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 */
374 #ifdef CONFIG_MEMCG_KMEM
375 static inline void memcg_kmem_set_active(struct mem_cgroup *memcg)
377 set_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
380 static bool memcg_kmem_is_active(struct mem_cgroup *memcg)
382 return test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
387 /* Stuffs for move charges at task migration. */
389 * Types of charges to be moved. "move_charge_at_immitgrate" and
390 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
393 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
394 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
398 /* "mc" and its members are protected by cgroup_mutex */
399 static struct move_charge_struct {
400 spinlock_t lock; /* for from, to */
401 struct mem_cgroup *from;
402 struct mem_cgroup *to;
403 unsigned long immigrate_flags;
404 unsigned long precharge;
405 unsigned long moved_charge;
406 unsigned long moved_swap;
407 struct task_struct *moving_task; /* a task moving charges */
408 wait_queue_head_t waitq; /* a waitq for other context */
410 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
411 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
414 static bool move_anon(void)
416 return test_bit(MOVE_CHARGE_TYPE_ANON, &mc.immigrate_flags);
419 static bool move_file(void)
421 return test_bit(MOVE_CHARGE_TYPE_FILE, &mc.immigrate_flags);
425 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
426 * limit reclaim to prevent infinite loops, if they ever occur.
428 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
429 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
432 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
433 MEM_CGROUP_CHARGE_TYPE_ANON,
434 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
435 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
439 /* for encoding cft->private value on file */
447 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
448 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
449 #define MEMFILE_ATTR(val) ((val) & 0xffff)
450 /* Used for OOM nofiier */
451 #define OOM_CONTROL (0)
454 * The memcg_create_mutex will be held whenever a new cgroup is created.
455 * As a consequence, any change that needs to protect against new child cgroups
456 * appearing has to hold it as well.
458 static DEFINE_MUTEX(memcg_create_mutex);
460 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
462 return s ? container_of(s, struct mem_cgroup, css) : NULL;
465 /* Some nice accessors for the vmpressure. */
466 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
469 memcg = root_mem_cgroup;
470 return &memcg->vmpressure;
473 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
475 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
478 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
480 return (memcg == root_mem_cgroup);
484 * We restrict the id in the range of [1, 65535], so it can fit into
487 #define MEM_CGROUP_ID_MAX USHRT_MAX
489 static inline unsigned short mem_cgroup_id(struct mem_cgroup *memcg)
491 return memcg->css.id;
494 static inline struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
496 struct cgroup_subsys_state *css;
498 css = css_from_id(id, &memory_cgrp_subsys);
499 return mem_cgroup_from_css(css);
502 /* Writing them here to avoid exposing memcg's inner layout */
503 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
505 void sock_update_memcg(struct sock *sk)
507 if (mem_cgroup_sockets_enabled) {
508 struct mem_cgroup *memcg;
509 struct cg_proto *cg_proto;
511 BUG_ON(!sk->sk_prot->proto_cgroup);
513 /* Socket cloning can throw us here with sk_cgrp already
514 * filled. It won't however, necessarily happen from
515 * process context. So the test for root memcg given
516 * the current task's memcg won't help us in this case.
518 * Respecting the original socket's memcg is a better
519 * decision in this case.
522 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
523 css_get(&sk->sk_cgrp->memcg->css);
528 memcg = mem_cgroup_from_task(current);
529 cg_proto = sk->sk_prot->proto_cgroup(memcg);
530 if (!mem_cgroup_is_root(memcg) &&
531 memcg_proto_active(cg_proto) &&
532 css_tryget_online(&memcg->css)) {
533 sk->sk_cgrp = cg_proto;
538 EXPORT_SYMBOL(sock_update_memcg);
540 void sock_release_memcg(struct sock *sk)
542 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
543 struct mem_cgroup *memcg;
544 WARN_ON(!sk->sk_cgrp->memcg);
545 memcg = sk->sk_cgrp->memcg;
546 css_put(&sk->sk_cgrp->memcg->css);
550 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
552 if (!memcg || mem_cgroup_is_root(memcg))
555 return &memcg->tcp_mem;
557 EXPORT_SYMBOL(tcp_proto_cgroup);
559 static void disarm_sock_keys(struct mem_cgroup *memcg)
561 if (!memcg_proto_activated(&memcg->tcp_mem))
563 static_key_slow_dec(&memcg_socket_limit_enabled);
566 static void disarm_sock_keys(struct mem_cgroup *memcg)
571 #ifdef CONFIG_MEMCG_KMEM
573 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
574 * The main reason for not using cgroup id for this:
575 * this works better in sparse environments, where we have a lot of memcgs,
576 * but only a few kmem-limited. Or also, if we have, for instance, 200
577 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
578 * 200 entry array for that.
580 * The current size of the caches array is stored in
581 * memcg_limited_groups_array_size. It will double each time we have to
584 static DEFINE_IDA(kmem_limited_groups);
585 int memcg_limited_groups_array_size;
588 * MIN_SIZE is different than 1, because we would like to avoid going through
589 * the alloc/free process all the time. In a small machine, 4 kmem-limited
590 * cgroups is a reasonable guess. In the future, it could be a parameter or
591 * tunable, but that is strictly not necessary.
593 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
594 * this constant directly from cgroup, but it is understandable that this is
595 * better kept as an internal representation in cgroup.c. In any case, the
596 * cgrp_id space is not getting any smaller, and we don't have to necessarily
597 * increase ours as well if it increases.
599 #define MEMCG_CACHES_MIN_SIZE 4
600 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
603 * A lot of the calls to the cache allocation functions are expected to be
604 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
605 * conditional to this static branch, we'll have to allow modules that does
606 * kmem_cache_alloc and the such to see this symbol as well
608 struct static_key memcg_kmem_enabled_key;
609 EXPORT_SYMBOL(memcg_kmem_enabled_key);
611 static void memcg_free_cache_id(int id);
613 static void disarm_kmem_keys(struct mem_cgroup *memcg)
615 if (memcg_kmem_is_active(memcg)) {
616 static_key_slow_dec(&memcg_kmem_enabled_key);
617 memcg_free_cache_id(memcg->kmemcg_id);
620 * This check can't live in kmem destruction function,
621 * since the charges will outlive the cgroup
623 WARN_ON(page_counter_read(&memcg->kmem));
626 static void disarm_kmem_keys(struct mem_cgroup *memcg)
629 #endif /* CONFIG_MEMCG_KMEM */
631 static void disarm_static_keys(struct mem_cgroup *memcg)
633 disarm_sock_keys(memcg);
634 disarm_kmem_keys(memcg);
637 static struct mem_cgroup_per_zone *
638 mem_cgroup_zone_zoneinfo(struct mem_cgroup *memcg, struct zone *zone)
640 int nid = zone_to_nid(zone);
641 int zid = zone_idx(zone);
643 return &memcg->nodeinfo[nid]->zoneinfo[zid];
646 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
651 static struct mem_cgroup_per_zone *
652 mem_cgroup_page_zoneinfo(struct mem_cgroup *memcg, struct page *page)
654 int nid = page_to_nid(page);
655 int zid = page_zonenum(page);
657 return &memcg->nodeinfo[nid]->zoneinfo[zid];
660 static struct mem_cgroup_tree_per_zone *
661 soft_limit_tree_node_zone(int nid, int zid)
663 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
666 static struct mem_cgroup_tree_per_zone *
667 soft_limit_tree_from_page(struct page *page)
669 int nid = page_to_nid(page);
670 int zid = page_zonenum(page);
672 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
675 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone *mz,
676 struct mem_cgroup_tree_per_zone *mctz,
677 unsigned long new_usage_in_excess)
679 struct rb_node **p = &mctz->rb_root.rb_node;
680 struct rb_node *parent = NULL;
681 struct mem_cgroup_per_zone *mz_node;
686 mz->usage_in_excess = new_usage_in_excess;
687 if (!mz->usage_in_excess)
691 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
693 if (mz->usage_in_excess < mz_node->usage_in_excess)
696 * We can't avoid mem cgroups that are over their soft
697 * limit by the same amount
699 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
702 rb_link_node(&mz->tree_node, parent, p);
703 rb_insert_color(&mz->tree_node, &mctz->rb_root);
707 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
708 struct mem_cgroup_tree_per_zone *mctz)
712 rb_erase(&mz->tree_node, &mctz->rb_root);
716 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
717 struct mem_cgroup_tree_per_zone *mctz)
721 spin_lock_irqsave(&mctz->lock, flags);
722 __mem_cgroup_remove_exceeded(mz, mctz);
723 spin_unlock_irqrestore(&mctz->lock, flags);
726 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
728 unsigned long nr_pages = page_counter_read(&memcg->memory);
729 unsigned long soft_limit = ACCESS_ONCE(memcg->soft_limit);
730 unsigned long excess = 0;
732 if (nr_pages > soft_limit)
733 excess = nr_pages - soft_limit;
738 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
740 unsigned long excess;
741 struct mem_cgroup_per_zone *mz;
742 struct mem_cgroup_tree_per_zone *mctz;
744 mctz = soft_limit_tree_from_page(page);
746 * Necessary to update all ancestors when hierarchy is used.
747 * because their event counter is not touched.
749 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
750 mz = mem_cgroup_page_zoneinfo(memcg, page);
751 excess = soft_limit_excess(memcg);
753 * We have to update the tree if mz is on RB-tree or
754 * mem is over its softlimit.
756 if (excess || mz->on_tree) {
759 spin_lock_irqsave(&mctz->lock, flags);
760 /* if on-tree, remove it */
762 __mem_cgroup_remove_exceeded(mz, mctz);
764 * Insert again. mz->usage_in_excess will be updated.
765 * If excess is 0, no tree ops.
767 __mem_cgroup_insert_exceeded(mz, mctz, excess);
768 spin_unlock_irqrestore(&mctz->lock, flags);
773 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
775 struct mem_cgroup_tree_per_zone *mctz;
776 struct mem_cgroup_per_zone *mz;
780 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
781 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
782 mctz = soft_limit_tree_node_zone(nid, zid);
783 mem_cgroup_remove_exceeded(mz, mctz);
788 static struct mem_cgroup_per_zone *
789 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
791 struct rb_node *rightmost = NULL;
792 struct mem_cgroup_per_zone *mz;
796 rightmost = rb_last(&mctz->rb_root);
798 goto done; /* Nothing to reclaim from */
800 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
802 * Remove the node now but someone else can add it back,
803 * we will to add it back at the end of reclaim to its correct
804 * position in the tree.
806 __mem_cgroup_remove_exceeded(mz, mctz);
807 if (!soft_limit_excess(mz->memcg) ||
808 !css_tryget_online(&mz->memcg->css))
814 static struct mem_cgroup_per_zone *
815 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
817 struct mem_cgroup_per_zone *mz;
819 spin_lock_irq(&mctz->lock);
820 mz = __mem_cgroup_largest_soft_limit_node(mctz);
821 spin_unlock_irq(&mctz->lock);
826 * Implementation Note: reading percpu statistics for memcg.
828 * Both of vmstat[] and percpu_counter has threshold and do periodic
829 * synchronization to implement "quick" read. There are trade-off between
830 * reading cost and precision of value. Then, we may have a chance to implement
831 * a periodic synchronizion of counter in memcg's counter.
833 * But this _read() function is used for user interface now. The user accounts
834 * memory usage by memory cgroup and he _always_ requires exact value because
835 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
836 * have to visit all online cpus and make sum. So, for now, unnecessary
837 * synchronization is not implemented. (just implemented for cpu hotplug)
839 * If there are kernel internal actions which can make use of some not-exact
840 * value, and reading all cpu value can be performance bottleneck in some
841 * common workload, threashold and synchonization as vmstat[] should be
844 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
845 enum mem_cgroup_stat_index idx)
851 for_each_online_cpu(cpu)
852 val += per_cpu(memcg->stat->count[idx], cpu);
853 #ifdef CONFIG_HOTPLUG_CPU
854 spin_lock(&memcg->pcp_counter_lock);
855 val += memcg->nocpu_base.count[idx];
856 spin_unlock(&memcg->pcp_counter_lock);
862 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
863 enum mem_cgroup_events_index idx)
865 unsigned long val = 0;
869 for_each_online_cpu(cpu)
870 val += per_cpu(memcg->stat->events[idx], cpu);
871 #ifdef CONFIG_HOTPLUG_CPU
872 spin_lock(&memcg->pcp_counter_lock);
873 val += memcg->nocpu_base.events[idx];
874 spin_unlock(&memcg->pcp_counter_lock);
880 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
885 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
886 * counted as CACHE even if it's on ANON LRU.
889 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
892 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
895 if (PageTransHuge(page))
896 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
899 /* pagein of a big page is an event. So, ignore page size */
901 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
903 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
904 nr_pages = -nr_pages; /* for event */
907 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
910 unsigned long mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
912 struct mem_cgroup_per_zone *mz;
914 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
915 return mz->lru_size[lru];
918 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
920 unsigned int lru_mask)
922 unsigned long nr = 0;
925 VM_BUG_ON((unsigned)nid >= nr_node_ids);
927 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
928 struct mem_cgroup_per_zone *mz;
932 if (!(BIT(lru) & lru_mask))
934 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
935 nr += mz->lru_size[lru];
941 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
942 unsigned int lru_mask)
944 unsigned long nr = 0;
947 for_each_node_state(nid, N_MEMORY)
948 nr += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
952 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
953 enum mem_cgroup_events_target target)
955 unsigned long val, next;
957 val = __this_cpu_read(memcg->stat->nr_page_events);
958 next = __this_cpu_read(memcg->stat->targets[target]);
959 /* from time_after() in jiffies.h */
960 if ((long)next - (long)val < 0) {
962 case MEM_CGROUP_TARGET_THRESH:
963 next = val + THRESHOLDS_EVENTS_TARGET;
965 case MEM_CGROUP_TARGET_SOFTLIMIT:
966 next = val + SOFTLIMIT_EVENTS_TARGET;
968 case MEM_CGROUP_TARGET_NUMAINFO:
969 next = val + NUMAINFO_EVENTS_TARGET;
974 __this_cpu_write(memcg->stat->targets[target], next);
981 * Check events in order.
984 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
986 /* threshold event is triggered in finer grain than soft limit */
987 if (unlikely(mem_cgroup_event_ratelimit(memcg,
988 MEM_CGROUP_TARGET_THRESH))) {
990 bool do_numainfo __maybe_unused;
992 do_softlimit = mem_cgroup_event_ratelimit(memcg,
993 MEM_CGROUP_TARGET_SOFTLIMIT);
995 do_numainfo = mem_cgroup_event_ratelimit(memcg,
996 MEM_CGROUP_TARGET_NUMAINFO);
998 mem_cgroup_threshold(memcg);
999 if (unlikely(do_softlimit))
1000 mem_cgroup_update_tree(memcg, page);
1001 #if MAX_NUMNODES > 1
1002 if (unlikely(do_numainfo))
1003 atomic_inc(&memcg->numainfo_events);
1008 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1011 * mm_update_next_owner() may clear mm->owner to NULL
1012 * if it races with swapoff, page migration, etc.
1013 * So this can be called with p == NULL.
1018 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
1021 static struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
1023 struct mem_cgroup *memcg = NULL;
1028 * Page cache insertions can happen withou an
1029 * actual mm context, e.g. during disk probing
1030 * on boot, loopback IO, acct() writes etc.
1033 memcg = root_mem_cgroup;
1035 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1036 if (unlikely(!memcg))
1037 memcg = root_mem_cgroup;
1039 } while (!css_tryget_online(&memcg->css));
1045 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1046 * @root: hierarchy root
1047 * @prev: previously returned memcg, NULL on first invocation
1048 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1050 * Returns references to children of the hierarchy below @root, or
1051 * @root itself, or %NULL after a full round-trip.
1053 * Caller must pass the return value in @prev on subsequent
1054 * invocations for reference counting, or use mem_cgroup_iter_break()
1055 * to cancel a hierarchy walk before the round-trip is complete.
1057 * Reclaimers can specify a zone and a priority level in @reclaim to
1058 * divide up the memcgs in the hierarchy among all concurrent
1059 * reclaimers operating on the same zone and priority.
1061 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1062 struct mem_cgroup *prev,
1063 struct mem_cgroup_reclaim_cookie *reclaim)
1065 struct reclaim_iter *uninitialized_var(iter);
1066 struct cgroup_subsys_state *css = NULL;
1067 struct mem_cgroup *memcg = NULL;
1068 struct mem_cgroup *pos = NULL;
1070 if (mem_cgroup_disabled())
1074 root = root_mem_cgroup;
1076 if (prev && !reclaim)
1079 if (!root->use_hierarchy && root != root_mem_cgroup) {
1088 struct mem_cgroup_per_zone *mz;
1090 mz = mem_cgroup_zone_zoneinfo(root, reclaim->zone);
1091 iter = &mz->iter[reclaim->priority];
1093 if (prev && reclaim->generation != iter->generation)
1097 pos = ACCESS_ONCE(iter->position);
1099 * A racing update may change the position and
1100 * put the last reference, hence css_tryget(),
1101 * or retry to see the updated position.
1103 } while (pos && !css_tryget(&pos->css));
1110 css = css_next_descendant_pre(css, &root->css);
1113 * Reclaimers share the hierarchy walk, and a
1114 * new one might jump in right at the end of
1115 * the hierarchy - make sure they see at least
1116 * one group and restart from the beginning.
1124 * Verify the css and acquire a reference. The root
1125 * is provided by the caller, so we know it's alive
1126 * and kicking, and don't take an extra reference.
1128 memcg = mem_cgroup_from_css(css);
1130 if (css == &root->css)
1133 if (css_tryget(css)) {
1135 * Make sure the memcg is initialized:
1136 * mem_cgroup_css_online() orders the the
1137 * initialization against setting the flag.
1139 if (smp_load_acquire(&memcg->initialized))
1149 if (cmpxchg(&iter->position, pos, memcg) == pos) {
1151 css_get(&memcg->css);
1157 * pairs with css_tryget when dereferencing iter->position
1166 reclaim->generation = iter->generation;
1172 if (prev && prev != root)
1173 css_put(&prev->css);
1179 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1180 * @root: hierarchy root
1181 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1183 void mem_cgroup_iter_break(struct mem_cgroup *root,
1184 struct mem_cgroup *prev)
1187 root = root_mem_cgroup;
1188 if (prev && prev != root)
1189 css_put(&prev->css);
1193 * Iteration constructs for visiting all cgroups (under a tree). If
1194 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1195 * be used for reference counting.
1197 #define for_each_mem_cgroup_tree(iter, root) \
1198 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1200 iter = mem_cgroup_iter(root, iter, NULL))
1202 #define for_each_mem_cgroup(iter) \
1203 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1205 iter = mem_cgroup_iter(NULL, iter, NULL))
1207 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1209 struct mem_cgroup *memcg;
1212 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1213 if (unlikely(!memcg))
1218 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1221 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1229 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1232 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1233 * @zone: zone of the wanted lruvec
1234 * @memcg: memcg of the wanted lruvec
1236 * Returns the lru list vector holding pages for the given @zone and
1237 * @mem. This can be the global zone lruvec, if the memory controller
1240 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1241 struct mem_cgroup *memcg)
1243 struct mem_cgroup_per_zone *mz;
1244 struct lruvec *lruvec;
1246 if (mem_cgroup_disabled()) {
1247 lruvec = &zone->lruvec;
1251 mz = mem_cgroup_zone_zoneinfo(memcg, zone);
1252 lruvec = &mz->lruvec;
1255 * Since a node can be onlined after the mem_cgroup was created,
1256 * we have to be prepared to initialize lruvec->zone here;
1257 * and if offlined then reonlined, we need to reinitialize it.
1259 if (unlikely(lruvec->zone != zone))
1260 lruvec->zone = zone;
1265 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1267 * @zone: zone of the page
1269 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1271 struct mem_cgroup_per_zone *mz;
1272 struct mem_cgroup *memcg;
1273 struct page_cgroup *pc;
1274 struct lruvec *lruvec;
1276 if (mem_cgroup_disabled()) {
1277 lruvec = &zone->lruvec;
1281 pc = lookup_page_cgroup(page);
1282 memcg = pc->mem_cgroup;
1285 * Surreptitiously switch any uncharged offlist page to root:
1286 * an uncharged page off lru does nothing to secure
1287 * its former mem_cgroup from sudden removal.
1289 * Our caller holds lru_lock, and PageCgroupUsed is updated
1290 * under page_cgroup lock: between them, they make all uses
1291 * of pc->mem_cgroup safe.
1293 if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1294 pc->mem_cgroup = memcg = root_mem_cgroup;
1296 mz = mem_cgroup_page_zoneinfo(memcg, page);
1297 lruvec = &mz->lruvec;
1300 * Since a node can be onlined after the mem_cgroup was created,
1301 * we have to be prepared to initialize lruvec->zone here;
1302 * and if offlined then reonlined, we need to reinitialize it.
1304 if (unlikely(lruvec->zone != zone))
1305 lruvec->zone = zone;
1310 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1311 * @lruvec: mem_cgroup per zone lru vector
1312 * @lru: index of lru list the page is sitting on
1313 * @nr_pages: positive when adding or negative when removing
1315 * This function must be called when a page is added to or removed from an
1318 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1321 struct mem_cgroup_per_zone *mz;
1322 unsigned long *lru_size;
1324 if (mem_cgroup_disabled())
1327 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1328 lru_size = mz->lru_size + lru;
1329 *lru_size += nr_pages;
1330 VM_BUG_ON((long)(*lru_size) < 0);
1334 * Checks whether given mem is same or in the root_mem_cgroup's
1337 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1338 struct mem_cgroup *memcg)
1340 if (root_memcg == memcg)
1342 if (!root_memcg->use_hierarchy || !memcg)
1344 return cgroup_is_descendant(memcg->css.cgroup, root_memcg->css.cgroup);
1347 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1348 struct mem_cgroup *memcg)
1353 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1358 bool task_in_mem_cgroup(struct task_struct *task,
1359 const struct mem_cgroup *memcg)
1361 struct mem_cgroup *curr = NULL;
1362 struct task_struct *p;
1365 p = find_lock_task_mm(task);
1367 curr = get_mem_cgroup_from_mm(p->mm);
1371 * All threads may have already detached their mm's, but the oom
1372 * killer still needs to detect if they have already been oom
1373 * killed to prevent needlessly killing additional tasks.
1376 curr = mem_cgroup_from_task(task);
1378 css_get(&curr->css);
1382 * We should check use_hierarchy of "memcg" not "curr". Because checking
1383 * use_hierarchy of "curr" here make this function true if hierarchy is
1384 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1385 * hierarchy(even if use_hierarchy is disabled in "memcg").
1387 ret = mem_cgroup_same_or_subtree(memcg, curr);
1388 css_put(&curr->css);
1392 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1394 unsigned long inactive_ratio;
1395 unsigned long inactive;
1396 unsigned long active;
1399 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1400 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1402 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1404 inactive_ratio = int_sqrt(10 * gb);
1408 return inactive * inactive_ratio < active;
1411 #define mem_cgroup_from_counter(counter, member) \
1412 container_of(counter, struct mem_cgroup, member)
1415 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1416 * @memcg: the memory cgroup
1418 * Returns the maximum amount of memory @mem can be charged with, in
1421 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1423 unsigned long margin = 0;
1424 unsigned long count;
1425 unsigned long limit;
1427 count = page_counter_read(&memcg->memory);
1428 limit = ACCESS_ONCE(memcg->memory.limit);
1430 margin = limit - count;
1432 if (do_swap_account) {
1433 count = page_counter_read(&memcg->memsw);
1434 limit = ACCESS_ONCE(memcg->memsw.limit);
1436 margin = min(margin, limit - count);
1442 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1445 if (mem_cgroup_disabled() || !memcg->css.parent)
1446 return vm_swappiness;
1448 return memcg->swappiness;
1452 * memcg->moving_account is used for checking possibility that some thread is
1453 * calling move_account(). When a thread on CPU-A starts moving pages under
1454 * a memcg, other threads should check memcg->moving_account under
1455 * rcu_read_lock(), like this:
1459 * memcg->moving_account+1 if (memcg->mocing_account)
1461 * synchronize_rcu() update something.
1466 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1468 atomic_inc(&memcg->moving_account);
1472 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1475 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1476 * We check NULL in callee rather than caller.
1479 atomic_dec(&memcg->moving_account);
1483 * A routine for checking "mem" is under move_account() or not.
1485 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1486 * moving cgroups. This is for waiting at high-memory pressure
1489 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1491 struct mem_cgroup *from;
1492 struct mem_cgroup *to;
1495 * Unlike task_move routines, we access mc.to, mc.from not under
1496 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1498 spin_lock(&mc.lock);
1504 ret = mem_cgroup_same_or_subtree(memcg, from)
1505 || mem_cgroup_same_or_subtree(memcg, to);
1507 spin_unlock(&mc.lock);
1511 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1513 if (mc.moving_task && current != mc.moving_task) {
1514 if (mem_cgroup_under_move(memcg)) {
1516 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1517 /* moving charge context might have finished. */
1520 finish_wait(&mc.waitq, &wait);
1528 * Take this lock when
1529 * - a code tries to modify page's memcg while it's USED.
1530 * - a code tries to modify page state accounting in a memcg.
1532 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1533 unsigned long *flags)
1535 spin_lock_irqsave(&memcg->move_lock, *flags);
1538 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1539 unsigned long *flags)
1541 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1544 #define K(x) ((x) << (PAGE_SHIFT-10))
1546 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1547 * @memcg: The memory cgroup that went over limit
1548 * @p: Task that is going to be killed
1550 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1553 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1555 /* oom_info_lock ensures that parallel ooms do not interleave */
1556 static DEFINE_MUTEX(oom_info_lock);
1557 struct mem_cgroup *iter;
1563 mutex_lock(&oom_info_lock);
1566 pr_info("Task in ");
1567 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1568 pr_info(" killed as a result of limit of ");
1569 pr_cont_cgroup_path(memcg->css.cgroup);
1574 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1575 K((u64)page_counter_read(&memcg->memory)),
1576 K((u64)memcg->memory.limit), memcg->memory.failcnt);
1577 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1578 K((u64)page_counter_read(&memcg->memsw)),
1579 K((u64)memcg->memsw.limit), memcg->memsw.failcnt);
1580 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1581 K((u64)page_counter_read(&memcg->kmem)),
1582 K((u64)memcg->kmem.limit), memcg->kmem.failcnt);
1584 for_each_mem_cgroup_tree(iter, memcg) {
1585 pr_info("Memory cgroup stats for ");
1586 pr_cont_cgroup_path(iter->css.cgroup);
1589 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1590 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1592 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1593 K(mem_cgroup_read_stat(iter, i)));
1596 for (i = 0; i < NR_LRU_LISTS; i++)
1597 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1598 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1602 mutex_unlock(&oom_info_lock);
1606 * This function returns the number of memcg under hierarchy tree. Returns
1607 * 1(self count) if no children.
1609 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1612 struct mem_cgroup *iter;
1614 for_each_mem_cgroup_tree(iter, memcg)
1620 * Return the memory (and swap, if configured) limit for a memcg.
1622 static unsigned long mem_cgroup_get_limit(struct mem_cgroup *memcg)
1624 unsigned long limit;
1626 limit = memcg->memory.limit;
1627 if (mem_cgroup_swappiness(memcg)) {
1628 unsigned long memsw_limit;
1630 memsw_limit = memcg->memsw.limit;
1631 limit = min(limit + total_swap_pages, memsw_limit);
1636 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1639 struct mem_cgroup *iter;
1640 unsigned long chosen_points = 0;
1641 unsigned long totalpages;
1642 unsigned int points = 0;
1643 struct task_struct *chosen = NULL;
1646 * If current has a pending SIGKILL or is exiting, then automatically
1647 * select it. The goal is to allow it to allocate so that it may
1648 * quickly exit and free its memory.
1650 if (fatal_signal_pending(current) || current->flags & PF_EXITING) {
1651 set_thread_flag(TIF_MEMDIE);
1655 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1656 totalpages = mem_cgroup_get_limit(memcg) ? : 1;
1657 for_each_mem_cgroup_tree(iter, memcg) {
1658 struct css_task_iter it;
1659 struct task_struct *task;
1661 css_task_iter_start(&iter->css, &it);
1662 while ((task = css_task_iter_next(&it))) {
1663 switch (oom_scan_process_thread(task, totalpages, NULL,
1665 case OOM_SCAN_SELECT:
1667 put_task_struct(chosen);
1669 chosen_points = ULONG_MAX;
1670 get_task_struct(chosen);
1672 case OOM_SCAN_CONTINUE:
1674 case OOM_SCAN_ABORT:
1675 css_task_iter_end(&it);
1676 mem_cgroup_iter_break(memcg, iter);
1678 put_task_struct(chosen);
1683 points = oom_badness(task, memcg, NULL, totalpages);
1684 if (!points || points < chosen_points)
1686 /* Prefer thread group leaders for display purposes */
1687 if (points == chosen_points &&
1688 thread_group_leader(chosen))
1692 put_task_struct(chosen);
1694 chosen_points = points;
1695 get_task_struct(chosen);
1697 css_task_iter_end(&it);
1702 points = chosen_points * 1000 / totalpages;
1703 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1704 NULL, "Memory cgroup out of memory");
1708 * test_mem_cgroup_node_reclaimable
1709 * @memcg: the target memcg
1710 * @nid: the node ID to be checked.
1711 * @noswap : specify true here if the user wants flle only information.
1713 * This function returns whether the specified memcg contains any
1714 * reclaimable pages on a node. Returns true if there are any reclaimable
1715 * pages in the node.
1717 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1718 int nid, bool noswap)
1720 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1722 if (noswap || !total_swap_pages)
1724 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1729 #if MAX_NUMNODES > 1
1732 * Always updating the nodemask is not very good - even if we have an empty
1733 * list or the wrong list here, we can start from some node and traverse all
1734 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1737 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1741 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1742 * pagein/pageout changes since the last update.
1744 if (!atomic_read(&memcg->numainfo_events))
1746 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1749 /* make a nodemask where this memcg uses memory from */
1750 memcg->scan_nodes = node_states[N_MEMORY];
1752 for_each_node_mask(nid, node_states[N_MEMORY]) {
1754 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1755 node_clear(nid, memcg->scan_nodes);
1758 atomic_set(&memcg->numainfo_events, 0);
1759 atomic_set(&memcg->numainfo_updating, 0);
1763 * Selecting a node where we start reclaim from. Because what we need is just
1764 * reducing usage counter, start from anywhere is O,K. Considering
1765 * memory reclaim from current node, there are pros. and cons.
1767 * Freeing memory from current node means freeing memory from a node which
1768 * we'll use or we've used. So, it may make LRU bad. And if several threads
1769 * hit limits, it will see a contention on a node. But freeing from remote
1770 * node means more costs for memory reclaim because of memory latency.
1772 * Now, we use round-robin. Better algorithm is welcomed.
1774 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1778 mem_cgroup_may_update_nodemask(memcg);
1779 node = memcg->last_scanned_node;
1781 node = next_node(node, memcg->scan_nodes);
1782 if (node == MAX_NUMNODES)
1783 node = first_node(memcg->scan_nodes);
1785 * We call this when we hit limit, not when pages are added to LRU.
1786 * No LRU may hold pages because all pages are UNEVICTABLE or
1787 * memcg is too small and all pages are not on LRU. In that case,
1788 * we use curret node.
1790 if (unlikely(node == MAX_NUMNODES))
1791 node = numa_node_id();
1793 memcg->last_scanned_node = node;
1798 * Check all nodes whether it contains reclaimable pages or not.
1799 * For quick scan, we make use of scan_nodes. This will allow us to skip
1800 * unused nodes. But scan_nodes is lazily updated and may not cotain
1801 * enough new information. We need to do double check.
1803 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1808 * quick check...making use of scan_node.
1809 * We can skip unused nodes.
1811 if (!nodes_empty(memcg->scan_nodes)) {
1812 for (nid = first_node(memcg->scan_nodes);
1814 nid = next_node(nid, memcg->scan_nodes)) {
1816 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1821 * Check rest of nodes.
1823 for_each_node_state(nid, N_MEMORY) {
1824 if (node_isset(nid, memcg->scan_nodes))
1826 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1833 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1838 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1840 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
1844 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1847 unsigned long *total_scanned)
1849 struct mem_cgroup *victim = NULL;
1852 unsigned long excess;
1853 unsigned long nr_scanned;
1854 struct mem_cgroup_reclaim_cookie reclaim = {
1859 excess = soft_limit_excess(root_memcg);
1862 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1867 * If we have not been able to reclaim
1868 * anything, it might because there are
1869 * no reclaimable pages under this hierarchy
1874 * We want to do more targeted reclaim.
1875 * excess >> 2 is not to excessive so as to
1876 * reclaim too much, nor too less that we keep
1877 * coming back to reclaim from this cgroup
1879 if (total >= (excess >> 2) ||
1880 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1885 if (!mem_cgroup_reclaimable(victim, false))
1887 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1889 *total_scanned += nr_scanned;
1890 if (!soft_limit_excess(root_memcg))
1893 mem_cgroup_iter_break(root_memcg, victim);
1897 #ifdef CONFIG_LOCKDEP
1898 static struct lockdep_map memcg_oom_lock_dep_map = {
1899 .name = "memcg_oom_lock",
1903 static DEFINE_SPINLOCK(memcg_oom_lock);
1906 * Check OOM-Killer is already running under our hierarchy.
1907 * If someone is running, return false.
1909 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1911 struct mem_cgroup *iter, *failed = NULL;
1913 spin_lock(&memcg_oom_lock);
1915 for_each_mem_cgroup_tree(iter, memcg) {
1916 if (iter->oom_lock) {
1918 * this subtree of our hierarchy is already locked
1919 * so we cannot give a lock.
1922 mem_cgroup_iter_break(memcg, iter);
1925 iter->oom_lock = true;
1930 * OK, we failed to lock the whole subtree so we have
1931 * to clean up what we set up to the failing subtree
1933 for_each_mem_cgroup_tree(iter, memcg) {
1934 if (iter == failed) {
1935 mem_cgroup_iter_break(memcg, iter);
1938 iter->oom_lock = false;
1941 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1943 spin_unlock(&memcg_oom_lock);
1948 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1950 struct mem_cgroup *iter;
1952 spin_lock(&memcg_oom_lock);
1953 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1954 for_each_mem_cgroup_tree(iter, memcg)
1955 iter->oom_lock = false;
1956 spin_unlock(&memcg_oom_lock);
1959 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1961 struct mem_cgroup *iter;
1963 for_each_mem_cgroup_tree(iter, memcg)
1964 atomic_inc(&iter->under_oom);
1967 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1969 struct mem_cgroup *iter;
1972 * When a new child is created while the hierarchy is under oom,
1973 * mem_cgroup_oom_lock() may not be called. We have to use
1974 * atomic_add_unless() here.
1976 for_each_mem_cgroup_tree(iter, memcg)
1977 atomic_add_unless(&iter->under_oom, -1, 0);
1980 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1982 struct oom_wait_info {
1983 struct mem_cgroup *memcg;
1987 static int memcg_oom_wake_function(wait_queue_t *wait,
1988 unsigned mode, int sync, void *arg)
1990 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1991 struct mem_cgroup *oom_wait_memcg;
1992 struct oom_wait_info *oom_wait_info;
1994 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1995 oom_wait_memcg = oom_wait_info->memcg;
1998 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
1999 * Then we can use css_is_ancestor without taking care of RCU.
2001 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
2002 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
2004 return autoremove_wake_function(wait, mode, sync, arg);
2007 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
2009 atomic_inc(&memcg->oom_wakeups);
2010 /* for filtering, pass "memcg" as argument. */
2011 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
2014 static void memcg_oom_recover(struct mem_cgroup *memcg)
2016 if (memcg && atomic_read(&memcg->under_oom))
2017 memcg_wakeup_oom(memcg);
2020 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
2022 if (!current->memcg_oom.may_oom)
2025 * We are in the middle of the charge context here, so we
2026 * don't want to block when potentially sitting on a callstack
2027 * that holds all kinds of filesystem and mm locks.
2029 * Also, the caller may handle a failed allocation gracefully
2030 * (like optional page cache readahead) and so an OOM killer
2031 * invocation might not even be necessary.
2033 * That's why we don't do anything here except remember the
2034 * OOM context and then deal with it at the end of the page
2035 * fault when the stack is unwound, the locks are released,
2036 * and when we know whether the fault was overall successful.
2038 css_get(&memcg->css);
2039 current->memcg_oom.memcg = memcg;
2040 current->memcg_oom.gfp_mask = mask;
2041 current->memcg_oom.order = order;
2045 * mem_cgroup_oom_synchronize - complete memcg OOM handling
2046 * @handle: actually kill/wait or just clean up the OOM state
2048 * This has to be called at the end of a page fault if the memcg OOM
2049 * handler was enabled.
2051 * Memcg supports userspace OOM handling where failed allocations must
2052 * sleep on a waitqueue until the userspace task resolves the
2053 * situation. Sleeping directly in the charge context with all kinds
2054 * of locks held is not a good idea, instead we remember an OOM state
2055 * in the task and mem_cgroup_oom_synchronize() has to be called at
2056 * the end of the page fault to complete the OOM handling.
2058 * Returns %true if an ongoing memcg OOM situation was detected and
2059 * completed, %false otherwise.
2061 bool mem_cgroup_oom_synchronize(bool handle)
2063 struct mem_cgroup *memcg = current->memcg_oom.memcg;
2064 struct oom_wait_info owait;
2067 /* OOM is global, do not handle */
2074 owait.memcg = memcg;
2075 owait.wait.flags = 0;
2076 owait.wait.func = memcg_oom_wake_function;
2077 owait.wait.private = current;
2078 INIT_LIST_HEAD(&owait.wait.task_list);
2080 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2081 mem_cgroup_mark_under_oom(memcg);
2083 locked = mem_cgroup_oom_trylock(memcg);
2086 mem_cgroup_oom_notify(memcg);
2088 if (locked && !memcg->oom_kill_disable) {
2089 mem_cgroup_unmark_under_oom(memcg);
2090 finish_wait(&memcg_oom_waitq, &owait.wait);
2091 mem_cgroup_out_of_memory(memcg, current->memcg_oom.gfp_mask,
2092 current->memcg_oom.order);
2095 mem_cgroup_unmark_under_oom(memcg);
2096 finish_wait(&memcg_oom_waitq, &owait.wait);
2100 mem_cgroup_oom_unlock(memcg);
2102 * There is no guarantee that an OOM-lock contender
2103 * sees the wakeups triggered by the OOM kill
2104 * uncharges. Wake any sleepers explicitely.
2106 memcg_oom_recover(memcg);
2109 current->memcg_oom.memcg = NULL;
2110 css_put(&memcg->css);
2115 * mem_cgroup_begin_page_stat - begin a page state statistics transaction
2116 * @page: page that is going to change accounted state
2117 * @locked: &memcg->move_lock slowpath was taken
2118 * @flags: IRQ-state flags for &memcg->move_lock
2120 * This function must mark the beginning of an accounted page state
2121 * change to prevent double accounting when the page is concurrently
2122 * being moved to another memcg:
2124 * memcg = mem_cgroup_begin_page_stat(page, &locked, &flags);
2125 * if (TestClearPageState(page))
2126 * mem_cgroup_update_page_stat(memcg, state, -1);
2127 * mem_cgroup_end_page_stat(memcg, locked, flags);
2129 * The RCU lock is held throughout the transaction. The fast path can
2130 * get away without acquiring the memcg->move_lock (@locked is false)
2131 * because page moving starts with an RCU grace period.
2133 * The RCU lock also protects the memcg from being freed when the page
2134 * state that is going to change is the only thing preventing the page
2135 * from being uncharged. E.g. end-writeback clearing PageWriteback(),
2136 * which allows migration to go ahead and uncharge the page before the
2137 * account transaction might be complete.
2139 struct mem_cgroup *mem_cgroup_begin_page_stat(struct page *page,
2141 unsigned long *flags)
2143 struct mem_cgroup *memcg;
2144 struct page_cgroup *pc;
2148 if (mem_cgroup_disabled())
2151 pc = lookup_page_cgroup(page);
2153 memcg = pc->mem_cgroup;
2154 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2158 if (atomic_read(&memcg->moving_account) <= 0)
2161 move_lock_mem_cgroup(memcg, flags);
2162 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2163 move_unlock_mem_cgroup(memcg, flags);
2172 * mem_cgroup_end_page_stat - finish a page state statistics transaction
2173 * @memcg: the memcg that was accounted against
2174 * @locked: value received from mem_cgroup_begin_page_stat()
2175 * @flags: value received from mem_cgroup_begin_page_stat()
2177 void mem_cgroup_end_page_stat(struct mem_cgroup *memcg, bool locked,
2178 unsigned long flags)
2180 if (memcg && locked)
2181 move_unlock_mem_cgroup(memcg, &flags);
2187 * mem_cgroup_update_page_stat - update page state statistics
2188 * @memcg: memcg to account against
2189 * @idx: page state item to account
2190 * @val: number of pages (positive or negative)
2192 * See mem_cgroup_begin_page_stat() for locking requirements.
2194 void mem_cgroup_update_page_stat(struct mem_cgroup *memcg,
2195 enum mem_cgroup_stat_index idx, int val)
2197 VM_BUG_ON(!rcu_read_lock_held());
2200 this_cpu_add(memcg->stat->count[idx], val);
2204 * size of first charge trial. "32" comes from vmscan.c's magic value.
2205 * TODO: maybe necessary to use big numbers in big irons.
2207 #define CHARGE_BATCH 32U
2208 struct memcg_stock_pcp {
2209 struct mem_cgroup *cached; /* this never be root cgroup */
2210 unsigned int nr_pages;
2211 struct work_struct work;
2212 unsigned long flags;
2213 #define FLUSHING_CACHED_CHARGE 0
2215 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2216 static DEFINE_MUTEX(percpu_charge_mutex);
2219 * consume_stock: Try to consume stocked charge on this cpu.
2220 * @memcg: memcg to consume from.
2221 * @nr_pages: how many pages to charge.
2223 * The charges will only happen if @memcg matches the current cpu's memcg
2224 * stock, and at least @nr_pages are available in that stock. Failure to
2225 * service an allocation will refill the stock.
2227 * returns true if successful, false otherwise.
2229 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2231 struct memcg_stock_pcp *stock;
2234 if (nr_pages > CHARGE_BATCH)
2237 stock = &get_cpu_var(memcg_stock);
2238 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2239 stock->nr_pages -= nr_pages;
2242 put_cpu_var(memcg_stock);
2247 * Returns stocks cached in percpu and reset cached information.
2249 static void drain_stock(struct memcg_stock_pcp *stock)
2251 struct mem_cgroup *old = stock->cached;
2253 if (stock->nr_pages) {
2254 page_counter_uncharge(&old->memory, stock->nr_pages);
2255 if (do_swap_account)
2256 page_counter_uncharge(&old->memsw, stock->nr_pages);
2257 css_put_many(&old->css, stock->nr_pages);
2258 stock->nr_pages = 0;
2260 stock->cached = NULL;
2264 * This must be called under preempt disabled or must be called by
2265 * a thread which is pinned to local cpu.
2267 static void drain_local_stock(struct work_struct *dummy)
2269 struct memcg_stock_pcp *stock = this_cpu_ptr(&memcg_stock);
2271 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2274 static void __init memcg_stock_init(void)
2278 for_each_possible_cpu(cpu) {
2279 struct memcg_stock_pcp *stock =
2280 &per_cpu(memcg_stock, cpu);
2281 INIT_WORK(&stock->work, drain_local_stock);
2286 * Cache charges(val) to local per_cpu area.
2287 * This will be consumed by consume_stock() function, later.
2289 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2291 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2293 if (stock->cached != memcg) { /* reset if necessary */
2295 stock->cached = memcg;
2297 stock->nr_pages += nr_pages;
2298 put_cpu_var(memcg_stock);
2302 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2303 * of the hierarchy under it.
2305 static void drain_all_stock(struct mem_cgroup *root_memcg)
2309 /* If someone's already draining, avoid adding running more workers. */
2310 if (!mutex_trylock(&percpu_charge_mutex))
2312 /* Notify other cpus that system-wide "drain" is running */
2315 for_each_online_cpu(cpu) {
2316 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2317 struct mem_cgroup *memcg;
2319 memcg = stock->cached;
2320 if (!memcg || !stock->nr_pages)
2322 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2324 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2326 drain_local_stock(&stock->work);
2328 schedule_work_on(cpu, &stock->work);
2333 mutex_unlock(&percpu_charge_mutex);
2337 * This function drains percpu counter value from DEAD cpu and
2338 * move it to local cpu. Note that this function can be preempted.
2340 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2344 spin_lock(&memcg->pcp_counter_lock);
2345 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2346 long x = per_cpu(memcg->stat->count[i], cpu);
2348 per_cpu(memcg->stat->count[i], cpu) = 0;
2349 memcg->nocpu_base.count[i] += x;
2351 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2352 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2354 per_cpu(memcg->stat->events[i], cpu) = 0;
2355 memcg->nocpu_base.events[i] += x;
2357 spin_unlock(&memcg->pcp_counter_lock);
2360 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
2361 unsigned long action,
2364 int cpu = (unsigned long)hcpu;
2365 struct memcg_stock_pcp *stock;
2366 struct mem_cgroup *iter;
2368 if (action == CPU_ONLINE)
2371 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2374 for_each_mem_cgroup(iter)
2375 mem_cgroup_drain_pcp_counter(iter, cpu);
2377 stock = &per_cpu(memcg_stock, cpu);
2382 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2383 unsigned int nr_pages)
2385 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2386 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2387 struct mem_cgroup *mem_over_limit;
2388 struct page_counter *counter;
2389 unsigned long nr_reclaimed;
2390 bool may_swap = true;
2391 bool drained = false;
2394 if (mem_cgroup_is_root(memcg))
2397 if (consume_stock(memcg, nr_pages))
2400 if (!do_swap_account ||
2401 !page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2402 if (!page_counter_try_charge(&memcg->memory, batch, &counter))
2404 if (do_swap_account)
2405 page_counter_uncharge(&memcg->memsw, batch);
2406 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2408 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2412 if (batch > nr_pages) {
2418 * Unlike in global OOM situations, memcg is not in a physical
2419 * memory shortage. Allow dying and OOM-killed tasks to
2420 * bypass the last charges so that they can exit quickly and
2421 * free their memory.
2423 if (unlikely(test_thread_flag(TIF_MEMDIE) ||
2424 fatal_signal_pending(current) ||
2425 current->flags & PF_EXITING))
2428 if (unlikely(task_in_memcg_oom(current)))
2431 if (!(gfp_mask & __GFP_WAIT))
2434 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2435 gfp_mask, may_swap);
2437 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2441 drain_all_stock(mem_over_limit);
2446 if (gfp_mask & __GFP_NORETRY)
2449 * Even though the limit is exceeded at this point, reclaim
2450 * may have been able to free some pages. Retry the charge
2451 * before killing the task.
2453 * Only for regular pages, though: huge pages are rather
2454 * unlikely to succeed so close to the limit, and we fall back
2455 * to regular pages anyway in case of failure.
2457 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2460 * At task move, charge accounts can be doubly counted. So, it's
2461 * better to wait until the end of task_move if something is going on.
2463 if (mem_cgroup_wait_acct_move(mem_over_limit))
2469 if (gfp_mask & __GFP_NOFAIL)
2472 if (fatal_signal_pending(current))
2475 mem_cgroup_oom(mem_over_limit, gfp_mask, get_order(nr_pages));
2477 if (!(gfp_mask & __GFP_NOFAIL))
2483 css_get_many(&memcg->css, batch);
2484 if (batch > nr_pages)
2485 refill_stock(memcg, batch - nr_pages);
2490 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2492 if (mem_cgroup_is_root(memcg))
2495 page_counter_uncharge(&memcg->memory, nr_pages);
2496 if (do_swap_account)
2497 page_counter_uncharge(&memcg->memsw, nr_pages);
2499 css_put_many(&memcg->css, nr_pages);
2503 * A helper function to get mem_cgroup from ID. must be called under
2504 * rcu_read_lock(). The caller is responsible for calling
2505 * css_tryget_online() if the mem_cgroup is used for charging. (dropping
2506 * refcnt from swap can be called against removed memcg.)
2508 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2510 /* ID 0 is unused ID */
2513 return mem_cgroup_from_id(id);
2517 * try_get_mem_cgroup_from_page - look up page's memcg association
2520 * Look up, get a css reference, and return the memcg that owns @page.
2522 * The page must be locked to prevent racing with swap-in and page
2523 * cache charges. If coming from an unlocked page table, the caller
2524 * must ensure the page is on the LRU or this can race with charging.
2526 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2528 struct mem_cgroup *memcg = NULL;
2529 struct page_cgroup *pc;
2533 VM_BUG_ON_PAGE(!PageLocked(page), page);
2535 pc = lookup_page_cgroup(page);
2536 if (PageCgroupUsed(pc)) {
2537 memcg = pc->mem_cgroup;
2538 if (memcg && !css_tryget_online(&memcg->css))
2540 } else if (PageSwapCache(page)) {
2541 ent.val = page_private(page);
2542 id = lookup_swap_cgroup_id(ent);
2544 memcg = mem_cgroup_lookup(id);
2545 if (memcg && !css_tryget_online(&memcg->css))
2552 static void lock_page_lru(struct page *page, int *isolated)
2554 struct zone *zone = page_zone(page);
2556 spin_lock_irq(&zone->lru_lock);
2557 if (PageLRU(page)) {
2558 struct lruvec *lruvec;
2560 lruvec = mem_cgroup_page_lruvec(page, zone);
2562 del_page_from_lru_list(page, lruvec, page_lru(page));
2568 static void unlock_page_lru(struct page *page, int isolated)
2570 struct zone *zone = page_zone(page);
2573 struct lruvec *lruvec;
2575 lruvec = mem_cgroup_page_lruvec(page, zone);
2576 VM_BUG_ON_PAGE(PageLRU(page), page);
2578 add_page_to_lru_list(page, lruvec, page_lru(page));
2580 spin_unlock_irq(&zone->lru_lock);
2583 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2586 struct page_cgroup *pc = lookup_page_cgroup(page);
2589 VM_BUG_ON_PAGE(PageCgroupUsed(pc), page);
2591 * we don't need page_cgroup_lock about tail pages, becase they are not
2592 * accessed by any other context at this point.
2596 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2597 * may already be on some other mem_cgroup's LRU. Take care of it.
2600 lock_page_lru(page, &isolated);
2603 * Nobody should be changing or seriously looking at
2604 * pc->mem_cgroup and pc->flags at this point:
2606 * - the page is uncharged
2608 * - the page is off-LRU
2610 * - an anonymous fault has exclusive page access, except for
2611 * a locked page table
2613 * - a page cache insertion, a swapin fault, or a migration
2614 * have the page locked
2616 pc->mem_cgroup = memcg;
2617 pc->flags = PCG_USED | PCG_MEM | (do_swap_account ? PCG_MEMSW : 0);
2620 unlock_page_lru(page, isolated);
2623 #ifdef CONFIG_MEMCG_KMEM
2625 * The memcg_slab_mutex is held whenever a per memcg kmem cache is created or
2626 * destroyed. It protects memcg_caches arrays and memcg_slab_caches lists.
2628 static DEFINE_MUTEX(memcg_slab_mutex);
2630 static DEFINE_MUTEX(activate_kmem_mutex);
2633 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2634 * in the memcg_cache_params struct.
2636 static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
2638 struct kmem_cache *cachep;
2640 VM_BUG_ON(p->is_root_cache);
2641 cachep = p->root_cache;
2642 return cache_from_memcg_idx(cachep, memcg_cache_id(p->memcg));
2645 #ifdef CONFIG_SLABINFO
2646 static int mem_cgroup_slabinfo_read(struct seq_file *m, void *v)
2648 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
2649 struct memcg_cache_params *params;
2651 if (!memcg_kmem_is_active(memcg))
2654 print_slabinfo_header(m);
2656 mutex_lock(&memcg_slab_mutex);
2657 list_for_each_entry(params, &memcg->memcg_slab_caches, list)
2658 cache_show(memcg_params_to_cache(params), m);
2659 mutex_unlock(&memcg_slab_mutex);
2665 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp,
2666 unsigned long nr_pages)
2668 struct page_counter *counter;
2671 ret = page_counter_try_charge(&memcg->kmem, nr_pages, &counter);
2675 ret = try_charge(memcg, gfp, nr_pages);
2676 if (ret == -EINTR) {
2678 * try_charge() chose to bypass to root due to OOM kill or
2679 * fatal signal. Since our only options are to either fail
2680 * the allocation or charge it to this cgroup, do it as a
2681 * temporary condition. But we can't fail. From a kmem/slab
2682 * perspective, the cache has already been selected, by
2683 * mem_cgroup_kmem_get_cache(), so it is too late to change
2686 * This condition will only trigger if the task entered
2687 * memcg_charge_kmem in a sane state, but was OOM-killed
2688 * during try_charge() above. Tasks that were already dying
2689 * when the allocation triggers should have been already
2690 * directed to the root cgroup in memcontrol.h
2692 page_counter_charge(&memcg->memory, nr_pages);
2693 if (do_swap_account)
2694 page_counter_charge(&memcg->memsw, nr_pages);
2695 css_get_many(&memcg->css, nr_pages);
2698 page_counter_uncharge(&memcg->kmem, nr_pages);
2703 static void memcg_uncharge_kmem(struct mem_cgroup *memcg,
2704 unsigned long nr_pages)
2706 page_counter_uncharge(&memcg->memory, nr_pages);
2707 if (do_swap_account)
2708 page_counter_uncharge(&memcg->memsw, nr_pages);
2710 page_counter_uncharge(&memcg->kmem, nr_pages);
2712 css_put_many(&memcg->css, nr_pages);
2716 * helper for acessing a memcg's index. It will be used as an index in the
2717 * child cache array in kmem_cache, and also to derive its name. This function
2718 * will return -1 when this is not a kmem-limited memcg.
2720 int memcg_cache_id(struct mem_cgroup *memcg)
2722 return memcg ? memcg->kmemcg_id : -1;
2725 static int memcg_alloc_cache_id(void)
2730 id = ida_simple_get(&kmem_limited_groups,
2731 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2735 if (id < memcg_limited_groups_array_size)
2739 * There's no space for the new id in memcg_caches arrays,
2740 * so we have to grow them.
2743 size = 2 * (id + 1);
2744 if (size < MEMCG_CACHES_MIN_SIZE)
2745 size = MEMCG_CACHES_MIN_SIZE;
2746 else if (size > MEMCG_CACHES_MAX_SIZE)
2747 size = MEMCG_CACHES_MAX_SIZE;
2749 mutex_lock(&memcg_slab_mutex);
2750 err = memcg_update_all_caches(size);
2751 mutex_unlock(&memcg_slab_mutex);
2754 ida_simple_remove(&kmem_limited_groups, id);
2760 static void memcg_free_cache_id(int id)
2762 ida_simple_remove(&kmem_limited_groups, id);
2766 * We should update the current array size iff all caches updates succeed. This
2767 * can only be done from the slab side. The slab mutex needs to be held when
2770 void memcg_update_array_size(int num)
2772 memcg_limited_groups_array_size = num;
2775 static void memcg_register_cache(struct mem_cgroup *memcg,
2776 struct kmem_cache *root_cache)
2778 static char memcg_name_buf[NAME_MAX + 1]; /* protected by
2780 struct kmem_cache *cachep;
2783 lockdep_assert_held(&memcg_slab_mutex);
2785 id = memcg_cache_id(memcg);
2788 * Since per-memcg caches are created asynchronously on first
2789 * allocation (see memcg_kmem_get_cache()), several threads can try to
2790 * create the same cache, but only one of them may succeed.
2792 if (cache_from_memcg_idx(root_cache, id))
2795 cgroup_name(memcg->css.cgroup, memcg_name_buf, NAME_MAX + 1);
2796 cachep = memcg_create_kmem_cache(memcg, root_cache, memcg_name_buf);
2798 * If we could not create a memcg cache, do not complain, because
2799 * that's not critical at all as we can always proceed with the root
2805 css_get(&memcg->css);
2806 list_add(&cachep->memcg_params->list, &memcg->memcg_slab_caches);
2809 * Since readers won't lock (see cache_from_memcg_idx()), we need a
2810 * barrier here to ensure nobody will see the kmem_cache partially
2815 BUG_ON(root_cache->memcg_params->memcg_caches[id]);
2816 root_cache->memcg_params->memcg_caches[id] = cachep;
2819 static void memcg_unregister_cache(struct kmem_cache *cachep)
2821 struct kmem_cache *root_cache;
2822 struct mem_cgroup *memcg;
2825 lockdep_assert_held(&memcg_slab_mutex);
2827 BUG_ON(is_root_cache(cachep));
2829 root_cache = cachep->memcg_params->root_cache;
2830 memcg = cachep->memcg_params->memcg;
2831 id = memcg_cache_id(memcg);
2833 BUG_ON(root_cache->memcg_params->memcg_caches[id] != cachep);
2834 root_cache->memcg_params->memcg_caches[id] = NULL;
2836 list_del(&cachep->memcg_params->list);
2838 kmem_cache_destroy(cachep);
2840 /* drop the reference taken in memcg_register_cache */
2841 css_put(&memcg->css);
2845 * During the creation a new cache, we need to disable our accounting mechanism
2846 * altogether. This is true even if we are not creating, but rather just
2847 * enqueing new caches to be created.
2849 * This is because that process will trigger allocations; some visible, like
2850 * explicit kmallocs to auxiliary data structures, name strings and internal
2851 * cache structures; some well concealed, like INIT_WORK() that can allocate
2852 * objects during debug.
2854 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
2855 * to it. This may not be a bounded recursion: since the first cache creation
2856 * failed to complete (waiting on the allocation), we'll just try to create the
2857 * cache again, failing at the same point.
2859 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
2860 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
2861 * inside the following two functions.
2863 static inline void memcg_stop_kmem_account(void)
2865 VM_BUG_ON(!current->mm);
2866 current->memcg_kmem_skip_account++;
2869 static inline void memcg_resume_kmem_account(void)
2871 VM_BUG_ON(!current->mm);
2872 current->memcg_kmem_skip_account--;
2875 int __memcg_cleanup_cache_params(struct kmem_cache *s)
2877 struct kmem_cache *c;
2880 mutex_lock(&memcg_slab_mutex);
2881 for_each_memcg_cache_index(i) {
2882 c = cache_from_memcg_idx(s, i);
2886 memcg_unregister_cache(c);
2888 if (cache_from_memcg_idx(s, i))
2891 mutex_unlock(&memcg_slab_mutex);
2895 static void memcg_unregister_all_caches(struct mem_cgroup *memcg)
2897 struct kmem_cache *cachep;
2898 struct memcg_cache_params *params, *tmp;
2900 if (!memcg_kmem_is_active(memcg))
2903 mutex_lock(&memcg_slab_mutex);
2904 list_for_each_entry_safe(params, tmp, &memcg->memcg_slab_caches, list) {
2905 cachep = memcg_params_to_cache(params);
2906 kmem_cache_shrink(cachep);
2907 if (atomic_read(&cachep->memcg_params->nr_pages) == 0)
2908 memcg_unregister_cache(cachep);
2910 mutex_unlock(&memcg_slab_mutex);
2913 struct memcg_register_cache_work {
2914 struct mem_cgroup *memcg;
2915 struct kmem_cache *cachep;
2916 struct work_struct work;
2919 static void memcg_register_cache_func(struct work_struct *w)
2921 struct memcg_register_cache_work *cw =
2922 container_of(w, struct memcg_register_cache_work, work);
2923 struct mem_cgroup *memcg = cw->memcg;
2924 struct kmem_cache *cachep = cw->cachep;
2926 mutex_lock(&memcg_slab_mutex);
2927 memcg_register_cache(memcg, cachep);
2928 mutex_unlock(&memcg_slab_mutex);
2930 css_put(&memcg->css);
2935 * Enqueue the creation of a per-memcg kmem_cache.
2937 static void __memcg_schedule_register_cache(struct mem_cgroup *memcg,
2938 struct kmem_cache *cachep)
2940 struct memcg_register_cache_work *cw;
2942 cw = kmalloc(sizeof(*cw), GFP_NOWAIT);
2944 css_put(&memcg->css);
2949 cw->cachep = cachep;
2951 INIT_WORK(&cw->work, memcg_register_cache_func);
2952 schedule_work(&cw->work);
2955 static void memcg_schedule_register_cache(struct mem_cgroup *memcg,
2956 struct kmem_cache *cachep)
2959 * We need to stop accounting when we kmalloc, because if the
2960 * corresponding kmalloc cache is not yet created, the first allocation
2961 * in __memcg_schedule_register_cache will recurse.
2963 * However, it is better to enclose the whole function. Depending on
2964 * the debugging options enabled, INIT_WORK(), for instance, can
2965 * trigger an allocation. This too, will make us recurse. Because at
2966 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2967 * the safest choice is to do it like this, wrapping the whole function.
2969 memcg_stop_kmem_account();
2970 __memcg_schedule_register_cache(memcg, cachep);
2971 memcg_resume_kmem_account();
2974 int __memcg_charge_slab(struct kmem_cache *cachep, gfp_t gfp, int order)
2976 unsigned int nr_pages = 1 << order;
2979 res = memcg_charge_kmem(cachep->memcg_params->memcg, gfp, nr_pages);
2981 atomic_add(nr_pages, &cachep->memcg_params->nr_pages);
2985 void __memcg_uncharge_slab(struct kmem_cache *cachep, int order)
2987 unsigned int nr_pages = 1 << order;
2989 memcg_uncharge_kmem(cachep->memcg_params->memcg, nr_pages);
2990 atomic_sub(nr_pages, &cachep->memcg_params->nr_pages);
2994 * Return the kmem_cache we're supposed to use for a slab allocation.
2995 * We try to use the current memcg's version of the cache.
2997 * If the cache does not exist yet, if we are the first user of it,
2998 * we either create it immediately, if possible, or create it asynchronously
3000 * In the latter case, we will let the current allocation go through with
3001 * the original cache.
3003 * Can't be called in interrupt context or from kernel threads.
3004 * This function needs to be called with rcu_read_lock() held.
3006 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
3009 struct mem_cgroup *memcg;
3010 struct kmem_cache *memcg_cachep;
3012 VM_BUG_ON(!cachep->memcg_params);
3013 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
3015 if (!current->mm || current->memcg_kmem_skip_account)
3019 memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
3021 if (!memcg_kmem_is_active(memcg))
3024 memcg_cachep = cache_from_memcg_idx(cachep, memcg_cache_id(memcg));
3025 if (likely(memcg_cachep)) {
3026 cachep = memcg_cachep;
3030 /* The corresponding put will be done in the workqueue. */
3031 if (!css_tryget_online(&memcg->css))
3036 * If we are in a safe context (can wait, and not in interrupt
3037 * context), we could be be predictable and return right away.
3038 * This would guarantee that the allocation being performed
3039 * already belongs in the new cache.
3041 * However, there are some clashes that can arrive from locking.
3042 * For instance, because we acquire the slab_mutex while doing
3043 * memcg_create_kmem_cache, this means no further allocation
3044 * could happen with the slab_mutex held. So it's better to
3047 memcg_schedule_register_cache(memcg, cachep);
3055 * We need to verify if the allocation against current->mm->owner's memcg is
3056 * possible for the given order. But the page is not allocated yet, so we'll
3057 * need a further commit step to do the final arrangements.
3059 * It is possible for the task to switch cgroups in this mean time, so at
3060 * commit time, we can't rely on task conversion any longer. We'll then use
3061 * the handle argument to return to the caller which cgroup we should commit
3062 * against. We could also return the memcg directly and avoid the pointer
3063 * passing, but a boolean return value gives better semantics considering
3064 * the compiled-out case as well.
3066 * Returning true means the allocation is possible.
3069 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3071 struct mem_cgroup *memcg;
3077 * Disabling accounting is only relevant for some specific memcg
3078 * internal allocations. Therefore we would initially not have such
3079 * check here, since direct calls to the page allocator that are
3080 * accounted to kmemcg (alloc_kmem_pages and friends) only happen
3081 * outside memcg core. We are mostly concerned with cache allocations,
3082 * and by having this test at memcg_kmem_get_cache, we are already able
3083 * to relay the allocation to the root cache and bypass the memcg cache
3086 * There is one exception, though: the SLUB allocator does not create
3087 * large order caches, but rather service large kmallocs directly from
3088 * the page allocator. Therefore, the following sequence when backed by
3089 * the SLUB allocator:
3091 * memcg_stop_kmem_account();
3092 * kmalloc(<large_number>)
3093 * memcg_resume_kmem_account();
3095 * would effectively ignore the fact that we should skip accounting,
3096 * since it will drive us directly to this function without passing
3097 * through the cache selector memcg_kmem_get_cache. Such large
3098 * allocations are extremely rare but can happen, for instance, for the
3099 * cache arrays. We bring this test here.
3101 if (!current->mm || current->memcg_kmem_skip_account)
3104 memcg = get_mem_cgroup_from_mm(current->mm);
3106 if (!memcg_kmem_is_active(memcg)) {
3107 css_put(&memcg->css);
3111 ret = memcg_charge_kmem(memcg, gfp, 1 << order);
3115 css_put(&memcg->css);
3119 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3122 struct page_cgroup *pc;
3124 VM_BUG_ON(mem_cgroup_is_root(memcg));
3126 /* The page allocation failed. Revert */
3128 memcg_uncharge_kmem(memcg, 1 << order);
3132 * The page is freshly allocated and not visible to any
3133 * outside callers yet. Set up pc non-atomically.
3135 pc = lookup_page_cgroup(page);
3136 pc->mem_cgroup = memcg;
3137 pc->flags = PCG_USED;
3140 void __memcg_kmem_uncharge_pages(struct page *page, int order)
3142 struct mem_cgroup *memcg = NULL;
3143 struct page_cgroup *pc;
3146 pc = lookup_page_cgroup(page);
3147 if (!PageCgroupUsed(pc))
3150 memcg = pc->mem_cgroup;
3154 * We trust that only if there is a memcg associated with the page, it
3155 * is a valid allocation
3160 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
3161 memcg_uncharge_kmem(memcg, 1 << order);
3164 static inline void memcg_unregister_all_caches(struct mem_cgroup *memcg)
3167 #endif /* CONFIG_MEMCG_KMEM */
3169 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3172 * Because tail pages are not marked as "used", set it. We're under
3173 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3174 * charge/uncharge will be never happen and move_account() is done under
3175 * compound_lock(), so we don't have to take care of races.
3177 void mem_cgroup_split_huge_fixup(struct page *head)
3179 struct page_cgroup *head_pc = lookup_page_cgroup(head);
3180 struct page_cgroup *pc;
3181 struct mem_cgroup *memcg;
3184 if (mem_cgroup_disabled())
3187 memcg = head_pc->mem_cgroup;
3188 for (i = 1; i < HPAGE_PMD_NR; i++) {
3190 pc->mem_cgroup = memcg;
3191 pc->flags = head_pc->flags;
3193 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
3196 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3199 * mem_cgroup_move_account - move account of the page
3201 * @nr_pages: number of regular pages (>1 for huge pages)
3202 * @pc: page_cgroup of the page.
3203 * @from: mem_cgroup which the page is moved from.
3204 * @to: mem_cgroup which the page is moved to. @from != @to.
3206 * The caller must confirm following.
3207 * - page is not on LRU (isolate_page() is useful.)
3208 * - compound_lock is held when nr_pages > 1
3210 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3213 static int mem_cgroup_move_account(struct page *page,
3214 unsigned int nr_pages,
3215 struct page_cgroup *pc,
3216 struct mem_cgroup *from,
3217 struct mem_cgroup *to)
3219 unsigned long flags;
3222 VM_BUG_ON(from == to);
3223 VM_BUG_ON_PAGE(PageLRU(page), page);
3225 * The page is isolated from LRU. So, collapse function
3226 * will not handle this page. But page splitting can happen.
3227 * Do this check under compound_page_lock(). The caller should
3231 if (nr_pages > 1 && !PageTransHuge(page))
3235 * Prevent mem_cgroup_migrate() from looking at pc->mem_cgroup
3236 * of its source page while we change it: page migration takes
3237 * both pages off the LRU, but page cache replacement doesn't.
3239 if (!trylock_page(page))
3243 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3246 move_lock_mem_cgroup(from, &flags);
3248 if (!PageAnon(page) && page_mapped(page)) {
3249 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
3251 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
3255 if (PageWriteback(page)) {
3256 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK],
3258 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK],
3263 * It is safe to change pc->mem_cgroup here because the page
3264 * is referenced, charged, and isolated - we can't race with
3265 * uncharging, charging, migration, or LRU putback.
3268 /* caller should have done css_get */
3269 pc->mem_cgroup = to;
3270 move_unlock_mem_cgroup(from, &flags);
3273 local_irq_disable();
3274 mem_cgroup_charge_statistics(to, page, nr_pages);
3275 memcg_check_events(to, page);
3276 mem_cgroup_charge_statistics(from, page, -nr_pages);
3277 memcg_check_events(from, page);
3285 #ifdef CONFIG_MEMCG_SWAP
3286 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
3289 int val = (charge) ? 1 : -1;
3290 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
3294 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3295 * @entry: swap entry to be moved
3296 * @from: mem_cgroup which the entry is moved from
3297 * @to: mem_cgroup which the entry is moved to
3299 * It succeeds only when the swap_cgroup's record for this entry is the same
3300 * as the mem_cgroup's id of @from.
3302 * Returns 0 on success, -EINVAL on failure.
3304 * The caller must have charged to @to, IOW, called page_counter_charge() about
3305 * both res and memsw, and called css_get().
3307 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3308 struct mem_cgroup *from, struct mem_cgroup *to)
3310 unsigned short old_id, new_id;
3312 old_id = mem_cgroup_id(from);
3313 new_id = mem_cgroup_id(to);
3315 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3316 mem_cgroup_swap_statistics(from, false);
3317 mem_cgroup_swap_statistics(to, true);
3319 * This function is only called from task migration context now.
3320 * It postpones page_counter and refcount handling till the end
3321 * of task migration(mem_cgroup_clear_mc()) for performance
3322 * improvement. But we cannot postpone css_get(to) because if
3323 * the process that has been moved to @to does swap-in, the
3324 * refcount of @to might be decreased to 0.
3326 * We are in attach() phase, so the cgroup is guaranteed to be
3327 * alive, so we can just call css_get().
3335 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3336 struct mem_cgroup *from, struct mem_cgroup *to)
3342 #ifdef CONFIG_DEBUG_VM
3343 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3345 struct page_cgroup *pc;
3347 pc = lookup_page_cgroup(page);
3349 * Can be NULL while feeding pages into the page allocator for
3350 * the first time, i.e. during boot or memory hotplug;
3351 * or when mem_cgroup_disabled().
3353 if (likely(pc) && PageCgroupUsed(pc))
3358 bool mem_cgroup_bad_page_check(struct page *page)
3360 if (mem_cgroup_disabled())
3363 return lookup_page_cgroup_used(page) != NULL;
3366 void mem_cgroup_print_bad_page(struct page *page)
3368 struct page_cgroup *pc;
3370 pc = lookup_page_cgroup_used(page);
3372 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
3373 pc, pc->flags, pc->mem_cgroup);
3378 static DEFINE_MUTEX(memcg_limit_mutex);
3380 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3381 unsigned long limit)
3383 unsigned long curusage;
3384 unsigned long oldusage;
3385 bool enlarge = false;
3390 * For keeping hierarchical_reclaim simple, how long we should retry
3391 * is depends on callers. We set our retry-count to be function
3392 * of # of children which we should visit in this loop.
3394 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
3395 mem_cgroup_count_children(memcg);
3397 oldusage = page_counter_read(&memcg->memory);
3400 if (signal_pending(current)) {
3405 mutex_lock(&memcg_limit_mutex);
3406 if (limit > memcg->memsw.limit) {
3407 mutex_unlock(&memcg_limit_mutex);
3411 if (limit > memcg->memory.limit)
3413 ret = page_counter_limit(&memcg->memory, limit);
3414 mutex_unlock(&memcg_limit_mutex);
3419 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, true);
3421 curusage = page_counter_read(&memcg->memory);
3422 /* Usage is reduced ? */
3423 if (curusage >= oldusage)
3426 oldusage = curusage;
3427 } while (retry_count);
3429 if (!ret && enlarge)
3430 memcg_oom_recover(memcg);
3435 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3436 unsigned long limit)
3438 unsigned long curusage;
3439 unsigned long oldusage;
3440 bool enlarge = false;
3444 /* see mem_cgroup_resize_res_limit */
3445 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
3446 mem_cgroup_count_children(memcg);
3448 oldusage = page_counter_read(&memcg->memsw);
3451 if (signal_pending(current)) {
3456 mutex_lock(&memcg_limit_mutex);
3457 if (limit < memcg->memory.limit) {
3458 mutex_unlock(&memcg_limit_mutex);
3462 if (limit > memcg->memsw.limit)
3464 ret = page_counter_limit(&memcg->memsw, limit);
3465 mutex_unlock(&memcg_limit_mutex);
3470 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, false);
3472 curusage = page_counter_read(&memcg->memsw);
3473 /* Usage is reduced ? */
3474 if (curusage >= oldusage)
3477 oldusage = curusage;
3478 } while (retry_count);
3480 if (!ret && enlarge)
3481 memcg_oom_recover(memcg);
3486 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3488 unsigned long *total_scanned)
3490 unsigned long nr_reclaimed = 0;
3491 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3492 unsigned long reclaimed;
3494 struct mem_cgroup_tree_per_zone *mctz;
3495 unsigned long excess;
3496 unsigned long nr_scanned;
3501 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3503 * This loop can run a while, specially if mem_cgroup's continuously
3504 * keep exceeding their soft limit and putting the system under
3511 mz = mem_cgroup_largest_soft_limit_node(mctz);
3516 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
3517 gfp_mask, &nr_scanned);
3518 nr_reclaimed += reclaimed;
3519 *total_scanned += nr_scanned;
3520 spin_lock_irq(&mctz->lock);
3523 * If we failed to reclaim anything from this memory cgroup
3524 * it is time to move on to the next cgroup
3530 * Loop until we find yet another one.
3532 * By the time we get the soft_limit lock
3533 * again, someone might have aded the
3534 * group back on the RB tree. Iterate to
3535 * make sure we get a different mem.
3536 * mem_cgroup_largest_soft_limit_node returns
3537 * NULL if no other cgroup is present on
3541 __mem_cgroup_largest_soft_limit_node(mctz);
3543 css_put(&next_mz->memcg->css);
3544 else /* next_mz == NULL or other memcg */
3548 __mem_cgroup_remove_exceeded(mz, mctz);
3549 excess = soft_limit_excess(mz->memcg);
3551 * One school of thought says that we should not add
3552 * back the node to the tree if reclaim returns 0.
3553 * But our reclaim could return 0, simply because due
3554 * to priority we are exposing a smaller subset of
3555 * memory to reclaim from. Consider this as a longer
3558 /* If excess == 0, no tree ops */
3559 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3560 spin_unlock_irq(&mctz->lock);
3561 css_put(&mz->memcg->css);
3564 * Could not reclaim anything and there are no more
3565 * mem cgroups to try or we seem to be looping without
3566 * reclaiming anything.
3568 if (!nr_reclaimed &&
3570 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3572 } while (!nr_reclaimed);
3574 css_put(&next_mz->memcg->css);
3575 return nr_reclaimed;
3579 * Test whether @memcg has children, dead or alive. Note that this
3580 * function doesn't care whether @memcg has use_hierarchy enabled and
3581 * returns %true if there are child csses according to the cgroup
3582 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
3584 static inline bool memcg_has_children(struct mem_cgroup *memcg)
3589 * The lock does not prevent addition or deletion of children, but
3590 * it prevents a new child from being initialized based on this
3591 * parent in css_online(), so it's enough to decide whether
3592 * hierarchically inherited attributes can still be changed or not.
3594 lockdep_assert_held(&memcg_create_mutex);
3597 ret = css_next_child(NULL, &memcg->css);
3603 * Reclaims as many pages from the given memcg as possible and moves
3604 * the rest to the parent.
3606 * Caller is responsible for holding css reference for memcg.
3608 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3610 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3612 /* we call try-to-free pages for make this cgroup empty */
3613 lru_add_drain_all();
3614 /* try to free all pages in this cgroup */
3615 while (nr_retries && page_counter_read(&memcg->memory)) {
3618 if (signal_pending(current))
3621 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3625 /* maybe some writeback is necessary */
3626 congestion_wait(BLK_RW_ASYNC, HZ/10);
3634 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3635 char *buf, size_t nbytes,
3638 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3640 if (mem_cgroup_is_root(memcg))
3642 return mem_cgroup_force_empty(memcg) ?: nbytes;
3645 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3648 return mem_cgroup_from_css(css)->use_hierarchy;
3651 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3652 struct cftype *cft, u64 val)
3655 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3656 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
3658 mutex_lock(&memcg_create_mutex);
3660 if (memcg->use_hierarchy == val)
3664 * If parent's use_hierarchy is set, we can't make any modifications
3665 * in the child subtrees. If it is unset, then the change can
3666 * occur, provided the current cgroup has no children.
3668 * For the root cgroup, parent_mem is NULL, we allow value to be
3669 * set if there are no children.
3671 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3672 (val == 1 || val == 0)) {
3673 if (!memcg_has_children(memcg))
3674 memcg->use_hierarchy = val;
3681 mutex_unlock(&memcg_create_mutex);
3686 static unsigned long tree_stat(struct mem_cgroup *memcg,
3687 enum mem_cgroup_stat_index idx)
3689 struct mem_cgroup *iter;
3692 /* Per-cpu values can be negative, use a signed accumulator */
3693 for_each_mem_cgroup_tree(iter, memcg)
3694 val += mem_cgroup_read_stat(iter, idx);
3696 if (val < 0) /* race ? */
3701 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3705 if (mem_cgroup_is_root(memcg)) {
3706 val = tree_stat(memcg, MEM_CGROUP_STAT_CACHE);
3707 val += tree_stat(memcg, MEM_CGROUP_STAT_RSS);
3709 val += tree_stat(memcg, MEM_CGROUP_STAT_SWAP);
3712 val = page_counter_read(&memcg->memory);
3714 val = page_counter_read(&memcg->memsw);
3716 return val << PAGE_SHIFT;
3727 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3730 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3731 struct page_counter *counter;
3733 switch (MEMFILE_TYPE(cft->private)) {
3735 counter = &memcg->memory;
3738 counter = &memcg->memsw;
3741 counter = &memcg->kmem;
3747 switch (MEMFILE_ATTR(cft->private)) {
3749 if (counter == &memcg->memory)
3750 return mem_cgroup_usage(memcg, false);
3751 if (counter == &memcg->memsw)
3752 return mem_cgroup_usage(memcg, true);
3753 return (u64)page_counter_read(counter) * PAGE_SIZE;
3755 return (u64)counter->limit * PAGE_SIZE;
3757 return (u64)counter->watermark * PAGE_SIZE;
3759 return counter->failcnt;
3760 case RES_SOFT_LIMIT:
3761 return (u64)memcg->soft_limit * PAGE_SIZE;
3767 #ifdef CONFIG_MEMCG_KMEM
3768 /* should be called with activate_kmem_mutex held */
3769 static int __memcg_activate_kmem(struct mem_cgroup *memcg,
3770 unsigned long nr_pages)
3775 if (memcg_kmem_is_active(memcg))
3779 * We are going to allocate memory for data shared by all memory
3780 * cgroups so let's stop accounting here.
3782 memcg_stop_kmem_account();
3785 * For simplicity, we won't allow this to be disabled. It also can't
3786 * be changed if the cgroup has children already, or if tasks had
3789 * If tasks join before we set the limit, a person looking at
3790 * kmem.usage_in_bytes will have no way to determine when it took
3791 * place, which makes the value quite meaningless.
3793 * After it first became limited, changes in the value of the limit are
3794 * of course permitted.
3796 mutex_lock(&memcg_create_mutex);
3797 if (cgroup_has_tasks(memcg->css.cgroup) ||
3798 (memcg->use_hierarchy && memcg_has_children(memcg)))
3800 mutex_unlock(&memcg_create_mutex);
3804 memcg_id = memcg_alloc_cache_id();
3810 memcg->kmemcg_id = memcg_id;
3811 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
3814 * We couldn't have accounted to this cgroup, because it hasn't got the
3815 * active bit set yet, so this should succeed.
3817 err = page_counter_limit(&memcg->kmem, nr_pages);
3820 static_key_slow_inc(&memcg_kmem_enabled_key);
3822 * Setting the active bit after enabling static branching will
3823 * guarantee no one starts accounting before all call sites are
3826 memcg_kmem_set_active(memcg);
3828 memcg_resume_kmem_account();
3832 static int memcg_activate_kmem(struct mem_cgroup *memcg,
3833 unsigned long nr_pages)
3837 mutex_lock(&activate_kmem_mutex);
3838 ret = __memcg_activate_kmem(memcg, nr_pages);
3839 mutex_unlock(&activate_kmem_mutex);
3843 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
3844 unsigned long limit)
3848 mutex_lock(&memcg_limit_mutex);
3849 if (!memcg_kmem_is_active(memcg))
3850 ret = memcg_activate_kmem(memcg, limit);
3852 ret = page_counter_limit(&memcg->kmem, limit);
3853 mutex_unlock(&memcg_limit_mutex);
3857 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
3860 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
3865 mutex_lock(&activate_kmem_mutex);
3867 * If the parent cgroup is not kmem-active now, it cannot be activated
3868 * after this point, because it has at least one child already.
3870 if (memcg_kmem_is_active(parent))
3871 ret = __memcg_activate_kmem(memcg, PAGE_COUNTER_MAX);
3872 mutex_unlock(&activate_kmem_mutex);
3876 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
3877 unsigned long limit)
3881 #endif /* CONFIG_MEMCG_KMEM */
3884 * The user of this function is...
3887 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3888 char *buf, size_t nbytes, loff_t off)
3890 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3891 unsigned long nr_pages;
3894 buf = strstrip(buf);
3895 ret = page_counter_memparse(buf, &nr_pages);
3899 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3901 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3905 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3907 ret = mem_cgroup_resize_limit(memcg, nr_pages);
3910 ret = mem_cgroup_resize_memsw_limit(memcg, nr_pages);
3913 ret = memcg_update_kmem_limit(memcg, nr_pages);
3917 case RES_SOFT_LIMIT:
3918 memcg->soft_limit = nr_pages;
3922 return ret ?: nbytes;
3925 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3926 size_t nbytes, loff_t off)
3928 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3929 struct page_counter *counter;
3931 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3933 counter = &memcg->memory;
3936 counter = &memcg->memsw;
3939 counter = &memcg->kmem;
3945 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3947 page_counter_reset_watermark(counter);
3950 counter->failcnt = 0;
3959 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3962 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3966 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3967 struct cftype *cft, u64 val)
3969 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3971 if (val >= (1 << NR_MOVE_TYPE))
3975 * No kind of locking is needed in here, because ->can_attach() will
3976 * check this value once in the beginning of the process, and then carry
3977 * on with stale data. This means that changes to this value will only
3978 * affect task migrations starting after the change.
3980 memcg->move_charge_at_immigrate = val;
3984 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3985 struct cftype *cft, u64 val)
3992 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3996 unsigned int lru_mask;
3999 static const struct numa_stat stats[] = {
4000 { "total", LRU_ALL },
4001 { "file", LRU_ALL_FILE },
4002 { "anon", LRU_ALL_ANON },
4003 { "unevictable", BIT(LRU_UNEVICTABLE) },
4005 const struct numa_stat *stat;
4008 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
4010 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
4011 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
4012 seq_printf(m, "%s=%lu", stat->name, nr);
4013 for_each_node_state(nid, N_MEMORY) {
4014 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4016 seq_printf(m, " N%d=%lu", nid, nr);
4021 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
4022 struct mem_cgroup *iter;
4025 for_each_mem_cgroup_tree(iter, memcg)
4026 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
4027 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
4028 for_each_node_state(nid, N_MEMORY) {
4030 for_each_mem_cgroup_tree(iter, memcg)
4031 nr += mem_cgroup_node_nr_lru_pages(
4032 iter, nid, stat->lru_mask);
4033 seq_printf(m, " N%d=%lu", nid, nr);
4040 #endif /* CONFIG_NUMA */
4042 static inline void mem_cgroup_lru_names_not_uptodate(void)
4044 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
4047 static int memcg_stat_show(struct seq_file *m, void *v)
4049 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
4050 unsigned long memory, memsw;
4051 struct mem_cgroup *mi;
4054 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
4055 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
4057 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
4058 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
4061 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
4062 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
4063 mem_cgroup_read_events(memcg, i));
4065 for (i = 0; i < NR_LRU_LISTS; i++)
4066 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
4067 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
4069 /* Hierarchical information */
4070 memory = memsw = PAGE_COUNTER_MAX;
4071 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
4072 memory = min(memory, mi->memory.limit);
4073 memsw = min(memsw, mi->memsw.limit);
4075 seq_printf(m, "hierarchical_memory_limit %llu\n",
4076 (u64)memory * PAGE_SIZE);
4077 if (do_swap_account)
4078 seq_printf(m, "hierarchical_memsw_limit %llu\n",
4079 (u64)memsw * PAGE_SIZE);
4081 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
4084 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
4086 for_each_mem_cgroup_tree(mi, memcg)
4087 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
4088 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
4091 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
4092 unsigned long long val = 0;
4094 for_each_mem_cgroup_tree(mi, memcg)
4095 val += mem_cgroup_read_events(mi, i);
4096 seq_printf(m, "total_%s %llu\n",
4097 mem_cgroup_events_names[i], val);
4100 for (i = 0; i < NR_LRU_LISTS; i++) {
4101 unsigned long long val = 0;
4103 for_each_mem_cgroup_tree(mi, memcg)
4104 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
4105 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
4108 #ifdef CONFIG_DEBUG_VM
4111 struct mem_cgroup_per_zone *mz;
4112 struct zone_reclaim_stat *rstat;
4113 unsigned long recent_rotated[2] = {0, 0};
4114 unsigned long recent_scanned[2] = {0, 0};
4116 for_each_online_node(nid)
4117 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4118 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
4119 rstat = &mz->lruvec.reclaim_stat;
4121 recent_rotated[0] += rstat->recent_rotated[0];
4122 recent_rotated[1] += rstat->recent_rotated[1];
4123 recent_scanned[0] += rstat->recent_scanned[0];
4124 recent_scanned[1] += rstat->recent_scanned[1];
4126 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
4127 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
4128 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
4129 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
4136 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
4139 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4141 return mem_cgroup_swappiness(memcg);
4144 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
4145 struct cftype *cft, u64 val)
4147 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4153 memcg->swappiness = val;
4155 vm_swappiness = val;
4160 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4162 struct mem_cgroup_threshold_ary *t;
4163 unsigned long usage;
4168 t = rcu_dereference(memcg->thresholds.primary);
4170 t = rcu_dereference(memcg->memsw_thresholds.primary);
4175 usage = mem_cgroup_usage(memcg, swap);
4178 * current_threshold points to threshold just below or equal to usage.
4179 * If it's not true, a threshold was crossed after last
4180 * call of __mem_cgroup_threshold().
4182 i = t->current_threshold;
4185 * Iterate backward over array of thresholds starting from
4186 * current_threshold and check if a threshold is crossed.
4187 * If none of thresholds below usage is crossed, we read
4188 * only one element of the array here.
4190 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4191 eventfd_signal(t->entries[i].eventfd, 1);
4193 /* i = current_threshold + 1 */
4197 * Iterate forward over array of thresholds starting from
4198 * current_threshold+1 and check if a threshold is crossed.
4199 * If none of thresholds above usage is crossed, we read
4200 * only one element of the array here.
4202 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4203 eventfd_signal(t->entries[i].eventfd, 1);
4205 /* Update current_threshold */
4206 t->current_threshold = i - 1;
4211 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4214 __mem_cgroup_threshold(memcg, false);
4215 if (do_swap_account)
4216 __mem_cgroup_threshold(memcg, true);
4218 memcg = parent_mem_cgroup(memcg);
4222 static int compare_thresholds(const void *a, const void *b)
4224 const struct mem_cgroup_threshold *_a = a;
4225 const struct mem_cgroup_threshold *_b = b;
4227 if (_a->threshold > _b->threshold)
4230 if (_a->threshold < _b->threshold)
4236 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4238 struct mem_cgroup_eventfd_list *ev;
4240 spin_lock(&memcg_oom_lock);
4242 list_for_each_entry(ev, &memcg->oom_notify, list)
4243 eventfd_signal(ev->eventfd, 1);
4245 spin_unlock(&memcg_oom_lock);
4249 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4251 struct mem_cgroup *iter;
4253 for_each_mem_cgroup_tree(iter, memcg)
4254 mem_cgroup_oom_notify_cb(iter);
4257 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4258 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4260 struct mem_cgroup_thresholds *thresholds;
4261 struct mem_cgroup_threshold_ary *new;
4262 unsigned long threshold;
4263 unsigned long usage;
4266 ret = page_counter_memparse(args, &threshold);
4270 mutex_lock(&memcg->thresholds_lock);
4273 thresholds = &memcg->thresholds;
4274 usage = mem_cgroup_usage(memcg, false);
4275 } else if (type == _MEMSWAP) {
4276 thresholds = &memcg->memsw_thresholds;
4277 usage = mem_cgroup_usage(memcg, true);
4281 /* Check if a threshold crossed before adding a new one */
4282 if (thresholds->primary)
4283 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4285 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4287 /* Allocate memory for new array of thresholds */
4288 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4296 /* Copy thresholds (if any) to new array */
4297 if (thresholds->primary) {
4298 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4299 sizeof(struct mem_cgroup_threshold));
4302 /* Add new threshold */
4303 new->entries[size - 1].eventfd = eventfd;
4304 new->entries[size - 1].threshold = threshold;
4306 /* Sort thresholds. Registering of new threshold isn't time-critical */
4307 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4308 compare_thresholds, NULL);
4310 /* Find current threshold */
4311 new->current_threshold = -1;
4312 for (i = 0; i < size; i++) {
4313 if (new->entries[i].threshold <= usage) {
4315 * new->current_threshold will not be used until
4316 * rcu_assign_pointer(), so it's safe to increment
4319 ++new->current_threshold;
4324 /* Free old spare buffer and save old primary buffer as spare */
4325 kfree(thresholds->spare);
4326 thresholds->spare = thresholds->primary;
4328 rcu_assign_pointer(thresholds->primary, new);
4330 /* To be sure that nobody uses thresholds */
4334 mutex_unlock(&memcg->thresholds_lock);
4339 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4340 struct eventfd_ctx *eventfd, const char *args)
4342 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4345 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4346 struct eventfd_ctx *eventfd, const char *args)
4348 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4351 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4352 struct eventfd_ctx *eventfd, enum res_type type)
4354 struct mem_cgroup_thresholds *thresholds;
4355 struct mem_cgroup_threshold_ary *new;
4356 unsigned long usage;
4359 mutex_lock(&memcg->thresholds_lock);
4362 thresholds = &memcg->thresholds;
4363 usage = mem_cgroup_usage(memcg, false);
4364 } else if (type == _MEMSWAP) {
4365 thresholds = &memcg->memsw_thresholds;
4366 usage = mem_cgroup_usage(memcg, true);
4370 if (!thresholds->primary)
4373 /* Check if a threshold crossed before removing */
4374 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4376 /* Calculate new number of threshold */
4378 for (i = 0; i < thresholds->primary->size; i++) {
4379 if (thresholds->primary->entries[i].eventfd != eventfd)
4383 new = thresholds->spare;
4385 /* Set thresholds array to NULL if we don't have thresholds */
4394 /* Copy thresholds and find current threshold */
4395 new->current_threshold = -1;
4396 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4397 if (thresholds->primary->entries[i].eventfd == eventfd)
4400 new->entries[j] = thresholds->primary->entries[i];
4401 if (new->entries[j].threshold <= usage) {
4403 * new->current_threshold will not be used
4404 * until rcu_assign_pointer(), so it's safe to increment
4407 ++new->current_threshold;
4413 /* Swap primary and spare array */
4414 thresholds->spare = thresholds->primary;
4415 /* If all events are unregistered, free the spare array */
4417 kfree(thresholds->spare);
4418 thresholds->spare = NULL;
4421 rcu_assign_pointer(thresholds->primary, new);
4423 /* To be sure that nobody uses thresholds */
4426 mutex_unlock(&memcg->thresholds_lock);
4429 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4430 struct eventfd_ctx *eventfd)
4432 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4435 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4436 struct eventfd_ctx *eventfd)
4438 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4441 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4442 struct eventfd_ctx *eventfd, const char *args)
4444 struct mem_cgroup_eventfd_list *event;
4446 event = kmalloc(sizeof(*event), GFP_KERNEL);
4450 spin_lock(&memcg_oom_lock);
4452 event->eventfd = eventfd;
4453 list_add(&event->list, &memcg->oom_notify);
4455 /* already in OOM ? */
4456 if (atomic_read(&memcg->under_oom))
4457 eventfd_signal(eventfd, 1);
4458 spin_unlock(&memcg_oom_lock);
4463 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4464 struct eventfd_ctx *eventfd)
4466 struct mem_cgroup_eventfd_list *ev, *tmp;
4468 spin_lock(&memcg_oom_lock);
4470 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4471 if (ev->eventfd == eventfd) {
4472 list_del(&ev->list);
4477 spin_unlock(&memcg_oom_lock);
4480 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4482 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
4484 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4485 seq_printf(sf, "under_oom %d\n", (bool)atomic_read(&memcg->under_oom));
4489 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4490 struct cftype *cft, u64 val)
4492 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4494 /* cannot set to root cgroup and only 0 and 1 are allowed */
4495 if (!css->parent || !((val == 0) || (val == 1)))
4498 memcg->oom_kill_disable = val;
4500 memcg_oom_recover(memcg);
4505 #ifdef CONFIG_MEMCG_KMEM
4506 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
4510 memcg->kmemcg_id = -1;
4511 ret = memcg_propagate_kmem(memcg);
4515 return mem_cgroup_sockets_init(memcg, ss);
4518 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
4520 mem_cgroup_sockets_destroy(memcg);
4523 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
4528 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
4534 * DO NOT USE IN NEW FILES.
4536 * "cgroup.event_control" implementation.
4538 * This is way over-engineered. It tries to support fully configurable
4539 * events for each user. Such level of flexibility is completely
4540 * unnecessary especially in the light of the planned unified hierarchy.
4542 * Please deprecate this and replace with something simpler if at all
4547 * Unregister event and free resources.
4549 * Gets called from workqueue.
4551 static void memcg_event_remove(struct work_struct *work)
4553 struct mem_cgroup_event *event =
4554 container_of(work, struct mem_cgroup_event, remove);
4555 struct mem_cgroup *memcg = event->memcg;
4557 remove_wait_queue(event->wqh, &event->wait);
4559 event->unregister_event(memcg, event->eventfd);
4561 /* Notify userspace the event is going away. */
4562 eventfd_signal(event->eventfd, 1);
4564 eventfd_ctx_put(event->eventfd);
4566 css_put(&memcg->css);
4570 * Gets called on POLLHUP on eventfd when user closes it.
4572 * Called with wqh->lock held and interrupts disabled.
4574 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
4575 int sync, void *key)
4577 struct mem_cgroup_event *event =
4578 container_of(wait, struct mem_cgroup_event, wait);
4579 struct mem_cgroup *memcg = event->memcg;
4580 unsigned long flags = (unsigned long)key;
4582 if (flags & POLLHUP) {
4584 * If the event has been detached at cgroup removal, we
4585 * can simply return knowing the other side will cleanup
4588 * We can't race against event freeing since the other
4589 * side will require wqh->lock via remove_wait_queue(),
4592 spin_lock(&memcg->event_list_lock);
4593 if (!list_empty(&event->list)) {
4594 list_del_init(&event->list);
4596 * We are in atomic context, but cgroup_event_remove()
4597 * may sleep, so we have to call it in workqueue.
4599 schedule_work(&event->remove);
4601 spin_unlock(&memcg->event_list_lock);
4607 static void memcg_event_ptable_queue_proc(struct file *file,
4608 wait_queue_head_t *wqh, poll_table *pt)
4610 struct mem_cgroup_event *event =
4611 container_of(pt, struct mem_cgroup_event, pt);
4614 add_wait_queue(wqh, &event->wait);
4618 * DO NOT USE IN NEW FILES.
4620 * Parse input and register new cgroup event handler.
4622 * Input must be in format '<event_fd> <control_fd> <args>'.
4623 * Interpretation of args is defined by control file implementation.
4625 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4626 char *buf, size_t nbytes, loff_t off)
4628 struct cgroup_subsys_state *css = of_css(of);
4629 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4630 struct mem_cgroup_event *event;
4631 struct cgroup_subsys_state *cfile_css;
4632 unsigned int efd, cfd;
4639 buf = strstrip(buf);
4641 efd = simple_strtoul(buf, &endp, 10);
4646 cfd = simple_strtoul(buf, &endp, 10);
4647 if ((*endp != ' ') && (*endp != '\0'))
4651 event = kzalloc(sizeof(*event), GFP_KERNEL);
4655 event->memcg = memcg;
4656 INIT_LIST_HEAD(&event->list);
4657 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4658 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4659 INIT_WORK(&event->remove, memcg_event_remove);
4667 event->eventfd = eventfd_ctx_fileget(efile.file);
4668 if (IS_ERR(event->eventfd)) {
4669 ret = PTR_ERR(event->eventfd);
4676 goto out_put_eventfd;
4679 /* the process need read permission on control file */
4680 /* AV: shouldn't we check that it's been opened for read instead? */
4681 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4686 * Determine the event callbacks and set them in @event. This used
4687 * to be done via struct cftype but cgroup core no longer knows
4688 * about these events. The following is crude but the whole thing
4689 * is for compatibility anyway.
4691 * DO NOT ADD NEW FILES.
4693 name = cfile.file->f_dentry->d_name.name;
4695 if (!strcmp(name, "memory.usage_in_bytes")) {
4696 event->register_event = mem_cgroup_usage_register_event;
4697 event->unregister_event = mem_cgroup_usage_unregister_event;
4698 } else if (!strcmp(name, "memory.oom_control")) {
4699 event->register_event = mem_cgroup_oom_register_event;
4700 event->unregister_event = mem_cgroup_oom_unregister_event;
4701 } else if (!strcmp(name, "memory.pressure_level")) {
4702 event->register_event = vmpressure_register_event;
4703 event->unregister_event = vmpressure_unregister_event;
4704 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4705 event->register_event = memsw_cgroup_usage_register_event;
4706 event->unregister_event = memsw_cgroup_usage_unregister_event;
4713 * Verify @cfile should belong to @css. Also, remaining events are
4714 * automatically removed on cgroup destruction but the removal is
4715 * asynchronous, so take an extra ref on @css.
4717 cfile_css = css_tryget_online_from_dir(cfile.file->f_dentry->d_parent,
4718 &memory_cgrp_subsys);
4720 if (IS_ERR(cfile_css))
4722 if (cfile_css != css) {
4727 ret = event->register_event(memcg, event->eventfd, buf);
4731 efile.file->f_op->poll(efile.file, &event->pt);
4733 spin_lock(&memcg->event_list_lock);
4734 list_add(&event->list, &memcg->event_list);
4735 spin_unlock(&memcg->event_list_lock);
4747 eventfd_ctx_put(event->eventfd);
4756 static struct cftype mem_cgroup_files[] = {
4758 .name = "usage_in_bytes",
4759 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4760 .read_u64 = mem_cgroup_read_u64,
4763 .name = "max_usage_in_bytes",
4764 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4765 .write = mem_cgroup_reset,
4766 .read_u64 = mem_cgroup_read_u64,
4769 .name = "limit_in_bytes",
4770 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4771 .write = mem_cgroup_write,
4772 .read_u64 = mem_cgroup_read_u64,
4775 .name = "soft_limit_in_bytes",
4776 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4777 .write = mem_cgroup_write,
4778 .read_u64 = mem_cgroup_read_u64,
4782 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4783 .write = mem_cgroup_reset,
4784 .read_u64 = mem_cgroup_read_u64,
4788 .seq_show = memcg_stat_show,
4791 .name = "force_empty",
4792 .write = mem_cgroup_force_empty_write,
4795 .name = "use_hierarchy",
4796 .write_u64 = mem_cgroup_hierarchy_write,
4797 .read_u64 = mem_cgroup_hierarchy_read,
4800 .name = "cgroup.event_control", /* XXX: for compat */
4801 .write = memcg_write_event_control,
4802 .flags = CFTYPE_NO_PREFIX,
4806 .name = "swappiness",
4807 .read_u64 = mem_cgroup_swappiness_read,
4808 .write_u64 = mem_cgroup_swappiness_write,
4811 .name = "move_charge_at_immigrate",
4812 .read_u64 = mem_cgroup_move_charge_read,
4813 .write_u64 = mem_cgroup_move_charge_write,
4816 .name = "oom_control",
4817 .seq_show = mem_cgroup_oom_control_read,
4818 .write_u64 = mem_cgroup_oom_control_write,
4819 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4822 .name = "pressure_level",
4826 .name = "numa_stat",
4827 .seq_show = memcg_numa_stat_show,
4830 #ifdef CONFIG_MEMCG_KMEM
4832 .name = "kmem.limit_in_bytes",
4833 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4834 .write = mem_cgroup_write,
4835 .read_u64 = mem_cgroup_read_u64,
4838 .name = "kmem.usage_in_bytes",
4839 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4840 .read_u64 = mem_cgroup_read_u64,
4843 .name = "kmem.failcnt",
4844 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4845 .write = mem_cgroup_reset,
4846 .read_u64 = mem_cgroup_read_u64,
4849 .name = "kmem.max_usage_in_bytes",
4850 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4851 .write = mem_cgroup_reset,
4852 .read_u64 = mem_cgroup_read_u64,
4854 #ifdef CONFIG_SLABINFO
4856 .name = "kmem.slabinfo",
4857 .seq_show = mem_cgroup_slabinfo_read,
4861 { }, /* terminate */
4864 #ifdef CONFIG_MEMCG_SWAP
4865 static struct cftype memsw_cgroup_files[] = {
4867 .name = "memsw.usage_in_bytes",
4868 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4869 .read_u64 = mem_cgroup_read_u64,
4872 .name = "memsw.max_usage_in_bytes",
4873 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4874 .write = mem_cgroup_reset,
4875 .read_u64 = mem_cgroup_read_u64,
4878 .name = "memsw.limit_in_bytes",
4879 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4880 .write = mem_cgroup_write,
4881 .read_u64 = mem_cgroup_read_u64,
4884 .name = "memsw.failcnt",
4885 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4886 .write = mem_cgroup_reset,
4887 .read_u64 = mem_cgroup_read_u64,
4889 { }, /* terminate */
4892 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4894 struct mem_cgroup_per_node *pn;
4895 struct mem_cgroup_per_zone *mz;
4896 int zone, tmp = node;
4898 * This routine is called against possible nodes.
4899 * But it's BUG to call kmalloc() against offline node.
4901 * TODO: this routine can waste much memory for nodes which will
4902 * never be onlined. It's better to use memory hotplug callback
4905 if (!node_state(node, N_NORMAL_MEMORY))
4907 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4911 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4912 mz = &pn->zoneinfo[zone];
4913 lruvec_init(&mz->lruvec);
4914 mz->usage_in_excess = 0;
4915 mz->on_tree = false;
4918 memcg->nodeinfo[node] = pn;
4922 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4924 kfree(memcg->nodeinfo[node]);
4927 static struct mem_cgroup *mem_cgroup_alloc(void)
4929 struct mem_cgroup *memcg;
4932 size = sizeof(struct mem_cgroup);
4933 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4935 memcg = kzalloc(size, GFP_KERNEL);
4939 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4942 spin_lock_init(&memcg->pcp_counter_lock);
4951 * At destroying mem_cgroup, references from swap_cgroup can remain.
4952 * (scanning all at force_empty is too costly...)
4954 * Instead of clearing all references at force_empty, we remember
4955 * the number of reference from swap_cgroup and free mem_cgroup when
4956 * it goes down to 0.
4958 * Removal of cgroup itself succeeds regardless of refs from swap.
4961 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4965 mem_cgroup_remove_from_trees(memcg);
4968 free_mem_cgroup_per_zone_info(memcg, node);
4970 free_percpu(memcg->stat);
4973 * We need to make sure that (at least for now), the jump label
4974 * destruction code runs outside of the cgroup lock. This is because
4975 * get_online_cpus(), which is called from the static_branch update,
4976 * can't be called inside the cgroup_lock. cpusets are the ones
4977 * enforcing this dependency, so if they ever change, we might as well.
4979 * schedule_work() will guarantee this happens. Be careful if you need
4980 * to move this code around, and make sure it is outside
4983 disarm_static_keys(memcg);
4988 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4990 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4992 if (!memcg->memory.parent)
4994 return mem_cgroup_from_counter(memcg->memory.parent, memory);
4996 EXPORT_SYMBOL(parent_mem_cgroup);
4998 static void __init mem_cgroup_soft_limit_tree_init(void)
5000 struct mem_cgroup_tree_per_node *rtpn;
5001 struct mem_cgroup_tree_per_zone *rtpz;
5002 int tmp, node, zone;
5004 for_each_node(node) {
5006 if (!node_state(node, N_NORMAL_MEMORY))
5008 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
5011 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5013 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5014 rtpz = &rtpn->rb_tree_per_zone[zone];
5015 rtpz->rb_root = RB_ROOT;
5016 spin_lock_init(&rtpz->lock);
5021 static struct cgroup_subsys_state * __ref
5022 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5024 struct mem_cgroup *memcg;
5025 long error = -ENOMEM;
5028 memcg = mem_cgroup_alloc();
5030 return ERR_PTR(error);
5033 if (alloc_mem_cgroup_per_zone_info(memcg, node))
5037 if (parent_css == NULL) {
5038 root_mem_cgroup = memcg;
5039 page_counter_init(&memcg->memory, NULL);
5040 page_counter_init(&memcg->memsw, NULL);
5041 page_counter_init(&memcg->kmem, NULL);
5044 memcg->last_scanned_node = MAX_NUMNODES;
5045 INIT_LIST_HEAD(&memcg->oom_notify);
5046 memcg->move_charge_at_immigrate = 0;
5047 mutex_init(&memcg->thresholds_lock);
5048 spin_lock_init(&memcg->move_lock);
5049 vmpressure_init(&memcg->vmpressure);
5050 INIT_LIST_HEAD(&memcg->event_list);
5051 spin_lock_init(&memcg->event_list_lock);
5056 __mem_cgroup_free(memcg);
5057 return ERR_PTR(error);
5061 mem_cgroup_css_online(struct cgroup_subsys_state *css)
5063 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5064 struct mem_cgroup *parent = mem_cgroup_from_css(css->parent);
5067 if (css->id > MEM_CGROUP_ID_MAX)
5073 mutex_lock(&memcg_create_mutex);
5075 memcg->use_hierarchy = parent->use_hierarchy;
5076 memcg->oom_kill_disable = parent->oom_kill_disable;
5077 memcg->swappiness = mem_cgroup_swappiness(parent);
5079 if (parent->use_hierarchy) {
5080 page_counter_init(&memcg->memory, &parent->memory);
5081 page_counter_init(&memcg->memsw, &parent->memsw);
5082 page_counter_init(&memcg->kmem, &parent->kmem);
5085 * No need to take a reference to the parent because cgroup
5086 * core guarantees its existence.
5089 page_counter_init(&memcg->memory, NULL);
5090 page_counter_init(&memcg->memsw, NULL);
5091 page_counter_init(&memcg->kmem, NULL);
5093 * Deeper hierachy with use_hierarchy == false doesn't make
5094 * much sense so let cgroup subsystem know about this
5095 * unfortunate state in our controller.
5097 if (parent != root_mem_cgroup)
5098 memory_cgrp_subsys.broken_hierarchy = true;
5100 mutex_unlock(&memcg_create_mutex);
5102 ret = memcg_init_kmem(memcg, &memory_cgrp_subsys);
5107 * Make sure the memcg is initialized: mem_cgroup_iter()
5108 * orders reading memcg->initialized against its callers
5109 * reading the memcg members.
5111 smp_store_release(&memcg->initialized, 1);
5116 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5118 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5119 struct mem_cgroup_event *event, *tmp;
5122 * Unregister events and notify userspace.
5123 * Notify userspace about cgroup removing only after rmdir of cgroup
5124 * directory to avoid race between userspace and kernelspace.
5126 spin_lock(&memcg->event_list_lock);
5127 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5128 list_del_init(&event->list);
5129 schedule_work(&event->remove);
5131 spin_unlock(&memcg->event_list_lock);
5133 memcg_unregister_all_caches(memcg);
5134 vmpressure_cleanup(&memcg->vmpressure);
5137 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5139 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5141 memcg_destroy_kmem(memcg);
5142 __mem_cgroup_free(memcg);
5146 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5147 * @css: the target css
5149 * Reset the states of the mem_cgroup associated with @css. This is
5150 * invoked when the userland requests disabling on the default hierarchy
5151 * but the memcg is pinned through dependency. The memcg should stop
5152 * applying policies and should revert to the vanilla state as it may be
5153 * made visible again.
5155 * The current implementation only resets the essential configurations.
5156 * This needs to be expanded to cover all the visible parts.
5158 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5160 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5162 mem_cgroup_resize_limit(memcg, PAGE_COUNTER_MAX);
5163 mem_cgroup_resize_memsw_limit(memcg, PAGE_COUNTER_MAX);
5164 memcg_update_kmem_limit(memcg, PAGE_COUNTER_MAX);
5165 memcg->soft_limit = 0;
5169 /* Handlers for move charge at task migration. */
5170 static int mem_cgroup_do_precharge(unsigned long count)
5174 /* Try a single bulk charge without reclaim first */
5175 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_WAIT, count);
5177 mc.precharge += count;
5180 if (ret == -EINTR) {
5181 cancel_charge(root_mem_cgroup, count);
5185 /* Try charges one by one with reclaim */
5187 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_NORETRY, 1);
5189 * In case of failure, any residual charges against
5190 * mc.to will be dropped by mem_cgroup_clear_mc()
5191 * later on. However, cancel any charges that are
5192 * bypassed to root right away or they'll be lost.
5195 cancel_charge(root_mem_cgroup, 1);
5205 * get_mctgt_type - get target type of moving charge
5206 * @vma: the vma the pte to be checked belongs
5207 * @addr: the address corresponding to the pte to be checked
5208 * @ptent: the pte to be checked
5209 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5212 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5213 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5214 * move charge. if @target is not NULL, the page is stored in target->page
5215 * with extra refcnt got(Callers should handle it).
5216 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5217 * target for charge migration. if @target is not NULL, the entry is stored
5220 * Called with pte lock held.
5227 enum mc_target_type {
5233 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5234 unsigned long addr, pte_t ptent)
5236 struct page *page = vm_normal_page(vma, addr, ptent);
5238 if (!page || !page_mapped(page))
5240 if (PageAnon(page)) {
5241 /* we don't move shared anon */
5244 } else if (!move_file())
5245 /* we ignore mapcount for file pages */
5247 if (!get_page_unless_zero(page))
5254 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5255 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5257 struct page *page = NULL;
5258 swp_entry_t ent = pte_to_swp_entry(ptent);
5260 if (!move_anon() || non_swap_entry(ent))
5263 * Because lookup_swap_cache() updates some statistics counter,
5264 * we call find_get_page() with swapper_space directly.
5266 page = find_get_page(swap_address_space(ent), ent.val);
5267 if (do_swap_account)
5268 entry->val = ent.val;
5273 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5274 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5280 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5281 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5283 struct page *page = NULL;
5284 struct address_space *mapping;
5287 if (!vma->vm_file) /* anonymous vma */
5292 mapping = vma->vm_file->f_mapping;
5293 if (pte_none(ptent))
5294 pgoff = linear_page_index(vma, addr);
5295 else /* pte_file(ptent) is true */
5296 pgoff = pte_to_pgoff(ptent);
5298 /* page is moved even if it's not RSS of this task(page-faulted). */
5300 /* shmem/tmpfs may report page out on swap: account for that too. */
5301 if (shmem_mapping(mapping)) {
5302 page = find_get_entry(mapping, pgoff);
5303 if (radix_tree_exceptional_entry(page)) {
5304 swp_entry_t swp = radix_to_swp_entry(page);
5305 if (do_swap_account)
5307 page = find_get_page(swap_address_space(swp), swp.val);
5310 page = find_get_page(mapping, pgoff);
5312 page = find_get_page(mapping, pgoff);
5317 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5318 unsigned long addr, pte_t ptent, union mc_target *target)
5320 struct page *page = NULL;
5321 struct page_cgroup *pc;
5322 enum mc_target_type ret = MC_TARGET_NONE;
5323 swp_entry_t ent = { .val = 0 };
5325 if (pte_present(ptent))
5326 page = mc_handle_present_pte(vma, addr, ptent);
5327 else if (is_swap_pte(ptent))
5328 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5329 else if (pte_none(ptent) || pte_file(ptent))
5330 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5332 if (!page && !ent.val)
5335 pc = lookup_page_cgroup(page);
5337 * Do only loose check w/o serialization.
5338 * mem_cgroup_move_account() checks the pc is valid or
5339 * not under LRU exclusion.
5341 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5342 ret = MC_TARGET_PAGE;
5344 target->page = page;
5346 if (!ret || !target)
5349 /* There is a swap entry and a page doesn't exist or isn't charged */
5350 if (ent.val && !ret &&
5351 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5352 ret = MC_TARGET_SWAP;
5359 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5361 * We don't consider swapping or file mapped pages because THP does not
5362 * support them for now.
5363 * Caller should make sure that pmd_trans_huge(pmd) is true.
5365 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5366 unsigned long addr, pmd_t pmd, union mc_target *target)
5368 struct page *page = NULL;
5369 struct page_cgroup *pc;
5370 enum mc_target_type ret = MC_TARGET_NONE;
5372 page = pmd_page(pmd);
5373 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5376 pc = lookup_page_cgroup(page);
5377 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5378 ret = MC_TARGET_PAGE;
5381 target->page = page;
5387 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5388 unsigned long addr, pmd_t pmd, union mc_target *target)
5390 return MC_TARGET_NONE;
5394 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5395 unsigned long addr, unsigned long end,
5396 struct mm_walk *walk)
5398 struct vm_area_struct *vma = walk->private;
5402 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
5403 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5404 mc.precharge += HPAGE_PMD_NR;
5409 if (pmd_trans_unstable(pmd))
5411 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5412 for (; addr != end; pte++, addr += PAGE_SIZE)
5413 if (get_mctgt_type(vma, addr, *pte, NULL))
5414 mc.precharge++; /* increment precharge temporarily */
5415 pte_unmap_unlock(pte - 1, ptl);
5421 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5423 unsigned long precharge;
5424 struct vm_area_struct *vma;
5426 down_read(&mm->mmap_sem);
5427 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5428 struct mm_walk mem_cgroup_count_precharge_walk = {
5429 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5433 if (is_vm_hugetlb_page(vma))
5435 walk_page_range(vma->vm_start, vma->vm_end,
5436 &mem_cgroup_count_precharge_walk);
5438 up_read(&mm->mmap_sem);
5440 precharge = mc.precharge;
5446 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5448 unsigned long precharge = mem_cgroup_count_precharge(mm);
5450 VM_BUG_ON(mc.moving_task);
5451 mc.moving_task = current;
5452 return mem_cgroup_do_precharge(precharge);
5455 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5456 static void __mem_cgroup_clear_mc(void)
5458 struct mem_cgroup *from = mc.from;
5459 struct mem_cgroup *to = mc.to;
5461 /* we must uncharge all the leftover precharges from mc.to */
5463 cancel_charge(mc.to, mc.precharge);
5467 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5468 * we must uncharge here.
5470 if (mc.moved_charge) {
5471 cancel_charge(mc.from, mc.moved_charge);
5472 mc.moved_charge = 0;
5474 /* we must fixup refcnts and charges */
5475 if (mc.moved_swap) {
5476 /* uncharge swap account from the old cgroup */
5477 if (!mem_cgroup_is_root(mc.from))
5478 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5481 * we charged both to->memory and to->memsw, so we
5482 * should uncharge to->memory.
5484 if (!mem_cgroup_is_root(mc.to))
5485 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5487 css_put_many(&mc.from->css, mc.moved_swap);
5489 /* we've already done css_get(mc.to) */
5492 memcg_oom_recover(from);
5493 memcg_oom_recover(to);
5494 wake_up_all(&mc.waitq);
5497 static void mem_cgroup_clear_mc(void)
5499 struct mem_cgroup *from = mc.from;
5502 * we must clear moving_task before waking up waiters at the end of
5505 mc.moving_task = NULL;
5506 __mem_cgroup_clear_mc();
5507 spin_lock(&mc.lock);
5510 spin_unlock(&mc.lock);
5511 mem_cgroup_end_move(from);
5514 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
5515 struct cgroup_taskset *tset)
5517 struct task_struct *p = cgroup_taskset_first(tset);
5519 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5520 unsigned long move_charge_at_immigrate;
5523 * We are now commited to this value whatever it is. Changes in this
5524 * tunable will only affect upcoming migrations, not the current one.
5525 * So we need to save it, and keep it going.
5527 move_charge_at_immigrate = memcg->move_charge_at_immigrate;
5528 if (move_charge_at_immigrate) {
5529 struct mm_struct *mm;
5530 struct mem_cgroup *from = mem_cgroup_from_task(p);
5532 VM_BUG_ON(from == memcg);
5534 mm = get_task_mm(p);
5537 /* We move charges only when we move a owner of the mm */
5538 if (mm->owner == p) {
5541 VM_BUG_ON(mc.precharge);
5542 VM_BUG_ON(mc.moved_charge);
5543 VM_BUG_ON(mc.moved_swap);
5544 mem_cgroup_start_move(from);
5545 spin_lock(&mc.lock);
5548 mc.immigrate_flags = move_charge_at_immigrate;
5549 spin_unlock(&mc.lock);
5550 /* We set mc.moving_task later */
5552 ret = mem_cgroup_precharge_mc(mm);
5554 mem_cgroup_clear_mc();
5561 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
5562 struct cgroup_taskset *tset)
5564 mem_cgroup_clear_mc();
5567 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5568 unsigned long addr, unsigned long end,
5569 struct mm_walk *walk)
5572 struct vm_area_struct *vma = walk->private;
5575 enum mc_target_type target_type;
5576 union mc_target target;
5578 struct page_cgroup *pc;
5581 * We don't take compound_lock() here but no race with splitting thp
5583 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
5584 * under splitting, which means there's no concurrent thp split,
5585 * - if another thread runs into split_huge_page() just after we
5586 * entered this if-block, the thread must wait for page table lock
5587 * to be unlocked in __split_huge_page_splitting(), where the main
5588 * part of thp split is not executed yet.
5590 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
5591 if (mc.precharge < HPAGE_PMD_NR) {
5595 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5596 if (target_type == MC_TARGET_PAGE) {
5598 if (!isolate_lru_page(page)) {
5599 pc = lookup_page_cgroup(page);
5600 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
5601 pc, mc.from, mc.to)) {
5602 mc.precharge -= HPAGE_PMD_NR;
5603 mc.moved_charge += HPAGE_PMD_NR;
5605 putback_lru_page(page);
5613 if (pmd_trans_unstable(pmd))
5616 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5617 for (; addr != end; addr += PAGE_SIZE) {
5618 pte_t ptent = *(pte++);
5624 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5625 case MC_TARGET_PAGE:
5627 if (isolate_lru_page(page))
5629 pc = lookup_page_cgroup(page);
5630 if (!mem_cgroup_move_account(page, 1, pc,
5633 /* we uncharge from mc.from later. */
5636 putback_lru_page(page);
5637 put: /* get_mctgt_type() gets the page */
5640 case MC_TARGET_SWAP:
5642 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5644 /* we fixup refcnts and charges later. */
5652 pte_unmap_unlock(pte - 1, ptl);
5657 * We have consumed all precharges we got in can_attach().
5658 * We try charge one by one, but don't do any additional
5659 * charges to mc.to if we have failed in charge once in attach()
5662 ret = mem_cgroup_do_precharge(1);
5670 static void mem_cgroup_move_charge(struct mm_struct *mm)
5672 struct vm_area_struct *vma;
5674 lru_add_drain_all();
5676 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5678 * Someone who are holding the mmap_sem might be waiting in
5679 * waitq. So we cancel all extra charges, wake up all waiters,
5680 * and retry. Because we cancel precharges, we might not be able
5681 * to move enough charges, but moving charge is a best-effort
5682 * feature anyway, so it wouldn't be a big problem.
5684 __mem_cgroup_clear_mc();
5688 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5690 struct mm_walk mem_cgroup_move_charge_walk = {
5691 .pmd_entry = mem_cgroup_move_charge_pte_range,
5695 if (is_vm_hugetlb_page(vma))
5697 ret = walk_page_range(vma->vm_start, vma->vm_end,
5698 &mem_cgroup_move_charge_walk);
5701 * means we have consumed all precharges and failed in
5702 * doing additional charge. Just abandon here.
5706 up_read(&mm->mmap_sem);
5709 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
5710 struct cgroup_taskset *tset)
5712 struct task_struct *p = cgroup_taskset_first(tset);
5713 struct mm_struct *mm = get_task_mm(p);
5717 mem_cgroup_move_charge(mm);
5721 mem_cgroup_clear_mc();
5723 #else /* !CONFIG_MMU */
5724 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
5725 struct cgroup_taskset *tset)
5729 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
5730 struct cgroup_taskset *tset)
5733 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
5734 struct cgroup_taskset *tset)
5740 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5741 * to verify whether we're attached to the default hierarchy on each mount
5744 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5747 * use_hierarchy is forced on the default hierarchy. cgroup core
5748 * guarantees that @root doesn't have any children, so turning it
5749 * on for the root memcg is enough.
5751 if (cgroup_on_dfl(root_css->cgroup))
5752 mem_cgroup_from_css(root_css)->use_hierarchy = true;
5755 struct cgroup_subsys memory_cgrp_subsys = {
5756 .css_alloc = mem_cgroup_css_alloc,
5757 .css_online = mem_cgroup_css_online,
5758 .css_offline = mem_cgroup_css_offline,
5759 .css_free = mem_cgroup_css_free,
5760 .css_reset = mem_cgroup_css_reset,
5761 .can_attach = mem_cgroup_can_attach,
5762 .cancel_attach = mem_cgroup_cancel_attach,
5763 .attach = mem_cgroup_move_task,
5764 .bind = mem_cgroup_bind,
5765 .legacy_cftypes = mem_cgroup_files,
5769 #ifdef CONFIG_MEMCG_SWAP
5770 static int __init enable_swap_account(char *s)
5772 if (!strcmp(s, "1"))
5773 really_do_swap_account = 1;
5774 else if (!strcmp(s, "0"))
5775 really_do_swap_account = 0;
5778 __setup("swapaccount=", enable_swap_account);
5780 static void __init memsw_file_init(void)
5782 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
5783 memsw_cgroup_files));
5786 static void __init enable_swap_cgroup(void)
5788 if (!mem_cgroup_disabled() && really_do_swap_account) {
5789 do_swap_account = 1;
5795 static void __init enable_swap_cgroup(void)
5800 #ifdef CONFIG_MEMCG_SWAP
5802 * mem_cgroup_swapout - transfer a memsw charge to swap
5803 * @page: page whose memsw charge to transfer
5804 * @entry: swap entry to move the charge to
5806 * Transfer the memsw charge of @page to @entry.
5808 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
5810 struct page_cgroup *pc;
5811 unsigned short oldid;
5813 VM_BUG_ON_PAGE(PageLRU(page), page);
5814 VM_BUG_ON_PAGE(page_count(page), page);
5816 if (!do_swap_account)
5819 pc = lookup_page_cgroup(page);
5821 /* Readahead page, never charged */
5822 if (!PageCgroupUsed(pc))
5825 VM_BUG_ON_PAGE(!(pc->flags & PCG_MEMSW), page);
5827 oldid = swap_cgroup_record(entry, mem_cgroup_id(pc->mem_cgroup));
5828 VM_BUG_ON_PAGE(oldid, page);
5830 pc->flags &= ~PCG_MEMSW;
5831 css_get(&pc->mem_cgroup->css);
5832 mem_cgroup_swap_statistics(pc->mem_cgroup, true);
5836 * mem_cgroup_uncharge_swap - uncharge a swap entry
5837 * @entry: swap entry to uncharge
5839 * Drop the memsw charge associated with @entry.
5841 void mem_cgroup_uncharge_swap(swp_entry_t entry)
5843 struct mem_cgroup *memcg;
5846 if (!do_swap_account)
5849 id = swap_cgroup_record(entry, 0);
5851 memcg = mem_cgroup_lookup(id);
5853 if (!mem_cgroup_is_root(memcg))
5854 page_counter_uncharge(&memcg->memsw, 1);
5855 mem_cgroup_swap_statistics(memcg, false);
5856 css_put(&memcg->css);
5863 * mem_cgroup_try_charge - try charging a page
5864 * @page: page to charge
5865 * @mm: mm context of the victim
5866 * @gfp_mask: reclaim mode
5867 * @memcgp: charged memcg return
5869 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5870 * pages according to @gfp_mask if necessary.
5872 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5873 * Otherwise, an error code is returned.
5875 * After page->mapping has been set up, the caller must finalize the
5876 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5877 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5879 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
5880 gfp_t gfp_mask, struct mem_cgroup **memcgp)
5882 struct mem_cgroup *memcg = NULL;
5883 unsigned int nr_pages = 1;
5886 if (mem_cgroup_disabled())
5889 if (PageSwapCache(page)) {
5890 struct page_cgroup *pc = lookup_page_cgroup(page);
5892 * Every swap fault against a single page tries to charge the
5893 * page, bail as early as possible. shmem_unuse() encounters
5894 * already charged pages, too. The USED bit is protected by
5895 * the page lock, which serializes swap cache removal, which
5896 * in turn serializes uncharging.
5898 if (PageCgroupUsed(pc))
5902 if (PageTransHuge(page)) {
5903 nr_pages <<= compound_order(page);
5904 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5907 if (do_swap_account && PageSwapCache(page))
5908 memcg = try_get_mem_cgroup_from_page(page);
5910 memcg = get_mem_cgroup_from_mm(mm);
5912 ret = try_charge(memcg, gfp_mask, nr_pages);
5914 css_put(&memcg->css);
5916 if (ret == -EINTR) {
5917 memcg = root_mem_cgroup;
5926 * mem_cgroup_commit_charge - commit a page charge
5927 * @page: page to charge
5928 * @memcg: memcg to charge the page to
5929 * @lrucare: page might be on LRU already
5931 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5932 * after page->mapping has been set up. This must happen atomically
5933 * as part of the page instantiation, i.e. under the page table lock
5934 * for anonymous pages, under the page lock for page and swap cache.
5936 * In addition, the page must not be on the LRU during the commit, to
5937 * prevent racing with task migration. If it might be, use @lrucare.
5939 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5941 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
5944 unsigned int nr_pages = 1;
5946 VM_BUG_ON_PAGE(!page->mapping, page);
5947 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
5949 if (mem_cgroup_disabled())
5952 * Swap faults will attempt to charge the same page multiple
5953 * times. But reuse_swap_page() might have removed the page
5954 * from swapcache already, so we can't check PageSwapCache().
5959 commit_charge(page, memcg, lrucare);
5961 if (PageTransHuge(page)) {
5962 nr_pages <<= compound_order(page);
5963 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5966 local_irq_disable();
5967 mem_cgroup_charge_statistics(memcg, page, nr_pages);
5968 memcg_check_events(memcg, page);
5971 if (do_swap_account && PageSwapCache(page)) {
5972 swp_entry_t entry = { .val = page_private(page) };
5974 * The swap entry might not get freed for a long time,
5975 * let's not wait for it. The page already received a
5976 * memory+swap charge, drop the swap entry duplicate.
5978 mem_cgroup_uncharge_swap(entry);
5983 * mem_cgroup_cancel_charge - cancel a page charge
5984 * @page: page to charge
5985 * @memcg: memcg to charge the page to
5987 * Cancel a charge transaction started by mem_cgroup_try_charge().
5989 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg)
5991 unsigned int nr_pages = 1;
5993 if (mem_cgroup_disabled())
5996 * Swap faults will attempt to charge the same page multiple
5997 * times. But reuse_swap_page() might have removed the page
5998 * from swapcache already, so we can't check PageSwapCache().
6003 if (PageTransHuge(page)) {
6004 nr_pages <<= compound_order(page);
6005 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
6008 cancel_charge(memcg, nr_pages);
6011 static void uncharge_batch(struct mem_cgroup *memcg, unsigned long pgpgout,
6012 unsigned long nr_mem, unsigned long nr_memsw,
6013 unsigned long nr_anon, unsigned long nr_file,
6014 unsigned long nr_huge, struct page *dummy_page)
6016 unsigned long flags;
6018 if (!mem_cgroup_is_root(memcg)) {
6020 page_counter_uncharge(&memcg->memory, nr_mem);
6022 page_counter_uncharge(&memcg->memsw, nr_memsw);
6023 memcg_oom_recover(memcg);
6026 local_irq_save(flags);
6027 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS], nr_anon);
6028 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_CACHE], nr_file);
6029 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE], nr_huge);
6030 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT], pgpgout);
6031 __this_cpu_add(memcg->stat->nr_page_events, nr_anon + nr_file);
6032 memcg_check_events(memcg, dummy_page);
6033 local_irq_restore(flags);
6035 if (!mem_cgroup_is_root(memcg))
6036 css_put_many(&memcg->css, max(nr_mem, nr_memsw));
6039 static void uncharge_list(struct list_head *page_list)
6041 struct mem_cgroup *memcg = NULL;
6042 unsigned long nr_memsw = 0;
6043 unsigned long nr_anon = 0;
6044 unsigned long nr_file = 0;
6045 unsigned long nr_huge = 0;
6046 unsigned long pgpgout = 0;
6047 unsigned long nr_mem = 0;
6048 struct list_head *next;
6051 next = page_list->next;
6053 unsigned int nr_pages = 1;
6054 struct page_cgroup *pc;
6056 page = list_entry(next, struct page, lru);
6057 next = page->lru.next;
6059 VM_BUG_ON_PAGE(PageLRU(page), page);
6060 VM_BUG_ON_PAGE(page_count(page), page);
6062 pc = lookup_page_cgroup(page);
6063 if (!PageCgroupUsed(pc))
6067 * Nobody should be changing or seriously looking at
6068 * pc->mem_cgroup and pc->flags at this point, we have
6069 * fully exclusive access to the page.
6072 if (memcg != pc->mem_cgroup) {
6074 uncharge_batch(memcg, pgpgout, nr_mem, nr_memsw,
6075 nr_anon, nr_file, nr_huge, page);
6076 pgpgout = nr_mem = nr_memsw = 0;
6077 nr_anon = nr_file = nr_huge = 0;
6079 memcg = pc->mem_cgroup;
6082 if (PageTransHuge(page)) {
6083 nr_pages <<= compound_order(page);
6084 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
6085 nr_huge += nr_pages;
6089 nr_anon += nr_pages;
6091 nr_file += nr_pages;
6093 if (pc->flags & PCG_MEM)
6095 if (pc->flags & PCG_MEMSW)
6096 nr_memsw += nr_pages;
6100 } while (next != page_list);
6103 uncharge_batch(memcg, pgpgout, nr_mem, nr_memsw,
6104 nr_anon, nr_file, nr_huge, page);
6108 * mem_cgroup_uncharge - uncharge a page
6109 * @page: page to uncharge
6111 * Uncharge a page previously charged with mem_cgroup_try_charge() and
6112 * mem_cgroup_commit_charge().
6114 void mem_cgroup_uncharge(struct page *page)
6116 struct page_cgroup *pc;
6118 if (mem_cgroup_disabled())
6121 /* Don't touch page->lru of any random page, pre-check: */
6122 pc = lookup_page_cgroup(page);
6123 if (!PageCgroupUsed(pc))
6126 INIT_LIST_HEAD(&page->lru);
6127 uncharge_list(&page->lru);
6131 * mem_cgroup_uncharge_list - uncharge a list of page
6132 * @page_list: list of pages to uncharge
6134 * Uncharge a list of pages previously charged with
6135 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
6137 void mem_cgroup_uncharge_list(struct list_head *page_list)
6139 if (mem_cgroup_disabled())
6142 if (!list_empty(page_list))
6143 uncharge_list(page_list);
6147 * mem_cgroup_migrate - migrate a charge to another page
6148 * @oldpage: currently charged page
6149 * @newpage: page to transfer the charge to
6150 * @lrucare: both pages might be on the LRU already
6152 * Migrate the charge from @oldpage to @newpage.
6154 * Both pages must be locked, @newpage->mapping must be set up.
6156 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage,
6159 struct page_cgroup *pc;
6162 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6163 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6164 VM_BUG_ON_PAGE(!lrucare && PageLRU(oldpage), oldpage);
6165 VM_BUG_ON_PAGE(!lrucare && PageLRU(newpage), newpage);
6166 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6167 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6170 if (mem_cgroup_disabled())
6173 /* Page cache replacement: new page already charged? */
6174 pc = lookup_page_cgroup(newpage);
6175 if (PageCgroupUsed(pc))
6178 /* Re-entrant migration: old page already uncharged? */
6179 pc = lookup_page_cgroup(oldpage);
6180 if (!PageCgroupUsed(pc))
6183 VM_BUG_ON_PAGE(!(pc->flags & PCG_MEM), oldpage);
6184 VM_BUG_ON_PAGE(do_swap_account && !(pc->flags & PCG_MEMSW), oldpage);
6187 lock_page_lru(oldpage, &isolated);
6192 unlock_page_lru(oldpage, isolated);
6194 commit_charge(newpage, pc->mem_cgroup, lrucare);
6198 * subsys_initcall() for memory controller.
6200 * Some parts like hotcpu_notifier() have to be initialized from this context
6201 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
6202 * everything that doesn't depend on a specific mem_cgroup structure should
6203 * be initialized from here.
6205 static int __init mem_cgroup_init(void)
6207 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
6208 enable_swap_cgroup();
6209 mem_cgroup_soft_limit_tree_init();
6213 subsys_initcall(mem_cgroup_init);