1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * Kernel Memory Controller
14 * Copyright (C) 2012 Parallels Inc. and Google Inc.
15 * Authors: Glauber Costa and Suleiman Souhlal
17 * This program is free software; you can redistribute it and/or modify
18 * it under the terms of the GNU General Public License as published by
19 * the Free Software Foundation; either version 2 of the License, or
20 * (at your option) any later version.
22 * This program is distributed in the hope that it will be useful,
23 * but WITHOUT ANY WARRANTY; without even the implied warranty of
24 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
25 * GNU General Public License for more details.
28 #include <linux/page_counter.h>
29 #include <linux/memcontrol.h>
30 #include <linux/cgroup.h>
32 #include <linux/hugetlb.h>
33 #include <linux/pagemap.h>
34 #include <linux/smp.h>
35 #include <linux/page-flags.h>
36 #include <linux/backing-dev.h>
37 #include <linux/bit_spinlock.h>
38 #include <linux/rcupdate.h>
39 #include <linux/limits.h>
40 #include <linux/export.h>
41 #include <linux/mutex.h>
42 #include <linux/rbtree.h>
43 #include <linux/slab.h>
44 #include <linux/swap.h>
45 #include <linux/swapops.h>
46 #include <linux/spinlock.h>
47 #include <linux/eventfd.h>
48 #include <linux/poll.h>
49 #include <linux/sort.h>
51 #include <linux/seq_file.h>
52 #include <linux/vmpressure.h>
53 #include <linux/mm_inline.h>
54 #include <linux/swap_cgroup.h>
55 #include <linux/cpu.h>
56 #include <linux/oom.h>
57 #include <linux/lockdep.h>
58 #include <linux/file.h>
62 #include <net/tcp_memcontrol.h>
65 #include <asm/uaccess.h>
67 #include <trace/events/vmscan.h>
69 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
70 EXPORT_SYMBOL(memory_cgrp_subsys);
72 #define MEM_CGROUP_RECLAIM_RETRIES 5
73 static struct mem_cgroup *root_mem_cgroup __read_mostly;
75 /* Whether the swap controller is active */
76 #ifdef CONFIG_MEMCG_SWAP
77 int do_swap_account __read_mostly;
79 #define do_swap_account 0
82 static const char * const mem_cgroup_stat_names[] = {
91 static const char * const mem_cgroup_events_names[] = {
98 static const char * const mem_cgroup_lru_names[] = {
107 * Per memcg event counter is incremented at every pagein/pageout. With THP,
108 * it will be incremated by the number of pages. This counter is used for
109 * for trigger some periodic events. This is straightforward and better
110 * than using jiffies etc. to handle periodic memcg event.
112 enum mem_cgroup_events_target {
113 MEM_CGROUP_TARGET_THRESH,
114 MEM_CGROUP_TARGET_SOFTLIMIT,
115 MEM_CGROUP_TARGET_NUMAINFO,
118 #define THRESHOLDS_EVENTS_TARGET 128
119 #define SOFTLIMIT_EVENTS_TARGET 1024
120 #define NUMAINFO_EVENTS_TARGET 1024
122 struct mem_cgroup_stat_cpu {
123 long count[MEM_CGROUP_STAT_NSTATS];
124 unsigned long events[MEMCG_NR_EVENTS];
125 unsigned long nr_page_events;
126 unsigned long targets[MEM_CGROUP_NTARGETS];
129 struct reclaim_iter {
130 struct mem_cgroup *position;
131 /* scan generation, increased every round-trip */
132 unsigned int generation;
136 * per-zone information in memory controller.
138 struct mem_cgroup_per_zone {
139 struct lruvec lruvec;
140 unsigned long lru_size[NR_LRU_LISTS];
142 struct reclaim_iter iter[DEF_PRIORITY + 1];
144 struct rb_node tree_node; /* RB tree node */
145 unsigned long usage_in_excess;/* Set to the value by which */
146 /* the soft limit is exceeded*/
148 struct mem_cgroup *memcg; /* Back pointer, we cannot */
149 /* use container_of */
152 struct mem_cgroup_per_node {
153 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
157 * Cgroups above their limits are maintained in a RB-Tree, independent of
158 * their hierarchy representation
161 struct mem_cgroup_tree_per_zone {
162 struct rb_root rb_root;
166 struct mem_cgroup_tree_per_node {
167 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
170 struct mem_cgroup_tree {
171 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
174 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
176 struct mem_cgroup_threshold {
177 struct eventfd_ctx *eventfd;
178 unsigned long threshold;
182 struct mem_cgroup_threshold_ary {
183 /* An array index points to threshold just below or equal to usage. */
184 int current_threshold;
185 /* Size of entries[] */
187 /* Array of thresholds */
188 struct mem_cgroup_threshold entries[0];
191 struct mem_cgroup_thresholds {
192 /* Primary thresholds array */
193 struct mem_cgroup_threshold_ary *primary;
195 * Spare threshold array.
196 * This is needed to make mem_cgroup_unregister_event() "never fail".
197 * It must be able to store at least primary->size - 1 entries.
199 struct mem_cgroup_threshold_ary *spare;
203 struct mem_cgroup_eventfd_list {
204 struct list_head list;
205 struct eventfd_ctx *eventfd;
209 * cgroup_event represents events which userspace want to receive.
211 struct mem_cgroup_event {
213 * memcg which the event belongs to.
215 struct mem_cgroup *memcg;
217 * eventfd to signal userspace about the event.
219 struct eventfd_ctx *eventfd;
221 * Each of these stored in a list by the cgroup.
223 struct list_head list;
225 * register_event() callback will be used to add new userspace
226 * waiter for changes related to this event. Use eventfd_signal()
227 * on eventfd to send notification to userspace.
229 int (*register_event)(struct mem_cgroup *memcg,
230 struct eventfd_ctx *eventfd, const char *args);
232 * unregister_event() callback will be called when userspace closes
233 * the eventfd or on cgroup removing. This callback must be set,
234 * if you want provide notification functionality.
236 void (*unregister_event)(struct mem_cgroup *memcg,
237 struct eventfd_ctx *eventfd);
239 * All fields below needed to unregister event when
240 * userspace closes eventfd.
243 wait_queue_head_t *wqh;
245 struct work_struct remove;
248 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
249 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
252 * The memory controller data structure. The memory controller controls both
253 * page cache and RSS per cgroup. We would eventually like to provide
254 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
255 * to help the administrator determine what knobs to tune.
257 * TODO: Add a water mark for the memory controller. Reclaim will begin when
258 * we hit the water mark. May be even add a low water mark, such that
259 * no reclaim occurs from a cgroup at it's low water mark, this is
260 * a feature that will be implemented much later in the future.
263 struct cgroup_subsys_state css;
265 /* Accounted resources */
266 struct page_counter memory;
267 struct page_counter memsw;
268 struct page_counter kmem;
270 /* Normal memory consumption range */
274 unsigned long soft_limit;
276 /* vmpressure notifications */
277 struct vmpressure vmpressure;
279 /* css_online() has been completed */
283 * Should the accounting and control be hierarchical, per subtree?
289 atomic_t oom_wakeups;
292 /* OOM-Killer disable */
293 int oom_kill_disable;
295 /* protect arrays of thresholds */
296 struct mutex thresholds_lock;
298 /* thresholds for memory usage. RCU-protected */
299 struct mem_cgroup_thresholds thresholds;
301 /* thresholds for mem+swap usage. RCU-protected */
302 struct mem_cgroup_thresholds memsw_thresholds;
304 /* For oom notifier event fd */
305 struct list_head oom_notify;
308 * Should we move charges of a task when a task is moved into this
309 * mem_cgroup ? And what type of charges should we move ?
311 unsigned long move_charge_at_immigrate;
313 * set > 0 if pages under this cgroup are moving to other cgroup.
315 atomic_t moving_account;
316 /* taken only while moving_account > 0 */
317 spinlock_t move_lock;
318 struct task_struct *move_lock_task;
319 unsigned long move_lock_flags;
323 struct mem_cgroup_stat_cpu __percpu *stat;
325 * used when a cpu is offlined or other synchronizations
326 * See mem_cgroup_read_stat().
328 struct mem_cgroup_stat_cpu nocpu_base;
329 spinlock_t pcp_counter_lock;
331 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
332 struct cg_proto tcp_mem;
334 #if defined(CONFIG_MEMCG_KMEM)
335 /* Index in the kmem_cache->memcg_params->memcg_caches array */
339 int last_scanned_node;
341 nodemask_t scan_nodes;
342 atomic_t numainfo_events;
343 atomic_t numainfo_updating;
346 /* List of events which userspace want to receive */
347 struct list_head event_list;
348 spinlock_t event_list_lock;
350 struct mem_cgroup_per_node *nodeinfo[0];
351 /* WARNING: nodeinfo must be the last member here */
354 #ifdef CONFIG_MEMCG_KMEM
355 static bool memcg_kmem_is_active(struct mem_cgroup *memcg)
357 return memcg->kmemcg_id >= 0;
361 /* Stuffs for move charges at task migration. */
363 * Types of charges to be moved.
365 #define MOVE_ANON 0x1U
366 #define MOVE_FILE 0x2U
367 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
369 /* "mc" and its members are protected by cgroup_mutex */
370 static struct move_charge_struct {
371 spinlock_t lock; /* for from, to */
372 struct mem_cgroup *from;
373 struct mem_cgroup *to;
375 unsigned long precharge;
376 unsigned long moved_charge;
377 unsigned long moved_swap;
378 struct task_struct *moving_task; /* a task moving charges */
379 wait_queue_head_t waitq; /* a waitq for other context */
381 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
382 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
386 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
387 * limit reclaim to prevent infinite loops, if they ever occur.
389 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
390 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
393 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
394 MEM_CGROUP_CHARGE_TYPE_ANON,
395 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
396 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
400 /* for encoding cft->private value on file */
408 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
409 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
410 #define MEMFILE_ATTR(val) ((val) & 0xffff)
411 /* Used for OOM nofiier */
412 #define OOM_CONTROL (0)
415 * The memcg_create_mutex will be held whenever a new cgroup is created.
416 * As a consequence, any change that needs to protect against new child cgroups
417 * appearing has to hold it as well.
419 static DEFINE_MUTEX(memcg_create_mutex);
421 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
423 return s ? container_of(s, struct mem_cgroup, css) : NULL;
426 /* Some nice accessors for the vmpressure. */
427 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
430 memcg = root_mem_cgroup;
431 return &memcg->vmpressure;
434 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
436 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
439 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
441 return (memcg == root_mem_cgroup);
445 * We restrict the id in the range of [1, 65535], so it can fit into
448 #define MEM_CGROUP_ID_MAX USHRT_MAX
450 static inline unsigned short mem_cgroup_id(struct mem_cgroup *memcg)
452 return memcg->css.id;
455 static inline struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
457 struct cgroup_subsys_state *css;
459 css = css_from_id(id, &memory_cgrp_subsys);
460 return mem_cgroup_from_css(css);
463 /* Writing them here to avoid exposing memcg's inner layout */
464 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
466 void sock_update_memcg(struct sock *sk)
468 if (mem_cgroup_sockets_enabled) {
469 struct mem_cgroup *memcg;
470 struct cg_proto *cg_proto;
472 BUG_ON(!sk->sk_prot->proto_cgroup);
474 /* Socket cloning can throw us here with sk_cgrp already
475 * filled. It won't however, necessarily happen from
476 * process context. So the test for root memcg given
477 * the current task's memcg won't help us in this case.
479 * Respecting the original socket's memcg is a better
480 * decision in this case.
483 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
484 css_get(&sk->sk_cgrp->memcg->css);
489 memcg = mem_cgroup_from_task(current);
490 cg_proto = sk->sk_prot->proto_cgroup(memcg);
491 if (!mem_cgroup_is_root(memcg) &&
492 memcg_proto_active(cg_proto) &&
493 css_tryget_online(&memcg->css)) {
494 sk->sk_cgrp = cg_proto;
499 EXPORT_SYMBOL(sock_update_memcg);
501 void sock_release_memcg(struct sock *sk)
503 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
504 struct mem_cgroup *memcg;
505 WARN_ON(!sk->sk_cgrp->memcg);
506 memcg = sk->sk_cgrp->memcg;
507 css_put(&sk->sk_cgrp->memcg->css);
511 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
513 if (!memcg || mem_cgroup_is_root(memcg))
516 return &memcg->tcp_mem;
518 EXPORT_SYMBOL(tcp_proto_cgroup);
520 static void disarm_sock_keys(struct mem_cgroup *memcg)
522 if (!memcg_proto_activated(&memcg->tcp_mem))
524 static_key_slow_dec(&memcg_socket_limit_enabled);
527 static void disarm_sock_keys(struct mem_cgroup *memcg)
532 #ifdef CONFIG_MEMCG_KMEM
534 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
535 * The main reason for not using cgroup id for this:
536 * this works better in sparse environments, where we have a lot of memcgs,
537 * but only a few kmem-limited. Or also, if we have, for instance, 200
538 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
539 * 200 entry array for that.
541 * The current size of the caches array is stored in
542 * memcg_limited_groups_array_size. It will double each time we have to
545 static DEFINE_IDA(kmem_limited_groups);
546 int memcg_limited_groups_array_size;
549 * MIN_SIZE is different than 1, because we would like to avoid going through
550 * the alloc/free process all the time. In a small machine, 4 kmem-limited
551 * cgroups is a reasonable guess. In the future, it could be a parameter or
552 * tunable, but that is strictly not necessary.
554 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
555 * this constant directly from cgroup, but it is understandable that this is
556 * better kept as an internal representation in cgroup.c. In any case, the
557 * cgrp_id space is not getting any smaller, and we don't have to necessarily
558 * increase ours as well if it increases.
560 #define MEMCG_CACHES_MIN_SIZE 4
561 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
564 * A lot of the calls to the cache allocation functions are expected to be
565 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
566 * conditional to this static branch, we'll have to allow modules that does
567 * kmem_cache_alloc and the such to see this symbol as well
569 struct static_key memcg_kmem_enabled_key;
570 EXPORT_SYMBOL(memcg_kmem_enabled_key);
572 static void memcg_free_cache_id(int id);
574 static void disarm_kmem_keys(struct mem_cgroup *memcg)
576 if (memcg_kmem_is_active(memcg)) {
577 static_key_slow_dec(&memcg_kmem_enabled_key);
578 memcg_free_cache_id(memcg->kmemcg_id);
581 * This check can't live in kmem destruction function,
582 * since the charges will outlive the cgroup
584 WARN_ON(page_counter_read(&memcg->kmem));
587 static void disarm_kmem_keys(struct mem_cgroup *memcg)
590 #endif /* CONFIG_MEMCG_KMEM */
592 static void disarm_static_keys(struct mem_cgroup *memcg)
594 disarm_sock_keys(memcg);
595 disarm_kmem_keys(memcg);
598 static struct mem_cgroup_per_zone *
599 mem_cgroup_zone_zoneinfo(struct mem_cgroup *memcg, struct zone *zone)
601 int nid = zone_to_nid(zone);
602 int zid = zone_idx(zone);
604 return &memcg->nodeinfo[nid]->zoneinfo[zid];
607 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
612 static struct mem_cgroup_per_zone *
613 mem_cgroup_page_zoneinfo(struct mem_cgroup *memcg, struct page *page)
615 int nid = page_to_nid(page);
616 int zid = page_zonenum(page);
618 return &memcg->nodeinfo[nid]->zoneinfo[zid];
621 static struct mem_cgroup_tree_per_zone *
622 soft_limit_tree_node_zone(int nid, int zid)
624 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
627 static struct mem_cgroup_tree_per_zone *
628 soft_limit_tree_from_page(struct page *page)
630 int nid = page_to_nid(page);
631 int zid = page_zonenum(page);
633 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
636 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone *mz,
637 struct mem_cgroup_tree_per_zone *mctz,
638 unsigned long new_usage_in_excess)
640 struct rb_node **p = &mctz->rb_root.rb_node;
641 struct rb_node *parent = NULL;
642 struct mem_cgroup_per_zone *mz_node;
647 mz->usage_in_excess = new_usage_in_excess;
648 if (!mz->usage_in_excess)
652 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
654 if (mz->usage_in_excess < mz_node->usage_in_excess)
657 * We can't avoid mem cgroups that are over their soft
658 * limit by the same amount
660 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
663 rb_link_node(&mz->tree_node, parent, p);
664 rb_insert_color(&mz->tree_node, &mctz->rb_root);
668 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
669 struct mem_cgroup_tree_per_zone *mctz)
673 rb_erase(&mz->tree_node, &mctz->rb_root);
677 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
678 struct mem_cgroup_tree_per_zone *mctz)
682 spin_lock_irqsave(&mctz->lock, flags);
683 __mem_cgroup_remove_exceeded(mz, mctz);
684 spin_unlock_irqrestore(&mctz->lock, flags);
687 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
689 unsigned long nr_pages = page_counter_read(&memcg->memory);
690 unsigned long soft_limit = ACCESS_ONCE(memcg->soft_limit);
691 unsigned long excess = 0;
693 if (nr_pages > soft_limit)
694 excess = nr_pages - soft_limit;
699 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
701 unsigned long excess;
702 struct mem_cgroup_per_zone *mz;
703 struct mem_cgroup_tree_per_zone *mctz;
705 mctz = soft_limit_tree_from_page(page);
707 * Necessary to update all ancestors when hierarchy is used.
708 * because their event counter is not touched.
710 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
711 mz = mem_cgroup_page_zoneinfo(memcg, page);
712 excess = soft_limit_excess(memcg);
714 * We have to update the tree if mz is on RB-tree or
715 * mem is over its softlimit.
717 if (excess || mz->on_tree) {
720 spin_lock_irqsave(&mctz->lock, flags);
721 /* if on-tree, remove it */
723 __mem_cgroup_remove_exceeded(mz, mctz);
725 * Insert again. mz->usage_in_excess will be updated.
726 * If excess is 0, no tree ops.
728 __mem_cgroup_insert_exceeded(mz, mctz, excess);
729 spin_unlock_irqrestore(&mctz->lock, flags);
734 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
736 struct mem_cgroup_tree_per_zone *mctz;
737 struct mem_cgroup_per_zone *mz;
741 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
742 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
743 mctz = soft_limit_tree_node_zone(nid, zid);
744 mem_cgroup_remove_exceeded(mz, mctz);
749 static struct mem_cgroup_per_zone *
750 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
752 struct rb_node *rightmost = NULL;
753 struct mem_cgroup_per_zone *mz;
757 rightmost = rb_last(&mctz->rb_root);
759 goto done; /* Nothing to reclaim from */
761 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
763 * Remove the node now but someone else can add it back,
764 * we will to add it back at the end of reclaim to its correct
765 * position in the tree.
767 __mem_cgroup_remove_exceeded(mz, mctz);
768 if (!soft_limit_excess(mz->memcg) ||
769 !css_tryget_online(&mz->memcg->css))
775 static struct mem_cgroup_per_zone *
776 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
778 struct mem_cgroup_per_zone *mz;
780 spin_lock_irq(&mctz->lock);
781 mz = __mem_cgroup_largest_soft_limit_node(mctz);
782 spin_unlock_irq(&mctz->lock);
787 * Implementation Note: reading percpu statistics for memcg.
789 * Both of vmstat[] and percpu_counter has threshold and do periodic
790 * synchronization to implement "quick" read. There are trade-off between
791 * reading cost and precision of value. Then, we may have a chance to implement
792 * a periodic synchronizion of counter in memcg's counter.
794 * But this _read() function is used for user interface now. The user accounts
795 * memory usage by memory cgroup and he _always_ requires exact value because
796 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
797 * have to visit all online cpus and make sum. So, for now, unnecessary
798 * synchronization is not implemented. (just implemented for cpu hotplug)
800 * If there are kernel internal actions which can make use of some not-exact
801 * value, and reading all cpu value can be performance bottleneck in some
802 * common workload, threashold and synchonization as vmstat[] should be
805 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
806 enum mem_cgroup_stat_index idx)
812 for_each_online_cpu(cpu)
813 val += per_cpu(memcg->stat->count[idx], cpu);
814 #ifdef CONFIG_HOTPLUG_CPU
815 spin_lock(&memcg->pcp_counter_lock);
816 val += memcg->nocpu_base.count[idx];
817 spin_unlock(&memcg->pcp_counter_lock);
823 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
824 enum mem_cgroup_events_index idx)
826 unsigned long val = 0;
830 for_each_online_cpu(cpu)
831 val += per_cpu(memcg->stat->events[idx], cpu);
832 #ifdef CONFIG_HOTPLUG_CPU
833 spin_lock(&memcg->pcp_counter_lock);
834 val += memcg->nocpu_base.events[idx];
835 spin_unlock(&memcg->pcp_counter_lock);
841 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
846 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
847 * counted as CACHE even if it's on ANON LRU.
850 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
853 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
856 if (PageTransHuge(page))
857 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
860 /* pagein of a big page is an event. So, ignore page size */
862 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
864 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
865 nr_pages = -nr_pages; /* for event */
868 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
871 unsigned long mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
873 struct mem_cgroup_per_zone *mz;
875 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
876 return mz->lru_size[lru];
879 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
881 unsigned int lru_mask)
883 unsigned long nr = 0;
886 VM_BUG_ON((unsigned)nid >= nr_node_ids);
888 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
889 struct mem_cgroup_per_zone *mz;
893 if (!(BIT(lru) & lru_mask))
895 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
896 nr += mz->lru_size[lru];
902 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
903 unsigned int lru_mask)
905 unsigned long nr = 0;
908 for_each_node_state(nid, N_MEMORY)
909 nr += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
913 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
914 enum mem_cgroup_events_target target)
916 unsigned long val, next;
918 val = __this_cpu_read(memcg->stat->nr_page_events);
919 next = __this_cpu_read(memcg->stat->targets[target]);
920 /* from time_after() in jiffies.h */
921 if ((long)next - (long)val < 0) {
923 case MEM_CGROUP_TARGET_THRESH:
924 next = val + THRESHOLDS_EVENTS_TARGET;
926 case MEM_CGROUP_TARGET_SOFTLIMIT:
927 next = val + SOFTLIMIT_EVENTS_TARGET;
929 case MEM_CGROUP_TARGET_NUMAINFO:
930 next = val + NUMAINFO_EVENTS_TARGET;
935 __this_cpu_write(memcg->stat->targets[target], next);
942 * Check events in order.
945 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
947 /* threshold event is triggered in finer grain than soft limit */
948 if (unlikely(mem_cgroup_event_ratelimit(memcg,
949 MEM_CGROUP_TARGET_THRESH))) {
951 bool do_numainfo __maybe_unused;
953 do_softlimit = mem_cgroup_event_ratelimit(memcg,
954 MEM_CGROUP_TARGET_SOFTLIMIT);
956 do_numainfo = mem_cgroup_event_ratelimit(memcg,
957 MEM_CGROUP_TARGET_NUMAINFO);
959 mem_cgroup_threshold(memcg);
960 if (unlikely(do_softlimit))
961 mem_cgroup_update_tree(memcg, page);
963 if (unlikely(do_numainfo))
964 atomic_inc(&memcg->numainfo_events);
969 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
972 * mm_update_next_owner() may clear mm->owner to NULL
973 * if it races with swapoff, page migration, etc.
974 * So this can be called with p == NULL.
979 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
982 static struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
984 struct mem_cgroup *memcg = NULL;
989 * Page cache insertions can happen withou an
990 * actual mm context, e.g. during disk probing
991 * on boot, loopback IO, acct() writes etc.
994 memcg = root_mem_cgroup;
996 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
997 if (unlikely(!memcg))
998 memcg = root_mem_cgroup;
1000 } while (!css_tryget_online(&memcg->css));
1006 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1007 * @root: hierarchy root
1008 * @prev: previously returned memcg, NULL on first invocation
1009 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1011 * Returns references to children of the hierarchy below @root, or
1012 * @root itself, or %NULL after a full round-trip.
1014 * Caller must pass the return value in @prev on subsequent
1015 * invocations for reference counting, or use mem_cgroup_iter_break()
1016 * to cancel a hierarchy walk before the round-trip is complete.
1018 * Reclaimers can specify a zone and a priority level in @reclaim to
1019 * divide up the memcgs in the hierarchy among all concurrent
1020 * reclaimers operating on the same zone and priority.
1022 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1023 struct mem_cgroup *prev,
1024 struct mem_cgroup_reclaim_cookie *reclaim)
1026 struct reclaim_iter *uninitialized_var(iter);
1027 struct cgroup_subsys_state *css = NULL;
1028 struct mem_cgroup *memcg = NULL;
1029 struct mem_cgroup *pos = NULL;
1031 if (mem_cgroup_disabled())
1035 root = root_mem_cgroup;
1037 if (prev && !reclaim)
1040 if (!root->use_hierarchy && root != root_mem_cgroup) {
1049 struct mem_cgroup_per_zone *mz;
1051 mz = mem_cgroup_zone_zoneinfo(root, reclaim->zone);
1052 iter = &mz->iter[reclaim->priority];
1054 if (prev && reclaim->generation != iter->generation)
1058 pos = ACCESS_ONCE(iter->position);
1060 * A racing update may change the position and
1061 * put the last reference, hence css_tryget(),
1062 * or retry to see the updated position.
1064 } while (pos && !css_tryget(&pos->css));
1071 css = css_next_descendant_pre(css, &root->css);
1074 * Reclaimers share the hierarchy walk, and a
1075 * new one might jump in right at the end of
1076 * the hierarchy - make sure they see at least
1077 * one group and restart from the beginning.
1085 * Verify the css and acquire a reference. The root
1086 * is provided by the caller, so we know it's alive
1087 * and kicking, and don't take an extra reference.
1089 memcg = mem_cgroup_from_css(css);
1091 if (css == &root->css)
1094 if (css_tryget(css)) {
1096 * Make sure the memcg is initialized:
1097 * mem_cgroup_css_online() orders the the
1098 * initialization against setting the flag.
1100 if (smp_load_acquire(&memcg->initialized))
1110 if (cmpxchg(&iter->position, pos, memcg) == pos) {
1112 css_get(&memcg->css);
1118 * pairs with css_tryget when dereferencing iter->position
1127 reclaim->generation = iter->generation;
1133 if (prev && prev != root)
1134 css_put(&prev->css);
1140 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1141 * @root: hierarchy root
1142 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1144 void mem_cgroup_iter_break(struct mem_cgroup *root,
1145 struct mem_cgroup *prev)
1148 root = root_mem_cgroup;
1149 if (prev && prev != root)
1150 css_put(&prev->css);
1154 * Iteration constructs for visiting all cgroups (under a tree). If
1155 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1156 * be used for reference counting.
1158 #define for_each_mem_cgroup_tree(iter, root) \
1159 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1161 iter = mem_cgroup_iter(root, iter, NULL))
1163 #define for_each_mem_cgroup(iter) \
1164 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1166 iter = mem_cgroup_iter(NULL, iter, NULL))
1168 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1170 struct mem_cgroup *memcg;
1173 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1174 if (unlikely(!memcg))
1179 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1182 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1190 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1193 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1194 * @zone: zone of the wanted lruvec
1195 * @memcg: memcg of the wanted lruvec
1197 * Returns the lru list vector holding pages for the given @zone and
1198 * @mem. This can be the global zone lruvec, if the memory controller
1201 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1202 struct mem_cgroup *memcg)
1204 struct mem_cgroup_per_zone *mz;
1205 struct lruvec *lruvec;
1207 if (mem_cgroup_disabled()) {
1208 lruvec = &zone->lruvec;
1212 mz = mem_cgroup_zone_zoneinfo(memcg, zone);
1213 lruvec = &mz->lruvec;
1216 * Since a node can be onlined after the mem_cgroup was created,
1217 * we have to be prepared to initialize lruvec->zone here;
1218 * and if offlined then reonlined, we need to reinitialize it.
1220 if (unlikely(lruvec->zone != zone))
1221 lruvec->zone = zone;
1226 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1228 * @zone: zone of the page
1230 * This function is only safe when following the LRU page isolation
1231 * and putback protocol: the LRU lock must be held, and the page must
1232 * either be PageLRU() or the caller must have isolated/allocated it.
1234 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1236 struct mem_cgroup_per_zone *mz;
1237 struct mem_cgroup *memcg;
1238 struct lruvec *lruvec;
1240 if (mem_cgroup_disabled()) {
1241 lruvec = &zone->lruvec;
1245 memcg = page->mem_cgroup;
1247 * Swapcache readahead pages are added to the LRU - and
1248 * possibly migrated - before they are charged.
1251 memcg = root_mem_cgroup;
1253 mz = mem_cgroup_page_zoneinfo(memcg, page);
1254 lruvec = &mz->lruvec;
1257 * Since a node can be onlined after the mem_cgroup was created,
1258 * we have to be prepared to initialize lruvec->zone here;
1259 * and if offlined then reonlined, we need to reinitialize it.
1261 if (unlikely(lruvec->zone != zone))
1262 lruvec->zone = zone;
1267 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1268 * @lruvec: mem_cgroup per zone lru vector
1269 * @lru: index of lru list the page is sitting on
1270 * @nr_pages: positive when adding or negative when removing
1272 * This function must be called when a page is added to or removed from an
1275 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1278 struct mem_cgroup_per_zone *mz;
1279 unsigned long *lru_size;
1281 if (mem_cgroup_disabled())
1284 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1285 lru_size = mz->lru_size + lru;
1286 *lru_size += nr_pages;
1287 VM_BUG_ON((long)(*lru_size) < 0);
1290 bool mem_cgroup_is_descendant(struct mem_cgroup *memcg, struct mem_cgroup *root)
1294 if (!root->use_hierarchy)
1296 return cgroup_is_descendant(memcg->css.cgroup, root->css.cgroup);
1299 bool task_in_mem_cgroup(struct task_struct *task, struct mem_cgroup *memcg)
1301 struct mem_cgroup *task_memcg;
1302 struct task_struct *p;
1305 p = find_lock_task_mm(task);
1307 task_memcg = get_mem_cgroup_from_mm(p->mm);
1311 * All threads may have already detached their mm's, but the oom
1312 * killer still needs to detect if they have already been oom
1313 * killed to prevent needlessly killing additional tasks.
1316 task_memcg = mem_cgroup_from_task(task);
1317 css_get(&task_memcg->css);
1320 ret = mem_cgroup_is_descendant(task_memcg, memcg);
1321 css_put(&task_memcg->css);
1325 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1327 unsigned long inactive_ratio;
1328 unsigned long inactive;
1329 unsigned long active;
1332 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1333 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1335 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1337 inactive_ratio = int_sqrt(10 * gb);
1341 return inactive * inactive_ratio < active;
1344 bool mem_cgroup_lruvec_online(struct lruvec *lruvec)
1346 struct mem_cgroup_per_zone *mz;
1347 struct mem_cgroup *memcg;
1349 if (mem_cgroup_disabled())
1352 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1355 return !!(memcg->css.flags & CSS_ONLINE);
1358 #define mem_cgroup_from_counter(counter, member) \
1359 container_of(counter, struct mem_cgroup, member)
1362 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1363 * @memcg: the memory cgroup
1365 * Returns the maximum amount of memory @mem can be charged with, in
1368 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1370 unsigned long margin = 0;
1371 unsigned long count;
1372 unsigned long limit;
1374 count = page_counter_read(&memcg->memory);
1375 limit = ACCESS_ONCE(memcg->memory.limit);
1377 margin = limit - count;
1379 if (do_swap_account) {
1380 count = page_counter_read(&memcg->memsw);
1381 limit = ACCESS_ONCE(memcg->memsw.limit);
1383 margin = min(margin, limit - count);
1389 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1392 if (mem_cgroup_disabled() || !memcg->css.parent)
1393 return vm_swappiness;
1395 return memcg->swappiness;
1399 * A routine for checking "mem" is under move_account() or not.
1401 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1402 * moving cgroups. This is for waiting at high-memory pressure
1405 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1407 struct mem_cgroup *from;
1408 struct mem_cgroup *to;
1411 * Unlike task_move routines, we access mc.to, mc.from not under
1412 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1414 spin_lock(&mc.lock);
1420 ret = mem_cgroup_is_descendant(from, memcg) ||
1421 mem_cgroup_is_descendant(to, memcg);
1423 spin_unlock(&mc.lock);
1427 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1429 if (mc.moving_task && current != mc.moving_task) {
1430 if (mem_cgroup_under_move(memcg)) {
1432 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1433 /* moving charge context might have finished. */
1436 finish_wait(&mc.waitq, &wait);
1443 #define K(x) ((x) << (PAGE_SHIFT-10))
1445 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1446 * @memcg: The memory cgroup that went over limit
1447 * @p: Task that is going to be killed
1449 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1452 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1454 /* oom_info_lock ensures that parallel ooms do not interleave */
1455 static DEFINE_MUTEX(oom_info_lock);
1456 struct mem_cgroup *iter;
1462 mutex_lock(&oom_info_lock);
1465 pr_info("Task in ");
1466 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1467 pr_cont(" killed as a result of limit of ");
1468 pr_cont_cgroup_path(memcg->css.cgroup);
1473 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1474 K((u64)page_counter_read(&memcg->memory)),
1475 K((u64)memcg->memory.limit), memcg->memory.failcnt);
1476 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1477 K((u64)page_counter_read(&memcg->memsw)),
1478 K((u64)memcg->memsw.limit), memcg->memsw.failcnt);
1479 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1480 K((u64)page_counter_read(&memcg->kmem)),
1481 K((u64)memcg->kmem.limit), memcg->kmem.failcnt);
1483 for_each_mem_cgroup_tree(iter, memcg) {
1484 pr_info("Memory cgroup stats for ");
1485 pr_cont_cgroup_path(iter->css.cgroup);
1488 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1489 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1491 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1492 K(mem_cgroup_read_stat(iter, i)));
1495 for (i = 0; i < NR_LRU_LISTS; i++)
1496 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1497 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1501 mutex_unlock(&oom_info_lock);
1505 * This function returns the number of memcg under hierarchy tree. Returns
1506 * 1(self count) if no children.
1508 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1511 struct mem_cgroup *iter;
1513 for_each_mem_cgroup_tree(iter, memcg)
1519 * Return the memory (and swap, if configured) limit for a memcg.
1521 static unsigned long mem_cgroup_get_limit(struct mem_cgroup *memcg)
1523 unsigned long limit;
1525 limit = memcg->memory.limit;
1526 if (mem_cgroup_swappiness(memcg)) {
1527 unsigned long memsw_limit;
1529 memsw_limit = memcg->memsw.limit;
1530 limit = min(limit + total_swap_pages, memsw_limit);
1535 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1538 struct mem_cgroup *iter;
1539 unsigned long chosen_points = 0;
1540 unsigned long totalpages;
1541 unsigned int points = 0;
1542 struct task_struct *chosen = NULL;
1545 * If current has a pending SIGKILL or is exiting, then automatically
1546 * select it. The goal is to allow it to allocate so that it may
1547 * quickly exit and free its memory.
1549 if (fatal_signal_pending(current) || task_will_free_mem(current)) {
1550 mark_tsk_oom_victim(current);
1554 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1555 totalpages = mem_cgroup_get_limit(memcg) ? : 1;
1556 for_each_mem_cgroup_tree(iter, memcg) {
1557 struct css_task_iter it;
1558 struct task_struct *task;
1560 css_task_iter_start(&iter->css, &it);
1561 while ((task = css_task_iter_next(&it))) {
1562 switch (oom_scan_process_thread(task, totalpages, NULL,
1564 case OOM_SCAN_SELECT:
1566 put_task_struct(chosen);
1568 chosen_points = ULONG_MAX;
1569 get_task_struct(chosen);
1571 case OOM_SCAN_CONTINUE:
1573 case OOM_SCAN_ABORT:
1574 css_task_iter_end(&it);
1575 mem_cgroup_iter_break(memcg, iter);
1577 put_task_struct(chosen);
1582 points = oom_badness(task, memcg, NULL, totalpages);
1583 if (!points || points < chosen_points)
1585 /* Prefer thread group leaders for display purposes */
1586 if (points == chosen_points &&
1587 thread_group_leader(chosen))
1591 put_task_struct(chosen);
1593 chosen_points = points;
1594 get_task_struct(chosen);
1596 css_task_iter_end(&it);
1601 points = chosen_points * 1000 / totalpages;
1602 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1603 NULL, "Memory cgroup out of memory");
1606 #if MAX_NUMNODES > 1
1609 * test_mem_cgroup_node_reclaimable
1610 * @memcg: the target memcg
1611 * @nid: the node ID to be checked.
1612 * @noswap : specify true here if the user wants flle only information.
1614 * This function returns whether the specified memcg contains any
1615 * reclaimable pages on a node. Returns true if there are any reclaimable
1616 * pages in the node.
1618 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1619 int nid, bool noswap)
1621 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1623 if (noswap || !total_swap_pages)
1625 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1632 * Always updating the nodemask is not very good - even if we have an empty
1633 * list or the wrong list here, we can start from some node and traverse all
1634 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1637 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1641 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1642 * pagein/pageout changes since the last update.
1644 if (!atomic_read(&memcg->numainfo_events))
1646 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1649 /* make a nodemask where this memcg uses memory from */
1650 memcg->scan_nodes = node_states[N_MEMORY];
1652 for_each_node_mask(nid, node_states[N_MEMORY]) {
1654 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1655 node_clear(nid, memcg->scan_nodes);
1658 atomic_set(&memcg->numainfo_events, 0);
1659 atomic_set(&memcg->numainfo_updating, 0);
1663 * Selecting a node where we start reclaim from. Because what we need is just
1664 * reducing usage counter, start from anywhere is O,K. Considering
1665 * memory reclaim from current node, there are pros. and cons.
1667 * Freeing memory from current node means freeing memory from a node which
1668 * we'll use or we've used. So, it may make LRU bad. And if several threads
1669 * hit limits, it will see a contention on a node. But freeing from remote
1670 * node means more costs for memory reclaim because of memory latency.
1672 * Now, we use round-robin. Better algorithm is welcomed.
1674 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1678 mem_cgroup_may_update_nodemask(memcg);
1679 node = memcg->last_scanned_node;
1681 node = next_node(node, memcg->scan_nodes);
1682 if (node == MAX_NUMNODES)
1683 node = first_node(memcg->scan_nodes);
1685 * We call this when we hit limit, not when pages are added to LRU.
1686 * No LRU may hold pages because all pages are UNEVICTABLE or
1687 * memcg is too small and all pages are not on LRU. In that case,
1688 * we use curret node.
1690 if (unlikely(node == MAX_NUMNODES))
1691 node = numa_node_id();
1693 memcg->last_scanned_node = node;
1697 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1703 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1706 unsigned long *total_scanned)
1708 struct mem_cgroup *victim = NULL;
1711 unsigned long excess;
1712 unsigned long nr_scanned;
1713 struct mem_cgroup_reclaim_cookie reclaim = {
1718 excess = soft_limit_excess(root_memcg);
1721 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1726 * If we have not been able to reclaim
1727 * anything, it might because there are
1728 * no reclaimable pages under this hierarchy
1733 * We want to do more targeted reclaim.
1734 * excess >> 2 is not to excessive so as to
1735 * reclaim too much, nor too less that we keep
1736 * coming back to reclaim from this cgroup
1738 if (total >= (excess >> 2) ||
1739 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1744 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1746 *total_scanned += nr_scanned;
1747 if (!soft_limit_excess(root_memcg))
1750 mem_cgroup_iter_break(root_memcg, victim);
1754 #ifdef CONFIG_LOCKDEP
1755 static struct lockdep_map memcg_oom_lock_dep_map = {
1756 .name = "memcg_oom_lock",
1760 static DEFINE_SPINLOCK(memcg_oom_lock);
1763 * Check OOM-Killer is already running under our hierarchy.
1764 * If someone is running, return false.
1766 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1768 struct mem_cgroup *iter, *failed = NULL;
1770 spin_lock(&memcg_oom_lock);
1772 for_each_mem_cgroup_tree(iter, memcg) {
1773 if (iter->oom_lock) {
1775 * this subtree of our hierarchy is already locked
1776 * so we cannot give a lock.
1779 mem_cgroup_iter_break(memcg, iter);
1782 iter->oom_lock = true;
1787 * OK, we failed to lock the whole subtree so we have
1788 * to clean up what we set up to the failing subtree
1790 for_each_mem_cgroup_tree(iter, memcg) {
1791 if (iter == failed) {
1792 mem_cgroup_iter_break(memcg, iter);
1795 iter->oom_lock = false;
1798 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1800 spin_unlock(&memcg_oom_lock);
1805 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1807 struct mem_cgroup *iter;
1809 spin_lock(&memcg_oom_lock);
1810 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1811 for_each_mem_cgroup_tree(iter, memcg)
1812 iter->oom_lock = false;
1813 spin_unlock(&memcg_oom_lock);
1816 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1818 struct mem_cgroup *iter;
1820 for_each_mem_cgroup_tree(iter, memcg)
1821 atomic_inc(&iter->under_oom);
1824 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1826 struct mem_cgroup *iter;
1829 * When a new child is created while the hierarchy is under oom,
1830 * mem_cgroup_oom_lock() may not be called. We have to use
1831 * atomic_add_unless() here.
1833 for_each_mem_cgroup_tree(iter, memcg)
1834 atomic_add_unless(&iter->under_oom, -1, 0);
1837 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1839 struct oom_wait_info {
1840 struct mem_cgroup *memcg;
1844 static int memcg_oom_wake_function(wait_queue_t *wait,
1845 unsigned mode, int sync, void *arg)
1847 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1848 struct mem_cgroup *oom_wait_memcg;
1849 struct oom_wait_info *oom_wait_info;
1851 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1852 oom_wait_memcg = oom_wait_info->memcg;
1854 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1855 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1857 return autoremove_wake_function(wait, mode, sync, arg);
1860 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
1862 atomic_inc(&memcg->oom_wakeups);
1863 /* for filtering, pass "memcg" as argument. */
1864 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1867 static void memcg_oom_recover(struct mem_cgroup *memcg)
1869 if (memcg && atomic_read(&memcg->under_oom))
1870 memcg_wakeup_oom(memcg);
1873 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1875 if (!current->memcg_oom.may_oom)
1878 * We are in the middle of the charge context here, so we
1879 * don't want to block when potentially sitting on a callstack
1880 * that holds all kinds of filesystem and mm locks.
1882 * Also, the caller may handle a failed allocation gracefully
1883 * (like optional page cache readahead) and so an OOM killer
1884 * invocation might not even be necessary.
1886 * That's why we don't do anything here except remember the
1887 * OOM context and then deal with it at the end of the page
1888 * fault when the stack is unwound, the locks are released,
1889 * and when we know whether the fault was overall successful.
1891 css_get(&memcg->css);
1892 current->memcg_oom.memcg = memcg;
1893 current->memcg_oom.gfp_mask = mask;
1894 current->memcg_oom.order = order;
1898 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1899 * @handle: actually kill/wait or just clean up the OOM state
1901 * This has to be called at the end of a page fault if the memcg OOM
1902 * handler was enabled.
1904 * Memcg supports userspace OOM handling where failed allocations must
1905 * sleep on a waitqueue until the userspace task resolves the
1906 * situation. Sleeping directly in the charge context with all kinds
1907 * of locks held is not a good idea, instead we remember an OOM state
1908 * in the task and mem_cgroup_oom_synchronize() has to be called at
1909 * the end of the page fault to complete the OOM handling.
1911 * Returns %true if an ongoing memcg OOM situation was detected and
1912 * completed, %false otherwise.
1914 bool mem_cgroup_oom_synchronize(bool handle)
1916 struct mem_cgroup *memcg = current->memcg_oom.memcg;
1917 struct oom_wait_info owait;
1920 /* OOM is global, do not handle */
1924 if (!handle || oom_killer_disabled)
1927 owait.memcg = memcg;
1928 owait.wait.flags = 0;
1929 owait.wait.func = memcg_oom_wake_function;
1930 owait.wait.private = current;
1931 INIT_LIST_HEAD(&owait.wait.task_list);
1933 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1934 mem_cgroup_mark_under_oom(memcg);
1936 locked = mem_cgroup_oom_trylock(memcg);
1939 mem_cgroup_oom_notify(memcg);
1941 if (locked && !memcg->oom_kill_disable) {
1942 mem_cgroup_unmark_under_oom(memcg);
1943 finish_wait(&memcg_oom_waitq, &owait.wait);
1944 mem_cgroup_out_of_memory(memcg, current->memcg_oom.gfp_mask,
1945 current->memcg_oom.order);
1948 mem_cgroup_unmark_under_oom(memcg);
1949 finish_wait(&memcg_oom_waitq, &owait.wait);
1953 mem_cgroup_oom_unlock(memcg);
1955 * There is no guarantee that an OOM-lock contender
1956 * sees the wakeups triggered by the OOM kill
1957 * uncharges. Wake any sleepers explicitely.
1959 memcg_oom_recover(memcg);
1962 current->memcg_oom.memcg = NULL;
1963 css_put(&memcg->css);
1968 * mem_cgroup_begin_page_stat - begin a page state statistics transaction
1969 * @page: page that is going to change accounted state
1971 * This function must mark the beginning of an accounted page state
1972 * change to prevent double accounting when the page is concurrently
1973 * being moved to another memcg:
1975 * memcg = mem_cgroup_begin_page_stat(page);
1976 * if (TestClearPageState(page))
1977 * mem_cgroup_update_page_stat(memcg, state, -1);
1978 * mem_cgroup_end_page_stat(memcg);
1980 struct mem_cgroup *mem_cgroup_begin_page_stat(struct page *page)
1982 struct mem_cgroup *memcg;
1983 unsigned long flags;
1986 * The RCU lock is held throughout the transaction. The fast
1987 * path can get away without acquiring the memcg->move_lock
1988 * because page moving starts with an RCU grace period.
1990 * The RCU lock also protects the memcg from being freed when
1991 * the page state that is going to change is the only thing
1992 * preventing the page from being uncharged.
1993 * E.g. end-writeback clearing PageWriteback(), which allows
1994 * migration to go ahead and uncharge the page before the
1995 * account transaction might be complete.
1999 if (mem_cgroup_disabled())
2002 memcg = page->mem_cgroup;
2003 if (unlikely(!memcg))
2006 if (atomic_read(&memcg->moving_account) <= 0)
2009 spin_lock_irqsave(&memcg->move_lock, flags);
2010 if (memcg != page->mem_cgroup) {
2011 spin_unlock_irqrestore(&memcg->move_lock, flags);
2016 * When charge migration first begins, we can have locked and
2017 * unlocked page stat updates happening concurrently. Track
2018 * the task who has the lock for mem_cgroup_end_page_stat().
2020 memcg->move_lock_task = current;
2021 memcg->move_lock_flags = flags;
2027 * mem_cgroup_end_page_stat - finish a page state statistics transaction
2028 * @memcg: the memcg that was accounted against
2030 void mem_cgroup_end_page_stat(struct mem_cgroup *memcg)
2032 if (memcg && memcg->move_lock_task == current) {
2033 unsigned long flags = memcg->move_lock_flags;
2035 memcg->move_lock_task = NULL;
2036 memcg->move_lock_flags = 0;
2038 spin_unlock_irqrestore(&memcg->move_lock, flags);
2045 * mem_cgroup_update_page_stat - update page state statistics
2046 * @memcg: memcg to account against
2047 * @idx: page state item to account
2048 * @val: number of pages (positive or negative)
2050 * See mem_cgroup_begin_page_stat() for locking requirements.
2052 void mem_cgroup_update_page_stat(struct mem_cgroup *memcg,
2053 enum mem_cgroup_stat_index idx, int val)
2055 VM_BUG_ON(!rcu_read_lock_held());
2058 this_cpu_add(memcg->stat->count[idx], val);
2062 * size of first charge trial. "32" comes from vmscan.c's magic value.
2063 * TODO: maybe necessary to use big numbers in big irons.
2065 #define CHARGE_BATCH 32U
2066 struct memcg_stock_pcp {
2067 struct mem_cgroup *cached; /* this never be root cgroup */
2068 unsigned int nr_pages;
2069 struct work_struct work;
2070 unsigned long flags;
2071 #define FLUSHING_CACHED_CHARGE 0
2073 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2074 static DEFINE_MUTEX(percpu_charge_mutex);
2077 * consume_stock: Try to consume stocked charge on this cpu.
2078 * @memcg: memcg to consume from.
2079 * @nr_pages: how many pages to charge.
2081 * The charges will only happen if @memcg matches the current cpu's memcg
2082 * stock, and at least @nr_pages are available in that stock. Failure to
2083 * service an allocation will refill the stock.
2085 * returns true if successful, false otherwise.
2087 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2089 struct memcg_stock_pcp *stock;
2092 if (nr_pages > CHARGE_BATCH)
2095 stock = &get_cpu_var(memcg_stock);
2096 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2097 stock->nr_pages -= nr_pages;
2100 put_cpu_var(memcg_stock);
2105 * Returns stocks cached in percpu and reset cached information.
2107 static void drain_stock(struct memcg_stock_pcp *stock)
2109 struct mem_cgroup *old = stock->cached;
2111 if (stock->nr_pages) {
2112 page_counter_uncharge(&old->memory, stock->nr_pages);
2113 if (do_swap_account)
2114 page_counter_uncharge(&old->memsw, stock->nr_pages);
2115 css_put_many(&old->css, stock->nr_pages);
2116 stock->nr_pages = 0;
2118 stock->cached = NULL;
2122 * This must be called under preempt disabled or must be called by
2123 * a thread which is pinned to local cpu.
2125 static void drain_local_stock(struct work_struct *dummy)
2127 struct memcg_stock_pcp *stock = this_cpu_ptr(&memcg_stock);
2129 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2133 * Cache charges(val) to local per_cpu area.
2134 * This will be consumed by consume_stock() function, later.
2136 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2138 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2140 if (stock->cached != memcg) { /* reset if necessary */
2142 stock->cached = memcg;
2144 stock->nr_pages += nr_pages;
2145 put_cpu_var(memcg_stock);
2149 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2150 * of the hierarchy under it.
2152 static void drain_all_stock(struct mem_cgroup *root_memcg)
2156 /* If someone's already draining, avoid adding running more workers. */
2157 if (!mutex_trylock(&percpu_charge_mutex))
2159 /* Notify other cpus that system-wide "drain" is running */
2162 for_each_online_cpu(cpu) {
2163 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2164 struct mem_cgroup *memcg;
2166 memcg = stock->cached;
2167 if (!memcg || !stock->nr_pages)
2169 if (!mem_cgroup_is_descendant(memcg, root_memcg))
2171 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2173 drain_local_stock(&stock->work);
2175 schedule_work_on(cpu, &stock->work);
2180 mutex_unlock(&percpu_charge_mutex);
2184 * This function drains percpu counter value from DEAD cpu and
2185 * move it to local cpu. Note that this function can be preempted.
2187 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2191 spin_lock(&memcg->pcp_counter_lock);
2192 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2193 long x = per_cpu(memcg->stat->count[i], cpu);
2195 per_cpu(memcg->stat->count[i], cpu) = 0;
2196 memcg->nocpu_base.count[i] += x;
2198 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2199 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2201 per_cpu(memcg->stat->events[i], cpu) = 0;
2202 memcg->nocpu_base.events[i] += x;
2204 spin_unlock(&memcg->pcp_counter_lock);
2207 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
2208 unsigned long action,
2211 int cpu = (unsigned long)hcpu;
2212 struct memcg_stock_pcp *stock;
2213 struct mem_cgroup *iter;
2215 if (action == CPU_ONLINE)
2218 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2221 for_each_mem_cgroup(iter)
2222 mem_cgroup_drain_pcp_counter(iter, cpu);
2224 stock = &per_cpu(memcg_stock, cpu);
2229 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2230 unsigned int nr_pages)
2232 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2233 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2234 struct mem_cgroup *mem_over_limit;
2235 struct page_counter *counter;
2236 unsigned long nr_reclaimed;
2237 bool may_swap = true;
2238 bool drained = false;
2241 if (mem_cgroup_is_root(memcg))
2244 if (consume_stock(memcg, nr_pages))
2247 if (!do_swap_account ||
2248 !page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2249 if (!page_counter_try_charge(&memcg->memory, batch, &counter))
2251 if (do_swap_account)
2252 page_counter_uncharge(&memcg->memsw, batch);
2253 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2255 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2259 if (batch > nr_pages) {
2265 * Unlike in global OOM situations, memcg is not in a physical
2266 * memory shortage. Allow dying and OOM-killed tasks to
2267 * bypass the last charges so that they can exit quickly and
2268 * free their memory.
2270 if (unlikely(test_thread_flag(TIF_MEMDIE) ||
2271 fatal_signal_pending(current) ||
2272 current->flags & PF_EXITING))
2275 if (unlikely(task_in_memcg_oom(current)))
2278 if (!(gfp_mask & __GFP_WAIT))
2281 mem_cgroup_events(mem_over_limit, MEMCG_MAX, 1);
2283 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2284 gfp_mask, may_swap);
2286 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2290 drain_all_stock(mem_over_limit);
2295 if (gfp_mask & __GFP_NORETRY)
2298 * Even though the limit is exceeded at this point, reclaim
2299 * may have been able to free some pages. Retry the charge
2300 * before killing the task.
2302 * Only for regular pages, though: huge pages are rather
2303 * unlikely to succeed so close to the limit, and we fall back
2304 * to regular pages anyway in case of failure.
2306 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2309 * At task move, charge accounts can be doubly counted. So, it's
2310 * better to wait until the end of task_move if something is going on.
2312 if (mem_cgroup_wait_acct_move(mem_over_limit))
2318 if (gfp_mask & __GFP_NOFAIL)
2321 if (fatal_signal_pending(current))
2324 mem_cgroup_events(mem_over_limit, MEMCG_OOM, 1);
2326 mem_cgroup_oom(mem_over_limit, gfp_mask, get_order(nr_pages));
2328 if (!(gfp_mask & __GFP_NOFAIL))
2334 css_get_many(&memcg->css, batch);
2335 if (batch > nr_pages)
2336 refill_stock(memcg, batch - nr_pages);
2338 * If the hierarchy is above the normal consumption range,
2339 * make the charging task trim their excess contribution.
2342 if (page_counter_read(&memcg->memory) <= memcg->high)
2344 mem_cgroup_events(memcg, MEMCG_HIGH, 1);
2345 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
2346 } while ((memcg = parent_mem_cgroup(memcg)));
2351 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2353 if (mem_cgroup_is_root(memcg))
2356 page_counter_uncharge(&memcg->memory, nr_pages);
2357 if (do_swap_account)
2358 page_counter_uncharge(&memcg->memsw, nr_pages);
2360 css_put_many(&memcg->css, nr_pages);
2364 * A helper function to get mem_cgroup from ID. must be called under
2365 * rcu_read_lock(). The caller is responsible for calling
2366 * css_tryget_online() if the mem_cgroup is used for charging. (dropping
2367 * refcnt from swap can be called against removed memcg.)
2369 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2371 /* ID 0 is unused ID */
2374 return mem_cgroup_from_id(id);
2378 * try_get_mem_cgroup_from_page - look up page's memcg association
2381 * Look up, get a css reference, and return the memcg that owns @page.
2383 * The page must be locked to prevent racing with swap-in and page
2384 * cache charges. If coming from an unlocked page table, the caller
2385 * must ensure the page is on the LRU or this can race with charging.
2387 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2389 struct mem_cgroup *memcg;
2393 VM_BUG_ON_PAGE(!PageLocked(page), page);
2395 memcg = page->mem_cgroup;
2397 if (!css_tryget_online(&memcg->css))
2399 } else if (PageSwapCache(page)) {
2400 ent.val = page_private(page);
2401 id = lookup_swap_cgroup_id(ent);
2403 memcg = mem_cgroup_lookup(id);
2404 if (memcg && !css_tryget_online(&memcg->css))
2411 static void lock_page_lru(struct page *page, int *isolated)
2413 struct zone *zone = page_zone(page);
2415 spin_lock_irq(&zone->lru_lock);
2416 if (PageLRU(page)) {
2417 struct lruvec *lruvec;
2419 lruvec = mem_cgroup_page_lruvec(page, zone);
2421 del_page_from_lru_list(page, lruvec, page_lru(page));
2427 static void unlock_page_lru(struct page *page, int isolated)
2429 struct zone *zone = page_zone(page);
2432 struct lruvec *lruvec;
2434 lruvec = mem_cgroup_page_lruvec(page, zone);
2435 VM_BUG_ON_PAGE(PageLRU(page), page);
2437 add_page_to_lru_list(page, lruvec, page_lru(page));
2439 spin_unlock_irq(&zone->lru_lock);
2442 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2447 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2450 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2451 * may already be on some other mem_cgroup's LRU. Take care of it.
2454 lock_page_lru(page, &isolated);
2457 * Nobody should be changing or seriously looking at
2458 * page->mem_cgroup at this point:
2460 * - the page is uncharged
2462 * - the page is off-LRU
2464 * - an anonymous fault has exclusive page access, except for
2465 * a locked page table
2467 * - a page cache insertion, a swapin fault, or a migration
2468 * have the page locked
2470 page->mem_cgroup = memcg;
2473 unlock_page_lru(page, isolated);
2476 #ifdef CONFIG_MEMCG_KMEM
2477 int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp,
2478 unsigned long nr_pages)
2480 struct page_counter *counter;
2483 ret = page_counter_try_charge(&memcg->kmem, nr_pages, &counter);
2487 ret = try_charge(memcg, gfp, nr_pages);
2488 if (ret == -EINTR) {
2490 * try_charge() chose to bypass to root due to OOM kill or
2491 * fatal signal. Since our only options are to either fail
2492 * the allocation or charge it to this cgroup, do it as a
2493 * temporary condition. But we can't fail. From a kmem/slab
2494 * perspective, the cache has already been selected, by
2495 * mem_cgroup_kmem_get_cache(), so it is too late to change
2498 * This condition will only trigger if the task entered
2499 * memcg_charge_kmem in a sane state, but was OOM-killed
2500 * during try_charge() above. Tasks that were already dying
2501 * when the allocation triggers should have been already
2502 * directed to the root cgroup in memcontrol.h
2504 page_counter_charge(&memcg->memory, nr_pages);
2505 if (do_swap_account)
2506 page_counter_charge(&memcg->memsw, nr_pages);
2507 css_get_many(&memcg->css, nr_pages);
2510 page_counter_uncharge(&memcg->kmem, nr_pages);
2515 void memcg_uncharge_kmem(struct mem_cgroup *memcg, unsigned long nr_pages)
2517 page_counter_uncharge(&memcg->memory, nr_pages);
2518 if (do_swap_account)
2519 page_counter_uncharge(&memcg->memsw, nr_pages);
2521 page_counter_uncharge(&memcg->kmem, nr_pages);
2523 css_put_many(&memcg->css, nr_pages);
2527 * helper for acessing a memcg's index. It will be used as an index in the
2528 * child cache array in kmem_cache, and also to derive its name. This function
2529 * will return -1 when this is not a kmem-limited memcg.
2531 int memcg_cache_id(struct mem_cgroup *memcg)
2533 return memcg ? memcg->kmemcg_id : -1;
2536 static int memcg_alloc_cache_id(void)
2541 id = ida_simple_get(&kmem_limited_groups,
2542 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2546 if (id < memcg_limited_groups_array_size)
2550 * There's no space for the new id in memcg_caches arrays,
2551 * so we have to grow them.
2554 size = 2 * (id + 1);
2555 if (size < MEMCG_CACHES_MIN_SIZE)
2556 size = MEMCG_CACHES_MIN_SIZE;
2557 else if (size > MEMCG_CACHES_MAX_SIZE)
2558 size = MEMCG_CACHES_MAX_SIZE;
2560 err = memcg_update_all_caches(size);
2562 ida_simple_remove(&kmem_limited_groups, id);
2568 static void memcg_free_cache_id(int id)
2570 ida_simple_remove(&kmem_limited_groups, id);
2574 * We should update the current array size iff all caches updates succeed. This
2575 * can only be done from the slab side. The slab mutex needs to be held when
2578 void memcg_update_array_size(int num)
2580 memcg_limited_groups_array_size = num;
2583 struct memcg_kmem_cache_create_work {
2584 struct mem_cgroup *memcg;
2585 struct kmem_cache *cachep;
2586 struct work_struct work;
2589 static void memcg_kmem_cache_create_func(struct work_struct *w)
2591 struct memcg_kmem_cache_create_work *cw =
2592 container_of(w, struct memcg_kmem_cache_create_work, work);
2593 struct mem_cgroup *memcg = cw->memcg;
2594 struct kmem_cache *cachep = cw->cachep;
2596 memcg_create_kmem_cache(memcg, cachep);
2598 css_put(&memcg->css);
2603 * Enqueue the creation of a per-memcg kmem_cache.
2605 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2606 struct kmem_cache *cachep)
2608 struct memcg_kmem_cache_create_work *cw;
2610 cw = kmalloc(sizeof(*cw), GFP_NOWAIT);
2614 css_get(&memcg->css);
2617 cw->cachep = cachep;
2618 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2620 schedule_work(&cw->work);
2623 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2624 struct kmem_cache *cachep)
2627 * We need to stop accounting when we kmalloc, because if the
2628 * corresponding kmalloc cache is not yet created, the first allocation
2629 * in __memcg_schedule_kmem_cache_create will recurse.
2631 * However, it is better to enclose the whole function. Depending on
2632 * the debugging options enabled, INIT_WORK(), for instance, can
2633 * trigger an allocation. This too, will make us recurse. Because at
2634 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2635 * the safest choice is to do it like this, wrapping the whole function.
2637 current->memcg_kmem_skip_account = 1;
2638 __memcg_schedule_kmem_cache_create(memcg, cachep);
2639 current->memcg_kmem_skip_account = 0;
2643 * Return the kmem_cache we're supposed to use for a slab allocation.
2644 * We try to use the current memcg's version of the cache.
2646 * If the cache does not exist yet, if we are the first user of it,
2647 * we either create it immediately, if possible, or create it asynchronously
2649 * In the latter case, we will let the current allocation go through with
2650 * the original cache.
2652 * Can't be called in interrupt context or from kernel threads.
2653 * This function needs to be called with rcu_read_lock() held.
2655 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep)
2657 struct mem_cgroup *memcg;
2658 struct kmem_cache *memcg_cachep;
2660 VM_BUG_ON(!cachep->memcg_params);
2661 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
2663 if (current->memcg_kmem_skip_account)
2666 memcg = get_mem_cgroup_from_mm(current->mm);
2667 if (!memcg_kmem_is_active(memcg))
2670 memcg_cachep = cache_from_memcg_idx(cachep, memcg_cache_id(memcg));
2671 if (likely(memcg_cachep))
2672 return memcg_cachep;
2675 * If we are in a safe context (can wait, and not in interrupt
2676 * context), we could be be predictable and return right away.
2677 * This would guarantee that the allocation being performed
2678 * already belongs in the new cache.
2680 * However, there are some clashes that can arrive from locking.
2681 * For instance, because we acquire the slab_mutex while doing
2682 * memcg_create_kmem_cache, this means no further allocation
2683 * could happen with the slab_mutex held. So it's better to
2686 memcg_schedule_kmem_cache_create(memcg, cachep);
2688 css_put(&memcg->css);
2692 void __memcg_kmem_put_cache(struct kmem_cache *cachep)
2694 if (!is_root_cache(cachep))
2695 css_put(&cachep->memcg_params->memcg->css);
2699 * We need to verify if the allocation against current->mm->owner's memcg is
2700 * possible for the given order. But the page is not allocated yet, so we'll
2701 * need a further commit step to do the final arrangements.
2703 * It is possible for the task to switch cgroups in this mean time, so at
2704 * commit time, we can't rely on task conversion any longer. We'll then use
2705 * the handle argument to return to the caller which cgroup we should commit
2706 * against. We could also return the memcg directly and avoid the pointer
2707 * passing, but a boolean return value gives better semantics considering
2708 * the compiled-out case as well.
2710 * Returning true means the allocation is possible.
2713 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
2715 struct mem_cgroup *memcg;
2720 memcg = get_mem_cgroup_from_mm(current->mm);
2722 if (!memcg_kmem_is_active(memcg)) {
2723 css_put(&memcg->css);
2727 ret = memcg_charge_kmem(memcg, gfp, 1 << order);
2731 css_put(&memcg->css);
2735 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
2738 VM_BUG_ON(mem_cgroup_is_root(memcg));
2740 /* The page allocation failed. Revert */
2742 memcg_uncharge_kmem(memcg, 1 << order);
2745 page->mem_cgroup = memcg;
2748 void __memcg_kmem_uncharge_pages(struct page *page, int order)
2750 struct mem_cgroup *memcg = page->mem_cgroup;
2755 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2757 memcg_uncharge_kmem(memcg, 1 << order);
2758 page->mem_cgroup = NULL;
2760 #endif /* CONFIG_MEMCG_KMEM */
2762 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2765 * Because tail pages are not marked as "used", set it. We're under
2766 * zone->lru_lock, 'splitting on pmd' and compound_lock.
2767 * charge/uncharge will be never happen and move_account() is done under
2768 * compound_lock(), so we don't have to take care of races.
2770 void mem_cgroup_split_huge_fixup(struct page *head)
2774 if (mem_cgroup_disabled())
2777 for (i = 1; i < HPAGE_PMD_NR; i++)
2778 head[i].mem_cgroup = head->mem_cgroup;
2780 __this_cpu_sub(head->mem_cgroup->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
2783 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2786 * mem_cgroup_move_account - move account of the page
2788 * @nr_pages: number of regular pages (>1 for huge pages)
2789 * @from: mem_cgroup which the page is moved from.
2790 * @to: mem_cgroup which the page is moved to. @from != @to.
2792 * The caller must confirm following.
2793 * - page is not on LRU (isolate_page() is useful.)
2794 * - compound_lock is held when nr_pages > 1
2796 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
2799 static int mem_cgroup_move_account(struct page *page,
2800 unsigned int nr_pages,
2801 struct mem_cgroup *from,
2802 struct mem_cgroup *to)
2804 unsigned long flags;
2807 VM_BUG_ON(from == to);
2808 VM_BUG_ON_PAGE(PageLRU(page), page);
2810 * The page is isolated from LRU. So, collapse function
2811 * will not handle this page. But page splitting can happen.
2812 * Do this check under compound_page_lock(). The caller should
2816 if (nr_pages > 1 && !PageTransHuge(page))
2820 * Prevent mem_cgroup_migrate() from looking at page->mem_cgroup
2821 * of its source page while we change it: page migration takes
2822 * both pages off the LRU, but page cache replacement doesn't.
2824 if (!trylock_page(page))
2828 if (page->mem_cgroup != from)
2831 spin_lock_irqsave(&from->move_lock, flags);
2833 if (!PageAnon(page) && page_mapped(page)) {
2834 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
2836 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
2840 if (PageWriteback(page)) {
2841 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK],
2843 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK],
2848 * It is safe to change page->mem_cgroup here because the page
2849 * is referenced, charged, and isolated - we can't race with
2850 * uncharging, charging, migration, or LRU putback.
2853 /* caller should have done css_get */
2854 page->mem_cgroup = to;
2855 spin_unlock_irqrestore(&from->move_lock, flags);
2859 local_irq_disable();
2860 mem_cgroup_charge_statistics(to, page, nr_pages);
2861 memcg_check_events(to, page);
2862 mem_cgroup_charge_statistics(from, page, -nr_pages);
2863 memcg_check_events(from, page);
2871 #ifdef CONFIG_MEMCG_SWAP
2872 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
2875 int val = (charge) ? 1 : -1;
2876 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
2880 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2881 * @entry: swap entry to be moved
2882 * @from: mem_cgroup which the entry is moved from
2883 * @to: mem_cgroup which the entry is moved to
2885 * It succeeds only when the swap_cgroup's record for this entry is the same
2886 * as the mem_cgroup's id of @from.
2888 * Returns 0 on success, -EINVAL on failure.
2890 * The caller must have charged to @to, IOW, called page_counter_charge() about
2891 * both res and memsw, and called css_get().
2893 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2894 struct mem_cgroup *from, struct mem_cgroup *to)
2896 unsigned short old_id, new_id;
2898 old_id = mem_cgroup_id(from);
2899 new_id = mem_cgroup_id(to);
2901 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2902 mem_cgroup_swap_statistics(from, false);
2903 mem_cgroup_swap_statistics(to, true);
2909 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2910 struct mem_cgroup *from, struct mem_cgroup *to)
2916 static DEFINE_MUTEX(memcg_limit_mutex);
2918 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
2919 unsigned long limit)
2921 unsigned long curusage;
2922 unsigned long oldusage;
2923 bool enlarge = false;
2928 * For keeping hierarchical_reclaim simple, how long we should retry
2929 * is depends on callers. We set our retry-count to be function
2930 * of # of children which we should visit in this loop.
2932 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2933 mem_cgroup_count_children(memcg);
2935 oldusage = page_counter_read(&memcg->memory);
2938 if (signal_pending(current)) {
2943 mutex_lock(&memcg_limit_mutex);
2944 if (limit > memcg->memsw.limit) {
2945 mutex_unlock(&memcg_limit_mutex);
2949 if (limit > memcg->memory.limit)
2951 ret = page_counter_limit(&memcg->memory, limit);
2952 mutex_unlock(&memcg_limit_mutex);
2957 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, true);
2959 curusage = page_counter_read(&memcg->memory);
2960 /* Usage is reduced ? */
2961 if (curusage >= oldusage)
2964 oldusage = curusage;
2965 } while (retry_count);
2967 if (!ret && enlarge)
2968 memcg_oom_recover(memcg);
2973 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
2974 unsigned long limit)
2976 unsigned long curusage;
2977 unsigned long oldusage;
2978 bool enlarge = false;
2982 /* see mem_cgroup_resize_res_limit */
2983 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2984 mem_cgroup_count_children(memcg);
2986 oldusage = page_counter_read(&memcg->memsw);
2989 if (signal_pending(current)) {
2994 mutex_lock(&memcg_limit_mutex);
2995 if (limit < memcg->memory.limit) {
2996 mutex_unlock(&memcg_limit_mutex);
3000 if (limit > memcg->memsw.limit)
3002 ret = page_counter_limit(&memcg->memsw, limit);
3003 mutex_unlock(&memcg_limit_mutex);
3008 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, false);
3010 curusage = page_counter_read(&memcg->memsw);
3011 /* Usage is reduced ? */
3012 if (curusage >= oldusage)
3015 oldusage = curusage;
3016 } while (retry_count);
3018 if (!ret && enlarge)
3019 memcg_oom_recover(memcg);
3024 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3026 unsigned long *total_scanned)
3028 unsigned long nr_reclaimed = 0;
3029 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3030 unsigned long reclaimed;
3032 struct mem_cgroup_tree_per_zone *mctz;
3033 unsigned long excess;
3034 unsigned long nr_scanned;
3039 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3041 * This loop can run a while, specially if mem_cgroup's continuously
3042 * keep exceeding their soft limit and putting the system under
3049 mz = mem_cgroup_largest_soft_limit_node(mctz);
3054 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
3055 gfp_mask, &nr_scanned);
3056 nr_reclaimed += reclaimed;
3057 *total_scanned += nr_scanned;
3058 spin_lock_irq(&mctz->lock);
3059 __mem_cgroup_remove_exceeded(mz, mctz);
3062 * If we failed to reclaim anything from this memory cgroup
3063 * it is time to move on to the next cgroup
3067 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3069 excess = soft_limit_excess(mz->memcg);
3071 * One school of thought says that we should not add
3072 * back the node to the tree if reclaim returns 0.
3073 * But our reclaim could return 0, simply because due
3074 * to priority we are exposing a smaller subset of
3075 * memory to reclaim from. Consider this as a longer
3078 /* If excess == 0, no tree ops */
3079 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3080 spin_unlock_irq(&mctz->lock);
3081 css_put(&mz->memcg->css);
3084 * Could not reclaim anything and there are no more
3085 * mem cgroups to try or we seem to be looping without
3086 * reclaiming anything.
3088 if (!nr_reclaimed &&
3090 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3092 } while (!nr_reclaimed);
3094 css_put(&next_mz->memcg->css);
3095 return nr_reclaimed;
3099 * Test whether @memcg has children, dead or alive. Note that this
3100 * function doesn't care whether @memcg has use_hierarchy enabled and
3101 * returns %true if there are child csses according to the cgroup
3102 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
3104 static inline bool memcg_has_children(struct mem_cgroup *memcg)
3109 * The lock does not prevent addition or deletion of children, but
3110 * it prevents a new child from being initialized based on this
3111 * parent in css_online(), so it's enough to decide whether
3112 * hierarchically inherited attributes can still be changed or not.
3114 lockdep_assert_held(&memcg_create_mutex);
3117 ret = css_next_child(NULL, &memcg->css);
3123 * Reclaims as many pages from the given memcg as possible and moves
3124 * the rest to the parent.
3126 * Caller is responsible for holding css reference for memcg.
3128 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3130 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3132 /* we call try-to-free pages for make this cgroup empty */
3133 lru_add_drain_all();
3134 /* try to free all pages in this cgroup */
3135 while (nr_retries && page_counter_read(&memcg->memory)) {
3138 if (signal_pending(current))
3141 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3145 /* maybe some writeback is necessary */
3146 congestion_wait(BLK_RW_ASYNC, HZ/10);
3154 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3155 char *buf, size_t nbytes,
3158 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3160 if (mem_cgroup_is_root(memcg))
3162 return mem_cgroup_force_empty(memcg) ?: nbytes;
3165 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3168 return mem_cgroup_from_css(css)->use_hierarchy;
3171 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3172 struct cftype *cft, u64 val)
3175 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3176 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
3178 mutex_lock(&memcg_create_mutex);
3180 if (memcg->use_hierarchy == val)
3184 * If parent's use_hierarchy is set, we can't make any modifications
3185 * in the child subtrees. If it is unset, then the change can
3186 * occur, provided the current cgroup has no children.
3188 * For the root cgroup, parent_mem is NULL, we allow value to be
3189 * set if there are no children.
3191 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3192 (val == 1 || val == 0)) {
3193 if (!memcg_has_children(memcg))
3194 memcg->use_hierarchy = val;
3201 mutex_unlock(&memcg_create_mutex);
3206 static unsigned long tree_stat(struct mem_cgroup *memcg,
3207 enum mem_cgroup_stat_index idx)
3209 struct mem_cgroup *iter;
3212 /* Per-cpu values can be negative, use a signed accumulator */
3213 for_each_mem_cgroup_tree(iter, memcg)
3214 val += mem_cgroup_read_stat(iter, idx);
3216 if (val < 0) /* race ? */
3221 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3225 if (mem_cgroup_is_root(memcg)) {
3226 val = tree_stat(memcg, MEM_CGROUP_STAT_CACHE);
3227 val += tree_stat(memcg, MEM_CGROUP_STAT_RSS);
3229 val += tree_stat(memcg, MEM_CGROUP_STAT_SWAP);
3232 val = page_counter_read(&memcg->memory);
3234 val = page_counter_read(&memcg->memsw);
3236 return val << PAGE_SHIFT;
3247 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3250 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3251 struct page_counter *counter;
3253 switch (MEMFILE_TYPE(cft->private)) {
3255 counter = &memcg->memory;
3258 counter = &memcg->memsw;
3261 counter = &memcg->kmem;
3267 switch (MEMFILE_ATTR(cft->private)) {
3269 if (counter == &memcg->memory)
3270 return mem_cgroup_usage(memcg, false);
3271 if (counter == &memcg->memsw)
3272 return mem_cgroup_usage(memcg, true);
3273 return (u64)page_counter_read(counter) * PAGE_SIZE;
3275 return (u64)counter->limit * PAGE_SIZE;
3277 return (u64)counter->watermark * PAGE_SIZE;
3279 return counter->failcnt;
3280 case RES_SOFT_LIMIT:
3281 return (u64)memcg->soft_limit * PAGE_SIZE;
3287 #ifdef CONFIG_MEMCG_KMEM
3288 static int memcg_activate_kmem(struct mem_cgroup *memcg,
3289 unsigned long nr_pages)
3294 if (memcg_kmem_is_active(memcg))
3298 * For simplicity, we won't allow this to be disabled. It also can't
3299 * be changed if the cgroup has children already, or if tasks had
3302 * If tasks join before we set the limit, a person looking at
3303 * kmem.usage_in_bytes will have no way to determine when it took
3304 * place, which makes the value quite meaningless.
3306 * After it first became limited, changes in the value of the limit are
3307 * of course permitted.
3309 mutex_lock(&memcg_create_mutex);
3310 if (cgroup_has_tasks(memcg->css.cgroup) ||
3311 (memcg->use_hierarchy && memcg_has_children(memcg)))
3313 mutex_unlock(&memcg_create_mutex);
3317 memcg_id = memcg_alloc_cache_id();
3324 * We couldn't have accounted to this cgroup, because it hasn't got
3325 * activated yet, so this should succeed.
3327 err = page_counter_limit(&memcg->kmem, nr_pages);
3330 static_key_slow_inc(&memcg_kmem_enabled_key);
3332 * A memory cgroup is considered kmem-active as soon as it gets
3333 * kmemcg_id. Setting the id after enabling static branching will
3334 * guarantee no one starts accounting before all call sites are
3337 memcg->kmemcg_id = memcg_id;
3342 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
3343 unsigned long limit)
3347 mutex_lock(&memcg_limit_mutex);
3348 if (!memcg_kmem_is_active(memcg))
3349 ret = memcg_activate_kmem(memcg, limit);
3351 ret = page_counter_limit(&memcg->kmem, limit);
3352 mutex_unlock(&memcg_limit_mutex);
3356 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
3359 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
3364 mutex_lock(&memcg_limit_mutex);
3366 * If the parent cgroup is not kmem-active now, it cannot be activated
3367 * after this point, because it has at least one child already.
3369 if (memcg_kmem_is_active(parent))
3370 ret = memcg_activate_kmem(memcg, PAGE_COUNTER_MAX);
3371 mutex_unlock(&memcg_limit_mutex);
3375 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
3376 unsigned long limit)
3380 #endif /* CONFIG_MEMCG_KMEM */
3383 * The user of this function is...
3386 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3387 char *buf, size_t nbytes, loff_t off)
3389 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3390 unsigned long nr_pages;
3393 buf = strstrip(buf);
3394 ret = page_counter_memparse(buf, "-1", &nr_pages);
3398 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3400 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3404 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3406 ret = mem_cgroup_resize_limit(memcg, nr_pages);
3409 ret = mem_cgroup_resize_memsw_limit(memcg, nr_pages);
3412 ret = memcg_update_kmem_limit(memcg, nr_pages);
3416 case RES_SOFT_LIMIT:
3417 memcg->soft_limit = nr_pages;
3421 return ret ?: nbytes;
3424 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3425 size_t nbytes, loff_t off)
3427 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3428 struct page_counter *counter;
3430 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3432 counter = &memcg->memory;
3435 counter = &memcg->memsw;
3438 counter = &memcg->kmem;
3444 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3446 page_counter_reset_watermark(counter);
3449 counter->failcnt = 0;
3458 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3461 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3465 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3466 struct cftype *cft, u64 val)
3468 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3470 if (val & ~MOVE_MASK)
3474 * No kind of locking is needed in here, because ->can_attach() will
3475 * check this value once in the beginning of the process, and then carry
3476 * on with stale data. This means that changes to this value will only
3477 * affect task migrations starting after the change.
3479 memcg->move_charge_at_immigrate = val;
3483 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3484 struct cftype *cft, u64 val)
3491 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3495 unsigned int lru_mask;
3498 static const struct numa_stat stats[] = {
3499 { "total", LRU_ALL },
3500 { "file", LRU_ALL_FILE },
3501 { "anon", LRU_ALL_ANON },
3502 { "unevictable", BIT(LRU_UNEVICTABLE) },
3504 const struct numa_stat *stat;
3507 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3509 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3510 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3511 seq_printf(m, "%s=%lu", stat->name, nr);
3512 for_each_node_state(nid, N_MEMORY) {
3513 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3515 seq_printf(m, " N%d=%lu", nid, nr);
3520 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3521 struct mem_cgroup *iter;
3524 for_each_mem_cgroup_tree(iter, memcg)
3525 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3526 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3527 for_each_node_state(nid, N_MEMORY) {
3529 for_each_mem_cgroup_tree(iter, memcg)
3530 nr += mem_cgroup_node_nr_lru_pages(
3531 iter, nid, stat->lru_mask);
3532 seq_printf(m, " N%d=%lu", nid, nr);
3539 #endif /* CONFIG_NUMA */
3541 static int memcg_stat_show(struct seq_file *m, void *v)
3543 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3544 unsigned long memory, memsw;
3545 struct mem_cgroup *mi;
3548 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_stat_names) !=
3549 MEM_CGROUP_STAT_NSTATS);
3550 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_events_names) !=
3551 MEM_CGROUP_EVENTS_NSTATS);
3552 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3554 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3555 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
3557 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
3558 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
3561 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
3562 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
3563 mem_cgroup_read_events(memcg, i));
3565 for (i = 0; i < NR_LRU_LISTS; i++)
3566 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3567 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
3569 /* Hierarchical information */
3570 memory = memsw = PAGE_COUNTER_MAX;
3571 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3572 memory = min(memory, mi->memory.limit);
3573 memsw = min(memsw, mi->memsw.limit);
3575 seq_printf(m, "hierarchical_memory_limit %llu\n",
3576 (u64)memory * PAGE_SIZE);
3577 if (do_swap_account)
3578 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3579 (u64)memsw * PAGE_SIZE);
3581 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3584 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
3586 for_each_mem_cgroup_tree(mi, memcg)
3587 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
3588 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
3591 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
3592 unsigned long long val = 0;
3594 for_each_mem_cgroup_tree(mi, memcg)
3595 val += mem_cgroup_read_events(mi, i);
3596 seq_printf(m, "total_%s %llu\n",
3597 mem_cgroup_events_names[i], val);
3600 for (i = 0; i < NR_LRU_LISTS; i++) {
3601 unsigned long long val = 0;
3603 for_each_mem_cgroup_tree(mi, memcg)
3604 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
3605 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
3608 #ifdef CONFIG_DEBUG_VM
3611 struct mem_cgroup_per_zone *mz;
3612 struct zone_reclaim_stat *rstat;
3613 unsigned long recent_rotated[2] = {0, 0};
3614 unsigned long recent_scanned[2] = {0, 0};
3616 for_each_online_node(nid)
3617 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3618 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
3619 rstat = &mz->lruvec.reclaim_stat;
3621 recent_rotated[0] += rstat->recent_rotated[0];
3622 recent_rotated[1] += rstat->recent_rotated[1];
3623 recent_scanned[0] += rstat->recent_scanned[0];
3624 recent_scanned[1] += rstat->recent_scanned[1];
3626 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3627 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3628 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3629 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3636 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3639 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3641 return mem_cgroup_swappiness(memcg);
3644 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3645 struct cftype *cft, u64 val)
3647 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3653 memcg->swappiness = val;
3655 vm_swappiness = val;
3660 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3662 struct mem_cgroup_threshold_ary *t;
3663 unsigned long usage;
3668 t = rcu_dereference(memcg->thresholds.primary);
3670 t = rcu_dereference(memcg->memsw_thresholds.primary);
3675 usage = mem_cgroup_usage(memcg, swap);
3678 * current_threshold points to threshold just below or equal to usage.
3679 * If it's not true, a threshold was crossed after last
3680 * call of __mem_cgroup_threshold().
3682 i = t->current_threshold;
3685 * Iterate backward over array of thresholds starting from
3686 * current_threshold and check if a threshold is crossed.
3687 * If none of thresholds below usage is crossed, we read
3688 * only one element of the array here.
3690 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3691 eventfd_signal(t->entries[i].eventfd, 1);
3693 /* i = current_threshold + 1 */
3697 * Iterate forward over array of thresholds starting from
3698 * current_threshold+1 and check if a threshold is crossed.
3699 * If none of thresholds above usage is crossed, we read
3700 * only one element of the array here.
3702 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3703 eventfd_signal(t->entries[i].eventfd, 1);
3705 /* Update current_threshold */
3706 t->current_threshold = i - 1;
3711 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3714 __mem_cgroup_threshold(memcg, false);
3715 if (do_swap_account)
3716 __mem_cgroup_threshold(memcg, true);
3718 memcg = parent_mem_cgroup(memcg);
3722 static int compare_thresholds(const void *a, const void *b)
3724 const struct mem_cgroup_threshold *_a = a;
3725 const struct mem_cgroup_threshold *_b = b;
3727 if (_a->threshold > _b->threshold)
3730 if (_a->threshold < _b->threshold)
3736 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3738 struct mem_cgroup_eventfd_list *ev;
3740 spin_lock(&memcg_oom_lock);
3742 list_for_each_entry(ev, &memcg->oom_notify, list)
3743 eventfd_signal(ev->eventfd, 1);
3745 spin_unlock(&memcg_oom_lock);
3749 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3751 struct mem_cgroup *iter;
3753 for_each_mem_cgroup_tree(iter, memcg)
3754 mem_cgroup_oom_notify_cb(iter);
3757 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3758 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3760 struct mem_cgroup_thresholds *thresholds;
3761 struct mem_cgroup_threshold_ary *new;
3762 unsigned long threshold;
3763 unsigned long usage;
3766 ret = page_counter_memparse(args, "-1", &threshold);
3770 mutex_lock(&memcg->thresholds_lock);
3773 thresholds = &memcg->thresholds;
3774 usage = mem_cgroup_usage(memcg, false);
3775 } else if (type == _MEMSWAP) {
3776 thresholds = &memcg->memsw_thresholds;
3777 usage = mem_cgroup_usage(memcg, true);
3781 /* Check if a threshold crossed before adding a new one */
3782 if (thresholds->primary)
3783 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3785 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3787 /* Allocate memory for new array of thresholds */
3788 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3796 /* Copy thresholds (if any) to new array */
3797 if (thresholds->primary) {
3798 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3799 sizeof(struct mem_cgroup_threshold));
3802 /* Add new threshold */
3803 new->entries[size - 1].eventfd = eventfd;
3804 new->entries[size - 1].threshold = threshold;
3806 /* Sort thresholds. Registering of new threshold isn't time-critical */
3807 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3808 compare_thresholds, NULL);
3810 /* Find current threshold */
3811 new->current_threshold = -1;
3812 for (i = 0; i < size; i++) {
3813 if (new->entries[i].threshold <= usage) {
3815 * new->current_threshold will not be used until
3816 * rcu_assign_pointer(), so it's safe to increment
3819 ++new->current_threshold;
3824 /* Free old spare buffer and save old primary buffer as spare */
3825 kfree(thresholds->spare);
3826 thresholds->spare = thresholds->primary;
3828 rcu_assign_pointer(thresholds->primary, new);
3830 /* To be sure that nobody uses thresholds */
3834 mutex_unlock(&memcg->thresholds_lock);
3839 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3840 struct eventfd_ctx *eventfd, const char *args)
3842 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
3845 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
3846 struct eventfd_ctx *eventfd, const char *args)
3848 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
3851 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3852 struct eventfd_ctx *eventfd, enum res_type type)
3854 struct mem_cgroup_thresholds *thresholds;
3855 struct mem_cgroup_threshold_ary *new;
3856 unsigned long usage;
3859 mutex_lock(&memcg->thresholds_lock);
3862 thresholds = &memcg->thresholds;
3863 usage = mem_cgroup_usage(memcg, false);
3864 } else if (type == _MEMSWAP) {
3865 thresholds = &memcg->memsw_thresholds;
3866 usage = mem_cgroup_usage(memcg, true);
3870 if (!thresholds->primary)
3873 /* Check if a threshold crossed before removing */
3874 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3876 /* Calculate new number of threshold */
3878 for (i = 0; i < thresholds->primary->size; i++) {
3879 if (thresholds->primary->entries[i].eventfd != eventfd)
3883 new = thresholds->spare;
3885 /* Set thresholds array to NULL if we don't have thresholds */
3894 /* Copy thresholds and find current threshold */
3895 new->current_threshold = -1;
3896 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3897 if (thresholds->primary->entries[i].eventfd == eventfd)
3900 new->entries[j] = thresholds->primary->entries[i];
3901 if (new->entries[j].threshold <= usage) {
3903 * new->current_threshold will not be used
3904 * until rcu_assign_pointer(), so it's safe to increment
3907 ++new->current_threshold;
3913 /* Swap primary and spare array */
3914 thresholds->spare = thresholds->primary;
3915 /* If all events are unregistered, free the spare array */
3917 kfree(thresholds->spare);
3918 thresholds->spare = NULL;
3921 rcu_assign_pointer(thresholds->primary, new);
3923 /* To be sure that nobody uses thresholds */
3926 mutex_unlock(&memcg->thresholds_lock);
3929 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3930 struct eventfd_ctx *eventfd)
3932 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
3935 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3936 struct eventfd_ctx *eventfd)
3938 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
3941 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
3942 struct eventfd_ctx *eventfd, const char *args)
3944 struct mem_cgroup_eventfd_list *event;
3946 event = kmalloc(sizeof(*event), GFP_KERNEL);
3950 spin_lock(&memcg_oom_lock);
3952 event->eventfd = eventfd;
3953 list_add(&event->list, &memcg->oom_notify);
3955 /* already in OOM ? */
3956 if (atomic_read(&memcg->under_oom))
3957 eventfd_signal(eventfd, 1);
3958 spin_unlock(&memcg_oom_lock);
3963 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
3964 struct eventfd_ctx *eventfd)
3966 struct mem_cgroup_eventfd_list *ev, *tmp;
3968 spin_lock(&memcg_oom_lock);
3970 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
3971 if (ev->eventfd == eventfd) {
3972 list_del(&ev->list);
3977 spin_unlock(&memcg_oom_lock);
3980 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
3982 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
3984 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
3985 seq_printf(sf, "under_oom %d\n", (bool)atomic_read(&memcg->under_oom));
3989 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
3990 struct cftype *cft, u64 val)
3992 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3994 /* cannot set to root cgroup and only 0 and 1 are allowed */
3995 if (!css->parent || !((val == 0) || (val == 1)))
3998 memcg->oom_kill_disable = val;
4000 memcg_oom_recover(memcg);
4005 #ifdef CONFIG_MEMCG_KMEM
4006 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
4010 ret = memcg_propagate_kmem(memcg);
4014 return mem_cgroup_sockets_init(memcg, ss);
4017 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
4019 memcg_destroy_kmem_caches(memcg);
4020 mem_cgroup_sockets_destroy(memcg);
4023 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
4028 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
4034 * DO NOT USE IN NEW FILES.
4036 * "cgroup.event_control" implementation.
4038 * This is way over-engineered. It tries to support fully configurable
4039 * events for each user. Such level of flexibility is completely
4040 * unnecessary especially in the light of the planned unified hierarchy.
4042 * Please deprecate this and replace with something simpler if at all
4047 * Unregister event and free resources.
4049 * Gets called from workqueue.
4051 static void memcg_event_remove(struct work_struct *work)
4053 struct mem_cgroup_event *event =
4054 container_of(work, struct mem_cgroup_event, remove);
4055 struct mem_cgroup *memcg = event->memcg;
4057 remove_wait_queue(event->wqh, &event->wait);
4059 event->unregister_event(memcg, event->eventfd);
4061 /* Notify userspace the event is going away. */
4062 eventfd_signal(event->eventfd, 1);
4064 eventfd_ctx_put(event->eventfd);
4066 css_put(&memcg->css);
4070 * Gets called on POLLHUP on eventfd when user closes it.
4072 * Called with wqh->lock held and interrupts disabled.
4074 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
4075 int sync, void *key)
4077 struct mem_cgroup_event *event =
4078 container_of(wait, struct mem_cgroup_event, wait);
4079 struct mem_cgroup *memcg = event->memcg;
4080 unsigned long flags = (unsigned long)key;
4082 if (flags & POLLHUP) {
4084 * If the event has been detached at cgroup removal, we
4085 * can simply return knowing the other side will cleanup
4088 * We can't race against event freeing since the other
4089 * side will require wqh->lock via remove_wait_queue(),
4092 spin_lock(&memcg->event_list_lock);
4093 if (!list_empty(&event->list)) {
4094 list_del_init(&event->list);
4096 * We are in atomic context, but cgroup_event_remove()
4097 * may sleep, so we have to call it in workqueue.
4099 schedule_work(&event->remove);
4101 spin_unlock(&memcg->event_list_lock);
4107 static void memcg_event_ptable_queue_proc(struct file *file,
4108 wait_queue_head_t *wqh, poll_table *pt)
4110 struct mem_cgroup_event *event =
4111 container_of(pt, struct mem_cgroup_event, pt);
4114 add_wait_queue(wqh, &event->wait);
4118 * DO NOT USE IN NEW FILES.
4120 * Parse input and register new cgroup event handler.
4122 * Input must be in format '<event_fd> <control_fd> <args>'.
4123 * Interpretation of args is defined by control file implementation.
4125 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4126 char *buf, size_t nbytes, loff_t off)
4128 struct cgroup_subsys_state *css = of_css(of);
4129 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4130 struct mem_cgroup_event *event;
4131 struct cgroup_subsys_state *cfile_css;
4132 unsigned int efd, cfd;
4139 buf = strstrip(buf);
4141 efd = simple_strtoul(buf, &endp, 10);
4146 cfd = simple_strtoul(buf, &endp, 10);
4147 if ((*endp != ' ') && (*endp != '\0'))
4151 event = kzalloc(sizeof(*event), GFP_KERNEL);
4155 event->memcg = memcg;
4156 INIT_LIST_HEAD(&event->list);
4157 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4158 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4159 INIT_WORK(&event->remove, memcg_event_remove);
4167 event->eventfd = eventfd_ctx_fileget(efile.file);
4168 if (IS_ERR(event->eventfd)) {
4169 ret = PTR_ERR(event->eventfd);
4176 goto out_put_eventfd;
4179 /* the process need read permission on control file */
4180 /* AV: shouldn't we check that it's been opened for read instead? */
4181 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4186 * Determine the event callbacks and set them in @event. This used
4187 * to be done via struct cftype but cgroup core no longer knows
4188 * about these events. The following is crude but the whole thing
4189 * is for compatibility anyway.
4191 * DO NOT ADD NEW FILES.
4193 name = cfile.file->f_path.dentry->d_name.name;
4195 if (!strcmp(name, "memory.usage_in_bytes")) {
4196 event->register_event = mem_cgroup_usage_register_event;
4197 event->unregister_event = mem_cgroup_usage_unregister_event;
4198 } else if (!strcmp(name, "memory.oom_control")) {
4199 event->register_event = mem_cgroup_oom_register_event;
4200 event->unregister_event = mem_cgroup_oom_unregister_event;
4201 } else if (!strcmp(name, "memory.pressure_level")) {
4202 event->register_event = vmpressure_register_event;
4203 event->unregister_event = vmpressure_unregister_event;
4204 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4205 event->register_event = memsw_cgroup_usage_register_event;
4206 event->unregister_event = memsw_cgroup_usage_unregister_event;
4213 * Verify @cfile should belong to @css. Also, remaining events are
4214 * automatically removed on cgroup destruction but the removal is
4215 * asynchronous, so take an extra ref on @css.
4217 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4218 &memory_cgrp_subsys);
4220 if (IS_ERR(cfile_css))
4222 if (cfile_css != css) {
4227 ret = event->register_event(memcg, event->eventfd, buf);
4231 efile.file->f_op->poll(efile.file, &event->pt);
4233 spin_lock(&memcg->event_list_lock);
4234 list_add(&event->list, &memcg->event_list);
4235 spin_unlock(&memcg->event_list_lock);
4247 eventfd_ctx_put(event->eventfd);
4256 static struct cftype mem_cgroup_legacy_files[] = {
4258 .name = "usage_in_bytes",
4259 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4260 .read_u64 = mem_cgroup_read_u64,
4263 .name = "max_usage_in_bytes",
4264 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4265 .write = mem_cgroup_reset,
4266 .read_u64 = mem_cgroup_read_u64,
4269 .name = "limit_in_bytes",
4270 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4271 .write = mem_cgroup_write,
4272 .read_u64 = mem_cgroup_read_u64,
4275 .name = "soft_limit_in_bytes",
4276 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4277 .write = mem_cgroup_write,
4278 .read_u64 = mem_cgroup_read_u64,
4282 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4283 .write = mem_cgroup_reset,
4284 .read_u64 = mem_cgroup_read_u64,
4288 .seq_show = memcg_stat_show,
4291 .name = "force_empty",
4292 .write = mem_cgroup_force_empty_write,
4295 .name = "use_hierarchy",
4296 .write_u64 = mem_cgroup_hierarchy_write,
4297 .read_u64 = mem_cgroup_hierarchy_read,
4300 .name = "cgroup.event_control", /* XXX: for compat */
4301 .write = memcg_write_event_control,
4302 .flags = CFTYPE_NO_PREFIX,
4306 .name = "swappiness",
4307 .read_u64 = mem_cgroup_swappiness_read,
4308 .write_u64 = mem_cgroup_swappiness_write,
4311 .name = "move_charge_at_immigrate",
4312 .read_u64 = mem_cgroup_move_charge_read,
4313 .write_u64 = mem_cgroup_move_charge_write,
4316 .name = "oom_control",
4317 .seq_show = mem_cgroup_oom_control_read,
4318 .write_u64 = mem_cgroup_oom_control_write,
4319 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4322 .name = "pressure_level",
4326 .name = "numa_stat",
4327 .seq_show = memcg_numa_stat_show,
4330 #ifdef CONFIG_MEMCG_KMEM
4332 .name = "kmem.limit_in_bytes",
4333 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4334 .write = mem_cgroup_write,
4335 .read_u64 = mem_cgroup_read_u64,
4338 .name = "kmem.usage_in_bytes",
4339 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4340 .read_u64 = mem_cgroup_read_u64,
4343 .name = "kmem.failcnt",
4344 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4345 .write = mem_cgroup_reset,
4346 .read_u64 = mem_cgroup_read_u64,
4349 .name = "kmem.max_usage_in_bytes",
4350 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4351 .write = mem_cgroup_reset,
4352 .read_u64 = mem_cgroup_read_u64,
4354 #ifdef CONFIG_SLABINFO
4356 .name = "kmem.slabinfo",
4357 .seq_start = slab_start,
4358 .seq_next = slab_next,
4359 .seq_stop = slab_stop,
4360 .seq_show = memcg_slab_show,
4364 { }, /* terminate */
4367 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4369 struct mem_cgroup_per_node *pn;
4370 struct mem_cgroup_per_zone *mz;
4371 int zone, tmp = node;
4373 * This routine is called against possible nodes.
4374 * But it's BUG to call kmalloc() against offline node.
4376 * TODO: this routine can waste much memory for nodes which will
4377 * never be onlined. It's better to use memory hotplug callback
4380 if (!node_state(node, N_NORMAL_MEMORY))
4382 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4386 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4387 mz = &pn->zoneinfo[zone];
4388 lruvec_init(&mz->lruvec);
4389 mz->usage_in_excess = 0;
4390 mz->on_tree = false;
4393 memcg->nodeinfo[node] = pn;
4397 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4399 kfree(memcg->nodeinfo[node]);
4402 static struct mem_cgroup *mem_cgroup_alloc(void)
4404 struct mem_cgroup *memcg;
4407 size = sizeof(struct mem_cgroup);
4408 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4410 memcg = kzalloc(size, GFP_KERNEL);
4414 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4417 spin_lock_init(&memcg->pcp_counter_lock);
4426 * At destroying mem_cgroup, references from swap_cgroup can remain.
4427 * (scanning all at force_empty is too costly...)
4429 * Instead of clearing all references at force_empty, we remember
4430 * the number of reference from swap_cgroup and free mem_cgroup when
4431 * it goes down to 0.
4433 * Removal of cgroup itself succeeds regardless of refs from swap.
4436 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4440 mem_cgroup_remove_from_trees(memcg);
4443 free_mem_cgroup_per_zone_info(memcg, node);
4445 free_percpu(memcg->stat);
4447 disarm_static_keys(memcg);
4452 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4454 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4456 if (!memcg->memory.parent)
4458 return mem_cgroup_from_counter(memcg->memory.parent, memory);
4460 EXPORT_SYMBOL(parent_mem_cgroup);
4462 static struct cgroup_subsys_state * __ref
4463 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4465 struct mem_cgroup *memcg;
4466 long error = -ENOMEM;
4469 memcg = mem_cgroup_alloc();
4471 return ERR_PTR(error);
4474 if (alloc_mem_cgroup_per_zone_info(memcg, node))
4478 if (parent_css == NULL) {
4479 root_mem_cgroup = memcg;
4480 page_counter_init(&memcg->memory, NULL);
4481 memcg->high = PAGE_COUNTER_MAX;
4482 memcg->soft_limit = PAGE_COUNTER_MAX;
4483 page_counter_init(&memcg->memsw, NULL);
4484 page_counter_init(&memcg->kmem, NULL);
4487 memcg->last_scanned_node = MAX_NUMNODES;
4488 INIT_LIST_HEAD(&memcg->oom_notify);
4489 memcg->move_charge_at_immigrate = 0;
4490 mutex_init(&memcg->thresholds_lock);
4491 spin_lock_init(&memcg->move_lock);
4492 vmpressure_init(&memcg->vmpressure);
4493 INIT_LIST_HEAD(&memcg->event_list);
4494 spin_lock_init(&memcg->event_list_lock);
4495 #ifdef CONFIG_MEMCG_KMEM
4496 memcg->kmemcg_id = -1;
4502 __mem_cgroup_free(memcg);
4503 return ERR_PTR(error);
4507 mem_cgroup_css_online(struct cgroup_subsys_state *css)
4509 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4510 struct mem_cgroup *parent = mem_cgroup_from_css(css->parent);
4513 if (css->id > MEM_CGROUP_ID_MAX)
4519 mutex_lock(&memcg_create_mutex);
4521 memcg->use_hierarchy = parent->use_hierarchy;
4522 memcg->oom_kill_disable = parent->oom_kill_disable;
4523 memcg->swappiness = mem_cgroup_swappiness(parent);
4525 if (parent->use_hierarchy) {
4526 page_counter_init(&memcg->memory, &parent->memory);
4527 memcg->high = PAGE_COUNTER_MAX;
4528 memcg->soft_limit = PAGE_COUNTER_MAX;
4529 page_counter_init(&memcg->memsw, &parent->memsw);
4530 page_counter_init(&memcg->kmem, &parent->kmem);
4533 * No need to take a reference to the parent because cgroup
4534 * core guarantees its existence.
4537 page_counter_init(&memcg->memory, NULL);
4538 memcg->high = PAGE_COUNTER_MAX;
4539 memcg->soft_limit = PAGE_COUNTER_MAX;
4540 page_counter_init(&memcg->memsw, NULL);
4541 page_counter_init(&memcg->kmem, NULL);
4543 * Deeper hierachy with use_hierarchy == false doesn't make
4544 * much sense so let cgroup subsystem know about this
4545 * unfortunate state in our controller.
4547 if (parent != root_mem_cgroup)
4548 memory_cgrp_subsys.broken_hierarchy = true;
4550 mutex_unlock(&memcg_create_mutex);
4552 ret = memcg_init_kmem(memcg, &memory_cgrp_subsys);
4557 * Make sure the memcg is initialized: mem_cgroup_iter()
4558 * orders reading memcg->initialized against its callers
4559 * reading the memcg members.
4561 smp_store_release(&memcg->initialized, 1);
4566 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4568 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4569 struct mem_cgroup_event *event, *tmp;
4572 * Unregister events and notify userspace.
4573 * Notify userspace about cgroup removing only after rmdir of cgroup
4574 * directory to avoid race between userspace and kernelspace.
4576 spin_lock(&memcg->event_list_lock);
4577 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4578 list_del_init(&event->list);
4579 schedule_work(&event->remove);
4581 spin_unlock(&memcg->event_list_lock);
4583 vmpressure_cleanup(&memcg->vmpressure);
4586 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4588 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4590 memcg_destroy_kmem(memcg);
4591 __mem_cgroup_free(memcg);
4595 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4596 * @css: the target css
4598 * Reset the states of the mem_cgroup associated with @css. This is
4599 * invoked when the userland requests disabling on the default hierarchy
4600 * but the memcg is pinned through dependency. The memcg should stop
4601 * applying policies and should revert to the vanilla state as it may be
4602 * made visible again.
4604 * The current implementation only resets the essential configurations.
4605 * This needs to be expanded to cover all the visible parts.
4607 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4609 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4611 mem_cgroup_resize_limit(memcg, PAGE_COUNTER_MAX);
4612 mem_cgroup_resize_memsw_limit(memcg, PAGE_COUNTER_MAX);
4613 memcg_update_kmem_limit(memcg, PAGE_COUNTER_MAX);
4615 memcg->high = PAGE_COUNTER_MAX;
4616 memcg->soft_limit = PAGE_COUNTER_MAX;
4620 /* Handlers for move charge at task migration. */
4621 static int mem_cgroup_do_precharge(unsigned long count)
4625 /* Try a single bulk charge without reclaim first */
4626 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_WAIT, count);
4628 mc.precharge += count;
4631 if (ret == -EINTR) {
4632 cancel_charge(root_mem_cgroup, count);
4636 /* Try charges one by one with reclaim */
4638 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_NORETRY, 1);
4640 * In case of failure, any residual charges against
4641 * mc.to will be dropped by mem_cgroup_clear_mc()
4642 * later on. However, cancel any charges that are
4643 * bypassed to root right away or they'll be lost.
4646 cancel_charge(root_mem_cgroup, 1);
4656 * get_mctgt_type - get target type of moving charge
4657 * @vma: the vma the pte to be checked belongs
4658 * @addr: the address corresponding to the pte to be checked
4659 * @ptent: the pte to be checked
4660 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4663 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4664 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4665 * move charge. if @target is not NULL, the page is stored in target->page
4666 * with extra refcnt got(Callers should handle it).
4667 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4668 * target for charge migration. if @target is not NULL, the entry is stored
4671 * Called with pte lock held.
4678 enum mc_target_type {
4684 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4685 unsigned long addr, pte_t ptent)
4687 struct page *page = vm_normal_page(vma, addr, ptent);
4689 if (!page || !page_mapped(page))
4691 if (PageAnon(page)) {
4692 if (!(mc.flags & MOVE_ANON))
4695 if (!(mc.flags & MOVE_FILE))
4698 if (!get_page_unless_zero(page))
4705 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4706 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4708 struct page *page = NULL;
4709 swp_entry_t ent = pte_to_swp_entry(ptent);
4711 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
4714 * Because lookup_swap_cache() updates some statistics counter,
4715 * we call find_get_page() with swapper_space directly.
4717 page = find_get_page(swap_address_space(ent), ent.val);
4718 if (do_swap_account)
4719 entry->val = ent.val;
4724 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4725 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4731 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4732 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4734 struct page *page = NULL;
4735 struct address_space *mapping;
4738 if (!vma->vm_file) /* anonymous vma */
4740 if (!(mc.flags & MOVE_FILE))
4743 mapping = vma->vm_file->f_mapping;
4744 pgoff = linear_page_index(vma, addr);
4746 /* page is moved even if it's not RSS of this task(page-faulted). */
4748 /* shmem/tmpfs may report page out on swap: account for that too. */
4749 if (shmem_mapping(mapping)) {
4750 page = find_get_entry(mapping, pgoff);
4751 if (radix_tree_exceptional_entry(page)) {
4752 swp_entry_t swp = radix_to_swp_entry(page);
4753 if (do_swap_account)
4755 page = find_get_page(swap_address_space(swp), swp.val);
4758 page = find_get_page(mapping, pgoff);
4760 page = find_get_page(mapping, pgoff);
4765 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
4766 unsigned long addr, pte_t ptent, union mc_target *target)
4768 struct page *page = NULL;
4769 enum mc_target_type ret = MC_TARGET_NONE;
4770 swp_entry_t ent = { .val = 0 };
4772 if (pte_present(ptent))
4773 page = mc_handle_present_pte(vma, addr, ptent);
4774 else if (is_swap_pte(ptent))
4775 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
4776 else if (pte_none(ptent))
4777 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4779 if (!page && !ent.val)
4783 * Do only loose check w/o serialization.
4784 * mem_cgroup_move_account() checks the page is valid or
4785 * not under LRU exclusion.
4787 if (page->mem_cgroup == mc.from) {
4788 ret = MC_TARGET_PAGE;
4790 target->page = page;
4792 if (!ret || !target)
4795 /* There is a swap entry and a page doesn't exist or isn't charged */
4796 if (ent.val && !ret &&
4797 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
4798 ret = MC_TARGET_SWAP;
4805 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4807 * We don't consider swapping or file mapped pages because THP does not
4808 * support them for now.
4809 * Caller should make sure that pmd_trans_huge(pmd) is true.
4811 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4812 unsigned long addr, pmd_t pmd, union mc_target *target)
4814 struct page *page = NULL;
4815 enum mc_target_type ret = MC_TARGET_NONE;
4817 page = pmd_page(pmd);
4818 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
4819 if (!(mc.flags & MOVE_ANON))
4821 if (page->mem_cgroup == mc.from) {
4822 ret = MC_TARGET_PAGE;
4825 target->page = page;
4831 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4832 unsigned long addr, pmd_t pmd, union mc_target *target)
4834 return MC_TARGET_NONE;
4838 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4839 unsigned long addr, unsigned long end,
4840 struct mm_walk *walk)
4842 struct vm_area_struct *vma = walk->private;
4846 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
4847 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
4848 mc.precharge += HPAGE_PMD_NR;
4853 if (pmd_trans_unstable(pmd))
4855 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4856 for (; addr != end; pte++, addr += PAGE_SIZE)
4857 if (get_mctgt_type(vma, addr, *pte, NULL))
4858 mc.precharge++; /* increment precharge temporarily */
4859 pte_unmap_unlock(pte - 1, ptl);
4865 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4867 unsigned long precharge;
4868 struct vm_area_struct *vma;
4870 down_read(&mm->mmap_sem);
4871 for (vma = mm->mmap; vma; vma = vma->vm_next) {
4872 struct mm_walk mem_cgroup_count_precharge_walk = {
4873 .pmd_entry = mem_cgroup_count_precharge_pte_range,
4877 if (is_vm_hugetlb_page(vma))
4879 walk_page_range(vma->vm_start, vma->vm_end,
4880 &mem_cgroup_count_precharge_walk);
4882 up_read(&mm->mmap_sem);
4884 precharge = mc.precharge;
4890 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4892 unsigned long precharge = mem_cgroup_count_precharge(mm);
4894 VM_BUG_ON(mc.moving_task);
4895 mc.moving_task = current;
4896 return mem_cgroup_do_precharge(precharge);
4899 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4900 static void __mem_cgroup_clear_mc(void)
4902 struct mem_cgroup *from = mc.from;
4903 struct mem_cgroup *to = mc.to;
4905 /* we must uncharge all the leftover precharges from mc.to */
4907 cancel_charge(mc.to, mc.precharge);
4911 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4912 * we must uncharge here.
4914 if (mc.moved_charge) {
4915 cancel_charge(mc.from, mc.moved_charge);
4916 mc.moved_charge = 0;
4918 /* we must fixup refcnts and charges */
4919 if (mc.moved_swap) {
4920 /* uncharge swap account from the old cgroup */
4921 if (!mem_cgroup_is_root(mc.from))
4922 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
4925 * we charged both to->memory and to->memsw, so we
4926 * should uncharge to->memory.
4928 if (!mem_cgroup_is_root(mc.to))
4929 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
4931 css_put_many(&mc.from->css, mc.moved_swap);
4933 /* we've already done css_get(mc.to) */
4936 memcg_oom_recover(from);
4937 memcg_oom_recover(to);
4938 wake_up_all(&mc.waitq);
4941 static void mem_cgroup_clear_mc(void)
4944 * we must clear moving_task before waking up waiters at the end of
4947 mc.moving_task = NULL;
4948 __mem_cgroup_clear_mc();
4949 spin_lock(&mc.lock);
4952 spin_unlock(&mc.lock);
4955 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
4956 struct cgroup_taskset *tset)
4958 struct task_struct *p = cgroup_taskset_first(tset);
4960 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4961 unsigned long move_flags;
4964 * We are now commited to this value whatever it is. Changes in this
4965 * tunable will only affect upcoming migrations, not the current one.
4966 * So we need to save it, and keep it going.
4968 move_flags = ACCESS_ONCE(memcg->move_charge_at_immigrate);
4970 struct mm_struct *mm;
4971 struct mem_cgroup *from = mem_cgroup_from_task(p);
4973 VM_BUG_ON(from == memcg);
4975 mm = get_task_mm(p);
4978 /* We move charges only when we move a owner of the mm */
4979 if (mm->owner == p) {
4982 VM_BUG_ON(mc.precharge);
4983 VM_BUG_ON(mc.moved_charge);
4984 VM_BUG_ON(mc.moved_swap);
4986 spin_lock(&mc.lock);
4989 mc.flags = move_flags;
4990 spin_unlock(&mc.lock);
4991 /* We set mc.moving_task later */
4993 ret = mem_cgroup_precharge_mc(mm);
4995 mem_cgroup_clear_mc();
5002 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
5003 struct cgroup_taskset *tset)
5006 mem_cgroup_clear_mc();
5009 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5010 unsigned long addr, unsigned long end,
5011 struct mm_walk *walk)
5014 struct vm_area_struct *vma = walk->private;
5017 enum mc_target_type target_type;
5018 union mc_target target;
5022 * We don't take compound_lock() here but no race with splitting thp
5024 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
5025 * under splitting, which means there's no concurrent thp split,
5026 * - if another thread runs into split_huge_page() just after we
5027 * entered this if-block, the thread must wait for page table lock
5028 * to be unlocked in __split_huge_page_splitting(), where the main
5029 * part of thp split is not executed yet.
5031 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
5032 if (mc.precharge < HPAGE_PMD_NR) {
5036 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5037 if (target_type == MC_TARGET_PAGE) {
5039 if (!isolate_lru_page(page)) {
5040 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
5042 mc.precharge -= HPAGE_PMD_NR;
5043 mc.moved_charge += HPAGE_PMD_NR;
5045 putback_lru_page(page);
5053 if (pmd_trans_unstable(pmd))
5056 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5057 for (; addr != end; addr += PAGE_SIZE) {
5058 pte_t ptent = *(pte++);
5064 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5065 case MC_TARGET_PAGE:
5067 if (isolate_lru_page(page))
5069 if (!mem_cgroup_move_account(page, 1, mc.from, mc.to)) {
5071 /* we uncharge from mc.from later. */
5074 putback_lru_page(page);
5075 put: /* get_mctgt_type() gets the page */
5078 case MC_TARGET_SWAP:
5080 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5082 /* we fixup refcnts and charges later. */
5090 pte_unmap_unlock(pte - 1, ptl);
5095 * We have consumed all precharges we got in can_attach().
5096 * We try charge one by one, but don't do any additional
5097 * charges to mc.to if we have failed in charge once in attach()
5100 ret = mem_cgroup_do_precharge(1);
5108 static void mem_cgroup_move_charge(struct mm_struct *mm)
5110 struct vm_area_struct *vma;
5112 lru_add_drain_all();
5114 * Signal mem_cgroup_begin_page_stat() to take the memcg's
5115 * move_lock while we're moving its pages to another memcg.
5116 * Then wait for already started RCU-only updates to finish.
5118 atomic_inc(&mc.from->moving_account);
5121 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5123 * Someone who are holding the mmap_sem might be waiting in
5124 * waitq. So we cancel all extra charges, wake up all waiters,
5125 * and retry. Because we cancel precharges, we might not be able
5126 * to move enough charges, but moving charge is a best-effort
5127 * feature anyway, so it wouldn't be a big problem.
5129 __mem_cgroup_clear_mc();
5133 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5135 struct mm_walk mem_cgroup_move_charge_walk = {
5136 .pmd_entry = mem_cgroup_move_charge_pte_range,
5140 if (is_vm_hugetlb_page(vma))
5142 ret = walk_page_range(vma->vm_start, vma->vm_end,
5143 &mem_cgroup_move_charge_walk);
5146 * means we have consumed all precharges and failed in
5147 * doing additional charge. Just abandon here.
5151 up_read(&mm->mmap_sem);
5152 atomic_dec(&mc.from->moving_account);
5155 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
5156 struct cgroup_taskset *tset)
5158 struct task_struct *p = cgroup_taskset_first(tset);
5159 struct mm_struct *mm = get_task_mm(p);
5163 mem_cgroup_move_charge(mm);
5167 mem_cgroup_clear_mc();
5169 #else /* !CONFIG_MMU */
5170 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
5171 struct cgroup_taskset *tset)
5175 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
5176 struct cgroup_taskset *tset)
5179 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
5180 struct cgroup_taskset *tset)
5186 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5187 * to verify whether we're attached to the default hierarchy on each mount
5190 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5193 * use_hierarchy is forced on the default hierarchy. cgroup core
5194 * guarantees that @root doesn't have any children, so turning it
5195 * on for the root memcg is enough.
5197 if (cgroup_on_dfl(root_css->cgroup))
5198 mem_cgroup_from_css(root_css)->use_hierarchy = true;
5201 static u64 memory_current_read(struct cgroup_subsys_state *css,
5204 return mem_cgroup_usage(mem_cgroup_from_css(css), false);
5207 static int memory_low_show(struct seq_file *m, void *v)
5209 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5210 unsigned long low = ACCESS_ONCE(memcg->low);
5212 if (low == PAGE_COUNTER_MAX)
5213 seq_puts(m, "infinity\n");
5215 seq_printf(m, "%llu\n", (u64)low * PAGE_SIZE);
5220 static ssize_t memory_low_write(struct kernfs_open_file *of,
5221 char *buf, size_t nbytes, loff_t off)
5223 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5227 buf = strstrip(buf);
5228 err = page_counter_memparse(buf, "infinity", &low);
5237 static int memory_high_show(struct seq_file *m, void *v)
5239 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5240 unsigned long high = ACCESS_ONCE(memcg->high);
5242 if (high == PAGE_COUNTER_MAX)
5243 seq_puts(m, "infinity\n");
5245 seq_printf(m, "%llu\n", (u64)high * PAGE_SIZE);
5250 static ssize_t memory_high_write(struct kernfs_open_file *of,
5251 char *buf, size_t nbytes, loff_t off)
5253 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5257 buf = strstrip(buf);
5258 err = page_counter_memparse(buf, "infinity", &high);
5267 static int memory_max_show(struct seq_file *m, void *v)
5269 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5270 unsigned long max = ACCESS_ONCE(memcg->memory.limit);
5272 if (max == PAGE_COUNTER_MAX)
5273 seq_puts(m, "infinity\n");
5275 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
5280 static ssize_t memory_max_write(struct kernfs_open_file *of,
5281 char *buf, size_t nbytes, loff_t off)
5283 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5287 buf = strstrip(buf);
5288 err = page_counter_memparse(buf, "infinity", &max);
5292 err = mem_cgroup_resize_limit(memcg, max);
5299 static int memory_events_show(struct seq_file *m, void *v)
5301 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5303 seq_printf(m, "low %lu\n", mem_cgroup_read_events(memcg, MEMCG_LOW));
5304 seq_printf(m, "high %lu\n", mem_cgroup_read_events(memcg, MEMCG_HIGH));
5305 seq_printf(m, "max %lu\n", mem_cgroup_read_events(memcg, MEMCG_MAX));
5306 seq_printf(m, "oom %lu\n", mem_cgroup_read_events(memcg, MEMCG_OOM));
5311 static struct cftype memory_files[] = {
5314 .read_u64 = memory_current_read,
5318 .flags = CFTYPE_NOT_ON_ROOT,
5319 .seq_show = memory_low_show,
5320 .write = memory_low_write,
5324 .flags = CFTYPE_NOT_ON_ROOT,
5325 .seq_show = memory_high_show,
5326 .write = memory_high_write,
5330 .flags = CFTYPE_NOT_ON_ROOT,
5331 .seq_show = memory_max_show,
5332 .write = memory_max_write,
5336 .flags = CFTYPE_NOT_ON_ROOT,
5337 .seq_show = memory_events_show,
5342 struct cgroup_subsys memory_cgrp_subsys = {
5343 .css_alloc = mem_cgroup_css_alloc,
5344 .css_online = mem_cgroup_css_online,
5345 .css_offline = mem_cgroup_css_offline,
5346 .css_free = mem_cgroup_css_free,
5347 .css_reset = mem_cgroup_css_reset,
5348 .can_attach = mem_cgroup_can_attach,
5349 .cancel_attach = mem_cgroup_cancel_attach,
5350 .attach = mem_cgroup_move_task,
5351 .bind = mem_cgroup_bind,
5352 .dfl_cftypes = memory_files,
5353 .legacy_cftypes = mem_cgroup_legacy_files,
5358 * mem_cgroup_events - count memory events against a cgroup
5359 * @memcg: the memory cgroup
5360 * @idx: the event index
5361 * @nr: the number of events to account for
5363 void mem_cgroup_events(struct mem_cgroup *memcg,
5364 enum mem_cgroup_events_index idx,
5367 this_cpu_add(memcg->stat->events[idx], nr);
5371 * mem_cgroup_low - check if memory consumption is below the normal range
5372 * @root: the highest ancestor to consider
5373 * @memcg: the memory cgroup to check
5375 * Returns %true if memory consumption of @memcg, and that of all
5376 * configurable ancestors up to @root, is below the normal range.
5378 bool mem_cgroup_low(struct mem_cgroup *root, struct mem_cgroup *memcg)
5380 if (mem_cgroup_disabled())
5384 * The toplevel group doesn't have a configurable range, so
5385 * it's never low when looked at directly, and it is not
5386 * considered an ancestor when assessing the hierarchy.
5389 if (memcg == root_mem_cgroup)
5392 if (page_counter_read(&memcg->memory) > memcg->low)
5395 while (memcg != root) {
5396 memcg = parent_mem_cgroup(memcg);
5398 if (memcg == root_mem_cgroup)
5401 if (page_counter_read(&memcg->memory) > memcg->low)
5408 * mem_cgroup_try_charge - try charging a page
5409 * @page: page to charge
5410 * @mm: mm context of the victim
5411 * @gfp_mask: reclaim mode
5412 * @memcgp: charged memcg return
5414 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5415 * pages according to @gfp_mask if necessary.
5417 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5418 * Otherwise, an error code is returned.
5420 * After page->mapping has been set up, the caller must finalize the
5421 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5422 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5424 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
5425 gfp_t gfp_mask, struct mem_cgroup **memcgp)
5427 struct mem_cgroup *memcg = NULL;
5428 unsigned int nr_pages = 1;
5431 if (mem_cgroup_disabled())
5434 if (PageSwapCache(page)) {
5436 * Every swap fault against a single page tries to charge the
5437 * page, bail as early as possible. shmem_unuse() encounters
5438 * already charged pages, too. The USED bit is protected by
5439 * the page lock, which serializes swap cache removal, which
5440 * in turn serializes uncharging.
5442 if (page->mem_cgroup)
5446 if (PageTransHuge(page)) {
5447 nr_pages <<= compound_order(page);
5448 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5451 if (do_swap_account && PageSwapCache(page))
5452 memcg = try_get_mem_cgroup_from_page(page);
5454 memcg = get_mem_cgroup_from_mm(mm);
5456 ret = try_charge(memcg, gfp_mask, nr_pages);
5458 css_put(&memcg->css);
5460 if (ret == -EINTR) {
5461 memcg = root_mem_cgroup;
5470 * mem_cgroup_commit_charge - commit a page charge
5471 * @page: page to charge
5472 * @memcg: memcg to charge the page to
5473 * @lrucare: page might be on LRU already
5475 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5476 * after page->mapping has been set up. This must happen atomically
5477 * as part of the page instantiation, i.e. under the page table lock
5478 * for anonymous pages, under the page lock for page and swap cache.
5480 * In addition, the page must not be on the LRU during the commit, to
5481 * prevent racing with task migration. If it might be, use @lrucare.
5483 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5485 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
5488 unsigned int nr_pages = 1;
5490 VM_BUG_ON_PAGE(!page->mapping, page);
5491 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
5493 if (mem_cgroup_disabled())
5496 * Swap faults will attempt to charge the same page multiple
5497 * times. But reuse_swap_page() might have removed the page
5498 * from swapcache already, so we can't check PageSwapCache().
5503 commit_charge(page, memcg, lrucare);
5505 if (PageTransHuge(page)) {
5506 nr_pages <<= compound_order(page);
5507 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5510 local_irq_disable();
5511 mem_cgroup_charge_statistics(memcg, page, nr_pages);
5512 memcg_check_events(memcg, page);
5515 if (do_swap_account && PageSwapCache(page)) {
5516 swp_entry_t entry = { .val = page_private(page) };
5518 * The swap entry might not get freed for a long time,
5519 * let's not wait for it. The page already received a
5520 * memory+swap charge, drop the swap entry duplicate.
5522 mem_cgroup_uncharge_swap(entry);
5527 * mem_cgroup_cancel_charge - cancel a page charge
5528 * @page: page to charge
5529 * @memcg: memcg to charge the page to
5531 * Cancel a charge transaction started by mem_cgroup_try_charge().
5533 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg)
5535 unsigned int nr_pages = 1;
5537 if (mem_cgroup_disabled())
5540 * Swap faults will attempt to charge the same page multiple
5541 * times. But reuse_swap_page() might have removed the page
5542 * from swapcache already, so we can't check PageSwapCache().
5547 if (PageTransHuge(page)) {
5548 nr_pages <<= compound_order(page);
5549 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5552 cancel_charge(memcg, nr_pages);
5555 static void uncharge_batch(struct mem_cgroup *memcg, unsigned long pgpgout,
5556 unsigned long nr_anon, unsigned long nr_file,
5557 unsigned long nr_huge, struct page *dummy_page)
5559 unsigned long nr_pages = nr_anon + nr_file;
5560 unsigned long flags;
5562 if (!mem_cgroup_is_root(memcg)) {
5563 page_counter_uncharge(&memcg->memory, nr_pages);
5564 if (do_swap_account)
5565 page_counter_uncharge(&memcg->memsw, nr_pages);
5566 memcg_oom_recover(memcg);
5569 local_irq_save(flags);
5570 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS], nr_anon);
5571 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_CACHE], nr_file);
5572 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE], nr_huge);
5573 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT], pgpgout);
5574 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
5575 memcg_check_events(memcg, dummy_page);
5576 local_irq_restore(flags);
5578 if (!mem_cgroup_is_root(memcg))
5579 css_put_many(&memcg->css, nr_pages);
5582 static void uncharge_list(struct list_head *page_list)
5584 struct mem_cgroup *memcg = NULL;
5585 unsigned long nr_anon = 0;
5586 unsigned long nr_file = 0;
5587 unsigned long nr_huge = 0;
5588 unsigned long pgpgout = 0;
5589 struct list_head *next;
5592 next = page_list->next;
5594 unsigned int nr_pages = 1;
5596 page = list_entry(next, struct page, lru);
5597 next = page->lru.next;
5599 VM_BUG_ON_PAGE(PageLRU(page), page);
5600 VM_BUG_ON_PAGE(page_count(page), page);
5602 if (!page->mem_cgroup)
5606 * Nobody should be changing or seriously looking at
5607 * page->mem_cgroup at this point, we have fully
5608 * exclusive access to the page.
5611 if (memcg != page->mem_cgroup) {
5613 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5615 pgpgout = nr_anon = nr_file = nr_huge = 0;
5617 memcg = page->mem_cgroup;
5620 if (PageTransHuge(page)) {
5621 nr_pages <<= compound_order(page);
5622 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5623 nr_huge += nr_pages;
5627 nr_anon += nr_pages;
5629 nr_file += nr_pages;
5631 page->mem_cgroup = NULL;
5634 } while (next != page_list);
5637 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5642 * mem_cgroup_uncharge - uncharge a page
5643 * @page: page to uncharge
5645 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5646 * mem_cgroup_commit_charge().
5648 void mem_cgroup_uncharge(struct page *page)
5650 if (mem_cgroup_disabled())
5653 /* Don't touch page->lru of any random page, pre-check: */
5654 if (!page->mem_cgroup)
5657 INIT_LIST_HEAD(&page->lru);
5658 uncharge_list(&page->lru);
5662 * mem_cgroup_uncharge_list - uncharge a list of page
5663 * @page_list: list of pages to uncharge
5665 * Uncharge a list of pages previously charged with
5666 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5668 void mem_cgroup_uncharge_list(struct list_head *page_list)
5670 if (mem_cgroup_disabled())
5673 if (!list_empty(page_list))
5674 uncharge_list(page_list);
5678 * mem_cgroup_migrate - migrate a charge to another page
5679 * @oldpage: currently charged page
5680 * @newpage: page to transfer the charge to
5681 * @lrucare: either or both pages might be on the LRU already
5683 * Migrate the charge from @oldpage to @newpage.
5685 * Both pages must be locked, @newpage->mapping must be set up.
5687 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage,
5690 struct mem_cgroup *memcg;
5693 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
5694 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
5695 VM_BUG_ON_PAGE(!lrucare && PageLRU(oldpage), oldpage);
5696 VM_BUG_ON_PAGE(!lrucare && PageLRU(newpage), newpage);
5697 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
5698 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
5701 if (mem_cgroup_disabled())
5704 /* Page cache replacement: new page already charged? */
5705 if (newpage->mem_cgroup)
5709 * Swapcache readahead pages can get migrated before being
5710 * charged, and migration from compaction can happen to an
5711 * uncharged page when the PFN walker finds a page that
5712 * reclaim just put back on the LRU but has not released yet.
5714 memcg = oldpage->mem_cgroup;
5719 lock_page_lru(oldpage, &isolated);
5721 oldpage->mem_cgroup = NULL;
5724 unlock_page_lru(oldpage, isolated);
5726 commit_charge(newpage, memcg, lrucare);
5730 * subsys_initcall() for memory controller.
5732 * Some parts like hotcpu_notifier() have to be initialized from this context
5733 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5734 * everything that doesn't depend on a specific mem_cgroup structure should
5735 * be initialized from here.
5737 static int __init mem_cgroup_init(void)
5741 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5743 for_each_possible_cpu(cpu)
5744 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
5747 for_each_node(node) {
5748 struct mem_cgroup_tree_per_node *rtpn;
5751 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
5752 node_online(node) ? node : NUMA_NO_NODE);
5754 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5755 struct mem_cgroup_tree_per_zone *rtpz;
5757 rtpz = &rtpn->rb_tree_per_zone[zone];
5758 rtpz->rb_root = RB_ROOT;
5759 spin_lock_init(&rtpz->lock);
5761 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5766 subsys_initcall(mem_cgroup_init);
5768 #ifdef CONFIG_MEMCG_SWAP
5770 * mem_cgroup_swapout - transfer a memsw charge to swap
5771 * @page: page whose memsw charge to transfer
5772 * @entry: swap entry to move the charge to
5774 * Transfer the memsw charge of @page to @entry.
5776 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
5778 struct mem_cgroup *memcg;
5779 unsigned short oldid;
5781 VM_BUG_ON_PAGE(PageLRU(page), page);
5782 VM_BUG_ON_PAGE(page_count(page), page);
5784 if (!do_swap_account)
5787 memcg = page->mem_cgroup;
5789 /* Readahead page, never charged */
5793 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg));
5794 VM_BUG_ON_PAGE(oldid, page);
5795 mem_cgroup_swap_statistics(memcg, true);
5797 page->mem_cgroup = NULL;
5799 if (!mem_cgroup_is_root(memcg))
5800 page_counter_uncharge(&memcg->memory, 1);
5802 /* XXX: caller holds IRQ-safe mapping->tree_lock */
5803 VM_BUG_ON(!irqs_disabled());
5805 mem_cgroup_charge_statistics(memcg, page, -1);
5806 memcg_check_events(memcg, page);
5810 * mem_cgroup_uncharge_swap - uncharge a swap entry
5811 * @entry: swap entry to uncharge
5813 * Drop the memsw charge associated with @entry.
5815 void mem_cgroup_uncharge_swap(swp_entry_t entry)
5817 struct mem_cgroup *memcg;
5820 if (!do_swap_account)
5823 id = swap_cgroup_record(entry, 0);
5825 memcg = mem_cgroup_lookup(id);
5827 if (!mem_cgroup_is_root(memcg))
5828 page_counter_uncharge(&memcg->memsw, 1);
5829 mem_cgroup_swap_statistics(memcg, false);
5830 css_put(&memcg->css);
5835 /* for remember boot option*/
5836 #ifdef CONFIG_MEMCG_SWAP_ENABLED
5837 static int really_do_swap_account __initdata = 1;
5839 static int really_do_swap_account __initdata;
5842 static int __init enable_swap_account(char *s)
5844 if (!strcmp(s, "1"))
5845 really_do_swap_account = 1;
5846 else if (!strcmp(s, "0"))
5847 really_do_swap_account = 0;
5850 __setup("swapaccount=", enable_swap_account);
5852 static struct cftype memsw_cgroup_files[] = {
5854 .name = "memsw.usage_in_bytes",
5855 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
5856 .read_u64 = mem_cgroup_read_u64,
5859 .name = "memsw.max_usage_in_bytes",
5860 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
5861 .write = mem_cgroup_reset,
5862 .read_u64 = mem_cgroup_read_u64,
5865 .name = "memsw.limit_in_bytes",
5866 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
5867 .write = mem_cgroup_write,
5868 .read_u64 = mem_cgroup_read_u64,
5871 .name = "memsw.failcnt",
5872 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
5873 .write = mem_cgroup_reset,
5874 .read_u64 = mem_cgroup_read_u64,
5876 { }, /* terminate */
5879 static int __init mem_cgroup_swap_init(void)
5881 if (!mem_cgroup_disabled() && really_do_swap_account) {
5882 do_swap_account = 1;
5883 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
5884 memsw_cgroup_files));
5888 subsys_initcall(mem_cgroup_swap_init);
5890 #endif /* CONFIG_MEMCG_SWAP */