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 stock->nr_pages = 0;
2278 stock->cached = NULL;
2282 * This must be called under preempt disabled or must be called by
2283 * a thread which is pinned to local cpu.
2285 static void drain_local_stock(struct work_struct *dummy)
2287 struct memcg_stock_pcp *stock = this_cpu_ptr(&memcg_stock);
2289 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2292 static void __init memcg_stock_init(void)
2296 for_each_possible_cpu(cpu) {
2297 struct memcg_stock_pcp *stock =
2298 &per_cpu(memcg_stock, cpu);
2299 INIT_WORK(&stock->work, drain_local_stock);
2304 * Cache charges(val) to local per_cpu area.
2305 * This will be consumed by consume_stock() function, later.
2307 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2309 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2311 if (stock->cached != memcg) { /* reset if necessary */
2313 stock->cached = memcg;
2315 stock->nr_pages += nr_pages;
2316 put_cpu_var(memcg_stock);
2320 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2321 * of the hierarchy under it. sync flag says whether we should block
2322 * until the work is done.
2324 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2328 /* Notify other cpus that system-wide "drain" is running */
2331 for_each_online_cpu(cpu) {
2332 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2333 struct mem_cgroup *memcg;
2335 memcg = stock->cached;
2336 if (!memcg || !stock->nr_pages)
2338 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2340 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2342 drain_local_stock(&stock->work);
2344 schedule_work_on(cpu, &stock->work);
2352 for_each_online_cpu(cpu) {
2353 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2354 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2355 flush_work(&stock->work);
2362 * Tries to drain stocked charges in other cpus. This function is asynchronous
2363 * and just put a work per cpu for draining localy on each cpu. Caller can
2364 * expects some charges will be back later but cannot wait for it.
2366 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2369 * If someone calls draining, avoid adding more kworker runs.
2371 if (!mutex_trylock(&percpu_charge_mutex))
2373 drain_all_stock(root_memcg, false);
2374 mutex_unlock(&percpu_charge_mutex);
2377 /* This is a synchronous drain interface. */
2378 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2380 /* called when force_empty is called */
2381 mutex_lock(&percpu_charge_mutex);
2382 drain_all_stock(root_memcg, true);
2383 mutex_unlock(&percpu_charge_mutex);
2387 * This function drains percpu counter value from DEAD cpu and
2388 * move it to local cpu. Note that this function can be preempted.
2390 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2394 spin_lock(&memcg->pcp_counter_lock);
2395 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2396 long x = per_cpu(memcg->stat->count[i], cpu);
2398 per_cpu(memcg->stat->count[i], cpu) = 0;
2399 memcg->nocpu_base.count[i] += x;
2401 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2402 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2404 per_cpu(memcg->stat->events[i], cpu) = 0;
2405 memcg->nocpu_base.events[i] += x;
2407 spin_unlock(&memcg->pcp_counter_lock);
2410 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
2411 unsigned long action,
2414 int cpu = (unsigned long)hcpu;
2415 struct memcg_stock_pcp *stock;
2416 struct mem_cgroup *iter;
2418 if (action == CPU_ONLINE)
2421 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2424 for_each_mem_cgroup(iter)
2425 mem_cgroup_drain_pcp_counter(iter, cpu);
2427 stock = &per_cpu(memcg_stock, cpu);
2432 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2433 unsigned int nr_pages)
2435 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2436 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2437 struct mem_cgroup *mem_over_limit;
2438 struct page_counter *counter;
2439 unsigned long nr_reclaimed;
2440 bool may_swap = true;
2441 bool drained = false;
2444 if (mem_cgroup_is_root(memcg))
2447 if (consume_stock(memcg, nr_pages))
2450 if (!do_swap_account ||
2451 !page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2452 if (!page_counter_try_charge(&memcg->memory, batch, &counter))
2454 if (do_swap_account)
2455 page_counter_uncharge(&memcg->memsw, batch);
2456 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2458 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2462 if (batch > nr_pages) {
2468 * Unlike in global OOM situations, memcg is not in a physical
2469 * memory shortage. Allow dying and OOM-killed tasks to
2470 * bypass the last charges so that they can exit quickly and
2471 * free their memory.
2473 if (unlikely(test_thread_flag(TIF_MEMDIE) ||
2474 fatal_signal_pending(current) ||
2475 current->flags & PF_EXITING))
2478 if (unlikely(task_in_memcg_oom(current)))
2481 if (!(gfp_mask & __GFP_WAIT))
2484 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2485 gfp_mask, may_swap);
2487 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2491 drain_all_stock_async(mem_over_limit);
2496 if (gfp_mask & __GFP_NORETRY)
2499 * Even though the limit is exceeded at this point, reclaim
2500 * may have been able to free some pages. Retry the charge
2501 * before killing the task.
2503 * Only for regular pages, though: huge pages are rather
2504 * unlikely to succeed so close to the limit, and we fall back
2505 * to regular pages anyway in case of failure.
2507 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2510 * At task move, charge accounts can be doubly counted. So, it's
2511 * better to wait until the end of task_move if something is going on.
2513 if (mem_cgroup_wait_acct_move(mem_over_limit))
2519 if (gfp_mask & __GFP_NOFAIL)
2522 if (fatal_signal_pending(current))
2525 mem_cgroup_oom(mem_over_limit, gfp_mask, get_order(nr_pages));
2527 if (!(gfp_mask & __GFP_NOFAIL))
2533 if (batch > nr_pages)
2534 refill_stock(memcg, batch - nr_pages);
2539 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2541 if (mem_cgroup_is_root(memcg))
2544 page_counter_uncharge(&memcg->memory, nr_pages);
2545 if (do_swap_account)
2546 page_counter_uncharge(&memcg->memsw, nr_pages);
2550 * A helper function to get mem_cgroup from ID. must be called under
2551 * rcu_read_lock(). The caller is responsible for calling
2552 * css_tryget_online() if the mem_cgroup is used for charging. (dropping
2553 * refcnt from swap can be called against removed memcg.)
2555 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2557 /* ID 0 is unused ID */
2560 return mem_cgroup_from_id(id);
2564 * try_get_mem_cgroup_from_page - look up page's memcg association
2567 * Look up, get a css reference, and return the memcg that owns @page.
2569 * The page must be locked to prevent racing with swap-in and page
2570 * cache charges. If coming from an unlocked page table, the caller
2571 * must ensure the page is on the LRU or this can race with charging.
2573 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2575 struct mem_cgroup *memcg = NULL;
2576 struct page_cgroup *pc;
2580 VM_BUG_ON_PAGE(!PageLocked(page), page);
2582 pc = lookup_page_cgroup(page);
2583 if (PageCgroupUsed(pc)) {
2584 memcg = pc->mem_cgroup;
2585 if (memcg && !css_tryget_online(&memcg->css))
2587 } else if (PageSwapCache(page)) {
2588 ent.val = page_private(page);
2589 id = lookup_swap_cgroup_id(ent);
2591 memcg = mem_cgroup_lookup(id);
2592 if (memcg && !css_tryget_online(&memcg->css))
2599 static void lock_page_lru(struct page *page, int *isolated)
2601 struct zone *zone = page_zone(page);
2603 spin_lock_irq(&zone->lru_lock);
2604 if (PageLRU(page)) {
2605 struct lruvec *lruvec;
2607 lruvec = mem_cgroup_page_lruvec(page, zone);
2609 del_page_from_lru_list(page, lruvec, page_lru(page));
2615 static void unlock_page_lru(struct page *page, int isolated)
2617 struct zone *zone = page_zone(page);
2620 struct lruvec *lruvec;
2622 lruvec = mem_cgroup_page_lruvec(page, zone);
2623 VM_BUG_ON_PAGE(PageLRU(page), page);
2625 add_page_to_lru_list(page, lruvec, page_lru(page));
2627 spin_unlock_irq(&zone->lru_lock);
2630 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2633 struct page_cgroup *pc = lookup_page_cgroup(page);
2636 VM_BUG_ON_PAGE(PageCgroupUsed(pc), page);
2638 * we don't need page_cgroup_lock about tail pages, becase they are not
2639 * accessed by any other context at this point.
2643 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2644 * may already be on some other mem_cgroup's LRU. Take care of it.
2647 lock_page_lru(page, &isolated);
2650 * Nobody should be changing or seriously looking at
2651 * pc->mem_cgroup and pc->flags at this point:
2653 * - the page is uncharged
2655 * - the page is off-LRU
2657 * - an anonymous fault has exclusive page access, except for
2658 * a locked page table
2660 * - a page cache insertion, a swapin fault, or a migration
2661 * have the page locked
2663 pc->mem_cgroup = memcg;
2664 pc->flags = PCG_USED | PCG_MEM | (do_swap_account ? PCG_MEMSW : 0);
2667 unlock_page_lru(page, isolated);
2670 #ifdef CONFIG_MEMCG_KMEM
2672 * The memcg_slab_mutex is held whenever a per memcg kmem cache is created or
2673 * destroyed. It protects memcg_caches arrays and memcg_slab_caches lists.
2675 static DEFINE_MUTEX(memcg_slab_mutex);
2677 static DEFINE_MUTEX(activate_kmem_mutex);
2680 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2681 * in the memcg_cache_params struct.
2683 static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
2685 struct kmem_cache *cachep;
2687 VM_BUG_ON(p->is_root_cache);
2688 cachep = p->root_cache;
2689 return cache_from_memcg_idx(cachep, memcg_cache_id(p->memcg));
2692 #ifdef CONFIG_SLABINFO
2693 static int mem_cgroup_slabinfo_read(struct seq_file *m, void *v)
2695 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
2696 struct memcg_cache_params *params;
2698 if (!memcg_kmem_is_active(memcg))
2701 print_slabinfo_header(m);
2703 mutex_lock(&memcg_slab_mutex);
2704 list_for_each_entry(params, &memcg->memcg_slab_caches, list)
2705 cache_show(memcg_params_to_cache(params), m);
2706 mutex_unlock(&memcg_slab_mutex);
2712 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp,
2713 unsigned long nr_pages)
2715 struct page_counter *counter;
2718 ret = page_counter_try_charge(&memcg->kmem, nr_pages, &counter);
2722 ret = try_charge(memcg, gfp, nr_pages);
2723 if (ret == -EINTR) {
2725 * try_charge() chose to bypass to root due to OOM kill or
2726 * fatal signal. Since our only options are to either fail
2727 * the allocation or charge it to this cgroup, do it as a
2728 * temporary condition. But we can't fail. From a kmem/slab
2729 * perspective, the cache has already been selected, by
2730 * mem_cgroup_kmem_get_cache(), so it is too late to change
2733 * This condition will only trigger if the task entered
2734 * memcg_charge_kmem in a sane state, but was OOM-killed
2735 * during try_charge() above. Tasks that were already dying
2736 * when the allocation triggers should have been already
2737 * directed to the root cgroup in memcontrol.h
2739 page_counter_charge(&memcg->memory, nr_pages);
2740 if (do_swap_account)
2741 page_counter_charge(&memcg->memsw, nr_pages);
2744 page_counter_uncharge(&memcg->kmem, nr_pages);
2749 static void memcg_uncharge_kmem(struct mem_cgroup *memcg,
2750 unsigned long nr_pages)
2752 page_counter_uncharge(&memcg->memory, nr_pages);
2753 if (do_swap_account)
2754 page_counter_uncharge(&memcg->memsw, nr_pages);
2757 if (page_counter_uncharge(&memcg->kmem, nr_pages))
2761 * Releases a reference taken in kmem_cgroup_css_offline in case
2762 * this last uncharge is racing with the offlining code or it is
2763 * outliving the memcg existence.
2765 * The memory barrier imposed by test&clear is paired with the
2766 * explicit one in memcg_kmem_mark_dead().
2768 if (memcg_kmem_test_and_clear_dead(memcg))
2769 css_put(&memcg->css);
2773 * helper for acessing a memcg's index. It will be used as an index in the
2774 * child cache array in kmem_cache, and also to derive its name. This function
2775 * will return -1 when this is not a kmem-limited memcg.
2777 int memcg_cache_id(struct mem_cgroup *memcg)
2779 return memcg ? memcg->kmemcg_id : -1;
2782 static int memcg_alloc_cache_id(void)
2787 id = ida_simple_get(&kmem_limited_groups,
2788 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2792 if (id < memcg_limited_groups_array_size)
2796 * There's no space for the new id in memcg_caches arrays,
2797 * so we have to grow them.
2800 size = 2 * (id + 1);
2801 if (size < MEMCG_CACHES_MIN_SIZE)
2802 size = MEMCG_CACHES_MIN_SIZE;
2803 else if (size > MEMCG_CACHES_MAX_SIZE)
2804 size = MEMCG_CACHES_MAX_SIZE;
2806 mutex_lock(&memcg_slab_mutex);
2807 err = memcg_update_all_caches(size);
2808 mutex_unlock(&memcg_slab_mutex);
2811 ida_simple_remove(&kmem_limited_groups, id);
2817 static void memcg_free_cache_id(int id)
2819 ida_simple_remove(&kmem_limited_groups, id);
2823 * We should update the current array size iff all caches updates succeed. This
2824 * can only be done from the slab side. The slab mutex needs to be held when
2827 void memcg_update_array_size(int num)
2829 memcg_limited_groups_array_size = num;
2832 static void memcg_register_cache(struct mem_cgroup *memcg,
2833 struct kmem_cache *root_cache)
2835 static char memcg_name_buf[NAME_MAX + 1]; /* protected by
2837 struct kmem_cache *cachep;
2840 lockdep_assert_held(&memcg_slab_mutex);
2842 id = memcg_cache_id(memcg);
2845 * Since per-memcg caches are created asynchronously on first
2846 * allocation (see memcg_kmem_get_cache()), several threads can try to
2847 * create the same cache, but only one of them may succeed.
2849 if (cache_from_memcg_idx(root_cache, id))
2852 cgroup_name(memcg->css.cgroup, memcg_name_buf, NAME_MAX + 1);
2853 cachep = memcg_create_kmem_cache(memcg, root_cache, memcg_name_buf);
2855 * If we could not create a memcg cache, do not complain, because
2856 * that's not critical at all as we can always proceed with the root
2862 css_get(&memcg->css);
2863 list_add(&cachep->memcg_params->list, &memcg->memcg_slab_caches);
2866 * Since readers won't lock (see cache_from_memcg_idx()), we need a
2867 * barrier here to ensure nobody will see the kmem_cache partially
2872 BUG_ON(root_cache->memcg_params->memcg_caches[id]);
2873 root_cache->memcg_params->memcg_caches[id] = cachep;
2876 static void memcg_unregister_cache(struct kmem_cache *cachep)
2878 struct kmem_cache *root_cache;
2879 struct mem_cgroup *memcg;
2882 lockdep_assert_held(&memcg_slab_mutex);
2884 BUG_ON(is_root_cache(cachep));
2886 root_cache = cachep->memcg_params->root_cache;
2887 memcg = cachep->memcg_params->memcg;
2888 id = memcg_cache_id(memcg);
2890 BUG_ON(root_cache->memcg_params->memcg_caches[id] != cachep);
2891 root_cache->memcg_params->memcg_caches[id] = NULL;
2893 list_del(&cachep->memcg_params->list);
2895 kmem_cache_destroy(cachep);
2897 /* drop the reference taken in memcg_register_cache */
2898 css_put(&memcg->css);
2902 * During the creation a new cache, we need to disable our accounting mechanism
2903 * altogether. This is true even if we are not creating, but rather just
2904 * enqueing new caches to be created.
2906 * This is because that process will trigger allocations; some visible, like
2907 * explicit kmallocs to auxiliary data structures, name strings and internal
2908 * cache structures; some well concealed, like INIT_WORK() that can allocate
2909 * objects during debug.
2911 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
2912 * to it. This may not be a bounded recursion: since the first cache creation
2913 * failed to complete (waiting on the allocation), we'll just try to create the
2914 * cache again, failing at the same point.
2916 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
2917 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
2918 * inside the following two functions.
2920 static inline void memcg_stop_kmem_account(void)
2922 VM_BUG_ON(!current->mm);
2923 current->memcg_kmem_skip_account++;
2926 static inline void memcg_resume_kmem_account(void)
2928 VM_BUG_ON(!current->mm);
2929 current->memcg_kmem_skip_account--;
2932 int __memcg_cleanup_cache_params(struct kmem_cache *s)
2934 struct kmem_cache *c;
2937 mutex_lock(&memcg_slab_mutex);
2938 for_each_memcg_cache_index(i) {
2939 c = cache_from_memcg_idx(s, i);
2943 memcg_unregister_cache(c);
2945 if (cache_from_memcg_idx(s, i))
2948 mutex_unlock(&memcg_slab_mutex);
2952 static void memcg_unregister_all_caches(struct mem_cgroup *memcg)
2954 struct kmem_cache *cachep;
2955 struct memcg_cache_params *params, *tmp;
2957 if (!memcg_kmem_is_active(memcg))
2960 mutex_lock(&memcg_slab_mutex);
2961 list_for_each_entry_safe(params, tmp, &memcg->memcg_slab_caches, list) {
2962 cachep = memcg_params_to_cache(params);
2963 kmem_cache_shrink(cachep);
2964 if (atomic_read(&cachep->memcg_params->nr_pages) == 0)
2965 memcg_unregister_cache(cachep);
2967 mutex_unlock(&memcg_slab_mutex);
2970 struct memcg_register_cache_work {
2971 struct mem_cgroup *memcg;
2972 struct kmem_cache *cachep;
2973 struct work_struct work;
2976 static void memcg_register_cache_func(struct work_struct *w)
2978 struct memcg_register_cache_work *cw =
2979 container_of(w, struct memcg_register_cache_work, work);
2980 struct mem_cgroup *memcg = cw->memcg;
2981 struct kmem_cache *cachep = cw->cachep;
2983 mutex_lock(&memcg_slab_mutex);
2984 memcg_register_cache(memcg, cachep);
2985 mutex_unlock(&memcg_slab_mutex);
2987 css_put(&memcg->css);
2992 * Enqueue the creation of a per-memcg kmem_cache.
2994 static void __memcg_schedule_register_cache(struct mem_cgroup *memcg,
2995 struct kmem_cache *cachep)
2997 struct memcg_register_cache_work *cw;
2999 cw = kmalloc(sizeof(*cw), GFP_NOWAIT);
3001 css_put(&memcg->css);
3006 cw->cachep = cachep;
3008 INIT_WORK(&cw->work, memcg_register_cache_func);
3009 schedule_work(&cw->work);
3012 static void memcg_schedule_register_cache(struct mem_cgroup *memcg,
3013 struct kmem_cache *cachep)
3016 * We need to stop accounting when we kmalloc, because if the
3017 * corresponding kmalloc cache is not yet created, the first allocation
3018 * in __memcg_schedule_register_cache will recurse.
3020 * However, it is better to enclose the whole function. Depending on
3021 * the debugging options enabled, INIT_WORK(), for instance, can
3022 * trigger an allocation. This too, will make us recurse. Because at
3023 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3024 * the safest choice is to do it like this, wrapping the whole function.
3026 memcg_stop_kmem_account();
3027 __memcg_schedule_register_cache(memcg, cachep);
3028 memcg_resume_kmem_account();
3031 int __memcg_charge_slab(struct kmem_cache *cachep, gfp_t gfp, int order)
3033 unsigned int nr_pages = 1 << order;
3036 res = memcg_charge_kmem(cachep->memcg_params->memcg, gfp, nr_pages);
3038 atomic_add(nr_pages, &cachep->memcg_params->nr_pages);
3042 void __memcg_uncharge_slab(struct kmem_cache *cachep, int order)
3044 unsigned int nr_pages = 1 << order;
3046 memcg_uncharge_kmem(cachep->memcg_params->memcg, nr_pages);
3047 atomic_sub(nr_pages, &cachep->memcg_params->nr_pages);
3051 * Return the kmem_cache we're supposed to use for a slab allocation.
3052 * We try to use the current memcg's version of the cache.
3054 * If the cache does not exist yet, if we are the first user of it,
3055 * we either create it immediately, if possible, or create it asynchronously
3057 * In the latter case, we will let the current allocation go through with
3058 * the original cache.
3060 * Can't be called in interrupt context or from kernel threads.
3061 * This function needs to be called with rcu_read_lock() held.
3063 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
3066 struct mem_cgroup *memcg;
3067 struct kmem_cache *memcg_cachep;
3069 VM_BUG_ON(!cachep->memcg_params);
3070 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
3072 if (!current->mm || current->memcg_kmem_skip_account)
3076 memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
3078 if (!memcg_kmem_is_active(memcg))
3081 memcg_cachep = cache_from_memcg_idx(cachep, memcg_cache_id(memcg));
3082 if (likely(memcg_cachep)) {
3083 cachep = memcg_cachep;
3087 /* The corresponding put will be done in the workqueue. */
3088 if (!css_tryget_online(&memcg->css))
3093 * If we are in a safe context (can wait, and not in interrupt
3094 * context), we could be be predictable and return right away.
3095 * This would guarantee that the allocation being performed
3096 * already belongs in the new cache.
3098 * However, there are some clashes that can arrive from locking.
3099 * For instance, because we acquire the slab_mutex while doing
3100 * memcg_create_kmem_cache, this means no further allocation
3101 * could happen with the slab_mutex held. So it's better to
3104 memcg_schedule_register_cache(memcg, cachep);
3112 * We need to verify if the allocation against current->mm->owner's memcg is
3113 * possible for the given order. But the page is not allocated yet, so we'll
3114 * need a further commit step to do the final arrangements.
3116 * It is possible for the task to switch cgroups in this mean time, so at
3117 * commit time, we can't rely on task conversion any longer. We'll then use
3118 * the handle argument to return to the caller which cgroup we should commit
3119 * against. We could also return the memcg directly and avoid the pointer
3120 * passing, but a boolean return value gives better semantics considering
3121 * the compiled-out case as well.
3123 * Returning true means the allocation is possible.
3126 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3128 struct mem_cgroup *memcg;
3134 * Disabling accounting is only relevant for some specific memcg
3135 * internal allocations. Therefore we would initially not have such
3136 * check here, since direct calls to the page allocator that are
3137 * accounted to kmemcg (alloc_kmem_pages and friends) only happen
3138 * outside memcg core. We are mostly concerned with cache allocations,
3139 * and by having this test at memcg_kmem_get_cache, we are already able
3140 * to relay the allocation to the root cache and bypass the memcg cache
3143 * There is one exception, though: the SLUB allocator does not create
3144 * large order caches, but rather service large kmallocs directly from
3145 * the page allocator. Therefore, the following sequence when backed by
3146 * the SLUB allocator:
3148 * memcg_stop_kmem_account();
3149 * kmalloc(<large_number>)
3150 * memcg_resume_kmem_account();
3152 * would effectively ignore the fact that we should skip accounting,
3153 * since it will drive us directly to this function without passing
3154 * through the cache selector memcg_kmem_get_cache. Such large
3155 * allocations are extremely rare but can happen, for instance, for the
3156 * cache arrays. We bring this test here.
3158 if (!current->mm || current->memcg_kmem_skip_account)
3161 memcg = get_mem_cgroup_from_mm(current->mm);
3163 if (!memcg_kmem_is_active(memcg)) {
3164 css_put(&memcg->css);
3168 ret = memcg_charge_kmem(memcg, gfp, 1 << order);
3172 css_put(&memcg->css);
3176 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3179 struct page_cgroup *pc;
3181 VM_BUG_ON(mem_cgroup_is_root(memcg));
3183 /* The page allocation failed. Revert */
3185 memcg_uncharge_kmem(memcg, 1 << order);
3189 * The page is freshly allocated and not visible to any
3190 * outside callers yet. Set up pc non-atomically.
3192 pc = lookup_page_cgroup(page);
3193 pc->mem_cgroup = memcg;
3194 pc->flags = PCG_USED;
3197 void __memcg_kmem_uncharge_pages(struct page *page, int order)
3199 struct mem_cgroup *memcg = NULL;
3200 struct page_cgroup *pc;
3203 pc = lookup_page_cgroup(page);
3204 if (!PageCgroupUsed(pc))
3207 memcg = pc->mem_cgroup;
3211 * We trust that only if there is a memcg associated with the page, it
3212 * is a valid allocation
3217 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
3218 memcg_uncharge_kmem(memcg, 1 << order);
3221 static inline void memcg_unregister_all_caches(struct mem_cgroup *memcg)
3224 #endif /* CONFIG_MEMCG_KMEM */
3226 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3229 * Because tail pages are not marked as "used", set it. We're under
3230 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3231 * charge/uncharge will be never happen and move_account() is done under
3232 * compound_lock(), so we don't have to take care of races.
3234 void mem_cgroup_split_huge_fixup(struct page *head)
3236 struct page_cgroup *head_pc = lookup_page_cgroup(head);
3237 struct page_cgroup *pc;
3238 struct mem_cgroup *memcg;
3241 if (mem_cgroup_disabled())
3244 memcg = head_pc->mem_cgroup;
3245 for (i = 1; i < HPAGE_PMD_NR; i++) {
3247 pc->mem_cgroup = memcg;
3248 pc->flags = head_pc->flags;
3250 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
3253 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3256 * mem_cgroup_move_account - move account of the page
3258 * @nr_pages: number of regular pages (>1 for huge pages)
3259 * @pc: page_cgroup of the page.
3260 * @from: mem_cgroup which the page is moved from.
3261 * @to: mem_cgroup which the page is moved to. @from != @to.
3263 * The caller must confirm following.
3264 * - page is not on LRU (isolate_page() is useful.)
3265 * - compound_lock is held when nr_pages > 1
3267 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3270 static int mem_cgroup_move_account(struct page *page,
3271 unsigned int nr_pages,
3272 struct page_cgroup *pc,
3273 struct mem_cgroup *from,
3274 struct mem_cgroup *to)
3276 unsigned long flags;
3279 VM_BUG_ON(from == to);
3280 VM_BUG_ON_PAGE(PageLRU(page), page);
3282 * The page is isolated from LRU. So, collapse function
3283 * will not handle this page. But page splitting can happen.
3284 * Do this check under compound_page_lock(). The caller should
3288 if (nr_pages > 1 && !PageTransHuge(page))
3292 * Prevent mem_cgroup_migrate() from looking at pc->mem_cgroup
3293 * of its source page while we change it: page migration takes
3294 * both pages off the LRU, but page cache replacement doesn't.
3296 if (!trylock_page(page))
3300 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3303 move_lock_mem_cgroup(from, &flags);
3305 if (!PageAnon(page) && page_mapped(page)) {
3306 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
3308 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
3312 if (PageWriteback(page)) {
3313 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK],
3315 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK],
3320 * It is safe to change pc->mem_cgroup here because the page
3321 * is referenced, charged, and isolated - we can't race with
3322 * uncharging, charging, migration, or LRU putback.
3325 /* caller should have done css_get */
3326 pc->mem_cgroup = to;
3327 move_unlock_mem_cgroup(from, &flags);
3330 local_irq_disable();
3331 mem_cgroup_charge_statistics(to, page, nr_pages);
3332 memcg_check_events(to, page);
3333 mem_cgroup_charge_statistics(from, page, -nr_pages);
3334 memcg_check_events(from, page);
3343 * mem_cgroup_move_parent - moves page to the parent group
3344 * @page: the page to move
3345 * @pc: page_cgroup of the page
3346 * @child: page's cgroup
3348 * move charges to its parent or the root cgroup if the group has no
3349 * parent (aka use_hierarchy==0).
3350 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3351 * mem_cgroup_move_account fails) the failure is always temporary and
3352 * it signals a race with a page removal/uncharge or migration. In the
3353 * first case the page is on the way out and it will vanish from the LRU
3354 * on the next attempt and the call should be retried later.
3355 * Isolation from the LRU fails only if page has been isolated from
3356 * the LRU since we looked at it and that usually means either global
3357 * reclaim or migration going on. The page will either get back to the
3359 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3360 * (!PageCgroupUsed) or moved to a different group. The page will
3361 * disappear in the next attempt.
3363 static int mem_cgroup_move_parent(struct page *page,
3364 struct page_cgroup *pc,
3365 struct mem_cgroup *child)
3367 struct mem_cgroup *parent;
3368 unsigned int nr_pages;
3369 unsigned long uninitialized_var(flags);
3372 VM_BUG_ON(mem_cgroup_is_root(child));
3375 if (!get_page_unless_zero(page))
3377 if (isolate_lru_page(page))
3380 nr_pages = hpage_nr_pages(page);
3382 parent = parent_mem_cgroup(child);
3384 * If no parent, move charges to root cgroup.
3387 parent = root_mem_cgroup;
3390 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
3391 flags = compound_lock_irqsave(page);
3394 ret = mem_cgroup_move_account(page, nr_pages,
3397 /* Take charge off the local counters */
3398 page_counter_cancel(&child->memory, nr_pages);
3399 if (do_swap_account)
3400 page_counter_cancel(&child->memsw, nr_pages);
3404 compound_unlock_irqrestore(page, flags);
3405 putback_lru_page(page);
3412 #ifdef CONFIG_MEMCG_SWAP
3413 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
3416 int val = (charge) ? 1 : -1;
3417 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
3421 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3422 * @entry: swap entry to be moved
3423 * @from: mem_cgroup which the entry is moved from
3424 * @to: mem_cgroup which the entry is moved to
3426 * It succeeds only when the swap_cgroup's record for this entry is the same
3427 * as the mem_cgroup's id of @from.
3429 * Returns 0 on success, -EINVAL on failure.
3431 * The caller must have charged to @to, IOW, called page_counter_charge() about
3432 * both res and memsw, and called css_get().
3434 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3435 struct mem_cgroup *from, struct mem_cgroup *to)
3437 unsigned short old_id, new_id;
3439 old_id = mem_cgroup_id(from);
3440 new_id = mem_cgroup_id(to);
3442 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3443 mem_cgroup_swap_statistics(from, false);
3444 mem_cgroup_swap_statistics(to, true);
3446 * This function is only called from task migration context now.
3447 * It postpones page_counter and refcount handling till the end
3448 * of task migration(mem_cgroup_clear_mc()) for performance
3449 * improvement. But we cannot postpone css_get(to) because if
3450 * the process that has been moved to @to does swap-in, the
3451 * refcount of @to might be decreased to 0.
3453 * We are in attach() phase, so the cgroup is guaranteed to be
3454 * alive, so we can just call css_get().
3462 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3463 struct mem_cgroup *from, struct mem_cgroup *to)
3469 #ifdef CONFIG_DEBUG_VM
3470 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3472 struct page_cgroup *pc;
3474 pc = lookup_page_cgroup(page);
3476 * Can be NULL while feeding pages into the page allocator for
3477 * the first time, i.e. during boot or memory hotplug;
3478 * or when mem_cgroup_disabled().
3480 if (likely(pc) && PageCgroupUsed(pc))
3485 bool mem_cgroup_bad_page_check(struct page *page)
3487 if (mem_cgroup_disabled())
3490 return lookup_page_cgroup_used(page) != NULL;
3493 void mem_cgroup_print_bad_page(struct page *page)
3495 struct page_cgroup *pc;
3497 pc = lookup_page_cgroup_used(page);
3499 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
3500 pc, pc->flags, pc->mem_cgroup);
3505 static DEFINE_MUTEX(memcg_limit_mutex);
3507 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3508 unsigned long limit)
3510 unsigned long curusage;
3511 unsigned long oldusage;
3512 bool enlarge = false;
3517 * For keeping hierarchical_reclaim simple, how long we should retry
3518 * is depends on callers. We set our retry-count to be function
3519 * of # of children which we should visit in this loop.
3521 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
3522 mem_cgroup_count_children(memcg);
3524 oldusage = page_counter_read(&memcg->memory);
3527 if (signal_pending(current)) {
3532 mutex_lock(&memcg_limit_mutex);
3533 if (limit > memcg->memsw.limit) {
3534 mutex_unlock(&memcg_limit_mutex);
3538 if (limit > memcg->memory.limit)
3540 ret = page_counter_limit(&memcg->memory, limit);
3541 mutex_unlock(&memcg_limit_mutex);
3546 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, true);
3548 curusage = page_counter_read(&memcg->memory);
3549 /* Usage is reduced ? */
3550 if (curusage >= oldusage)
3553 oldusage = curusage;
3554 } while (retry_count);
3556 if (!ret && enlarge)
3557 memcg_oom_recover(memcg);
3562 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3563 unsigned long limit)
3565 unsigned long curusage;
3566 unsigned long oldusage;
3567 bool enlarge = false;
3571 /* see mem_cgroup_resize_res_limit */
3572 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
3573 mem_cgroup_count_children(memcg);
3575 oldusage = page_counter_read(&memcg->memsw);
3578 if (signal_pending(current)) {
3583 mutex_lock(&memcg_limit_mutex);
3584 if (limit < memcg->memory.limit) {
3585 mutex_unlock(&memcg_limit_mutex);
3589 if (limit > memcg->memsw.limit)
3591 ret = page_counter_limit(&memcg->memsw, limit);
3592 mutex_unlock(&memcg_limit_mutex);
3597 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, false);
3599 curusage = page_counter_read(&memcg->memsw);
3600 /* Usage is reduced ? */
3601 if (curusage >= oldusage)
3604 oldusage = curusage;
3605 } while (retry_count);
3607 if (!ret && enlarge)
3608 memcg_oom_recover(memcg);
3613 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3615 unsigned long *total_scanned)
3617 unsigned long nr_reclaimed = 0;
3618 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3619 unsigned long reclaimed;
3621 struct mem_cgroup_tree_per_zone *mctz;
3622 unsigned long excess;
3623 unsigned long nr_scanned;
3628 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3630 * This loop can run a while, specially if mem_cgroup's continuously
3631 * keep exceeding their soft limit and putting the system under
3638 mz = mem_cgroup_largest_soft_limit_node(mctz);
3643 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
3644 gfp_mask, &nr_scanned);
3645 nr_reclaimed += reclaimed;
3646 *total_scanned += nr_scanned;
3647 spin_lock_irq(&mctz->lock);
3650 * If we failed to reclaim anything from this memory cgroup
3651 * it is time to move on to the next cgroup
3657 * Loop until we find yet another one.
3659 * By the time we get the soft_limit lock
3660 * again, someone might have aded the
3661 * group back on the RB tree. Iterate to
3662 * make sure we get a different mem.
3663 * mem_cgroup_largest_soft_limit_node returns
3664 * NULL if no other cgroup is present on
3668 __mem_cgroup_largest_soft_limit_node(mctz);
3670 css_put(&next_mz->memcg->css);
3671 else /* next_mz == NULL or other memcg */
3675 __mem_cgroup_remove_exceeded(mz, mctz);
3676 excess = soft_limit_excess(mz->memcg);
3678 * One school of thought says that we should not add
3679 * back the node to the tree if reclaim returns 0.
3680 * But our reclaim could return 0, simply because due
3681 * to priority we are exposing a smaller subset of
3682 * memory to reclaim from. Consider this as a longer
3685 /* If excess == 0, no tree ops */
3686 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3687 spin_unlock_irq(&mctz->lock);
3688 css_put(&mz->memcg->css);
3691 * Could not reclaim anything and there are no more
3692 * mem cgroups to try or we seem to be looping without
3693 * reclaiming anything.
3695 if (!nr_reclaimed &&
3697 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3699 } while (!nr_reclaimed);
3701 css_put(&next_mz->memcg->css);
3702 return nr_reclaimed;
3706 * mem_cgroup_force_empty_list - clears LRU of a group
3707 * @memcg: group to clear
3710 * @lru: lru to to clear
3712 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
3713 * reclaim the pages page themselves - pages are moved to the parent (or root)
3716 static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
3717 int node, int zid, enum lru_list lru)
3719 struct lruvec *lruvec;
3720 unsigned long flags;
3721 struct list_head *list;
3725 zone = &NODE_DATA(node)->node_zones[zid];
3726 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
3727 list = &lruvec->lists[lru];
3731 struct page_cgroup *pc;
3734 spin_lock_irqsave(&zone->lru_lock, flags);
3735 if (list_empty(list)) {
3736 spin_unlock_irqrestore(&zone->lru_lock, flags);
3739 page = list_entry(list->prev, struct page, lru);
3741 list_move(&page->lru, list);
3743 spin_unlock_irqrestore(&zone->lru_lock, flags);
3746 spin_unlock_irqrestore(&zone->lru_lock, flags);
3748 pc = lookup_page_cgroup(page);
3750 if (mem_cgroup_move_parent(page, pc, memcg)) {
3751 /* found lock contention or "pc" is obsolete. */
3756 } while (!list_empty(list));
3760 * make mem_cgroup's charge to be 0 if there is no task by moving
3761 * all the charges and pages to the parent.
3762 * This enables deleting this mem_cgroup.
3764 * Caller is responsible for holding css reference on the memcg.
3766 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
3771 /* This is for making all *used* pages to be on LRU. */
3772 lru_add_drain_all();
3773 drain_all_stock_sync(memcg);
3774 mem_cgroup_start_move(memcg);
3775 for_each_node_state(node, N_MEMORY) {
3776 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3779 mem_cgroup_force_empty_list(memcg,
3784 mem_cgroup_end_move(memcg);
3785 memcg_oom_recover(memcg);
3789 * Kernel memory may not necessarily be trackable to a specific
3790 * process. So they are not migrated, and therefore we can't
3791 * expect their value to drop to 0 here.
3792 * Having res filled up with kmem only is enough.
3794 * This is a safety check because mem_cgroup_force_empty_list
3795 * could have raced with mem_cgroup_replace_page_cache callers
3796 * so the lru seemed empty but the page could have been added
3797 * right after the check. RES_USAGE should be safe as we always
3798 * charge before adding to the LRU.
3800 } while (page_counter_read(&memcg->memory) -
3801 page_counter_read(&memcg->kmem) > 0);
3805 * Test whether @memcg has children, dead or alive. Note that this
3806 * function doesn't care whether @memcg has use_hierarchy enabled and
3807 * returns %true if there are child csses according to the cgroup
3808 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
3810 static inline bool memcg_has_children(struct mem_cgroup *memcg)
3815 * The lock does not prevent addition or deletion of children, but
3816 * it prevents a new child from being initialized based on this
3817 * parent in css_online(), so it's enough to decide whether
3818 * hierarchically inherited attributes can still be changed or not.
3820 lockdep_assert_held(&memcg_create_mutex);
3823 ret = css_next_child(NULL, &memcg->css);
3829 * Reclaims as many pages from the given memcg as possible and moves
3830 * the rest to the parent.
3832 * Caller is responsible for holding css reference for memcg.
3834 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3836 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3838 /* we call try-to-free pages for make this cgroup empty */
3839 lru_add_drain_all();
3840 /* try to free all pages in this cgroup */
3841 while (nr_retries && page_counter_read(&memcg->memory)) {
3844 if (signal_pending(current))
3847 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3851 /* maybe some writeback is necessary */
3852 congestion_wait(BLK_RW_ASYNC, HZ/10);
3860 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3861 char *buf, size_t nbytes,
3864 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3866 if (mem_cgroup_is_root(memcg))
3868 return mem_cgroup_force_empty(memcg) ?: nbytes;
3871 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3874 return mem_cgroup_from_css(css)->use_hierarchy;
3877 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3878 struct cftype *cft, u64 val)
3881 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3882 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
3884 mutex_lock(&memcg_create_mutex);
3886 if (memcg->use_hierarchy == val)
3890 * If parent's use_hierarchy is set, we can't make any modifications
3891 * in the child subtrees. If it is unset, then the change can
3892 * occur, provided the current cgroup has no children.
3894 * For the root cgroup, parent_mem is NULL, we allow value to be
3895 * set if there are no children.
3897 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3898 (val == 1 || val == 0)) {
3899 if (!memcg_has_children(memcg))
3900 memcg->use_hierarchy = val;
3907 mutex_unlock(&memcg_create_mutex);
3912 static unsigned long tree_stat(struct mem_cgroup *memcg,
3913 enum mem_cgroup_stat_index idx)
3915 struct mem_cgroup *iter;
3918 /* Per-cpu values can be negative, use a signed accumulator */
3919 for_each_mem_cgroup_tree(iter, memcg)
3920 val += mem_cgroup_read_stat(iter, idx);
3922 if (val < 0) /* race ? */
3927 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3931 if (mem_cgroup_is_root(memcg)) {
3932 val = tree_stat(memcg, MEM_CGROUP_STAT_CACHE);
3933 val += tree_stat(memcg, MEM_CGROUP_STAT_RSS);
3935 val += tree_stat(memcg, MEM_CGROUP_STAT_SWAP);
3938 val = page_counter_read(&memcg->memory);
3940 val = page_counter_read(&memcg->memsw);
3942 return val << PAGE_SHIFT;
3953 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3956 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3957 struct page_counter *counter;
3959 switch (MEMFILE_TYPE(cft->private)) {
3961 counter = &memcg->memory;
3964 counter = &memcg->memsw;
3967 counter = &memcg->kmem;
3973 switch (MEMFILE_ATTR(cft->private)) {
3975 if (counter == &memcg->memory)
3976 return mem_cgroup_usage(memcg, false);
3977 if (counter == &memcg->memsw)
3978 return mem_cgroup_usage(memcg, true);
3979 return (u64)page_counter_read(counter) * PAGE_SIZE;
3981 return (u64)counter->limit * PAGE_SIZE;
3983 return (u64)counter->watermark * PAGE_SIZE;
3985 return counter->failcnt;
3986 case RES_SOFT_LIMIT:
3987 return (u64)memcg->soft_limit * PAGE_SIZE;
3993 #ifdef CONFIG_MEMCG_KMEM
3994 /* should be called with activate_kmem_mutex held */
3995 static int __memcg_activate_kmem(struct mem_cgroup *memcg,
3996 unsigned long nr_pages)
4001 if (memcg_kmem_is_active(memcg))
4005 * We are going to allocate memory for data shared by all memory
4006 * cgroups so let's stop accounting here.
4008 memcg_stop_kmem_account();
4011 * For simplicity, we won't allow this to be disabled. It also can't
4012 * be changed if the cgroup has children already, or if tasks had
4015 * If tasks join before we set the limit, a person looking at
4016 * kmem.usage_in_bytes will have no way to determine when it took
4017 * place, which makes the value quite meaningless.
4019 * After it first became limited, changes in the value of the limit are
4020 * of course permitted.
4022 mutex_lock(&memcg_create_mutex);
4023 if (cgroup_has_tasks(memcg->css.cgroup) ||
4024 (memcg->use_hierarchy && memcg_has_children(memcg)))
4026 mutex_unlock(&memcg_create_mutex);
4030 memcg_id = memcg_alloc_cache_id();
4036 memcg->kmemcg_id = memcg_id;
4037 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
4040 * We couldn't have accounted to this cgroup, because it hasn't got the
4041 * active bit set yet, so this should succeed.
4043 err = page_counter_limit(&memcg->kmem, nr_pages);
4046 static_key_slow_inc(&memcg_kmem_enabled_key);
4048 * Setting the active bit after enabling static branching will
4049 * guarantee no one starts accounting before all call sites are
4052 memcg_kmem_set_active(memcg);
4054 memcg_resume_kmem_account();
4058 static int memcg_activate_kmem(struct mem_cgroup *memcg,
4059 unsigned long nr_pages)
4063 mutex_lock(&activate_kmem_mutex);
4064 ret = __memcg_activate_kmem(memcg, nr_pages);
4065 mutex_unlock(&activate_kmem_mutex);
4069 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
4070 unsigned long limit)
4074 mutex_lock(&memcg_limit_mutex);
4075 if (!memcg_kmem_is_active(memcg))
4076 ret = memcg_activate_kmem(memcg, limit);
4078 ret = page_counter_limit(&memcg->kmem, limit);
4079 mutex_unlock(&memcg_limit_mutex);
4083 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
4086 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4091 mutex_lock(&activate_kmem_mutex);
4093 * If the parent cgroup is not kmem-active now, it cannot be activated
4094 * after this point, because it has at least one child already.
4096 if (memcg_kmem_is_active(parent))
4097 ret = __memcg_activate_kmem(memcg, PAGE_COUNTER_MAX);
4098 mutex_unlock(&activate_kmem_mutex);
4102 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
4103 unsigned long limit)
4107 #endif /* CONFIG_MEMCG_KMEM */
4110 * The user of this function is...
4113 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
4114 char *buf, size_t nbytes, loff_t off)
4116 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
4117 unsigned long nr_pages;
4120 buf = strstrip(buf);
4121 ret = page_counter_memparse(buf, &nr_pages);
4125 switch (MEMFILE_ATTR(of_cft(of)->private)) {
4127 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
4131 switch (MEMFILE_TYPE(of_cft(of)->private)) {
4133 ret = mem_cgroup_resize_limit(memcg, nr_pages);
4136 ret = mem_cgroup_resize_memsw_limit(memcg, nr_pages);
4139 ret = memcg_update_kmem_limit(memcg, nr_pages);
4143 case RES_SOFT_LIMIT:
4144 memcg->soft_limit = nr_pages;
4148 return ret ?: nbytes;
4151 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
4152 size_t nbytes, loff_t off)
4154 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
4155 struct page_counter *counter;
4157 switch (MEMFILE_TYPE(of_cft(of)->private)) {
4159 counter = &memcg->memory;
4162 counter = &memcg->memsw;
4165 counter = &memcg->kmem;
4171 switch (MEMFILE_ATTR(of_cft(of)->private)) {
4173 page_counter_reset_watermark(counter);
4176 counter->failcnt = 0;
4185 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
4188 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
4192 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
4193 struct cftype *cft, u64 val)
4195 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4197 if (val >= (1 << NR_MOVE_TYPE))
4201 * No kind of locking is needed in here, because ->can_attach() will
4202 * check this value once in the beginning of the process, and then carry
4203 * on with stale data. This means that changes to this value will only
4204 * affect task migrations starting after the change.
4206 memcg->move_charge_at_immigrate = val;
4210 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
4211 struct cftype *cft, u64 val)
4218 static int memcg_numa_stat_show(struct seq_file *m, void *v)
4222 unsigned int lru_mask;
4225 static const struct numa_stat stats[] = {
4226 { "total", LRU_ALL },
4227 { "file", LRU_ALL_FILE },
4228 { "anon", LRU_ALL_ANON },
4229 { "unevictable", BIT(LRU_UNEVICTABLE) },
4231 const struct numa_stat *stat;
4234 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
4236 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
4237 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
4238 seq_printf(m, "%s=%lu", stat->name, nr);
4239 for_each_node_state(nid, N_MEMORY) {
4240 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4242 seq_printf(m, " N%d=%lu", nid, nr);
4247 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
4248 struct mem_cgroup *iter;
4251 for_each_mem_cgroup_tree(iter, memcg)
4252 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
4253 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
4254 for_each_node_state(nid, N_MEMORY) {
4256 for_each_mem_cgroup_tree(iter, memcg)
4257 nr += mem_cgroup_node_nr_lru_pages(
4258 iter, nid, stat->lru_mask);
4259 seq_printf(m, " N%d=%lu", nid, nr);
4266 #endif /* CONFIG_NUMA */
4268 static inline void mem_cgroup_lru_names_not_uptodate(void)
4270 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
4273 static int memcg_stat_show(struct seq_file *m, void *v)
4275 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
4276 unsigned long memory, memsw;
4277 struct mem_cgroup *mi;
4280 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
4281 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
4283 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
4284 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
4287 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
4288 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
4289 mem_cgroup_read_events(memcg, i));
4291 for (i = 0; i < NR_LRU_LISTS; i++)
4292 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
4293 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
4295 /* Hierarchical information */
4296 memory = memsw = PAGE_COUNTER_MAX;
4297 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
4298 memory = min(memory, mi->memory.limit);
4299 memsw = min(memsw, mi->memsw.limit);
4301 seq_printf(m, "hierarchical_memory_limit %llu\n",
4302 (u64)memory * PAGE_SIZE);
4303 if (do_swap_account)
4304 seq_printf(m, "hierarchical_memsw_limit %llu\n",
4305 (u64)memsw * PAGE_SIZE);
4307 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
4310 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
4312 for_each_mem_cgroup_tree(mi, memcg)
4313 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
4314 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
4317 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
4318 unsigned long long val = 0;
4320 for_each_mem_cgroup_tree(mi, memcg)
4321 val += mem_cgroup_read_events(mi, i);
4322 seq_printf(m, "total_%s %llu\n",
4323 mem_cgroup_events_names[i], val);
4326 for (i = 0; i < NR_LRU_LISTS; i++) {
4327 unsigned long long val = 0;
4329 for_each_mem_cgroup_tree(mi, memcg)
4330 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
4331 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
4334 #ifdef CONFIG_DEBUG_VM
4337 struct mem_cgroup_per_zone *mz;
4338 struct zone_reclaim_stat *rstat;
4339 unsigned long recent_rotated[2] = {0, 0};
4340 unsigned long recent_scanned[2] = {0, 0};
4342 for_each_online_node(nid)
4343 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4344 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
4345 rstat = &mz->lruvec.reclaim_stat;
4347 recent_rotated[0] += rstat->recent_rotated[0];
4348 recent_rotated[1] += rstat->recent_rotated[1];
4349 recent_scanned[0] += rstat->recent_scanned[0];
4350 recent_scanned[1] += rstat->recent_scanned[1];
4352 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
4353 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
4354 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
4355 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
4362 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
4365 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4367 return mem_cgroup_swappiness(memcg);
4370 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
4371 struct cftype *cft, u64 val)
4373 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4379 memcg->swappiness = val;
4381 vm_swappiness = val;
4386 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4388 struct mem_cgroup_threshold_ary *t;
4389 unsigned long usage;
4394 t = rcu_dereference(memcg->thresholds.primary);
4396 t = rcu_dereference(memcg->memsw_thresholds.primary);
4401 usage = mem_cgroup_usage(memcg, swap);
4404 * current_threshold points to threshold just below or equal to usage.
4405 * If it's not true, a threshold was crossed after last
4406 * call of __mem_cgroup_threshold().
4408 i = t->current_threshold;
4411 * Iterate backward over array of thresholds starting from
4412 * current_threshold and check if a threshold is crossed.
4413 * If none of thresholds below usage is crossed, we read
4414 * only one element of the array here.
4416 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4417 eventfd_signal(t->entries[i].eventfd, 1);
4419 /* i = current_threshold + 1 */
4423 * Iterate forward over array of thresholds starting from
4424 * current_threshold+1 and check if a threshold is crossed.
4425 * If none of thresholds above usage is crossed, we read
4426 * only one element of the array here.
4428 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4429 eventfd_signal(t->entries[i].eventfd, 1);
4431 /* Update current_threshold */
4432 t->current_threshold = i - 1;
4437 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4440 __mem_cgroup_threshold(memcg, false);
4441 if (do_swap_account)
4442 __mem_cgroup_threshold(memcg, true);
4444 memcg = parent_mem_cgroup(memcg);
4448 static int compare_thresholds(const void *a, const void *b)
4450 const struct mem_cgroup_threshold *_a = a;
4451 const struct mem_cgroup_threshold *_b = b;
4453 if (_a->threshold > _b->threshold)
4456 if (_a->threshold < _b->threshold)
4462 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4464 struct mem_cgroup_eventfd_list *ev;
4466 spin_lock(&memcg_oom_lock);
4468 list_for_each_entry(ev, &memcg->oom_notify, list)
4469 eventfd_signal(ev->eventfd, 1);
4471 spin_unlock(&memcg_oom_lock);
4475 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4477 struct mem_cgroup *iter;
4479 for_each_mem_cgroup_tree(iter, memcg)
4480 mem_cgroup_oom_notify_cb(iter);
4483 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4484 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4486 struct mem_cgroup_thresholds *thresholds;
4487 struct mem_cgroup_threshold_ary *new;
4488 unsigned long threshold;
4489 unsigned long usage;
4492 ret = page_counter_memparse(args, &threshold);
4496 mutex_lock(&memcg->thresholds_lock);
4499 thresholds = &memcg->thresholds;
4500 usage = mem_cgroup_usage(memcg, false);
4501 } else if (type == _MEMSWAP) {
4502 thresholds = &memcg->memsw_thresholds;
4503 usage = mem_cgroup_usage(memcg, true);
4507 /* Check if a threshold crossed before adding a new one */
4508 if (thresholds->primary)
4509 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4511 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4513 /* Allocate memory for new array of thresholds */
4514 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4522 /* Copy thresholds (if any) to new array */
4523 if (thresholds->primary) {
4524 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4525 sizeof(struct mem_cgroup_threshold));
4528 /* Add new threshold */
4529 new->entries[size - 1].eventfd = eventfd;
4530 new->entries[size - 1].threshold = threshold;
4532 /* Sort thresholds. Registering of new threshold isn't time-critical */
4533 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4534 compare_thresholds, NULL);
4536 /* Find current threshold */
4537 new->current_threshold = -1;
4538 for (i = 0; i < size; i++) {
4539 if (new->entries[i].threshold <= usage) {
4541 * new->current_threshold will not be used until
4542 * rcu_assign_pointer(), so it's safe to increment
4545 ++new->current_threshold;
4550 /* Free old spare buffer and save old primary buffer as spare */
4551 kfree(thresholds->spare);
4552 thresholds->spare = thresholds->primary;
4554 rcu_assign_pointer(thresholds->primary, new);
4556 /* To be sure that nobody uses thresholds */
4560 mutex_unlock(&memcg->thresholds_lock);
4565 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4566 struct eventfd_ctx *eventfd, const char *args)
4568 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4571 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4572 struct eventfd_ctx *eventfd, const char *args)
4574 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4577 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4578 struct eventfd_ctx *eventfd, enum res_type type)
4580 struct mem_cgroup_thresholds *thresholds;
4581 struct mem_cgroup_threshold_ary *new;
4582 unsigned long usage;
4585 mutex_lock(&memcg->thresholds_lock);
4588 thresholds = &memcg->thresholds;
4589 usage = mem_cgroup_usage(memcg, false);
4590 } else if (type == _MEMSWAP) {
4591 thresholds = &memcg->memsw_thresholds;
4592 usage = mem_cgroup_usage(memcg, true);
4596 if (!thresholds->primary)
4599 /* Check if a threshold crossed before removing */
4600 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4602 /* Calculate new number of threshold */
4604 for (i = 0; i < thresholds->primary->size; i++) {
4605 if (thresholds->primary->entries[i].eventfd != eventfd)
4609 new = thresholds->spare;
4611 /* Set thresholds array to NULL if we don't have thresholds */
4620 /* Copy thresholds and find current threshold */
4621 new->current_threshold = -1;
4622 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4623 if (thresholds->primary->entries[i].eventfd == eventfd)
4626 new->entries[j] = thresholds->primary->entries[i];
4627 if (new->entries[j].threshold <= usage) {
4629 * new->current_threshold will not be used
4630 * until rcu_assign_pointer(), so it's safe to increment
4633 ++new->current_threshold;
4639 /* Swap primary and spare array */
4640 thresholds->spare = thresholds->primary;
4641 /* If all events are unregistered, free the spare array */
4643 kfree(thresholds->spare);
4644 thresholds->spare = NULL;
4647 rcu_assign_pointer(thresholds->primary, new);
4649 /* To be sure that nobody uses thresholds */
4652 mutex_unlock(&memcg->thresholds_lock);
4655 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4656 struct eventfd_ctx *eventfd)
4658 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4661 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4662 struct eventfd_ctx *eventfd)
4664 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4667 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4668 struct eventfd_ctx *eventfd, const char *args)
4670 struct mem_cgroup_eventfd_list *event;
4672 event = kmalloc(sizeof(*event), GFP_KERNEL);
4676 spin_lock(&memcg_oom_lock);
4678 event->eventfd = eventfd;
4679 list_add(&event->list, &memcg->oom_notify);
4681 /* already in OOM ? */
4682 if (atomic_read(&memcg->under_oom))
4683 eventfd_signal(eventfd, 1);
4684 spin_unlock(&memcg_oom_lock);
4689 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4690 struct eventfd_ctx *eventfd)
4692 struct mem_cgroup_eventfd_list *ev, *tmp;
4694 spin_lock(&memcg_oom_lock);
4696 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4697 if (ev->eventfd == eventfd) {
4698 list_del(&ev->list);
4703 spin_unlock(&memcg_oom_lock);
4706 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4708 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
4710 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4711 seq_printf(sf, "under_oom %d\n", (bool)atomic_read(&memcg->under_oom));
4715 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4716 struct cftype *cft, u64 val)
4718 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4720 /* cannot set to root cgroup and only 0 and 1 are allowed */
4721 if (!css->parent || !((val == 0) || (val == 1)))
4724 memcg->oom_kill_disable = val;
4726 memcg_oom_recover(memcg);
4731 #ifdef CONFIG_MEMCG_KMEM
4732 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
4736 memcg->kmemcg_id = -1;
4737 ret = memcg_propagate_kmem(memcg);
4741 return mem_cgroup_sockets_init(memcg, ss);
4744 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
4746 mem_cgroup_sockets_destroy(memcg);
4749 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
4751 if (!memcg_kmem_is_active(memcg))
4755 * kmem charges can outlive the cgroup. In the case of slab
4756 * pages, for instance, a page contain objects from various
4757 * processes. As we prevent from taking a reference for every
4758 * such allocation we have to be careful when doing uncharge
4759 * (see memcg_uncharge_kmem) and here during offlining.
4761 * The idea is that that only the _last_ uncharge which sees
4762 * the dead memcg will drop the last reference. An additional
4763 * reference is taken here before the group is marked dead
4764 * which is then paired with css_put during uncharge resp. here.
4766 * Although this might sound strange as this path is called from
4767 * css_offline() when the referencemight have dropped down to 0 and
4768 * shouldn't be incremented anymore (css_tryget_online() would
4769 * fail) we do not have other options because of the kmem
4770 * allocations lifetime.
4772 css_get(&memcg->css);
4774 memcg_kmem_mark_dead(memcg);
4776 if (page_counter_read(&memcg->kmem))
4779 if (memcg_kmem_test_and_clear_dead(memcg))
4780 css_put(&memcg->css);
4783 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
4788 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
4792 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
4798 * DO NOT USE IN NEW FILES.
4800 * "cgroup.event_control" implementation.
4802 * This is way over-engineered. It tries to support fully configurable
4803 * events for each user. Such level of flexibility is completely
4804 * unnecessary especially in the light of the planned unified hierarchy.
4806 * Please deprecate this and replace with something simpler if at all
4811 * Unregister event and free resources.
4813 * Gets called from workqueue.
4815 static void memcg_event_remove(struct work_struct *work)
4817 struct mem_cgroup_event *event =
4818 container_of(work, struct mem_cgroup_event, remove);
4819 struct mem_cgroup *memcg = event->memcg;
4821 remove_wait_queue(event->wqh, &event->wait);
4823 event->unregister_event(memcg, event->eventfd);
4825 /* Notify userspace the event is going away. */
4826 eventfd_signal(event->eventfd, 1);
4828 eventfd_ctx_put(event->eventfd);
4830 css_put(&memcg->css);
4834 * Gets called on POLLHUP on eventfd when user closes it.
4836 * Called with wqh->lock held and interrupts disabled.
4838 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
4839 int sync, void *key)
4841 struct mem_cgroup_event *event =
4842 container_of(wait, struct mem_cgroup_event, wait);
4843 struct mem_cgroup *memcg = event->memcg;
4844 unsigned long flags = (unsigned long)key;
4846 if (flags & POLLHUP) {
4848 * If the event has been detached at cgroup removal, we
4849 * can simply return knowing the other side will cleanup
4852 * We can't race against event freeing since the other
4853 * side will require wqh->lock via remove_wait_queue(),
4856 spin_lock(&memcg->event_list_lock);
4857 if (!list_empty(&event->list)) {
4858 list_del_init(&event->list);
4860 * We are in atomic context, but cgroup_event_remove()
4861 * may sleep, so we have to call it in workqueue.
4863 schedule_work(&event->remove);
4865 spin_unlock(&memcg->event_list_lock);
4871 static void memcg_event_ptable_queue_proc(struct file *file,
4872 wait_queue_head_t *wqh, poll_table *pt)
4874 struct mem_cgroup_event *event =
4875 container_of(pt, struct mem_cgroup_event, pt);
4878 add_wait_queue(wqh, &event->wait);
4882 * DO NOT USE IN NEW FILES.
4884 * Parse input and register new cgroup event handler.
4886 * Input must be in format '<event_fd> <control_fd> <args>'.
4887 * Interpretation of args is defined by control file implementation.
4889 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4890 char *buf, size_t nbytes, loff_t off)
4892 struct cgroup_subsys_state *css = of_css(of);
4893 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4894 struct mem_cgroup_event *event;
4895 struct cgroup_subsys_state *cfile_css;
4896 unsigned int efd, cfd;
4903 buf = strstrip(buf);
4905 efd = simple_strtoul(buf, &endp, 10);
4910 cfd = simple_strtoul(buf, &endp, 10);
4911 if ((*endp != ' ') && (*endp != '\0'))
4915 event = kzalloc(sizeof(*event), GFP_KERNEL);
4919 event->memcg = memcg;
4920 INIT_LIST_HEAD(&event->list);
4921 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4922 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4923 INIT_WORK(&event->remove, memcg_event_remove);
4931 event->eventfd = eventfd_ctx_fileget(efile.file);
4932 if (IS_ERR(event->eventfd)) {
4933 ret = PTR_ERR(event->eventfd);
4940 goto out_put_eventfd;
4943 /* the process need read permission on control file */
4944 /* AV: shouldn't we check that it's been opened for read instead? */
4945 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4950 * Determine the event callbacks and set them in @event. This used
4951 * to be done via struct cftype but cgroup core no longer knows
4952 * about these events. The following is crude but the whole thing
4953 * is for compatibility anyway.
4955 * DO NOT ADD NEW FILES.
4957 name = cfile.file->f_dentry->d_name.name;
4959 if (!strcmp(name, "memory.usage_in_bytes")) {
4960 event->register_event = mem_cgroup_usage_register_event;
4961 event->unregister_event = mem_cgroup_usage_unregister_event;
4962 } else if (!strcmp(name, "memory.oom_control")) {
4963 event->register_event = mem_cgroup_oom_register_event;
4964 event->unregister_event = mem_cgroup_oom_unregister_event;
4965 } else if (!strcmp(name, "memory.pressure_level")) {
4966 event->register_event = vmpressure_register_event;
4967 event->unregister_event = vmpressure_unregister_event;
4968 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4969 event->register_event = memsw_cgroup_usage_register_event;
4970 event->unregister_event = memsw_cgroup_usage_unregister_event;
4977 * Verify @cfile should belong to @css. Also, remaining events are
4978 * automatically removed on cgroup destruction but the removal is
4979 * asynchronous, so take an extra ref on @css.
4981 cfile_css = css_tryget_online_from_dir(cfile.file->f_dentry->d_parent,
4982 &memory_cgrp_subsys);
4984 if (IS_ERR(cfile_css))
4986 if (cfile_css != css) {
4991 ret = event->register_event(memcg, event->eventfd, buf);
4995 efile.file->f_op->poll(efile.file, &event->pt);
4997 spin_lock(&memcg->event_list_lock);
4998 list_add(&event->list, &memcg->event_list);
4999 spin_unlock(&memcg->event_list_lock);
5011 eventfd_ctx_put(event->eventfd);
5020 static struct cftype mem_cgroup_files[] = {
5022 .name = "usage_in_bytes",
5023 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
5024 .read_u64 = mem_cgroup_read_u64,
5027 .name = "max_usage_in_bytes",
5028 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
5029 .write = mem_cgroup_reset,
5030 .read_u64 = mem_cgroup_read_u64,
5033 .name = "limit_in_bytes",
5034 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
5035 .write = mem_cgroup_write,
5036 .read_u64 = mem_cgroup_read_u64,
5039 .name = "soft_limit_in_bytes",
5040 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
5041 .write = mem_cgroup_write,
5042 .read_u64 = mem_cgroup_read_u64,
5046 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
5047 .write = mem_cgroup_reset,
5048 .read_u64 = mem_cgroup_read_u64,
5052 .seq_show = memcg_stat_show,
5055 .name = "force_empty",
5056 .write = mem_cgroup_force_empty_write,
5059 .name = "use_hierarchy",
5060 .write_u64 = mem_cgroup_hierarchy_write,
5061 .read_u64 = mem_cgroup_hierarchy_read,
5064 .name = "cgroup.event_control", /* XXX: for compat */
5065 .write = memcg_write_event_control,
5066 .flags = CFTYPE_NO_PREFIX,
5070 .name = "swappiness",
5071 .read_u64 = mem_cgroup_swappiness_read,
5072 .write_u64 = mem_cgroup_swappiness_write,
5075 .name = "move_charge_at_immigrate",
5076 .read_u64 = mem_cgroup_move_charge_read,
5077 .write_u64 = mem_cgroup_move_charge_write,
5080 .name = "oom_control",
5081 .seq_show = mem_cgroup_oom_control_read,
5082 .write_u64 = mem_cgroup_oom_control_write,
5083 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
5086 .name = "pressure_level",
5090 .name = "numa_stat",
5091 .seq_show = memcg_numa_stat_show,
5094 #ifdef CONFIG_MEMCG_KMEM
5096 .name = "kmem.limit_in_bytes",
5097 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
5098 .write = mem_cgroup_write,
5099 .read_u64 = mem_cgroup_read_u64,
5102 .name = "kmem.usage_in_bytes",
5103 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5104 .read_u64 = mem_cgroup_read_u64,
5107 .name = "kmem.failcnt",
5108 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5109 .write = mem_cgroup_reset,
5110 .read_u64 = mem_cgroup_read_u64,
5113 .name = "kmem.max_usage_in_bytes",
5114 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5115 .write = mem_cgroup_reset,
5116 .read_u64 = mem_cgroup_read_u64,
5118 #ifdef CONFIG_SLABINFO
5120 .name = "kmem.slabinfo",
5121 .seq_show = mem_cgroup_slabinfo_read,
5125 { }, /* terminate */
5128 #ifdef CONFIG_MEMCG_SWAP
5129 static struct cftype memsw_cgroup_files[] = {
5131 .name = "memsw.usage_in_bytes",
5132 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
5133 .read_u64 = mem_cgroup_read_u64,
5136 .name = "memsw.max_usage_in_bytes",
5137 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
5138 .write = mem_cgroup_reset,
5139 .read_u64 = mem_cgroup_read_u64,
5142 .name = "memsw.limit_in_bytes",
5143 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
5144 .write = mem_cgroup_write,
5145 .read_u64 = mem_cgroup_read_u64,
5148 .name = "memsw.failcnt",
5149 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
5150 .write = mem_cgroup_reset,
5151 .read_u64 = mem_cgroup_read_u64,
5153 { }, /* terminate */
5156 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
5158 struct mem_cgroup_per_node *pn;
5159 struct mem_cgroup_per_zone *mz;
5160 int zone, tmp = node;
5162 * This routine is called against possible nodes.
5163 * But it's BUG to call kmalloc() against offline node.
5165 * TODO: this routine can waste much memory for nodes which will
5166 * never be onlined. It's better to use memory hotplug callback
5169 if (!node_state(node, N_NORMAL_MEMORY))
5171 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5175 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5176 mz = &pn->zoneinfo[zone];
5177 lruvec_init(&mz->lruvec);
5178 mz->usage_in_excess = 0;
5179 mz->on_tree = false;
5182 memcg->nodeinfo[node] = pn;
5186 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
5188 kfree(memcg->nodeinfo[node]);
5191 static struct mem_cgroup *mem_cgroup_alloc(void)
5193 struct mem_cgroup *memcg;
5196 size = sizeof(struct mem_cgroup);
5197 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
5199 memcg = kzalloc(size, GFP_KERNEL);
5203 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
5206 spin_lock_init(&memcg->pcp_counter_lock);
5215 * At destroying mem_cgroup, references from swap_cgroup can remain.
5216 * (scanning all at force_empty is too costly...)
5218 * Instead of clearing all references at force_empty, we remember
5219 * the number of reference from swap_cgroup and free mem_cgroup when
5220 * it goes down to 0.
5222 * Removal of cgroup itself succeeds regardless of refs from swap.
5225 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5229 mem_cgroup_remove_from_trees(memcg);
5232 free_mem_cgroup_per_zone_info(memcg, node);
5234 free_percpu(memcg->stat);
5237 * We need to make sure that (at least for now), the jump label
5238 * destruction code runs outside of the cgroup lock. This is because
5239 * get_online_cpus(), which is called from the static_branch update,
5240 * can't be called inside the cgroup_lock. cpusets are the ones
5241 * enforcing this dependency, so if they ever change, we might as well.
5243 * schedule_work() will guarantee this happens. Be careful if you need
5244 * to move this code around, and make sure it is outside
5247 disarm_static_keys(memcg);
5252 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
5254 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
5256 if (!memcg->memory.parent)
5258 return mem_cgroup_from_counter(memcg->memory.parent, memory);
5260 EXPORT_SYMBOL(parent_mem_cgroup);
5262 static void __init mem_cgroup_soft_limit_tree_init(void)
5264 struct mem_cgroup_tree_per_node *rtpn;
5265 struct mem_cgroup_tree_per_zone *rtpz;
5266 int tmp, node, zone;
5268 for_each_node(node) {
5270 if (!node_state(node, N_NORMAL_MEMORY))
5272 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
5275 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5277 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5278 rtpz = &rtpn->rb_tree_per_zone[zone];
5279 rtpz->rb_root = RB_ROOT;
5280 spin_lock_init(&rtpz->lock);
5285 static struct cgroup_subsys_state * __ref
5286 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5288 struct mem_cgroup *memcg;
5289 long error = -ENOMEM;
5292 memcg = mem_cgroup_alloc();
5294 return ERR_PTR(error);
5297 if (alloc_mem_cgroup_per_zone_info(memcg, node))
5301 if (parent_css == NULL) {
5302 root_mem_cgroup = memcg;
5303 page_counter_init(&memcg->memory, NULL);
5304 page_counter_init(&memcg->memsw, NULL);
5305 page_counter_init(&memcg->kmem, NULL);
5308 memcg->last_scanned_node = MAX_NUMNODES;
5309 INIT_LIST_HEAD(&memcg->oom_notify);
5310 memcg->move_charge_at_immigrate = 0;
5311 mutex_init(&memcg->thresholds_lock);
5312 spin_lock_init(&memcg->move_lock);
5313 vmpressure_init(&memcg->vmpressure);
5314 INIT_LIST_HEAD(&memcg->event_list);
5315 spin_lock_init(&memcg->event_list_lock);
5320 __mem_cgroup_free(memcg);
5321 return ERR_PTR(error);
5325 mem_cgroup_css_online(struct cgroup_subsys_state *css)
5327 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5328 struct mem_cgroup *parent = mem_cgroup_from_css(css->parent);
5331 if (css->id > MEM_CGROUP_ID_MAX)
5337 mutex_lock(&memcg_create_mutex);
5339 memcg->use_hierarchy = parent->use_hierarchy;
5340 memcg->oom_kill_disable = parent->oom_kill_disable;
5341 memcg->swappiness = mem_cgroup_swappiness(parent);
5343 if (parent->use_hierarchy) {
5344 page_counter_init(&memcg->memory, &parent->memory);
5345 page_counter_init(&memcg->memsw, &parent->memsw);
5346 page_counter_init(&memcg->kmem, &parent->kmem);
5349 * No need to take a reference to the parent because cgroup
5350 * core guarantees its existence.
5353 page_counter_init(&memcg->memory, NULL);
5354 page_counter_init(&memcg->memsw, NULL);
5355 page_counter_init(&memcg->kmem, NULL);
5357 * Deeper hierachy with use_hierarchy == false doesn't make
5358 * much sense so let cgroup subsystem know about this
5359 * unfortunate state in our controller.
5361 if (parent != root_mem_cgroup)
5362 memory_cgrp_subsys.broken_hierarchy = true;
5364 mutex_unlock(&memcg_create_mutex);
5366 ret = memcg_init_kmem(memcg, &memory_cgrp_subsys);
5371 * Make sure the memcg is initialized: mem_cgroup_iter()
5372 * orders reading memcg->initialized against its callers
5373 * reading the memcg members.
5375 smp_store_release(&memcg->initialized, 1);
5380 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5382 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5383 struct mem_cgroup_event *event, *tmp;
5384 struct cgroup_subsys_state *iter;
5387 * Unregister events and notify userspace.
5388 * Notify userspace about cgroup removing only after rmdir of cgroup
5389 * directory to avoid race between userspace and kernelspace.
5391 spin_lock(&memcg->event_list_lock);
5392 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5393 list_del_init(&event->list);
5394 schedule_work(&event->remove);
5396 spin_unlock(&memcg->event_list_lock);
5398 kmem_cgroup_css_offline(memcg);
5401 * This requires that offlining is serialized. Right now that is
5402 * guaranteed because css_killed_work_fn() holds the cgroup_mutex.
5404 css_for_each_descendant_post(iter, css)
5405 mem_cgroup_reparent_charges(mem_cgroup_from_css(iter));
5407 memcg_unregister_all_caches(memcg);
5408 vmpressure_cleanup(&memcg->vmpressure);
5411 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5413 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5415 * XXX: css_offline() would be where we should reparent all
5416 * memory to prepare the cgroup for destruction. However,
5417 * memcg does not do css_tryget_online() and page_counter charging
5418 * under the same RCU lock region, which means that charging
5419 * could race with offlining. Offlining only happens to
5420 * cgroups with no tasks in them but charges can show up
5421 * without any tasks from the swapin path when the target
5422 * memcg is looked up from the swapout record and not from the
5423 * current task as it usually is. A race like this can leak
5424 * charges and put pages with stale cgroup pointers into
5428 * lookup_swap_cgroup_id()
5430 * mem_cgroup_lookup()
5431 * css_tryget_online()
5433 * disable css_tryget_online()
5436 * reparent_charges()
5437 * page_counter_try_charge()
5440 * pc->mem_cgroup = dead memcg
5443 * The bulk of the charges are still moved in offline_css() to
5444 * avoid pinning a lot of pages in case a long-term reference
5445 * like a swapout record is deferring the css_free() to long
5446 * after offlining. But this makes sure we catch any charges
5447 * made after offlining:
5449 mem_cgroup_reparent_charges(memcg);
5451 memcg_destroy_kmem(memcg);
5452 __mem_cgroup_free(memcg);
5456 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5457 * @css: the target css
5459 * Reset the states of the mem_cgroup associated with @css. This is
5460 * invoked when the userland requests disabling on the default hierarchy
5461 * but the memcg is pinned through dependency. The memcg should stop
5462 * applying policies and should revert to the vanilla state as it may be
5463 * made visible again.
5465 * The current implementation only resets the essential configurations.
5466 * This needs to be expanded to cover all the visible parts.
5468 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5470 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5472 mem_cgroup_resize_limit(memcg, PAGE_COUNTER_MAX);
5473 mem_cgroup_resize_memsw_limit(memcg, PAGE_COUNTER_MAX);
5474 memcg_update_kmem_limit(memcg, PAGE_COUNTER_MAX);
5475 memcg->soft_limit = 0;
5479 /* Handlers for move charge at task migration. */
5480 static int mem_cgroup_do_precharge(unsigned long count)
5484 /* Try a single bulk charge without reclaim first */
5485 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_WAIT, count);
5487 mc.precharge += count;
5490 if (ret == -EINTR) {
5491 cancel_charge(root_mem_cgroup, count);
5495 /* Try charges one by one with reclaim */
5497 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_NORETRY, 1);
5499 * In case of failure, any residual charges against
5500 * mc.to will be dropped by mem_cgroup_clear_mc()
5501 * later on. However, cancel any charges that are
5502 * bypassed to root right away or they'll be lost.
5505 cancel_charge(root_mem_cgroup, 1);
5515 * get_mctgt_type - get target type of moving charge
5516 * @vma: the vma the pte to be checked belongs
5517 * @addr: the address corresponding to the pte to be checked
5518 * @ptent: the pte to be checked
5519 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5522 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5523 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5524 * move charge. if @target is not NULL, the page is stored in target->page
5525 * with extra refcnt got(Callers should handle it).
5526 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5527 * target for charge migration. if @target is not NULL, the entry is stored
5530 * Called with pte lock held.
5537 enum mc_target_type {
5543 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5544 unsigned long addr, pte_t ptent)
5546 struct page *page = vm_normal_page(vma, addr, ptent);
5548 if (!page || !page_mapped(page))
5550 if (PageAnon(page)) {
5551 /* we don't move shared anon */
5554 } else if (!move_file())
5555 /* we ignore mapcount for file pages */
5557 if (!get_page_unless_zero(page))
5564 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5565 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5567 struct page *page = NULL;
5568 swp_entry_t ent = pte_to_swp_entry(ptent);
5570 if (!move_anon() || non_swap_entry(ent))
5573 * Because lookup_swap_cache() updates some statistics counter,
5574 * we call find_get_page() with swapper_space directly.
5576 page = find_get_page(swap_address_space(ent), ent.val);
5577 if (do_swap_account)
5578 entry->val = ent.val;
5583 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5584 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5590 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5591 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5593 struct page *page = NULL;
5594 struct address_space *mapping;
5597 if (!vma->vm_file) /* anonymous vma */
5602 mapping = vma->vm_file->f_mapping;
5603 if (pte_none(ptent))
5604 pgoff = linear_page_index(vma, addr);
5605 else /* pte_file(ptent) is true */
5606 pgoff = pte_to_pgoff(ptent);
5608 /* page is moved even if it's not RSS of this task(page-faulted). */
5610 /* shmem/tmpfs may report page out on swap: account for that too. */
5611 if (shmem_mapping(mapping)) {
5612 page = find_get_entry(mapping, pgoff);
5613 if (radix_tree_exceptional_entry(page)) {
5614 swp_entry_t swp = radix_to_swp_entry(page);
5615 if (do_swap_account)
5617 page = find_get_page(swap_address_space(swp), swp.val);
5620 page = find_get_page(mapping, pgoff);
5622 page = find_get_page(mapping, pgoff);
5627 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5628 unsigned long addr, pte_t ptent, union mc_target *target)
5630 struct page *page = NULL;
5631 struct page_cgroup *pc;
5632 enum mc_target_type ret = MC_TARGET_NONE;
5633 swp_entry_t ent = { .val = 0 };
5635 if (pte_present(ptent))
5636 page = mc_handle_present_pte(vma, addr, ptent);
5637 else if (is_swap_pte(ptent))
5638 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5639 else if (pte_none(ptent) || pte_file(ptent))
5640 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5642 if (!page && !ent.val)
5645 pc = lookup_page_cgroup(page);
5647 * Do only loose check w/o serialization.
5648 * mem_cgroup_move_account() checks the pc is valid or
5649 * not under LRU exclusion.
5651 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5652 ret = MC_TARGET_PAGE;
5654 target->page = page;
5656 if (!ret || !target)
5659 /* There is a swap entry and a page doesn't exist or isn't charged */
5660 if (ent.val && !ret &&
5661 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5662 ret = MC_TARGET_SWAP;
5669 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5671 * We don't consider swapping or file mapped pages because THP does not
5672 * support them for now.
5673 * Caller should make sure that pmd_trans_huge(pmd) is true.
5675 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5676 unsigned long addr, pmd_t pmd, union mc_target *target)
5678 struct page *page = NULL;
5679 struct page_cgroup *pc;
5680 enum mc_target_type ret = MC_TARGET_NONE;
5682 page = pmd_page(pmd);
5683 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5686 pc = lookup_page_cgroup(page);
5687 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5688 ret = MC_TARGET_PAGE;
5691 target->page = page;
5697 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5698 unsigned long addr, pmd_t pmd, union mc_target *target)
5700 return MC_TARGET_NONE;
5704 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5705 unsigned long addr, unsigned long end,
5706 struct mm_walk *walk)
5708 struct vm_area_struct *vma = walk->private;
5712 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
5713 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5714 mc.precharge += HPAGE_PMD_NR;
5719 if (pmd_trans_unstable(pmd))
5721 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5722 for (; addr != end; pte++, addr += PAGE_SIZE)
5723 if (get_mctgt_type(vma, addr, *pte, NULL))
5724 mc.precharge++; /* increment precharge temporarily */
5725 pte_unmap_unlock(pte - 1, ptl);
5731 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5733 unsigned long precharge;
5734 struct vm_area_struct *vma;
5736 down_read(&mm->mmap_sem);
5737 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5738 struct mm_walk mem_cgroup_count_precharge_walk = {
5739 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5743 if (is_vm_hugetlb_page(vma))
5745 walk_page_range(vma->vm_start, vma->vm_end,
5746 &mem_cgroup_count_precharge_walk);
5748 up_read(&mm->mmap_sem);
5750 precharge = mc.precharge;
5756 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5758 unsigned long precharge = mem_cgroup_count_precharge(mm);
5760 VM_BUG_ON(mc.moving_task);
5761 mc.moving_task = current;
5762 return mem_cgroup_do_precharge(precharge);
5765 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5766 static void __mem_cgroup_clear_mc(void)
5768 struct mem_cgroup *from = mc.from;
5769 struct mem_cgroup *to = mc.to;
5772 /* we must uncharge all the leftover precharges from mc.to */
5774 cancel_charge(mc.to, mc.precharge);
5778 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5779 * we must uncharge here.
5781 if (mc.moved_charge) {
5782 cancel_charge(mc.from, mc.moved_charge);
5783 mc.moved_charge = 0;
5785 /* we must fixup refcnts and charges */
5786 if (mc.moved_swap) {
5787 /* uncharge swap account from the old cgroup */
5788 if (!mem_cgroup_is_root(mc.from))
5789 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5792 * we charged both to->memory and to->memsw, so we
5793 * should uncharge to->memory.
5795 if (!mem_cgroup_is_root(mc.to))
5796 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5798 for (i = 0; i < mc.moved_swap; i++)
5799 css_put(&mc.from->css);
5801 /* we've already done css_get(mc.to) */
5804 memcg_oom_recover(from);
5805 memcg_oom_recover(to);
5806 wake_up_all(&mc.waitq);
5809 static void mem_cgroup_clear_mc(void)
5811 struct mem_cgroup *from = mc.from;
5814 * we must clear moving_task before waking up waiters at the end of
5817 mc.moving_task = NULL;
5818 __mem_cgroup_clear_mc();
5819 spin_lock(&mc.lock);
5822 spin_unlock(&mc.lock);
5823 mem_cgroup_end_move(from);
5826 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
5827 struct cgroup_taskset *tset)
5829 struct task_struct *p = cgroup_taskset_first(tset);
5831 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5832 unsigned long move_charge_at_immigrate;
5835 * We are now commited to this value whatever it is. Changes in this
5836 * tunable will only affect upcoming migrations, not the current one.
5837 * So we need to save it, and keep it going.
5839 move_charge_at_immigrate = memcg->move_charge_at_immigrate;
5840 if (move_charge_at_immigrate) {
5841 struct mm_struct *mm;
5842 struct mem_cgroup *from = mem_cgroup_from_task(p);
5844 VM_BUG_ON(from == memcg);
5846 mm = get_task_mm(p);
5849 /* We move charges only when we move a owner of the mm */
5850 if (mm->owner == p) {
5853 VM_BUG_ON(mc.precharge);
5854 VM_BUG_ON(mc.moved_charge);
5855 VM_BUG_ON(mc.moved_swap);
5856 mem_cgroup_start_move(from);
5857 spin_lock(&mc.lock);
5860 mc.immigrate_flags = move_charge_at_immigrate;
5861 spin_unlock(&mc.lock);
5862 /* We set mc.moving_task later */
5864 ret = mem_cgroup_precharge_mc(mm);
5866 mem_cgroup_clear_mc();
5873 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
5874 struct cgroup_taskset *tset)
5876 mem_cgroup_clear_mc();
5879 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5880 unsigned long addr, unsigned long end,
5881 struct mm_walk *walk)
5884 struct vm_area_struct *vma = walk->private;
5887 enum mc_target_type target_type;
5888 union mc_target target;
5890 struct page_cgroup *pc;
5893 * We don't take compound_lock() here but no race with splitting thp
5895 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
5896 * under splitting, which means there's no concurrent thp split,
5897 * - if another thread runs into split_huge_page() just after we
5898 * entered this if-block, the thread must wait for page table lock
5899 * to be unlocked in __split_huge_page_splitting(), where the main
5900 * part of thp split is not executed yet.
5902 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
5903 if (mc.precharge < HPAGE_PMD_NR) {
5907 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5908 if (target_type == MC_TARGET_PAGE) {
5910 if (!isolate_lru_page(page)) {
5911 pc = lookup_page_cgroup(page);
5912 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
5913 pc, mc.from, mc.to)) {
5914 mc.precharge -= HPAGE_PMD_NR;
5915 mc.moved_charge += HPAGE_PMD_NR;
5917 putback_lru_page(page);
5925 if (pmd_trans_unstable(pmd))
5928 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5929 for (; addr != end; addr += PAGE_SIZE) {
5930 pte_t ptent = *(pte++);
5936 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5937 case MC_TARGET_PAGE:
5939 if (isolate_lru_page(page))
5941 pc = lookup_page_cgroup(page);
5942 if (!mem_cgroup_move_account(page, 1, pc,
5945 /* we uncharge from mc.from later. */
5948 putback_lru_page(page);
5949 put: /* get_mctgt_type() gets the page */
5952 case MC_TARGET_SWAP:
5954 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5956 /* we fixup refcnts and charges later. */
5964 pte_unmap_unlock(pte - 1, ptl);
5969 * We have consumed all precharges we got in can_attach().
5970 * We try charge one by one, but don't do any additional
5971 * charges to mc.to if we have failed in charge once in attach()
5974 ret = mem_cgroup_do_precharge(1);
5982 static void mem_cgroup_move_charge(struct mm_struct *mm)
5984 struct vm_area_struct *vma;
5986 lru_add_drain_all();
5988 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5990 * Someone who are holding the mmap_sem might be waiting in
5991 * waitq. So we cancel all extra charges, wake up all waiters,
5992 * and retry. Because we cancel precharges, we might not be able
5993 * to move enough charges, but moving charge is a best-effort
5994 * feature anyway, so it wouldn't be a big problem.
5996 __mem_cgroup_clear_mc();
6000 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6002 struct mm_walk mem_cgroup_move_charge_walk = {
6003 .pmd_entry = mem_cgroup_move_charge_pte_range,
6007 if (is_vm_hugetlb_page(vma))
6009 ret = walk_page_range(vma->vm_start, vma->vm_end,
6010 &mem_cgroup_move_charge_walk);
6013 * means we have consumed all precharges and failed in
6014 * doing additional charge. Just abandon here.
6018 up_read(&mm->mmap_sem);
6021 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
6022 struct cgroup_taskset *tset)
6024 struct task_struct *p = cgroup_taskset_first(tset);
6025 struct mm_struct *mm = get_task_mm(p);
6029 mem_cgroup_move_charge(mm);
6033 mem_cgroup_clear_mc();
6035 #else /* !CONFIG_MMU */
6036 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
6037 struct cgroup_taskset *tset)
6041 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
6042 struct cgroup_taskset *tset)
6045 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
6046 struct cgroup_taskset *tset)
6052 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6053 * to verify whether we're attached to the default hierarchy on each mount
6056 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
6059 * use_hierarchy is forced on the default hierarchy. cgroup core
6060 * guarantees that @root doesn't have any children, so turning it
6061 * on for the root memcg is enough.
6063 if (cgroup_on_dfl(root_css->cgroup))
6064 mem_cgroup_from_css(root_css)->use_hierarchy = true;
6067 struct cgroup_subsys memory_cgrp_subsys = {
6068 .css_alloc = mem_cgroup_css_alloc,
6069 .css_online = mem_cgroup_css_online,
6070 .css_offline = mem_cgroup_css_offline,
6071 .css_free = mem_cgroup_css_free,
6072 .css_reset = mem_cgroup_css_reset,
6073 .can_attach = mem_cgroup_can_attach,
6074 .cancel_attach = mem_cgroup_cancel_attach,
6075 .attach = mem_cgroup_move_task,
6076 .bind = mem_cgroup_bind,
6077 .legacy_cftypes = mem_cgroup_files,
6081 #ifdef CONFIG_MEMCG_SWAP
6082 static int __init enable_swap_account(char *s)
6084 if (!strcmp(s, "1"))
6085 really_do_swap_account = 1;
6086 else if (!strcmp(s, "0"))
6087 really_do_swap_account = 0;
6090 __setup("swapaccount=", enable_swap_account);
6092 static void __init memsw_file_init(void)
6094 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
6095 memsw_cgroup_files));
6098 static void __init enable_swap_cgroup(void)
6100 if (!mem_cgroup_disabled() && really_do_swap_account) {
6101 do_swap_account = 1;
6107 static void __init enable_swap_cgroup(void)
6112 #ifdef CONFIG_MEMCG_SWAP
6114 * mem_cgroup_swapout - transfer a memsw charge to swap
6115 * @page: page whose memsw charge to transfer
6116 * @entry: swap entry to move the charge to
6118 * Transfer the memsw charge of @page to @entry.
6120 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
6122 struct page_cgroup *pc;
6123 unsigned short oldid;
6125 VM_BUG_ON_PAGE(PageLRU(page), page);
6126 VM_BUG_ON_PAGE(page_count(page), page);
6128 if (!do_swap_account)
6131 pc = lookup_page_cgroup(page);
6133 /* Readahead page, never charged */
6134 if (!PageCgroupUsed(pc))
6137 VM_BUG_ON_PAGE(!(pc->flags & PCG_MEMSW), page);
6139 oldid = swap_cgroup_record(entry, mem_cgroup_id(pc->mem_cgroup));
6140 VM_BUG_ON_PAGE(oldid, page);
6142 pc->flags &= ~PCG_MEMSW;
6143 css_get(&pc->mem_cgroup->css);
6144 mem_cgroup_swap_statistics(pc->mem_cgroup, true);
6148 * mem_cgroup_uncharge_swap - uncharge a swap entry
6149 * @entry: swap entry to uncharge
6151 * Drop the memsw charge associated with @entry.
6153 void mem_cgroup_uncharge_swap(swp_entry_t entry)
6155 struct mem_cgroup *memcg;
6158 if (!do_swap_account)
6161 id = swap_cgroup_record(entry, 0);
6163 memcg = mem_cgroup_lookup(id);
6165 if (!mem_cgroup_is_root(memcg))
6166 page_counter_uncharge(&memcg->memsw, 1);
6167 mem_cgroup_swap_statistics(memcg, false);
6168 css_put(&memcg->css);
6175 * mem_cgroup_try_charge - try charging a page
6176 * @page: page to charge
6177 * @mm: mm context of the victim
6178 * @gfp_mask: reclaim mode
6179 * @memcgp: charged memcg return
6181 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6182 * pages according to @gfp_mask if necessary.
6184 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
6185 * Otherwise, an error code is returned.
6187 * After page->mapping has been set up, the caller must finalize the
6188 * charge with mem_cgroup_commit_charge(). Or abort the transaction
6189 * with mem_cgroup_cancel_charge() in case page instantiation fails.
6191 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
6192 gfp_t gfp_mask, struct mem_cgroup **memcgp)
6194 struct mem_cgroup *memcg = NULL;
6195 unsigned int nr_pages = 1;
6198 if (mem_cgroup_disabled())
6201 if (PageSwapCache(page)) {
6202 struct page_cgroup *pc = lookup_page_cgroup(page);
6204 * Every swap fault against a single page tries to charge the
6205 * page, bail as early as possible. shmem_unuse() encounters
6206 * already charged pages, too. The USED bit is protected by
6207 * the page lock, which serializes swap cache removal, which
6208 * in turn serializes uncharging.
6210 if (PageCgroupUsed(pc))
6214 if (PageTransHuge(page)) {
6215 nr_pages <<= compound_order(page);
6216 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
6219 if (do_swap_account && PageSwapCache(page))
6220 memcg = try_get_mem_cgroup_from_page(page);
6222 memcg = get_mem_cgroup_from_mm(mm);
6224 ret = try_charge(memcg, gfp_mask, nr_pages);
6226 css_put(&memcg->css);
6228 if (ret == -EINTR) {
6229 memcg = root_mem_cgroup;
6238 * mem_cgroup_commit_charge - commit a page charge
6239 * @page: page to charge
6240 * @memcg: memcg to charge the page to
6241 * @lrucare: page might be on LRU already
6243 * Finalize a charge transaction started by mem_cgroup_try_charge(),
6244 * after page->mapping has been set up. This must happen atomically
6245 * as part of the page instantiation, i.e. under the page table lock
6246 * for anonymous pages, under the page lock for page and swap cache.
6248 * In addition, the page must not be on the LRU during the commit, to
6249 * prevent racing with task migration. If it might be, use @lrucare.
6251 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
6253 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
6256 unsigned int nr_pages = 1;
6258 VM_BUG_ON_PAGE(!page->mapping, page);
6259 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
6261 if (mem_cgroup_disabled())
6264 * Swap faults will attempt to charge the same page multiple
6265 * times. But reuse_swap_page() might have removed the page
6266 * from swapcache already, so we can't check PageSwapCache().
6271 commit_charge(page, memcg, lrucare);
6273 if (PageTransHuge(page)) {
6274 nr_pages <<= compound_order(page);
6275 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
6278 local_irq_disable();
6279 mem_cgroup_charge_statistics(memcg, page, nr_pages);
6280 memcg_check_events(memcg, page);
6283 if (do_swap_account && PageSwapCache(page)) {
6284 swp_entry_t entry = { .val = page_private(page) };
6286 * The swap entry might not get freed for a long time,
6287 * let's not wait for it. The page already received a
6288 * memory+swap charge, drop the swap entry duplicate.
6290 mem_cgroup_uncharge_swap(entry);
6295 * mem_cgroup_cancel_charge - cancel a page charge
6296 * @page: page to charge
6297 * @memcg: memcg to charge the page to
6299 * Cancel a charge transaction started by mem_cgroup_try_charge().
6301 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg)
6303 unsigned int nr_pages = 1;
6305 if (mem_cgroup_disabled())
6308 * Swap faults will attempt to charge the same page multiple
6309 * times. But reuse_swap_page() might have removed the page
6310 * from swapcache already, so we can't check PageSwapCache().
6315 if (PageTransHuge(page)) {
6316 nr_pages <<= compound_order(page);
6317 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
6320 cancel_charge(memcg, nr_pages);
6323 static void uncharge_batch(struct mem_cgroup *memcg, unsigned long pgpgout,
6324 unsigned long nr_mem, unsigned long nr_memsw,
6325 unsigned long nr_anon, unsigned long nr_file,
6326 unsigned long nr_huge, struct page *dummy_page)
6328 unsigned long flags;
6330 if (!mem_cgroup_is_root(memcg)) {
6332 page_counter_uncharge(&memcg->memory, nr_mem);
6334 page_counter_uncharge(&memcg->memsw, nr_memsw);
6335 memcg_oom_recover(memcg);
6338 local_irq_save(flags);
6339 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS], nr_anon);
6340 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_CACHE], nr_file);
6341 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE], nr_huge);
6342 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT], pgpgout);
6343 __this_cpu_add(memcg->stat->nr_page_events, nr_anon + nr_file);
6344 memcg_check_events(memcg, dummy_page);
6345 local_irq_restore(flags);
6348 static void uncharge_list(struct list_head *page_list)
6350 struct mem_cgroup *memcg = NULL;
6351 unsigned long nr_memsw = 0;
6352 unsigned long nr_anon = 0;
6353 unsigned long nr_file = 0;
6354 unsigned long nr_huge = 0;
6355 unsigned long pgpgout = 0;
6356 unsigned long nr_mem = 0;
6357 struct list_head *next;
6360 next = page_list->next;
6362 unsigned int nr_pages = 1;
6363 struct page_cgroup *pc;
6365 page = list_entry(next, struct page, lru);
6366 next = page->lru.next;
6368 VM_BUG_ON_PAGE(PageLRU(page), page);
6369 VM_BUG_ON_PAGE(page_count(page), page);
6371 pc = lookup_page_cgroup(page);
6372 if (!PageCgroupUsed(pc))
6376 * Nobody should be changing or seriously looking at
6377 * pc->mem_cgroup and pc->flags at this point, we have
6378 * fully exclusive access to the page.
6381 if (memcg != pc->mem_cgroup) {
6383 uncharge_batch(memcg, pgpgout, nr_mem, nr_memsw,
6384 nr_anon, nr_file, nr_huge, page);
6385 pgpgout = nr_mem = nr_memsw = 0;
6386 nr_anon = nr_file = nr_huge = 0;
6388 memcg = pc->mem_cgroup;
6391 if (PageTransHuge(page)) {
6392 nr_pages <<= compound_order(page);
6393 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
6394 nr_huge += nr_pages;
6398 nr_anon += nr_pages;
6400 nr_file += nr_pages;
6402 if (pc->flags & PCG_MEM)
6404 if (pc->flags & PCG_MEMSW)
6405 nr_memsw += nr_pages;
6409 } while (next != page_list);
6412 uncharge_batch(memcg, pgpgout, nr_mem, nr_memsw,
6413 nr_anon, nr_file, nr_huge, page);
6417 * mem_cgroup_uncharge - uncharge a page
6418 * @page: page to uncharge
6420 * Uncharge a page previously charged with mem_cgroup_try_charge() and
6421 * mem_cgroup_commit_charge().
6423 void mem_cgroup_uncharge(struct page *page)
6425 struct page_cgroup *pc;
6427 if (mem_cgroup_disabled())
6430 /* Don't touch page->lru of any random page, pre-check: */
6431 pc = lookup_page_cgroup(page);
6432 if (!PageCgroupUsed(pc))
6435 INIT_LIST_HEAD(&page->lru);
6436 uncharge_list(&page->lru);
6440 * mem_cgroup_uncharge_list - uncharge a list of page
6441 * @page_list: list of pages to uncharge
6443 * Uncharge a list of pages previously charged with
6444 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
6446 void mem_cgroup_uncharge_list(struct list_head *page_list)
6448 if (mem_cgroup_disabled())
6451 if (!list_empty(page_list))
6452 uncharge_list(page_list);
6456 * mem_cgroup_migrate - migrate a charge to another page
6457 * @oldpage: currently charged page
6458 * @newpage: page to transfer the charge to
6459 * @lrucare: both pages might be on the LRU already
6461 * Migrate the charge from @oldpage to @newpage.
6463 * Both pages must be locked, @newpage->mapping must be set up.
6465 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage,
6468 struct page_cgroup *pc;
6471 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6472 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6473 VM_BUG_ON_PAGE(!lrucare && PageLRU(oldpage), oldpage);
6474 VM_BUG_ON_PAGE(!lrucare && PageLRU(newpage), newpage);
6475 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6476 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6479 if (mem_cgroup_disabled())
6482 /* Page cache replacement: new page already charged? */
6483 pc = lookup_page_cgroup(newpage);
6484 if (PageCgroupUsed(pc))
6487 /* Re-entrant migration: old page already uncharged? */
6488 pc = lookup_page_cgroup(oldpage);
6489 if (!PageCgroupUsed(pc))
6492 VM_BUG_ON_PAGE(!(pc->flags & PCG_MEM), oldpage);
6493 VM_BUG_ON_PAGE(do_swap_account && !(pc->flags & PCG_MEMSW), oldpage);
6496 lock_page_lru(oldpage, &isolated);
6501 unlock_page_lru(oldpage, isolated);
6503 commit_charge(newpage, pc->mem_cgroup, lrucare);
6507 * subsys_initcall() for memory controller.
6509 * Some parts like hotcpu_notifier() have to be initialized from this context
6510 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
6511 * everything that doesn't depend on a specific mem_cgroup structure should
6512 * be initialized from here.
6514 static int __init mem_cgroup_init(void)
6516 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
6517 enable_swap_cgroup();
6518 mem_cgroup_soft_limit_tree_init();
6522 subsys_initcall(mem_cgroup_init);