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 #ifdef CONFIG_MEMCG_SWAP
76 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
77 int do_swap_account __read_mostly;
79 /* for remember boot option*/
80 #ifdef CONFIG_MEMCG_SWAP_ENABLED
81 static int really_do_swap_account __initdata = 1;
83 static int really_do_swap_account __initdata;
87 #define do_swap_account 0
91 static const char * const mem_cgroup_stat_names[] = {
100 static const char * const mem_cgroup_events_names[] = {
107 static const char * const mem_cgroup_lru_names[] = {
116 * Per memcg event counter is incremented at every pagein/pageout. With THP,
117 * it will be incremated by the number of pages. This counter is used for
118 * for trigger some periodic events. This is straightforward and better
119 * than using jiffies etc. to handle periodic memcg event.
121 enum mem_cgroup_events_target {
122 MEM_CGROUP_TARGET_THRESH,
123 MEM_CGROUP_TARGET_SOFTLIMIT,
124 MEM_CGROUP_TARGET_NUMAINFO,
127 #define THRESHOLDS_EVENTS_TARGET 128
128 #define SOFTLIMIT_EVENTS_TARGET 1024
129 #define NUMAINFO_EVENTS_TARGET 1024
131 struct mem_cgroup_stat_cpu {
132 long count[MEM_CGROUP_STAT_NSTATS];
133 unsigned long events[MEMCG_NR_EVENTS];
134 unsigned long nr_page_events;
135 unsigned long targets[MEM_CGROUP_NTARGETS];
138 struct reclaim_iter {
139 struct mem_cgroup *position;
140 /* scan generation, increased every round-trip */
141 unsigned int generation;
145 * per-zone information in memory controller.
147 struct mem_cgroup_per_zone {
148 struct lruvec lruvec;
149 unsigned long lru_size[NR_LRU_LISTS];
151 struct reclaim_iter iter[DEF_PRIORITY + 1];
153 struct rb_node tree_node; /* RB tree node */
154 unsigned long usage_in_excess;/* Set to the value by which */
155 /* the soft limit is exceeded*/
157 struct mem_cgroup *memcg; /* Back pointer, we cannot */
158 /* use container_of */
161 struct mem_cgroup_per_node {
162 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
166 * Cgroups above their limits are maintained in a RB-Tree, independent of
167 * their hierarchy representation
170 struct mem_cgroup_tree_per_zone {
171 struct rb_root rb_root;
175 struct mem_cgroup_tree_per_node {
176 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
179 struct mem_cgroup_tree {
180 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
183 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
185 struct mem_cgroup_threshold {
186 struct eventfd_ctx *eventfd;
187 unsigned long threshold;
191 struct mem_cgroup_threshold_ary {
192 /* An array index points to threshold just below or equal to usage. */
193 int current_threshold;
194 /* Size of entries[] */
196 /* Array of thresholds */
197 struct mem_cgroup_threshold entries[0];
200 struct mem_cgroup_thresholds {
201 /* Primary thresholds array */
202 struct mem_cgroup_threshold_ary *primary;
204 * Spare threshold array.
205 * This is needed to make mem_cgroup_unregister_event() "never fail".
206 * It must be able to store at least primary->size - 1 entries.
208 struct mem_cgroup_threshold_ary *spare;
212 struct mem_cgroup_eventfd_list {
213 struct list_head list;
214 struct eventfd_ctx *eventfd;
218 * cgroup_event represents events which userspace want to receive.
220 struct mem_cgroup_event {
222 * memcg which the event belongs to.
224 struct mem_cgroup *memcg;
226 * eventfd to signal userspace about the event.
228 struct eventfd_ctx *eventfd;
230 * Each of these stored in a list by the cgroup.
232 struct list_head list;
234 * register_event() callback will be used to add new userspace
235 * waiter for changes related to this event. Use eventfd_signal()
236 * on eventfd to send notification to userspace.
238 int (*register_event)(struct mem_cgroup *memcg,
239 struct eventfd_ctx *eventfd, const char *args);
241 * unregister_event() callback will be called when userspace closes
242 * the eventfd or on cgroup removing. This callback must be set,
243 * if you want provide notification functionality.
245 void (*unregister_event)(struct mem_cgroup *memcg,
246 struct eventfd_ctx *eventfd);
248 * All fields below needed to unregister event when
249 * userspace closes eventfd.
252 wait_queue_head_t *wqh;
254 struct work_struct remove;
257 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
258 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
261 * The memory controller data structure. The memory controller controls both
262 * page cache and RSS per cgroup. We would eventually like to provide
263 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
264 * to help the administrator determine what knobs to tune.
266 * TODO: Add a water mark for the memory controller. Reclaim will begin when
267 * we hit the water mark. May be even add a low water mark, such that
268 * no reclaim occurs from a cgroup at it's low water mark, this is
269 * a feature that will be implemented much later in the future.
272 struct cgroup_subsys_state css;
274 /* Accounted resources */
275 struct page_counter memory;
276 struct page_counter memsw;
277 struct page_counter kmem;
279 /* Normal memory consumption range */
283 unsigned long soft_limit;
285 /* vmpressure notifications */
286 struct vmpressure vmpressure;
288 /* css_online() has been completed */
292 * Should the accounting and control be hierarchical, per subtree?
298 atomic_t oom_wakeups;
301 /* OOM-Killer disable */
302 int oom_kill_disable;
304 /* protect arrays of thresholds */
305 struct mutex thresholds_lock;
307 /* thresholds for memory usage. RCU-protected */
308 struct mem_cgroup_thresholds thresholds;
310 /* thresholds for mem+swap usage. RCU-protected */
311 struct mem_cgroup_thresholds memsw_thresholds;
313 /* For oom notifier event fd */
314 struct list_head oom_notify;
317 * Should we move charges of a task when a task is moved into this
318 * mem_cgroup ? And what type of charges should we move ?
320 unsigned long move_charge_at_immigrate;
322 * set > 0 if pages under this cgroup are moving to other cgroup.
324 atomic_t moving_account;
325 /* taken only while moving_account > 0 */
326 spinlock_t move_lock;
327 struct task_struct *move_lock_task;
328 unsigned long move_lock_flags;
332 struct mem_cgroup_stat_cpu __percpu *stat;
334 * used when a cpu is offlined or other synchronizations
335 * See mem_cgroup_read_stat().
337 struct mem_cgroup_stat_cpu nocpu_base;
338 spinlock_t pcp_counter_lock;
340 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
341 struct cg_proto tcp_mem;
343 #if defined(CONFIG_MEMCG_KMEM)
344 /* Index in the kmem_cache->memcg_params->memcg_caches array */
348 int last_scanned_node;
350 nodemask_t scan_nodes;
351 atomic_t numainfo_events;
352 atomic_t numainfo_updating;
355 /* List of events which userspace want to receive */
356 struct list_head event_list;
357 spinlock_t event_list_lock;
359 struct mem_cgroup_per_node *nodeinfo[0];
360 /* WARNING: nodeinfo must be the last member here */
363 #ifdef CONFIG_MEMCG_KMEM
364 static bool memcg_kmem_is_active(struct mem_cgroup *memcg)
366 return memcg->kmemcg_id >= 0;
370 /* Stuffs for move charges at task migration. */
372 * Types of charges to be moved.
374 #define MOVE_ANON 0x1U
375 #define MOVE_FILE 0x2U
376 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
378 /* "mc" and its members are protected by cgroup_mutex */
379 static struct move_charge_struct {
380 spinlock_t lock; /* for from, to */
381 struct mem_cgroup *from;
382 struct mem_cgroup *to;
384 unsigned long precharge;
385 unsigned long moved_charge;
386 unsigned long moved_swap;
387 struct task_struct *moving_task; /* a task moving charges */
388 wait_queue_head_t waitq; /* a waitq for other context */
390 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
391 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
395 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
396 * limit reclaim to prevent infinite loops, if they ever occur.
398 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
399 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
402 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
403 MEM_CGROUP_CHARGE_TYPE_ANON,
404 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
405 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
409 /* for encoding cft->private value on file */
417 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
418 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
419 #define MEMFILE_ATTR(val) ((val) & 0xffff)
420 /* Used for OOM nofiier */
421 #define OOM_CONTROL (0)
424 * The memcg_create_mutex will be held whenever a new cgroup is created.
425 * As a consequence, any change that needs to protect against new child cgroups
426 * appearing has to hold it as well.
428 static DEFINE_MUTEX(memcg_create_mutex);
430 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
432 return s ? container_of(s, struct mem_cgroup, css) : NULL;
435 /* Some nice accessors for the vmpressure. */
436 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
439 memcg = root_mem_cgroup;
440 return &memcg->vmpressure;
443 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
445 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
448 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
450 return (memcg == root_mem_cgroup);
454 * We restrict the id in the range of [1, 65535], so it can fit into
457 #define MEM_CGROUP_ID_MAX USHRT_MAX
459 static inline unsigned short mem_cgroup_id(struct mem_cgroup *memcg)
461 return memcg->css.id;
464 static inline struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
466 struct cgroup_subsys_state *css;
468 css = css_from_id(id, &memory_cgrp_subsys);
469 return mem_cgroup_from_css(css);
472 /* Writing them here to avoid exposing memcg's inner layout */
473 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
475 void sock_update_memcg(struct sock *sk)
477 if (mem_cgroup_sockets_enabled) {
478 struct mem_cgroup *memcg;
479 struct cg_proto *cg_proto;
481 BUG_ON(!sk->sk_prot->proto_cgroup);
483 /* Socket cloning can throw us here with sk_cgrp already
484 * filled. It won't however, necessarily happen from
485 * process context. So the test for root memcg given
486 * the current task's memcg won't help us in this case.
488 * Respecting the original socket's memcg is a better
489 * decision in this case.
492 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
493 css_get(&sk->sk_cgrp->memcg->css);
498 memcg = mem_cgroup_from_task(current);
499 cg_proto = sk->sk_prot->proto_cgroup(memcg);
500 if (!mem_cgroup_is_root(memcg) &&
501 memcg_proto_active(cg_proto) &&
502 css_tryget_online(&memcg->css)) {
503 sk->sk_cgrp = cg_proto;
508 EXPORT_SYMBOL(sock_update_memcg);
510 void sock_release_memcg(struct sock *sk)
512 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
513 struct mem_cgroup *memcg;
514 WARN_ON(!sk->sk_cgrp->memcg);
515 memcg = sk->sk_cgrp->memcg;
516 css_put(&sk->sk_cgrp->memcg->css);
520 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
522 if (!memcg || mem_cgroup_is_root(memcg))
525 return &memcg->tcp_mem;
527 EXPORT_SYMBOL(tcp_proto_cgroup);
529 static void disarm_sock_keys(struct mem_cgroup *memcg)
531 if (!memcg_proto_activated(&memcg->tcp_mem))
533 static_key_slow_dec(&memcg_socket_limit_enabled);
536 static void disarm_sock_keys(struct mem_cgroup *memcg)
541 #ifdef CONFIG_MEMCG_KMEM
543 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
544 * The main reason for not using cgroup id for this:
545 * this works better in sparse environments, where we have a lot of memcgs,
546 * but only a few kmem-limited. Or also, if we have, for instance, 200
547 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
548 * 200 entry array for that.
550 * The current size of the caches array is stored in
551 * memcg_limited_groups_array_size. It will double each time we have to
554 static DEFINE_IDA(kmem_limited_groups);
555 int memcg_limited_groups_array_size;
558 * MIN_SIZE is different than 1, because we would like to avoid going through
559 * the alloc/free process all the time. In a small machine, 4 kmem-limited
560 * cgroups is a reasonable guess. In the future, it could be a parameter or
561 * tunable, but that is strictly not necessary.
563 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
564 * this constant directly from cgroup, but it is understandable that this is
565 * better kept as an internal representation in cgroup.c. In any case, the
566 * cgrp_id space is not getting any smaller, and we don't have to necessarily
567 * increase ours as well if it increases.
569 #define MEMCG_CACHES_MIN_SIZE 4
570 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
573 * A lot of the calls to the cache allocation functions are expected to be
574 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
575 * conditional to this static branch, we'll have to allow modules that does
576 * kmem_cache_alloc and the such to see this symbol as well
578 struct static_key memcg_kmem_enabled_key;
579 EXPORT_SYMBOL(memcg_kmem_enabled_key);
581 static void memcg_free_cache_id(int id);
583 static void disarm_kmem_keys(struct mem_cgroup *memcg)
585 if (memcg_kmem_is_active(memcg)) {
586 static_key_slow_dec(&memcg_kmem_enabled_key);
587 memcg_free_cache_id(memcg->kmemcg_id);
590 * This check can't live in kmem destruction function,
591 * since the charges will outlive the cgroup
593 WARN_ON(page_counter_read(&memcg->kmem));
596 static void disarm_kmem_keys(struct mem_cgroup *memcg)
599 #endif /* CONFIG_MEMCG_KMEM */
601 static void disarm_static_keys(struct mem_cgroup *memcg)
603 disarm_sock_keys(memcg);
604 disarm_kmem_keys(memcg);
607 static struct mem_cgroup_per_zone *
608 mem_cgroup_zone_zoneinfo(struct mem_cgroup *memcg, struct zone *zone)
610 int nid = zone_to_nid(zone);
611 int zid = zone_idx(zone);
613 return &memcg->nodeinfo[nid]->zoneinfo[zid];
616 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
621 static struct mem_cgroup_per_zone *
622 mem_cgroup_page_zoneinfo(struct mem_cgroup *memcg, struct page *page)
624 int nid = page_to_nid(page);
625 int zid = page_zonenum(page);
627 return &memcg->nodeinfo[nid]->zoneinfo[zid];
630 static struct mem_cgroup_tree_per_zone *
631 soft_limit_tree_node_zone(int nid, int zid)
633 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
636 static struct mem_cgroup_tree_per_zone *
637 soft_limit_tree_from_page(struct page *page)
639 int nid = page_to_nid(page);
640 int zid = page_zonenum(page);
642 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
645 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone *mz,
646 struct mem_cgroup_tree_per_zone *mctz,
647 unsigned long new_usage_in_excess)
649 struct rb_node **p = &mctz->rb_root.rb_node;
650 struct rb_node *parent = NULL;
651 struct mem_cgroup_per_zone *mz_node;
656 mz->usage_in_excess = new_usage_in_excess;
657 if (!mz->usage_in_excess)
661 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
663 if (mz->usage_in_excess < mz_node->usage_in_excess)
666 * We can't avoid mem cgroups that are over their soft
667 * limit by the same amount
669 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
672 rb_link_node(&mz->tree_node, parent, p);
673 rb_insert_color(&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 rb_erase(&mz->tree_node, &mctz->rb_root);
686 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
687 struct mem_cgroup_tree_per_zone *mctz)
691 spin_lock_irqsave(&mctz->lock, flags);
692 __mem_cgroup_remove_exceeded(mz, mctz);
693 spin_unlock_irqrestore(&mctz->lock, flags);
696 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
698 unsigned long nr_pages = page_counter_read(&memcg->memory);
699 unsigned long soft_limit = ACCESS_ONCE(memcg->soft_limit);
700 unsigned long excess = 0;
702 if (nr_pages > soft_limit)
703 excess = nr_pages - soft_limit;
708 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
710 unsigned long excess;
711 struct mem_cgroup_per_zone *mz;
712 struct mem_cgroup_tree_per_zone *mctz;
714 mctz = soft_limit_tree_from_page(page);
716 * Necessary to update all ancestors when hierarchy is used.
717 * because their event counter is not touched.
719 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
720 mz = mem_cgroup_page_zoneinfo(memcg, page);
721 excess = soft_limit_excess(memcg);
723 * We have to update the tree if mz is on RB-tree or
724 * mem is over its softlimit.
726 if (excess || mz->on_tree) {
729 spin_lock_irqsave(&mctz->lock, flags);
730 /* if on-tree, remove it */
732 __mem_cgroup_remove_exceeded(mz, mctz);
734 * Insert again. mz->usage_in_excess will be updated.
735 * If excess is 0, no tree ops.
737 __mem_cgroup_insert_exceeded(mz, mctz, excess);
738 spin_unlock_irqrestore(&mctz->lock, flags);
743 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
745 struct mem_cgroup_tree_per_zone *mctz;
746 struct mem_cgroup_per_zone *mz;
750 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
751 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
752 mctz = soft_limit_tree_node_zone(nid, zid);
753 mem_cgroup_remove_exceeded(mz, mctz);
758 static struct mem_cgroup_per_zone *
759 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
761 struct rb_node *rightmost = NULL;
762 struct mem_cgroup_per_zone *mz;
766 rightmost = rb_last(&mctz->rb_root);
768 goto done; /* Nothing to reclaim from */
770 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
772 * Remove the node now but someone else can add it back,
773 * we will to add it back at the end of reclaim to its correct
774 * position in the tree.
776 __mem_cgroup_remove_exceeded(mz, mctz);
777 if (!soft_limit_excess(mz->memcg) ||
778 !css_tryget_online(&mz->memcg->css))
784 static struct mem_cgroup_per_zone *
785 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
787 struct mem_cgroup_per_zone *mz;
789 spin_lock_irq(&mctz->lock);
790 mz = __mem_cgroup_largest_soft_limit_node(mctz);
791 spin_unlock_irq(&mctz->lock);
796 * Implementation Note: reading percpu statistics for memcg.
798 * Both of vmstat[] and percpu_counter has threshold and do periodic
799 * synchronization to implement "quick" read. There are trade-off between
800 * reading cost and precision of value. Then, we may have a chance to implement
801 * a periodic synchronizion of counter in memcg's counter.
803 * But this _read() function is used for user interface now. The user accounts
804 * memory usage by memory cgroup and he _always_ requires exact value because
805 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
806 * have to visit all online cpus and make sum. So, for now, unnecessary
807 * synchronization is not implemented. (just implemented for cpu hotplug)
809 * If there are kernel internal actions which can make use of some not-exact
810 * value, and reading all cpu value can be performance bottleneck in some
811 * common workload, threashold and synchonization as vmstat[] should be
814 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
815 enum mem_cgroup_stat_index idx)
821 for_each_online_cpu(cpu)
822 val += per_cpu(memcg->stat->count[idx], cpu);
823 #ifdef CONFIG_HOTPLUG_CPU
824 spin_lock(&memcg->pcp_counter_lock);
825 val += memcg->nocpu_base.count[idx];
826 spin_unlock(&memcg->pcp_counter_lock);
832 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
833 enum mem_cgroup_events_index idx)
835 unsigned long val = 0;
839 for_each_online_cpu(cpu)
840 val += per_cpu(memcg->stat->events[idx], cpu);
841 #ifdef CONFIG_HOTPLUG_CPU
842 spin_lock(&memcg->pcp_counter_lock);
843 val += memcg->nocpu_base.events[idx];
844 spin_unlock(&memcg->pcp_counter_lock);
850 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
855 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
856 * counted as CACHE even if it's on ANON LRU.
859 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
862 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
865 if (PageTransHuge(page))
866 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
869 /* pagein of a big page is an event. So, ignore page size */
871 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
873 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
874 nr_pages = -nr_pages; /* for event */
877 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
880 unsigned long mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
882 struct mem_cgroup_per_zone *mz;
884 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
885 return mz->lru_size[lru];
888 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
890 unsigned int lru_mask)
892 unsigned long nr = 0;
895 VM_BUG_ON((unsigned)nid >= nr_node_ids);
897 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
898 struct mem_cgroup_per_zone *mz;
902 if (!(BIT(lru) & lru_mask))
904 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
905 nr += mz->lru_size[lru];
911 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
912 unsigned int lru_mask)
914 unsigned long nr = 0;
917 for_each_node_state(nid, N_MEMORY)
918 nr += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
922 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
923 enum mem_cgroup_events_target target)
925 unsigned long val, next;
927 val = __this_cpu_read(memcg->stat->nr_page_events);
928 next = __this_cpu_read(memcg->stat->targets[target]);
929 /* from time_after() in jiffies.h */
930 if ((long)next - (long)val < 0) {
932 case MEM_CGROUP_TARGET_THRESH:
933 next = val + THRESHOLDS_EVENTS_TARGET;
935 case MEM_CGROUP_TARGET_SOFTLIMIT:
936 next = val + SOFTLIMIT_EVENTS_TARGET;
938 case MEM_CGROUP_TARGET_NUMAINFO:
939 next = val + NUMAINFO_EVENTS_TARGET;
944 __this_cpu_write(memcg->stat->targets[target], next);
951 * Check events in order.
954 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
956 /* threshold event is triggered in finer grain than soft limit */
957 if (unlikely(mem_cgroup_event_ratelimit(memcg,
958 MEM_CGROUP_TARGET_THRESH))) {
960 bool do_numainfo __maybe_unused;
962 do_softlimit = mem_cgroup_event_ratelimit(memcg,
963 MEM_CGROUP_TARGET_SOFTLIMIT);
965 do_numainfo = mem_cgroup_event_ratelimit(memcg,
966 MEM_CGROUP_TARGET_NUMAINFO);
968 mem_cgroup_threshold(memcg);
969 if (unlikely(do_softlimit))
970 mem_cgroup_update_tree(memcg, page);
972 if (unlikely(do_numainfo))
973 atomic_inc(&memcg->numainfo_events);
978 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
981 * mm_update_next_owner() may clear mm->owner to NULL
982 * if it races with swapoff, page migration, etc.
983 * So this can be called with p == NULL.
988 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
991 static struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
993 struct mem_cgroup *memcg = NULL;
998 * Page cache insertions can happen withou an
999 * actual mm context, e.g. during disk probing
1000 * on boot, loopback IO, acct() writes etc.
1003 memcg = root_mem_cgroup;
1005 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1006 if (unlikely(!memcg))
1007 memcg = root_mem_cgroup;
1009 } while (!css_tryget_online(&memcg->css));
1015 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1016 * @root: hierarchy root
1017 * @prev: previously returned memcg, NULL on first invocation
1018 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1020 * Returns references to children of the hierarchy below @root, or
1021 * @root itself, or %NULL after a full round-trip.
1023 * Caller must pass the return value in @prev on subsequent
1024 * invocations for reference counting, or use mem_cgroup_iter_break()
1025 * to cancel a hierarchy walk before the round-trip is complete.
1027 * Reclaimers can specify a zone and a priority level in @reclaim to
1028 * divide up the memcgs in the hierarchy among all concurrent
1029 * reclaimers operating on the same zone and priority.
1031 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1032 struct mem_cgroup *prev,
1033 struct mem_cgroup_reclaim_cookie *reclaim)
1035 struct reclaim_iter *uninitialized_var(iter);
1036 struct cgroup_subsys_state *css = NULL;
1037 struct mem_cgroup *memcg = NULL;
1038 struct mem_cgroup *pos = NULL;
1040 if (mem_cgroup_disabled())
1044 root = root_mem_cgroup;
1046 if (prev && !reclaim)
1049 if (!root->use_hierarchy && root != root_mem_cgroup) {
1058 struct mem_cgroup_per_zone *mz;
1060 mz = mem_cgroup_zone_zoneinfo(root, reclaim->zone);
1061 iter = &mz->iter[reclaim->priority];
1063 if (prev && reclaim->generation != iter->generation)
1067 pos = ACCESS_ONCE(iter->position);
1069 * A racing update may change the position and
1070 * put the last reference, hence css_tryget(),
1071 * or retry to see the updated position.
1073 } while (pos && !css_tryget(&pos->css));
1080 css = css_next_descendant_pre(css, &root->css);
1083 * Reclaimers share the hierarchy walk, and a
1084 * new one might jump in right at the end of
1085 * the hierarchy - make sure they see at least
1086 * one group and restart from the beginning.
1094 * Verify the css and acquire a reference. The root
1095 * is provided by the caller, so we know it's alive
1096 * and kicking, and don't take an extra reference.
1098 memcg = mem_cgroup_from_css(css);
1100 if (css == &root->css)
1103 if (css_tryget(css)) {
1105 * Make sure the memcg is initialized:
1106 * mem_cgroup_css_online() orders the the
1107 * initialization against setting the flag.
1109 if (smp_load_acquire(&memcg->initialized))
1119 if (cmpxchg(&iter->position, pos, memcg) == pos) {
1121 css_get(&memcg->css);
1127 * pairs with css_tryget when dereferencing iter->position
1136 reclaim->generation = iter->generation;
1142 if (prev && prev != root)
1143 css_put(&prev->css);
1149 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1150 * @root: hierarchy root
1151 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1153 void mem_cgroup_iter_break(struct mem_cgroup *root,
1154 struct mem_cgroup *prev)
1157 root = root_mem_cgroup;
1158 if (prev && prev != root)
1159 css_put(&prev->css);
1163 * Iteration constructs for visiting all cgroups (under a tree). If
1164 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1165 * be used for reference counting.
1167 #define for_each_mem_cgroup_tree(iter, root) \
1168 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1170 iter = mem_cgroup_iter(root, iter, NULL))
1172 #define for_each_mem_cgroup(iter) \
1173 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1175 iter = mem_cgroup_iter(NULL, iter, NULL))
1177 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1179 struct mem_cgroup *memcg;
1182 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1183 if (unlikely(!memcg))
1188 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1191 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1199 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1202 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1203 * @zone: zone of the wanted lruvec
1204 * @memcg: memcg of the wanted lruvec
1206 * Returns the lru list vector holding pages for the given @zone and
1207 * @mem. This can be the global zone lruvec, if the memory controller
1210 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1211 struct mem_cgroup *memcg)
1213 struct mem_cgroup_per_zone *mz;
1214 struct lruvec *lruvec;
1216 if (mem_cgroup_disabled()) {
1217 lruvec = &zone->lruvec;
1221 mz = mem_cgroup_zone_zoneinfo(memcg, zone);
1222 lruvec = &mz->lruvec;
1225 * Since a node can be onlined after the mem_cgroup was created,
1226 * we have to be prepared to initialize lruvec->zone here;
1227 * and if offlined then reonlined, we need to reinitialize it.
1229 if (unlikely(lruvec->zone != zone))
1230 lruvec->zone = zone;
1235 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1237 * @zone: zone of the page
1239 * This function is only safe when following the LRU page isolation
1240 * and putback protocol: the LRU lock must be held, and the page must
1241 * either be PageLRU() or the caller must have isolated/allocated it.
1243 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1245 struct mem_cgroup_per_zone *mz;
1246 struct mem_cgroup *memcg;
1247 struct lruvec *lruvec;
1249 if (mem_cgroup_disabled()) {
1250 lruvec = &zone->lruvec;
1254 memcg = page->mem_cgroup;
1256 * Swapcache readahead pages are added to the LRU - and
1257 * possibly migrated - before they are charged.
1260 memcg = root_mem_cgroup;
1262 mz = mem_cgroup_page_zoneinfo(memcg, page);
1263 lruvec = &mz->lruvec;
1266 * Since a node can be onlined after the mem_cgroup was created,
1267 * we have to be prepared to initialize lruvec->zone here;
1268 * and if offlined then reonlined, we need to reinitialize it.
1270 if (unlikely(lruvec->zone != zone))
1271 lruvec->zone = zone;
1276 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1277 * @lruvec: mem_cgroup per zone lru vector
1278 * @lru: index of lru list the page is sitting on
1279 * @nr_pages: positive when adding or negative when removing
1281 * This function must be called when a page is added to or removed from an
1284 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1287 struct mem_cgroup_per_zone *mz;
1288 unsigned long *lru_size;
1290 if (mem_cgroup_disabled())
1293 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1294 lru_size = mz->lru_size + lru;
1295 *lru_size += nr_pages;
1296 VM_BUG_ON((long)(*lru_size) < 0);
1299 bool mem_cgroup_is_descendant(struct mem_cgroup *memcg, struct mem_cgroup *root)
1303 if (!root->use_hierarchy)
1305 return cgroup_is_descendant(memcg->css.cgroup, root->css.cgroup);
1308 bool task_in_mem_cgroup(struct task_struct *task, struct mem_cgroup *memcg)
1310 struct mem_cgroup *task_memcg;
1311 struct task_struct *p;
1314 p = find_lock_task_mm(task);
1316 task_memcg = get_mem_cgroup_from_mm(p->mm);
1320 * All threads may have already detached their mm's, but the oom
1321 * killer still needs to detect if they have already been oom
1322 * killed to prevent needlessly killing additional tasks.
1325 task_memcg = mem_cgroup_from_task(task);
1326 css_get(&task_memcg->css);
1329 ret = mem_cgroup_is_descendant(task_memcg, memcg);
1330 css_put(&task_memcg->css);
1334 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1336 unsigned long inactive_ratio;
1337 unsigned long inactive;
1338 unsigned long active;
1341 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1342 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1344 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1346 inactive_ratio = int_sqrt(10 * gb);
1350 return inactive * inactive_ratio < active;
1353 bool mem_cgroup_lruvec_online(struct lruvec *lruvec)
1355 struct mem_cgroup_per_zone *mz;
1356 struct mem_cgroup *memcg;
1358 if (mem_cgroup_disabled())
1361 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1364 return !!(memcg->css.flags & CSS_ONLINE);
1367 #define mem_cgroup_from_counter(counter, member) \
1368 container_of(counter, struct mem_cgroup, member)
1371 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1372 * @memcg: the memory cgroup
1374 * Returns the maximum amount of memory @mem can be charged with, in
1377 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1379 unsigned long margin = 0;
1380 unsigned long count;
1381 unsigned long limit;
1383 count = page_counter_read(&memcg->memory);
1384 limit = ACCESS_ONCE(memcg->memory.limit);
1386 margin = limit - count;
1388 if (do_swap_account) {
1389 count = page_counter_read(&memcg->memsw);
1390 limit = ACCESS_ONCE(memcg->memsw.limit);
1392 margin = min(margin, limit - count);
1398 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1401 if (mem_cgroup_disabled() || !memcg->css.parent)
1402 return vm_swappiness;
1404 return memcg->swappiness;
1408 * A routine for checking "mem" is under move_account() or not.
1410 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1411 * moving cgroups. This is for waiting at high-memory pressure
1414 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1416 struct mem_cgroup *from;
1417 struct mem_cgroup *to;
1420 * Unlike task_move routines, we access mc.to, mc.from not under
1421 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1423 spin_lock(&mc.lock);
1429 ret = mem_cgroup_is_descendant(from, memcg) ||
1430 mem_cgroup_is_descendant(to, memcg);
1432 spin_unlock(&mc.lock);
1436 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1438 if (mc.moving_task && current != mc.moving_task) {
1439 if (mem_cgroup_under_move(memcg)) {
1441 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1442 /* moving charge context might have finished. */
1445 finish_wait(&mc.waitq, &wait);
1452 #define K(x) ((x) << (PAGE_SHIFT-10))
1454 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1455 * @memcg: The memory cgroup that went over limit
1456 * @p: Task that is going to be killed
1458 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1461 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1463 /* oom_info_lock ensures that parallel ooms do not interleave */
1464 static DEFINE_MUTEX(oom_info_lock);
1465 struct mem_cgroup *iter;
1471 mutex_lock(&oom_info_lock);
1474 pr_info("Task in ");
1475 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1476 pr_cont(" killed as a result of limit of ");
1477 pr_cont_cgroup_path(memcg->css.cgroup);
1482 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1483 K((u64)page_counter_read(&memcg->memory)),
1484 K((u64)memcg->memory.limit), memcg->memory.failcnt);
1485 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1486 K((u64)page_counter_read(&memcg->memsw)),
1487 K((u64)memcg->memsw.limit), memcg->memsw.failcnt);
1488 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1489 K((u64)page_counter_read(&memcg->kmem)),
1490 K((u64)memcg->kmem.limit), memcg->kmem.failcnt);
1492 for_each_mem_cgroup_tree(iter, memcg) {
1493 pr_info("Memory cgroup stats for ");
1494 pr_cont_cgroup_path(iter->css.cgroup);
1497 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1498 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1500 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1501 K(mem_cgroup_read_stat(iter, i)));
1504 for (i = 0; i < NR_LRU_LISTS; i++)
1505 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1506 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1510 mutex_unlock(&oom_info_lock);
1514 * This function returns the number of memcg under hierarchy tree. Returns
1515 * 1(self count) if no children.
1517 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1520 struct mem_cgroup *iter;
1522 for_each_mem_cgroup_tree(iter, memcg)
1528 * Return the memory (and swap, if configured) limit for a memcg.
1530 static unsigned long mem_cgroup_get_limit(struct mem_cgroup *memcg)
1532 unsigned long limit;
1534 limit = memcg->memory.limit;
1535 if (mem_cgroup_swappiness(memcg)) {
1536 unsigned long memsw_limit;
1538 memsw_limit = memcg->memsw.limit;
1539 limit = min(limit + total_swap_pages, memsw_limit);
1544 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1547 struct mem_cgroup *iter;
1548 unsigned long chosen_points = 0;
1549 unsigned long totalpages;
1550 unsigned int points = 0;
1551 struct task_struct *chosen = NULL;
1554 * If current has a pending SIGKILL or is exiting, then automatically
1555 * select it. The goal is to allow it to allocate so that it may
1556 * quickly exit and free its memory.
1558 if (fatal_signal_pending(current) || task_will_free_mem(current)) {
1559 mark_tsk_oom_victim(current);
1563 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1564 totalpages = mem_cgroup_get_limit(memcg) ? : 1;
1565 for_each_mem_cgroup_tree(iter, memcg) {
1566 struct css_task_iter it;
1567 struct task_struct *task;
1569 css_task_iter_start(&iter->css, &it);
1570 while ((task = css_task_iter_next(&it))) {
1571 switch (oom_scan_process_thread(task, totalpages, NULL,
1573 case OOM_SCAN_SELECT:
1575 put_task_struct(chosen);
1577 chosen_points = ULONG_MAX;
1578 get_task_struct(chosen);
1580 case OOM_SCAN_CONTINUE:
1582 case OOM_SCAN_ABORT:
1583 css_task_iter_end(&it);
1584 mem_cgroup_iter_break(memcg, iter);
1586 put_task_struct(chosen);
1591 points = oom_badness(task, memcg, NULL, totalpages);
1592 if (!points || points < chosen_points)
1594 /* Prefer thread group leaders for display purposes */
1595 if (points == chosen_points &&
1596 thread_group_leader(chosen))
1600 put_task_struct(chosen);
1602 chosen_points = points;
1603 get_task_struct(chosen);
1605 css_task_iter_end(&it);
1610 points = chosen_points * 1000 / totalpages;
1611 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1612 NULL, "Memory cgroup out of memory");
1615 #if MAX_NUMNODES > 1
1618 * test_mem_cgroup_node_reclaimable
1619 * @memcg: the target memcg
1620 * @nid: the node ID to be checked.
1621 * @noswap : specify true here if the user wants flle only information.
1623 * This function returns whether the specified memcg contains any
1624 * reclaimable pages on a node. Returns true if there are any reclaimable
1625 * pages in the node.
1627 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1628 int nid, bool noswap)
1630 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1632 if (noswap || !total_swap_pages)
1634 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1641 * Always updating the nodemask is not very good - even if we have an empty
1642 * list or the wrong list here, we can start from some node and traverse all
1643 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1646 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1650 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1651 * pagein/pageout changes since the last update.
1653 if (!atomic_read(&memcg->numainfo_events))
1655 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1658 /* make a nodemask where this memcg uses memory from */
1659 memcg->scan_nodes = node_states[N_MEMORY];
1661 for_each_node_mask(nid, node_states[N_MEMORY]) {
1663 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1664 node_clear(nid, memcg->scan_nodes);
1667 atomic_set(&memcg->numainfo_events, 0);
1668 atomic_set(&memcg->numainfo_updating, 0);
1672 * Selecting a node where we start reclaim from. Because what we need is just
1673 * reducing usage counter, start from anywhere is O,K. Considering
1674 * memory reclaim from current node, there are pros. and cons.
1676 * Freeing memory from current node means freeing memory from a node which
1677 * we'll use or we've used. So, it may make LRU bad. And if several threads
1678 * hit limits, it will see a contention on a node. But freeing from remote
1679 * node means more costs for memory reclaim because of memory latency.
1681 * Now, we use round-robin. Better algorithm is welcomed.
1683 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1687 mem_cgroup_may_update_nodemask(memcg);
1688 node = memcg->last_scanned_node;
1690 node = next_node(node, memcg->scan_nodes);
1691 if (node == MAX_NUMNODES)
1692 node = first_node(memcg->scan_nodes);
1694 * We call this when we hit limit, not when pages are added to LRU.
1695 * No LRU may hold pages because all pages are UNEVICTABLE or
1696 * memcg is too small and all pages are not on LRU. In that case,
1697 * we use curret node.
1699 if (unlikely(node == MAX_NUMNODES))
1700 node = numa_node_id();
1702 memcg->last_scanned_node = node;
1706 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1712 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1715 unsigned long *total_scanned)
1717 struct mem_cgroup *victim = NULL;
1720 unsigned long excess;
1721 unsigned long nr_scanned;
1722 struct mem_cgroup_reclaim_cookie reclaim = {
1727 excess = soft_limit_excess(root_memcg);
1730 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1735 * If we have not been able to reclaim
1736 * anything, it might because there are
1737 * no reclaimable pages under this hierarchy
1742 * We want to do more targeted reclaim.
1743 * excess >> 2 is not to excessive so as to
1744 * reclaim too much, nor too less that we keep
1745 * coming back to reclaim from this cgroup
1747 if (total >= (excess >> 2) ||
1748 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1753 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1755 *total_scanned += nr_scanned;
1756 if (!soft_limit_excess(root_memcg))
1759 mem_cgroup_iter_break(root_memcg, victim);
1763 #ifdef CONFIG_LOCKDEP
1764 static struct lockdep_map memcg_oom_lock_dep_map = {
1765 .name = "memcg_oom_lock",
1769 static DEFINE_SPINLOCK(memcg_oom_lock);
1772 * Check OOM-Killer is already running under our hierarchy.
1773 * If someone is running, return false.
1775 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1777 struct mem_cgroup *iter, *failed = NULL;
1779 spin_lock(&memcg_oom_lock);
1781 for_each_mem_cgroup_tree(iter, memcg) {
1782 if (iter->oom_lock) {
1784 * this subtree of our hierarchy is already locked
1785 * so we cannot give a lock.
1788 mem_cgroup_iter_break(memcg, iter);
1791 iter->oom_lock = true;
1796 * OK, we failed to lock the whole subtree so we have
1797 * to clean up what we set up to the failing subtree
1799 for_each_mem_cgroup_tree(iter, memcg) {
1800 if (iter == failed) {
1801 mem_cgroup_iter_break(memcg, iter);
1804 iter->oom_lock = false;
1807 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1809 spin_unlock(&memcg_oom_lock);
1814 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1816 struct mem_cgroup *iter;
1818 spin_lock(&memcg_oom_lock);
1819 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1820 for_each_mem_cgroup_tree(iter, memcg)
1821 iter->oom_lock = false;
1822 spin_unlock(&memcg_oom_lock);
1825 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1827 struct mem_cgroup *iter;
1829 for_each_mem_cgroup_tree(iter, memcg)
1830 atomic_inc(&iter->under_oom);
1833 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1835 struct mem_cgroup *iter;
1838 * When a new child is created while the hierarchy is under oom,
1839 * mem_cgroup_oom_lock() may not be called. We have to use
1840 * atomic_add_unless() here.
1842 for_each_mem_cgroup_tree(iter, memcg)
1843 atomic_add_unless(&iter->under_oom, -1, 0);
1846 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1848 struct oom_wait_info {
1849 struct mem_cgroup *memcg;
1853 static int memcg_oom_wake_function(wait_queue_t *wait,
1854 unsigned mode, int sync, void *arg)
1856 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1857 struct mem_cgroup *oom_wait_memcg;
1858 struct oom_wait_info *oom_wait_info;
1860 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1861 oom_wait_memcg = oom_wait_info->memcg;
1863 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1864 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1866 return autoremove_wake_function(wait, mode, sync, arg);
1869 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
1871 atomic_inc(&memcg->oom_wakeups);
1872 /* for filtering, pass "memcg" as argument. */
1873 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1876 static void memcg_oom_recover(struct mem_cgroup *memcg)
1878 if (memcg && atomic_read(&memcg->under_oom))
1879 memcg_wakeup_oom(memcg);
1882 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1884 if (!current->memcg_oom.may_oom)
1887 * We are in the middle of the charge context here, so we
1888 * don't want to block when potentially sitting on a callstack
1889 * that holds all kinds of filesystem and mm locks.
1891 * Also, the caller may handle a failed allocation gracefully
1892 * (like optional page cache readahead) and so an OOM killer
1893 * invocation might not even be necessary.
1895 * That's why we don't do anything here except remember the
1896 * OOM context and then deal with it at the end of the page
1897 * fault when the stack is unwound, the locks are released,
1898 * and when we know whether the fault was overall successful.
1900 css_get(&memcg->css);
1901 current->memcg_oom.memcg = memcg;
1902 current->memcg_oom.gfp_mask = mask;
1903 current->memcg_oom.order = order;
1907 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1908 * @handle: actually kill/wait or just clean up the OOM state
1910 * This has to be called at the end of a page fault if the memcg OOM
1911 * handler was enabled.
1913 * Memcg supports userspace OOM handling where failed allocations must
1914 * sleep on a waitqueue until the userspace task resolves the
1915 * situation. Sleeping directly in the charge context with all kinds
1916 * of locks held is not a good idea, instead we remember an OOM state
1917 * in the task and mem_cgroup_oom_synchronize() has to be called at
1918 * the end of the page fault to complete the OOM handling.
1920 * Returns %true if an ongoing memcg OOM situation was detected and
1921 * completed, %false otherwise.
1923 bool mem_cgroup_oom_synchronize(bool handle)
1925 struct mem_cgroup *memcg = current->memcg_oom.memcg;
1926 struct oom_wait_info owait;
1929 /* OOM is global, do not handle */
1933 if (!handle || oom_killer_disabled)
1936 owait.memcg = memcg;
1937 owait.wait.flags = 0;
1938 owait.wait.func = memcg_oom_wake_function;
1939 owait.wait.private = current;
1940 INIT_LIST_HEAD(&owait.wait.task_list);
1942 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1943 mem_cgroup_mark_under_oom(memcg);
1945 locked = mem_cgroup_oom_trylock(memcg);
1948 mem_cgroup_oom_notify(memcg);
1950 if (locked && !memcg->oom_kill_disable) {
1951 mem_cgroup_unmark_under_oom(memcg);
1952 finish_wait(&memcg_oom_waitq, &owait.wait);
1953 mem_cgroup_out_of_memory(memcg, current->memcg_oom.gfp_mask,
1954 current->memcg_oom.order);
1957 mem_cgroup_unmark_under_oom(memcg);
1958 finish_wait(&memcg_oom_waitq, &owait.wait);
1962 mem_cgroup_oom_unlock(memcg);
1964 * There is no guarantee that an OOM-lock contender
1965 * sees the wakeups triggered by the OOM kill
1966 * uncharges. Wake any sleepers explicitely.
1968 memcg_oom_recover(memcg);
1971 current->memcg_oom.memcg = NULL;
1972 css_put(&memcg->css);
1977 * mem_cgroup_begin_page_stat - begin a page state statistics transaction
1978 * @page: page that is going to change accounted state
1980 * This function must mark the beginning of an accounted page state
1981 * change to prevent double accounting when the page is concurrently
1982 * being moved to another memcg:
1984 * memcg = mem_cgroup_begin_page_stat(page);
1985 * if (TestClearPageState(page))
1986 * mem_cgroup_update_page_stat(memcg, state, -1);
1987 * mem_cgroup_end_page_stat(memcg);
1989 struct mem_cgroup *mem_cgroup_begin_page_stat(struct page *page)
1991 struct mem_cgroup *memcg;
1992 unsigned long flags;
1995 * The RCU lock is held throughout the transaction. The fast
1996 * path can get away without acquiring the memcg->move_lock
1997 * because page moving starts with an RCU grace period.
1999 * The RCU lock also protects the memcg from being freed when
2000 * the page state that is going to change is the only thing
2001 * preventing the page from being uncharged.
2002 * E.g. end-writeback clearing PageWriteback(), which allows
2003 * migration to go ahead and uncharge the page before the
2004 * account transaction might be complete.
2008 if (mem_cgroup_disabled())
2011 memcg = page->mem_cgroup;
2012 if (unlikely(!memcg))
2015 if (atomic_read(&memcg->moving_account) <= 0)
2018 spin_lock_irqsave(&memcg->move_lock, flags);
2019 if (memcg != page->mem_cgroup) {
2020 spin_unlock_irqrestore(&memcg->move_lock, flags);
2025 * When charge migration first begins, we can have locked and
2026 * unlocked page stat updates happening concurrently. Track
2027 * the task who has the lock for mem_cgroup_end_page_stat().
2029 memcg->move_lock_task = current;
2030 memcg->move_lock_flags = flags;
2036 * mem_cgroup_end_page_stat - finish a page state statistics transaction
2037 * @memcg: the memcg that was accounted against
2039 void mem_cgroup_end_page_stat(struct mem_cgroup *memcg)
2041 if (memcg && memcg->move_lock_task == current) {
2042 unsigned long flags = memcg->move_lock_flags;
2044 memcg->move_lock_task = NULL;
2045 memcg->move_lock_flags = 0;
2047 spin_unlock_irqrestore(&memcg->move_lock, flags);
2054 * mem_cgroup_update_page_stat - update page state statistics
2055 * @memcg: memcg to account against
2056 * @idx: page state item to account
2057 * @val: number of pages (positive or negative)
2059 * See mem_cgroup_begin_page_stat() for locking requirements.
2061 void mem_cgroup_update_page_stat(struct mem_cgroup *memcg,
2062 enum mem_cgroup_stat_index idx, int val)
2064 VM_BUG_ON(!rcu_read_lock_held());
2067 this_cpu_add(memcg->stat->count[idx], val);
2071 * size of first charge trial. "32" comes from vmscan.c's magic value.
2072 * TODO: maybe necessary to use big numbers in big irons.
2074 #define CHARGE_BATCH 32U
2075 struct memcg_stock_pcp {
2076 struct mem_cgroup *cached; /* this never be root cgroup */
2077 unsigned int nr_pages;
2078 struct work_struct work;
2079 unsigned long flags;
2080 #define FLUSHING_CACHED_CHARGE 0
2082 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2083 static DEFINE_MUTEX(percpu_charge_mutex);
2086 * consume_stock: Try to consume stocked charge on this cpu.
2087 * @memcg: memcg to consume from.
2088 * @nr_pages: how many pages to charge.
2090 * The charges will only happen if @memcg matches the current cpu's memcg
2091 * stock, and at least @nr_pages are available in that stock. Failure to
2092 * service an allocation will refill the stock.
2094 * returns true if successful, false otherwise.
2096 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2098 struct memcg_stock_pcp *stock;
2101 if (nr_pages > CHARGE_BATCH)
2104 stock = &get_cpu_var(memcg_stock);
2105 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2106 stock->nr_pages -= nr_pages;
2109 put_cpu_var(memcg_stock);
2114 * Returns stocks cached in percpu and reset cached information.
2116 static void drain_stock(struct memcg_stock_pcp *stock)
2118 struct mem_cgroup *old = stock->cached;
2120 if (stock->nr_pages) {
2121 page_counter_uncharge(&old->memory, stock->nr_pages);
2122 if (do_swap_account)
2123 page_counter_uncharge(&old->memsw, stock->nr_pages);
2124 css_put_many(&old->css, stock->nr_pages);
2125 stock->nr_pages = 0;
2127 stock->cached = NULL;
2131 * This must be called under preempt disabled or must be called by
2132 * a thread which is pinned to local cpu.
2134 static void drain_local_stock(struct work_struct *dummy)
2136 struct memcg_stock_pcp *stock = this_cpu_ptr(&memcg_stock);
2138 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2141 static void __init memcg_stock_init(void)
2145 for_each_possible_cpu(cpu) {
2146 struct memcg_stock_pcp *stock =
2147 &per_cpu(memcg_stock, cpu);
2148 INIT_WORK(&stock->work, drain_local_stock);
2153 * Cache charges(val) to local per_cpu area.
2154 * This will be consumed by consume_stock() function, later.
2156 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2158 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2160 if (stock->cached != memcg) { /* reset if necessary */
2162 stock->cached = memcg;
2164 stock->nr_pages += nr_pages;
2165 put_cpu_var(memcg_stock);
2169 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2170 * of the hierarchy under it.
2172 static void drain_all_stock(struct mem_cgroup *root_memcg)
2176 /* If someone's already draining, avoid adding running more workers. */
2177 if (!mutex_trylock(&percpu_charge_mutex))
2179 /* Notify other cpus that system-wide "drain" is running */
2182 for_each_online_cpu(cpu) {
2183 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2184 struct mem_cgroup *memcg;
2186 memcg = stock->cached;
2187 if (!memcg || !stock->nr_pages)
2189 if (!mem_cgroup_is_descendant(memcg, root_memcg))
2191 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2193 drain_local_stock(&stock->work);
2195 schedule_work_on(cpu, &stock->work);
2200 mutex_unlock(&percpu_charge_mutex);
2204 * This function drains percpu counter value from DEAD cpu and
2205 * move it to local cpu. Note that this function can be preempted.
2207 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2211 spin_lock(&memcg->pcp_counter_lock);
2212 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2213 long x = per_cpu(memcg->stat->count[i], cpu);
2215 per_cpu(memcg->stat->count[i], cpu) = 0;
2216 memcg->nocpu_base.count[i] += x;
2218 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2219 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2221 per_cpu(memcg->stat->events[i], cpu) = 0;
2222 memcg->nocpu_base.events[i] += x;
2224 spin_unlock(&memcg->pcp_counter_lock);
2227 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
2228 unsigned long action,
2231 int cpu = (unsigned long)hcpu;
2232 struct memcg_stock_pcp *stock;
2233 struct mem_cgroup *iter;
2235 if (action == CPU_ONLINE)
2238 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2241 for_each_mem_cgroup(iter)
2242 mem_cgroup_drain_pcp_counter(iter, cpu);
2244 stock = &per_cpu(memcg_stock, cpu);
2249 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2250 unsigned int nr_pages)
2252 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2253 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2254 struct mem_cgroup *mem_over_limit;
2255 struct page_counter *counter;
2256 unsigned long nr_reclaimed;
2257 bool may_swap = true;
2258 bool drained = false;
2261 if (mem_cgroup_is_root(memcg))
2264 if (consume_stock(memcg, nr_pages))
2267 if (!do_swap_account ||
2268 !page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2269 if (!page_counter_try_charge(&memcg->memory, batch, &counter))
2271 if (do_swap_account)
2272 page_counter_uncharge(&memcg->memsw, batch);
2273 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2275 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2279 if (batch > nr_pages) {
2285 * Unlike in global OOM situations, memcg is not in a physical
2286 * memory shortage. Allow dying and OOM-killed tasks to
2287 * bypass the last charges so that they can exit quickly and
2288 * free their memory.
2290 if (unlikely(test_thread_flag(TIF_MEMDIE) ||
2291 fatal_signal_pending(current) ||
2292 current->flags & PF_EXITING))
2295 if (unlikely(task_in_memcg_oom(current)))
2298 if (!(gfp_mask & __GFP_WAIT))
2301 mem_cgroup_events(mem_over_limit, MEMCG_MAX, 1);
2303 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2304 gfp_mask, may_swap);
2306 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2310 drain_all_stock(mem_over_limit);
2315 if (gfp_mask & __GFP_NORETRY)
2318 * Even though the limit is exceeded at this point, reclaim
2319 * may have been able to free some pages. Retry the charge
2320 * before killing the task.
2322 * Only for regular pages, though: huge pages are rather
2323 * unlikely to succeed so close to the limit, and we fall back
2324 * to regular pages anyway in case of failure.
2326 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2329 * At task move, charge accounts can be doubly counted. So, it's
2330 * better to wait until the end of task_move if something is going on.
2332 if (mem_cgroup_wait_acct_move(mem_over_limit))
2338 if (gfp_mask & __GFP_NOFAIL)
2341 if (fatal_signal_pending(current))
2344 mem_cgroup_events(mem_over_limit, MEMCG_OOM, 1);
2346 mem_cgroup_oom(mem_over_limit, gfp_mask, get_order(nr_pages));
2348 if (!(gfp_mask & __GFP_NOFAIL))
2354 css_get_many(&memcg->css, batch);
2355 if (batch > nr_pages)
2356 refill_stock(memcg, batch - nr_pages);
2358 * If the hierarchy is above the normal consumption range,
2359 * make the charging task trim their excess contribution.
2362 if (page_counter_read(&memcg->memory) <= memcg->high)
2364 mem_cgroup_events(memcg, MEMCG_HIGH, 1);
2365 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
2366 } while ((memcg = parent_mem_cgroup(memcg)));
2371 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2373 if (mem_cgroup_is_root(memcg))
2376 page_counter_uncharge(&memcg->memory, nr_pages);
2377 if (do_swap_account)
2378 page_counter_uncharge(&memcg->memsw, nr_pages);
2380 css_put_many(&memcg->css, nr_pages);
2384 * A helper function to get mem_cgroup from ID. must be called under
2385 * rcu_read_lock(). The caller is responsible for calling
2386 * css_tryget_online() if the mem_cgroup is used for charging. (dropping
2387 * refcnt from swap can be called against removed memcg.)
2389 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2391 /* ID 0 is unused ID */
2394 return mem_cgroup_from_id(id);
2398 * try_get_mem_cgroup_from_page - look up page's memcg association
2401 * Look up, get a css reference, and return the memcg that owns @page.
2403 * The page must be locked to prevent racing with swap-in and page
2404 * cache charges. If coming from an unlocked page table, the caller
2405 * must ensure the page is on the LRU or this can race with charging.
2407 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2409 struct mem_cgroup *memcg;
2413 VM_BUG_ON_PAGE(!PageLocked(page), page);
2415 memcg = page->mem_cgroup;
2417 if (!css_tryget_online(&memcg->css))
2419 } else if (PageSwapCache(page)) {
2420 ent.val = page_private(page);
2421 id = lookup_swap_cgroup_id(ent);
2423 memcg = mem_cgroup_lookup(id);
2424 if (memcg && !css_tryget_online(&memcg->css))
2431 static void lock_page_lru(struct page *page, int *isolated)
2433 struct zone *zone = page_zone(page);
2435 spin_lock_irq(&zone->lru_lock);
2436 if (PageLRU(page)) {
2437 struct lruvec *lruvec;
2439 lruvec = mem_cgroup_page_lruvec(page, zone);
2441 del_page_from_lru_list(page, lruvec, page_lru(page));
2447 static void unlock_page_lru(struct page *page, int isolated)
2449 struct zone *zone = page_zone(page);
2452 struct lruvec *lruvec;
2454 lruvec = mem_cgroup_page_lruvec(page, zone);
2455 VM_BUG_ON_PAGE(PageLRU(page), page);
2457 add_page_to_lru_list(page, lruvec, page_lru(page));
2459 spin_unlock_irq(&zone->lru_lock);
2462 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2467 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2470 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2471 * may already be on some other mem_cgroup's LRU. Take care of it.
2474 lock_page_lru(page, &isolated);
2477 * Nobody should be changing or seriously looking at
2478 * page->mem_cgroup at this point:
2480 * - the page is uncharged
2482 * - the page is off-LRU
2484 * - an anonymous fault has exclusive page access, except for
2485 * a locked page table
2487 * - a page cache insertion, a swapin fault, or a migration
2488 * have the page locked
2490 page->mem_cgroup = memcg;
2493 unlock_page_lru(page, isolated);
2496 #ifdef CONFIG_MEMCG_KMEM
2497 int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp,
2498 unsigned long nr_pages)
2500 struct page_counter *counter;
2503 ret = page_counter_try_charge(&memcg->kmem, nr_pages, &counter);
2507 ret = try_charge(memcg, gfp, nr_pages);
2508 if (ret == -EINTR) {
2510 * try_charge() chose to bypass to root due to OOM kill or
2511 * fatal signal. Since our only options are to either fail
2512 * the allocation or charge it to this cgroup, do it as a
2513 * temporary condition. But we can't fail. From a kmem/slab
2514 * perspective, the cache has already been selected, by
2515 * mem_cgroup_kmem_get_cache(), so it is too late to change
2518 * This condition will only trigger if the task entered
2519 * memcg_charge_kmem in a sane state, but was OOM-killed
2520 * during try_charge() above. Tasks that were already dying
2521 * when the allocation triggers should have been already
2522 * directed to the root cgroup in memcontrol.h
2524 page_counter_charge(&memcg->memory, nr_pages);
2525 if (do_swap_account)
2526 page_counter_charge(&memcg->memsw, nr_pages);
2527 css_get_many(&memcg->css, nr_pages);
2530 page_counter_uncharge(&memcg->kmem, nr_pages);
2535 void memcg_uncharge_kmem(struct mem_cgroup *memcg, unsigned long nr_pages)
2537 page_counter_uncharge(&memcg->memory, nr_pages);
2538 if (do_swap_account)
2539 page_counter_uncharge(&memcg->memsw, nr_pages);
2541 page_counter_uncharge(&memcg->kmem, nr_pages);
2543 css_put_many(&memcg->css, nr_pages);
2547 * helper for acessing a memcg's index. It will be used as an index in the
2548 * child cache array in kmem_cache, and also to derive its name. This function
2549 * will return -1 when this is not a kmem-limited memcg.
2551 int memcg_cache_id(struct mem_cgroup *memcg)
2553 return memcg ? memcg->kmemcg_id : -1;
2556 static int memcg_alloc_cache_id(void)
2561 id = ida_simple_get(&kmem_limited_groups,
2562 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2566 if (id < memcg_limited_groups_array_size)
2570 * There's no space for the new id in memcg_caches arrays,
2571 * so we have to grow them.
2574 size = 2 * (id + 1);
2575 if (size < MEMCG_CACHES_MIN_SIZE)
2576 size = MEMCG_CACHES_MIN_SIZE;
2577 else if (size > MEMCG_CACHES_MAX_SIZE)
2578 size = MEMCG_CACHES_MAX_SIZE;
2580 err = memcg_update_all_caches(size);
2582 ida_simple_remove(&kmem_limited_groups, id);
2588 static void memcg_free_cache_id(int id)
2590 ida_simple_remove(&kmem_limited_groups, id);
2594 * We should update the current array size iff all caches updates succeed. This
2595 * can only be done from the slab side. The slab mutex needs to be held when
2598 void memcg_update_array_size(int num)
2600 memcg_limited_groups_array_size = num;
2603 struct memcg_kmem_cache_create_work {
2604 struct mem_cgroup *memcg;
2605 struct kmem_cache *cachep;
2606 struct work_struct work;
2609 static void memcg_kmem_cache_create_func(struct work_struct *w)
2611 struct memcg_kmem_cache_create_work *cw =
2612 container_of(w, struct memcg_kmem_cache_create_work, work);
2613 struct mem_cgroup *memcg = cw->memcg;
2614 struct kmem_cache *cachep = cw->cachep;
2616 memcg_create_kmem_cache(memcg, cachep);
2618 css_put(&memcg->css);
2623 * Enqueue the creation of a per-memcg kmem_cache.
2625 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2626 struct kmem_cache *cachep)
2628 struct memcg_kmem_cache_create_work *cw;
2630 cw = kmalloc(sizeof(*cw), GFP_NOWAIT);
2634 css_get(&memcg->css);
2637 cw->cachep = cachep;
2638 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2640 schedule_work(&cw->work);
2643 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2644 struct kmem_cache *cachep)
2647 * We need to stop accounting when we kmalloc, because if the
2648 * corresponding kmalloc cache is not yet created, the first allocation
2649 * in __memcg_schedule_kmem_cache_create will recurse.
2651 * However, it is better to enclose the whole function. Depending on
2652 * the debugging options enabled, INIT_WORK(), for instance, can
2653 * trigger an allocation. This too, will make us recurse. Because at
2654 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2655 * the safest choice is to do it like this, wrapping the whole function.
2657 current->memcg_kmem_skip_account = 1;
2658 __memcg_schedule_kmem_cache_create(memcg, cachep);
2659 current->memcg_kmem_skip_account = 0;
2663 * Return the kmem_cache we're supposed to use for a slab allocation.
2664 * We try to use the current memcg's version of the cache.
2666 * If the cache does not exist yet, if we are the first user of it,
2667 * we either create it immediately, if possible, or create it asynchronously
2669 * In the latter case, we will let the current allocation go through with
2670 * the original cache.
2672 * Can't be called in interrupt context or from kernel threads.
2673 * This function needs to be called with rcu_read_lock() held.
2675 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep)
2677 struct mem_cgroup *memcg;
2678 struct kmem_cache *memcg_cachep;
2680 VM_BUG_ON(!cachep->memcg_params);
2681 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
2683 if (current->memcg_kmem_skip_account)
2686 memcg = get_mem_cgroup_from_mm(current->mm);
2687 if (!memcg_kmem_is_active(memcg))
2690 memcg_cachep = cache_from_memcg_idx(cachep, memcg_cache_id(memcg));
2691 if (likely(memcg_cachep))
2692 return memcg_cachep;
2695 * If we are in a safe context (can wait, and not in interrupt
2696 * context), we could be be predictable and return right away.
2697 * This would guarantee that the allocation being performed
2698 * already belongs in the new cache.
2700 * However, there are some clashes that can arrive from locking.
2701 * For instance, because we acquire the slab_mutex while doing
2702 * memcg_create_kmem_cache, this means no further allocation
2703 * could happen with the slab_mutex held. So it's better to
2706 memcg_schedule_kmem_cache_create(memcg, cachep);
2708 css_put(&memcg->css);
2712 void __memcg_kmem_put_cache(struct kmem_cache *cachep)
2714 if (!is_root_cache(cachep))
2715 css_put(&cachep->memcg_params->memcg->css);
2719 * We need to verify if the allocation against current->mm->owner's memcg is
2720 * possible for the given order. But the page is not allocated yet, so we'll
2721 * need a further commit step to do the final arrangements.
2723 * It is possible for the task to switch cgroups in this mean time, so at
2724 * commit time, we can't rely on task conversion any longer. We'll then use
2725 * the handle argument to return to the caller which cgroup we should commit
2726 * against. We could also return the memcg directly and avoid the pointer
2727 * passing, but a boolean return value gives better semantics considering
2728 * the compiled-out case as well.
2730 * Returning true means the allocation is possible.
2733 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
2735 struct mem_cgroup *memcg;
2740 memcg = get_mem_cgroup_from_mm(current->mm);
2742 if (!memcg_kmem_is_active(memcg)) {
2743 css_put(&memcg->css);
2747 ret = memcg_charge_kmem(memcg, gfp, 1 << order);
2751 css_put(&memcg->css);
2755 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
2758 VM_BUG_ON(mem_cgroup_is_root(memcg));
2760 /* The page allocation failed. Revert */
2762 memcg_uncharge_kmem(memcg, 1 << order);
2765 page->mem_cgroup = memcg;
2768 void __memcg_kmem_uncharge_pages(struct page *page, int order)
2770 struct mem_cgroup *memcg = page->mem_cgroup;
2775 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2777 memcg_uncharge_kmem(memcg, 1 << order);
2778 page->mem_cgroup = NULL;
2780 #endif /* CONFIG_MEMCG_KMEM */
2782 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2785 * Because tail pages are not marked as "used", set it. We're under
2786 * zone->lru_lock, 'splitting on pmd' and compound_lock.
2787 * charge/uncharge will be never happen and move_account() is done under
2788 * compound_lock(), so we don't have to take care of races.
2790 void mem_cgroup_split_huge_fixup(struct page *head)
2794 if (mem_cgroup_disabled())
2797 for (i = 1; i < HPAGE_PMD_NR; i++)
2798 head[i].mem_cgroup = head->mem_cgroup;
2800 __this_cpu_sub(head->mem_cgroup->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
2803 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2806 * mem_cgroup_move_account - move account of the page
2808 * @nr_pages: number of regular pages (>1 for huge pages)
2809 * @from: mem_cgroup which the page is moved from.
2810 * @to: mem_cgroup which the page is moved to. @from != @to.
2812 * The caller must confirm following.
2813 * - page is not on LRU (isolate_page() is useful.)
2814 * - compound_lock is held when nr_pages > 1
2816 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
2819 static int mem_cgroup_move_account(struct page *page,
2820 unsigned int nr_pages,
2821 struct mem_cgroup *from,
2822 struct mem_cgroup *to)
2824 unsigned long flags;
2827 VM_BUG_ON(from == to);
2828 VM_BUG_ON_PAGE(PageLRU(page), page);
2830 * The page is isolated from LRU. So, collapse function
2831 * will not handle this page. But page splitting can happen.
2832 * Do this check under compound_page_lock(). The caller should
2836 if (nr_pages > 1 && !PageTransHuge(page))
2840 * Prevent mem_cgroup_migrate() from looking at page->mem_cgroup
2841 * of its source page while we change it: page migration takes
2842 * both pages off the LRU, but page cache replacement doesn't.
2844 if (!trylock_page(page))
2848 if (page->mem_cgroup != from)
2851 spin_lock_irqsave(&from->move_lock, flags);
2853 if (!PageAnon(page) && page_mapped(page)) {
2854 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
2856 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
2860 if (PageWriteback(page)) {
2861 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK],
2863 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK],
2868 * It is safe to change page->mem_cgroup here because the page
2869 * is referenced, charged, and isolated - we can't race with
2870 * uncharging, charging, migration, or LRU putback.
2873 /* caller should have done css_get */
2874 page->mem_cgroup = to;
2875 spin_unlock_irqrestore(&from->move_lock, flags);
2879 local_irq_disable();
2880 mem_cgroup_charge_statistics(to, page, nr_pages);
2881 memcg_check_events(to, page);
2882 mem_cgroup_charge_statistics(from, page, -nr_pages);
2883 memcg_check_events(from, page);
2891 #ifdef CONFIG_MEMCG_SWAP
2892 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
2895 int val = (charge) ? 1 : -1;
2896 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
2900 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2901 * @entry: swap entry to be moved
2902 * @from: mem_cgroup which the entry is moved from
2903 * @to: mem_cgroup which the entry is moved to
2905 * It succeeds only when the swap_cgroup's record for this entry is the same
2906 * as the mem_cgroup's id of @from.
2908 * Returns 0 on success, -EINVAL on failure.
2910 * The caller must have charged to @to, IOW, called page_counter_charge() about
2911 * both res and memsw, and called css_get().
2913 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2914 struct mem_cgroup *from, struct mem_cgroup *to)
2916 unsigned short old_id, new_id;
2918 old_id = mem_cgroup_id(from);
2919 new_id = mem_cgroup_id(to);
2921 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2922 mem_cgroup_swap_statistics(from, false);
2923 mem_cgroup_swap_statistics(to, true);
2929 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2930 struct mem_cgroup *from, struct mem_cgroup *to)
2936 static DEFINE_MUTEX(memcg_limit_mutex);
2938 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
2939 unsigned long limit)
2941 unsigned long curusage;
2942 unsigned long oldusage;
2943 bool enlarge = false;
2948 * For keeping hierarchical_reclaim simple, how long we should retry
2949 * is depends on callers. We set our retry-count to be function
2950 * of # of children which we should visit in this loop.
2952 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2953 mem_cgroup_count_children(memcg);
2955 oldusage = page_counter_read(&memcg->memory);
2958 if (signal_pending(current)) {
2963 mutex_lock(&memcg_limit_mutex);
2964 if (limit > memcg->memsw.limit) {
2965 mutex_unlock(&memcg_limit_mutex);
2969 if (limit > memcg->memory.limit)
2971 ret = page_counter_limit(&memcg->memory, limit);
2972 mutex_unlock(&memcg_limit_mutex);
2977 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, true);
2979 curusage = page_counter_read(&memcg->memory);
2980 /* Usage is reduced ? */
2981 if (curusage >= oldusage)
2984 oldusage = curusage;
2985 } while (retry_count);
2987 if (!ret && enlarge)
2988 memcg_oom_recover(memcg);
2993 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
2994 unsigned long limit)
2996 unsigned long curusage;
2997 unsigned long oldusage;
2998 bool enlarge = false;
3002 /* see mem_cgroup_resize_res_limit */
3003 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
3004 mem_cgroup_count_children(memcg);
3006 oldusage = page_counter_read(&memcg->memsw);
3009 if (signal_pending(current)) {
3014 mutex_lock(&memcg_limit_mutex);
3015 if (limit < memcg->memory.limit) {
3016 mutex_unlock(&memcg_limit_mutex);
3020 if (limit > memcg->memsw.limit)
3022 ret = page_counter_limit(&memcg->memsw, limit);
3023 mutex_unlock(&memcg_limit_mutex);
3028 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, false);
3030 curusage = page_counter_read(&memcg->memsw);
3031 /* Usage is reduced ? */
3032 if (curusage >= oldusage)
3035 oldusage = curusage;
3036 } while (retry_count);
3038 if (!ret && enlarge)
3039 memcg_oom_recover(memcg);
3044 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3046 unsigned long *total_scanned)
3048 unsigned long nr_reclaimed = 0;
3049 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3050 unsigned long reclaimed;
3052 struct mem_cgroup_tree_per_zone *mctz;
3053 unsigned long excess;
3054 unsigned long nr_scanned;
3059 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3061 * This loop can run a while, specially if mem_cgroup's continuously
3062 * keep exceeding their soft limit and putting the system under
3069 mz = mem_cgroup_largest_soft_limit_node(mctz);
3074 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
3075 gfp_mask, &nr_scanned);
3076 nr_reclaimed += reclaimed;
3077 *total_scanned += nr_scanned;
3078 spin_lock_irq(&mctz->lock);
3079 __mem_cgroup_remove_exceeded(mz, mctz);
3082 * If we failed to reclaim anything from this memory cgroup
3083 * it is time to move on to the next cgroup
3087 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3089 excess = soft_limit_excess(mz->memcg);
3091 * One school of thought says that we should not add
3092 * back the node to the tree if reclaim returns 0.
3093 * But our reclaim could return 0, simply because due
3094 * to priority we are exposing a smaller subset of
3095 * memory to reclaim from. Consider this as a longer
3098 /* If excess == 0, no tree ops */
3099 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3100 spin_unlock_irq(&mctz->lock);
3101 css_put(&mz->memcg->css);
3104 * Could not reclaim anything and there are no more
3105 * mem cgroups to try or we seem to be looping without
3106 * reclaiming anything.
3108 if (!nr_reclaimed &&
3110 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3112 } while (!nr_reclaimed);
3114 css_put(&next_mz->memcg->css);
3115 return nr_reclaimed;
3119 * Test whether @memcg has children, dead or alive. Note that this
3120 * function doesn't care whether @memcg has use_hierarchy enabled and
3121 * returns %true if there are child csses according to the cgroup
3122 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
3124 static inline bool memcg_has_children(struct mem_cgroup *memcg)
3129 * The lock does not prevent addition or deletion of children, but
3130 * it prevents a new child from being initialized based on this
3131 * parent in css_online(), so it's enough to decide whether
3132 * hierarchically inherited attributes can still be changed or not.
3134 lockdep_assert_held(&memcg_create_mutex);
3137 ret = css_next_child(NULL, &memcg->css);
3143 * Reclaims as many pages from the given memcg as possible and moves
3144 * the rest to the parent.
3146 * Caller is responsible for holding css reference for memcg.
3148 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3150 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3152 /* we call try-to-free pages for make this cgroup empty */
3153 lru_add_drain_all();
3154 /* try to free all pages in this cgroup */
3155 while (nr_retries && page_counter_read(&memcg->memory)) {
3158 if (signal_pending(current))
3161 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3165 /* maybe some writeback is necessary */
3166 congestion_wait(BLK_RW_ASYNC, HZ/10);
3174 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3175 char *buf, size_t nbytes,
3178 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3180 if (mem_cgroup_is_root(memcg))
3182 return mem_cgroup_force_empty(memcg) ?: nbytes;
3185 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3188 return mem_cgroup_from_css(css)->use_hierarchy;
3191 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3192 struct cftype *cft, u64 val)
3195 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3196 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
3198 mutex_lock(&memcg_create_mutex);
3200 if (memcg->use_hierarchy == val)
3204 * If parent's use_hierarchy is set, we can't make any modifications
3205 * in the child subtrees. If it is unset, then the change can
3206 * occur, provided the current cgroup has no children.
3208 * For the root cgroup, parent_mem is NULL, we allow value to be
3209 * set if there are no children.
3211 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3212 (val == 1 || val == 0)) {
3213 if (!memcg_has_children(memcg))
3214 memcg->use_hierarchy = val;
3221 mutex_unlock(&memcg_create_mutex);
3226 static unsigned long tree_stat(struct mem_cgroup *memcg,
3227 enum mem_cgroup_stat_index idx)
3229 struct mem_cgroup *iter;
3232 /* Per-cpu values can be negative, use a signed accumulator */
3233 for_each_mem_cgroup_tree(iter, memcg)
3234 val += mem_cgroup_read_stat(iter, idx);
3236 if (val < 0) /* race ? */
3241 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3245 if (mem_cgroup_is_root(memcg)) {
3246 val = tree_stat(memcg, MEM_CGROUP_STAT_CACHE);
3247 val += tree_stat(memcg, MEM_CGROUP_STAT_RSS);
3249 val += tree_stat(memcg, MEM_CGROUP_STAT_SWAP);
3252 val = page_counter_read(&memcg->memory);
3254 val = page_counter_read(&memcg->memsw);
3256 return val << PAGE_SHIFT;
3267 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3270 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3271 struct page_counter *counter;
3273 switch (MEMFILE_TYPE(cft->private)) {
3275 counter = &memcg->memory;
3278 counter = &memcg->memsw;
3281 counter = &memcg->kmem;
3287 switch (MEMFILE_ATTR(cft->private)) {
3289 if (counter == &memcg->memory)
3290 return mem_cgroup_usage(memcg, false);
3291 if (counter == &memcg->memsw)
3292 return mem_cgroup_usage(memcg, true);
3293 return (u64)page_counter_read(counter) * PAGE_SIZE;
3295 return (u64)counter->limit * PAGE_SIZE;
3297 return (u64)counter->watermark * PAGE_SIZE;
3299 return counter->failcnt;
3300 case RES_SOFT_LIMIT:
3301 return (u64)memcg->soft_limit * PAGE_SIZE;
3307 #ifdef CONFIG_MEMCG_KMEM
3308 static int memcg_activate_kmem(struct mem_cgroup *memcg,
3309 unsigned long nr_pages)
3314 if (memcg_kmem_is_active(memcg))
3318 * For simplicity, we won't allow this to be disabled. It also can't
3319 * be changed if the cgroup has children already, or if tasks had
3322 * If tasks join before we set the limit, a person looking at
3323 * kmem.usage_in_bytes will have no way to determine when it took
3324 * place, which makes the value quite meaningless.
3326 * After it first became limited, changes in the value of the limit are
3327 * of course permitted.
3329 mutex_lock(&memcg_create_mutex);
3330 if (cgroup_has_tasks(memcg->css.cgroup) ||
3331 (memcg->use_hierarchy && memcg_has_children(memcg)))
3333 mutex_unlock(&memcg_create_mutex);
3337 memcg_id = memcg_alloc_cache_id();
3344 * We couldn't have accounted to this cgroup, because it hasn't got
3345 * activated yet, so this should succeed.
3347 err = page_counter_limit(&memcg->kmem, nr_pages);
3350 static_key_slow_inc(&memcg_kmem_enabled_key);
3352 * A memory cgroup is considered kmem-active as soon as it gets
3353 * kmemcg_id. Setting the id after enabling static branching will
3354 * guarantee no one starts accounting before all call sites are
3357 memcg->kmemcg_id = memcg_id;
3362 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
3363 unsigned long limit)
3367 mutex_lock(&memcg_limit_mutex);
3368 if (!memcg_kmem_is_active(memcg))
3369 ret = memcg_activate_kmem(memcg, limit);
3371 ret = page_counter_limit(&memcg->kmem, limit);
3372 mutex_unlock(&memcg_limit_mutex);
3376 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
3379 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
3384 mutex_lock(&memcg_limit_mutex);
3386 * If the parent cgroup is not kmem-active now, it cannot be activated
3387 * after this point, because it has at least one child already.
3389 if (memcg_kmem_is_active(parent))
3390 ret = memcg_activate_kmem(memcg, PAGE_COUNTER_MAX);
3391 mutex_unlock(&memcg_limit_mutex);
3395 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
3396 unsigned long limit)
3400 #endif /* CONFIG_MEMCG_KMEM */
3403 * The user of this function is...
3406 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3407 char *buf, size_t nbytes, loff_t off)
3409 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3410 unsigned long nr_pages;
3413 buf = strstrip(buf);
3414 ret = page_counter_memparse(buf, "-1", &nr_pages);
3418 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3420 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3424 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3426 ret = mem_cgroup_resize_limit(memcg, nr_pages);
3429 ret = mem_cgroup_resize_memsw_limit(memcg, nr_pages);
3432 ret = memcg_update_kmem_limit(memcg, nr_pages);
3436 case RES_SOFT_LIMIT:
3437 memcg->soft_limit = nr_pages;
3441 return ret ?: nbytes;
3444 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3445 size_t nbytes, loff_t off)
3447 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3448 struct page_counter *counter;
3450 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3452 counter = &memcg->memory;
3455 counter = &memcg->memsw;
3458 counter = &memcg->kmem;
3464 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3466 page_counter_reset_watermark(counter);
3469 counter->failcnt = 0;
3478 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3481 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3485 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3486 struct cftype *cft, u64 val)
3488 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3490 if (val & ~MOVE_MASK)
3494 * No kind of locking is needed in here, because ->can_attach() will
3495 * check this value once in the beginning of the process, and then carry
3496 * on with stale data. This means that changes to this value will only
3497 * affect task migrations starting after the change.
3499 memcg->move_charge_at_immigrate = val;
3503 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3504 struct cftype *cft, u64 val)
3511 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3515 unsigned int lru_mask;
3518 static const struct numa_stat stats[] = {
3519 { "total", LRU_ALL },
3520 { "file", LRU_ALL_FILE },
3521 { "anon", LRU_ALL_ANON },
3522 { "unevictable", BIT(LRU_UNEVICTABLE) },
3524 const struct numa_stat *stat;
3527 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3529 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3530 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3531 seq_printf(m, "%s=%lu", stat->name, nr);
3532 for_each_node_state(nid, N_MEMORY) {
3533 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3535 seq_printf(m, " N%d=%lu", nid, nr);
3540 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3541 struct mem_cgroup *iter;
3544 for_each_mem_cgroup_tree(iter, memcg)
3545 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3546 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3547 for_each_node_state(nid, N_MEMORY) {
3549 for_each_mem_cgroup_tree(iter, memcg)
3550 nr += mem_cgroup_node_nr_lru_pages(
3551 iter, nid, stat->lru_mask);
3552 seq_printf(m, " N%d=%lu", nid, nr);
3559 #endif /* CONFIG_NUMA */
3561 static int memcg_stat_show(struct seq_file *m, void *v)
3563 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3564 unsigned long memory, memsw;
3565 struct mem_cgroup *mi;
3568 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_stat_names) !=
3569 MEM_CGROUP_STAT_NSTATS);
3570 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_events_names) !=
3571 MEM_CGROUP_EVENTS_NSTATS);
3572 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3574 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3575 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
3577 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
3578 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
3581 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
3582 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
3583 mem_cgroup_read_events(memcg, i));
3585 for (i = 0; i < NR_LRU_LISTS; i++)
3586 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3587 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
3589 /* Hierarchical information */
3590 memory = memsw = PAGE_COUNTER_MAX;
3591 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3592 memory = min(memory, mi->memory.limit);
3593 memsw = min(memsw, mi->memsw.limit);
3595 seq_printf(m, "hierarchical_memory_limit %llu\n",
3596 (u64)memory * PAGE_SIZE);
3597 if (do_swap_account)
3598 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3599 (u64)memsw * PAGE_SIZE);
3601 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3604 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
3606 for_each_mem_cgroup_tree(mi, memcg)
3607 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
3608 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
3611 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
3612 unsigned long long val = 0;
3614 for_each_mem_cgroup_tree(mi, memcg)
3615 val += mem_cgroup_read_events(mi, i);
3616 seq_printf(m, "total_%s %llu\n",
3617 mem_cgroup_events_names[i], val);
3620 for (i = 0; i < NR_LRU_LISTS; i++) {
3621 unsigned long long val = 0;
3623 for_each_mem_cgroup_tree(mi, memcg)
3624 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
3625 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
3628 #ifdef CONFIG_DEBUG_VM
3631 struct mem_cgroup_per_zone *mz;
3632 struct zone_reclaim_stat *rstat;
3633 unsigned long recent_rotated[2] = {0, 0};
3634 unsigned long recent_scanned[2] = {0, 0};
3636 for_each_online_node(nid)
3637 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3638 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
3639 rstat = &mz->lruvec.reclaim_stat;
3641 recent_rotated[0] += rstat->recent_rotated[0];
3642 recent_rotated[1] += rstat->recent_rotated[1];
3643 recent_scanned[0] += rstat->recent_scanned[0];
3644 recent_scanned[1] += rstat->recent_scanned[1];
3646 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3647 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3648 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3649 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3656 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3659 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3661 return mem_cgroup_swappiness(memcg);
3664 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3665 struct cftype *cft, u64 val)
3667 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3673 memcg->swappiness = val;
3675 vm_swappiness = val;
3680 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3682 struct mem_cgroup_threshold_ary *t;
3683 unsigned long usage;
3688 t = rcu_dereference(memcg->thresholds.primary);
3690 t = rcu_dereference(memcg->memsw_thresholds.primary);
3695 usage = mem_cgroup_usage(memcg, swap);
3698 * current_threshold points to threshold just below or equal to usage.
3699 * If it's not true, a threshold was crossed after last
3700 * call of __mem_cgroup_threshold().
3702 i = t->current_threshold;
3705 * Iterate backward over array of thresholds starting from
3706 * current_threshold and check if a threshold is crossed.
3707 * If none of thresholds below usage is crossed, we read
3708 * only one element of the array here.
3710 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3711 eventfd_signal(t->entries[i].eventfd, 1);
3713 /* i = current_threshold + 1 */
3717 * Iterate forward over array of thresholds starting from
3718 * current_threshold+1 and check if a threshold is crossed.
3719 * If none of thresholds above usage is crossed, we read
3720 * only one element of the array here.
3722 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3723 eventfd_signal(t->entries[i].eventfd, 1);
3725 /* Update current_threshold */
3726 t->current_threshold = i - 1;
3731 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3734 __mem_cgroup_threshold(memcg, false);
3735 if (do_swap_account)
3736 __mem_cgroup_threshold(memcg, true);
3738 memcg = parent_mem_cgroup(memcg);
3742 static int compare_thresholds(const void *a, const void *b)
3744 const struct mem_cgroup_threshold *_a = a;
3745 const struct mem_cgroup_threshold *_b = b;
3747 if (_a->threshold > _b->threshold)
3750 if (_a->threshold < _b->threshold)
3756 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3758 struct mem_cgroup_eventfd_list *ev;
3760 spin_lock(&memcg_oom_lock);
3762 list_for_each_entry(ev, &memcg->oom_notify, list)
3763 eventfd_signal(ev->eventfd, 1);
3765 spin_unlock(&memcg_oom_lock);
3769 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3771 struct mem_cgroup *iter;
3773 for_each_mem_cgroup_tree(iter, memcg)
3774 mem_cgroup_oom_notify_cb(iter);
3777 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3778 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3780 struct mem_cgroup_thresholds *thresholds;
3781 struct mem_cgroup_threshold_ary *new;
3782 unsigned long threshold;
3783 unsigned long usage;
3786 ret = page_counter_memparse(args, "-1", &threshold);
3790 mutex_lock(&memcg->thresholds_lock);
3793 thresholds = &memcg->thresholds;
3794 usage = mem_cgroup_usage(memcg, false);
3795 } else if (type == _MEMSWAP) {
3796 thresholds = &memcg->memsw_thresholds;
3797 usage = mem_cgroup_usage(memcg, true);
3801 /* Check if a threshold crossed before adding a new one */
3802 if (thresholds->primary)
3803 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3805 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3807 /* Allocate memory for new array of thresholds */
3808 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3816 /* Copy thresholds (if any) to new array */
3817 if (thresholds->primary) {
3818 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3819 sizeof(struct mem_cgroup_threshold));
3822 /* Add new threshold */
3823 new->entries[size - 1].eventfd = eventfd;
3824 new->entries[size - 1].threshold = threshold;
3826 /* Sort thresholds. Registering of new threshold isn't time-critical */
3827 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3828 compare_thresholds, NULL);
3830 /* Find current threshold */
3831 new->current_threshold = -1;
3832 for (i = 0; i < size; i++) {
3833 if (new->entries[i].threshold <= usage) {
3835 * new->current_threshold will not be used until
3836 * rcu_assign_pointer(), so it's safe to increment
3839 ++new->current_threshold;
3844 /* Free old spare buffer and save old primary buffer as spare */
3845 kfree(thresholds->spare);
3846 thresholds->spare = thresholds->primary;
3848 rcu_assign_pointer(thresholds->primary, new);
3850 /* To be sure that nobody uses thresholds */
3854 mutex_unlock(&memcg->thresholds_lock);
3859 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3860 struct eventfd_ctx *eventfd, const char *args)
3862 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
3865 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
3866 struct eventfd_ctx *eventfd, const char *args)
3868 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
3871 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3872 struct eventfd_ctx *eventfd, enum res_type type)
3874 struct mem_cgroup_thresholds *thresholds;
3875 struct mem_cgroup_threshold_ary *new;
3876 unsigned long usage;
3879 mutex_lock(&memcg->thresholds_lock);
3882 thresholds = &memcg->thresholds;
3883 usage = mem_cgroup_usage(memcg, false);
3884 } else if (type == _MEMSWAP) {
3885 thresholds = &memcg->memsw_thresholds;
3886 usage = mem_cgroup_usage(memcg, true);
3890 if (!thresholds->primary)
3893 /* Check if a threshold crossed before removing */
3894 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3896 /* Calculate new number of threshold */
3898 for (i = 0; i < thresholds->primary->size; i++) {
3899 if (thresholds->primary->entries[i].eventfd != eventfd)
3903 new = thresholds->spare;
3905 /* Set thresholds array to NULL if we don't have thresholds */
3914 /* Copy thresholds and find current threshold */
3915 new->current_threshold = -1;
3916 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3917 if (thresholds->primary->entries[i].eventfd == eventfd)
3920 new->entries[j] = thresholds->primary->entries[i];
3921 if (new->entries[j].threshold <= usage) {
3923 * new->current_threshold will not be used
3924 * until rcu_assign_pointer(), so it's safe to increment
3927 ++new->current_threshold;
3933 /* Swap primary and spare array */
3934 thresholds->spare = thresholds->primary;
3935 /* If all events are unregistered, free the spare array */
3937 kfree(thresholds->spare);
3938 thresholds->spare = NULL;
3941 rcu_assign_pointer(thresholds->primary, new);
3943 /* To be sure that nobody uses thresholds */
3946 mutex_unlock(&memcg->thresholds_lock);
3949 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3950 struct eventfd_ctx *eventfd)
3952 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
3955 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3956 struct eventfd_ctx *eventfd)
3958 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
3961 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
3962 struct eventfd_ctx *eventfd, const char *args)
3964 struct mem_cgroup_eventfd_list *event;
3966 event = kmalloc(sizeof(*event), GFP_KERNEL);
3970 spin_lock(&memcg_oom_lock);
3972 event->eventfd = eventfd;
3973 list_add(&event->list, &memcg->oom_notify);
3975 /* already in OOM ? */
3976 if (atomic_read(&memcg->under_oom))
3977 eventfd_signal(eventfd, 1);
3978 spin_unlock(&memcg_oom_lock);
3983 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
3984 struct eventfd_ctx *eventfd)
3986 struct mem_cgroup_eventfd_list *ev, *tmp;
3988 spin_lock(&memcg_oom_lock);
3990 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
3991 if (ev->eventfd == eventfd) {
3992 list_del(&ev->list);
3997 spin_unlock(&memcg_oom_lock);
4000 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4002 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
4004 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4005 seq_printf(sf, "under_oom %d\n", (bool)atomic_read(&memcg->under_oom));
4009 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4010 struct cftype *cft, u64 val)
4012 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4014 /* cannot set to root cgroup and only 0 and 1 are allowed */
4015 if (!css->parent || !((val == 0) || (val == 1)))
4018 memcg->oom_kill_disable = val;
4020 memcg_oom_recover(memcg);
4025 #ifdef CONFIG_MEMCG_KMEM
4026 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
4030 ret = memcg_propagate_kmem(memcg);
4034 return mem_cgroup_sockets_init(memcg, ss);
4037 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
4039 memcg_destroy_kmem_caches(memcg);
4040 mem_cgroup_sockets_destroy(memcg);
4043 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
4048 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
4054 * DO NOT USE IN NEW FILES.
4056 * "cgroup.event_control" implementation.
4058 * This is way over-engineered. It tries to support fully configurable
4059 * events for each user. Such level of flexibility is completely
4060 * unnecessary especially in the light of the planned unified hierarchy.
4062 * Please deprecate this and replace with something simpler if at all
4067 * Unregister event and free resources.
4069 * Gets called from workqueue.
4071 static void memcg_event_remove(struct work_struct *work)
4073 struct mem_cgroup_event *event =
4074 container_of(work, struct mem_cgroup_event, remove);
4075 struct mem_cgroup *memcg = event->memcg;
4077 remove_wait_queue(event->wqh, &event->wait);
4079 event->unregister_event(memcg, event->eventfd);
4081 /* Notify userspace the event is going away. */
4082 eventfd_signal(event->eventfd, 1);
4084 eventfd_ctx_put(event->eventfd);
4086 css_put(&memcg->css);
4090 * Gets called on POLLHUP on eventfd when user closes it.
4092 * Called with wqh->lock held and interrupts disabled.
4094 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
4095 int sync, void *key)
4097 struct mem_cgroup_event *event =
4098 container_of(wait, struct mem_cgroup_event, wait);
4099 struct mem_cgroup *memcg = event->memcg;
4100 unsigned long flags = (unsigned long)key;
4102 if (flags & POLLHUP) {
4104 * If the event has been detached at cgroup removal, we
4105 * can simply return knowing the other side will cleanup
4108 * We can't race against event freeing since the other
4109 * side will require wqh->lock via remove_wait_queue(),
4112 spin_lock(&memcg->event_list_lock);
4113 if (!list_empty(&event->list)) {
4114 list_del_init(&event->list);
4116 * We are in atomic context, but cgroup_event_remove()
4117 * may sleep, so we have to call it in workqueue.
4119 schedule_work(&event->remove);
4121 spin_unlock(&memcg->event_list_lock);
4127 static void memcg_event_ptable_queue_proc(struct file *file,
4128 wait_queue_head_t *wqh, poll_table *pt)
4130 struct mem_cgroup_event *event =
4131 container_of(pt, struct mem_cgroup_event, pt);
4134 add_wait_queue(wqh, &event->wait);
4138 * DO NOT USE IN NEW FILES.
4140 * Parse input and register new cgroup event handler.
4142 * Input must be in format '<event_fd> <control_fd> <args>'.
4143 * Interpretation of args is defined by control file implementation.
4145 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4146 char *buf, size_t nbytes, loff_t off)
4148 struct cgroup_subsys_state *css = of_css(of);
4149 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4150 struct mem_cgroup_event *event;
4151 struct cgroup_subsys_state *cfile_css;
4152 unsigned int efd, cfd;
4159 buf = strstrip(buf);
4161 efd = simple_strtoul(buf, &endp, 10);
4166 cfd = simple_strtoul(buf, &endp, 10);
4167 if ((*endp != ' ') && (*endp != '\0'))
4171 event = kzalloc(sizeof(*event), GFP_KERNEL);
4175 event->memcg = memcg;
4176 INIT_LIST_HEAD(&event->list);
4177 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4178 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4179 INIT_WORK(&event->remove, memcg_event_remove);
4187 event->eventfd = eventfd_ctx_fileget(efile.file);
4188 if (IS_ERR(event->eventfd)) {
4189 ret = PTR_ERR(event->eventfd);
4196 goto out_put_eventfd;
4199 /* the process need read permission on control file */
4200 /* AV: shouldn't we check that it's been opened for read instead? */
4201 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4206 * Determine the event callbacks and set them in @event. This used
4207 * to be done via struct cftype but cgroup core no longer knows
4208 * about these events. The following is crude but the whole thing
4209 * is for compatibility anyway.
4211 * DO NOT ADD NEW FILES.
4213 name = cfile.file->f_path.dentry->d_name.name;
4215 if (!strcmp(name, "memory.usage_in_bytes")) {
4216 event->register_event = mem_cgroup_usage_register_event;
4217 event->unregister_event = mem_cgroup_usage_unregister_event;
4218 } else if (!strcmp(name, "memory.oom_control")) {
4219 event->register_event = mem_cgroup_oom_register_event;
4220 event->unregister_event = mem_cgroup_oom_unregister_event;
4221 } else if (!strcmp(name, "memory.pressure_level")) {
4222 event->register_event = vmpressure_register_event;
4223 event->unregister_event = vmpressure_unregister_event;
4224 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4225 event->register_event = memsw_cgroup_usage_register_event;
4226 event->unregister_event = memsw_cgroup_usage_unregister_event;
4233 * Verify @cfile should belong to @css. Also, remaining events are
4234 * automatically removed on cgroup destruction but the removal is
4235 * asynchronous, so take an extra ref on @css.
4237 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4238 &memory_cgrp_subsys);
4240 if (IS_ERR(cfile_css))
4242 if (cfile_css != css) {
4247 ret = event->register_event(memcg, event->eventfd, buf);
4251 efile.file->f_op->poll(efile.file, &event->pt);
4253 spin_lock(&memcg->event_list_lock);
4254 list_add(&event->list, &memcg->event_list);
4255 spin_unlock(&memcg->event_list_lock);
4267 eventfd_ctx_put(event->eventfd);
4276 static struct cftype mem_cgroup_legacy_files[] = {
4278 .name = "usage_in_bytes",
4279 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4280 .read_u64 = mem_cgroup_read_u64,
4283 .name = "max_usage_in_bytes",
4284 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4285 .write = mem_cgroup_reset,
4286 .read_u64 = mem_cgroup_read_u64,
4289 .name = "limit_in_bytes",
4290 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4291 .write = mem_cgroup_write,
4292 .read_u64 = mem_cgroup_read_u64,
4295 .name = "soft_limit_in_bytes",
4296 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4297 .write = mem_cgroup_write,
4298 .read_u64 = mem_cgroup_read_u64,
4302 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4303 .write = mem_cgroup_reset,
4304 .read_u64 = mem_cgroup_read_u64,
4308 .seq_show = memcg_stat_show,
4311 .name = "force_empty",
4312 .write = mem_cgroup_force_empty_write,
4315 .name = "use_hierarchy",
4316 .write_u64 = mem_cgroup_hierarchy_write,
4317 .read_u64 = mem_cgroup_hierarchy_read,
4320 .name = "cgroup.event_control", /* XXX: for compat */
4321 .write = memcg_write_event_control,
4322 .flags = CFTYPE_NO_PREFIX,
4326 .name = "swappiness",
4327 .read_u64 = mem_cgroup_swappiness_read,
4328 .write_u64 = mem_cgroup_swappiness_write,
4331 .name = "move_charge_at_immigrate",
4332 .read_u64 = mem_cgroup_move_charge_read,
4333 .write_u64 = mem_cgroup_move_charge_write,
4336 .name = "oom_control",
4337 .seq_show = mem_cgroup_oom_control_read,
4338 .write_u64 = mem_cgroup_oom_control_write,
4339 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4342 .name = "pressure_level",
4346 .name = "numa_stat",
4347 .seq_show = memcg_numa_stat_show,
4350 #ifdef CONFIG_MEMCG_KMEM
4352 .name = "kmem.limit_in_bytes",
4353 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4354 .write = mem_cgroup_write,
4355 .read_u64 = mem_cgroup_read_u64,
4358 .name = "kmem.usage_in_bytes",
4359 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4360 .read_u64 = mem_cgroup_read_u64,
4363 .name = "kmem.failcnt",
4364 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4365 .write = mem_cgroup_reset,
4366 .read_u64 = mem_cgroup_read_u64,
4369 .name = "kmem.max_usage_in_bytes",
4370 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4371 .write = mem_cgroup_reset,
4372 .read_u64 = mem_cgroup_read_u64,
4374 #ifdef CONFIG_SLABINFO
4376 .name = "kmem.slabinfo",
4377 .seq_start = slab_start,
4378 .seq_next = slab_next,
4379 .seq_stop = slab_stop,
4380 .seq_show = memcg_slab_show,
4384 { }, /* terminate */
4387 #ifdef CONFIG_MEMCG_SWAP
4388 static struct cftype memsw_cgroup_files[] = {
4390 .name = "memsw.usage_in_bytes",
4391 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4392 .read_u64 = mem_cgroup_read_u64,
4395 .name = "memsw.max_usage_in_bytes",
4396 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4397 .write = mem_cgroup_reset,
4398 .read_u64 = mem_cgroup_read_u64,
4401 .name = "memsw.limit_in_bytes",
4402 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4403 .write = mem_cgroup_write,
4404 .read_u64 = mem_cgroup_read_u64,
4407 .name = "memsw.failcnt",
4408 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4409 .write = mem_cgroup_reset,
4410 .read_u64 = mem_cgroup_read_u64,
4412 { }, /* terminate */
4415 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4417 struct mem_cgroup_per_node *pn;
4418 struct mem_cgroup_per_zone *mz;
4419 int zone, tmp = node;
4421 * This routine is called against possible nodes.
4422 * But it's BUG to call kmalloc() against offline node.
4424 * TODO: this routine can waste much memory for nodes which will
4425 * never be onlined. It's better to use memory hotplug callback
4428 if (!node_state(node, N_NORMAL_MEMORY))
4430 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4434 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4435 mz = &pn->zoneinfo[zone];
4436 lruvec_init(&mz->lruvec);
4437 mz->usage_in_excess = 0;
4438 mz->on_tree = false;
4441 memcg->nodeinfo[node] = pn;
4445 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4447 kfree(memcg->nodeinfo[node]);
4450 static struct mem_cgroup *mem_cgroup_alloc(void)
4452 struct mem_cgroup *memcg;
4455 size = sizeof(struct mem_cgroup);
4456 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4458 memcg = kzalloc(size, GFP_KERNEL);
4462 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4465 spin_lock_init(&memcg->pcp_counter_lock);
4474 * At destroying mem_cgroup, references from swap_cgroup can remain.
4475 * (scanning all at force_empty is too costly...)
4477 * Instead of clearing all references at force_empty, we remember
4478 * the number of reference from swap_cgroup and free mem_cgroup when
4479 * it goes down to 0.
4481 * Removal of cgroup itself succeeds regardless of refs from swap.
4484 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4488 mem_cgroup_remove_from_trees(memcg);
4491 free_mem_cgroup_per_zone_info(memcg, node);
4493 free_percpu(memcg->stat);
4495 disarm_static_keys(memcg);
4500 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4502 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4504 if (!memcg->memory.parent)
4506 return mem_cgroup_from_counter(memcg->memory.parent, memory);
4508 EXPORT_SYMBOL(parent_mem_cgroup);
4510 static void __init mem_cgroup_soft_limit_tree_init(void)
4514 for_each_node(node) {
4515 struct mem_cgroup_tree_per_node *rtpn;
4518 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
4519 node_online(node) ? node : NUMA_NO_NODE);
4521 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4522 struct mem_cgroup_tree_per_zone *rtpz;
4524 rtpz = &rtpn->rb_tree_per_zone[zone];
4525 rtpz->rb_root = RB_ROOT;
4526 spin_lock_init(&rtpz->lock);
4528 soft_limit_tree.rb_tree_per_node[node] = rtpn;
4532 static struct cgroup_subsys_state * __ref
4533 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4535 struct mem_cgroup *memcg;
4536 long error = -ENOMEM;
4539 memcg = mem_cgroup_alloc();
4541 return ERR_PTR(error);
4544 if (alloc_mem_cgroup_per_zone_info(memcg, node))
4548 if (parent_css == NULL) {
4549 root_mem_cgroup = memcg;
4550 page_counter_init(&memcg->memory, NULL);
4551 memcg->high = PAGE_COUNTER_MAX;
4552 memcg->soft_limit = PAGE_COUNTER_MAX;
4553 page_counter_init(&memcg->memsw, NULL);
4554 page_counter_init(&memcg->kmem, NULL);
4557 memcg->last_scanned_node = MAX_NUMNODES;
4558 INIT_LIST_HEAD(&memcg->oom_notify);
4559 memcg->move_charge_at_immigrate = 0;
4560 mutex_init(&memcg->thresholds_lock);
4561 spin_lock_init(&memcg->move_lock);
4562 vmpressure_init(&memcg->vmpressure);
4563 INIT_LIST_HEAD(&memcg->event_list);
4564 spin_lock_init(&memcg->event_list_lock);
4565 #ifdef CONFIG_MEMCG_KMEM
4566 memcg->kmemcg_id = -1;
4572 __mem_cgroup_free(memcg);
4573 return ERR_PTR(error);
4577 mem_cgroup_css_online(struct cgroup_subsys_state *css)
4579 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4580 struct mem_cgroup *parent = mem_cgroup_from_css(css->parent);
4583 if (css->id > MEM_CGROUP_ID_MAX)
4589 mutex_lock(&memcg_create_mutex);
4591 memcg->use_hierarchy = parent->use_hierarchy;
4592 memcg->oom_kill_disable = parent->oom_kill_disable;
4593 memcg->swappiness = mem_cgroup_swappiness(parent);
4595 if (parent->use_hierarchy) {
4596 page_counter_init(&memcg->memory, &parent->memory);
4597 memcg->high = PAGE_COUNTER_MAX;
4598 memcg->soft_limit = PAGE_COUNTER_MAX;
4599 page_counter_init(&memcg->memsw, &parent->memsw);
4600 page_counter_init(&memcg->kmem, &parent->kmem);
4603 * No need to take a reference to the parent because cgroup
4604 * core guarantees its existence.
4607 page_counter_init(&memcg->memory, NULL);
4608 memcg->high = PAGE_COUNTER_MAX;
4609 memcg->soft_limit = PAGE_COUNTER_MAX;
4610 page_counter_init(&memcg->memsw, NULL);
4611 page_counter_init(&memcg->kmem, NULL);
4613 * Deeper hierachy with use_hierarchy == false doesn't make
4614 * much sense so let cgroup subsystem know about this
4615 * unfortunate state in our controller.
4617 if (parent != root_mem_cgroup)
4618 memory_cgrp_subsys.broken_hierarchy = true;
4620 mutex_unlock(&memcg_create_mutex);
4622 ret = memcg_init_kmem(memcg, &memory_cgrp_subsys);
4627 * Make sure the memcg is initialized: mem_cgroup_iter()
4628 * orders reading memcg->initialized against its callers
4629 * reading the memcg members.
4631 smp_store_release(&memcg->initialized, 1);
4636 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4638 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4639 struct mem_cgroup_event *event, *tmp;
4642 * Unregister events and notify userspace.
4643 * Notify userspace about cgroup removing only after rmdir of cgroup
4644 * directory to avoid race between userspace and kernelspace.
4646 spin_lock(&memcg->event_list_lock);
4647 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4648 list_del_init(&event->list);
4649 schedule_work(&event->remove);
4651 spin_unlock(&memcg->event_list_lock);
4653 vmpressure_cleanup(&memcg->vmpressure);
4656 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4658 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4660 memcg_destroy_kmem(memcg);
4661 __mem_cgroup_free(memcg);
4665 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4666 * @css: the target css
4668 * Reset the states of the mem_cgroup associated with @css. This is
4669 * invoked when the userland requests disabling on the default hierarchy
4670 * but the memcg is pinned through dependency. The memcg should stop
4671 * applying policies and should revert to the vanilla state as it may be
4672 * made visible again.
4674 * The current implementation only resets the essential configurations.
4675 * This needs to be expanded to cover all the visible parts.
4677 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4679 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4681 mem_cgroup_resize_limit(memcg, PAGE_COUNTER_MAX);
4682 mem_cgroup_resize_memsw_limit(memcg, PAGE_COUNTER_MAX);
4683 memcg_update_kmem_limit(memcg, PAGE_COUNTER_MAX);
4685 memcg->high = PAGE_COUNTER_MAX;
4686 memcg->soft_limit = PAGE_COUNTER_MAX;
4690 /* Handlers for move charge at task migration. */
4691 static int mem_cgroup_do_precharge(unsigned long count)
4695 /* Try a single bulk charge without reclaim first */
4696 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_WAIT, count);
4698 mc.precharge += count;
4701 if (ret == -EINTR) {
4702 cancel_charge(root_mem_cgroup, count);
4706 /* Try charges one by one with reclaim */
4708 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_NORETRY, 1);
4710 * In case of failure, any residual charges against
4711 * mc.to will be dropped by mem_cgroup_clear_mc()
4712 * later on. However, cancel any charges that are
4713 * bypassed to root right away or they'll be lost.
4716 cancel_charge(root_mem_cgroup, 1);
4726 * get_mctgt_type - get target type of moving charge
4727 * @vma: the vma the pte to be checked belongs
4728 * @addr: the address corresponding to the pte to be checked
4729 * @ptent: the pte to be checked
4730 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4733 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4734 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4735 * move charge. if @target is not NULL, the page is stored in target->page
4736 * with extra refcnt got(Callers should handle it).
4737 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4738 * target for charge migration. if @target is not NULL, the entry is stored
4741 * Called with pte lock held.
4748 enum mc_target_type {
4754 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4755 unsigned long addr, pte_t ptent)
4757 struct page *page = vm_normal_page(vma, addr, ptent);
4759 if (!page || !page_mapped(page))
4761 if (PageAnon(page)) {
4762 if (!(mc.flags & MOVE_ANON))
4765 if (!(mc.flags & MOVE_FILE))
4768 if (!get_page_unless_zero(page))
4775 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4776 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4778 struct page *page = NULL;
4779 swp_entry_t ent = pte_to_swp_entry(ptent);
4781 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
4784 * Because lookup_swap_cache() updates some statistics counter,
4785 * we call find_get_page() with swapper_space directly.
4787 page = find_get_page(swap_address_space(ent), ent.val);
4788 if (do_swap_account)
4789 entry->val = ent.val;
4794 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4795 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4801 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4802 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4804 struct page *page = NULL;
4805 struct address_space *mapping;
4808 if (!vma->vm_file) /* anonymous vma */
4810 if (!(mc.flags & MOVE_FILE))
4813 mapping = vma->vm_file->f_mapping;
4814 pgoff = linear_page_index(vma, addr);
4816 /* page is moved even if it's not RSS of this task(page-faulted). */
4818 /* shmem/tmpfs may report page out on swap: account for that too. */
4819 if (shmem_mapping(mapping)) {
4820 page = find_get_entry(mapping, pgoff);
4821 if (radix_tree_exceptional_entry(page)) {
4822 swp_entry_t swp = radix_to_swp_entry(page);
4823 if (do_swap_account)
4825 page = find_get_page(swap_address_space(swp), swp.val);
4828 page = find_get_page(mapping, pgoff);
4830 page = find_get_page(mapping, pgoff);
4835 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
4836 unsigned long addr, pte_t ptent, union mc_target *target)
4838 struct page *page = NULL;
4839 enum mc_target_type ret = MC_TARGET_NONE;
4840 swp_entry_t ent = { .val = 0 };
4842 if (pte_present(ptent))
4843 page = mc_handle_present_pte(vma, addr, ptent);
4844 else if (is_swap_pte(ptent))
4845 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
4846 else if (pte_none(ptent))
4847 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4849 if (!page && !ent.val)
4853 * Do only loose check w/o serialization.
4854 * mem_cgroup_move_account() checks the page is valid or
4855 * not under LRU exclusion.
4857 if (page->mem_cgroup == mc.from) {
4858 ret = MC_TARGET_PAGE;
4860 target->page = page;
4862 if (!ret || !target)
4865 /* There is a swap entry and a page doesn't exist or isn't charged */
4866 if (ent.val && !ret &&
4867 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
4868 ret = MC_TARGET_SWAP;
4875 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4877 * We don't consider swapping or file mapped pages because THP does not
4878 * support them for now.
4879 * Caller should make sure that pmd_trans_huge(pmd) is true.
4881 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4882 unsigned long addr, pmd_t pmd, union mc_target *target)
4884 struct page *page = NULL;
4885 enum mc_target_type ret = MC_TARGET_NONE;
4887 page = pmd_page(pmd);
4888 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
4889 if (!(mc.flags & MOVE_ANON))
4891 if (page->mem_cgroup == mc.from) {
4892 ret = MC_TARGET_PAGE;
4895 target->page = page;
4901 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4902 unsigned long addr, pmd_t pmd, union mc_target *target)
4904 return MC_TARGET_NONE;
4908 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4909 unsigned long addr, unsigned long end,
4910 struct mm_walk *walk)
4912 struct vm_area_struct *vma = walk->private;
4916 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
4917 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
4918 mc.precharge += HPAGE_PMD_NR;
4923 if (pmd_trans_unstable(pmd))
4925 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4926 for (; addr != end; pte++, addr += PAGE_SIZE)
4927 if (get_mctgt_type(vma, addr, *pte, NULL))
4928 mc.precharge++; /* increment precharge temporarily */
4929 pte_unmap_unlock(pte - 1, ptl);
4935 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4937 unsigned long precharge;
4938 struct vm_area_struct *vma;
4940 down_read(&mm->mmap_sem);
4941 for (vma = mm->mmap; vma; vma = vma->vm_next) {
4942 struct mm_walk mem_cgroup_count_precharge_walk = {
4943 .pmd_entry = mem_cgroup_count_precharge_pte_range,
4947 if (is_vm_hugetlb_page(vma))
4949 walk_page_range(vma->vm_start, vma->vm_end,
4950 &mem_cgroup_count_precharge_walk);
4952 up_read(&mm->mmap_sem);
4954 precharge = mc.precharge;
4960 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4962 unsigned long precharge = mem_cgroup_count_precharge(mm);
4964 VM_BUG_ON(mc.moving_task);
4965 mc.moving_task = current;
4966 return mem_cgroup_do_precharge(precharge);
4969 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4970 static void __mem_cgroup_clear_mc(void)
4972 struct mem_cgroup *from = mc.from;
4973 struct mem_cgroup *to = mc.to;
4975 /* we must uncharge all the leftover precharges from mc.to */
4977 cancel_charge(mc.to, mc.precharge);
4981 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4982 * we must uncharge here.
4984 if (mc.moved_charge) {
4985 cancel_charge(mc.from, mc.moved_charge);
4986 mc.moved_charge = 0;
4988 /* we must fixup refcnts and charges */
4989 if (mc.moved_swap) {
4990 /* uncharge swap account from the old cgroup */
4991 if (!mem_cgroup_is_root(mc.from))
4992 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
4995 * we charged both to->memory and to->memsw, so we
4996 * should uncharge to->memory.
4998 if (!mem_cgroup_is_root(mc.to))
4999 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5001 css_put_many(&mc.from->css, mc.moved_swap);
5003 /* we've already done css_get(mc.to) */
5006 memcg_oom_recover(from);
5007 memcg_oom_recover(to);
5008 wake_up_all(&mc.waitq);
5011 static void mem_cgroup_clear_mc(void)
5014 * we must clear moving_task before waking up waiters at the end of
5017 mc.moving_task = NULL;
5018 __mem_cgroup_clear_mc();
5019 spin_lock(&mc.lock);
5022 spin_unlock(&mc.lock);
5025 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
5026 struct cgroup_taskset *tset)
5028 struct task_struct *p = cgroup_taskset_first(tset);
5030 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5031 unsigned long move_flags;
5034 * We are now commited to this value whatever it is. Changes in this
5035 * tunable will only affect upcoming migrations, not the current one.
5036 * So we need to save it, and keep it going.
5038 move_flags = ACCESS_ONCE(memcg->move_charge_at_immigrate);
5040 struct mm_struct *mm;
5041 struct mem_cgroup *from = mem_cgroup_from_task(p);
5043 VM_BUG_ON(from == memcg);
5045 mm = get_task_mm(p);
5048 /* We move charges only when we move a owner of the mm */
5049 if (mm->owner == p) {
5052 VM_BUG_ON(mc.precharge);
5053 VM_BUG_ON(mc.moved_charge);
5054 VM_BUG_ON(mc.moved_swap);
5056 spin_lock(&mc.lock);
5059 mc.flags = move_flags;
5060 spin_unlock(&mc.lock);
5061 /* We set mc.moving_task later */
5063 ret = mem_cgroup_precharge_mc(mm);
5065 mem_cgroup_clear_mc();
5072 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
5073 struct cgroup_taskset *tset)
5076 mem_cgroup_clear_mc();
5079 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5080 unsigned long addr, unsigned long end,
5081 struct mm_walk *walk)
5084 struct vm_area_struct *vma = walk->private;
5087 enum mc_target_type target_type;
5088 union mc_target target;
5092 * We don't take compound_lock() here but no race with splitting thp
5094 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
5095 * under splitting, which means there's no concurrent thp split,
5096 * - if another thread runs into split_huge_page() just after we
5097 * entered this if-block, the thread must wait for page table lock
5098 * to be unlocked in __split_huge_page_splitting(), where the main
5099 * part of thp split is not executed yet.
5101 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
5102 if (mc.precharge < HPAGE_PMD_NR) {
5106 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5107 if (target_type == MC_TARGET_PAGE) {
5109 if (!isolate_lru_page(page)) {
5110 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
5112 mc.precharge -= HPAGE_PMD_NR;
5113 mc.moved_charge += HPAGE_PMD_NR;
5115 putback_lru_page(page);
5123 if (pmd_trans_unstable(pmd))
5126 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5127 for (; addr != end; addr += PAGE_SIZE) {
5128 pte_t ptent = *(pte++);
5134 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5135 case MC_TARGET_PAGE:
5137 if (isolate_lru_page(page))
5139 if (!mem_cgroup_move_account(page, 1, mc.from, mc.to)) {
5141 /* we uncharge from mc.from later. */
5144 putback_lru_page(page);
5145 put: /* get_mctgt_type() gets the page */
5148 case MC_TARGET_SWAP:
5150 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5152 /* we fixup refcnts and charges later. */
5160 pte_unmap_unlock(pte - 1, ptl);
5165 * We have consumed all precharges we got in can_attach().
5166 * We try charge one by one, but don't do any additional
5167 * charges to mc.to if we have failed in charge once in attach()
5170 ret = mem_cgroup_do_precharge(1);
5178 static void mem_cgroup_move_charge(struct mm_struct *mm)
5180 struct vm_area_struct *vma;
5182 lru_add_drain_all();
5184 * Signal mem_cgroup_begin_page_stat() to take the memcg's
5185 * move_lock while we're moving its pages to another memcg.
5186 * Then wait for already started RCU-only updates to finish.
5188 atomic_inc(&mc.from->moving_account);
5191 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5193 * Someone who are holding the mmap_sem might be waiting in
5194 * waitq. So we cancel all extra charges, wake up all waiters,
5195 * and retry. Because we cancel precharges, we might not be able
5196 * to move enough charges, but moving charge is a best-effort
5197 * feature anyway, so it wouldn't be a big problem.
5199 __mem_cgroup_clear_mc();
5203 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5205 struct mm_walk mem_cgroup_move_charge_walk = {
5206 .pmd_entry = mem_cgroup_move_charge_pte_range,
5210 if (is_vm_hugetlb_page(vma))
5212 ret = walk_page_range(vma->vm_start, vma->vm_end,
5213 &mem_cgroup_move_charge_walk);
5216 * means we have consumed all precharges and failed in
5217 * doing additional charge. Just abandon here.
5221 up_read(&mm->mmap_sem);
5222 atomic_dec(&mc.from->moving_account);
5225 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
5226 struct cgroup_taskset *tset)
5228 struct task_struct *p = cgroup_taskset_first(tset);
5229 struct mm_struct *mm = get_task_mm(p);
5233 mem_cgroup_move_charge(mm);
5237 mem_cgroup_clear_mc();
5239 #else /* !CONFIG_MMU */
5240 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
5241 struct cgroup_taskset *tset)
5245 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
5246 struct cgroup_taskset *tset)
5249 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
5250 struct cgroup_taskset *tset)
5256 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5257 * to verify whether we're attached to the default hierarchy on each mount
5260 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5263 * use_hierarchy is forced on the default hierarchy. cgroup core
5264 * guarantees that @root doesn't have any children, so turning it
5265 * on for the root memcg is enough.
5267 if (cgroup_on_dfl(root_css->cgroup))
5268 mem_cgroup_from_css(root_css)->use_hierarchy = true;
5271 static u64 memory_current_read(struct cgroup_subsys_state *css,
5274 return mem_cgroup_usage(mem_cgroup_from_css(css), false);
5277 static int memory_low_show(struct seq_file *m, void *v)
5279 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5280 unsigned long low = ACCESS_ONCE(memcg->low);
5282 if (low == PAGE_COUNTER_MAX)
5283 seq_puts(m, "infinity\n");
5285 seq_printf(m, "%llu\n", (u64)low * PAGE_SIZE);
5290 static ssize_t memory_low_write(struct kernfs_open_file *of,
5291 char *buf, size_t nbytes, loff_t off)
5293 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5297 buf = strstrip(buf);
5298 err = page_counter_memparse(buf, "infinity", &low);
5307 static int memory_high_show(struct seq_file *m, void *v)
5309 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5310 unsigned long high = ACCESS_ONCE(memcg->high);
5312 if (high == PAGE_COUNTER_MAX)
5313 seq_puts(m, "infinity\n");
5315 seq_printf(m, "%llu\n", (u64)high * PAGE_SIZE);
5320 static ssize_t memory_high_write(struct kernfs_open_file *of,
5321 char *buf, size_t nbytes, loff_t off)
5323 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5327 buf = strstrip(buf);
5328 err = page_counter_memparse(buf, "infinity", &high);
5337 static int memory_max_show(struct seq_file *m, void *v)
5339 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5340 unsigned long max = ACCESS_ONCE(memcg->memory.limit);
5342 if (max == PAGE_COUNTER_MAX)
5343 seq_puts(m, "infinity\n");
5345 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
5350 static ssize_t memory_max_write(struct kernfs_open_file *of,
5351 char *buf, size_t nbytes, loff_t off)
5353 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5357 buf = strstrip(buf);
5358 err = page_counter_memparse(buf, "infinity", &max);
5362 err = mem_cgroup_resize_limit(memcg, max);
5369 static int memory_events_show(struct seq_file *m, void *v)
5371 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5373 seq_printf(m, "low %lu\n", mem_cgroup_read_events(memcg, MEMCG_LOW));
5374 seq_printf(m, "high %lu\n", mem_cgroup_read_events(memcg, MEMCG_HIGH));
5375 seq_printf(m, "max %lu\n", mem_cgroup_read_events(memcg, MEMCG_MAX));
5376 seq_printf(m, "oom %lu\n", mem_cgroup_read_events(memcg, MEMCG_OOM));
5381 static struct cftype memory_files[] = {
5384 .read_u64 = memory_current_read,
5388 .flags = CFTYPE_NOT_ON_ROOT,
5389 .seq_show = memory_low_show,
5390 .write = memory_low_write,
5394 .flags = CFTYPE_NOT_ON_ROOT,
5395 .seq_show = memory_high_show,
5396 .write = memory_high_write,
5400 .flags = CFTYPE_NOT_ON_ROOT,
5401 .seq_show = memory_max_show,
5402 .write = memory_max_write,
5406 .flags = CFTYPE_NOT_ON_ROOT,
5407 .seq_show = memory_events_show,
5412 struct cgroup_subsys memory_cgrp_subsys = {
5413 .css_alloc = mem_cgroup_css_alloc,
5414 .css_online = mem_cgroup_css_online,
5415 .css_offline = mem_cgroup_css_offline,
5416 .css_free = mem_cgroup_css_free,
5417 .css_reset = mem_cgroup_css_reset,
5418 .can_attach = mem_cgroup_can_attach,
5419 .cancel_attach = mem_cgroup_cancel_attach,
5420 .attach = mem_cgroup_move_task,
5421 .bind = mem_cgroup_bind,
5422 .dfl_cftypes = memory_files,
5423 .legacy_cftypes = mem_cgroup_legacy_files,
5427 #ifdef CONFIG_MEMCG_SWAP
5428 static int __init enable_swap_account(char *s)
5430 if (!strcmp(s, "1"))
5431 really_do_swap_account = 1;
5432 else if (!strcmp(s, "0"))
5433 really_do_swap_account = 0;
5436 __setup("swapaccount=", enable_swap_account);
5438 static void __init memsw_file_init(void)
5440 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
5441 memsw_cgroup_files));
5444 static void __init enable_swap_cgroup(void)
5446 if (!mem_cgroup_disabled() && really_do_swap_account) {
5447 do_swap_account = 1;
5453 static void __init enable_swap_cgroup(void)
5459 * mem_cgroup_events - count memory events against a cgroup
5460 * @memcg: the memory cgroup
5461 * @idx: the event index
5462 * @nr: the number of events to account for
5464 void mem_cgroup_events(struct mem_cgroup *memcg,
5465 enum mem_cgroup_events_index idx,
5468 this_cpu_add(memcg->stat->events[idx], nr);
5472 * mem_cgroup_low - check if memory consumption is below the normal range
5473 * @root: the highest ancestor to consider
5474 * @memcg: the memory cgroup to check
5476 * Returns %true if memory consumption of @memcg, and that of all
5477 * configurable ancestors up to @root, is below the normal range.
5479 bool mem_cgroup_low(struct mem_cgroup *root, struct mem_cgroup *memcg)
5481 if (mem_cgroup_disabled())
5485 * The toplevel group doesn't have a configurable range, so
5486 * it's never low when looked at directly, and it is not
5487 * considered an ancestor when assessing the hierarchy.
5490 if (memcg == root_mem_cgroup)
5493 if (page_counter_read(&memcg->memory) > memcg->low)
5496 while (memcg != root) {
5497 memcg = parent_mem_cgroup(memcg);
5499 if (memcg == root_mem_cgroup)
5502 if (page_counter_read(&memcg->memory) > memcg->low)
5508 #ifdef CONFIG_MEMCG_SWAP
5510 * mem_cgroup_swapout - transfer a memsw charge to swap
5511 * @page: page whose memsw charge to transfer
5512 * @entry: swap entry to move the charge to
5514 * Transfer the memsw charge of @page to @entry.
5516 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
5518 struct mem_cgroup *memcg;
5519 unsigned short oldid;
5521 VM_BUG_ON_PAGE(PageLRU(page), page);
5522 VM_BUG_ON_PAGE(page_count(page), page);
5524 if (!do_swap_account)
5527 memcg = page->mem_cgroup;
5529 /* Readahead page, never charged */
5533 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg));
5534 VM_BUG_ON_PAGE(oldid, page);
5535 mem_cgroup_swap_statistics(memcg, true);
5537 page->mem_cgroup = NULL;
5539 if (!mem_cgroup_is_root(memcg))
5540 page_counter_uncharge(&memcg->memory, 1);
5542 /* XXX: caller holds IRQ-safe mapping->tree_lock */
5543 VM_BUG_ON(!irqs_disabled());
5545 mem_cgroup_charge_statistics(memcg, page, -1);
5546 memcg_check_events(memcg, page);
5550 * mem_cgroup_uncharge_swap - uncharge a swap entry
5551 * @entry: swap entry to uncharge
5553 * Drop the memsw charge associated with @entry.
5555 void mem_cgroup_uncharge_swap(swp_entry_t entry)
5557 struct mem_cgroup *memcg;
5560 if (!do_swap_account)
5563 id = swap_cgroup_record(entry, 0);
5565 memcg = mem_cgroup_lookup(id);
5567 if (!mem_cgroup_is_root(memcg))
5568 page_counter_uncharge(&memcg->memsw, 1);
5569 mem_cgroup_swap_statistics(memcg, false);
5570 css_put(&memcg->css);
5577 * mem_cgroup_try_charge - try charging a page
5578 * @page: page to charge
5579 * @mm: mm context of the victim
5580 * @gfp_mask: reclaim mode
5581 * @memcgp: charged memcg return
5583 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5584 * pages according to @gfp_mask if necessary.
5586 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5587 * Otherwise, an error code is returned.
5589 * After page->mapping has been set up, the caller must finalize the
5590 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5591 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5593 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
5594 gfp_t gfp_mask, struct mem_cgroup **memcgp)
5596 struct mem_cgroup *memcg = NULL;
5597 unsigned int nr_pages = 1;
5600 if (mem_cgroup_disabled())
5603 if (PageSwapCache(page)) {
5605 * Every swap fault against a single page tries to charge the
5606 * page, bail as early as possible. shmem_unuse() encounters
5607 * already charged pages, too. The USED bit is protected by
5608 * the page lock, which serializes swap cache removal, which
5609 * in turn serializes uncharging.
5611 if (page->mem_cgroup)
5615 if (PageTransHuge(page)) {
5616 nr_pages <<= compound_order(page);
5617 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5620 if (do_swap_account && PageSwapCache(page))
5621 memcg = try_get_mem_cgroup_from_page(page);
5623 memcg = get_mem_cgroup_from_mm(mm);
5625 ret = try_charge(memcg, gfp_mask, nr_pages);
5627 css_put(&memcg->css);
5629 if (ret == -EINTR) {
5630 memcg = root_mem_cgroup;
5639 * mem_cgroup_commit_charge - commit a page charge
5640 * @page: page to charge
5641 * @memcg: memcg to charge the page to
5642 * @lrucare: page might be on LRU already
5644 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5645 * after page->mapping has been set up. This must happen atomically
5646 * as part of the page instantiation, i.e. under the page table lock
5647 * for anonymous pages, under the page lock for page and swap cache.
5649 * In addition, the page must not be on the LRU during the commit, to
5650 * prevent racing with task migration. If it might be, use @lrucare.
5652 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5654 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
5657 unsigned int nr_pages = 1;
5659 VM_BUG_ON_PAGE(!page->mapping, page);
5660 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
5662 if (mem_cgroup_disabled())
5665 * Swap faults will attempt to charge the same page multiple
5666 * times. But reuse_swap_page() might have removed the page
5667 * from swapcache already, so we can't check PageSwapCache().
5672 commit_charge(page, memcg, lrucare);
5674 if (PageTransHuge(page)) {
5675 nr_pages <<= compound_order(page);
5676 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5679 local_irq_disable();
5680 mem_cgroup_charge_statistics(memcg, page, nr_pages);
5681 memcg_check_events(memcg, page);
5684 if (do_swap_account && PageSwapCache(page)) {
5685 swp_entry_t entry = { .val = page_private(page) };
5687 * The swap entry might not get freed for a long time,
5688 * let's not wait for it. The page already received a
5689 * memory+swap charge, drop the swap entry duplicate.
5691 mem_cgroup_uncharge_swap(entry);
5696 * mem_cgroup_cancel_charge - cancel a page charge
5697 * @page: page to charge
5698 * @memcg: memcg to charge the page to
5700 * Cancel a charge transaction started by mem_cgroup_try_charge().
5702 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg)
5704 unsigned int nr_pages = 1;
5706 if (mem_cgroup_disabled())
5709 * Swap faults will attempt to charge the same page multiple
5710 * times. But reuse_swap_page() might have removed the page
5711 * from swapcache already, so we can't check PageSwapCache().
5716 if (PageTransHuge(page)) {
5717 nr_pages <<= compound_order(page);
5718 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5721 cancel_charge(memcg, nr_pages);
5724 static void uncharge_batch(struct mem_cgroup *memcg, unsigned long pgpgout,
5725 unsigned long nr_anon, unsigned long nr_file,
5726 unsigned long nr_huge, struct page *dummy_page)
5728 unsigned long nr_pages = nr_anon + nr_file;
5729 unsigned long flags;
5731 if (!mem_cgroup_is_root(memcg)) {
5732 page_counter_uncharge(&memcg->memory, nr_pages);
5733 if (do_swap_account)
5734 page_counter_uncharge(&memcg->memsw, nr_pages);
5735 memcg_oom_recover(memcg);
5738 local_irq_save(flags);
5739 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS], nr_anon);
5740 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_CACHE], nr_file);
5741 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE], nr_huge);
5742 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT], pgpgout);
5743 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
5744 memcg_check_events(memcg, dummy_page);
5745 local_irq_restore(flags);
5747 if (!mem_cgroup_is_root(memcg))
5748 css_put_many(&memcg->css, nr_pages);
5751 static void uncharge_list(struct list_head *page_list)
5753 struct mem_cgroup *memcg = NULL;
5754 unsigned long nr_anon = 0;
5755 unsigned long nr_file = 0;
5756 unsigned long nr_huge = 0;
5757 unsigned long pgpgout = 0;
5758 struct list_head *next;
5761 next = page_list->next;
5763 unsigned int nr_pages = 1;
5765 page = list_entry(next, struct page, lru);
5766 next = page->lru.next;
5768 VM_BUG_ON_PAGE(PageLRU(page), page);
5769 VM_BUG_ON_PAGE(page_count(page), page);
5771 if (!page->mem_cgroup)
5775 * Nobody should be changing or seriously looking at
5776 * page->mem_cgroup at this point, we have fully
5777 * exclusive access to the page.
5780 if (memcg != page->mem_cgroup) {
5782 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5784 pgpgout = nr_anon = nr_file = nr_huge = 0;
5786 memcg = page->mem_cgroup;
5789 if (PageTransHuge(page)) {
5790 nr_pages <<= compound_order(page);
5791 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5792 nr_huge += nr_pages;
5796 nr_anon += nr_pages;
5798 nr_file += nr_pages;
5800 page->mem_cgroup = NULL;
5803 } while (next != page_list);
5806 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5811 * mem_cgroup_uncharge - uncharge a page
5812 * @page: page to uncharge
5814 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5815 * mem_cgroup_commit_charge().
5817 void mem_cgroup_uncharge(struct page *page)
5819 if (mem_cgroup_disabled())
5822 /* Don't touch page->lru of any random page, pre-check: */
5823 if (!page->mem_cgroup)
5826 INIT_LIST_HEAD(&page->lru);
5827 uncharge_list(&page->lru);
5831 * mem_cgroup_uncharge_list - uncharge a list of page
5832 * @page_list: list of pages to uncharge
5834 * Uncharge a list of pages previously charged with
5835 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5837 void mem_cgroup_uncharge_list(struct list_head *page_list)
5839 if (mem_cgroup_disabled())
5842 if (!list_empty(page_list))
5843 uncharge_list(page_list);
5847 * mem_cgroup_migrate - migrate a charge to another page
5848 * @oldpage: currently charged page
5849 * @newpage: page to transfer the charge to
5850 * @lrucare: either or both pages might be on the LRU already
5852 * Migrate the charge from @oldpage to @newpage.
5854 * Both pages must be locked, @newpage->mapping must be set up.
5856 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage,
5859 struct mem_cgroup *memcg;
5862 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
5863 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
5864 VM_BUG_ON_PAGE(!lrucare && PageLRU(oldpage), oldpage);
5865 VM_BUG_ON_PAGE(!lrucare && PageLRU(newpage), newpage);
5866 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
5867 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
5870 if (mem_cgroup_disabled())
5873 /* Page cache replacement: new page already charged? */
5874 if (newpage->mem_cgroup)
5878 * Swapcache readahead pages can get migrated before being
5879 * charged, and migration from compaction can happen to an
5880 * uncharged page when the PFN walker finds a page that
5881 * reclaim just put back on the LRU but has not released yet.
5883 memcg = oldpage->mem_cgroup;
5888 lock_page_lru(oldpage, &isolated);
5890 oldpage->mem_cgroup = NULL;
5893 unlock_page_lru(oldpage, isolated);
5895 commit_charge(newpage, memcg, lrucare);
5899 * subsys_initcall() for memory controller.
5901 * Some parts like hotcpu_notifier() have to be initialized from this context
5902 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5903 * everything that doesn't depend on a specific mem_cgroup structure should
5904 * be initialized from here.
5906 static int __init mem_cgroup_init(void)
5908 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5909 enable_swap_cgroup();
5910 mem_cgroup_soft_limit_tree_init();
5914 subsys_initcall(mem_cgroup_init);