 if (unlikely(!memcg))
990 memcg = root_mem_cgroup;
992 } while (!css_tryget_online(&memcg->css));
998 * mem_cgroup_iter - iterate over memory cgroup hierarchy
999 * @root: hierarchy root
1000 * @prev: previously returned memcg, NULL on first invocation
1001 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1003 * Returns references to children of the hierarchy below @root, or
1004 * @root itself, or %NULL after a full round-trip.
1006 * Caller must pass the return value in @prev on subsequent
1007 * invocations for reference counting, or use mem_cgroup_iter_break()
1008 * to cancel a hierarchy walk before the round-trip is complete.
1010 * Reclaimers can specify a zone and a priority level in @reclaim to
1011 * divide up the memcgs in the hierarchy among all concurrent
1012 * reclaimers operating on the same zone and priority.
1014 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1015 struct mem_cgroup *prev,
1016 struct mem_cgroup_reclaim_cookie *reclaim)
1018 struct reclaim_iter *uninitialized_var(iter);
1019 struct cgroup_subsys_state *css = NULL;
1020 struct mem_cgroup *memcg = NULL;
1021 struct mem_cgroup *pos = NULL;
1023 if (mem_cgroup_disabled())
1027 root = root_mem_cgroup;
1029 if (prev && !reclaim)
1032 if (!root->use_hierarchy && root != root_mem_cgroup) {
1041 struct mem_cgroup_per_zone *mz;
1043 mz = mem_cgroup_zone_zoneinfo(root, reclaim->zone);
1044 iter = &mz->iter[reclaim->priority];
1046 if (prev && reclaim->generation != iter->generation)
1050 pos = ACCESS_ONCE(iter->position);
1052 * A racing update may change the position and
1053 * put the last reference, hence css_tryget(),
1054 * or retry to see the updated position.
1056 } while (pos && !css_tryget(&pos->css));
1063 css = css_next_descendant_pre(css, &root->css);
1066 * Reclaimers share the hierarchy walk, and a
1067 * new one might jump in right at the end of
1068 * the hierarchy - make sure they see at least
1069 * one group and restart from the beginning.
1077 * Verify the css and acquire a reference. The root
1078 * is provided by the caller, so we know it's alive
1079 * and kicking, and don't take an extra reference.
1081 memcg = mem_cgroup_from_css(css);
1083 if (css == &root->css)
1086 if (css_tryget(css)) {
1088 * Make sure the memcg is initialized:
1089 * mem_cgroup_css_online() orders the the
1090 * initialization against setting the flag.
1092 if (smp_load_acquire(&memcg->initialized))
1102 if (cmpxchg(&iter->position, pos, memcg) == pos) {
1104 css_get(&memcg->css);
1110 * pairs with css_tryget when dereferencing iter->position
1119 reclaim->generation = iter->generation;
1125 if (prev && prev != root)
1126 css_put(&prev->css);
1132 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1133 * @root: hierarchy root
1134 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1136 void mem_cgroup_iter_break(struct mem_cgroup *root,
1137 struct mem_cgroup *prev)
1140 root = root_mem_cgroup;
1141 if (prev && prev != root)
1142 css_put(&prev->css);
1146 * Iteration constructs for visiting all cgroups (under a tree). If
1147 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1148 * be used for reference counting.
1150 #define for_each_mem_cgroup_tree(iter, root) \
1151 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1153 iter = mem_cgroup_iter(root, iter, NULL))
1155 #define for_each_mem_cgroup(iter) \
1156 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1158 iter = mem_cgroup_iter(NULL, iter, NULL))
1160 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1162 struct mem_cgroup *memcg;
1165 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1166 if (unlikely(!memcg))
1171 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1174 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1182 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1185 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1186 * @zone: zone of the wanted lruvec
1187 * @memcg: memcg of the wanted lruvec
1189 * Returns the lru list vector holding pages for the given @zone and
1190 * @mem. This can be the global zone lruvec, if the memory controller
1193 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1194 struct mem_cgroup *memcg)
1196 struct mem_cgroup_per_zone *mz;
1197 struct lruvec *lruvec;
1199 if (mem_cgroup_disabled()) {
1200 lruvec = &zone->lruvec;
1204 mz = mem_cgroup_zone_zoneinfo(memcg, zone);
1205 lruvec = &mz->lruvec;
1208 * Since a node can be onlined after the mem_cgroup was created,
1209 * we have to be prepared to initialize lruvec->zone here;
1210 * and if offlined then reonlined, we need to reinitialize it.
1212 if (unlikely(lruvec->zone != zone))
1213 lruvec->zone = zone;
1218 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1220 * @zone: zone of the page
1222 * This function is only safe when following the LRU page isolation
1223 * and putback protocol: the LRU lock must be held, and the page must
1224 * either be PageLRU() or the caller must have isolated/allocated it.
1226 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1228 struct mem_cgroup_per_zone *mz;
1229 struct mem_cgroup *memcg;
1230 struct lruvec *lruvec;
1232 if (mem_cgroup_disabled()) {
1233 lruvec = &zone->lruvec;
1237 memcg = page->mem_cgroup;
1239 * Swapcache readahead pages are added to the LRU - and
1240 * possibly migrated - before they are charged.
1243 memcg = root_mem_cgroup;
1245 mz = mem_cgroup_page_zoneinfo(memcg, page);
1246 lruvec = &mz->lruvec;
1249 * Since a node can be onlined after the mem_cgroup was created,
1250 * we have to be prepared to initialize lruvec->zone here;
1251 * and if offlined then reonlined, we need to reinitialize it.
1253 if (unlikely(lruvec->zone != zone))
1254 lruvec->zone = zone;
1259 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1260 * @lruvec: mem_cgroup per zone lru vector
1261 * @lru: index of lru list the page is sitting on
1262 * @nr_pages: positive when adding or negative when removing
1264 * This function must be called when a page is added to or removed from an
1267 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1270 struct mem_cgroup_per_zone *mz;
1271 unsigned long *lru_size;
1273 if (mem_cgroup_disabled())
1276 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1277 lru_size = mz->lru_size + lru;
1278 *lru_size += nr_pages;
1279 VM_BUG_ON((long)(*lru_size) < 0);
1282 bool mem_cgroup_is_descendant(struct mem_cgroup *memcg, struct mem_cgroup *root)
1286 if (!root->use_hierarchy)
1288 return cgroup_is_descendant(memcg->css.cgroup, root->css.cgroup);
1291 bool task_in_mem_cgroup(struct task_struct *task, struct mem_cgroup *memcg)
1293 struct mem_cgroup *task_memcg;
1294 struct task_struct *p;
1297 p = find_lock_task_mm(task);
1299 task_memcg = get_mem_cgroup_from_mm(p->mm);
1303 * All threads may have already detached their mm's, but the oom
1304 * killer still needs to detect if they have already been oom
1305 * killed to prevent needlessly killing additional tasks.
1308 task_memcg = mem_cgroup_from_task(task);
1309 css_get(&task_memcg->css);
1312 ret = mem_cgroup_is_descendant(task_memcg, memcg);
1313 css_put(&task_memcg->css);
1317 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1319 unsigned long inactive_ratio;
1320 unsigned long inactive;
1321 unsigned long active;
1324 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1325 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1327 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1329 inactive_ratio = int_sqrt(10 * gb);
1333 return inactive * inactive_ratio < active;
1336 bool mem_cgroup_lruvec_online(struct lruvec *lruvec)
1338 struct mem_cgroup_per_zone *mz;
1339 struct mem_cgroup *memcg;
1341 if (mem_cgroup_disabled())
1344 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1347 return !!(memcg->css.flags & CSS_ONLINE);
1350 #define mem_cgroup_from_counter(counter, member) \
1351 container_of(counter, struct mem_cgroup, member)
1354 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1355 * @memcg: the memory cgroup
1357 * Returns the maximum amount of memory @mem can be charged with, in
1360 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1362 unsigned long margin = 0;
1363 unsigned long count;
1364 unsigned long limit;
1366 count = page_counter_read(&memcg->memory);
1367 limit = ACCESS_ONCE(memcg->memory.limit);
1369 margin = limit - count;
1371 if (do_swap_account) {
1372 count = page_counter_read(&memcg->memsw);
1373 limit = ACCESS_ONCE(memcg->memsw.limit);
1375 margin = min(margin, limit - count);
1381 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1384 if (mem_cgroup_disabled() || !memcg->css.parent)
1385 return vm_swappiness;
1387 return memcg->swappiness;
1391 * A routine for checking "mem" is under move_account() or not.
1393 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1394 * moving cgroups. This is for waiting at high-memory pressure
1397 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1399 struct mem_cgroup *from;
1400 struct mem_cgroup *to;
1403 * Unlike task_move routines, we access mc.to, mc.from not under
1404 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1406 spin_lock(&mc.lock);
1412 ret = mem_cgroup_is_descendant(from, memcg) ||
1413 mem_cgroup_is_descendant(to, memcg);
1415 spin_unlock(&mc.lock);
1419 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1421 if (mc.moving_task && current != mc.moving_task) {
1422 if (mem_cgroup_under_move(memcg)) {
1424 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1425 /* moving charge context might have finished. */
1428 finish_wait(&mc.waitq, &wait);
1435 #define K(x) ((x) << (PAGE_SHIFT-10))
1437 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1438 * @memcg: The memory cgroup that went over limit
1439 * @p: Task that is going to be killed
1441 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1444 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1446 /* oom_info_lock ensures that parallel ooms do not interleave */
1447 static DEFINE_MUTEX(oom_info_lock);
1448 struct mem_cgroup *iter;
1451 mutex_lock(&oom_info_lock);
1455 pr_info("Task in ");
1456 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1457 pr_cont(" killed as a result of limit of ");
1459 pr_info("Memory limit reached of cgroup ");
1462 pr_cont_cgroup_path(memcg->css.cgroup);
1467 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1468 K((u64)page_counter_read(&memcg->memory)),
1469 K((u64)memcg->memory.limit), memcg->memory.failcnt);
1470 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1471 K((u64)page_counter_read(&memcg->memsw)),
1472 K((u64)memcg->memsw.limit), memcg->memsw.failcnt);
1473 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1474 K((u64)page_counter_read(&memcg->kmem)),
1475 K((u64)memcg->kmem.limit), memcg->kmem.failcnt);
1477 for_each_mem_cgroup_tree(iter, memcg) {
1478 pr_info("Memory cgroup stats for ");
1479 pr_cont_cgroup_path(iter->css.cgroup);
1482 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1483 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1485 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1486 K(mem_cgroup_read_stat(iter, i)));
1489 for (i = 0; i < NR_LRU_LISTS; i++)
1490 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1491 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1495 mutex_unlock(&oom_info_lock);
1499 * This function returns the number of memcg under hierarchy tree. Returns
1500 * 1(self count) if no children.
1502 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1505 struct mem_cgroup *iter;
1507 for_each_mem_cgroup_tree(iter, memcg)
1513 * Return the memory (and swap, if configured) limit for a memcg.
1515 static unsigned long mem_cgroup_get_limit(struct mem_cgroup *memcg)
1517 unsigned long limit;
1519 limit = memcg->memory.limit;
1520 if (mem_cgroup_swappiness(memcg)) {
1521 unsigned long memsw_limit;
1523 memsw_limit = memcg->memsw.limit;
1524 limit = min(limit + total_swap_pages, memsw_limit);
1529 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1532 struct mem_cgroup *iter;
1533 unsigned long chosen_points = 0;
1534 unsigned long totalpages;
1535 unsigned int points = 0;
1536 struct task_struct *chosen = NULL;
1539 * If current has a pending SIGKILL or is exiting, then automatically
1540 * select it. The goal is to allow it to allocate so that it may
1541 * quickly exit and free its memory.
1543 if (fatal_signal_pending(current) || task_will_free_mem(current)) {
1544 mark_tsk_oom_victim(current);
1548 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL, memcg);
1549 totalpages = mem_cgroup_get_limit(memcg) ? : 1;
1550 for_each_mem_cgroup_tree(iter, memcg) {
1551 struct css_task_iter it;
1552 struct task_struct *task;
1554 css_task_iter_start(&iter->css, &it);
1555 while ((task = css_task_iter_next(&it))) {
1556 switch (oom_scan_process_thread(task, totalpages, NULL,
1558 case OOM_SCAN_SELECT:
1560 put_task_struct(chosen);
1562 chosen_points = ULONG_MAX;
1563 get_task_struct(chosen);
1565 case OOM_SCAN_CONTINUE:
1567 case OOM_SCAN_ABORT:
1568 css_task_iter_end(&it);
1569 mem_cgroup_iter_break(memcg, iter);
1571 put_task_struct(chosen);
1576 points = oom_badness(task, memcg, NULL, totalpages);
1577 if (!points || points < chosen_points)
1579 /* Prefer thread group leaders for display purposes */
1580 if (points == chosen_points &&
1581 thread_group_leader(chosen))
1585 put_task_struct(chosen);
1587 chosen_points = points;
1588 get_task_struct(chosen);
1590 css_task_iter_end(&it);
1595 points = chosen_points * 1000 / totalpages;
1596 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1597 NULL, "Memory cgroup out of memory");
1600 #if MAX_NUMNODES > 1
1603 * test_mem_cgroup_node_reclaimable
1604 * @memcg: the target memcg
1605 * @nid: the node ID to be checked.
1606 * @noswap : specify true here if the user wants flle only information.
1608 * This function returns whether the specified memcg contains any
1609 * reclaimable pages on a node. Returns true if there are any reclaimable
1610 * pages in the node.
1612 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1613 int nid, bool noswap)
1615 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1617 if (noswap || !total_swap_pages)
1619 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1626 * Always updating the nodemask is not very good - even if we have an empty
1627 * list or the wrong list here, we can start from some node and traverse all
1628 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1631 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1635 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1636 * pagein/pageout changes since the last update.
1638 if (!atomic_read(&memcg->numainfo_events))
1640 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1643 /* make a nodemask where this memcg uses memory from */
1644 memcg->scan_nodes = node_states[N_MEMORY];
1646 for_each_node_mask(nid, node_states[N_MEMORY]) {
1648 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1649 node_clear(nid, memcg->scan_nodes);
1652 atomic_set(&memcg->numainfo_events, 0);
1653 atomic_set(&memcg->numainfo_updating, 0);
1657 * Selecting a node where we start reclaim from. Because what we need is just
1658 * reducing usage counter, start from anywhere is O,K. Considering
1659 * memory reclaim from current node, there are pros. and cons.
1661 * Freeing memory from current node means freeing memory from a node which
1662 * we'll use or we've used. So, it may make LRU bad. And if several threads
1663 * hit limits, it will see a contention on a node. But freeing from remote
1664 * node means more costs for memory reclaim because of memory latency.
1666 * Now, we use round-robin. Better algorithm is welcomed.
1668 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1672 mem_cgroup_may_update_nodemask(memcg);
1673 node = memcg->last_scanned_node;
1675 node = next_node(node, memcg->scan_nodes);
1676 if (node == MAX_NUMNODES)
1677 node = first_node(memcg->scan_nodes);
1679 * We call this when we hit limit, not when pages are added to LRU.
1680 * No LRU may hold pages because all pages are UNEVICTABLE or
1681 * memcg is too small and all pages are not on LRU. In that case,
1682 * we use curret node.
1684 if (unlikely(node == MAX_NUMNODES))
1685 node = numa_node_id();
1687 memcg->last_scanned_node = node;
1691 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1697 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1700 unsigned long *total_scanned)
1702 struct mem_cgroup *victim = NULL;
1705 unsigned long excess;
1706 unsigned long nr_scanned;
1707 struct mem_cgroup_reclaim_cookie reclaim = {
1712 excess = soft_limit_excess(root_memcg);
1715 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1720 * If we have not been able to reclaim
1721 * anything, it might because there are
1722 * no reclaimable pages under this hierarchy
1727 * We want to do more targeted reclaim.
1728 * excess >> 2 is not to excessive so as to
1729 * reclaim too much, nor too less that we keep
1730 * coming back to reclaim from this cgroup
1732 if (total >= (excess >> 2) ||
1733 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1738 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1740 *total_scanned += nr_scanned;
1741 if (!soft_limit_excess(root_memcg))
1744 mem_cgroup_iter_break(root_memcg, victim);
1748 #ifdef CONFIG_LOCKDEP
1749 static struct lockdep_map memcg_oom_lock_dep_map = {
1750 .name = "memcg_oom_lock",
1754 static DEFINE_SPINLOCK(memcg_oom_lock);
1757 * Check OOM-Killer is already running under our hierarchy.
1758 * If someone is running, return false.
1760 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1762 struct mem_cgroup *iter, *failed = NULL;
1764 spin_lock(&memcg_oom_lock);
1766 for_each_mem_cgroup_tree(iter, memcg) {
1767 if (iter->oom_lock) {
1769 * this subtree of our hierarchy is already locked
1770 * so we cannot give a lock.
1773 mem_cgroup_iter_break(memcg, iter);
1776 iter->oom_lock = true;
1781 * OK, we failed to lock the whole subtree so we have
1782 * to clean up what we set up to the failing subtree
1784 for_each_mem_cgroup_tree(iter, memcg) {
1785 if (iter == failed) {
1786 mem_cgroup_iter_break(memcg, iter);
1789 iter->oom_lock = false;
1792 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1794 spin_unlock(&memcg_oom_lock);
1799 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1801 struct mem_cgroup *iter;
1803 spin_lock(&memcg_oom_lock);
1804 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1805 for_each_mem_cgroup_tree(iter, memcg)
1806 iter->oom_lock = false;
1807 spin_unlock(&memcg_oom_lock);
1810 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1812 struct mem_cgroup *iter;
1814 for_each_mem_cgroup_tree(iter, memcg)
1815 atomic_inc(&iter->under_oom);
1818 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1820 struct mem_cgroup *iter;
1823 * When a new child is created while the hierarchy is under oom,
1824 * mem_cgroup_oom_lock() may not be called. We have to use
1825 * atomic_add_unless() here.
1827 for_each_mem_cgroup_tree(iter, memcg)
1828 atomic_add_unless(&iter->under_oom, -1, 0);
1831 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1833 struct oom_wait_info {
1834 struct mem_cgroup *memcg;
1838 static int memcg_oom_wake_function(wait_queue_t *wait,
1839 unsigned mode, int sync, void *arg)
1841 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1842 struct mem_cgroup *oom_wait_memcg;
1843 struct oom_wait_info *oom_wait_info;
1845 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1846 oom_wait_memcg = oom_wait_info->memcg;
1848 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1849 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1851 return autoremove_wake_function(wait, mode, sync, arg);
1854 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
1856 atomic_inc(&memcg->oom_wakeups);
1857 /* for filtering, pass "memcg" as argument. */
1858 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1861 static void memcg_oom_recover(struct mem_cgroup *memcg)
1863 if (memcg && atomic_read(&memcg->under_oom))
1864 memcg_wakeup_oom(memcg);
1867 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1869 if (!current->memcg_oom.may_oom)
1872 * We are in the middle of the charge context here, so we
1873 * don't want to block when potentially sitting on a callstack
1874 * that holds all kinds of filesystem and mm locks.
1876 * Also, the caller may handle a failed allocation gracefully
1877 * (like optional page cache readahead) and so an OOM killer
1878 * invocation might not even be necessary.
1880 * That's why we don't do anything here except remember the
1881 * OOM context and then deal with it at the end of the page
1882 * fault when the stack is unwound, the locks are released,
1883 * and when we know whether the fault was overall successful.
1885 css_get(&memcg->css);
1886 current->memcg_oom.memcg = memcg;
1887 current->memcg_oom.gfp_mask = mask;
1888 current->memcg_oom.order = order;
1892 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1893 * @handle: actually kill/wait or just clean up the OOM state
1895 * This has to be called at the end of a page fault if the memcg OOM
1896 * handler was enabled.
1898 * Memcg supports userspace OOM handling where failed allocations must
1899 * sleep on a waitqueue until the userspace task resolves the
1900 * situation. Sleeping directly in the charge context with all kinds
1901 * of locks held is not a good idea, instead we remember an OOM state
1902 * in the task and mem_cgroup_oom_synchronize() has to be called at
1903 * the end of the page fault to complete the OOM handling.
1905 * Returns %true if an ongoing memcg OOM situation was detected and
1906 * completed, %false otherwise.
1908 bool mem_cgroup_oom_synchronize(bool handle)
1910 struct mem_cgroup *memcg = current->memcg_oom.memcg;
1911 struct oom_wait_info owait;
1914 /* OOM is global, do not handle */
1918 if (!handle || oom_killer_disabled)
1921 owait.memcg = memcg;
1922 owait.wait.flags = 0;
1923 owait.wait.func = memcg_oom_wake_function;
1924 owait.wait.private = current;
1925 INIT_LIST_HEAD(&owait.wait.task_list);
1927 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1928 mem_cgroup_mark_under_oom(memcg);
1930 locked = mem_cgroup_oom_trylock(memcg);
1933 mem_cgroup_oom_notify(memcg);
1935 if (locked && !memcg->oom_kill_disable) {
1936 mem_cgroup_unmark_under_oom(memcg);
1937 finish_wait(&memcg_oom_waitq, &owait.wait);
1938 mem_cgroup_out_of_memory(memcg, current->memcg_oom.gfp_mask,
1939 current->memcg_oom.order);
1942 mem_cgroup_unmark_under_oom(memcg);
1943 finish_wait(&memcg_oom_waitq, &owait.wait);
1947 mem_cgroup_oom_unlock(memcg);
1949 * There is no guarantee that an OOM-lock contender
1950 * sees the wakeups triggered by the OOM kill
1951 * uncharges. Wake any sleepers explicitely.
1953 memcg_oom_recover(memcg);
1956 current->memcg_oom.memcg = NULL;
1957 css_put(&memcg->css);
1962 * mem_cgroup_begin_page_stat - begin a page state statistics transaction
1963 * @page: page that is going to change accounted state
1965 * This function must mark the beginning of an accounted page state
1966 * change to prevent double accounting when the page is concurrently
1967 * being moved to another memcg:
1969 * memcg = mem_cgroup_begin_page_stat(page);
1970 * if (TestClearPageState(page))
1971 * mem_cgroup_update_page_stat(memcg, state, -1);
1972 * mem_cgroup_end_page_stat(memcg);
1974 struct mem_cgroup *mem_cgroup_begin_page_stat(struct page *page)
1976 struct mem_cgroup *memcg;
1977 unsigned long flags;
1980 * The RCU lock is held throughout the transaction. The fast
1981 * path can get away without acquiring the memcg->move_lock
1982 * because page moving starts with an RCU grace period.
1984 * The RCU lock also protects the memcg from being freed when
1985 * the page state that is going to change is the only thing
1986 * preventing the page from being uncharged.
1987 * E.g. end-writeback clearing PageWriteback(), which allows
1988 * migration to go ahead and uncharge the page before the
1989 * account transaction might be complete.
1993 if (mem_cgroup_disabled())
1996 memcg = page->mem_cgroup;
1997 if (unlikely(!memcg))
2000 if (atomic_read(&memcg->moving_account) <= 0)
2003 spin_lock_irqsave(&memcg->move_lock, flags);
2004 if (memcg != page->mem_cgroup) {
2005 spin_unlock_irqrestore(&memcg->move_lock, flags);
2010 * When charge migration first begins, we can have locked and
2011 * unlocked page stat updates happening concurrently. Track
2012 * the task who has the lock for mem_cgroup_end_page_stat().
2014 memcg->move_lock_task = current;
2015 memcg->move_lock_flags = flags;
2021 * mem_cgroup_end_page_stat - finish a page state statistics transaction
2022 * @memcg: the memcg that was accounted against
2024 void mem_cgroup_end_page_stat(struct mem_cgroup *memcg)
2026 if (memcg && memcg->move_lock_task == current) {
2027 unsigned long flags = memcg->move_lock_flags;
2029 memcg->move_lock_task = NULL;
2030 memcg->move_lock_flags = 0;
2032 spin_unlock_irqrestore(&memcg->move_lock, flags);
2039 * mem_cgroup_update_page_stat - update page state statistics
2040 * @memcg: memcg to account against
2041 * @idx: page state item to account
2042 * @val: number of pages (positive or negative)
2044 * See mem_cgroup_begin_page_stat() for locking requirements.
2046 void mem_cgroup_update_page_stat(struct mem_cgroup *memcg,
2047 enum mem_cgroup_stat_index idx, int val)
2049 VM_BUG_ON(!rcu_read_lock_held());
2052 this_cpu_add(memcg->stat->count[idx], val);
2056 * size of first charge trial. "32" comes from vmscan.c's magic value.
2057 * TODO: maybe necessary to use big numbers in big irons.
2059 #define CHARGE_BATCH 32U
2060 struct memcg_stock_pcp {
2061 struct mem_cgroup *cached; /* this never be root cgroup */
2062 unsigned int nr_pages;
2063 struct work_struct work;
2064 unsigned long flags;
2065 #define FLUSHING_CACHED_CHARGE 0
2067 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2068 static DEFINE_MUTEX(percpu_charge_mutex);
2071 * consume_stock: Try to consume stocked charge on this cpu.
2072 * @memcg: memcg to consume from.
2073 * @nr_pages: how many pages to charge.
2075 * The charges will only happen if @memcg matches the current cpu's memcg
2076 * stock, and at least @nr_pages are available in that stock. Failure to
2077 * service an allocation will refill the stock.
2079 * returns true if successful, false otherwise.
2081 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2083 struct memcg_stock_pcp *stock;
2086 if (nr_pages > CHARGE_BATCH)
2089 stock = &get_cpu_var(memcg_stock);
2090 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2091 stock->nr_pages -= nr_pages;
2094 put_cpu_var(memcg_stock);
2099 * Returns stocks cached in percpu and reset cached information.
2101 static void drain_stock(struct memcg_stock_pcp *stock)
2103 struct mem_cgroup *old = stock->cached;
2105 if (stock->nr_pages) {
2106 page_counter_uncharge(&old->memory, stock->nr_pages);
2107 if (do_swap_account)
2108 page_counter_uncharge(&old->memsw, stock->nr_pages);
2109 css_put_many(&old->css, stock->nr_pages);
2110 stock->nr_pages = 0;
2112 stock->cached = NULL;
2116 * This must be called under preempt disabled or must be called by
2117 * a thread which is pinned to local cpu.
2119 static void drain_local_stock(struct work_struct *dummy)
2121 struct memcg_stock_pcp *stock = this_cpu_ptr(&memcg_stock);
2123 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2127 * Cache charges(val) to local per_cpu area.
2128 * This will be consumed by consume_stock() function, later.
2130 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2132 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2134 if (stock->cached != memcg) { /* reset if necessary */
2136 stock->cached = memcg;
2138 stock->nr_pages += nr_pages;
2139 put_cpu_var(memcg_stock);
2143 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2144 * of the hierarchy under it.
2146 static void drain_all_stock(struct mem_cgroup *root_memcg)
2150 /* If someone's already draining, avoid adding running more workers. */
2151 if (!mutex_trylock(&percpu_charge_mutex))
2153 /* Notify other cpus that system-wide "drain" is running */
2156 for_each_online_cpu(cpu) {
2157 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2158 struct mem_cgroup *memcg;
2160 memcg = stock->cached;
2161 if (!memcg || !stock->nr_pages)
2163 if (!mem_cgroup_is_descendant(memcg, root_memcg))
2165 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2167 drain_local_stock(&stock->work);
2169 schedule_work_on(cpu, &stock->work);
2174 mutex_unlock(&percpu_charge_mutex);
2178 * This function drains percpu counter value from DEAD cpu and
2179 * move it to local cpu. Note that this function can be preempted.
2181 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2185 spin_lock(&memcg->pcp_counter_lock);
2186 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2187 long x = per_cpu(memcg->stat->count[i], cpu);
2189 per_cpu(memcg->stat->count[i], cpu) = 0;
2190 memcg->nocpu_base.count[i] += x;
2192 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2193 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2195 per_cpu(memcg->stat->events[i], cpu) = 0;
2196 memcg->nocpu_base.events[i] += x;
2198 spin_unlock(&memcg->pcp_counter_lock);
2201 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
2202 unsigned long action,
2205 int cpu = (unsigned long)hcpu;
2206 struct memcg_stock_pcp *stock;
2207 struct mem_cgroup *iter;
2209 if (action == CPU_ONLINE)
2212 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2215 for_each_mem_cgroup(iter)
2216 mem_cgroup_drain_pcp_counter(iter, cpu);
2218 stock = &per_cpu(memcg_stock, cpu);
2223 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2224 unsigned int nr_pages)
2226 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2227 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2228 struct mem_cgroup *mem_over_limit;
2229 struct page_counter *counter;
2230 unsigned long nr_reclaimed;
2231 bool may_swap = true;
2232 bool drained = false;
2235 if (mem_cgroup_is_root(memcg))
2238 if (consume_stock(memcg, nr_pages))
2241 if (!do_swap_account ||
2242 !page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2243 if (!page_counter_try_charge(&memcg->memory, batch, &counter))
2245 if (do_swap_account)
2246 page_counter_uncharge(&memcg->memsw, batch);
2247 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2249 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2253 if (batch > nr_pages) {
2259 * Unlike in global OOM situations, memcg is not in a physical
2260 * memory shortage. Allow dying and OOM-killed tasks to
2261 * bypass the last charges so that they can exit quickly and
2262 * free their memory.
2264 if (unlikely(test_thread_flag(TIF_MEMDIE) ||
2265 fatal_signal_pending(current) ||
2266 current->flags & PF_EXITING))
2269 if (unlikely(task_in_memcg_oom(current)))
2272 if (!(gfp_mask & __GFP_WAIT))
2275 mem_cgroup_events(mem_over_limit, MEMCG_MAX, 1);
2277 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2278 gfp_mask, may_swap);
2280 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2284 drain_all_stock(mem_over_limit);
2289 if (gfp_mask & __GFP_NORETRY)
2292 * Even though the limit is exceeded at this point, reclaim
2293 * may have been able to free some pages. Retry the charge
2294 * before killing the task.
2296 * Only for regular pages, though: huge pages are rather
2297 * unlikely to succeed so close to the limit, and we fall back
2298 * to regular pages anyway in case of failure.
2300 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2303 * At task move, charge accounts can be doubly counted. So, it's
2304 * better to wait until the end of task_move if something is going on.
2306 if (mem_cgroup_wait_acct_move(mem_over_limit))
2312 if (gfp_mask & __GFP_NOFAIL)
2315 if (fatal_signal_pending(current))
2318 mem_cgroup_events(mem_over_limit, MEMCG_OOM, 1);
2320 mem_cgroup_oom(mem_over_limit, gfp_mask, get_order(nr_pages));
2322 if (!(gfp_mask & __GFP_NOFAIL))
2328 css_get_many(&memcg->css, batch);
2329 if (batch > nr_pages)
2330 refill_stock(memcg, batch - nr_pages);
2332 * If the hierarchy is above the normal consumption range,
2333 * make the charging task trim their excess contribution.
2336 if (page_counter_read(&memcg->memory) <= memcg->high)
2338 mem_cgroup_events(memcg, MEMCG_HIGH, 1);
2339 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
2340 } while ((memcg = parent_mem_cgroup(memcg)));
2345 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2347 if (mem_cgroup_is_root(memcg))
2350 page_counter_uncharge(&memcg->memory, nr_pages);
2351 if (do_swap_account)
2352 page_counter_uncharge(&memcg->memsw, nr_pages);
2354 css_put_many(&memcg->css, nr_pages);
2358 * try_get_mem_cgroup_from_page - look up page's memcg association
2361 * Look up, get a css reference, and return the memcg that owns @page.
2363 * The page must be locked to prevent racing with swap-in and page
2364 * cache charges. If coming from an unlocked page table, the caller
2365 * must ensure the page is on the LRU or this can race with charging.
2367 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2369 struct mem_cgroup *memcg;
2373 VM_BUG_ON_PAGE(!PageLocked(page), page);
2375 memcg = page->mem_cgroup;
2377 if (!css_tryget_online(&memcg->css))
2379 } else if (PageSwapCache(page)) {
2380 ent.val = page_private(page);
2381 id = lookup_swap_cgroup_id(ent);
2383 memcg = mem_cgroup_from_id(id);
2384 if (memcg && !css_tryget_online(&memcg->css))
2391 static void lock_page_lru(struct page *page, int *isolated)
2393 struct zone *zone = page_zone(page);
2395 spin_lock_irq(&zone->lru_lock);
2396 if (PageLRU(page)) {
2397 struct lruvec *lruvec;
2399 lruvec = mem_cgroup_page_lruvec(page, zone);
2401 del_page_from_lru_list(page, lruvec, page_lru(page));
2407 static void unlock_page_lru(struct page *page, int isolated)
2409 struct zone *zone = page_zone(page);
2412 struct lruvec *lruvec;
2414 lruvec = mem_cgroup_page_lruvec(page, zone);
2415 VM_BUG_ON_PAGE(PageLRU(page), page);
2417 add_page_to_lru_list(page, lruvec, page_lru(page));
2419 spin_unlock_irq(&zone->lru_lock);
2422 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2427 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2430 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2431 * may already be on some other mem_cgroup's LRU. Take care of it.
2434 lock_page_lru(page, &isolated);
2437 * Nobody should be changing or seriously looking at
2438 * page->mem_cgroup at this point:
2440 * - the page is uncharged
2442 * - the page is off-LRU
2444 * - an anonymous fault has exclusive page access, except for
2445 * a locked page table
2447 * - a page cache insertion, a swapin fault, or a migration
2448 * have the page locked
2450 page->mem_cgroup = memcg;
2453 unlock_page_lru(page, isolated);
2456 #ifdef CONFIG_MEMCG_KMEM
2457 int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp,
2458 unsigned long nr_pages)
2460 struct page_counter *counter;
2463 ret = page_counter_try_charge(&memcg->kmem, nr_pages, &counter);
2467 ret = try_charge(memcg, gfp, nr_pages);
2468 if (ret == -EINTR) {
2470 * try_charge() chose to bypass to root due to OOM kill or
2471 * fatal signal. Since our only options are to either fail
2472 * the allocation or charge it to this cgroup, do it as a
2473 * temporary condition. But we can't fail. From a kmem/slab
2474 * perspective, the cache has already been selected, by
2475 * mem_cgroup_kmem_get_cache(), so it is too late to change
2478 * This condition will only trigger if the task entered
2479 * memcg_charge_kmem in a sane state, but was OOM-killed
2480 * during try_charge() above. Tasks that were already dying
2481 * when the allocation triggers should have been already
2482 * directed to the root cgroup in memcontrol.h
2484 page_counter_charge(&memcg->memory, nr_pages);
2485 if (do_swap_account)
2486 page_counter_charge(&memcg->memsw, nr_pages);
2487 css_get_many(&memcg->css, nr_pages);
2490 page_counter_uncharge(&memcg->kmem, nr_pages);
2495 void memcg_uncharge_kmem(struct mem_cgroup *memcg, unsigned long nr_pages)
2497 page_counter_uncharge(&memcg->memory, nr_pages);
2498 if (do_swap_account)
2499 page_counter_uncharge(&memcg->memsw, nr_pages);
2501 page_counter_uncharge(&memcg->kmem, nr_pages);
2503 css_put_many(&memcg->css, nr_pages);
2507 * helper for acessing a memcg's index. It will be used as an index in the
2508 * child cache array in kmem_cache, and also to derive its name. This function
2509 * will return -1 when this is not a kmem-limited memcg.
2511 int memcg_cache_id(struct mem_cgroup *memcg)
2513 return memcg ? memcg->kmemcg_id : -1;
2516 static int memcg_alloc_cache_id(void)
2521 id = ida_simple_get(&memcg_cache_ida,
2522 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2526 if (id < memcg_nr_cache_ids)
2530 * There's no space for the new id in memcg_caches arrays,
2531 * so we have to grow them.
2533 down_write(&memcg_cache_ids_sem);
2535 size = 2 * (id + 1);
2536 if (size < MEMCG_CACHES_MIN_SIZE)
2537 size = MEMCG_CACHES_MIN_SIZE;
2538 else if (size > MEMCG_CACHES_MAX_SIZE)
2539 size = MEMCG_CACHES_MAX_SIZE;
2541 err = memcg_update_all_caches(size);
2543 err = memcg_update_all_list_lrus(size);
2545 memcg_nr_cache_ids = size;
2547 up_write(&memcg_cache_ids_sem);
2550 ida_simple_remove(&memcg_cache_ida, id);
2556 static void memcg_free_cache_id(int id)
2558 ida_simple_remove(&memcg_cache_ida, id);
2561 struct memcg_kmem_cache_create_work {
2562 struct mem_cgroup *memcg;
2563 struct kmem_cache *cachep;
2564 struct work_struct work;
2567 static void memcg_kmem_cache_create_func(struct work_struct *w)
2569 struct memcg_kmem_cache_create_work *cw =
2570 container_of(w, struct memcg_kmem_cache_create_work, work);
2571 struct mem_cgroup *memcg = cw->memcg;
2572 struct kmem_cache *cachep = cw->cachep;
2574 memcg_create_kmem_cache(memcg, cachep);
2576 css_put(&memcg->css);
2581 * Enqueue the creation of a per-memcg kmem_cache.
2583 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2584 struct kmem_cache *cachep)
2586 struct memcg_kmem_cache_create_work *cw;
2588 cw = kmalloc(sizeof(*cw), GFP_NOWAIT);
2592 css_get(&memcg->css);
2595 cw->cachep = cachep;
2596 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2598 schedule_work(&cw->work);
2601 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2602 struct kmem_cache *cachep)
2605 * We need to stop accounting when we kmalloc, because if the
2606 * corresponding kmalloc cache is not yet created, the first allocation
2607 * in __memcg_schedule_kmem_cache_create will recurse.
2609 * However, it is better to enclose the whole function. Depending on
2610 * the debugging options enabled, INIT_WORK(), for instance, can
2611 * trigger an allocation. This too, will make us recurse. Because at
2612 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2613 * the safest choice is to do it like this, wrapping the whole function.
2615 current->memcg_kmem_skip_account = 1;
2616 __memcg_schedule_kmem_cache_create(memcg, cachep);
2617 current->memcg_kmem_skip_account = 0;
2621 * Return the kmem_cache we're supposed to use for a slab allocation.
2622 * We try to use the current memcg's version of the cache.
2624 * If the cache does not exist yet, if we are the first user of it,
2625 * we either create it immediately, if possible, or create it asynchronously
2627 * In the latter case, we will let the current allocation go through with
2628 * the original cache.
2630 * Can't be called in interrupt context or from kernel threads.
2631 * This function needs to be called with rcu_read_lock() held.
2633 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep)
2635 struct mem_cgroup *memcg;
2636 struct kmem_cache *memcg_cachep;
2639 VM_BUG_ON(!is_root_cache(cachep));
2641 if (current->memcg_kmem_skip_account)
2644 memcg = get_mem_cgroup_from_mm(current->mm);
2645 kmemcg_id = ACCESS_ONCE(memcg->kmemcg_id);
2649 memcg_cachep = cache_from_memcg_idx(cachep, kmemcg_id);
2650 if (likely(memcg_cachep))
2651 return memcg_cachep;
2654 * If we are in a safe context (can wait, and not in interrupt
2655 * context), we could be be predictable and return right away.
2656 * This would guarantee that the allocation being performed
2657 * already belongs in the new cache.
2659 * However, there are some clashes that can arrive from locking.
2660 * For instance, because we acquire the slab_mutex while doing
2661 * memcg_create_kmem_cache, this means no further allocation
2662 * could happen with the slab_mutex held. So it's better to
2665 memcg_schedule_kmem_cache_create(memcg, cachep);
2667 css_put(&memcg->css);
2671 void __memcg_kmem_put_cache(struct kmem_cache *cachep)
2673 if (!is_root_cache(cachep))
2674 css_put(&cachep->memcg_params.memcg->css);
2678 * We need to verify if the allocation against current->mm->owner's memcg is
2679 * possible for the given order. But the page is not allocated yet, so we'll
2680 * need a further commit step to do the final arrangements.
2682 * It is possible for the task to switch cgroups in this mean time, so at
2683 * commit time, we can't rely on task conversion any longer. We'll then use
2684 * the handle argument to return to the caller which cgroup we should commit
2685 * against. We could also return the memcg directly and avoid the pointer
2686 * passing, but a boolean return value gives better semantics considering
2687 * the compiled-out case as well.
2689 * Returning true means the allocation is possible.
2692 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
2694 struct mem_cgroup *memcg;
2699 memcg = get_mem_cgroup_from_mm(current->mm);
2701 if (!memcg_kmem_is_active(memcg)) {
2702 css_put(&memcg->css);
2706 ret = memcg_charge_kmem(memcg, gfp, 1 << order);
2710 css_put(&memcg->css);
2714 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
2717 VM_BUG_ON(mem_cgroup_is_root(memcg));
2719 /* The page allocation failed. Revert */
2721 memcg_uncharge_kmem(memcg, 1 << order);
2724 page->mem_cgroup = memcg;
2727 void __memcg_kmem_uncharge_pages(struct page *page, int order)
2729 struct mem_cgroup *memcg = page->mem_cgroup;
2734 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2736 memcg_uncharge_kmem(memcg, 1 << order);
2737 page->mem_cgroup = NULL;
2740 struct mem_cgroup *__mem_cgroup_from_kmem(void *ptr)
2742 struct mem_cgroup *memcg = NULL;
2743 struct kmem_cache *cachep;
2746 page = virt_to_head_page(ptr);
2747 if (PageSlab(page)) {
2748 cachep = page->slab_cache;
2749 if (!is_root_cache(cachep))
2750 memcg = cachep->memcg_params.memcg;
2752 /* page allocated by alloc_kmem_pages */
2753 memcg = page->mem_cgroup;
2757 #endif /* CONFIG_MEMCG_KMEM */
2759 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2762 * Because tail pages are not marked as "used", set it. We're under
2763 * zone->lru_lock, 'splitting on pmd' and compound_lock.
2764 * charge/uncharge will be never happen and move_account() is done under
2765 * compound_lock(), so we don't have to take care of races.
2767 void mem_cgroup_split_huge_fixup(struct page *head)
2771 if (mem_cgroup_disabled())
2774 for (i = 1; i < HPAGE_PMD_NR; i++)
2775 head[i].mem_cgroup = head->mem_cgroup;
2777 __this_cpu_sub(head->mem_cgroup->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
2780 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2782 #ifdef CONFIG_MEMCG_SWAP
2783 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
2786 int val = (charge) ? 1 : -1;
2787 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
2791 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2792 * @entry: swap entry to be moved
2793 * @from: mem_cgroup which the entry is moved from
2794 * @to: mem_cgroup which the entry is moved to
2796 * It succeeds only when the swap_cgroup's record for this entry is the same
2797 * as the mem_cgroup's id of @from.
2799 * Returns 0 on success, -EINVAL on failure.
2801 * The caller must have charged to @to, IOW, called page_counter_charge() about
2802 * both res and memsw, and called css_get().
2804 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2805 struct mem_cgroup *from, struct mem_cgroup *to)
2807 unsigned short old_id, new_id;
2809 old_id = mem_cgroup_id(from);
2810 new_id = mem_cgroup_id(to);
2812 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2813 mem_cgroup_swap_statistics(from, false);
2814 mem_cgroup_swap_statistics(to, true);
2820 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2821 struct mem_cgroup *from, struct mem_cgroup *to)
2827 static DEFINE_MUTEX(memcg_limit_mutex);
2829 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
2830 unsigned long limit)
2832 unsigned long curusage;
2833 unsigned long oldusage;
2834 bool enlarge = false;
2839 * For keeping hierarchical_reclaim simple, how long we should retry
2840 * is depends on callers. We set our retry-count to be function
2841 * of # of children which we should visit in this loop.
2843 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2844 mem_cgroup_count_children(memcg);
2846 oldusage = page_counter_read(&memcg->memory);
2849 if (signal_pending(current)) {
2854 mutex_lock(&memcg_limit_mutex);
2855 if (limit > memcg->memsw.limit) {
2856 mutex_unlock(&memcg_limit_mutex);
2860 if (limit > memcg->memory.limit)
2862 ret = page_counter_limit(&memcg->memory, limit);
2863 mutex_unlock(&memcg_limit_mutex);
2868 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, true);
2870 curusage = page_counter_read(&memcg->memory);
2871 /* Usage is reduced ? */
2872 if (curusage >= oldusage)
2875 oldusage = curusage;
2876 } while (retry_count);
2878 if (!ret && enlarge)
2879 memcg_oom_recover(memcg);
2884 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
2885 unsigned long limit)
2887 unsigned long curusage;
2888 unsigned long oldusage;
2889 bool enlarge = false;
2893 /* see mem_cgroup_resize_res_limit */
2894 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2895 mem_cgroup_count_children(memcg);
2897 oldusage = page_counter_read(&memcg->memsw);
2900 if (signal_pending(current)) {
2905 mutex_lock(&memcg_limit_mutex);
2906 if (limit < memcg->memory.limit) {
2907 mutex_unlock(&memcg_limit_mutex);
2911 if (limit > memcg->memsw.limit)
2913 ret = page_counter_limit(&memcg->memsw, limit);
2914 mutex_unlock(&memcg_limit_mutex);
2919 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, false);
2921 curusage = page_counter_read(&memcg->memsw);
2922 /* Usage is reduced ? */
2923 if (curusage >= oldusage)
2926 oldusage = curusage;
2927 } while (retry_count);
2929 if (!ret && enlarge)
2930 memcg_oom_recover(memcg);
2935 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
2937 unsigned long *total_scanned)
2939 unsigned long nr_reclaimed = 0;
2940 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
2941 unsigned long reclaimed;
2943 struct mem_cgroup_tree_per_zone *mctz;
2944 unsigned long excess;
2945 unsigned long nr_scanned;
2950 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
2952 * This loop can run a while, specially if mem_cgroup's continuously
2953 * keep exceeding their soft limit and putting the system under
2960 mz = mem_cgroup_largest_soft_limit_node(mctz);
2965 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
2966 gfp_mask, &nr_scanned);
2967 nr_reclaimed += reclaimed;
2968 *total_scanned += nr_scanned;
2969 spin_lock_irq(&mctz->lock);
2970 __mem_cgroup_remove_exceeded(mz, mctz);
2973 * If we failed to reclaim anything from this memory cgroup
2974 * it is time to move on to the next cgroup
2978 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
2980 excess = soft_limit_excess(mz->memcg);
2982 * One school of thought says that we should not add
2983 * back the node to the tree if reclaim returns 0.
2984 * But our reclaim could return 0, simply because due
2985 * to priority we are exposing a smaller subset of
2986 * memory to reclaim from. Consider this as a longer
2989 /* If excess == 0, no tree ops */
2990 __mem_cgroup_insert_exceeded(mz, mctz, excess);
2991 spin_unlock_irq(&mctz->lock);
2992 css_put(&mz->memcg->css);
2995 * Could not reclaim anything and there are no more
2996 * mem cgroups to try or we seem to be looping without
2997 * reclaiming anything.
2999 if (!nr_reclaimed &&
3001 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3003 } while (!nr_reclaimed);
3005 css_put(&next_mz->memcg->css);
3006 return nr_reclaimed;
3010 * Test whether @memcg has children, dead or alive. Note that this
3011 * function doesn't care whether @memcg has use_hierarchy enabled and
3012 * returns %true if there are child csses according to the cgroup
3013 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
3015 static inline bool memcg_has_children(struct mem_cgroup *memcg)
3020 * The lock does not prevent addition or deletion of children, but
3021 * it prevents a new child from being initialized based on this
3022 * parent in css_online(), so it's enough to decide whether
3023 * hierarchically inherited attributes can still be changed or not.
3025 lockdep_assert_held(&memcg_create_mutex);
3028 ret = css_next_child(NULL, &memcg->css);
3034 * Reclaims as many pages from the given memcg as possible and moves
3035 * the rest to the parent.
3037 * Caller is responsible for holding css reference for memcg.
3039 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3041 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3043 /* we call try-to-free pages for make this cgroup empty */
3044 lru_add_drain_all();
3045 /* try to free all pages in this cgroup */
3046 while (nr_retries && page_counter_read(&memcg->memory)) {
3049 if (signal_pending(current))
3052 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3056 /* maybe some writeback is necessary */
3057 congestion_wait(BLK_RW_ASYNC, HZ/10);
3065 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3066 char *buf, size_t nbytes,
3069 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3071 if (mem_cgroup_is_root(memcg))
3073 return mem_cgroup_force_empty(memcg) ?: nbytes;
3076 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3079 return mem_cgroup_from_css(css)->use_hierarchy;
3082 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3083 struct cftype *cft, u64 val)
3086 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3087 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
3089 mutex_lock(&memcg_create_mutex);
3091 if (memcg->use_hierarchy == val)
3095 * If parent's use_hierarchy is set, we can't make any modifications
3096 * in the child subtrees. If it is unset, then the change can
3097 * occur, provided the current cgroup has no children.
3099 * For the root cgroup, parent_mem is NULL, we allow value to be
3100 * set if there are no children.
3102 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3103 (val == 1 || val == 0)) {
3104 if (!memcg_has_children(memcg))
3105 memcg->use_hierarchy = val;
3112 mutex_unlock(&memcg_create_mutex);
3117 static unsigned long tree_stat(struct mem_cgroup *memcg,
3118 enum mem_cgroup_stat_index idx)
3120 struct mem_cgroup *iter;
3123 /* Per-cpu values can be negative, use a signed accumulator */
3124 for_each_mem_cgroup_tree(iter, memcg)
3125 val += mem_cgroup_read_stat(iter, idx);
3127 if (val < 0) /* race ? */
3132 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3136 if (mem_cgroup_is_root(memcg)) {
3137 val = tree_stat(memcg, MEM_CGROUP_STAT_CACHE);
3138 val += tree_stat(memcg, MEM_CGROUP_STAT_RSS);
3140 val += tree_stat(memcg, MEM_CGROUP_STAT_SWAP);
3143 val = page_counter_read(&memcg->memory);
3145 val = page_counter_read(&memcg->memsw);
3147 return val << PAGE_SHIFT;
3158 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3161 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3162 struct page_counter *counter;
3164 switch (MEMFILE_TYPE(cft->private)) {
3166 counter = &memcg->memory;
3169 counter = &memcg->memsw;
3172 counter = &memcg->kmem;
3178 switch (MEMFILE_ATTR(cft->private)) {
3180 if (counter == &memcg->memory)
3181 return mem_cgroup_usage(memcg, false);
3182 if (counter == &memcg->memsw)
3183 return mem_cgroup_usage(memcg, true);
3184 return (u64)page_counter_read(counter) * PAGE_SIZE;
3186 return (u64)counter->limit * PAGE_SIZE;
3188 return (u64)counter->watermark * PAGE_SIZE;
3190 return counter->failcnt;
3191 case RES_SOFT_LIMIT:
3192 return (u64)memcg->soft_limit * PAGE_SIZE;
3198 #ifdef CONFIG_MEMCG_KMEM
3199 static int memcg_activate_kmem(struct mem_cgroup *memcg,
3200 unsigned long nr_pages)
3205 BUG_ON(memcg->kmemcg_id >= 0);
3206 BUG_ON(memcg->kmem_acct_activated);
3207 BUG_ON(memcg->kmem_acct_active);
3210 * For simplicity, we won't allow this to be disabled. It also can't
3211 * be changed if the cgroup has children already, or if tasks had
3214 * If tasks join before we set the limit, a person looking at
3215 * kmem.usage_in_bytes will have no way to determine when it took
3216 * place, which makes the value quite meaningless.
3218 * After it first became limited, changes in the value of the limit are
3219 * of course permitted.
3221 mutex_lock(&memcg_create_mutex);
3222 if (cgroup_has_tasks(memcg->css.cgroup) ||
3223 (memcg->use_hierarchy && memcg_has_children(memcg)))
3225 mutex_unlock(&memcg_create_mutex);
3229 memcg_id = memcg_alloc_cache_id();
3236 * We couldn't have accounted to this cgroup, because it hasn't got
3237 * activated yet, so this should succeed.
3239 err = page_counter_limit(&memcg->kmem, nr_pages);
3242 static_key_slow_inc(&memcg_kmem_enabled_key);
3244 * A memory cgroup is considered kmem-active as soon as it gets
3245 * kmemcg_id. Setting the id after enabling static branching will
3246 * guarantee no one starts accounting before all call sites are
3249 memcg->kmemcg_id = memcg_id;
3250 memcg->kmem_acct_activated = true;
3251 memcg->kmem_acct_active = true;
3256 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
3257 unsigned long limit)
3261 mutex_lock(&memcg_limit_mutex);
3262 if (!memcg_kmem_is_active(memcg))
3263 ret = memcg_activate_kmem(memcg, limit);
3265 ret = page_counter_limit(&memcg->kmem, limit);
3266 mutex_unlock(&memcg_limit_mutex);
3270 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
3273 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
3278 mutex_lock(&memcg_limit_mutex);
3280 * If the parent cgroup is not kmem-active now, it cannot be activated
3281 * after this point, because it has at least one child already.
3283 if (memcg_kmem_is_active(parent))
3284 ret = memcg_activate_kmem(memcg, PAGE_COUNTER_MAX);
3285 mutex_unlock(&memcg_limit_mutex);
3289 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
3290 unsigned long limit)
3294 #endif /* CONFIG_MEMCG_KMEM */
3297 * The user of this function is...
3300 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3301 char *buf, size_t nbytes, loff_t off)
3303 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3304 unsigned long nr_pages;
3307 buf = strstrip(buf);
3308 ret = page_counter_memparse(buf, "-1", &nr_pages);
3312 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3314 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3318 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3320 ret = mem_cgroup_resize_limit(memcg, nr_pages);
3323 ret = mem_cgroup_resize_memsw_limit(memcg, nr_pages);
3326 ret = memcg_update_kmem_limit(memcg, nr_pages);
3330 case RES_SOFT_LIMIT:
3331 memcg->soft_limit = nr_pages;
3335 return ret ?: nbytes;
3338 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3339 size_t nbytes, loff_t off)
3341 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3342 struct page_counter *counter;
3344 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3346 counter = &memcg->memory;
3349 counter = &memcg->memsw;
3352 counter = &memcg->kmem;
3358 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3360 page_counter_reset_watermark(counter);
3363 counter->failcnt = 0;
3372 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3375 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3379 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3380 struct cftype *cft, u64 val)
3382 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3384 if (val & ~MOVE_MASK)
3388 * No kind of locking is needed in here, because ->can_attach() will
3389 * check this value once in the beginning of the process, and then carry
3390 * on with stale data. This means that changes to this value will only
3391 * affect task migrations starting after the change.
3393 memcg->move_charge_at_immigrate = val;
3397 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3398 struct cftype *cft, u64 val)
3405 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3409 unsigned int lru_mask;
3412 static const struct numa_stat stats[] = {
3413 { "total", LRU_ALL },
3414 { "file", LRU_ALL_FILE },
3415 { "anon", LRU_ALL_ANON },
3416 { "unevictable", BIT(LRU_UNEVICTABLE) },
3418 const struct numa_stat *stat;
3421 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3423 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3424 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3425 seq_printf(m, "%s=%lu", stat->name, nr);
3426 for_each_node_state(nid, N_MEMORY) {
3427 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3429 seq_printf(m, " N%d=%lu", nid, nr);
3434 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3435 struct mem_cgroup *iter;
3438 for_each_mem_cgroup_tree(iter, memcg)
3439 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3440 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3441 for_each_node_state(nid, N_MEMORY) {
3443 for_each_mem_cgroup_tree(iter, memcg)
3444 nr += mem_cgroup_node_nr_lru_pages(
3445 iter, nid, stat->lru_mask);
3446 seq_printf(m, " N%d=%lu", nid, nr);
3453 #endif /* CONFIG_NUMA */
3455 static int memcg_stat_show(struct seq_file *m, void *v)
3457 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3458 unsigned long memory, memsw;
3459 struct mem_cgroup *mi;
3462 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_stat_names) !=
3463 MEM_CGROUP_STAT_NSTATS);
3464 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_events_names) !=
3465 MEM_CGROUP_EVENTS_NSTATS);
3466 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3468 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3469 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
3471 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
3472 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
3475 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
3476 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
3477 mem_cgroup_read_events(memcg, i));
3479 for (i = 0; i < NR_LRU_LISTS; i++)
3480 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3481 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
3483 /* Hierarchical information */
3484 memory = memsw = PAGE_COUNTER_MAX;
3485 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3486 memory = min(memory, mi->memory.limit);
3487 memsw = min(memsw, mi->memsw.limit);
3489 seq_printf(m, "hierarchical_memory_limit %llu\n",
3490 (u64)memory * PAGE_SIZE);
3491 if (do_swap_account)
3492 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3493 (u64)memsw * PAGE_SIZE);
3495 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3498 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
3500 for_each_mem_cgroup_tree(mi, memcg)
3501 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
3502 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
3505 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
3506 unsigned long long val = 0;
3508 for_each_mem_cgroup_tree(mi, memcg)
3509 val += mem_cgroup_read_events(mi, i);
3510 seq_printf(m, "total_%s %llu\n",
3511 mem_cgroup_events_names[i], val);
3514 for (i = 0; i < NR_LRU_LISTS; i++) {
3515 unsigned long long val = 0;
3517 for_each_mem_cgroup_tree(mi, memcg)
3518 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
3519 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
3522 #ifdef CONFIG_DEBUG_VM
3525 struct mem_cgroup_per_zone *mz;
3526 struct zone_reclaim_stat *rstat;
3527 unsigned long recent_rotated[2] = {0, 0};
3528 unsigned long recent_scanned[2] = {0, 0};
3530 for_each_online_node(nid)
3531 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3532 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
3533 rstat = &mz->lruvec.reclaim_stat;
3535 recent_rotated[0] += rstat->recent_rotated[0];
3536 recent_rotated[1] += rstat->recent_rotated[1];
3537 recent_scanned[0] += rstat->recent_scanned[0];
3538 recent_scanned[1] += rstat->recent_scanned[1];
3540 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3541 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3542 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3543 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3550 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3553 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3555 return mem_cgroup_swappiness(memcg);
3558 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3559 struct cftype *cft, u64 val)
3561 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3567 memcg->swappiness = val;
3569 vm_swappiness = val;
3574 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3576 struct mem_cgroup_threshold_ary *t;
3577 unsigned long usage;
3582 t = rcu_dereference(memcg->thresholds.primary);
3584 t = rcu_dereference(memcg->memsw_thresholds.primary);
3589 usage = mem_cgroup_usage(memcg, swap);
3592 * current_threshold points to threshold just below or equal to usage.
3593 * If it's not true, a threshold was crossed after last
3594 * call of __mem_cgroup_threshold().
3596 i = t->current_threshold;
3599 * Iterate backward over array of thresholds starting from
3600 * current_threshold and check if a threshold is crossed.
3601 * If none of thresholds below usage is crossed, we read
3602 * only one element of the array here.
3604 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3605 eventfd_signal(t->entries[i].eventfd, 1);
3607 /* i = current_threshold + 1 */
3611 * Iterate forward over array of thresholds starting from
3612 * current_threshold+1 and check if a threshold is crossed.
3613 * If none of thresholds above usage is crossed, we read
3614 * only one element of the array here.
3616 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3617 eventfd_signal(t->entries[i].eventfd, 1);
3619 /* Update current_threshold */
3620 t->current_threshold = i - 1;
3625 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3628 __mem_cgroup_threshold(memcg, false);
3629 if (do_swap_account)
3630 __mem_cgroup_threshold(memcg, true);
3632 memcg = parent_mem_cgroup(memcg);
3636 static int compare_thresholds(const void *a, const void *b)
3638 const struct mem_cgroup_threshold *_a = a;
3639 const struct mem_cgroup_threshold *_b = b;
3641 if (_a->threshold > _b->threshold)
3644 if (_a->threshold < _b->threshold)
3650 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3652 struct mem_cgroup_eventfd_list *ev;
3654 spin_lock(&memcg_oom_lock);
3656 list_for_each_entry(ev, &memcg->oom_notify, list)
3657 eventfd_signal(ev->eventfd, 1);
3659 spin_unlock(&memcg_oom_lock);
3663 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3665 struct mem_cgroup *iter;
3667 for_each_mem_cgroup_tree(iter, memcg)
3668 mem_cgroup_oom_notify_cb(iter);
3671 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3672 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3674 struct mem_cgroup_thresholds *thresholds;
3675 struct mem_cgroup_threshold_ary *new;
3676 unsigned long threshold;
3677 unsigned long usage;
3680 ret = page_counter_memparse(args, "-1", &threshold);
3684 mutex_lock(&memcg->thresholds_lock);
3687 thresholds = &memcg->thresholds;
3688 usage = mem_cgroup_usage(memcg, false);
3689 } else if (type == _MEMSWAP) {
3690 thresholds = &memcg->memsw_thresholds;
3691 usage = mem_cgroup_usage(memcg, true);
3695 /* Check if a threshold crossed before adding a new one */
3696 if (thresholds->primary)
3697 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3699 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3701 /* Allocate memory for new array of thresholds */
3702 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3710 /* Copy thresholds (if any) to new array */
3711 if (thresholds->primary) {
3712 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3713 sizeof(struct mem_cgroup_threshold));
3716 /* Add new threshold */
3717 new->entries[size - 1].eventfd = eventfd;
3718 new->entries[size - 1].threshold = threshold;
3720 /* Sort thresholds. Registering of new threshold isn't time-critical */
3721 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3722 compare_thresholds, NULL);
3724 /* Find current threshold */
3725 new->current_threshold = -1;
3726 for (i = 0; i < size; i++) {
3727 if (new->entries[i].threshold <= usage) {
3729 * new->current_threshold will not be used until
3730 * rcu_assign_pointer(), so it's safe to increment
3733 ++new->current_threshold;
3738 /* Free old spare buffer and save old primary buffer as spare */
3739 kfree(thresholds->spare);
3740 thresholds->spare = thresholds->primary;
3742 rcu_assign_pointer(thresholds->primary, new);
3744 /* To be sure that nobody uses thresholds */
3748 mutex_unlock(&memcg->thresholds_lock);
3753 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3754 struct eventfd_ctx *eventfd, const char *args)
3756 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
3759 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
3760 struct eventfd_ctx *eventfd, const char *args)
3762 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
3765 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3766 struct eventfd_ctx *eventfd, enum res_type type)
3768 struct mem_cgroup_thresholds *thresholds;
3769 struct mem_cgroup_threshold_ary *new;
3770 unsigned long usage;
3773 mutex_lock(&memcg->thresholds_lock);
3776 thresholds = &memcg->thresholds;
3777 usage = mem_cgroup_usage(memcg, false);
3778 } else if (type == _MEMSWAP) {
3779 thresholds = &memcg->memsw_thresholds;
3780 usage = mem_cgroup_usage(memcg, true);
3784 if (!thresholds->primary)
3787 /* Check if a threshold crossed before removing */
3788 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3790 /* Calculate new number of threshold */
3792 for (i = 0; i < thresholds->primary->size; i++) {
3793 if (thresholds->primary->entries[i].eventfd != eventfd)
3797 new = thresholds->spare;
3799 /* Set thresholds array to NULL if we don't have thresholds */
3808 /* Copy thresholds and find current threshold */
3809 new->current_threshold = -1;
3810 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3811 if (thresholds->primary->entries[i].eventfd == eventfd)
3814 new->entries[j] = thresholds->primary->entries[i];
3815 if (new->entries[j].threshold <= usage) {
3817 * new->current_threshold will not be used
3818 * until rcu_assign_pointer(), so it's safe to increment
3821 ++new->current_threshold;
3827 /* Swap primary and spare array */
3828 thresholds->spare = thresholds->primary;
3829 /* If all events are unregistered, free the spare array */
3831 kfree(thresholds->spare);
3832 thresholds->spare = NULL;
3835 rcu_assign_pointer(thresholds->primary, new);
3837 /* To be sure that nobody uses thresholds */
3840 mutex_unlock(&memcg->thresholds_lock);
3843 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3844 struct eventfd_ctx *eventfd)
3846 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
3849 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3850 struct eventfd_ctx *eventfd)
3852 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
3855 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
3856 struct eventfd_ctx *eventfd, const char *args)
3858 struct mem_cgroup_eventfd_list *event;
3860 event = kmalloc(sizeof(*event), GFP_KERNEL);
3864 spin_lock(&memcg_oom_lock);
3866 event->eventfd = eventfd;
3867 list_add(&event->list, &memcg->oom_notify);
3869 /* already in OOM ? */
3870 if (atomic_read(&memcg->under_oom))
3871 eventfd_signal(eventfd, 1);
3872 spin_unlock(&memcg_oom_lock);
3877 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
3878 struct eventfd_ctx *eventfd)
3880 struct mem_cgroup_eventfd_list *ev, *tmp;
3882 spin_lock(&memcg_oom_lock);
3884 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
3885 if (ev->eventfd == eventfd) {
3886 list_del(&ev->list);
3891 spin_unlock(&memcg_oom_lock);
3894 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
3896 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
3898 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
3899 seq_printf(sf, "under_oom %d\n", (bool)atomic_read(&memcg->under_oom));
3903 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
3904 struct cftype *cft, u64 val)
3906 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3908 /* cannot set to root cgroup and only 0 and 1 are allowed */
3909 if (!css->parent || !((val == 0) || (val == 1)))
3912 memcg->oom_kill_disable = val;
3914 memcg_oom_recover(memcg);
3919 #ifdef CONFIG_MEMCG_KMEM
3920 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
3924 ret = memcg_propagate_kmem(memcg);
3928 return mem_cgroup_sockets_init(memcg, ss);
3931 static void memcg_deactivate_kmem(struct mem_cgroup *memcg)
3933 struct cgroup_subsys_state *css;
3934 struct mem_cgroup *parent, *child;
3937 if (!memcg->kmem_acct_active)
3941 * Clear the 'active' flag before clearing memcg_caches arrays entries.
3942 * Since we take the slab_mutex in memcg_deactivate_kmem_caches(), it
3943 * guarantees no cache will be created for this cgroup after we are
3944 * done (see memcg_create_kmem_cache()).
3946 memcg->kmem_acct_active = false;
3948 memcg_deactivate_kmem_caches(memcg);
3950 kmemcg_id = memcg->kmemcg_id;
3951 BUG_ON(kmemcg_id < 0);
3953 parent = parent_mem_cgroup(memcg);
3955 parent = root_mem_cgroup;
3958 * Change kmemcg_id of this cgroup and all its descendants to the
3959 * parent's id, and then move all entries from this cgroup's list_lrus
3960 * to ones of the parent. After we have finished, all list_lrus
3961 * corresponding to this cgroup are guaranteed to remain empty. The
3962 * ordering is imposed by list_lru_node->lock taken by
3963 * memcg_drain_all_list_lrus().
3965 css_for_each_descendant_pre(css, &memcg->css) {
3966 child = mem_cgroup_from_css(css);
3967 BUG_ON(child->kmemcg_id != kmemcg_id);
3968 child->kmemcg_id = parent->kmemcg_id;
3969 if (!memcg->use_hierarchy)
3972 memcg_drain_all_list_lrus(kmemcg_id, parent->kmemcg_id);
3974 memcg_free_cache_id(kmemcg_id);
3977 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
3979 if (memcg->kmem_acct_activated) {
3980 memcg_destroy_kmem_caches(memcg);
3981 static_key_slow_dec(&memcg_kmem_enabled_key);
3982 WARN_ON(page_counter_read(&memcg->kmem));
3984 mem_cgroup_sockets_destroy(memcg);
3987 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
3992 static void memcg_deactivate_kmem(struct mem_cgroup *memcg)
3996 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
4002 * DO NOT USE IN NEW FILES.
4004 * "cgroup.event_control" implementation.
4006 * This is way over-engineered. It tries to support fully configurable
4007 * events for each user. Such level of flexibility is completely
4008 * unnecessary especially in the light of the planned unified hierarchy.
4010 * Please deprecate this and replace with something simpler if at all
4015 * Unregister event and free resources.
4017 * Gets called from workqueue.
4019 static void memcg_event_remove(struct work_struct *work)
4021 struct mem_cgroup_event *event =
4022 container_of(work, struct mem_cgroup_event, remove);
4023 struct mem_cgroup *memcg = event->memcg;
4025 remove_wait_queue(event->wqh, &event->wait);
4027 event->unregister_event(memcg, event->eventfd);
4029 /* Notify userspace the event is going away. */
4030 eventfd_signal(event->eventfd, 1);
4032 eventfd_ctx_put(event->eventfd);
4034 css_put(&memcg->css);
4038 * Gets called on POLLHUP on eventfd when user closes it.
4040 * Called with wqh->lock held and interrupts disabled.
4042 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
4043 int sync, void *key)
4045 struct mem_cgroup_event *event =
4046 container_of(wait, struct mem_cgroup_event, wait);
4047 struct mem_cgroup *memcg = event->memcg;
4048 unsigned long flags = (unsigned long)key;
4050 if (flags & POLLHUP) {
4052 * If the event has been detached at cgroup removal, we
4053 * can simply return knowing the other side will cleanup
4056 * We can't race against event freeing since the other
4057 * side will require wqh->lock via remove_wait_queue(),
4060 spin_lock(&memcg->event_list_lock);
4061 if (!list_empty(&event->list)) {
4062 list_del_init(&event->list);
4064 * We are in atomic context, but cgroup_event_remove()
4065 * may sleep, so we have to call it in workqueue.
4067 schedule_work(&event->remove);
4069 spin_unlock(&memcg->event_list_lock);
4075 static void memcg_event_ptable_queue_proc(struct file *file,
4076 wait_queue_head_t *wqh, poll_table *pt)
4078 struct mem_cgroup_event *event =
4079 container_of(pt, struct mem_cgroup_event, pt);
4082 add_wait_queue(wqh, &event->wait);
4086 * DO NOT USE IN NEW FILES.
4088 * Parse input and register new cgroup event handler.
4090 * Input must be in format '<event_fd> <control_fd> <args>'.
4091 * Interpretation of args is defined by control file implementation.
4093 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4094 char *buf, size_t nbytes, loff_t off)
4096 struct cgroup_subsys_state *css = of_css(of);
4097 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4098 struct mem_cgroup_event *event;
4099 struct cgroup_subsys_state *cfile_css;
4100 unsigned int efd, cfd;
4107 buf = strstrip(buf);
4109 efd = simple_strtoul(buf, &endp, 10);
4114 cfd = simple_strtoul(buf, &endp, 10);
4115 if ((*endp != ' ') && (*endp != '\0'))
4119 event = kzalloc(sizeof(*event), GFP_KERNEL);
4123 event->memcg = memcg;
4124 INIT_LIST_HEAD(&event->list);
4125 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4126 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4127 INIT_WORK(&event->remove, memcg_event_remove);
4135 event->eventfd = eventfd_ctx_fileget(efile.file);
4136 if (IS_ERR(event->eventfd)) {
4137 ret = PTR_ERR(event->eventfd);
4144 goto out_put_eventfd;
4147 /* the process need read permission on control file */
4148 /* AV: shouldn't we check that it's been opened for read instead? */
4149 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4154 * Determine the event callbacks and set them in @event. This used
4155 * to be done via struct cftype but cgroup core no longer knows
4156 * about these events. The following is crude but the whole thing
4157 * is for compatibility anyway.
4159 * DO NOT ADD NEW FILES.
4161 name = cfile.file->f_path.dentry->d_name.name;
4163 if (!strcmp(name, "memory.usage_in_bytes")) {
4164 event->register_event = mem_cgroup_usage_register_event;
4165 event->unregister_event = mem_cgroup_usage_unregister_event;
4166 } else if (!strcmp(name, "memory.oom_control")) {
4167 event->register_event = mem_cgroup_oom_register_event;
4168 event->unregister_event = mem_cgroup_oom_unregister_event;
4169 } else if (!strcmp(name, "memory.pressure_level")) {
4170 event->register_event = vmpressure_register_event;
4171 event->unregister_event = vmpressure_unregister_event;
4172 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4173 event->register_event = memsw_cgroup_usage_register_event;
4174 event->unregister_event = memsw_cgroup_usage_unregister_event;
4181 * Verify @cfile should belong to @css. Also, remaining events are
4182 * automatically removed on cgroup destruction but the removal is
4183 * asynchronous, so take an extra ref on @css.
4185 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4186 &memory_cgrp_subsys);
4188 if (IS_ERR(cfile_css))
4190 if (cfile_css != css) {
4195 ret = event->register_event(memcg, event->eventfd, buf);
4199 efile.file->f_op->poll(efile.file, &event->pt);
4201 spin_lock(&memcg->event_list_lock);
4202 list_add(&event->list, &memcg->event_list);
4203 spin_unlock(&memcg->event_list_lock);
4215 eventfd_ctx_put(event->eventfd);
4224 static struct cftype mem_cgroup_legacy_files[] = {
4226 .name = "usage_in_bytes",
4227 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4228 .read_u64 = mem_cgroup_read_u64,
4231 .name = "max_usage_in_bytes",
4232 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4233 .write = mem_cgroup_reset,
4234 .read_u64 = mem_cgroup_read_u64,
4237 .name = "limit_in_bytes",
4238 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4239 .write = mem_cgroup_write,
4240 .read_u64 = mem_cgroup_read_u64,
4243 .name = "soft_limit_in_bytes",
4244 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4245 .write = mem_cgroup_write,
4246 .read_u64 = mem_cgroup_read_u64,
4250 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4251 .write = mem_cgroup_reset,
4252 .read_u64 = mem_cgroup_read_u64,
4256 .seq_show = memcg_stat_show,
4259 .name = "force_empty",
4260 .write = mem_cgroup_force_empty_write,
4263 .name = "use_hierarchy",
4264 .write_u64 = mem_cgroup_hierarchy_write,
4265 .read_u64 = mem_cgroup_hierarchy_read,
4268 .name = "cgroup.event_control", /* XXX: for compat */
4269 .write = memcg_write_event_control,
4270 .flags = CFTYPE_NO_PREFIX,
4274 .name = "swappiness",
4275 .read_u64 = mem_cgroup_swappiness_read,
4276 .write_u64 = mem_cgroup_swappiness_write,
4279 .name = "move_charge_at_immigrate",
4280 .read_u64 = mem_cgroup_move_charge_read,
4281 .write_u64 = mem_cgroup_move_charge_write,
4284 .name = "oom_control",
4285 .seq_show = mem_cgroup_oom_control_read,
4286 .write_u64 = mem_cgroup_oom_control_write,
4287 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4290 .name = "pressure_level",
4294 .name = "numa_stat",
4295 .seq_show = memcg_numa_stat_show,
4298 #ifdef CONFIG_MEMCG_KMEM
4300 .name = "kmem.limit_in_bytes",
4301 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4302 .write = mem_cgroup_write,
4303 .read_u64 = mem_cgroup_read_u64,
4306 .name = "kmem.usage_in_bytes",
4307 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4308 .read_u64 = mem_cgroup_read_u64,
4311 .name = "kmem.failcnt",
4312 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4313 .write = mem_cgroup_reset,
4314 .read_u64 = mem_cgroup_read_u64,
4317 .name = "kmem.max_usage_in_bytes",
4318 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4319 .write = mem_cgroup_reset,
4320 .read_u64 = mem_cgroup_read_u64,
4322 #ifdef CONFIG_SLABINFO
4324 .name = "kmem.slabinfo",
4325 .seq_start = slab_start,
4326 .seq_next = slab_next,
4327 .seq_stop = slab_stop,
4328 .seq_show = memcg_slab_show,
4332 { }, /* terminate */
4335 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4337 struct mem_cgroup_per_node *pn;
4338 struct mem_cgroup_per_zone *mz;
4339 int zone, tmp = node;
4341 * This routine is called against possible nodes.
4342 * But it's BUG to call kmalloc() against offline node.
4344 * TODO: this routine can waste much memory for nodes which will
4345 * never be onlined. It's better to use memory hotplug callback
4348 if (!node_state(node, N_NORMAL_MEMORY))
4350 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4354 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4355 mz = &pn->zoneinfo[zone];
4356 lruvec_init(&mz->lruvec);
4357 mz->usage_in_excess = 0;
4358 mz->on_tree = false;
4361 memcg->nodeinfo[node] = pn;
4365 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4367 kfree(memcg->nodeinfo[node]);
4370 static struct mem_cgroup *mem_cgroup_alloc(void)
4372 struct mem_cgroup *memcg;
4375 size = sizeof(struct mem_cgroup);
4376 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4378 memcg = kzalloc(size, GFP_KERNEL);
4382 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4385 spin_lock_init(&memcg->pcp_counter_lock);
4394 * At destroying mem_cgroup, references from swap_cgroup can remain.
4395 * (scanning all at force_empty is too costly...)
4397 * Instead of clearing all references at force_empty, we remember
4398 * the number of reference from swap_cgroup and free mem_cgroup when
4399 * it goes down to 0.
4401 * Removal of cgroup itself succeeds regardless of refs from swap.
4404 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4408 mem_cgroup_remove_from_trees(memcg);
4411 free_mem_cgroup_per_zone_info(memcg, node);
4413 free_percpu(memcg->stat);
4418 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4420 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4422 if (!memcg->memory.parent)
4424 return mem_cgroup_from_counter(memcg->memory.parent, memory);
4426 EXPORT_SYMBOL(parent_mem_cgroup);
4428 static struct cgroup_subsys_state * __ref
4429 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4431 struct mem_cgroup *memcg;
4432 long error = -ENOMEM;
4435 memcg = mem_cgroup_alloc();
4437 return ERR_PTR(error);
4440 if (alloc_mem_cgroup_per_zone_info(memcg, node))
4444 if (parent_css == NULL) {
4445 root_mem_cgroup = memcg;
4446 page_counter_init(&memcg->memory, NULL);
4447 memcg->high = PAGE_COUNTER_MAX;
4448 memcg->soft_limit = PAGE_COUNTER_MAX;
4449 page_counter_init(&memcg->memsw, NULL);
4450 page_counter_init(&memcg->kmem, NULL);
4453 memcg->last_scanned_node = MAX_NUMNODES;
4454 INIT_LIST_HEAD(&memcg->oom_notify);
4455 memcg->move_charge_at_immigrate = 0;
4456 mutex_init(&memcg->thresholds_lock);
4457 spin_lock_init(&memcg->move_lock);
4458 vmpressure_init(&memcg->vmpressure);
4459 INIT_LIST_HEAD(&memcg->event_list);
4460 spin_lock_init(&memcg->event_list_lock);
4461 #ifdef CONFIG_MEMCG_KMEM
4462 memcg->kmemcg_id = -1;
4468 __mem_cgroup_free(memcg);
4469 return ERR_PTR(error);
4473 mem_cgroup_css_online(struct cgroup_subsys_state *css)
4475 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4476 struct mem_cgroup *parent = mem_cgroup_from_css(css->parent);
4479 if (css->id > MEM_CGROUP_ID_MAX)
4485 mutex_lock(&memcg_create_mutex);
4487 memcg->use_hierarchy = parent->use_hierarchy;
4488 memcg->oom_kill_disable = parent->oom_kill_disable;
4489 memcg->swappiness = mem_cgroup_swappiness(parent);
4491 if (parent->use_hierarchy) {
4492 page_counter_init(&memcg->memory, &parent->memory);
4493 memcg->high = PAGE_COUNTER_MAX;
4494 memcg->soft_limit = PAGE_COUNTER_MAX;
4495 page_counter_init(&memcg->memsw, &parent->memsw);
4496 page_counter_init(&memcg->kmem, &parent->kmem);
4499 * No need to take a reference to the parent because cgroup
4500 * core guarantees its existence.
4503 page_counter_init(&memcg->memory, NULL);
4504 memcg->high = PAGE_COUNTER_MAX;
4505 memcg->soft_limit = PAGE_COUNTER_MAX;
4506 page_counter_init(&memcg->memsw, NULL);
4507 page_counter_init(&memcg->kmem, NULL);
4509 * Deeper hierachy with use_hierarchy == false doesn't make
4510 * much sense so let cgroup subsystem know about this
4511 * unfortunate state in our controller.
4513 if (parent != root_mem_cgroup)
4514 memory_cgrp_subsys.broken_hierarchy = true;
4516 mutex_unlock(&memcg_create_mutex);
4518 ret = memcg_init_kmem(memcg, &memory_cgrp_subsys);
4523 * Make sure the memcg is initialized: mem_cgroup_iter()
4524 * orders reading memcg->initialized against its callers
4525 * reading the memcg members.
4527 smp_store_release(&memcg->initialized, 1);
4532 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4534 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4535 struct mem_cgroup_event *event, *tmp;
4538 * Unregister events and notify userspace.
4539 * Notify userspace about cgroup removing only after rmdir of cgroup
4540 * directory to avoid race between userspace and kernelspace.
4542 spin_lock(&memcg->event_list_lock);
4543 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4544 list_del_init(&event->list);
4545 schedule_work(&event->remove);
4547 spin_unlock(&memcg->event_list_lock);
4549 vmpressure_cleanup(&memcg->vmpressure);
4551 memcg_deactivate_kmem(memcg);
4554 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4556 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4558 memcg_destroy_kmem(memcg);
4559 __mem_cgroup_free(memcg);
4563 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4564 * @css: the target css
4566 * Reset the states of the mem_cgroup associated with @css. This is
4567 * invoked when the userland requests disabling on the default hierarchy
4568 * but the memcg is pinned through dependency. The memcg should stop
4569 * applying policies and should revert to the vanilla state as it may be
4570 * made visible again.
4572 * The current implementation only resets the essential configurations.
4573 * This needs to be expanded to cover all the visible parts.
4575 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4577 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4579 mem_cgroup_resize_limit(memcg, PAGE_COUNTER_MAX);
4580 mem_cgroup_resize_memsw_limit(memcg, PAGE_COUNTER_MAX);
4581 memcg_update_kmem_limit(memcg, PAGE_COUNTER_MAX);
4583 memcg->high = PAGE_COUNTER_MAX;
4584 memcg->soft_limit = PAGE_COUNTER_MAX;
4588 /* Handlers for move charge at task migration. */
4589 static int mem_cgroup_do_precharge(unsigned long count)
4593 /* Try a single bulk charge without reclaim first */
4594 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_WAIT, count);
4596 mc.precharge += count;
4599 if (ret == -EINTR) {
4600 cancel_charge(root_mem_cgroup, count);
4604 /* Try charges one by one with reclaim */
4606 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_NORETRY, 1);
4608 * In case of failure, any residual charges against
4609 * mc.to will be dropped by mem_cgroup_clear_mc()
4610 * later on. However, cancel any charges that are
4611 * bypassed to root right away or they'll be lost.
4614 cancel_charge(root_mem_cgroup, 1);
4624 * get_mctgt_type - get target type of moving charge
4625 * @vma: the vma the pte to be checked belongs
4626 * @addr: the address corresponding to the pte to be checked
4627 * @ptent: the pte to be checked
4628 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4631 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4632 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4633 * move charge. if @target is not NULL, the page is stored in target->page
4634 * with extra refcnt got(Callers should handle it).
4635 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4636 * target for charge migration. if @target is not NULL, the entry is stored
4639 * Called with pte lock held.
4646 enum mc_target_type {
4652 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4653 unsigned long addr, pte_t ptent)
4655 struct page *page = vm_normal_page(vma, addr, ptent);
4657 if (!page || !page_mapped(page))
4659 if (PageAnon(page)) {
4660 if (!(mc.flags & MOVE_ANON))
4663 if (!(mc.flags & MOVE_FILE))
4666 if (!get_page_unless_zero(page))
4673 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4674 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4676 struct page *page = NULL;
4677 swp_entry_t ent = pte_to_swp_entry(ptent);
4679 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
4682 * Because lookup_swap_cache() updates some statistics counter,
4683 * we call find_get_page() with swapper_space directly.
4685 page = find_get_page(swap_address_space(ent), ent.val);
4686 if (do_swap_account)
4687 entry->val = ent.val;
4692 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4693 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4699 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4700 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4702 struct page *page = NULL;
4703 struct address_space *mapping;
4706 if (!vma->vm_file) /* anonymous vma */
4708 if (!(mc.flags & MOVE_FILE))
4711 mapping = vma->vm_file->f_mapping;
4712 pgoff = linear_page_index(vma, addr);
4714 /* page is moved even if it's not RSS of this task(page-faulted). */
4716 /* shmem/tmpfs may report page out on swap: account for that too. */
4717 if (shmem_mapping(mapping)) {
4718 page = find_get_entry(mapping, pgoff);
4719 if (radix_tree_exceptional_entry(page)) {
4720 swp_entry_t swp = radix_to_swp_entry(page);
4721 if (do_swap_account)
4723 page = find_get_page(swap_address_space(swp), swp.val);
4726 page = find_get_page(mapping, pgoff);
4728 page = find_get_page(mapping, pgoff);
4734 * mem_cgroup_move_account - move account of the page
4736 * @nr_pages: number of regular pages (>1 for huge pages)
4737 * @from: mem_cgroup which the page is moved from.
4738 * @to: mem_cgroup which the page is moved to. @from != @to.
4740 * The caller must confirm following.
4741 * - page is not on LRU (isolate_page() is useful.)
4742 * - compound_lock is held when nr_pages > 1
4744 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4747 static int mem_cgroup_move_account(struct page *page,
4748 unsigned int nr_pages,
4749 struct mem_cgroup *from,
4750 struct mem_cgroup *to)
4752 unsigned long flags;
4755 VM_BUG_ON(from == to);
4756 VM_BUG_ON_PAGE(PageLRU(page), page);
4758 * The page is isolated from LRU. So, collapse function
4759 * will not handle this page. But page splitting can happen.
4760 * Do this check under compound_page_lock(). The caller should
4764 if (nr_pages > 1 && !PageTransHuge(page))
4768 * Prevent mem_cgroup_migrate() from looking at page->mem_cgroup
4769 * of its source page while we change it: page migration takes
4770 * both pages off the LRU, but page cache replacement doesn't.
4772 if (!trylock_page(page))
4776 if (page->mem_cgroup != from)
4779 spin_lock_irqsave(&from->move_lock, flags);
4781 if (!PageAnon(page) && page_mapped(page)) {
4782 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4784 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4788 if (PageWriteback(page)) {
4789 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4791 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4796 * It is safe to change page->mem_cgroup here because the page
4797 * is referenced, charged, and isolated - we can't race with
4798 * uncharging, charging, migration, or LRU putback.
4801 /* caller should have done css_get */
4802 page->mem_cgroup = to;
4803 spin_unlock_irqrestore(&from->move_lock, flags);
4807 local_irq_disable();
4808 mem_cgroup_charge_statistics(to, page, nr_pages);
4809 memcg_check_events(to, page);
4810 mem_cgroup_charge_statistics(from, page, -nr_pages);
4811 memcg_check_events(from, page);
4819 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
4820 unsigned long addr, pte_t ptent, union mc_target *target)
4822 struct page *page = NULL;
4823 enum mc_target_type ret = MC_TARGET_NONE;
4824 swp_entry_t ent = { .val = 0 };
4826 if (pte_present(ptent))
4827 page = mc_handle_present_pte(vma, addr, ptent);
4828 else if (is_swap_pte(ptent))
4829 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
4830 else if (pte_none(ptent))
4831 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4833 if (!page && !ent.val)
4837 * Do only loose check w/o serialization.
4838 * mem_cgroup_move_account() checks the page is valid or
4839 * not under LRU exclusion.
4841 if (page->mem_cgroup == mc.from) {
4842 ret = MC_TARGET_PAGE;
4844 target->page = page;
4846 if (!ret || !target)
4849 /* There is a swap entry and a page doesn't exist or isn't charged */
4850 if (ent.val && !ret &&
4851 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
4852 ret = MC_TARGET_SWAP;
4859 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4861 * We don't consider swapping or file mapped pages because THP does not
4862 * support them for now.
4863 * Caller should make sure that pmd_trans_huge(pmd) is true.
4865 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4866 unsigned long addr, pmd_t pmd, union mc_target *target)
4868 struct page *page = NULL;
4869 enum mc_target_type ret = MC_TARGET_NONE;
4871 page = pmd_page(pmd);
4872 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
4873 if (!(mc.flags & MOVE_ANON))
4875 if (page->mem_cgroup == mc.from) {
4876 ret = MC_TARGET_PAGE;
4879 target->page = page;
4885 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4886 unsigned long addr, pmd_t pmd, union mc_target *target)
4888 return MC_TARGET_NONE;
4892 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4893 unsigned long addr, unsigned long end,
4894 struct mm_walk *walk)
4896 struct vm_area_struct *vma = walk->vma;
4900 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
4901 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
4902 mc.precharge += HPAGE_PMD_NR;
4907 if (pmd_trans_unstable(pmd))
4909 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4910 for (; addr != end; pte++, addr += PAGE_SIZE)
4911 if (get_mctgt_type(vma, addr, *pte, NULL))
4912 mc.precharge++; /* increment precharge temporarily */
4913 pte_unmap_unlock(pte - 1, ptl);
4919 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4921 unsigned long precharge;
4923 struct mm_walk mem_cgroup_count_precharge_walk = {
4924 .pmd_entry = mem_cgroup_count_precharge_pte_range,
4927 down_read(&mm->mmap_sem);
4928 walk_page_range(0, ~0UL, &mem_cgroup_count_precharge_walk);
4929 up_read(&mm->mmap_sem);
4931 precharge = mc.precharge;
4937 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4939 unsigned long precharge = mem_cgroup_count_precharge(mm);
4941 VM_BUG_ON(mc.moving_task);
4942 mc.moving_task = current;
4943 return mem_cgroup_do_precharge(precharge);
4946 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4947 static void __mem_cgroup_clear_mc(void)
4949 struct mem_cgroup *from = mc.from;
4950 struct mem_cgroup *to = mc.to;
4952 /* we must uncharge all the leftover precharges from mc.to */
4954 cancel_charge(mc.to, mc.precharge);
4958 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4959 * we must uncharge here.
4961 if (mc.moved_charge) {
4962 cancel_charge(mc.from, mc.moved_charge);
4963 mc.moved_charge = 0;
4965 /* we must fixup refcnts and charges */
4966 if (mc.moved_swap) {
4967 /* uncharge swap account from the old cgroup */
4968 if (!mem_cgroup_is_root(mc.from))
4969 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
4972 * we charged both to->memory and to->memsw, so we
4973 * should uncharge to->memory.
4975 if (!mem_cgroup_is_root(mc.to))
4976 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
4978 css_put_many(&mc.from->css, mc.moved_swap);
4980 /* we've already done css_get(mc.to) */
4983 memcg_oom_recover(from);
4984 memcg_oom_recover(to);
4985 wake_up_all(&mc.waitq);
4988 static void mem_cgroup_clear_mc(void)
4991 * we must clear moving_task before waking up waiters at the end of
4994 mc.moving_task = NULL;
4995 __mem_cgroup_clear_mc();
4996 spin_lock(&mc.lock);
4999 spin_unlock(&mc.lock);
5002 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
5003 struct cgroup_taskset *tset)
5005 struct task_struct *p = cgroup_taskset_first(tset);
5007 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5008 unsigned long move_flags;
5011 * We are now commited to this value whatever it is. Changes in this
5012 * tunable will only affect upcoming migrations, not the current one.
5013 * So we need to save it, and keep it going.
5015 move_flags = ACCESS_ONCE(memcg->move_charge_at_immigrate);
5017 struct mm_struct *mm;
5018 struct mem_cgroup *from = mem_cgroup_from_task(p);
5020 VM_BUG_ON(from == memcg);
5022 mm = get_task_mm(p);
5025 /* We move charges only when we move a owner of the mm */
5026 if (mm->owner == p) {
5029 VM_BUG_ON(mc.precharge);
5030 VM_BUG_ON(mc.moved_charge);
5031 VM_BUG_ON(mc.moved_swap);
5033 spin_lock(&mc.lock);
5036 mc.flags = move_flags;
5037 spin_unlock(&mc.lock);
5038 /* We set mc.moving_task later */
5040 ret = mem_cgroup_precharge_mc(mm);
5042 mem_cgroup_clear_mc();
5049 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
5050 struct cgroup_taskset *tset)
5053 mem_cgroup_clear_mc();
5056 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5057 unsigned long addr, unsigned long end,
5058 struct mm_walk *walk)
5061 struct vm_area_struct *vma = walk->vma;
5064 enum mc_target_type target_type;
5065 union mc_target target;
5069 * We don't take compound_lock() here but no race with splitting thp
5071 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
5072 * under splitting, which means there's no concurrent thp split,
5073 * - if another thread runs into split_huge_page() just after we
5074 * entered this if-block, the thread must wait for page table lock
5075 * to be unlocked in __split_huge_page_splitting(), where the main
5076 * part of thp split is not executed yet.
5078 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
5079 if (mc.precharge < HPAGE_PMD_NR) {
5083 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5084 if (target_type == MC_TARGET_PAGE) {
5086 if (!isolate_lru_page(page)) {
5087 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
5089 mc.precharge -= HPAGE_PMD_NR;
5090 mc.moved_charge += HPAGE_PMD_NR;
5092 putback_lru_page(page);
5100 if (pmd_trans_unstable(pmd))
5103 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5104 for (; addr != end; addr += PAGE_SIZE) {
5105 pte_t ptent = *(pte++);
5111 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5112 case MC_TARGET_PAGE:
5114 if (isolate_lru_page(page))
5116 if (!mem_cgroup_move_account(page, 1, mc.from, mc.to)) {
5118 /* we uncharge from mc.from later. */
5121 putback_lru_page(page);
5122 put: /* get_mctgt_type() gets the page */
5125 case MC_TARGET_SWAP:
5127 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5129 /* we fixup refcnts and charges later. */
5137 pte_unmap_unlock(pte - 1, ptl);
5142 * We have consumed all precharges we got in can_attach().
5143 * We try charge one by one, but don't do any additional
5144 * charges to mc.to if we have failed in charge once in attach()
5147 ret = mem_cgroup_do_precharge(1);
5155 static void mem_cgroup_move_charge(struct mm_struct *mm)
5157 struct mm_walk mem_cgroup_move_charge_walk = {
5158 .pmd_entry = mem_cgroup_move_charge_pte_range,
5162 lru_add_drain_all();
5164 * Signal mem_cgroup_begin_page_stat() to take the memcg's
5165 * move_lock while we're moving its pages to another memcg.
5166 * Then wait for already started RCU-only updates to finish.
5168 atomic_inc(&mc.from->moving_account);
5171 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5173 * Someone who are holding the mmap_sem might be waiting in
5174 * waitq. So we cancel all extra charges, wake up all waiters,
5175 * and retry. Because we cancel precharges, we might not be able
5176 * to move enough charges, but moving charge is a best-effort
5177 * feature anyway, so it wouldn't be a big problem.
5179 __mem_cgroup_clear_mc();
5184 * When we have consumed all precharges and failed in doing
5185 * additional charge, the page walk just aborts.
5187 walk_page_range(0, ~0UL, &mem_cgroup_move_charge_walk);
5188 up_read(&mm->mmap_sem);
5189 atomic_dec(&mc.from->moving_account);
5192 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
5193 struct cgroup_taskset *tset)
5195 struct task_struct *p = cgroup_taskset_first(tset);
5196 struct mm_struct *mm = get_task_mm(p);
5200 mem_cgroup_move_charge(mm);
5204 mem_cgroup_clear_mc();
5206 #else /* !CONFIG_MMU */
5207 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
5208 struct cgroup_taskset *tset)
5212 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
5213 struct cgroup_taskset *tset)
5216 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
5217 struct cgroup_taskset *tset)
5223 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5224 * to verify whether we're attached to the default hierarchy on each mount
5227 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5230 * use_hierarchy is forced on the default hierarchy. cgroup core
5231 * guarantees that @root doesn't have any children, so turning it
5232 * on for the root memcg is enough.
5234 if (cgroup_on_dfl(root_css->cgroup))
5235 root_mem_cgroup->use_hierarchy = true;
5237 root_mem_cgroup->use_hierarchy = false;
5240 static u64 memory_current_read(struct cgroup_subsys_state *css,
5243 return mem_cgroup_usage(mem_cgroup_from_css(css), false);
5246 static int memory_low_show(struct seq_file *m, void *v)
5248 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5249 unsigned long low = ACCESS_ONCE(memcg->low);
5251 if (low == PAGE_COUNTER_MAX)
5252 seq_puts(m, "max\n");
5254 seq_printf(m, "%llu\n", (u64)low * PAGE_SIZE);
5259 static ssize_t memory_low_write(struct kernfs_open_file *of,
5260 char *buf, size_t nbytes, loff_t off)
5262 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5266 buf = strstrip(buf);
5267 err = page_counter_memparse(buf, "max", &low);
5276 static int memory_high_show(struct seq_file *m, void *v)
5278 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5279 unsigned long high = ACCESS_ONCE(memcg->high);
5281 if (high == PAGE_COUNTER_MAX)
5282 seq_puts(m, "max\n");
5284 seq_printf(m, "%llu\n", (u64)high * PAGE_SIZE);
5289 static ssize_t memory_high_write(struct kernfs_open_file *of,
5290 char *buf, size_t nbytes, loff_t off)
5292 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5296 buf = strstrip(buf);
5297 err = page_counter_memparse(buf, "max", &high);
5306 static int memory_max_show(struct seq_file *m, void *v)
5308 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5309 unsigned long max = ACCESS_ONCE(memcg->memory.limit);
5311 if (max == PAGE_COUNTER_MAX)
5312 seq_puts(m, "max\n");
5314 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
5319 static ssize_t memory_max_write(struct kernfs_open_file *of,
5320 char *buf, size_t nbytes, loff_t off)
5322 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5326 buf = strstrip(buf);
5327 err = page_counter_memparse(buf, "max", &max);
5331 err = mem_cgroup_resize_limit(memcg, max);
5338 static int memory_events_show(struct seq_file *m, void *v)
5340 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5342 seq_printf(m, "low %lu\n", mem_cgroup_read_events(memcg, MEMCG_LOW));
5343 seq_printf(m, "high %lu\n", mem_cgroup_read_events(memcg, MEMCG_HIGH));
5344 seq_printf(m, "max %lu\n", mem_cgroup_read_events(memcg, MEMCG_MAX));
5345 seq_printf(m, "oom %lu\n", mem_cgroup_read_events(memcg, MEMCG_OOM));
5350 static struct cftype memory_files[] = {
5353 .read_u64 = memory_current_read,
5357 .flags = CFTYPE_NOT_ON_ROOT,
5358 .seq_show = memory_low_show,
5359 .write = memory_low_write,
5363 .flags = CFTYPE_NOT_ON_ROOT,
5364 .seq_show = memory_high_show,
5365 .write = memory_high_write,
5369 .flags = CFTYPE_NOT_ON_ROOT,
5370 .seq_show = memory_max_show,
5371 .write = memory_max_write,
5375 .flags = CFTYPE_NOT_ON_ROOT,
5376 .seq_show = memory_events_show,
5381 struct cgroup_subsys memory_cgrp_subsys = {
5382 .css_alloc = mem_cgroup_css_alloc,
5383 .css_online = mem_cgroup_css_online,
5384 .css_offline = mem_cgroup_css_offline,
5385 .css_free = mem_cgroup_css_free,
5386 .css_reset = mem_cgroup_css_reset,
5387 .can_attach = mem_cgroup_can_attach,
5388 .cancel_attach = mem_cgroup_cancel_attach,
5389 .attach = mem_cgroup_move_task,
5390 .bind = mem_cgroup_bind,
5391 .dfl_cftypes = memory_files,
5392 .legacy_cftypes = mem_cgroup_legacy_files,
5397 * mem_cgroup_events - count memory events against a cgroup
5398 * @memcg: the memory cgroup
5399 * @idx: the event index
5400 * @nr: the number of events to account for
5402 void mem_cgroup_events(struct mem_cgroup *memcg,
5403 enum mem_cgroup_events_index idx,
5406 this_cpu_add(memcg->stat->events[idx], nr);
5410 * mem_cgroup_low - check if memory consumption is below the normal range
5411 * @root: the highest ancestor to consider
5412 * @memcg: the memory cgroup to check
5414 * Returns %true if memory consumption of @memcg, and that of all
5415 * configurable ancestors up to @root, is below the normal range.
5417 bool mem_cgroup_low(struct mem_cgroup *root, struct mem_cgroup *memcg)
5419 if (mem_cgroup_disabled())
5423 * The toplevel group doesn't have a configurable range, so
5424 * it's never low when looked at directly, and it is not
5425 * considered an ancestor when assessing the hierarchy.
5428 if (memcg == root_mem_cgroup)
5431 if (page_counter_read(&memcg->memory) >= memcg->low)
5434 while (memcg != root) {
5435 memcg = parent_mem_cgroup(memcg);
5437 if (memcg == root_mem_cgroup)
5440 if (page_counter_read(&memcg->memory) >= memcg->low)
5447 * mem_cgroup_try_charge - try charging a page
5448 * @page: page to charge
5449 * @mm: mm context of the victim
5450 * @gfp_mask: reclaim mode
5451 * @memcgp: charged memcg return
5453 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5454 * pages according to @gfp_mask if necessary.
5456 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5457 * Otherwise, an error code is returned.
5459 * After page->mapping has been set up, the caller must finalize the
5460 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5461 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5463 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
5464 gfp_t gfp_mask, struct mem_cgroup **memcgp)
5466 struct mem_cgroup *memcg = NULL;
5467 unsigned int nr_pages = 1;
5470 if (mem_cgroup_disabled())
5473 if (PageSwapCache(page)) {
5475 * Every swap fault against a single page tries to charge the
5476 * page, bail as early as possible. shmem_unuse() encounters
5477 * already charged pages, too. The USED bit is protected by
5478 * the page lock, which serializes swap cache removal, which
5479 * in turn serializes uncharging.
5481 if (page->mem_cgroup)
5485 if (PageTransHuge(page)) {
5486 nr_pages <<= compound_order(page);
5487 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5490 if (do_swap_account && PageSwapCache(page))
5491 memcg = try_get_mem_cgroup_from_page(page);
5493 memcg = get_mem_cgroup_from_mm(mm);
5495 ret = try_charge(memcg, gfp_mask, nr_pages);
5497 css_put(&memcg->css);
5499 if (ret == -EINTR) {
5500 memcg = root_mem_cgroup;
5509 * mem_cgroup_commit_charge - commit a page charge
5510 * @page: page to charge
5511 * @memcg: memcg to charge the page to
5512 * @lrucare: page might be on LRU already
5514 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5515 * after page->mapping has been set up. This must happen atomically
5516 * as part of the page instantiation, i.e. under the page table lock
5517 * for anonymous pages, under the page lock for page and swap cache.
5519 * In addition, the page must not be on the LRU during the commit, to
5520 * prevent racing with task migration. If it might be, use @lrucare.
5522 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5524 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
5527 unsigned int nr_pages = 1;
5529 VM_BUG_ON_PAGE(!page->mapping, page);
5530 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
5532 if (mem_cgroup_disabled())
5535 * Swap faults will attempt to charge the same page multiple
5536 * times. But reuse_swap_page() might have removed the page
5537 * from swapcache already, so we can't check PageSwapCache().
5542 commit_charge(page, memcg, lrucare);
5544 if (PageTransHuge(page)) {
5545 nr_pages <<= compound_order(page);
5546 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5549 local_irq_disable();
5550 mem_cgroup_charge_statistics(memcg, page, nr_pages);
5551 memcg_check_events(memcg, page);
5554 if (do_swap_account && PageSwapCache(page)) {
5555 swp_entry_t entry = { .val = page_private(page) };
5557 * The swap entry might not get freed for a long time,
5558 * let's not wait for it. The page already received a
5559 * memory+swap charge, drop the swap entry duplicate.
5561 mem_cgroup_uncharge_swap(entry);
5566 * mem_cgroup_cancel_charge - cancel a page charge
5567 * @page: page to charge
5568 * @memcg: memcg to charge the page to
5570 * Cancel a charge transaction started by mem_cgroup_try_charge().
5572 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg)
5574 unsigned int nr_pages = 1;
5576 if (mem_cgroup_disabled())
5579 * Swap faults will attempt to charge the same page multiple
5580 * times. But reuse_swap_page() might have removed the page
5581 * from swapcache already, so we can't check PageSwapCache().
5586 if (PageTransHuge(page)) {
5587 nr_pages <<= compound_order(page);
5588 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5591 cancel_charge(memcg, nr_pages);
5594 static void uncharge_batch(struct mem_cgroup *memcg, unsigned long pgpgout,
5595 unsigned long nr_anon, unsigned long nr_file,
5596 unsigned long nr_huge, struct page *dummy_page)
5598 unsigned long nr_pages = nr_anon + nr_file;
5599 unsigned long flags;
5601 if (!mem_cgroup_is_root(memcg)) {
5602 page_counter_uncharge(&memcg->memory, nr_pages);
5603 if (do_swap_account)
5604 page_counter_uncharge(&memcg->memsw, nr_pages);
5605 memcg_oom_recover(memcg);
5608 local_irq_save(flags);
5609 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS], nr_anon);
5610 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_CACHE], nr_file);
5611 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE], nr_huge);
5612 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT], pgpgout);
5613 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
5614 memcg_check_events(memcg, dummy_page);
5615 local_irq_restore(flags);
5617 if (!mem_cgroup_is_root(memcg))
5618 css_put_many(&memcg->css, nr_pages);
5621 static void uncharge_list(struct list_head *page_list)
5623 struct mem_cgroup *memcg = NULL;
5624 unsigned long nr_anon = 0;
5625 unsigned long nr_file = 0;
5626 unsigned long nr_huge = 0;
5627 unsigned long pgpgout = 0;
5628 struct list_head *next;
5631 next = page_list->next;
5633 unsigned int nr_pages = 1;
5635 page = list_entry(next, struct page, lru);
5636 next = page->lru.next;
5638 VM_BUG_ON_PAGE(PageLRU(page), page);
5639 VM_BUG_ON_PAGE(page_count(page), page);
5641 if (!page->mem_cgroup)
5645 * Nobody should be changing or seriously looking at
5646 * page->mem_cgroup at this point, we have fully
5647 * exclusive access to the page.
5650 if (memcg != page->mem_cgroup) {
5652 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5654 pgpgout = nr_anon = nr_file = nr_huge = 0;
5656 memcg = page->mem_cgroup;
5659 if (PageTransHuge(page)) {
5660 nr_pages <<= compound_order(page);
5661 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5662 nr_huge += nr_pages;
5666 nr_anon += nr_pages;
5668 nr_file += nr_pages;
5670 page->mem_cgroup = NULL;
5673 } while (next != page_list);
5676 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5681 * mem_cgroup_uncharge - uncharge a page
5682 * @page: page to uncharge
5684 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5685 * mem_cgroup_commit_charge().
5687 void mem_cgroup_uncharge(struct page *page)
5689 if (mem_cgroup_disabled())
5692 /* Don't touch page->lru of any random page, pre-check: */
5693 if (!page->mem_cgroup)
5696 INIT_LIST_HEAD(&page->lru);
5697 uncharge_list(&page->lru);
5701 * mem_cgroup_uncharge_list - uncharge a list of page
5702 * @page_list: list of pages to uncharge
5704 * Uncharge a list of pages previously charged with
5705 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5707 void mem_cgroup_uncharge_list(struct list_head *page_list)
5709 if (mem_cgroup_disabled())
5712 if (!list_empty(page_list))
5713 uncharge_list(page_list);
5717 * mem_cgroup_migrate - migrate a charge to another page
5718 * @oldpage: currently charged page
5719 * @newpage: page to transfer the charge to
5720 * @lrucare: either or both pages might be on the LRU already
5722 * Migrate the charge from @oldpage to @newpage.
5724 * Both pages must be locked, @newpage->mapping must be set up.
5726 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage,
5729 struct mem_cgroup *memcg;
5732 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
5733 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
5734 VM_BUG_ON_PAGE(!lrucare && PageLRU(oldpage), oldpage);
5735 VM_BUG_ON_PAGE(!lrucare && PageLRU(newpage), newpage);
5736 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
5737 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
5740 if (mem_cgroup_disabled())
5743 /* Page cache replacement: new page already charged? */
5744 if (newpage->mem_cgroup)
5748 * Swapcache readahead pages can get migrated before being
5749 * charged, and migration from compaction can happen to an
5750 * uncharged page when the PFN walker finds a page that
5751 * reclaim just put back on the LRU but has not released yet.
5753 memcg = oldpage->mem_cgroup;
5758 lock_page_lru(oldpage, &isolated);
5760 oldpage->mem_cgroup = NULL;
5763 unlock_page_lru(oldpage, isolated);
5765 commit_charge(newpage, memcg, lrucare);
5769 * subsys_initcall() for memory controller.
5771 * Some parts like hotcpu_notifier() have to be initialized from this context
5772 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5773 * everything that doesn't depend on a specific mem_cgroup structure should
5774 * be initialized from here.
5776 static int __init mem_cgroup_init(void)
5780 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5782 for_each_possible_cpu(cpu)
5783 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
5786 for_each_node(node) {
5787 struct mem_cgroup_tree_per_node *rtpn;
5790 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
5791 node_online(node) ? node : NUMA_NO_NODE);
5793 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5794 struct mem_cgroup_tree_per_zone *rtpz;
5796 rtpz = &rtpn->rb_tree_per_zone[zone];
5797 rtpz->rb_root = RB_ROOT;
5798 spin_lock_init(&rtpz->lock);
5800 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5805 subsys_initcall(mem_cgroup_init);
5807 #ifdef CONFIG_MEMCG_SWAP
5809 * mem_cgroup_swapout - transfer a memsw charge to swap
5810 * @page: page whose memsw charge to transfer
5811 * @entry: swap entry to move the charge to
5813 * Transfer the memsw charge of @page to @entry.
5815 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
5817 struct mem_cgroup *memcg;
5818 unsigned short oldid;
5820 VM_BUG_ON_PAGE(PageLRU(page), page);
5821 VM_BUG_ON_PAGE(page_count(page), page);
5823 if (!do_swap_account)
5826 memcg = page->mem_cgroup;
5828 /* Readahead page, never charged */
5832 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg));
5833 VM_BUG_ON_PAGE(oldid, page);
5834 mem_cgroup_swap_statistics(memcg, true);
5836 page->mem_cgroup = NULL;
5838 if (!mem_cgroup_is_root(memcg))
5839 page_counter_uncharge(&memcg->memory, 1);
5841 /* XXX: caller holds IRQ-safe mapping->tree_lock */
5842 VM_BUG_ON(!irqs_disabled());
5844 mem_cgroup_charge_statistics(memcg, page, -1);
5845 memcg_check_events(memcg, page);
5849 * mem_cgroup_uncharge_swap - uncharge a swap entry
5850 * @entry: swap entry to uncharge
5852 * Drop the memsw charge associated with @entry.
5854 void mem_cgroup_uncharge_swap(swp_entry_t entry)
5856 struct mem_cgroup *memcg;
5859 if (!do_swap_account)
5862 id = swap_cgroup_record(entry, 0);
5864 memcg = mem_cgroup_from_id(id);
5866 if (!mem_cgroup_is_root(memcg))
5867 page_counter_uncharge(&memcg->memsw, 1);
5868 mem_cgroup_swap_statistics(memcg, false);
5869 css_put(&memcg->css);
5874 /* for remember boot option*/
5875 #ifdef CONFIG_MEMCG_SWAP_ENABLED
5876 static int really_do_swap_account __initdata = 1;
5878 static int really_do_swap_account __initdata;
5881 static int __init enable_swap_account(char *s)
5883 if (!strcmp(s, "1"))
5884 really_do_swap_account = 1;
5885 else if (!strcmp(s, "0"))
5886 really_do_swap_account = 0;
5889 __setup("swapaccount=", enable_swap_account);
5891 static struct cftype memsw_cgroup_files[] = {
5893 .name = "memsw.usage_in_bytes",
5894 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
5895 .read_u64 = mem_cgroup_read_u64,
5898 .name = "memsw.max_usage_in_bytes",
5899 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
5900 .write = mem_cgroup_reset,
5901 .read_u64 = mem_cgroup_read_u64,
5904 .name = "memsw.limit_in_bytes",
5905 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
5906 .write = mem_cgroup_write,
5907 .read_u64 = mem_cgroup_read_u64,
5910 .name = "memsw.failcnt",
5911 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
5912 .write = mem_cgroup_reset,
5913 .read_u64 = mem_cgroup_read_u64,
5915 { }, /* terminate */
5918 static int __init mem_cgroup_swap_init(void)
5920 if (!mem_cgroup_disabled() && really_do_swap_account) {
5921 do_swap_account = 1;
5922 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
5923 memsw_cgroup_files));
5927 subsys_initcall(mem_cgroup_swap_init);
5929 #endif /* CONFIG_MEMCG_SWAP */