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/page_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 enum mem_cgroup_events_index {
101 MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */
102 MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */
103 MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */
104 MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */
105 MEM_CGROUP_EVENTS_NSTATS,
108 static const char * const mem_cgroup_events_names[] = {
115 static const char * const mem_cgroup_lru_names[] = {
124 * Per memcg event counter is incremented at every pagein/pageout. With THP,
125 * it will be incremated by the number of pages. This counter is used for
126 * for trigger some periodic events. This is straightforward and better
127 * than using jiffies etc. to handle periodic memcg event.
129 enum mem_cgroup_events_target {
130 MEM_CGROUP_TARGET_THRESH,
131 MEM_CGROUP_TARGET_SOFTLIMIT,
132 MEM_CGROUP_TARGET_NUMAINFO,
135 #define THRESHOLDS_EVENTS_TARGET 128
136 #define SOFTLIMIT_EVENTS_TARGET 1024
137 #define NUMAINFO_EVENTS_TARGET 1024
139 struct mem_cgroup_stat_cpu {
140 long count[MEM_CGROUP_STAT_NSTATS];
141 unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
142 unsigned long nr_page_events;
143 unsigned long targets[MEM_CGROUP_NTARGETS];
146 struct reclaim_iter {
147 struct mem_cgroup *position;
148 /* scan generation, increased every round-trip */
149 unsigned int generation;
153 * per-zone information in memory controller.
155 struct mem_cgroup_per_zone {
156 struct lruvec lruvec;
157 unsigned long lru_size[NR_LRU_LISTS];
159 struct reclaim_iter iter[DEF_PRIORITY + 1];
161 struct rb_node tree_node; /* RB tree node */
162 unsigned long usage_in_excess;/* Set to the value by which */
163 /* the soft limit is exceeded*/
165 struct mem_cgroup *memcg; /* Back pointer, we cannot */
166 /* use container_of */
169 struct mem_cgroup_per_node {
170 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
174 * Cgroups above their limits are maintained in a RB-Tree, independent of
175 * their hierarchy representation
178 struct mem_cgroup_tree_per_zone {
179 struct rb_root rb_root;
183 struct mem_cgroup_tree_per_node {
184 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
187 struct mem_cgroup_tree {
188 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
191 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
193 struct mem_cgroup_threshold {
194 struct eventfd_ctx *eventfd;
195 unsigned long threshold;
199 struct mem_cgroup_threshold_ary {
200 /* An array index points to threshold just below or equal to usage. */
201 int current_threshold;
202 /* Size of entries[] */
204 /* Array of thresholds */
205 struct mem_cgroup_threshold entries[0];
208 struct mem_cgroup_thresholds {
209 /* Primary thresholds array */
210 struct mem_cgroup_threshold_ary *primary;
212 * Spare threshold array.
213 * This is needed to make mem_cgroup_unregister_event() "never fail".
214 * It must be able to store at least primary->size - 1 entries.
216 struct mem_cgroup_threshold_ary *spare;
220 struct mem_cgroup_eventfd_list {
221 struct list_head list;
222 struct eventfd_ctx *eventfd;
226 * cgroup_event represents events which userspace want to receive.
228 struct mem_cgroup_event {
230 * memcg which the event belongs to.
232 struct mem_cgroup *memcg;
234 * eventfd to signal userspace about the event.
236 struct eventfd_ctx *eventfd;
238 * Each of these stored in a list by the cgroup.
240 struct list_head list;
242 * register_event() callback will be used to add new userspace
243 * waiter for changes related to this event. Use eventfd_signal()
244 * on eventfd to send notification to userspace.
246 int (*register_event)(struct mem_cgroup *memcg,
247 struct eventfd_ctx *eventfd, const char *args);
249 * unregister_event() callback will be called when userspace closes
250 * the eventfd or on cgroup removing. This callback must be set,
251 * if you want provide notification functionality.
253 void (*unregister_event)(struct mem_cgroup *memcg,
254 struct eventfd_ctx *eventfd);
256 * All fields below needed to unregister event when
257 * userspace closes eventfd.
260 wait_queue_head_t *wqh;
262 struct work_struct remove;
265 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
266 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
269 * The memory controller data structure. The memory controller controls both
270 * page cache and RSS per cgroup. We would eventually like to provide
271 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
272 * to help the administrator determine what knobs to tune.
274 * TODO: Add a water mark for the memory controller. Reclaim will begin when
275 * we hit the water mark. May be even add a low water mark, such that
276 * no reclaim occurs from a cgroup at it's low water mark, this is
277 * a feature that will be implemented much later in the future.
280 struct cgroup_subsys_state css;
282 /* Accounted resources */
283 struct page_counter memory;
284 struct page_counter memsw;
285 struct page_counter kmem;
287 unsigned long soft_limit;
289 /* vmpressure notifications */
290 struct vmpressure vmpressure;
292 /* css_online() has been completed */
296 * Should the accounting and control be hierarchical, per subtree?
299 unsigned long kmem_account_flags; /* See KMEM_ACCOUNTED_*, below */
303 atomic_t oom_wakeups;
306 /* OOM-Killer disable */
307 int oom_kill_disable;
309 /* protect arrays of thresholds */
310 struct mutex thresholds_lock;
312 /* thresholds for memory usage. RCU-protected */
313 struct mem_cgroup_thresholds thresholds;
315 /* thresholds for mem+swap usage. RCU-protected */
316 struct mem_cgroup_thresholds memsw_thresholds;
318 /* For oom notifier event fd */
319 struct list_head oom_notify;
322 * Should we move charges of a task when a task is moved into this
323 * mem_cgroup ? And what type of charges should we move ?
325 unsigned long move_charge_at_immigrate;
327 * set > 0 if pages under this cgroup are moving to other cgroup.
329 atomic_t moving_account;
330 /* taken only while moving_account > 0 */
331 spinlock_t move_lock;
335 struct mem_cgroup_stat_cpu __percpu *stat;
337 * used when a cpu is offlined or other synchronizations
338 * See mem_cgroup_read_stat().
340 struct mem_cgroup_stat_cpu nocpu_base;
341 spinlock_t pcp_counter_lock;
343 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
344 struct cg_proto tcp_mem;
346 #if defined(CONFIG_MEMCG_KMEM)
347 /* analogous to slab_common's slab_caches list, but per-memcg;
348 * protected by memcg_slab_mutex */
349 struct list_head memcg_slab_caches;
350 /* Index in the kmem_cache->memcg_params->memcg_caches array */
354 int last_scanned_node;
356 nodemask_t scan_nodes;
357 atomic_t numainfo_events;
358 atomic_t numainfo_updating;
361 /* List of events which userspace want to receive */
362 struct list_head event_list;
363 spinlock_t event_list_lock;
365 struct mem_cgroup_per_node *nodeinfo[0];
366 /* WARNING: nodeinfo must be the last member here */
369 /* internal only representation about the status of kmem accounting. */
371 KMEM_ACCOUNTED_ACTIVE, /* accounted by this cgroup itself */
372 KMEM_ACCOUNTED_DEAD, /* dead memcg with pending kmem charges */
375 #ifdef CONFIG_MEMCG_KMEM
376 static inline void memcg_kmem_set_active(struct mem_cgroup *memcg)
378 set_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
381 static bool memcg_kmem_is_active(struct mem_cgroup *memcg)
383 return test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
386 static void memcg_kmem_mark_dead(struct mem_cgroup *memcg)
389 * Our caller must use css_get() first, because memcg_uncharge_kmem()
390 * will call css_put() if it sees the memcg is dead.
393 if (test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags))
394 set_bit(KMEM_ACCOUNTED_DEAD, &memcg->kmem_account_flags);
397 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup *memcg)
399 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD,
400 &memcg->kmem_account_flags);
404 /* Stuffs for move charges at task migration. */
406 * Types of charges to be moved. "move_charge_at_immitgrate" and
407 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
410 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
411 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
415 /* "mc" and its members are protected by cgroup_mutex */
416 static struct move_charge_struct {
417 spinlock_t lock; /* for from, to */
418 struct mem_cgroup *from;
419 struct mem_cgroup *to;
420 unsigned long immigrate_flags;
421 unsigned long precharge;
422 unsigned long moved_charge;
423 unsigned long moved_swap;
424 struct task_struct *moving_task; /* a task moving charges */
425 wait_queue_head_t waitq; /* a waitq for other context */
427 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
428 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
431 static bool move_anon(void)
433 return test_bit(MOVE_CHARGE_TYPE_ANON, &mc.immigrate_flags);
436 static bool move_file(void)
438 return test_bit(MOVE_CHARGE_TYPE_FILE, &mc.immigrate_flags);
442 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
443 * limit reclaim to prevent infinite loops, if they ever occur.
445 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
446 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
449 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
450 MEM_CGROUP_CHARGE_TYPE_ANON,
451 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
452 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
456 /* for encoding cft->private value on file */
464 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
465 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
466 #define MEMFILE_ATTR(val) ((val) & 0xffff)
467 /* Used for OOM nofiier */
468 #define OOM_CONTROL (0)
471 * The memcg_create_mutex will be held whenever a new cgroup is created.
472 * As a consequence, any change that needs to protect against new child cgroups
473 * appearing has to hold it as well.
475 static DEFINE_MUTEX(memcg_create_mutex);
477 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
479 return s ? container_of(s, struct mem_cgroup, css) : NULL;
482 /* Some nice accessors for the vmpressure. */
483 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
486 memcg = root_mem_cgroup;
487 return &memcg->vmpressure;
490 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
492 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
495 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
497 return (memcg == root_mem_cgroup);
501 * We restrict the id in the range of [1, 65535], so it can fit into
504 #define MEM_CGROUP_ID_MAX USHRT_MAX
506 static inline unsigned short mem_cgroup_id(struct mem_cgroup *memcg)
508 return memcg->css.id;
511 static inline struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
513 struct cgroup_subsys_state *css;
515 css = css_from_id(id, &memory_cgrp_subsys);
516 return mem_cgroup_from_css(css);
519 /* Writing them here to avoid exposing memcg's inner layout */
520 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
522 void sock_update_memcg(struct sock *sk)
524 if (mem_cgroup_sockets_enabled) {
525 struct mem_cgroup *memcg;
526 struct cg_proto *cg_proto;
528 BUG_ON(!sk->sk_prot->proto_cgroup);
530 /* Socket cloning can throw us here with sk_cgrp already
531 * filled. It won't however, necessarily happen from
532 * process context. So the test for root memcg given
533 * the current task's memcg won't help us in this case.
535 * Respecting the original socket's memcg is a better
536 * decision in this case.
539 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
540 css_get(&sk->sk_cgrp->memcg->css);
545 memcg = mem_cgroup_from_task(current);
546 cg_proto = sk->sk_prot->proto_cgroup(memcg);
547 if (!mem_cgroup_is_root(memcg) &&
548 memcg_proto_active(cg_proto) &&
549 css_tryget_online(&memcg->css)) {
550 sk->sk_cgrp = cg_proto;
555 EXPORT_SYMBOL(sock_update_memcg);
557 void sock_release_memcg(struct sock *sk)
559 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
560 struct mem_cgroup *memcg;
561 WARN_ON(!sk->sk_cgrp->memcg);
562 memcg = sk->sk_cgrp->memcg;
563 css_put(&sk->sk_cgrp->memcg->css);
567 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
569 if (!memcg || mem_cgroup_is_root(memcg))
572 return &memcg->tcp_mem;
574 EXPORT_SYMBOL(tcp_proto_cgroup);
576 static void disarm_sock_keys(struct mem_cgroup *memcg)
578 if (!memcg_proto_activated(&memcg->tcp_mem))
580 static_key_slow_dec(&memcg_socket_limit_enabled);
583 static void disarm_sock_keys(struct mem_cgroup *memcg)
588 #ifdef CONFIG_MEMCG_KMEM
590 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
591 * The main reason for not using cgroup id for this:
592 * this works better in sparse environments, where we have a lot of memcgs,
593 * but only a few kmem-limited. Or also, if we have, for instance, 200
594 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
595 * 200 entry array for that.
597 * The current size of the caches array is stored in
598 * memcg_limited_groups_array_size. It will double each time we have to
601 static DEFINE_IDA(kmem_limited_groups);
602 int memcg_limited_groups_array_size;
605 * MIN_SIZE is different than 1, because we would like to avoid going through
606 * the alloc/free process all the time. In a small machine, 4 kmem-limited
607 * cgroups is a reasonable guess. In the future, it could be a parameter or
608 * tunable, but that is strictly not necessary.
610 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
611 * this constant directly from cgroup, but it is understandable that this is
612 * better kept as an internal representation in cgroup.c. In any case, the
613 * cgrp_id space is not getting any smaller, and we don't have to necessarily
614 * increase ours as well if it increases.
616 #define MEMCG_CACHES_MIN_SIZE 4
617 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
620 * A lot of the calls to the cache allocation functions are expected to be
621 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
622 * conditional to this static branch, we'll have to allow modules that does
623 * kmem_cache_alloc and the such to see this symbol as well
625 struct static_key memcg_kmem_enabled_key;
626 EXPORT_SYMBOL(memcg_kmem_enabled_key);
628 static void memcg_free_cache_id(int id);
630 static void disarm_kmem_keys(struct mem_cgroup *memcg)
632 if (memcg_kmem_is_active(memcg)) {
633 static_key_slow_dec(&memcg_kmem_enabled_key);
634 memcg_free_cache_id(memcg->kmemcg_id);
637 * This check can't live in kmem destruction function,
638 * since the charges will outlive the cgroup
640 WARN_ON(page_counter_read(&memcg->kmem));
643 static void disarm_kmem_keys(struct mem_cgroup *memcg)
646 #endif /* CONFIG_MEMCG_KMEM */
648 static void disarm_static_keys(struct mem_cgroup *memcg)
650 disarm_sock_keys(memcg);
651 disarm_kmem_keys(memcg);
654 static void drain_all_stock_async(struct mem_cgroup *memcg);
656 static struct mem_cgroup_per_zone *
657 mem_cgroup_zone_zoneinfo(struct mem_cgroup *memcg, struct zone *zone)
659 int nid = zone_to_nid(zone);
660 int zid = zone_idx(zone);
662 return &memcg->nodeinfo[nid]->zoneinfo[zid];
665 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
670 static struct mem_cgroup_per_zone *
671 mem_cgroup_page_zoneinfo(struct mem_cgroup *memcg, struct page *page)
673 int nid = page_to_nid(page);
674 int zid = page_zonenum(page);
676 return &memcg->nodeinfo[nid]->zoneinfo[zid];
679 static struct mem_cgroup_tree_per_zone *
680 soft_limit_tree_node_zone(int nid, int zid)
682 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
685 static struct mem_cgroup_tree_per_zone *
686 soft_limit_tree_from_page(struct page *page)
688 int nid = page_to_nid(page);
689 int zid = page_zonenum(page);
691 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
694 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone *mz,
695 struct mem_cgroup_tree_per_zone *mctz,
696 unsigned long new_usage_in_excess)
698 struct rb_node **p = &mctz->rb_root.rb_node;
699 struct rb_node *parent = NULL;
700 struct mem_cgroup_per_zone *mz_node;
705 mz->usage_in_excess = new_usage_in_excess;
706 if (!mz->usage_in_excess)
710 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
712 if (mz->usage_in_excess < mz_node->usage_in_excess)
715 * We can't avoid mem cgroups that are over their soft
716 * limit by the same amount
718 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
721 rb_link_node(&mz->tree_node, parent, p);
722 rb_insert_color(&mz->tree_node, &mctz->rb_root);
726 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
727 struct mem_cgroup_tree_per_zone *mctz)
731 rb_erase(&mz->tree_node, &mctz->rb_root);
735 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
736 struct mem_cgroup_tree_per_zone *mctz)
740 spin_lock_irqsave(&mctz->lock, flags);
741 __mem_cgroup_remove_exceeded(mz, mctz);
742 spin_unlock_irqrestore(&mctz->lock, flags);
745 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
747 unsigned long nr_pages = page_counter_read(&memcg->memory);
748 unsigned long soft_limit = ACCESS_ONCE(memcg->soft_limit);
749 unsigned long excess = 0;
751 if (nr_pages > soft_limit)
752 excess = nr_pages - soft_limit;
757 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
759 unsigned long excess;
760 struct mem_cgroup_per_zone *mz;
761 struct mem_cgroup_tree_per_zone *mctz;
763 mctz = soft_limit_tree_from_page(page);
765 * Necessary to update all ancestors when hierarchy is used.
766 * because their event counter is not touched.
768 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
769 mz = mem_cgroup_page_zoneinfo(memcg, page);
770 excess = soft_limit_excess(memcg);
772 * We have to update the tree if mz is on RB-tree or
773 * mem is over its softlimit.
775 if (excess || mz->on_tree) {
778 spin_lock_irqsave(&mctz->lock, flags);
779 /* if on-tree, remove it */
781 __mem_cgroup_remove_exceeded(mz, mctz);
783 * Insert again. mz->usage_in_excess will be updated.
784 * If excess is 0, no tree ops.
786 __mem_cgroup_insert_exceeded(mz, mctz, excess);
787 spin_unlock_irqrestore(&mctz->lock, flags);
792 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
794 struct mem_cgroup_tree_per_zone *mctz;
795 struct mem_cgroup_per_zone *mz;
799 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
800 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
801 mctz = soft_limit_tree_node_zone(nid, zid);
802 mem_cgroup_remove_exceeded(mz, mctz);
807 static struct mem_cgroup_per_zone *
808 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
810 struct rb_node *rightmost = NULL;
811 struct mem_cgroup_per_zone *mz;
815 rightmost = rb_last(&mctz->rb_root);
817 goto done; /* Nothing to reclaim from */
819 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
821 * Remove the node now but someone else can add it back,
822 * we will to add it back at the end of reclaim to its correct
823 * position in the tree.
825 __mem_cgroup_remove_exceeded(mz, mctz);
826 if (!soft_limit_excess(mz->memcg) ||
827 !css_tryget_online(&mz->memcg->css))
833 static struct mem_cgroup_per_zone *
834 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
836 struct mem_cgroup_per_zone *mz;
838 spin_lock_irq(&mctz->lock);
839 mz = __mem_cgroup_largest_soft_limit_node(mctz);
840 spin_unlock_irq(&mctz->lock);
845 * Implementation Note: reading percpu statistics for memcg.
847 * Both of vmstat[] and percpu_counter has threshold and do periodic
848 * synchronization to implement "quick" read. There are trade-off between
849 * reading cost and precision of value. Then, we may have a chance to implement
850 * a periodic synchronizion of counter in memcg's counter.
852 * But this _read() function is used for user interface now. The user accounts
853 * memory usage by memory cgroup and he _always_ requires exact value because
854 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
855 * have to visit all online cpus and make sum. So, for now, unnecessary
856 * synchronization is not implemented. (just implemented for cpu hotplug)
858 * If there are kernel internal actions which can make use of some not-exact
859 * value, and reading all cpu value can be performance bottleneck in some
860 * common workload, threashold and synchonization as vmstat[] should be
863 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
864 enum mem_cgroup_stat_index idx)
870 for_each_online_cpu(cpu)
871 val += per_cpu(memcg->stat->count[idx], cpu);
872 #ifdef CONFIG_HOTPLUG_CPU
873 spin_lock(&memcg->pcp_counter_lock);
874 val += memcg->nocpu_base.count[idx];
875 spin_unlock(&memcg->pcp_counter_lock);
881 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
882 enum mem_cgroup_events_index idx)
884 unsigned long val = 0;
888 for_each_online_cpu(cpu)
889 val += per_cpu(memcg->stat->events[idx], cpu);
890 #ifdef CONFIG_HOTPLUG_CPU
891 spin_lock(&memcg->pcp_counter_lock);
892 val += memcg->nocpu_base.events[idx];
893 spin_unlock(&memcg->pcp_counter_lock);
899 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
904 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
905 * counted as CACHE even if it's on ANON LRU.
908 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
911 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
914 if (PageTransHuge(page))
915 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
918 /* pagein of a big page is an event. So, ignore page size */
920 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
922 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
923 nr_pages = -nr_pages; /* for event */
926 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
929 unsigned long mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
931 struct mem_cgroup_per_zone *mz;
933 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
934 return mz->lru_size[lru];
937 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
939 unsigned int lru_mask)
941 unsigned long nr = 0;
944 VM_BUG_ON((unsigned)nid >= nr_node_ids);
946 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
947 struct mem_cgroup_per_zone *mz;
951 if (!(BIT(lru) & lru_mask))
953 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
954 nr += mz->lru_size[lru];
960 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
961 unsigned int lru_mask)
963 unsigned long nr = 0;
966 for_each_node_state(nid, N_MEMORY)
967 nr += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
971 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
972 enum mem_cgroup_events_target target)
974 unsigned long val, next;
976 val = __this_cpu_read(memcg->stat->nr_page_events);
977 next = __this_cpu_read(memcg->stat->targets[target]);
978 /* from time_after() in jiffies.h */
979 if ((long)next - (long)val < 0) {
981 case MEM_CGROUP_TARGET_THRESH:
982 next = val + THRESHOLDS_EVENTS_TARGET;
984 case MEM_CGROUP_TARGET_SOFTLIMIT:
985 next = val + SOFTLIMIT_EVENTS_TARGET;
987 case MEM_CGROUP_TARGET_NUMAINFO:
988 next = val + NUMAINFO_EVENTS_TARGET;
993 __this_cpu_write(memcg->stat->targets[target], next);
1000 * Check events in order.
1003 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
1005 /* threshold event is triggered in finer grain than soft limit */
1006 if (unlikely(mem_cgroup_event_ratelimit(memcg,
1007 MEM_CGROUP_TARGET_THRESH))) {
1009 bool do_numainfo __maybe_unused;
1011 do_softlimit = mem_cgroup_event_ratelimit(memcg,
1012 MEM_CGROUP_TARGET_SOFTLIMIT);
1013 #if MAX_NUMNODES > 1
1014 do_numainfo = mem_cgroup_event_ratelimit(memcg,
1015 MEM_CGROUP_TARGET_NUMAINFO);
1017 mem_cgroup_threshold(memcg);
1018 if (unlikely(do_softlimit))
1019 mem_cgroup_update_tree(memcg, page);
1020 #if MAX_NUMNODES > 1
1021 if (unlikely(do_numainfo))
1022 atomic_inc(&memcg->numainfo_events);
1027 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1030 * mm_update_next_owner() may clear mm->owner to NULL
1031 * if it races with swapoff, page migration, etc.
1032 * So this can be called with p == NULL.
1037 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
1040 static struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
1042 struct mem_cgroup *memcg = NULL;
1047 * Page cache insertions can happen withou an
1048 * actual mm context, e.g. during disk probing
1049 * on boot, loopback IO, acct() writes etc.
1052 memcg = root_mem_cgroup;
1054 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1055 if (unlikely(!memcg))
1056 memcg = root_mem_cgroup;
1058 } while (!css_tryget_online(&memcg->css));
1064 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1065 * @root: hierarchy root
1066 * @prev: previously returned memcg, NULL on first invocation
1067 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1069 * Returns references to children of the hierarchy below @root, or
1070 * @root itself, or %NULL after a full round-trip.
1072 * Caller must pass the return value in @prev on subsequent
1073 * invocations for reference counting, or use mem_cgroup_iter_break()
1074 * to cancel a hierarchy walk before the round-trip is complete.
1076 * Reclaimers can specify a zone and a priority level in @reclaim to
1077 * divide up the memcgs in the hierarchy among all concurrent
1078 * reclaimers operating on the same zone and priority.
1080 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1081 struct mem_cgroup *prev,
1082 struct mem_cgroup_reclaim_cookie *reclaim)
1084 struct reclaim_iter *uninitialized_var(iter);
1085 struct cgroup_subsys_state *css = NULL;
1086 struct mem_cgroup *memcg = NULL;
1087 struct mem_cgroup *pos = NULL;
1089 if (mem_cgroup_disabled())
1093 root = root_mem_cgroup;
1095 if (prev && !reclaim)
1098 if (!root->use_hierarchy && root != root_mem_cgroup) {
1107 struct mem_cgroup_per_zone *mz;
1109 mz = mem_cgroup_zone_zoneinfo(root, reclaim->zone);
1110 iter = &mz->iter[reclaim->priority];
1112 if (prev && reclaim->generation != iter->generation)
1116 pos = ACCESS_ONCE(iter->position);
1118 * A racing update may change the position and
1119 * put the last reference, hence css_tryget(),
1120 * or retry to see the updated position.
1122 } while (pos && !css_tryget(&pos->css));
1129 css = css_next_descendant_pre(css, &root->css);
1132 * Reclaimers share the hierarchy walk, and a
1133 * new one might jump in right at the end of
1134 * the hierarchy - make sure they see at least
1135 * one group and restart from the beginning.
1143 * Verify the css and acquire a reference. The root
1144 * is provided by the caller, so we know it's alive
1145 * and kicking, and don't take an extra reference.
1147 memcg = mem_cgroup_from_css(css);
1149 if (css == &root->css)
1152 if (css_tryget_online(css)) {
1154 * Make sure the memcg is initialized:
1155 * mem_cgroup_css_online() orders the the
1156 * initialization against setting the flag.
1158 if (smp_load_acquire(&memcg->initialized))
1168 if (cmpxchg(&iter->position, pos, memcg) == pos) {
1170 css_get(&memcg->css);
1176 * pairs with css_tryget when dereferencing iter->position
1185 reclaim->generation = iter->generation;
1191 if (prev && prev != root)
1192 css_put(&prev->css);
1198 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1199 * @root: hierarchy root
1200 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1202 void mem_cgroup_iter_break(struct mem_cgroup *root,
1203 struct mem_cgroup *prev)
1206 root = root_mem_cgroup;
1207 if (prev && prev != root)
1208 css_put(&prev->css);
1212 * Iteration constructs for visiting all cgroups (under a tree). If
1213 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1214 * be used for reference counting.
1216 #define for_each_mem_cgroup_tree(iter, root) \
1217 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1219 iter = mem_cgroup_iter(root, iter, NULL))
1221 #define for_each_mem_cgroup(iter) \
1222 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1224 iter = mem_cgroup_iter(NULL, iter, NULL))
1226 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1228 struct mem_cgroup *memcg;
1231 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1232 if (unlikely(!memcg))
1237 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1240 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1248 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1251 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1252 * @zone: zone of the wanted lruvec
1253 * @memcg: memcg of the wanted lruvec
1255 * Returns the lru list vector holding pages for the given @zone and
1256 * @mem. This can be the global zone lruvec, if the memory controller
1259 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1260 struct mem_cgroup *memcg)
1262 struct mem_cgroup_per_zone *mz;
1263 struct lruvec *lruvec;
1265 if (mem_cgroup_disabled()) {
1266 lruvec = &zone->lruvec;
1270 mz = mem_cgroup_zone_zoneinfo(memcg, zone);
1271 lruvec = &mz->lruvec;
1274 * Since a node can be onlined after the mem_cgroup was created,
1275 * we have to be prepared to initialize lruvec->zone here;
1276 * and if offlined then reonlined, we need to reinitialize it.
1278 if (unlikely(lruvec->zone != zone))
1279 lruvec->zone = zone;
1284 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1286 * @zone: zone of the page
1288 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1290 struct mem_cgroup_per_zone *mz;
1291 struct mem_cgroup *memcg;
1292 struct page_cgroup *pc;
1293 struct lruvec *lruvec;
1295 if (mem_cgroup_disabled()) {
1296 lruvec = &zone->lruvec;
1300 pc = lookup_page_cgroup(page);
1301 memcg = pc->mem_cgroup;
1304 * Surreptitiously switch any uncharged offlist page to root:
1305 * an uncharged page off lru does nothing to secure
1306 * its former mem_cgroup from sudden removal.
1308 * Our caller holds lru_lock, and PageCgroupUsed is updated
1309 * under page_cgroup lock: between them, they make all uses
1310 * of pc->mem_cgroup safe.
1312 if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1313 pc->mem_cgroup = memcg = root_mem_cgroup;
1315 mz = mem_cgroup_page_zoneinfo(memcg, page);
1316 lruvec = &mz->lruvec;
1319 * Since a node can be onlined after the mem_cgroup was created,
1320 * we have to be prepared to initialize lruvec->zone here;
1321 * and if offlined then reonlined, we need to reinitialize it.
1323 if (unlikely(lruvec->zone != zone))
1324 lruvec->zone = zone;
1329 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1330 * @lruvec: mem_cgroup per zone lru vector
1331 * @lru: index of lru list the page is sitting on
1332 * @nr_pages: positive when adding or negative when removing
1334 * This function must be called when a page is added to or removed from an
1337 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1340 struct mem_cgroup_per_zone *mz;
1341 unsigned long *lru_size;
1343 if (mem_cgroup_disabled())
1346 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1347 lru_size = mz->lru_size + lru;
1348 *lru_size += nr_pages;
1349 VM_BUG_ON((long)(*lru_size) < 0);
1353 * Checks whether given mem is same or in the root_mem_cgroup's
1356 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1357 struct mem_cgroup *memcg)
1359 if (root_memcg == memcg)
1361 if (!root_memcg->use_hierarchy || !memcg)
1363 return cgroup_is_descendant(memcg->css.cgroup, root_memcg->css.cgroup);
1366 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1367 struct mem_cgroup *memcg)
1372 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1377 bool task_in_mem_cgroup(struct task_struct *task,
1378 const struct mem_cgroup *memcg)
1380 struct mem_cgroup *curr = NULL;
1381 struct task_struct *p;
1384 p = find_lock_task_mm(task);
1386 curr = get_mem_cgroup_from_mm(p->mm);
1390 * All threads may have already detached their mm's, but the oom
1391 * killer still needs to detect if they have already been oom
1392 * killed to prevent needlessly killing additional tasks.
1395 curr = mem_cgroup_from_task(task);
1397 css_get(&curr->css);
1401 * We should check use_hierarchy of "memcg" not "curr". Because checking
1402 * use_hierarchy of "curr" here make this function true if hierarchy is
1403 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1404 * hierarchy(even if use_hierarchy is disabled in "memcg").
1406 ret = mem_cgroup_same_or_subtree(memcg, curr);
1407 css_put(&curr->css);
1411 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1413 unsigned long inactive_ratio;
1414 unsigned long inactive;
1415 unsigned long active;
1418 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1419 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1421 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1423 inactive_ratio = int_sqrt(10 * gb);
1427 return inactive * inactive_ratio < active;
1430 #define mem_cgroup_from_counter(counter, member) \
1431 container_of(counter, struct mem_cgroup, member)
1434 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1435 * @memcg: the memory cgroup
1437 * Returns the maximum amount of memory @mem can be charged with, in
1440 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1442 unsigned long margin = 0;
1443 unsigned long count;
1444 unsigned long limit;
1446 count = page_counter_read(&memcg->memory);
1447 limit = ACCESS_ONCE(memcg->memory.limit);
1449 margin = limit - count;
1451 if (do_swap_account) {
1452 count = page_counter_read(&memcg->memsw);
1453 limit = ACCESS_ONCE(memcg->memsw.limit);
1455 margin = min(margin, limit - count);
1461 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1464 if (mem_cgroup_disabled() || !memcg->css.parent)
1465 return vm_swappiness;
1467 return memcg->swappiness;
1471 * memcg->moving_account is used for checking possibility that some thread is
1472 * calling move_account(). When a thread on CPU-A starts moving pages under
1473 * a memcg, other threads should check memcg->moving_account under
1474 * rcu_read_lock(), like this:
1478 * memcg->moving_account+1 if (memcg->mocing_account)
1480 * synchronize_rcu() update something.
1485 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1487 atomic_inc(&memcg->moving_account);
1491 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1494 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1495 * We check NULL in callee rather than caller.
1498 atomic_dec(&memcg->moving_account);
1502 * A routine for checking "mem" is under move_account() or not.
1504 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1505 * moving cgroups. This is for waiting at high-memory pressure
1508 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1510 struct mem_cgroup *from;
1511 struct mem_cgroup *to;
1514 * Unlike task_move routines, we access mc.to, mc.from not under
1515 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1517 spin_lock(&mc.lock);
1523 ret = mem_cgroup_same_or_subtree(memcg, from)
1524 || mem_cgroup_same_or_subtree(memcg, to);
1526 spin_unlock(&mc.lock);
1530 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1532 if (mc.moving_task && current != mc.moving_task) {
1533 if (mem_cgroup_under_move(memcg)) {
1535 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1536 /* moving charge context might have finished. */
1539 finish_wait(&mc.waitq, &wait);
1547 * Take this lock when
1548 * - a code tries to modify page's memcg while it's USED.
1549 * - a code tries to modify page state accounting in a memcg.
1551 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1552 unsigned long *flags)
1554 spin_lock_irqsave(&memcg->move_lock, *flags);
1557 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1558 unsigned long *flags)
1560 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1563 #define K(x) ((x) << (PAGE_SHIFT-10))
1565 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1566 * @memcg: The memory cgroup that went over limit
1567 * @p: Task that is going to be killed
1569 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1572 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1574 /* oom_info_lock ensures that parallel ooms do not interleave */
1575 static DEFINE_MUTEX(oom_info_lock);
1576 struct mem_cgroup *iter;
1582 mutex_lock(&oom_info_lock);
1585 pr_info("Task in ");
1586 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1587 pr_info(" killed as a result of limit of ");
1588 pr_cont_cgroup_path(memcg->css.cgroup);
1593 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1594 K((u64)page_counter_read(&memcg->memory)),
1595 K((u64)memcg->memory.limit), memcg->memory.failcnt);
1596 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1597 K((u64)page_counter_read(&memcg->memsw)),
1598 K((u64)memcg->memsw.limit), memcg->memsw.failcnt);
1599 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1600 K((u64)page_counter_read(&memcg->kmem)),
1601 K((u64)memcg->kmem.limit), memcg->kmem.failcnt);
1603 for_each_mem_cgroup_tree(iter, memcg) {
1604 pr_info("Memory cgroup stats for ");
1605 pr_cont_cgroup_path(iter->css.cgroup);
1608 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1609 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1611 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1612 K(mem_cgroup_read_stat(iter, i)));
1615 for (i = 0; i < NR_LRU_LISTS; i++)
1616 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1617 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1621 mutex_unlock(&oom_info_lock);
1625 * This function returns the number of memcg under hierarchy tree. Returns
1626 * 1(self count) if no children.
1628 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1631 struct mem_cgroup *iter;
1633 for_each_mem_cgroup_tree(iter, memcg)
1639 * Return the memory (and swap, if configured) limit for a memcg.
1641 static unsigned long mem_cgroup_get_limit(struct mem_cgroup *memcg)
1643 unsigned long limit;
1645 limit = memcg->memory.limit;
1646 if (mem_cgroup_swappiness(memcg)) {
1647 unsigned long memsw_limit;
1649 memsw_limit = memcg->memsw.limit;
1650 limit = min(limit + total_swap_pages, memsw_limit);
1655 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1658 struct mem_cgroup *iter;
1659 unsigned long chosen_points = 0;
1660 unsigned long totalpages;
1661 unsigned int points = 0;
1662 struct task_struct *chosen = NULL;
1665 * If current has a pending SIGKILL or is exiting, then automatically
1666 * select it. The goal is to allow it to allocate so that it may
1667 * quickly exit and free its memory.
1669 if (fatal_signal_pending(current) || current->flags & PF_EXITING) {
1670 set_thread_flag(TIF_MEMDIE);
1674 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1675 totalpages = mem_cgroup_get_limit(memcg) ? : 1;
1676 for_each_mem_cgroup_tree(iter, memcg) {
1677 struct css_task_iter it;
1678 struct task_struct *task;
1680 css_task_iter_start(&iter->css, &it);
1681 while ((task = css_task_iter_next(&it))) {
1682 switch (oom_scan_process_thread(task, totalpages, NULL,
1684 case OOM_SCAN_SELECT:
1686 put_task_struct(chosen);
1688 chosen_points = ULONG_MAX;
1689 get_task_struct(chosen);
1691 case OOM_SCAN_CONTINUE:
1693 case OOM_SCAN_ABORT:
1694 css_task_iter_end(&it);
1695 mem_cgroup_iter_break(memcg, iter);
1697 put_task_struct(chosen);
1702 points = oom_badness(task, memcg, NULL, totalpages);
1703 if (!points || points < chosen_points)
1705 /* Prefer thread group leaders for display purposes */
1706 if (points == chosen_points &&
1707 thread_group_leader(chosen))
1711 put_task_struct(chosen);
1713 chosen_points = points;
1714 get_task_struct(chosen);
1716 css_task_iter_end(&it);
1721 points = chosen_points * 1000 / totalpages;
1722 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1723 NULL, "Memory cgroup out of memory");
1727 * test_mem_cgroup_node_reclaimable
1728 * @memcg: the target memcg
1729 * @nid: the node ID to be checked.
1730 * @noswap : specify true here if the user wants flle only information.
1732 * This function returns whether the specified memcg contains any
1733 * reclaimable pages on a node. Returns true if there are any reclaimable
1734 * pages in the node.
1736 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1737 int nid, bool noswap)
1739 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1741 if (noswap || !total_swap_pages)
1743 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1748 #if MAX_NUMNODES > 1
1751 * Always updating the nodemask is not very good - even if we have an empty
1752 * list or the wrong list here, we can start from some node and traverse all
1753 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1756 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1760 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1761 * pagein/pageout changes since the last update.
1763 if (!atomic_read(&memcg->numainfo_events))
1765 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1768 /* make a nodemask where this memcg uses memory from */
1769 memcg->scan_nodes = node_states[N_MEMORY];
1771 for_each_node_mask(nid, node_states[N_MEMORY]) {
1773 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1774 node_clear(nid, memcg->scan_nodes);
1777 atomic_set(&memcg->numainfo_events, 0);
1778 atomic_set(&memcg->numainfo_updating, 0);
1782 * Selecting a node where we start reclaim from. Because what we need is just
1783 * reducing usage counter, start from anywhere is O,K. Considering
1784 * memory reclaim from current node, there are pros. and cons.
1786 * Freeing memory from current node means freeing memory from a node which
1787 * we'll use or we've used. So, it may make LRU bad. And if several threads
1788 * hit limits, it will see a contention on a node. But freeing from remote
1789 * node means more costs for memory reclaim because of memory latency.
1791 * Now, we use round-robin. Better algorithm is welcomed.
1793 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1797 mem_cgroup_may_update_nodemask(memcg);
1798 node = memcg->last_scanned_node;
1800 node = next_node(node, memcg->scan_nodes);
1801 if (node == MAX_NUMNODES)
1802 node = first_node(memcg->scan_nodes);
1804 * We call this when we hit limit, not when pages are added to LRU.
1805 * No LRU may hold pages because all pages are UNEVICTABLE or
1806 * memcg is too small and all pages are not on LRU. In that case,
1807 * we use curret node.
1809 if (unlikely(node == MAX_NUMNODES))
1810 node = numa_node_id();
1812 memcg->last_scanned_node = node;
1817 * Check all nodes whether it contains reclaimable pages or not.
1818 * For quick scan, we make use of scan_nodes. This will allow us to skip
1819 * unused nodes. But scan_nodes is lazily updated and may not cotain
1820 * enough new information. We need to do double check.
1822 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1827 * quick check...making use of scan_node.
1828 * We can skip unused nodes.
1830 if (!nodes_empty(memcg->scan_nodes)) {
1831 for (nid = first_node(memcg->scan_nodes);
1833 nid = next_node(nid, memcg->scan_nodes)) {
1835 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1840 * Check rest of nodes.
1842 for_each_node_state(nid, N_MEMORY) {
1843 if (node_isset(nid, memcg->scan_nodes))
1845 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1852 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1857 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1859 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
1863 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1866 unsigned long *total_scanned)
1868 struct mem_cgroup *victim = NULL;
1871 unsigned long excess;
1872 unsigned long nr_scanned;
1873 struct mem_cgroup_reclaim_cookie reclaim = {
1878 excess = soft_limit_excess(root_memcg);
1881 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1886 * If we have not been able to reclaim
1887 * anything, it might because there are
1888 * no reclaimable pages under this hierarchy
1893 * We want to do more targeted reclaim.
1894 * excess >> 2 is not to excessive so as to
1895 * reclaim too much, nor too less that we keep
1896 * coming back to reclaim from this cgroup
1898 if (total >= (excess >> 2) ||
1899 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1904 if (!mem_cgroup_reclaimable(victim, false))
1906 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1908 *total_scanned += nr_scanned;
1909 if (!soft_limit_excess(root_memcg))
1912 mem_cgroup_iter_break(root_memcg, victim);
1916 #ifdef CONFIG_LOCKDEP
1917 static struct lockdep_map memcg_oom_lock_dep_map = {
1918 .name = "memcg_oom_lock",
1922 static DEFINE_SPINLOCK(memcg_oom_lock);
1925 * Check OOM-Killer is already running under our hierarchy.
1926 * If someone is running, return false.
1928 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1930 struct mem_cgroup *iter, *failed = NULL;
1932 spin_lock(&memcg_oom_lock);
1934 for_each_mem_cgroup_tree(iter, memcg) {
1935 if (iter->oom_lock) {
1937 * this subtree of our hierarchy is already locked
1938 * so we cannot give a lock.
1941 mem_cgroup_iter_break(memcg, iter);
1944 iter->oom_lock = true;
1949 * OK, we failed to lock the whole subtree so we have
1950 * to clean up what we set up to the failing subtree
1952 for_each_mem_cgroup_tree(iter, memcg) {
1953 if (iter == failed) {
1954 mem_cgroup_iter_break(memcg, iter);
1957 iter->oom_lock = false;
1960 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1962 spin_unlock(&memcg_oom_lock);
1967 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1969 struct mem_cgroup *iter;
1971 spin_lock(&memcg_oom_lock);
1972 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1973 for_each_mem_cgroup_tree(iter, memcg)
1974 iter->oom_lock = false;
1975 spin_unlock(&memcg_oom_lock);
1978 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1980 struct mem_cgroup *iter;
1982 for_each_mem_cgroup_tree(iter, memcg)
1983 atomic_inc(&iter->under_oom);
1986 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1988 struct mem_cgroup *iter;
1991 * When a new child is created while the hierarchy is under oom,
1992 * mem_cgroup_oom_lock() may not be called. We have to use
1993 * atomic_add_unless() here.
1995 for_each_mem_cgroup_tree(iter, memcg)
1996 atomic_add_unless(&iter->under_oom, -1, 0);
1999 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
2001 struct oom_wait_info {
2002 struct mem_cgroup *memcg;
2006 static int memcg_oom_wake_function(wait_queue_t *wait,
2007 unsigned mode, int sync, void *arg)
2009 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
2010 struct mem_cgroup *oom_wait_memcg;
2011 struct oom_wait_info *oom_wait_info;
2013 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
2014 oom_wait_memcg = oom_wait_info->memcg;
2017 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2018 * Then we can use css_is_ancestor without taking care of RCU.
2020 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
2021 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
2023 return autoremove_wake_function(wait, mode, sync, arg);
2026 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
2028 atomic_inc(&memcg->oom_wakeups);
2029 /* for filtering, pass "memcg" as argument. */
2030 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
2033 static void memcg_oom_recover(struct mem_cgroup *memcg)
2035 if (memcg && atomic_read(&memcg->under_oom))
2036 memcg_wakeup_oom(memcg);
2039 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
2041 if (!current->memcg_oom.may_oom)
2044 * We are in the middle of the charge context here, so we
2045 * don't want to block when potentially sitting on a callstack
2046 * that holds all kinds of filesystem and mm locks.
2048 * Also, the caller may handle a failed allocation gracefully
2049 * (like optional page cache readahead) and so an OOM killer
2050 * invocation might not even be necessary.
2052 * That's why we don't do anything here except remember the
2053 * OOM context and then deal with it at the end of the page
2054 * fault when the stack is unwound, the locks are released,
2055 * and when we know whether the fault was overall successful.
2057 css_get(&memcg->css);
2058 current->memcg_oom.memcg = memcg;
2059 current->memcg_oom.gfp_mask = mask;
2060 current->memcg_oom.order = order;
2064 * mem_cgroup_oom_synchronize - complete memcg OOM handling
2065 * @handle: actually kill/wait or just clean up the OOM state
2067 * This has to be called at the end of a page fault if the memcg OOM
2068 * handler was enabled.
2070 * Memcg supports userspace OOM handling where failed allocations must
2071 * sleep on a waitqueue until the userspace task resolves the
2072 * situation. Sleeping directly in the charge context with all kinds
2073 * of locks held is not a good idea, instead we remember an OOM state
2074 * in the task and mem_cgroup_oom_synchronize() has to be called at
2075 * the end of the page fault to complete the OOM handling.
2077 * Returns %true if an ongoing memcg OOM situation was detected and
2078 * completed, %false otherwise.
2080 bool mem_cgroup_oom_synchronize(bool handle)
2082 struct mem_cgroup *memcg = current->memcg_oom.memcg;
2083 struct oom_wait_info owait;
2086 /* OOM is global, do not handle */
2093 owait.memcg = memcg;
2094 owait.wait.flags = 0;
2095 owait.wait.func = memcg_oom_wake_function;
2096 owait.wait.private = current;
2097 INIT_LIST_HEAD(&owait.wait.task_list);
2099 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2100 mem_cgroup_mark_under_oom(memcg);
2102 locked = mem_cgroup_oom_trylock(memcg);
2105 mem_cgroup_oom_notify(memcg);
2107 if (locked && !memcg->oom_kill_disable) {
2108 mem_cgroup_unmark_under_oom(memcg);
2109 finish_wait(&memcg_oom_waitq, &owait.wait);
2110 mem_cgroup_out_of_memory(memcg, current->memcg_oom.gfp_mask,
2111 current->memcg_oom.order);
2114 mem_cgroup_unmark_under_oom(memcg);
2115 finish_wait(&memcg_oom_waitq, &owait.wait);
2119 mem_cgroup_oom_unlock(memcg);
2121 * There is no guarantee that an OOM-lock contender
2122 * sees the wakeups triggered by the OOM kill
2123 * uncharges. Wake any sleepers explicitely.
2125 memcg_oom_recover(memcg);
2128 current->memcg_oom.memcg = NULL;
2129 css_put(&memcg->css);
2134 * mem_cgroup_begin_page_stat - begin a page state statistics transaction
2135 * @page: page that is going to change accounted state
2136 * @locked: &memcg->move_lock slowpath was taken
2137 * @flags: IRQ-state flags for &memcg->move_lock
2139 * This function must mark the beginning of an accounted page state
2140 * change to prevent double accounting when the page is concurrently
2141 * being moved to another memcg:
2143 * memcg = mem_cgroup_begin_page_stat(page, &locked, &flags);
2144 * if (TestClearPageState(page))
2145 * mem_cgroup_update_page_stat(memcg, state, -1);
2146 * mem_cgroup_end_page_stat(memcg, locked, flags);
2148 * The RCU lock is held throughout the transaction. The fast path can
2149 * get away without acquiring the memcg->move_lock (@locked is false)
2150 * because page moving starts with an RCU grace period.
2152 * The RCU lock also protects the memcg from being freed when the page
2153 * state that is going to change is the only thing preventing the page
2154 * from being uncharged. E.g. end-writeback clearing PageWriteback(),
2155 * which allows migration to go ahead and uncharge the page before the
2156 * account transaction might be complete.
2158 struct mem_cgroup *mem_cgroup_begin_page_stat(struct page *page,
2160 unsigned long *flags)
2162 struct mem_cgroup *memcg;
2163 struct page_cgroup *pc;
2167 if (mem_cgroup_disabled())
2170 pc = lookup_page_cgroup(page);
2172 memcg = pc->mem_cgroup;
2173 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2177 if (atomic_read(&memcg->moving_account) <= 0)
2180 move_lock_mem_cgroup(memcg, flags);
2181 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2182 move_unlock_mem_cgroup(memcg, flags);
2191 * mem_cgroup_end_page_stat - finish a page state statistics transaction
2192 * @memcg: the memcg that was accounted against
2193 * @locked: value received from mem_cgroup_begin_page_stat()
2194 * @flags: value received from mem_cgroup_begin_page_stat()
2196 void mem_cgroup_end_page_stat(struct mem_cgroup *memcg, bool locked,
2197 unsigned long flags)
2199 if (memcg && locked)
2200 move_unlock_mem_cgroup(memcg, &flags);
2206 * mem_cgroup_update_page_stat - update page state statistics
2207 * @memcg: memcg to account against
2208 * @idx: page state item to account
2209 * @val: number of pages (positive or negative)
2211 * See mem_cgroup_begin_page_stat() for locking requirements.
2213 void mem_cgroup_update_page_stat(struct mem_cgroup *memcg,
2214 enum mem_cgroup_stat_index idx, int val)
2216 VM_BUG_ON(!rcu_read_lock_held());
2219 this_cpu_add(memcg->stat->count[idx], val);
2223 * size of first charge trial. "32" comes from vmscan.c's magic value.
2224 * TODO: maybe necessary to use big numbers in big irons.
2226 #define CHARGE_BATCH 32U
2227 struct memcg_stock_pcp {
2228 struct mem_cgroup *cached; /* this never be root cgroup */
2229 unsigned int nr_pages;
2230 struct work_struct work;
2231 unsigned long flags;
2232 #define FLUSHING_CACHED_CHARGE 0
2234 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2235 static DEFINE_MUTEX(percpu_charge_mutex);
2238 * consume_stock: Try to consume stocked charge on this cpu.
2239 * @memcg: memcg to consume from.
2240 * @nr_pages: how many pages to charge.
2242 * The charges will only happen if @memcg matches the current cpu's memcg
2243 * stock, and at least @nr_pages are available in that stock. Failure to
2244 * service an allocation will refill the stock.
2246 * returns true if successful, false otherwise.
2248 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2250 struct memcg_stock_pcp *stock;
2253 if (nr_pages > CHARGE_BATCH)
2256 stock = &get_cpu_var(memcg_stock);
2257 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2258 stock->nr_pages -= nr_pages;
2261 put_cpu_var(memcg_stock);
2266 * Returns stocks cached in percpu and reset cached information.
2268 static void drain_stock(struct memcg_stock_pcp *stock)
2270 struct mem_cgroup *old = stock->cached;
2272 if (stock->nr_pages) {
2273 page_counter_uncharge(&old->memory, stock->nr_pages);
2274 if (do_swap_account)
2275 page_counter_uncharge(&old->memsw, stock->nr_pages);
2276 css_put_many(&old->css, stock->nr_pages);
2277 stock->nr_pages = 0;
2279 stock->cached = NULL;
2283 * This must be called under preempt disabled or must be called by
2284 * a thread which is pinned to local cpu.
2286 static void drain_local_stock(struct work_struct *dummy)
2288 struct memcg_stock_pcp *stock = this_cpu_ptr(&memcg_stock);
2290 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2293 static void __init memcg_stock_init(void)
2297 for_each_possible_cpu(cpu) {
2298 struct memcg_stock_pcp *stock =
2299 &per_cpu(memcg_stock, cpu);
2300 INIT_WORK(&stock->work, drain_local_stock);
2305 * Cache charges(val) to local per_cpu area.
2306 * This will be consumed by consume_stock() function, later.
2308 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2310 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2312 if (stock->cached != memcg) { /* reset if necessary */
2314 stock->cached = memcg;
2316 stock->nr_pages += nr_pages;
2317 put_cpu_var(memcg_stock);
2321 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2322 * of the hierarchy under it. sync flag says whether we should block
2323 * until the work is done.
2325 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2329 /* Notify other cpus that system-wide "drain" is running */
2332 for_each_online_cpu(cpu) {
2333 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2334 struct mem_cgroup *memcg;
2336 memcg = stock->cached;
2337 if (!memcg || !stock->nr_pages)
2339 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2341 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2343 drain_local_stock(&stock->work);
2345 schedule_work_on(cpu, &stock->work);
2353 for_each_online_cpu(cpu) {
2354 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2355 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2356 flush_work(&stock->work);
2363 * Tries to drain stocked charges in other cpus. This function is asynchronous
2364 * and just put a work per cpu for draining localy on each cpu. Caller can
2365 * expects some charges will be back later but cannot wait for it.
2367 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2370 * If someone calls draining, avoid adding more kworker runs.
2372 if (!mutex_trylock(&percpu_charge_mutex))
2374 drain_all_stock(root_memcg, false);
2375 mutex_unlock(&percpu_charge_mutex);
2378 /* This is a synchronous drain interface. */
2379 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2381 /* called when force_empty is called */
2382 mutex_lock(&percpu_charge_mutex);
2383 drain_all_stock(root_memcg, true);
2384 mutex_unlock(&percpu_charge_mutex);
2388 * This function drains percpu counter value from DEAD cpu and
2389 * move it to local cpu. Note that this function can be preempted.
2391 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2395 spin_lock(&memcg->pcp_counter_lock);
2396 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2397 long x = per_cpu(memcg->stat->count[i], cpu);
2399 per_cpu(memcg->stat->count[i], cpu) = 0;
2400 memcg->nocpu_base.count[i] += x;
2402 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2403 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2405 per_cpu(memcg->stat->events[i], cpu) = 0;
2406 memcg->nocpu_base.events[i] += x;
2408 spin_unlock(&memcg->pcp_counter_lock);
2411 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
2412 unsigned long action,
2415 int cpu = (unsigned long)hcpu;
2416 struct memcg_stock_pcp *stock;
2417 struct mem_cgroup *iter;
2419 if (action == CPU_ONLINE)
2422 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2425 for_each_mem_cgroup(iter)
2426 mem_cgroup_drain_pcp_counter(iter, cpu);
2428 stock = &per_cpu(memcg_stock, cpu);
2433 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2434 unsigned int nr_pages)
2436 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2437 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2438 struct mem_cgroup *mem_over_limit;
2439 struct page_counter *counter;
2440 unsigned long nr_reclaimed;
2441 bool may_swap = true;
2442 bool drained = false;
2445 if (mem_cgroup_is_root(memcg))
2448 if (consume_stock(memcg, nr_pages))
2451 if (!do_swap_account ||
2452 !page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2453 if (!page_counter_try_charge(&memcg->memory, batch, &counter))
2455 if (do_swap_account)
2456 page_counter_uncharge(&memcg->memsw, batch);
2457 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2459 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2463 if (batch > nr_pages) {
2469 * Unlike in global OOM situations, memcg is not in a physical
2470 * memory shortage. Allow dying and OOM-killed tasks to
2471 * bypass the last charges so that they can exit quickly and
2472 * free their memory.
2474 if (unlikely(test_thread_flag(TIF_MEMDIE) ||
2475 fatal_signal_pending(current) ||
2476 current->flags & PF_EXITING))
2479 if (unlikely(task_in_memcg_oom(current)))
2482 if (!(gfp_mask & __GFP_WAIT))
2485 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2486 gfp_mask, may_swap);
2488 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2492 drain_all_stock_async(mem_over_limit);
2497 if (gfp_mask & __GFP_NORETRY)
2500 * Even though the limit is exceeded at this point, reclaim
2501 * may have been able to free some pages. Retry the charge
2502 * before killing the task.
2504 * Only for regular pages, though: huge pages are rather
2505 * unlikely to succeed so close to the limit, and we fall back
2506 * to regular pages anyway in case of failure.
2508 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2511 * At task move, charge accounts can be doubly counted. So, it's
2512 * better to wait until the end of task_move if something is going on.
2514 if (mem_cgroup_wait_acct_move(mem_over_limit))
2520 if (gfp_mask & __GFP_NOFAIL)
2523 if (fatal_signal_pending(current))
2526 mem_cgroup_oom(mem_over_limit, gfp_mask, get_order(nr_pages));
2528 if (!(gfp_mask & __GFP_NOFAIL))
2534 css_get_many(&memcg->css, batch);
2535 if (batch > nr_pages)
2536 refill_stock(memcg, batch - nr_pages);
2541 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2543 if (mem_cgroup_is_root(memcg))
2546 page_counter_uncharge(&memcg->memory, nr_pages);
2547 if (do_swap_account)
2548 page_counter_uncharge(&memcg->memsw, nr_pages);
2550 css_put_many(&memcg->css, nr_pages);
2554 * A helper function to get mem_cgroup from ID. must be called under
2555 * rcu_read_lock(). The caller is responsible for calling
2556 * css_tryget_online() if the mem_cgroup is used for charging. (dropping
2557 * refcnt from swap can be called against removed memcg.)
2559 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2561 /* ID 0 is unused ID */
2564 return mem_cgroup_from_id(id);
2568 * try_get_mem_cgroup_from_page - look up page's memcg association
2571 * Look up, get a css reference, and return the memcg that owns @page.
2573 * The page must be locked to prevent racing with swap-in and page
2574 * cache charges. If coming from an unlocked page table, the caller
2575 * must ensure the page is on the LRU or this can race with charging.
2577 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2579 struct mem_cgroup *memcg = NULL;
2580 struct page_cgroup *pc;
2584 VM_BUG_ON_PAGE(!PageLocked(page), page);
2586 pc = lookup_page_cgroup(page);
2587 if (PageCgroupUsed(pc)) {
2588 memcg = pc->mem_cgroup;
2589 if (memcg && !css_tryget_online(&memcg->css))
2591 } else if (PageSwapCache(page)) {
2592 ent.val = page_private(page);
2593 id = lookup_swap_cgroup_id(ent);
2595 memcg = mem_cgroup_lookup(id);
2596 if (memcg && !css_tryget_online(&memcg->css))
2603 static void lock_page_lru(struct page *page, int *isolated)
2605 struct zone *zone = page_zone(page);
2607 spin_lock_irq(&zone->lru_lock);
2608 if (PageLRU(page)) {
2609 struct lruvec *lruvec;
2611 lruvec = mem_cgroup_page_lruvec(page, zone);
2613 del_page_from_lru_list(page, lruvec, page_lru(page));
2619 static void unlock_page_lru(struct page *page, int isolated)
2621 struct zone *zone = page_zone(page);
2624 struct lruvec *lruvec;
2626 lruvec = mem_cgroup_page_lruvec(page, zone);
2627 VM_BUG_ON_PAGE(PageLRU(page), page);
2629 add_page_to_lru_list(page, lruvec, page_lru(page));
2631 spin_unlock_irq(&zone->lru_lock);
2634 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2637 struct page_cgroup *pc = lookup_page_cgroup(page);
2640 VM_BUG_ON_PAGE(PageCgroupUsed(pc), page);
2642 * we don't need page_cgroup_lock about tail pages, becase they are not
2643 * accessed by any other context at this point.
2647 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2648 * may already be on some other mem_cgroup's LRU. Take care of it.
2651 lock_page_lru(page, &isolated);
2654 * Nobody should be changing or seriously looking at
2655 * pc->mem_cgroup and pc->flags at this point:
2657 * - the page is uncharged
2659 * - the page is off-LRU
2661 * - an anonymous fault has exclusive page access, except for
2662 * a locked page table
2664 * - a page cache insertion, a swapin fault, or a migration
2665 * have the page locked
2667 pc->mem_cgroup = memcg;
2668 pc->flags = PCG_USED | PCG_MEM | (do_swap_account ? PCG_MEMSW : 0);
2671 unlock_page_lru(page, isolated);
2674 #ifdef CONFIG_MEMCG_KMEM
2676 * The memcg_slab_mutex is held whenever a per memcg kmem cache is created or
2677 * destroyed. It protects memcg_caches arrays and memcg_slab_caches lists.
2679 static DEFINE_MUTEX(memcg_slab_mutex);
2681 static DEFINE_MUTEX(activate_kmem_mutex);
2684 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2685 * in the memcg_cache_params struct.
2687 static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
2689 struct kmem_cache *cachep;
2691 VM_BUG_ON(p->is_root_cache);
2692 cachep = p->root_cache;
2693 return cache_from_memcg_idx(cachep, memcg_cache_id(p->memcg));
2696 #ifdef CONFIG_SLABINFO
2697 static int mem_cgroup_slabinfo_read(struct seq_file *m, void *v)
2699 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
2700 struct memcg_cache_params *params;
2702 if (!memcg_kmem_is_active(memcg))
2705 print_slabinfo_header(m);
2707 mutex_lock(&memcg_slab_mutex);
2708 list_for_each_entry(params, &memcg->memcg_slab_caches, list)
2709 cache_show(memcg_params_to_cache(params), m);
2710 mutex_unlock(&memcg_slab_mutex);
2716 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp,
2717 unsigned long nr_pages)
2719 struct page_counter *counter;
2722 ret = page_counter_try_charge(&memcg->kmem, nr_pages, &counter);
2726 ret = try_charge(memcg, gfp, nr_pages);
2727 if (ret == -EINTR) {
2729 * try_charge() chose to bypass to root due to OOM kill or
2730 * fatal signal. Since our only options are to either fail
2731 * the allocation or charge it to this cgroup, do it as a
2732 * temporary condition. But we can't fail. From a kmem/slab
2733 * perspective, the cache has already been selected, by
2734 * mem_cgroup_kmem_get_cache(), so it is too late to change
2737 * This condition will only trigger if the task entered
2738 * memcg_charge_kmem in a sane state, but was OOM-killed
2739 * during try_charge() above. Tasks that were already dying
2740 * when the allocation triggers should have been already
2741 * directed to the root cgroup in memcontrol.h
2743 page_counter_charge(&memcg->memory, nr_pages);
2744 if (do_swap_account)
2745 page_counter_charge(&memcg->memsw, nr_pages);
2746 css_get_many(&memcg->css, nr_pages);
2749 page_counter_uncharge(&memcg->kmem, nr_pages);
2754 static void memcg_uncharge_kmem(struct mem_cgroup *memcg,
2755 unsigned long nr_pages)
2757 page_counter_uncharge(&memcg->memory, nr_pages);
2758 if (do_swap_account)
2759 page_counter_uncharge(&memcg->memsw, nr_pages);
2762 if (page_counter_uncharge(&memcg->kmem, nr_pages)) {
2763 css_put_many(&memcg->css, nr_pages);
2768 * Releases a reference taken in kmem_cgroup_css_offline in case
2769 * this last uncharge is racing with the offlining code or it is
2770 * outliving the memcg existence.
2772 * The memory barrier imposed by test&clear is paired with the
2773 * explicit one in memcg_kmem_mark_dead().
2775 if (memcg_kmem_test_and_clear_dead(memcg))
2776 css_put(&memcg->css);
2778 css_put_many(&memcg->css, nr_pages);
2782 * helper for acessing a memcg's index. It will be used as an index in the
2783 * child cache array in kmem_cache, and also to derive its name. This function
2784 * will return -1 when this is not a kmem-limited memcg.
2786 int memcg_cache_id(struct mem_cgroup *memcg)
2788 return memcg ? memcg->kmemcg_id : -1;
2791 static int memcg_alloc_cache_id(void)
2796 id = ida_simple_get(&kmem_limited_groups,
2797 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2801 if (id < memcg_limited_groups_array_size)
2805 * There's no space for the new id in memcg_caches arrays,
2806 * so we have to grow them.
2809 size = 2 * (id + 1);
2810 if (size < MEMCG_CACHES_MIN_SIZE)
2811 size = MEMCG_CACHES_MIN_SIZE;
2812 else if (size > MEMCG_CACHES_MAX_SIZE)
2813 size = MEMCG_CACHES_MAX_SIZE;
2815 mutex_lock(&memcg_slab_mutex);
2816 err = memcg_update_all_caches(size);
2817 mutex_unlock(&memcg_slab_mutex);
2820 ida_simple_remove(&kmem_limited_groups, id);
2826 static void memcg_free_cache_id(int id)
2828 ida_simple_remove(&kmem_limited_groups, id);
2832 * We should update the current array size iff all caches updates succeed. This
2833 * can only be done from the slab side. The slab mutex needs to be held when
2836 void memcg_update_array_size(int num)
2838 memcg_limited_groups_array_size = num;
2841 static void memcg_register_cache(struct mem_cgroup *memcg,
2842 struct kmem_cache *root_cache)
2844 static char memcg_name_buf[NAME_MAX + 1]; /* protected by
2846 struct kmem_cache *cachep;
2849 lockdep_assert_held(&memcg_slab_mutex);
2851 id = memcg_cache_id(memcg);
2854 * Since per-memcg caches are created asynchronously on first
2855 * allocation (see memcg_kmem_get_cache()), several threads can try to
2856 * create the same cache, but only one of them may succeed.
2858 if (cache_from_memcg_idx(root_cache, id))
2861 cgroup_name(memcg->css.cgroup, memcg_name_buf, NAME_MAX + 1);
2862 cachep = memcg_create_kmem_cache(memcg, root_cache, memcg_name_buf);
2864 * If we could not create a memcg cache, do not complain, because
2865 * that's not critical at all as we can always proceed with the root
2871 css_get(&memcg->css);
2872 list_add(&cachep->memcg_params->list, &memcg->memcg_slab_caches);
2875 * Since readers won't lock (see cache_from_memcg_idx()), we need a
2876 * barrier here to ensure nobody will see the kmem_cache partially
2881 BUG_ON(root_cache->memcg_params->memcg_caches[id]);
2882 root_cache->memcg_params->memcg_caches[id] = cachep;
2885 static void memcg_unregister_cache(struct kmem_cache *cachep)
2887 struct kmem_cache *root_cache;
2888 struct mem_cgroup *memcg;
2891 lockdep_assert_held(&memcg_slab_mutex);
2893 BUG_ON(is_root_cache(cachep));
2895 root_cache = cachep->memcg_params->root_cache;
2896 memcg = cachep->memcg_params->memcg;
2897 id = memcg_cache_id(memcg);
2899 BUG_ON(root_cache->memcg_params->memcg_caches[id] != cachep);
2900 root_cache->memcg_params->memcg_caches[id] = NULL;
2902 list_del(&cachep->memcg_params->list);
2904 kmem_cache_destroy(cachep);
2906 /* drop the reference taken in memcg_register_cache */
2907 css_put(&memcg->css);
2911 * During the creation a new cache, we need to disable our accounting mechanism
2912 * altogether. This is true even if we are not creating, but rather just
2913 * enqueing new caches to be created.
2915 * This is because that process will trigger allocations; some visible, like
2916 * explicit kmallocs to auxiliary data structures, name strings and internal
2917 * cache structures; some well concealed, like INIT_WORK() that can allocate
2918 * objects during debug.
2920 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
2921 * to it. This may not be a bounded recursion: since the first cache creation
2922 * failed to complete (waiting on the allocation), we'll just try to create the
2923 * cache again, failing at the same point.
2925 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
2926 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
2927 * inside the following two functions.
2929 static inline void memcg_stop_kmem_account(void)
2931 VM_BUG_ON(!current->mm);
2932 current->memcg_kmem_skip_account++;
2935 static inline void memcg_resume_kmem_account(void)
2937 VM_BUG_ON(!current->mm);
2938 current->memcg_kmem_skip_account--;
2941 int __memcg_cleanup_cache_params(struct kmem_cache *s)
2943 struct kmem_cache *c;
2946 mutex_lock(&memcg_slab_mutex);
2947 for_each_memcg_cache_index(i) {
2948 c = cache_from_memcg_idx(s, i);
2952 memcg_unregister_cache(c);
2954 if (cache_from_memcg_idx(s, i))
2957 mutex_unlock(&memcg_slab_mutex);
2961 static void memcg_unregister_all_caches(struct mem_cgroup *memcg)
2963 struct kmem_cache *cachep;
2964 struct memcg_cache_params *params, *tmp;
2966 if (!memcg_kmem_is_active(memcg))
2969 mutex_lock(&memcg_slab_mutex);
2970 list_for_each_entry_safe(params, tmp, &memcg->memcg_slab_caches, list) {
2971 cachep = memcg_params_to_cache(params);
2972 kmem_cache_shrink(cachep);
2973 if (atomic_read(&cachep->memcg_params->nr_pages) == 0)
2974 memcg_unregister_cache(cachep);
2976 mutex_unlock(&memcg_slab_mutex);
2979 struct memcg_register_cache_work {
2980 struct mem_cgroup *memcg;
2981 struct kmem_cache *cachep;
2982 struct work_struct work;
2985 static void memcg_register_cache_func(struct work_struct *w)
2987 struct memcg_register_cache_work *cw =
2988 container_of(w, struct memcg_register_cache_work, work);
2989 struct mem_cgroup *memcg = cw->memcg;
2990 struct kmem_cache *cachep = cw->cachep;
2992 mutex_lock(&memcg_slab_mutex);
2993 memcg_register_cache(memcg, cachep);
2994 mutex_unlock(&memcg_slab_mutex);
2996 css_put(&memcg->css);
3001 * Enqueue the creation of a per-memcg kmem_cache.
3003 static void __memcg_schedule_register_cache(struct mem_cgroup *memcg,
3004 struct kmem_cache *cachep)
3006 struct memcg_register_cache_work *cw;
3008 cw = kmalloc(sizeof(*cw), GFP_NOWAIT);
3010 css_put(&memcg->css);
3015 cw->cachep = cachep;
3017 INIT_WORK(&cw->work, memcg_register_cache_func);
3018 schedule_work(&cw->work);
3021 static void memcg_schedule_register_cache(struct mem_cgroup *memcg,
3022 struct kmem_cache *cachep)
3025 * We need to stop accounting when we kmalloc, because if the
3026 * corresponding kmalloc cache is not yet created, the first allocation
3027 * in __memcg_schedule_register_cache will recurse.
3029 * However, it is better to enclose the whole function. Depending on
3030 * the debugging options enabled, INIT_WORK(), for instance, can
3031 * trigger an allocation. This too, will make us recurse. Because at
3032 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3033 * the safest choice is to do it like this, wrapping the whole function.
3035 memcg_stop_kmem_account();
3036 __memcg_schedule_register_cache(memcg, cachep);
3037 memcg_resume_kmem_account();
3040 int __memcg_charge_slab(struct kmem_cache *cachep, gfp_t gfp, int order)
3042 unsigned int nr_pages = 1 << order;
3045 res = memcg_charge_kmem(cachep->memcg_params->memcg, gfp, nr_pages);
3047 atomic_add(nr_pages, &cachep->memcg_params->nr_pages);
3051 void __memcg_uncharge_slab(struct kmem_cache *cachep, int order)
3053 unsigned int nr_pages = 1 << order;
3055 memcg_uncharge_kmem(cachep->memcg_params->memcg, nr_pages);
3056 atomic_sub(nr_pages, &cachep->memcg_params->nr_pages);
3060 * Return the kmem_cache we're supposed to use for a slab allocation.
3061 * We try to use the current memcg's version of the cache.
3063 * If the cache does not exist yet, if we are the first user of it,
3064 * we either create it immediately, if possible, or create it asynchronously
3066 * In the latter case, we will let the current allocation go through with
3067 * the original cache.
3069 * Can't be called in interrupt context or from kernel threads.
3070 * This function needs to be called with rcu_read_lock() held.
3072 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
3075 struct mem_cgroup *memcg;
3076 struct kmem_cache *memcg_cachep;
3078 VM_BUG_ON(!cachep->memcg_params);
3079 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
3081 if (!current->mm || current->memcg_kmem_skip_account)
3085 memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
3087 if (!memcg_kmem_is_active(memcg))
3090 memcg_cachep = cache_from_memcg_idx(cachep, memcg_cache_id(memcg));
3091 if (likely(memcg_cachep)) {
3092 cachep = memcg_cachep;
3096 /* The corresponding put will be done in the workqueue. */
3097 if (!css_tryget_online(&memcg->css))
3102 * If we are in a safe context (can wait, and not in interrupt
3103 * context), we could be be predictable and return right away.
3104 * This would guarantee that the allocation being performed
3105 * already belongs in the new cache.
3107 * However, there are some clashes that can arrive from locking.
3108 * For instance, because we acquire the slab_mutex while doing
3109 * memcg_create_kmem_cache, this means no further allocation
3110 * could happen with the slab_mutex held. So it's better to
3113 memcg_schedule_register_cache(memcg, cachep);
3121 * We need to verify if the allocation against current->mm->owner's memcg is
3122 * possible for the given order. But the page is not allocated yet, so we'll
3123 * need a further commit step to do the final arrangements.
3125 * It is possible for the task to switch cgroups in this mean time, so at
3126 * commit time, we can't rely on task conversion any longer. We'll then use
3127 * the handle argument to return to the caller which cgroup we should commit
3128 * against. We could also return the memcg directly and avoid the pointer
3129 * passing, but a boolean return value gives better semantics considering
3130 * the compiled-out case as well.
3132 * Returning true means the allocation is possible.
3135 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3137 struct mem_cgroup *memcg;
3143 * Disabling accounting is only relevant for some specific memcg
3144 * internal allocations. Therefore we would initially not have such
3145 * check here, since direct calls to the page allocator that are
3146 * accounted to kmemcg (alloc_kmem_pages and friends) only happen
3147 * outside memcg core. We are mostly concerned with cache allocations,
3148 * and by having this test at memcg_kmem_get_cache, we are already able
3149 * to relay the allocation to the root cache and bypass the memcg cache
3152 * There is one exception, though: the SLUB allocator does not create
3153 * large order caches, but rather service large kmallocs directly from
3154 * the page allocator. Therefore, the following sequence when backed by
3155 * the SLUB allocator:
3157 * memcg_stop_kmem_account();
3158 * kmalloc(<large_number>)
3159 * memcg_resume_kmem_account();
3161 * would effectively ignore the fact that we should skip accounting,
3162 * since it will drive us directly to this function without passing
3163 * through the cache selector memcg_kmem_get_cache. Such large
3164 * allocations are extremely rare but can happen, for instance, for the
3165 * cache arrays. We bring this test here.
3167 if (!current->mm || current->memcg_kmem_skip_account)
3170 memcg = get_mem_cgroup_from_mm(current->mm);
3172 if (!memcg_kmem_is_active(memcg)) {
3173 css_put(&memcg->css);
3177 ret = memcg_charge_kmem(memcg, gfp, 1 << order);
3181 css_put(&memcg->css);
3185 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3188 struct page_cgroup *pc;
3190 VM_BUG_ON(mem_cgroup_is_root(memcg));
3192 /* The page allocation failed. Revert */
3194 memcg_uncharge_kmem(memcg, 1 << order);
3198 * The page is freshly allocated and not visible to any
3199 * outside callers yet. Set up pc non-atomically.
3201 pc = lookup_page_cgroup(page);
3202 pc->mem_cgroup = memcg;
3203 pc->flags = PCG_USED;
3206 void __memcg_kmem_uncharge_pages(struct page *page, int order)
3208 struct mem_cgroup *memcg = NULL;
3209 struct page_cgroup *pc;
3212 pc = lookup_page_cgroup(page);
3213 if (!PageCgroupUsed(pc))
3216 memcg = pc->mem_cgroup;
3220 * We trust that only if there is a memcg associated with the page, it
3221 * is a valid allocation
3226 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
3227 memcg_uncharge_kmem(memcg, 1 << order);
3230 static inline void memcg_unregister_all_caches(struct mem_cgroup *memcg)
3233 #endif /* CONFIG_MEMCG_KMEM */
3235 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3238 * Because tail pages are not marked as "used", set it. We're under
3239 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3240 * charge/uncharge will be never happen and move_account() is done under
3241 * compound_lock(), so we don't have to take care of races.
3243 void mem_cgroup_split_huge_fixup(struct page *head)
3245 struct page_cgroup *head_pc = lookup_page_cgroup(head);
3246 struct page_cgroup *pc;
3247 struct mem_cgroup *memcg;
3250 if (mem_cgroup_disabled())
3253 memcg = head_pc->mem_cgroup;
3254 for (i = 1; i < HPAGE_PMD_NR; i++) {
3256 pc->mem_cgroup = memcg;
3257 pc->flags = head_pc->flags;
3259 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
3262 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3265 * mem_cgroup_move_account - move account of the page
3267 * @nr_pages: number of regular pages (>1 for huge pages)
3268 * @pc: page_cgroup of the page.
3269 * @from: mem_cgroup which the page is moved from.
3270 * @to: mem_cgroup which the page is moved to. @from != @to.
3272 * The caller must confirm following.
3273 * - page is not on LRU (isolate_page() is useful.)
3274 * - compound_lock is held when nr_pages > 1
3276 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3279 static int mem_cgroup_move_account(struct page *page,
3280 unsigned int nr_pages,
3281 struct page_cgroup *pc,
3282 struct mem_cgroup *from,
3283 struct mem_cgroup *to)
3285 unsigned long flags;
3288 VM_BUG_ON(from == to);
3289 VM_BUG_ON_PAGE(PageLRU(page), page);
3291 * The page is isolated from LRU. So, collapse function
3292 * will not handle this page. But page splitting can happen.
3293 * Do this check under compound_page_lock(). The caller should
3297 if (nr_pages > 1 && !PageTransHuge(page))
3301 * Prevent mem_cgroup_migrate() from looking at pc->mem_cgroup
3302 * of its source page while we change it: page migration takes
3303 * both pages off the LRU, but page cache replacement doesn't.
3305 if (!trylock_page(page))
3309 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3312 move_lock_mem_cgroup(from, &flags);
3314 if (!PageAnon(page) && page_mapped(page)) {
3315 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
3317 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
3321 if (PageWriteback(page)) {
3322 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK],
3324 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK],
3329 * It is safe to change pc->mem_cgroup here because the page
3330 * is referenced, charged, and isolated - we can't race with
3331 * uncharging, charging, migration, or LRU putback.
3334 /* caller should have done css_get */
3335 pc->mem_cgroup = to;
3336 move_unlock_mem_cgroup(from, &flags);
3339 local_irq_disable();
3340 mem_cgroup_charge_statistics(to, page, nr_pages);
3341 memcg_check_events(to, page);
3342 mem_cgroup_charge_statistics(from, page, -nr_pages);
3343 memcg_check_events(from, page);
3352 * mem_cgroup_move_parent - moves page to the parent group
3353 * @page: the page to move
3354 * @pc: page_cgroup of the page
3355 * @child: page's cgroup
3357 * move charges to its parent or the root cgroup if the group has no
3358 * parent (aka use_hierarchy==0).
3359 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3360 * mem_cgroup_move_account fails) the failure is always temporary and
3361 * it signals a race with a page removal/uncharge or migration. In the
3362 * first case the page is on the way out and it will vanish from the LRU
3363 * on the next attempt and the call should be retried later.
3364 * Isolation from the LRU fails only if page has been isolated from
3365 * the LRU since we looked at it and that usually means either global
3366 * reclaim or migration going on. The page will either get back to the
3368 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3369 * (!PageCgroupUsed) or moved to a different group. The page will
3370 * disappear in the next attempt.
3372 static int mem_cgroup_move_parent(struct page *page,
3373 struct page_cgroup *pc,
3374 struct mem_cgroup *child)
3376 struct mem_cgroup *parent;
3377 unsigned int nr_pages;
3378 unsigned long uninitialized_var(flags);
3381 VM_BUG_ON(mem_cgroup_is_root(child));
3384 if (!get_page_unless_zero(page))
3386 if (isolate_lru_page(page))
3389 nr_pages = hpage_nr_pages(page);
3391 parent = parent_mem_cgroup(child);
3393 * If no parent, move charges to root cgroup.
3396 parent = root_mem_cgroup;
3399 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
3400 flags = compound_lock_irqsave(page);
3403 ret = mem_cgroup_move_account(page, nr_pages,
3406 if (!mem_cgroup_is_root(parent))
3407 css_get_many(&parent->css, nr_pages);
3408 /* Take charge off the local counters */
3409 page_counter_cancel(&child->memory, nr_pages);
3410 if (do_swap_account)
3411 page_counter_cancel(&child->memsw, nr_pages);
3412 css_put_many(&child->css, nr_pages);
3416 compound_unlock_irqrestore(page, flags);
3417 putback_lru_page(page);
3424 #ifdef CONFIG_MEMCG_SWAP
3425 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
3428 int val = (charge) ? 1 : -1;
3429 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
3433 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3434 * @entry: swap entry to be moved
3435 * @from: mem_cgroup which the entry is moved from
3436 * @to: mem_cgroup which the entry is moved to
3438 * It succeeds only when the swap_cgroup's record for this entry is the same
3439 * as the mem_cgroup's id of @from.
3441 * Returns 0 on success, -EINVAL on failure.
3443 * The caller must have charged to @to, IOW, called page_counter_charge() about
3444 * both res and memsw, and called css_get().
3446 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3447 struct mem_cgroup *from, struct mem_cgroup *to)
3449 unsigned short old_id, new_id;
3451 old_id = mem_cgroup_id(from);
3452 new_id = mem_cgroup_id(to);
3454 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3455 mem_cgroup_swap_statistics(from, false);
3456 mem_cgroup_swap_statistics(to, true);
3458 * This function is only called from task migration context now.
3459 * It postpones page_counter and refcount handling till the end
3460 * of task migration(mem_cgroup_clear_mc()) for performance
3461 * improvement. But we cannot postpone css_get(to) because if
3462 * the process that has been moved to @to does swap-in, the
3463 * refcount of @to might be decreased to 0.
3465 * We are in attach() phase, so the cgroup is guaranteed to be
3466 * alive, so we can just call css_get().
3474 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3475 struct mem_cgroup *from, struct mem_cgroup *to)
3481 #ifdef CONFIG_DEBUG_VM
3482 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3484 struct page_cgroup *pc;
3486 pc = lookup_page_cgroup(page);
3488 * Can be NULL while feeding pages into the page allocator for
3489 * the first time, i.e. during boot or memory hotplug;
3490 * or when mem_cgroup_disabled().
3492 if (likely(pc) && PageCgroupUsed(pc))
3497 bool mem_cgroup_bad_page_check(struct page *page)
3499 if (mem_cgroup_disabled())
3502 return lookup_page_cgroup_used(page) != NULL;
3505 void mem_cgroup_print_bad_page(struct page *page)
3507 struct page_cgroup *pc;
3509 pc = lookup_page_cgroup_used(page);
3511 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
3512 pc, pc->flags, pc->mem_cgroup);
3517 static DEFINE_MUTEX(memcg_limit_mutex);
3519 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3520 unsigned long limit)
3522 unsigned long curusage;
3523 unsigned long oldusage;
3524 bool enlarge = false;
3529 * For keeping hierarchical_reclaim simple, how long we should retry
3530 * is depends on callers. We set our retry-count to be function
3531 * of # of children which we should visit in this loop.
3533 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
3534 mem_cgroup_count_children(memcg);
3536 oldusage = page_counter_read(&memcg->memory);
3539 if (signal_pending(current)) {
3544 mutex_lock(&memcg_limit_mutex);
3545 if (limit > memcg->memsw.limit) {
3546 mutex_unlock(&memcg_limit_mutex);
3550 if (limit > memcg->memory.limit)
3552 ret = page_counter_limit(&memcg->memory, limit);
3553 mutex_unlock(&memcg_limit_mutex);
3558 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, true);
3560 curusage = page_counter_read(&memcg->memory);
3561 /* Usage is reduced ? */
3562 if (curusage >= oldusage)
3565 oldusage = curusage;
3566 } while (retry_count);
3568 if (!ret && enlarge)
3569 memcg_oom_recover(memcg);
3574 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3575 unsigned long limit)
3577 unsigned long curusage;
3578 unsigned long oldusage;
3579 bool enlarge = false;
3583 /* see mem_cgroup_resize_res_limit */
3584 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
3585 mem_cgroup_count_children(memcg);
3587 oldusage = page_counter_read(&memcg->memsw);
3590 if (signal_pending(current)) {
3595 mutex_lock(&memcg_limit_mutex);
3596 if (limit < memcg->memory.limit) {
3597 mutex_unlock(&memcg_limit_mutex);
3601 if (limit > memcg->memsw.limit)
3603 ret = page_counter_limit(&memcg->memsw, limit);
3604 mutex_unlock(&memcg_limit_mutex);
3609 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, false);
3611 curusage = page_counter_read(&memcg->memsw);
3612 /* Usage is reduced ? */
3613 if (curusage >= oldusage)
3616 oldusage = curusage;
3617 } while (retry_count);
3619 if (!ret && enlarge)
3620 memcg_oom_recover(memcg);
3625 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3627 unsigned long *total_scanned)
3629 unsigned long nr_reclaimed = 0;
3630 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3631 unsigned long reclaimed;
3633 struct mem_cgroup_tree_per_zone *mctz;
3634 unsigned long excess;
3635 unsigned long nr_scanned;
3640 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3642 * This loop can run a while, specially if mem_cgroup's continuously
3643 * keep exceeding their soft limit and putting the system under
3650 mz = mem_cgroup_largest_soft_limit_node(mctz);
3655 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
3656 gfp_mask, &nr_scanned);
3657 nr_reclaimed += reclaimed;
3658 *total_scanned += nr_scanned;
3659 spin_lock_irq(&mctz->lock);
3662 * If we failed to reclaim anything from this memory cgroup
3663 * it is time to move on to the next cgroup
3669 * Loop until we find yet another one.
3671 * By the time we get the soft_limit lock
3672 * again, someone might have aded the
3673 * group back on the RB tree. Iterate to
3674 * make sure we get a different mem.
3675 * mem_cgroup_largest_soft_limit_node returns
3676 * NULL if no other cgroup is present on
3680 __mem_cgroup_largest_soft_limit_node(mctz);
3682 css_put(&next_mz->memcg->css);
3683 else /* next_mz == NULL or other memcg */
3687 __mem_cgroup_remove_exceeded(mz, mctz);
3688 excess = soft_limit_excess(mz->memcg);
3690 * One school of thought says that we should not add
3691 * back the node to the tree if reclaim returns 0.
3692 * But our reclaim could return 0, simply because due
3693 * to priority we are exposing a smaller subset of
3694 * memory to reclaim from. Consider this as a longer
3697 /* If excess == 0, no tree ops */
3698 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3699 spin_unlock_irq(&mctz->lock);
3700 css_put(&mz->memcg->css);
3703 * Could not reclaim anything and there are no more
3704 * mem cgroups to try or we seem to be looping without
3705 * reclaiming anything.
3707 if (!nr_reclaimed &&
3709 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3711 } while (!nr_reclaimed);
3713 css_put(&next_mz->memcg->css);
3714 return nr_reclaimed;
3718 * mem_cgroup_force_empty_list - clears LRU of a group
3719 * @memcg: group to clear
3722 * @lru: lru to to clear
3724 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
3725 * reclaim the pages page themselves - pages are moved to the parent (or root)
3728 static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
3729 int node, int zid, enum lru_list lru)
3731 struct lruvec *lruvec;
3732 unsigned long flags;
3733 struct list_head *list;
3737 zone = &NODE_DATA(node)->node_zones[zid];
3738 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
3739 list = &lruvec->lists[lru];
3743 struct page_cgroup *pc;
3746 spin_lock_irqsave(&zone->lru_lock, flags);
3747 if (list_empty(list)) {
3748 spin_unlock_irqrestore(&zone->lru_lock, flags);
3751 page = list_entry(list->prev, struct page, lru);
3753 list_move(&page->lru, list);
3755 spin_unlock_irqrestore(&zone->lru_lock, flags);
3758 spin_unlock_irqrestore(&zone->lru_lock, flags);
3760 pc = lookup_page_cgroup(page);
3762 if (mem_cgroup_move_parent(page, pc, memcg)) {
3763 /* found lock contention or "pc" is obsolete. */
3768 } while (!list_empty(list));
3772 * make mem_cgroup's charge to be 0 if there is no task by moving
3773 * all the charges and pages to the parent.
3774 * This enables deleting this mem_cgroup.
3776 * Caller is responsible for holding css reference on the memcg.
3778 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
3783 /* This is for making all *used* pages to be on LRU. */
3784 lru_add_drain_all();
3785 drain_all_stock_sync(memcg);
3786 mem_cgroup_start_move(memcg);
3787 for_each_node_state(node, N_MEMORY) {
3788 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3791 mem_cgroup_force_empty_list(memcg,
3796 mem_cgroup_end_move(memcg);
3797 memcg_oom_recover(memcg);
3801 * Kernel memory may not necessarily be trackable to a specific
3802 * process. So they are not migrated, and therefore we can't
3803 * expect their value to drop to 0 here.
3804 * Having res filled up with kmem only is enough.
3806 * This is a safety check because mem_cgroup_force_empty_list
3807 * could have raced with mem_cgroup_replace_page_cache callers
3808 * so the lru seemed empty but the page could have been added
3809 * right after the check. RES_USAGE should be safe as we always
3810 * charge before adding to the LRU.
3812 } while (page_counter_read(&memcg->memory) -
3813 page_counter_read(&memcg->kmem) > 0);
3817 * Test whether @memcg has children, dead or alive. Note that this
3818 * function doesn't care whether @memcg has use_hierarchy enabled and
3819 * returns %true if there are child csses according to the cgroup
3820 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
3822 static inline bool memcg_has_children(struct mem_cgroup *memcg)
3827 * The lock does not prevent addition or deletion of children, but
3828 * it prevents a new child from being initialized based on this
3829 * parent in css_online(), so it's enough to decide whether
3830 * hierarchically inherited attributes can still be changed or not.
3832 lockdep_assert_held(&memcg_create_mutex);
3835 ret = css_next_child(NULL, &memcg->css);
3841 * Reclaims as many pages from the given memcg as possible and moves
3842 * the rest to the parent.
3844 * Caller is responsible for holding css reference for memcg.
3846 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3848 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3850 /* we call try-to-free pages for make this cgroup empty */
3851 lru_add_drain_all();
3852 /* try to free all pages in this cgroup */
3853 while (nr_retries && page_counter_read(&memcg->memory)) {
3856 if (signal_pending(current))
3859 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3863 /* maybe some writeback is necessary */
3864 congestion_wait(BLK_RW_ASYNC, HZ/10);
3872 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3873 char *buf, size_t nbytes,
3876 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3878 if (mem_cgroup_is_root(memcg))
3880 return mem_cgroup_force_empty(memcg) ?: nbytes;
3883 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3886 return mem_cgroup_from_css(css)->use_hierarchy;
3889 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3890 struct cftype *cft, u64 val)
3893 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3894 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
3896 mutex_lock(&memcg_create_mutex);
3898 if (memcg->use_hierarchy == val)
3902 * If parent's use_hierarchy is set, we can't make any modifications
3903 * in the child subtrees. If it is unset, then the change can
3904 * occur, provided the current cgroup has no children.
3906 * For the root cgroup, parent_mem is NULL, we allow value to be
3907 * set if there are no children.
3909 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3910 (val == 1 || val == 0)) {
3911 if (!memcg_has_children(memcg))
3912 memcg->use_hierarchy = val;
3919 mutex_unlock(&memcg_create_mutex);
3924 static unsigned long tree_stat(struct mem_cgroup *memcg,
3925 enum mem_cgroup_stat_index idx)
3927 struct mem_cgroup *iter;
3930 /* Per-cpu values can be negative, use a signed accumulator */
3931 for_each_mem_cgroup_tree(iter, memcg)
3932 val += mem_cgroup_read_stat(iter, idx);
3934 if (val < 0) /* race ? */
3939 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3943 if (mem_cgroup_is_root(memcg)) {
3944 val = tree_stat(memcg, MEM_CGROUP_STAT_CACHE);
3945 val += tree_stat(memcg, MEM_CGROUP_STAT_RSS);
3947 val += tree_stat(memcg, MEM_CGROUP_STAT_SWAP);
3950 val = page_counter_read(&memcg->memory);
3952 val = page_counter_read(&memcg->memsw);
3954 return val << PAGE_SHIFT;
3965 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3968 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3969 struct page_counter *counter;
3971 switch (MEMFILE_TYPE(cft->private)) {
3973 counter = &memcg->memory;
3976 counter = &memcg->memsw;
3979 counter = &memcg->kmem;
3985 switch (MEMFILE_ATTR(cft->private)) {
3987 if (counter == &memcg->memory)
3988 return mem_cgroup_usage(memcg, false);
3989 if (counter == &memcg->memsw)
3990 return mem_cgroup_usage(memcg, true);
3991 return (u64)page_counter_read(counter) * PAGE_SIZE;
3993 return (u64)counter->limit * PAGE_SIZE;
3995 return (u64)counter->watermark * PAGE_SIZE;
3997 return counter->failcnt;
3998 case RES_SOFT_LIMIT:
3999 return (u64)memcg->soft_limit * PAGE_SIZE;
4005 #ifdef CONFIG_MEMCG_KMEM
4006 /* should be called with activate_kmem_mutex held */
4007 static int __memcg_activate_kmem(struct mem_cgroup *memcg,
4008 unsigned long nr_pages)
4013 if (memcg_kmem_is_active(memcg))
4017 * We are going to allocate memory for data shared by all memory
4018 * cgroups so let's stop accounting here.
4020 memcg_stop_kmem_account();
4023 * For simplicity, we won't allow this to be disabled. It also can't
4024 * be changed if the cgroup has children already, or if tasks had
4027 * If tasks join before we set the limit, a person looking at
4028 * kmem.usage_in_bytes will have no way to determine when it took
4029 * place, which makes the value quite meaningless.
4031 * After it first became limited, changes in the value of the limit are
4032 * of course permitted.
4034 mutex_lock(&memcg_create_mutex);
4035 if (cgroup_has_tasks(memcg->css.cgroup) ||
4036 (memcg->use_hierarchy && memcg_has_children(memcg)))
4038 mutex_unlock(&memcg_create_mutex);
4042 memcg_id = memcg_alloc_cache_id();
4048 memcg->kmemcg_id = memcg_id;
4049 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
4052 * We couldn't have accounted to this cgroup, because it hasn't got the
4053 * active bit set yet, so this should succeed.
4055 err = page_counter_limit(&memcg->kmem, nr_pages);
4058 static_key_slow_inc(&memcg_kmem_enabled_key);
4060 * Setting the active bit after enabling static branching will
4061 * guarantee no one starts accounting before all call sites are
4064 memcg_kmem_set_active(memcg);
4066 memcg_resume_kmem_account();
4070 static int memcg_activate_kmem(struct mem_cgroup *memcg,
4071 unsigned long nr_pages)
4075 mutex_lock(&activate_kmem_mutex);
4076 ret = __memcg_activate_kmem(memcg, nr_pages);
4077 mutex_unlock(&activate_kmem_mutex);
4081 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
4082 unsigned long limit)
4086 mutex_lock(&memcg_limit_mutex);
4087 if (!memcg_kmem_is_active(memcg))
4088 ret = memcg_activate_kmem(memcg, limit);
4090 ret = page_counter_limit(&memcg->kmem, limit);
4091 mutex_unlock(&memcg_limit_mutex);
4095 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
4098 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4103 mutex_lock(&activate_kmem_mutex);
4105 * If the parent cgroup is not kmem-active now, it cannot be activated
4106 * after this point, because it has at least one child already.
4108 if (memcg_kmem_is_active(parent))
4109 ret = __memcg_activate_kmem(memcg, PAGE_COUNTER_MAX);
4110 mutex_unlock(&activate_kmem_mutex);
4114 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
4115 unsigned long limit)
4119 #endif /* CONFIG_MEMCG_KMEM */
4122 * The user of this function is...
4125 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
4126 char *buf, size_t nbytes, loff_t off)
4128 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
4129 unsigned long nr_pages;
4132 buf = strstrip(buf);
4133 ret = page_counter_memparse(buf, &nr_pages);
4137 switch (MEMFILE_ATTR(of_cft(of)->private)) {
4139 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
4143 switch (MEMFILE_TYPE(of_cft(of)->private)) {
4145 ret = mem_cgroup_resize_limit(memcg, nr_pages);
4148 ret = mem_cgroup_resize_memsw_limit(memcg, nr_pages);
4151 ret = memcg_update_kmem_limit(memcg, nr_pages);
4155 case RES_SOFT_LIMIT:
4156 memcg->soft_limit = nr_pages;
4160 return ret ?: nbytes;
4163 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
4164 size_t nbytes, loff_t off)
4166 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
4167 struct page_counter *counter;
4169 switch (MEMFILE_TYPE(of_cft(of)->private)) {
4171 counter = &memcg->memory;
4174 counter = &memcg->memsw;
4177 counter = &memcg->kmem;
4183 switch (MEMFILE_ATTR(of_cft(of)->private)) {
4185 page_counter_reset_watermark(counter);
4188 counter->failcnt = 0;
4197 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
4200 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
4204 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
4205 struct cftype *cft, u64 val)
4207 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4209 if (val >= (1 << NR_MOVE_TYPE))
4213 * No kind of locking is needed in here, because ->can_attach() will
4214 * check this value once in the beginning of the process, and then carry
4215 * on with stale data. This means that changes to this value will only
4216 * affect task migrations starting after the change.
4218 memcg->move_charge_at_immigrate = val;
4222 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
4223 struct cftype *cft, u64 val)
4230 static int memcg_numa_stat_show(struct seq_file *m, void *v)
4234 unsigned int lru_mask;
4237 static const struct numa_stat stats[] = {
4238 { "total", LRU_ALL },
4239 { "file", LRU_ALL_FILE },
4240 { "anon", LRU_ALL_ANON },
4241 { "unevictable", BIT(LRU_UNEVICTABLE) },
4243 const struct numa_stat *stat;
4246 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
4248 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
4249 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
4250 seq_printf(m, "%s=%lu", stat->name, nr);
4251 for_each_node_state(nid, N_MEMORY) {
4252 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4254 seq_printf(m, " N%d=%lu", nid, nr);
4259 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
4260 struct mem_cgroup *iter;
4263 for_each_mem_cgroup_tree(iter, memcg)
4264 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
4265 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
4266 for_each_node_state(nid, N_MEMORY) {
4268 for_each_mem_cgroup_tree(iter, memcg)
4269 nr += mem_cgroup_node_nr_lru_pages(
4270 iter, nid, stat->lru_mask);
4271 seq_printf(m, " N%d=%lu", nid, nr);
4278 #endif /* CONFIG_NUMA */
4280 static inline void mem_cgroup_lru_names_not_uptodate(void)
4282 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
4285 static int memcg_stat_show(struct seq_file *m, void *v)
4287 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
4288 unsigned long memory, memsw;
4289 struct mem_cgroup *mi;
4292 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
4293 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
4295 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
4296 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
4299 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
4300 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
4301 mem_cgroup_read_events(memcg, i));
4303 for (i = 0; i < NR_LRU_LISTS; i++)
4304 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
4305 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
4307 /* Hierarchical information */
4308 memory = memsw = PAGE_COUNTER_MAX;
4309 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
4310 memory = min(memory, mi->memory.limit);
4311 memsw = min(memsw, mi->memsw.limit);
4313 seq_printf(m, "hierarchical_memory_limit %llu\n",
4314 (u64)memory * PAGE_SIZE);
4315 if (do_swap_account)
4316 seq_printf(m, "hierarchical_memsw_limit %llu\n",
4317 (u64)memsw * PAGE_SIZE);
4319 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
4322 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
4324 for_each_mem_cgroup_tree(mi, memcg)
4325 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
4326 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
4329 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
4330 unsigned long long val = 0;
4332 for_each_mem_cgroup_tree(mi, memcg)
4333 val += mem_cgroup_read_events(mi, i);
4334 seq_printf(m, "total_%s %llu\n",
4335 mem_cgroup_events_names[i], val);
4338 for (i = 0; i < NR_LRU_LISTS; i++) {
4339 unsigned long long val = 0;
4341 for_each_mem_cgroup_tree(mi, memcg)
4342 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
4343 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
4346 #ifdef CONFIG_DEBUG_VM
4349 struct mem_cgroup_per_zone *mz;
4350 struct zone_reclaim_stat *rstat;
4351 unsigned long recent_rotated[2] = {0, 0};
4352 unsigned long recent_scanned[2] = {0, 0};
4354 for_each_online_node(nid)
4355 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4356 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
4357 rstat = &mz->lruvec.reclaim_stat;
4359 recent_rotated[0] += rstat->recent_rotated[0];
4360 recent_rotated[1] += rstat->recent_rotated[1];
4361 recent_scanned[0] += rstat->recent_scanned[0];
4362 recent_scanned[1] += rstat->recent_scanned[1];
4364 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
4365 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
4366 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
4367 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
4374 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
4377 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4379 return mem_cgroup_swappiness(memcg);
4382 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
4383 struct cftype *cft, u64 val)
4385 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4391 memcg->swappiness = val;
4393 vm_swappiness = val;
4398 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4400 struct mem_cgroup_threshold_ary *t;
4401 unsigned long usage;
4406 t = rcu_dereference(memcg->thresholds.primary);
4408 t = rcu_dereference(memcg->memsw_thresholds.primary);
4413 usage = mem_cgroup_usage(memcg, swap);
4416 * current_threshold points to threshold just below or equal to usage.
4417 * If it's not true, a threshold was crossed after last
4418 * call of __mem_cgroup_threshold().
4420 i = t->current_threshold;
4423 * Iterate backward over array of thresholds starting from
4424 * current_threshold and check if a threshold is crossed.
4425 * If none of thresholds below usage is crossed, we read
4426 * only one element of the array here.
4428 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4429 eventfd_signal(t->entries[i].eventfd, 1);
4431 /* i = current_threshold + 1 */
4435 * Iterate forward over array of thresholds starting from
4436 * current_threshold+1 and check if a threshold is crossed.
4437 * If none of thresholds above usage is crossed, we read
4438 * only one element of the array here.
4440 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4441 eventfd_signal(t->entries[i].eventfd, 1);
4443 /* Update current_threshold */
4444 t->current_threshold = i - 1;
4449 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4452 __mem_cgroup_threshold(memcg, false);
4453 if (do_swap_account)
4454 __mem_cgroup_threshold(memcg, true);
4456 memcg = parent_mem_cgroup(memcg);
4460 static int compare_thresholds(const void *a, const void *b)
4462 const struct mem_cgroup_threshold *_a = a;
4463 const struct mem_cgroup_threshold *_b = b;
4465 if (_a->threshold > _b->threshold)
4468 if (_a->threshold < _b->threshold)
4474 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4476 struct mem_cgroup_eventfd_list *ev;
4478 spin_lock(&memcg_oom_lock);
4480 list_for_each_entry(ev, &memcg->oom_notify, list)
4481 eventfd_signal(ev->eventfd, 1);
4483 spin_unlock(&memcg_oom_lock);
4487 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4489 struct mem_cgroup *iter;
4491 for_each_mem_cgroup_tree(iter, memcg)
4492 mem_cgroup_oom_notify_cb(iter);
4495 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4496 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4498 struct mem_cgroup_thresholds *thresholds;
4499 struct mem_cgroup_threshold_ary *new;
4500 unsigned long threshold;
4501 unsigned long usage;
4504 ret = page_counter_memparse(args, &threshold);
4508 mutex_lock(&memcg->thresholds_lock);
4511 thresholds = &memcg->thresholds;
4512 usage = mem_cgroup_usage(memcg, false);
4513 } else if (type == _MEMSWAP) {
4514 thresholds = &memcg->memsw_thresholds;
4515 usage = mem_cgroup_usage(memcg, true);
4519 /* Check if a threshold crossed before adding a new one */
4520 if (thresholds->primary)
4521 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4523 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4525 /* Allocate memory for new array of thresholds */
4526 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4534 /* Copy thresholds (if any) to new array */
4535 if (thresholds->primary) {
4536 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4537 sizeof(struct mem_cgroup_threshold));
4540 /* Add new threshold */
4541 new->entries[size - 1].eventfd = eventfd;
4542 new->entries[size - 1].threshold = threshold;
4544 /* Sort thresholds. Registering of new threshold isn't time-critical */
4545 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4546 compare_thresholds, NULL);
4548 /* Find current threshold */
4549 new->current_threshold = -1;
4550 for (i = 0; i < size; i++) {
4551 if (new->entries[i].threshold <= usage) {
4553 * new->current_threshold will not be used until
4554 * rcu_assign_pointer(), so it's safe to increment
4557 ++new->current_threshold;
4562 /* Free old spare buffer and save old primary buffer as spare */
4563 kfree(thresholds->spare);
4564 thresholds->spare = thresholds->primary;
4566 rcu_assign_pointer(thresholds->primary, new);
4568 /* To be sure that nobody uses thresholds */
4572 mutex_unlock(&memcg->thresholds_lock);
4577 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4578 struct eventfd_ctx *eventfd, const char *args)
4580 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4583 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4584 struct eventfd_ctx *eventfd, const char *args)
4586 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4589 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4590 struct eventfd_ctx *eventfd, enum res_type type)
4592 struct mem_cgroup_thresholds *thresholds;
4593 struct mem_cgroup_threshold_ary *new;
4594 unsigned long usage;
4597 mutex_lock(&memcg->thresholds_lock);
4600 thresholds = &memcg->thresholds;
4601 usage = mem_cgroup_usage(memcg, false);
4602 } else if (type == _MEMSWAP) {
4603 thresholds = &memcg->memsw_thresholds;
4604 usage = mem_cgroup_usage(memcg, true);
4608 if (!thresholds->primary)
4611 /* Check if a threshold crossed before removing */
4612 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4614 /* Calculate new number of threshold */
4616 for (i = 0; i < thresholds->primary->size; i++) {
4617 if (thresholds->primary->entries[i].eventfd != eventfd)
4621 new = thresholds->spare;
4623 /* Set thresholds array to NULL if we don't have thresholds */
4632 /* Copy thresholds and find current threshold */
4633 new->current_threshold = -1;
4634 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4635 if (thresholds->primary->entries[i].eventfd == eventfd)
4638 new->entries[j] = thresholds->primary->entries[i];
4639 if (new->entries[j].threshold <= usage) {
4641 * new->current_threshold will not be used
4642 * until rcu_assign_pointer(), so it's safe to increment
4645 ++new->current_threshold;
4651 /* Swap primary and spare array */
4652 thresholds->spare = thresholds->primary;
4653 /* If all events are unregistered, free the spare array */
4655 kfree(thresholds->spare);
4656 thresholds->spare = NULL;
4659 rcu_assign_pointer(thresholds->primary, new);
4661 /* To be sure that nobody uses thresholds */
4664 mutex_unlock(&memcg->thresholds_lock);
4667 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4668 struct eventfd_ctx *eventfd)
4670 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4673 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4674 struct eventfd_ctx *eventfd)
4676 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4679 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4680 struct eventfd_ctx *eventfd, const char *args)
4682 struct mem_cgroup_eventfd_list *event;
4684 event = kmalloc(sizeof(*event), GFP_KERNEL);
4688 spin_lock(&memcg_oom_lock);
4690 event->eventfd = eventfd;
4691 list_add(&event->list, &memcg->oom_notify);
4693 /* already in OOM ? */
4694 if (atomic_read(&memcg->under_oom))
4695 eventfd_signal(eventfd, 1);
4696 spin_unlock(&memcg_oom_lock);
4701 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4702 struct eventfd_ctx *eventfd)
4704 struct mem_cgroup_eventfd_list *ev, *tmp;
4706 spin_lock(&memcg_oom_lock);
4708 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4709 if (ev->eventfd == eventfd) {
4710 list_del(&ev->list);
4715 spin_unlock(&memcg_oom_lock);
4718 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4720 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
4722 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4723 seq_printf(sf, "under_oom %d\n", (bool)atomic_read(&memcg->under_oom));
4727 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4728 struct cftype *cft, u64 val)
4730 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4732 /* cannot set to root cgroup and only 0 and 1 are allowed */
4733 if (!css->parent || !((val == 0) || (val == 1)))
4736 memcg->oom_kill_disable = val;
4738 memcg_oom_recover(memcg);
4743 #ifdef CONFIG_MEMCG_KMEM
4744 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
4748 memcg->kmemcg_id = -1;
4749 ret = memcg_propagate_kmem(memcg);
4753 return mem_cgroup_sockets_init(memcg, ss);
4756 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
4758 mem_cgroup_sockets_destroy(memcg);
4761 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
4763 if (!memcg_kmem_is_active(memcg))
4767 * kmem charges can outlive the cgroup. In the case of slab
4768 * pages, for instance, a page contain objects from various
4769 * processes. As we prevent from taking a reference for every
4770 * such allocation we have to be careful when doing uncharge
4771 * (see memcg_uncharge_kmem) and here during offlining.
4773 * The idea is that that only the _last_ uncharge which sees
4774 * the dead memcg will drop the last reference. An additional
4775 * reference is taken here before the group is marked dead
4776 * which is then paired with css_put during uncharge resp. here.
4778 * Although this might sound strange as this path is called from
4779 * css_offline() when the referencemight have dropped down to 0 and
4780 * shouldn't be incremented anymore (css_tryget_online() would
4781 * fail) we do not have other options because of the kmem
4782 * allocations lifetime.
4784 css_get(&memcg->css);
4786 memcg_kmem_mark_dead(memcg);
4788 if (page_counter_read(&memcg->kmem))
4791 if (memcg_kmem_test_and_clear_dead(memcg))
4792 css_put(&memcg->css);
4795 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
4800 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
4804 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
4810 * DO NOT USE IN NEW FILES.
4812 * "cgroup.event_control" implementation.
4814 * This is way over-engineered. It tries to support fully configurable
4815 * events for each user. Such level of flexibility is completely
4816 * unnecessary especially in the light of the planned unified hierarchy.
4818 * Please deprecate this and replace with something simpler if at all
4823 * Unregister event and free resources.
4825 * Gets called from workqueue.
4827 static void memcg_event_remove(struct work_struct *work)
4829 struct mem_cgroup_event *event =
4830 container_of(work, struct mem_cgroup_event, remove);
4831 struct mem_cgroup *memcg = event->memcg;
4833 remove_wait_queue(event->wqh, &event->wait);
4835 event->unregister_event(memcg, event->eventfd);
4837 /* Notify userspace the event is going away. */
4838 eventfd_signal(event->eventfd, 1);
4840 eventfd_ctx_put(event->eventfd);
4842 css_put(&memcg->css);
4846 * Gets called on POLLHUP on eventfd when user closes it.
4848 * Called with wqh->lock held and interrupts disabled.
4850 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
4851 int sync, void *key)
4853 struct mem_cgroup_event *event =
4854 container_of(wait, struct mem_cgroup_event, wait);
4855 struct mem_cgroup *memcg = event->memcg;
4856 unsigned long flags = (unsigned long)key;
4858 if (flags & POLLHUP) {
4860 * If the event has been detached at cgroup removal, we
4861 * can simply return knowing the other side will cleanup
4864 * We can't race against event freeing since the other
4865 * side will require wqh->lock via remove_wait_queue(),
4868 spin_lock(&memcg->event_list_lock);
4869 if (!list_empty(&event->list)) {
4870 list_del_init(&event->list);
4872 * We are in atomic context, but cgroup_event_remove()
4873 * may sleep, so we have to call it in workqueue.
4875 schedule_work(&event->remove);
4877 spin_unlock(&memcg->event_list_lock);
4883 static void memcg_event_ptable_queue_proc(struct file *file,
4884 wait_queue_head_t *wqh, poll_table *pt)
4886 struct mem_cgroup_event *event =
4887 container_of(pt, struct mem_cgroup_event, pt);
4890 add_wait_queue(wqh, &event->wait);
4894 * DO NOT USE IN NEW FILES.
4896 * Parse input and register new cgroup event handler.
4898 * Input must be in format '<event_fd> <control_fd> <args>'.
4899 * Interpretation of args is defined by control file implementation.
4901 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4902 char *buf, size_t nbytes, loff_t off)
4904 struct cgroup_subsys_state *css = of_css(of);
4905 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4906 struct mem_cgroup_event *event;
4907 struct cgroup_subsys_state *cfile_css;
4908 unsigned int efd, cfd;
4915 buf = strstrip(buf);
4917 efd = simple_strtoul(buf, &endp, 10);
4922 cfd = simple_strtoul(buf, &endp, 10);
4923 if ((*endp != ' ') && (*endp != '\0'))
4927 event = kzalloc(sizeof(*event), GFP_KERNEL);
4931 event->memcg = memcg;
4932 INIT_LIST_HEAD(&event->list);
4933 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4934 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4935 INIT_WORK(&event->remove, memcg_event_remove);
4943 event->eventfd = eventfd_ctx_fileget(efile.file);
4944 if (IS_ERR(event->eventfd)) {
4945 ret = PTR_ERR(event->eventfd);
4952 goto out_put_eventfd;
4955 /* the process need read permission on control file */
4956 /* AV: shouldn't we check that it's been opened for read instead? */
4957 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4962 * Determine the event callbacks and set them in @event. This used
4963 * to be done via struct cftype but cgroup core no longer knows
4964 * about these events. The following is crude but the whole thing
4965 * is for compatibility anyway.
4967 * DO NOT ADD NEW FILES.
4969 name = cfile.file->f_dentry->d_name.name;
4971 if (!strcmp(name, "memory.usage_in_bytes")) {
4972 event->register_event = mem_cgroup_usage_register_event;
4973 event->unregister_event = mem_cgroup_usage_unregister_event;
4974 } else if (!strcmp(name, "memory.oom_control")) {
4975 event->register_event = mem_cgroup_oom_register_event;
4976 event->unregister_event = mem_cgroup_oom_unregister_event;
4977 } else if (!strcmp(name, "memory.pressure_level")) {
4978 event->register_event = vmpressure_register_event;
4979 event->unregister_event = vmpressure_unregister_event;
4980 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4981 event->register_event = memsw_cgroup_usage_register_event;
4982 event->unregister_event = memsw_cgroup_usage_unregister_event;
4989 * Verify @cfile should belong to @css. Also, remaining events are
4990 * automatically removed on cgroup destruction but the removal is
4991 * asynchronous, so take an extra ref on @css.
4993 cfile_css = css_tryget_online_from_dir(cfile.file->f_dentry->d_parent,
4994 &memory_cgrp_subsys);
4996 if (IS_ERR(cfile_css))
4998 if (cfile_css != css) {
5003 ret = event->register_event(memcg, event->eventfd, buf);
5007 efile.file->f_op->poll(efile.file, &event->pt);
5009 spin_lock(&memcg->event_list_lock);
5010 list_add(&event->list, &memcg->event_list);
5011 spin_unlock(&memcg->event_list_lock);
5023 eventfd_ctx_put(event->eventfd);
5032 static struct cftype mem_cgroup_files[] = {
5034 .name = "usage_in_bytes",
5035 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
5036 .read_u64 = mem_cgroup_read_u64,
5039 .name = "max_usage_in_bytes",
5040 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
5041 .write = mem_cgroup_reset,
5042 .read_u64 = mem_cgroup_read_u64,
5045 .name = "limit_in_bytes",
5046 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
5047 .write = mem_cgroup_write,
5048 .read_u64 = mem_cgroup_read_u64,
5051 .name = "soft_limit_in_bytes",
5052 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
5053 .write = mem_cgroup_write,
5054 .read_u64 = mem_cgroup_read_u64,
5058 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
5059 .write = mem_cgroup_reset,
5060 .read_u64 = mem_cgroup_read_u64,
5064 .seq_show = memcg_stat_show,
5067 .name = "force_empty",
5068 .write = mem_cgroup_force_empty_write,
5071 .name = "use_hierarchy",
5072 .write_u64 = mem_cgroup_hierarchy_write,
5073 .read_u64 = mem_cgroup_hierarchy_read,
5076 .name = "cgroup.event_control", /* XXX: for compat */
5077 .write = memcg_write_event_control,
5078 .flags = CFTYPE_NO_PREFIX,
5082 .name = "swappiness",
5083 .read_u64 = mem_cgroup_swappiness_read,
5084 .write_u64 = mem_cgroup_swappiness_write,
5087 .name = "move_charge_at_immigrate",
5088 .read_u64 = mem_cgroup_move_charge_read,
5089 .write_u64 = mem_cgroup_move_charge_write,
5092 .name = "oom_control",
5093 .seq_show = mem_cgroup_oom_control_read,
5094 .write_u64 = mem_cgroup_oom_control_write,
5095 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
5098 .name = "pressure_level",
5102 .name = "numa_stat",
5103 .seq_show = memcg_numa_stat_show,
5106 #ifdef CONFIG_MEMCG_KMEM
5108 .name = "kmem.limit_in_bytes",
5109 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
5110 .write = mem_cgroup_write,
5111 .read_u64 = mem_cgroup_read_u64,
5114 .name = "kmem.usage_in_bytes",
5115 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5116 .read_u64 = mem_cgroup_read_u64,
5119 .name = "kmem.failcnt",
5120 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5121 .write = mem_cgroup_reset,
5122 .read_u64 = mem_cgroup_read_u64,
5125 .name = "kmem.max_usage_in_bytes",
5126 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5127 .write = mem_cgroup_reset,
5128 .read_u64 = mem_cgroup_read_u64,
5130 #ifdef CONFIG_SLABINFO
5132 .name = "kmem.slabinfo",
5133 .seq_show = mem_cgroup_slabinfo_read,
5137 { }, /* terminate */
5140 #ifdef CONFIG_MEMCG_SWAP
5141 static struct cftype memsw_cgroup_files[] = {
5143 .name = "memsw.usage_in_bytes",
5144 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
5145 .read_u64 = mem_cgroup_read_u64,
5148 .name = "memsw.max_usage_in_bytes",
5149 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
5150 .write = mem_cgroup_reset,
5151 .read_u64 = mem_cgroup_read_u64,
5154 .name = "memsw.limit_in_bytes",
5155 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
5156 .write = mem_cgroup_write,
5157 .read_u64 = mem_cgroup_read_u64,
5160 .name = "memsw.failcnt",
5161 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
5162 .write = mem_cgroup_reset,
5163 .read_u64 = mem_cgroup_read_u64,
5165 { }, /* terminate */
5168 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
5170 struct mem_cgroup_per_node *pn;
5171 struct mem_cgroup_per_zone *mz;
5172 int zone, tmp = node;
5174 * This routine is called against possible nodes.
5175 * But it's BUG to call kmalloc() against offline node.
5177 * TODO: this routine can waste much memory for nodes which will
5178 * never be onlined. It's better to use memory hotplug callback
5181 if (!node_state(node, N_NORMAL_MEMORY))
5183 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5187 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5188 mz = &pn->zoneinfo[zone];
5189 lruvec_init(&mz->lruvec);
5190 mz->usage_in_excess = 0;
5191 mz->on_tree = false;
5194 memcg->nodeinfo[node] = pn;
5198 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
5200 kfree(memcg->nodeinfo[node]);
5203 static struct mem_cgroup *mem_cgroup_alloc(void)
5205 struct mem_cgroup *memcg;
5208 size = sizeof(struct mem_cgroup);
5209 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
5211 memcg = kzalloc(size, GFP_KERNEL);
5215 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
5218 spin_lock_init(&memcg->pcp_counter_lock);
5227 * At destroying mem_cgroup, references from swap_cgroup can remain.
5228 * (scanning all at force_empty is too costly...)
5230 * Instead of clearing all references at force_empty, we remember
5231 * the number of reference from swap_cgroup and free mem_cgroup when
5232 * it goes down to 0.
5234 * Removal of cgroup itself succeeds regardless of refs from swap.
5237 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5241 mem_cgroup_remove_from_trees(memcg);
5244 free_mem_cgroup_per_zone_info(memcg, node);
5246 free_percpu(memcg->stat);
5249 * We need to make sure that (at least for now), the jump label
5250 * destruction code runs outside of the cgroup lock. This is because
5251 * get_online_cpus(), which is called from the static_branch update,
5252 * can't be called inside the cgroup_lock. cpusets are the ones
5253 * enforcing this dependency, so if they ever change, we might as well.
5255 * schedule_work() will guarantee this happens. Be careful if you need
5256 * to move this code around, and make sure it is outside
5259 disarm_static_keys(memcg);
5264 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
5266 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
5268 if (!memcg->memory.parent)
5270 return mem_cgroup_from_counter(memcg->memory.parent, memory);
5272 EXPORT_SYMBOL(parent_mem_cgroup);
5274 static void __init mem_cgroup_soft_limit_tree_init(void)
5276 struct mem_cgroup_tree_per_node *rtpn;
5277 struct mem_cgroup_tree_per_zone *rtpz;
5278 int tmp, node, zone;
5280 for_each_node(node) {
5282 if (!node_state(node, N_NORMAL_MEMORY))
5284 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
5287 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5289 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5290 rtpz = &rtpn->rb_tree_per_zone[zone];
5291 rtpz->rb_root = RB_ROOT;
5292 spin_lock_init(&rtpz->lock);
5297 static struct cgroup_subsys_state * __ref
5298 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5300 struct mem_cgroup *memcg;
5301 long error = -ENOMEM;
5304 memcg = mem_cgroup_alloc();
5306 return ERR_PTR(error);
5309 if (alloc_mem_cgroup_per_zone_info(memcg, node))
5313 if (parent_css == NULL) {
5314 root_mem_cgroup = memcg;
5315 page_counter_init(&memcg->memory, NULL);
5316 page_counter_init(&memcg->memsw, NULL);
5317 page_counter_init(&memcg->kmem, NULL);
5320 memcg->last_scanned_node = MAX_NUMNODES;
5321 INIT_LIST_HEAD(&memcg->oom_notify);
5322 memcg->move_charge_at_immigrate = 0;
5323 mutex_init(&memcg->thresholds_lock);
5324 spin_lock_init(&memcg->move_lock);
5325 vmpressure_init(&memcg->vmpressure);
5326 INIT_LIST_HEAD(&memcg->event_list);
5327 spin_lock_init(&memcg->event_list_lock);
5332 __mem_cgroup_free(memcg);
5333 return ERR_PTR(error);
5337 mem_cgroup_css_online(struct cgroup_subsys_state *css)
5339 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5340 struct mem_cgroup *parent = mem_cgroup_from_css(css->parent);
5343 if (css->id > MEM_CGROUP_ID_MAX)
5349 mutex_lock(&memcg_create_mutex);
5351 memcg->use_hierarchy = parent->use_hierarchy;
5352 memcg->oom_kill_disable = parent->oom_kill_disable;
5353 memcg->swappiness = mem_cgroup_swappiness(parent);
5355 if (parent->use_hierarchy) {
5356 page_counter_init(&memcg->memory, &parent->memory);
5357 page_counter_init(&memcg->memsw, &parent->memsw);
5358 page_counter_init(&memcg->kmem, &parent->kmem);
5361 * No need to take a reference to the parent because cgroup
5362 * core guarantees its existence.
5365 page_counter_init(&memcg->memory, NULL);
5366 page_counter_init(&memcg->memsw, NULL);
5367 page_counter_init(&memcg->kmem, NULL);
5369 * Deeper hierachy with use_hierarchy == false doesn't make
5370 * much sense so let cgroup subsystem know about this
5371 * unfortunate state in our controller.
5373 if (parent != root_mem_cgroup)
5374 memory_cgrp_subsys.broken_hierarchy = true;
5376 mutex_unlock(&memcg_create_mutex);
5378 ret = memcg_init_kmem(memcg, &memory_cgrp_subsys);
5383 * Make sure the memcg is initialized: mem_cgroup_iter()
5384 * orders reading memcg->initialized against its callers
5385 * reading the memcg members.
5387 smp_store_release(&memcg->initialized, 1);
5392 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5394 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5395 struct mem_cgroup_event *event, *tmp;
5396 struct cgroup_subsys_state *iter;
5399 * Unregister events and notify userspace.
5400 * Notify userspace about cgroup removing only after rmdir of cgroup
5401 * directory to avoid race between userspace and kernelspace.
5403 spin_lock(&memcg->event_list_lock);
5404 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5405 list_del_init(&event->list);
5406 schedule_work(&event->remove);
5408 spin_unlock(&memcg->event_list_lock);
5410 kmem_cgroup_css_offline(memcg);
5413 * This requires that offlining is serialized. Right now that is
5414 * guaranteed because css_killed_work_fn() holds the cgroup_mutex.
5416 css_for_each_descendant_post(iter, css)
5417 mem_cgroup_reparent_charges(mem_cgroup_from_css(iter));
5419 memcg_unregister_all_caches(memcg);
5420 vmpressure_cleanup(&memcg->vmpressure);
5423 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5425 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5427 * XXX: css_offline() would be where we should reparent all
5428 * memory to prepare the cgroup for destruction. However,
5429 * memcg does not do css_tryget_online() and page_counter charging
5430 * under the same RCU lock region, which means that charging
5431 * could race with offlining. Offlining only happens to
5432 * cgroups with no tasks in them but charges can show up
5433 * without any tasks from the swapin path when the target
5434 * memcg is looked up from the swapout record and not from the
5435 * current task as it usually is. A race like this can leak
5436 * charges and put pages with stale cgroup pointers into
5440 * lookup_swap_cgroup_id()
5442 * mem_cgroup_lookup()
5443 * css_tryget_online()
5445 * disable css_tryget_online()
5448 * reparent_charges()
5449 * page_counter_try_charge()
5452 * pc->mem_cgroup = dead memcg
5455 * The bulk of the charges are still moved in offline_css() to
5456 * avoid pinning a lot of pages in case a long-term reference
5457 * like a swapout record is deferring the css_free() to long
5458 * after offlining. But this makes sure we catch any charges
5459 * made after offlining:
5461 mem_cgroup_reparent_charges(memcg);
5463 memcg_destroy_kmem(memcg);
5464 __mem_cgroup_free(memcg);
5468 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5469 * @css: the target css
5471 * Reset the states of the mem_cgroup associated with @css. This is
5472 * invoked when the userland requests disabling on the default hierarchy
5473 * but the memcg is pinned through dependency. The memcg should stop
5474 * applying policies and should revert to the vanilla state as it may be
5475 * made visible again.
5477 * The current implementation only resets the essential configurations.
5478 * This needs to be expanded to cover all the visible parts.
5480 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5482 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5484 mem_cgroup_resize_limit(memcg, PAGE_COUNTER_MAX);
5485 mem_cgroup_resize_memsw_limit(memcg, PAGE_COUNTER_MAX);
5486 memcg_update_kmem_limit(memcg, PAGE_COUNTER_MAX);
5487 memcg->soft_limit = 0;
5491 /* Handlers for move charge at task migration. */
5492 static int mem_cgroup_do_precharge(unsigned long count)
5496 /* Try a single bulk charge without reclaim first */
5497 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_WAIT, count);
5499 mc.precharge += count;
5502 if (ret == -EINTR) {
5503 cancel_charge(root_mem_cgroup, count);
5507 /* Try charges one by one with reclaim */
5509 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_NORETRY, 1);
5511 * In case of failure, any residual charges against
5512 * mc.to will be dropped by mem_cgroup_clear_mc()
5513 * later on. However, cancel any charges that are
5514 * bypassed to root right away or they'll be lost.
5517 cancel_charge(root_mem_cgroup, 1);
5527 * get_mctgt_type - get target type of moving charge
5528 * @vma: the vma the pte to be checked belongs
5529 * @addr: the address corresponding to the pte to be checked
5530 * @ptent: the pte to be checked
5531 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5534 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5535 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5536 * move charge. if @target is not NULL, the page is stored in target->page
5537 * with extra refcnt got(Callers should handle it).
5538 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5539 * target for charge migration. if @target is not NULL, the entry is stored
5542 * Called with pte lock held.
5549 enum mc_target_type {
5555 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5556 unsigned long addr, pte_t ptent)
5558 struct page *page = vm_normal_page(vma, addr, ptent);
5560 if (!page || !page_mapped(page))
5562 if (PageAnon(page)) {
5563 /* we don't move shared anon */
5566 } else if (!move_file())
5567 /* we ignore mapcount for file pages */
5569 if (!get_page_unless_zero(page))
5576 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5577 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5579 struct page *page = NULL;
5580 swp_entry_t ent = pte_to_swp_entry(ptent);
5582 if (!move_anon() || non_swap_entry(ent))
5585 * Because lookup_swap_cache() updates some statistics counter,
5586 * we call find_get_page() with swapper_space directly.
5588 page = find_get_page(swap_address_space(ent), ent.val);
5589 if (do_swap_account)
5590 entry->val = ent.val;
5595 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5596 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5602 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5603 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5605 struct page *page = NULL;
5606 struct address_space *mapping;
5609 if (!vma->vm_file) /* anonymous vma */
5614 mapping = vma->vm_file->f_mapping;
5615 if (pte_none(ptent))
5616 pgoff = linear_page_index(vma, addr);
5617 else /* pte_file(ptent) is true */
5618 pgoff = pte_to_pgoff(ptent);
5620 /* page is moved even if it's not RSS of this task(page-faulted). */
5622 /* shmem/tmpfs may report page out on swap: account for that too. */
5623 if (shmem_mapping(mapping)) {
5624 page = find_get_entry(mapping, pgoff);
5625 if (radix_tree_exceptional_entry(page)) {
5626 swp_entry_t swp = radix_to_swp_entry(page);
5627 if (do_swap_account)
5629 page = find_get_page(swap_address_space(swp), swp.val);
5632 page = find_get_page(mapping, pgoff);
5634 page = find_get_page(mapping, pgoff);
5639 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5640 unsigned long addr, pte_t ptent, union mc_target *target)
5642 struct page *page = NULL;
5643 struct page_cgroup *pc;
5644 enum mc_target_type ret = MC_TARGET_NONE;
5645 swp_entry_t ent = { .val = 0 };
5647 if (pte_present(ptent))
5648 page = mc_handle_present_pte(vma, addr, ptent);
5649 else if (is_swap_pte(ptent))
5650 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5651 else if (pte_none(ptent) || pte_file(ptent))
5652 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5654 if (!page && !ent.val)
5657 pc = lookup_page_cgroup(page);
5659 * Do only loose check w/o serialization.
5660 * mem_cgroup_move_account() checks the pc is valid or
5661 * not under LRU exclusion.
5663 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5664 ret = MC_TARGET_PAGE;
5666 target->page = page;
5668 if (!ret || !target)
5671 /* There is a swap entry and a page doesn't exist or isn't charged */
5672 if (ent.val && !ret &&
5673 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5674 ret = MC_TARGET_SWAP;
5681 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5683 * We don't consider swapping or file mapped pages because THP does not
5684 * support them for now.
5685 * Caller should make sure that pmd_trans_huge(pmd) is true.
5687 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5688 unsigned long addr, pmd_t pmd, union mc_target *target)
5690 struct page *page = NULL;
5691 struct page_cgroup *pc;
5692 enum mc_target_type ret = MC_TARGET_NONE;
5694 page = pmd_page(pmd);
5695 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5698 pc = lookup_page_cgroup(page);
5699 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5700 ret = MC_TARGET_PAGE;
5703 target->page = page;
5709 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5710 unsigned long addr, pmd_t pmd, union mc_target *target)
5712 return MC_TARGET_NONE;
5716 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5717 unsigned long addr, unsigned long end,
5718 struct mm_walk *walk)
5720 struct vm_area_struct *vma = walk->private;
5724 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
5725 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5726 mc.precharge += HPAGE_PMD_NR;
5731 if (pmd_trans_unstable(pmd))
5733 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5734 for (; addr != end; pte++, addr += PAGE_SIZE)
5735 if (get_mctgt_type(vma, addr, *pte, NULL))
5736 mc.precharge++; /* increment precharge temporarily */
5737 pte_unmap_unlock(pte - 1, ptl);
5743 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5745 unsigned long precharge;
5746 struct vm_area_struct *vma;
5748 down_read(&mm->mmap_sem);
5749 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5750 struct mm_walk mem_cgroup_count_precharge_walk = {
5751 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5755 if (is_vm_hugetlb_page(vma))
5757 walk_page_range(vma->vm_start, vma->vm_end,
5758 &mem_cgroup_count_precharge_walk);
5760 up_read(&mm->mmap_sem);
5762 precharge = mc.precharge;
5768 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5770 unsigned long precharge = mem_cgroup_count_precharge(mm);
5772 VM_BUG_ON(mc.moving_task);
5773 mc.moving_task = current;
5774 return mem_cgroup_do_precharge(precharge);
5777 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5778 static void __mem_cgroup_clear_mc(void)
5780 struct mem_cgroup *from = mc.from;
5781 struct mem_cgroup *to = mc.to;
5783 /* we must uncharge all the leftover precharges from mc.to */
5785 cancel_charge(mc.to, mc.precharge);
5789 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5790 * we must uncharge here.
5792 if (mc.moved_charge) {
5793 cancel_charge(mc.from, mc.moved_charge);
5794 mc.moved_charge = 0;
5796 /* we must fixup refcnts and charges */
5797 if (mc.moved_swap) {
5798 /* uncharge swap account from the old cgroup */
5799 if (!mem_cgroup_is_root(mc.from))
5800 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5803 * we charged both to->memory and to->memsw, so we
5804 * should uncharge to->memory.
5806 if (!mem_cgroup_is_root(mc.to))
5807 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5809 css_put_many(&mc.from->css, mc.moved_swap);
5811 /* we've already done css_get(mc.to) */
5814 memcg_oom_recover(from);
5815 memcg_oom_recover(to);
5816 wake_up_all(&mc.waitq);
5819 static void mem_cgroup_clear_mc(void)
5821 struct mem_cgroup *from = mc.from;
5824 * we must clear moving_task before waking up waiters at the end of
5827 mc.moving_task = NULL;
5828 __mem_cgroup_clear_mc();
5829 spin_lock(&mc.lock);
5832 spin_unlock(&mc.lock);
5833 mem_cgroup_end_move(from);
5836 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
5837 struct cgroup_taskset *tset)
5839 struct task_struct *p = cgroup_taskset_first(tset);
5841 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5842 unsigned long move_charge_at_immigrate;
5845 * We are now commited to this value whatever it is. Changes in this
5846 * tunable will only affect upcoming migrations, not the current one.
5847 * So we need to save it, and keep it going.
5849 move_charge_at_immigrate = memcg->move_charge_at_immigrate;
5850 if (move_charge_at_immigrate) {
5851 struct mm_struct *mm;
5852 struct mem_cgroup *from = mem_cgroup_from_task(p);
5854 VM_BUG_ON(from == memcg);
5856 mm = get_task_mm(p);
5859 /* We move charges only when we move a owner of the mm */
5860 if (mm->owner == p) {
5863 VM_BUG_ON(mc.precharge);
5864 VM_BUG_ON(mc.moved_charge);
5865 VM_BUG_ON(mc.moved_swap);
5866 mem_cgroup_start_move(from);
5867 spin_lock(&mc.lock);
5870 mc.immigrate_flags = move_charge_at_immigrate;
5871 spin_unlock(&mc.lock);
5872 /* We set mc.moving_task later */
5874 ret = mem_cgroup_precharge_mc(mm);
5876 mem_cgroup_clear_mc();
5883 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
5884 struct cgroup_taskset *tset)
5886 mem_cgroup_clear_mc();
5889 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5890 unsigned long addr, unsigned long end,
5891 struct mm_walk *walk)
5894 struct vm_area_struct *vma = walk->private;
5897 enum mc_target_type target_type;
5898 union mc_target target;
5900 struct page_cgroup *pc;
5903 * We don't take compound_lock() here but no race with splitting thp
5905 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
5906 * under splitting, which means there's no concurrent thp split,
5907 * - if another thread runs into split_huge_page() just after we
5908 * entered this if-block, the thread must wait for page table lock
5909 * to be unlocked in __split_huge_page_splitting(), where the main
5910 * part of thp split is not executed yet.
5912 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
5913 if (mc.precharge < HPAGE_PMD_NR) {
5917 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5918 if (target_type == MC_TARGET_PAGE) {
5920 if (!isolate_lru_page(page)) {
5921 pc = lookup_page_cgroup(page);
5922 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
5923 pc, mc.from, mc.to)) {
5924 mc.precharge -= HPAGE_PMD_NR;
5925 mc.moved_charge += HPAGE_PMD_NR;
5927 putback_lru_page(page);
5935 if (pmd_trans_unstable(pmd))
5938 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5939 for (; addr != end; addr += PAGE_SIZE) {
5940 pte_t ptent = *(pte++);
5946 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5947 case MC_TARGET_PAGE:
5949 if (isolate_lru_page(page))
5951 pc = lookup_page_cgroup(page);
5952 if (!mem_cgroup_move_account(page, 1, pc,
5955 /* we uncharge from mc.from later. */
5958 putback_lru_page(page);
5959 put: /* get_mctgt_type() gets the page */
5962 case MC_TARGET_SWAP:
5964 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5966 /* we fixup refcnts and charges later. */
5974 pte_unmap_unlock(pte - 1, ptl);
5979 * We have consumed all precharges we got in can_attach().
5980 * We try charge one by one, but don't do any additional
5981 * charges to mc.to if we have failed in charge once in attach()
5984 ret = mem_cgroup_do_precharge(1);
5992 static void mem_cgroup_move_charge(struct mm_struct *mm)
5994 struct vm_area_struct *vma;
5996 lru_add_drain_all();
5998 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
6000 * Someone who are holding the mmap_sem might be waiting in
6001 * waitq. So we cancel all extra charges, wake up all waiters,
6002 * and retry. Because we cancel precharges, we might not be able
6003 * to move enough charges, but moving charge is a best-effort
6004 * feature anyway, so it wouldn't be a big problem.
6006 __mem_cgroup_clear_mc();
6010 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6012 struct mm_walk mem_cgroup_move_charge_walk = {
6013 .pmd_entry = mem_cgroup_move_charge_pte_range,
6017 if (is_vm_hugetlb_page(vma))
6019 ret = walk_page_range(vma->vm_start, vma->vm_end,
6020 &mem_cgroup_move_charge_walk);
6023 * means we have consumed all precharges and failed in
6024 * doing additional charge. Just abandon here.
6028 up_read(&mm->mmap_sem);
6031 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
6032 struct cgroup_taskset *tset)
6034 struct task_struct *p = cgroup_taskset_first(tset);
6035 struct mm_struct *mm = get_task_mm(p);
6039 mem_cgroup_move_charge(mm);
6043 mem_cgroup_clear_mc();
6045 #else /* !CONFIG_MMU */
6046 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
6047 struct cgroup_taskset *tset)
6051 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
6052 struct cgroup_taskset *tset)
6055 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
6056 struct cgroup_taskset *tset)
6062 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6063 * to verify whether we're attached to the default hierarchy on each mount
6066 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
6069 * use_hierarchy is forced on the default hierarchy. cgroup core
6070 * guarantees that @root doesn't have any children, so turning it
6071 * on for the root memcg is enough.
6073 if (cgroup_on_dfl(root_css->cgroup))
6074 mem_cgroup_from_css(root_css)->use_hierarchy = true;
6077 struct cgroup_subsys memory_cgrp_subsys = {
6078 .css_alloc = mem_cgroup_css_alloc,
6079 .css_online = mem_cgroup_css_online,
6080 .css_offline = mem_cgroup_css_offline,
6081 .css_free = mem_cgroup_css_free,
6082 .css_reset = mem_cgroup_css_reset,
6083 .can_attach = mem_cgroup_can_attach,
6084 .cancel_attach = mem_cgroup_cancel_attach,
6085 .attach = mem_cgroup_move_task,
6086 .bind = mem_cgroup_bind,
6087 .legacy_cftypes = mem_cgroup_files,
6091 #ifdef CONFIG_MEMCG_SWAP
6092 static int __init enable_swap_account(char *s)
6094 if (!strcmp(s, "1"))
6095 really_do_swap_account = 1;
6096 else if (!strcmp(s, "0"))
6097 really_do_swap_account = 0;
6100 __setup("swapaccount=", enable_swap_account);
6102 static void __init memsw_file_init(void)
6104 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
6105 memsw_cgroup_files));
6108 static void __init enable_swap_cgroup(void)
6110 if (!mem_cgroup_disabled() && really_do_swap_account) {
6111 do_swap_account = 1;
6117 static void __init enable_swap_cgroup(void)
6122 #ifdef CONFIG_MEMCG_SWAP
6124 * mem_cgroup_swapout - transfer a memsw charge to swap
6125 * @page: page whose memsw charge to transfer
6126 * @entry: swap entry to move the charge to
6128 * Transfer the memsw charge of @page to @entry.
6130 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
6132 struct page_cgroup *pc;
6133 unsigned short oldid;
6135 VM_BUG_ON_PAGE(PageLRU(page), page);
6136 VM_BUG_ON_PAGE(page_count(page), page);
6138 if (!do_swap_account)
6141 pc = lookup_page_cgroup(page);
6143 /* Readahead page, never charged */
6144 if (!PageCgroupUsed(pc))
6147 VM_BUG_ON_PAGE(!(pc->flags & PCG_MEMSW), page);
6149 oldid = swap_cgroup_record(entry, mem_cgroup_id(pc->mem_cgroup));
6150 VM_BUG_ON_PAGE(oldid, page);
6152 pc->flags &= ~PCG_MEMSW;
6153 css_get(&pc->mem_cgroup->css);
6154 mem_cgroup_swap_statistics(pc->mem_cgroup, true);
6158 * mem_cgroup_uncharge_swap - uncharge a swap entry
6159 * @entry: swap entry to uncharge
6161 * Drop the memsw charge associated with @entry.
6163 void mem_cgroup_uncharge_swap(swp_entry_t entry)
6165 struct mem_cgroup *memcg;
6168 if (!do_swap_account)
6171 id = swap_cgroup_record(entry, 0);
6173 memcg = mem_cgroup_lookup(id);
6175 if (!mem_cgroup_is_root(memcg))
6176 page_counter_uncharge(&memcg->memsw, 1);
6177 mem_cgroup_swap_statistics(memcg, false);
6178 css_put(&memcg->css);
6185 * mem_cgroup_try_charge - try charging a page
6186 * @page: page to charge
6187 * @mm: mm context of the victim
6188 * @gfp_mask: reclaim mode
6189 * @memcgp: charged memcg return
6191 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6192 * pages according to @gfp_mask if necessary.
6194 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
6195 * Otherwise, an error code is returned.
6197 * After page->mapping has been set up, the caller must finalize the
6198 * charge with mem_cgroup_commit_charge(). Or abort the transaction
6199 * with mem_cgroup_cancel_charge() in case page instantiation fails.
6201 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
6202 gfp_t gfp_mask, struct mem_cgroup **memcgp)
6204 struct mem_cgroup *memcg = NULL;
6205 unsigned int nr_pages = 1;
6208 if (mem_cgroup_disabled())
6211 if (PageSwapCache(page)) {
6212 struct page_cgroup *pc = lookup_page_cgroup(page);
6214 * Every swap fault against a single page tries to charge the
6215 * page, bail as early as possible. shmem_unuse() encounters
6216 * already charged pages, too. The USED bit is protected by
6217 * the page lock, which serializes swap cache removal, which
6218 * in turn serializes uncharging.
6220 if (PageCgroupUsed(pc))
6224 if (PageTransHuge(page)) {
6225 nr_pages <<= compound_order(page);
6226 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
6229 if (do_swap_account && PageSwapCache(page))
6230 memcg = try_get_mem_cgroup_from_page(page);
6232 memcg = get_mem_cgroup_from_mm(mm);
6234 ret = try_charge(memcg, gfp_mask, nr_pages);
6236 css_put(&memcg->css);
6238 if (ret == -EINTR) {
6239 memcg = root_mem_cgroup;
6248 * mem_cgroup_commit_charge - commit a page charge
6249 * @page: page to charge
6250 * @memcg: memcg to charge the page to
6251 * @lrucare: page might be on LRU already
6253 * Finalize a charge transaction started by mem_cgroup_try_charge(),
6254 * after page->mapping has been set up. This must happen atomically
6255 * as part of the page instantiation, i.e. under the page table lock
6256 * for anonymous pages, under the page lock for page and swap cache.
6258 * In addition, the page must not be on the LRU during the commit, to
6259 * prevent racing with task migration. If it might be, use @lrucare.
6261 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
6263 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
6266 unsigned int nr_pages = 1;
6268 VM_BUG_ON_PAGE(!page->mapping, page);
6269 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
6271 if (mem_cgroup_disabled())
6274 * Swap faults will attempt to charge the same page multiple
6275 * times. But reuse_swap_page() might have removed the page
6276 * from swapcache already, so we can't check PageSwapCache().
6281 commit_charge(page, memcg, lrucare);
6283 if (PageTransHuge(page)) {
6284 nr_pages <<= compound_order(page);
6285 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
6288 local_irq_disable();
6289 mem_cgroup_charge_statistics(memcg, page, nr_pages);
6290 memcg_check_events(memcg, page);
6293 if (do_swap_account && PageSwapCache(page)) {
6294 swp_entry_t entry = { .val = page_private(page) };
6296 * The swap entry might not get freed for a long time,
6297 * let's not wait for it. The page already received a
6298 * memory+swap charge, drop the swap entry duplicate.
6300 mem_cgroup_uncharge_swap(entry);
6305 * mem_cgroup_cancel_charge - cancel a page charge
6306 * @page: page to charge
6307 * @memcg: memcg to charge the page to
6309 * Cancel a charge transaction started by mem_cgroup_try_charge().
6311 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg)
6313 unsigned int nr_pages = 1;
6315 if (mem_cgroup_disabled())
6318 * Swap faults will attempt to charge the same page multiple
6319 * times. But reuse_swap_page() might have removed the page
6320 * from swapcache already, so we can't check PageSwapCache().
6325 if (PageTransHuge(page)) {
6326 nr_pages <<= compound_order(page);
6327 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
6330 cancel_charge(memcg, nr_pages);
6333 static void uncharge_batch(struct mem_cgroup *memcg, unsigned long pgpgout,
6334 unsigned long nr_mem, unsigned long nr_memsw,
6335 unsigned long nr_anon, unsigned long nr_file,
6336 unsigned long nr_huge, struct page *dummy_page)
6338 unsigned long flags;
6340 if (!mem_cgroup_is_root(memcg)) {
6342 page_counter_uncharge(&memcg->memory, nr_mem);
6344 page_counter_uncharge(&memcg->memsw, nr_memsw);
6345 memcg_oom_recover(memcg);
6348 local_irq_save(flags);
6349 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS], nr_anon);
6350 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_CACHE], nr_file);
6351 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE], nr_huge);
6352 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT], pgpgout);
6353 __this_cpu_add(memcg->stat->nr_page_events, nr_anon + nr_file);
6354 memcg_check_events(memcg, dummy_page);
6355 local_irq_restore(flags);
6357 if (!mem_cgroup_is_root(memcg))
6358 css_put_many(&memcg->css, max(nr_mem, nr_memsw));
6361 static void uncharge_list(struct list_head *page_list)
6363 struct mem_cgroup *memcg = NULL;
6364 unsigned long nr_memsw = 0;
6365 unsigned long nr_anon = 0;
6366 unsigned long nr_file = 0;
6367 unsigned long nr_huge = 0;
6368 unsigned long pgpgout = 0;
6369 unsigned long nr_mem = 0;
6370 struct list_head *next;
6373 next = page_list->next;
6375 unsigned int nr_pages = 1;
6376 struct page_cgroup *pc;
6378 page = list_entry(next, struct page, lru);
6379 next = page->lru.next;
6381 VM_BUG_ON_PAGE(PageLRU(page), page);
6382 VM_BUG_ON_PAGE(page_count(page), page);
6384 pc = lookup_page_cgroup(page);
6385 if (!PageCgroupUsed(pc))
6389 * Nobody should be changing or seriously looking at
6390 * pc->mem_cgroup and pc->flags at this point, we have
6391 * fully exclusive access to the page.
6394 if (memcg != pc->mem_cgroup) {
6396 uncharge_batch(memcg, pgpgout, nr_mem, nr_memsw,
6397 nr_anon, nr_file, nr_huge, page);
6398 pgpgout = nr_mem = nr_memsw = 0;
6399 nr_anon = nr_file = nr_huge = 0;
6401 memcg = pc->mem_cgroup;
6404 if (PageTransHuge(page)) {
6405 nr_pages <<= compound_order(page);
6406 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
6407 nr_huge += nr_pages;
6411 nr_anon += nr_pages;
6413 nr_file += nr_pages;
6415 if (pc->flags & PCG_MEM)
6417 if (pc->flags & PCG_MEMSW)
6418 nr_memsw += nr_pages;
6422 } while (next != page_list);
6425 uncharge_batch(memcg, pgpgout, nr_mem, nr_memsw,
6426 nr_anon, nr_file, nr_huge, page);
6430 * mem_cgroup_uncharge - uncharge a page
6431 * @page: page to uncharge
6433 * Uncharge a page previously charged with mem_cgroup_try_charge() and
6434 * mem_cgroup_commit_charge().
6436 void mem_cgroup_uncharge(struct page *page)
6438 struct page_cgroup *pc;
6440 if (mem_cgroup_disabled())
6443 /* Don't touch page->lru of any random page, pre-check: */
6444 pc = lookup_page_cgroup(page);
6445 if (!PageCgroupUsed(pc))
6448 INIT_LIST_HEAD(&page->lru);
6449 uncharge_list(&page->lru);
6453 * mem_cgroup_uncharge_list - uncharge a list of page
6454 * @page_list: list of pages to uncharge
6456 * Uncharge a list of pages previously charged with
6457 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
6459 void mem_cgroup_uncharge_list(struct list_head *page_list)
6461 if (mem_cgroup_disabled())
6464 if (!list_empty(page_list))
6465 uncharge_list(page_list);
6469 * mem_cgroup_migrate - migrate a charge to another page
6470 * @oldpage: currently charged page
6471 * @newpage: page to transfer the charge to
6472 * @lrucare: both pages might be on the LRU already
6474 * Migrate the charge from @oldpage to @newpage.
6476 * Both pages must be locked, @newpage->mapping must be set up.
6478 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage,
6481 struct page_cgroup *pc;
6484 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6485 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6486 VM_BUG_ON_PAGE(!lrucare && PageLRU(oldpage), oldpage);
6487 VM_BUG_ON_PAGE(!lrucare && PageLRU(newpage), newpage);
6488 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6489 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6492 if (mem_cgroup_disabled())
6495 /* Page cache replacement: new page already charged? */
6496 pc = lookup_page_cgroup(newpage);
6497 if (PageCgroupUsed(pc))
6500 /* Re-entrant migration: old page already uncharged? */
6501 pc = lookup_page_cgroup(oldpage);
6502 if (!PageCgroupUsed(pc))
6505 VM_BUG_ON_PAGE(!(pc->flags & PCG_MEM), oldpage);
6506 VM_BUG_ON_PAGE(do_swap_account && !(pc->flags & PCG_MEMSW), oldpage);
6509 lock_page_lru(oldpage, &isolated);
6514 unlock_page_lru(oldpage, isolated);
6516 commit_charge(newpage, pc->mem_cgroup, lrucare);
6520 * subsys_initcall() for memory controller.
6522 * Some parts like hotcpu_notifier() have to be initialized from this context
6523 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
6524 * everything that doesn't depend on a specific mem_cgroup structure should
6525 * be initialized from here.
6527 static int __init mem_cgroup_init(void)
6529 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
6530 enable_swap_cgroup();
6531 mem_cgroup_soft_limit_tree_init();
6535 subsys_initcall(mem_cgroup_init);