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 isolating/putting an LRU page
1267 * @zone: zone of the page
1269 * This function is only safe when following the LRU page isolation
1270 * and putback protocol: the LRU lock must be held, and the page must
1271 * either be PageLRU() or the caller must have isolated/allocated it.
1273 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1275 struct mem_cgroup_per_zone *mz;
1276 struct mem_cgroup *memcg;
1277 struct page_cgroup *pc;
1278 struct lruvec *lruvec;
1280 if (mem_cgroup_disabled()) {
1281 lruvec = &zone->lruvec;
1285 pc = lookup_page_cgroup(page);
1286 memcg = pc->mem_cgroup;
1288 * Swapcache readahead pages are added to the LRU - and
1289 * possibly migrated - before they are charged.
1292 memcg = root_mem_cgroup;
1294 mz = mem_cgroup_page_zoneinfo(memcg, page);
1295 lruvec = &mz->lruvec;
1298 * Since a node can be onlined after the mem_cgroup was created,
1299 * we have to be prepared to initialize lruvec->zone here;
1300 * and if offlined then reonlined, we need to reinitialize it.
1302 if (unlikely(lruvec->zone != zone))
1303 lruvec->zone = zone;
1308 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1309 * @lruvec: mem_cgroup per zone lru vector
1310 * @lru: index of lru list the page is sitting on
1311 * @nr_pages: positive when adding or negative when removing
1313 * This function must be called when a page is added to or removed from an
1316 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1319 struct mem_cgroup_per_zone *mz;
1320 unsigned long *lru_size;
1322 if (mem_cgroup_disabled())
1325 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1326 lru_size = mz->lru_size + lru;
1327 *lru_size += nr_pages;
1328 VM_BUG_ON((long)(*lru_size) < 0);
1332 * Checks whether given mem is same or in the root_mem_cgroup's
1335 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1336 struct mem_cgroup *memcg)
1338 if (root_memcg == memcg)
1340 if (!root_memcg->use_hierarchy || !memcg)
1342 return cgroup_is_descendant(memcg->css.cgroup, root_memcg->css.cgroup);
1345 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1346 struct mem_cgroup *memcg)
1351 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1356 bool task_in_mem_cgroup(struct task_struct *task,
1357 const struct mem_cgroup *memcg)
1359 struct mem_cgroup *curr = NULL;
1360 struct task_struct *p;
1363 p = find_lock_task_mm(task);
1365 curr = get_mem_cgroup_from_mm(p->mm);
1369 * All threads may have already detached their mm's, but the oom
1370 * killer still needs to detect if they have already been oom
1371 * killed to prevent needlessly killing additional tasks.
1374 curr = mem_cgroup_from_task(task);
1376 css_get(&curr->css);
1380 * We should check use_hierarchy of "memcg" not "curr". Because checking
1381 * use_hierarchy of "curr" here make this function true if hierarchy is
1382 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1383 * hierarchy(even if use_hierarchy is disabled in "memcg").
1385 ret = mem_cgroup_same_or_subtree(memcg, curr);
1386 css_put(&curr->css);
1390 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1392 unsigned long inactive_ratio;
1393 unsigned long inactive;
1394 unsigned long active;
1397 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1398 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1400 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1402 inactive_ratio = int_sqrt(10 * gb);
1406 return inactive * inactive_ratio < active;
1409 #define mem_cgroup_from_counter(counter, member) \
1410 container_of(counter, struct mem_cgroup, member)
1413 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1414 * @memcg: the memory cgroup
1416 * Returns the maximum amount of memory @mem can be charged with, in
1419 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1421 unsigned long margin = 0;
1422 unsigned long count;
1423 unsigned long limit;
1425 count = page_counter_read(&memcg->memory);
1426 limit = ACCESS_ONCE(memcg->memory.limit);
1428 margin = limit - count;
1430 if (do_swap_account) {
1431 count = page_counter_read(&memcg->memsw);
1432 limit = ACCESS_ONCE(memcg->memsw.limit);
1434 margin = min(margin, limit - count);
1440 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1443 if (mem_cgroup_disabled() || !memcg->css.parent)
1444 return vm_swappiness;
1446 return memcg->swappiness;
1450 * A routine for checking "mem" is under move_account() or not.
1452 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1453 * moving cgroups. This is for waiting at high-memory pressure
1456 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1458 struct mem_cgroup *from;
1459 struct mem_cgroup *to;
1462 * Unlike task_move routines, we access mc.to, mc.from not under
1463 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1465 spin_lock(&mc.lock);
1471 ret = mem_cgroup_same_or_subtree(memcg, from)
1472 || mem_cgroup_same_or_subtree(memcg, to);
1474 spin_unlock(&mc.lock);
1478 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1480 if (mc.moving_task && current != mc.moving_task) {
1481 if (mem_cgroup_under_move(memcg)) {
1483 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1484 /* moving charge context might have finished. */
1487 finish_wait(&mc.waitq, &wait);
1494 #define K(x) ((x) << (PAGE_SHIFT-10))
1496 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1497 * @memcg: The memory cgroup that went over limit
1498 * @p: Task that is going to be killed
1500 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1503 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1505 /* oom_info_lock ensures that parallel ooms do not interleave */
1506 static DEFINE_MUTEX(oom_info_lock);
1507 struct mem_cgroup *iter;
1513 mutex_lock(&oom_info_lock);
1516 pr_info("Task in ");
1517 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1518 pr_info(" killed as a result of limit of ");
1519 pr_cont_cgroup_path(memcg->css.cgroup);
1524 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1525 K((u64)page_counter_read(&memcg->memory)),
1526 K((u64)memcg->memory.limit), memcg->memory.failcnt);
1527 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1528 K((u64)page_counter_read(&memcg->memsw)),
1529 K((u64)memcg->memsw.limit), memcg->memsw.failcnt);
1530 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1531 K((u64)page_counter_read(&memcg->kmem)),
1532 K((u64)memcg->kmem.limit), memcg->kmem.failcnt);
1534 for_each_mem_cgroup_tree(iter, memcg) {
1535 pr_info("Memory cgroup stats for ");
1536 pr_cont_cgroup_path(iter->css.cgroup);
1539 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1540 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1542 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1543 K(mem_cgroup_read_stat(iter, i)));
1546 for (i = 0; i < NR_LRU_LISTS; i++)
1547 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1548 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1552 mutex_unlock(&oom_info_lock);
1556 * This function returns the number of memcg under hierarchy tree. Returns
1557 * 1(self count) if no children.
1559 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1562 struct mem_cgroup *iter;
1564 for_each_mem_cgroup_tree(iter, memcg)
1570 * Return the memory (and swap, if configured) limit for a memcg.
1572 static unsigned long mem_cgroup_get_limit(struct mem_cgroup *memcg)
1574 unsigned long limit;
1576 limit = memcg->memory.limit;
1577 if (mem_cgroup_swappiness(memcg)) {
1578 unsigned long memsw_limit;
1580 memsw_limit = memcg->memsw.limit;
1581 limit = min(limit + total_swap_pages, memsw_limit);
1586 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1589 struct mem_cgroup *iter;
1590 unsigned long chosen_points = 0;
1591 unsigned long totalpages;
1592 unsigned int points = 0;
1593 struct task_struct *chosen = NULL;
1596 * If current has a pending SIGKILL or is exiting, then automatically
1597 * select it. The goal is to allow it to allocate so that it may
1598 * quickly exit and free its memory.
1600 if (fatal_signal_pending(current) || current->flags & PF_EXITING) {
1601 set_thread_flag(TIF_MEMDIE);
1605 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1606 totalpages = mem_cgroup_get_limit(memcg) ? : 1;
1607 for_each_mem_cgroup_tree(iter, memcg) {
1608 struct css_task_iter it;
1609 struct task_struct *task;
1611 css_task_iter_start(&iter->css, &it);
1612 while ((task = css_task_iter_next(&it))) {
1613 switch (oom_scan_process_thread(task, totalpages, NULL,
1615 case OOM_SCAN_SELECT:
1617 put_task_struct(chosen);
1619 chosen_points = ULONG_MAX;
1620 get_task_struct(chosen);
1622 case OOM_SCAN_CONTINUE:
1624 case OOM_SCAN_ABORT:
1625 css_task_iter_end(&it);
1626 mem_cgroup_iter_break(memcg, iter);
1628 put_task_struct(chosen);
1633 points = oom_badness(task, memcg, NULL, totalpages);
1634 if (!points || points < chosen_points)
1636 /* Prefer thread group leaders for display purposes */
1637 if (points == chosen_points &&
1638 thread_group_leader(chosen))
1642 put_task_struct(chosen);
1644 chosen_points = points;
1645 get_task_struct(chosen);
1647 css_task_iter_end(&it);
1652 points = chosen_points * 1000 / totalpages;
1653 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1654 NULL, "Memory cgroup out of memory");
1658 * test_mem_cgroup_node_reclaimable
1659 * @memcg: the target memcg
1660 * @nid: the node ID to be checked.
1661 * @noswap : specify true here if the user wants flle only information.
1663 * This function returns whether the specified memcg contains any
1664 * reclaimable pages on a node. Returns true if there are any reclaimable
1665 * pages in the node.
1667 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1668 int nid, bool noswap)
1670 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1672 if (noswap || !total_swap_pages)
1674 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1679 #if MAX_NUMNODES > 1
1682 * Always updating the nodemask is not very good - even if we have an empty
1683 * list or the wrong list here, we can start from some node and traverse all
1684 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1687 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1691 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1692 * pagein/pageout changes since the last update.
1694 if (!atomic_read(&memcg->numainfo_events))
1696 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1699 /* make a nodemask where this memcg uses memory from */
1700 memcg->scan_nodes = node_states[N_MEMORY];
1702 for_each_node_mask(nid, node_states[N_MEMORY]) {
1704 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1705 node_clear(nid, memcg->scan_nodes);
1708 atomic_set(&memcg->numainfo_events, 0);
1709 atomic_set(&memcg->numainfo_updating, 0);
1713 * Selecting a node where we start reclaim from. Because what we need is just
1714 * reducing usage counter, start from anywhere is O,K. Considering
1715 * memory reclaim from current node, there are pros. and cons.
1717 * Freeing memory from current node means freeing memory from a node which
1718 * we'll use or we've used. So, it may make LRU bad. And if several threads
1719 * hit limits, it will see a contention on a node. But freeing from remote
1720 * node means more costs for memory reclaim because of memory latency.
1722 * Now, we use round-robin. Better algorithm is welcomed.
1724 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1728 mem_cgroup_may_update_nodemask(memcg);
1729 node = memcg->last_scanned_node;
1731 node = next_node(node, memcg->scan_nodes);
1732 if (node == MAX_NUMNODES)
1733 node = first_node(memcg->scan_nodes);
1735 * We call this when we hit limit, not when pages are added to LRU.
1736 * No LRU may hold pages because all pages are UNEVICTABLE or
1737 * memcg is too small and all pages are not on LRU. In that case,
1738 * we use curret node.
1740 if (unlikely(node == MAX_NUMNODES))
1741 node = numa_node_id();
1743 memcg->last_scanned_node = node;
1748 * Check all nodes whether it contains reclaimable pages or not.
1749 * For quick scan, we make use of scan_nodes. This will allow us to skip
1750 * unused nodes. But scan_nodes is lazily updated and may not cotain
1751 * enough new information. We need to do double check.
1753 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1758 * quick check...making use of scan_node.
1759 * We can skip unused nodes.
1761 if (!nodes_empty(memcg->scan_nodes)) {
1762 for (nid = first_node(memcg->scan_nodes);
1764 nid = next_node(nid, memcg->scan_nodes)) {
1766 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1771 * Check rest of nodes.
1773 for_each_node_state(nid, N_MEMORY) {
1774 if (node_isset(nid, memcg->scan_nodes))
1776 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1783 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1788 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1790 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
1794 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1797 unsigned long *total_scanned)
1799 struct mem_cgroup *victim = NULL;
1802 unsigned long excess;
1803 unsigned long nr_scanned;
1804 struct mem_cgroup_reclaim_cookie reclaim = {
1809 excess = soft_limit_excess(root_memcg);
1812 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1817 * If we have not been able to reclaim
1818 * anything, it might because there are
1819 * no reclaimable pages under this hierarchy
1824 * We want to do more targeted reclaim.
1825 * excess >> 2 is not to excessive so as to
1826 * reclaim too much, nor too less that we keep
1827 * coming back to reclaim from this cgroup
1829 if (total >= (excess >> 2) ||
1830 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1835 if (!mem_cgroup_reclaimable(victim, false))
1837 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1839 *total_scanned += nr_scanned;
1840 if (!soft_limit_excess(root_memcg))
1843 mem_cgroup_iter_break(root_memcg, victim);
1847 #ifdef CONFIG_LOCKDEP
1848 static struct lockdep_map memcg_oom_lock_dep_map = {
1849 .name = "memcg_oom_lock",
1853 static DEFINE_SPINLOCK(memcg_oom_lock);
1856 * Check OOM-Killer is already running under our hierarchy.
1857 * If someone is running, return false.
1859 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1861 struct mem_cgroup *iter, *failed = NULL;
1863 spin_lock(&memcg_oom_lock);
1865 for_each_mem_cgroup_tree(iter, memcg) {
1866 if (iter->oom_lock) {
1868 * this subtree of our hierarchy is already locked
1869 * so we cannot give a lock.
1872 mem_cgroup_iter_break(memcg, iter);
1875 iter->oom_lock = true;
1880 * OK, we failed to lock the whole subtree so we have
1881 * to clean up what we set up to the failing subtree
1883 for_each_mem_cgroup_tree(iter, memcg) {
1884 if (iter == failed) {
1885 mem_cgroup_iter_break(memcg, iter);
1888 iter->oom_lock = false;
1891 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1893 spin_unlock(&memcg_oom_lock);
1898 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1900 struct mem_cgroup *iter;
1902 spin_lock(&memcg_oom_lock);
1903 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1904 for_each_mem_cgroup_tree(iter, memcg)
1905 iter->oom_lock = false;
1906 spin_unlock(&memcg_oom_lock);
1909 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1911 struct mem_cgroup *iter;
1913 for_each_mem_cgroup_tree(iter, memcg)
1914 atomic_inc(&iter->under_oom);
1917 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1919 struct mem_cgroup *iter;
1922 * When a new child is created while the hierarchy is under oom,
1923 * mem_cgroup_oom_lock() may not be called. We have to use
1924 * atomic_add_unless() here.
1926 for_each_mem_cgroup_tree(iter, memcg)
1927 atomic_add_unless(&iter->under_oom, -1, 0);
1930 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1932 struct oom_wait_info {
1933 struct mem_cgroup *memcg;
1937 static int memcg_oom_wake_function(wait_queue_t *wait,
1938 unsigned mode, int sync, void *arg)
1940 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1941 struct mem_cgroup *oom_wait_memcg;
1942 struct oom_wait_info *oom_wait_info;
1944 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1945 oom_wait_memcg = oom_wait_info->memcg;
1948 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
1949 * Then we can use css_is_ancestor without taking care of RCU.
1951 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
1952 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
1954 return autoremove_wake_function(wait, mode, sync, arg);
1957 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
1959 atomic_inc(&memcg->oom_wakeups);
1960 /* for filtering, pass "memcg" as argument. */
1961 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1964 static void memcg_oom_recover(struct mem_cgroup *memcg)
1966 if (memcg && atomic_read(&memcg->under_oom))
1967 memcg_wakeup_oom(memcg);
1970 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1972 if (!current->memcg_oom.may_oom)
1975 * We are in the middle of the charge context here, so we
1976 * don't want to block when potentially sitting on a callstack
1977 * that holds all kinds of filesystem and mm locks.
1979 * Also, the caller may handle a failed allocation gracefully
1980 * (like optional page cache readahead) and so an OOM killer
1981 * invocation might not even be necessary.
1983 * That's why we don't do anything here except remember the
1984 * OOM context and then deal with it at the end of the page
1985 * fault when the stack is unwound, the locks are released,
1986 * and when we know whether the fault was overall successful.
1988 css_get(&memcg->css);
1989 current->memcg_oom.memcg = memcg;
1990 current->memcg_oom.gfp_mask = mask;
1991 current->memcg_oom.order = order;
1995 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1996 * @handle: actually kill/wait or just clean up the OOM state
1998 * This has to be called at the end of a page fault if the memcg OOM
1999 * handler was enabled.
2001 * Memcg supports userspace OOM handling where failed allocations must
2002 * sleep on a waitqueue until the userspace task resolves the
2003 * situation. Sleeping directly in the charge context with all kinds
2004 * of locks held is not a good idea, instead we remember an OOM state
2005 * in the task and mem_cgroup_oom_synchronize() has to be called at
2006 * the end of the page fault to complete the OOM handling.
2008 * Returns %true if an ongoing memcg OOM situation was detected and
2009 * completed, %false otherwise.
2011 bool mem_cgroup_oom_synchronize(bool handle)
2013 struct mem_cgroup *memcg = current->memcg_oom.memcg;
2014 struct oom_wait_info owait;
2017 /* OOM is global, do not handle */
2024 owait.memcg = memcg;
2025 owait.wait.flags = 0;
2026 owait.wait.func = memcg_oom_wake_function;
2027 owait.wait.private = current;
2028 INIT_LIST_HEAD(&owait.wait.task_list);
2030 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2031 mem_cgroup_mark_under_oom(memcg);
2033 locked = mem_cgroup_oom_trylock(memcg);
2036 mem_cgroup_oom_notify(memcg);
2038 if (locked && !memcg->oom_kill_disable) {
2039 mem_cgroup_unmark_under_oom(memcg);
2040 finish_wait(&memcg_oom_waitq, &owait.wait);
2041 mem_cgroup_out_of_memory(memcg, current->memcg_oom.gfp_mask,
2042 current->memcg_oom.order);
2045 mem_cgroup_unmark_under_oom(memcg);
2046 finish_wait(&memcg_oom_waitq, &owait.wait);
2050 mem_cgroup_oom_unlock(memcg);
2052 * There is no guarantee that an OOM-lock contender
2053 * sees the wakeups triggered by the OOM kill
2054 * uncharges. Wake any sleepers explicitely.
2056 memcg_oom_recover(memcg);
2059 current->memcg_oom.memcg = NULL;
2060 css_put(&memcg->css);
2065 * mem_cgroup_begin_page_stat - begin a page state statistics transaction
2066 * @page: page that is going to change accounted state
2067 * @locked: &memcg->move_lock slowpath was taken
2068 * @flags: IRQ-state flags for &memcg->move_lock
2070 * This function must mark the beginning of an accounted page state
2071 * change to prevent double accounting when the page is concurrently
2072 * being moved to another memcg:
2074 * memcg = mem_cgroup_begin_page_stat(page, &locked, &flags);
2075 * if (TestClearPageState(page))
2076 * mem_cgroup_update_page_stat(memcg, state, -1);
2077 * mem_cgroup_end_page_stat(memcg, locked, flags);
2079 * The RCU lock is held throughout the transaction. The fast path can
2080 * get away without acquiring the memcg->move_lock (@locked is false)
2081 * because page moving starts with an RCU grace period.
2083 * The RCU lock also protects the memcg from being freed when the page
2084 * state that is going to change is the only thing preventing the page
2085 * from being uncharged. E.g. end-writeback clearing PageWriteback(),
2086 * which allows migration to go ahead and uncharge the page before the
2087 * account transaction might be complete.
2089 struct mem_cgroup *mem_cgroup_begin_page_stat(struct page *page,
2091 unsigned long *flags)
2093 struct mem_cgroup *memcg;
2094 struct page_cgroup *pc;
2098 if (mem_cgroup_disabled())
2101 pc = lookup_page_cgroup(page);
2103 memcg = pc->mem_cgroup;
2104 if (unlikely(!memcg))
2108 if (atomic_read(&memcg->moving_account) <= 0)
2111 spin_lock_irqsave(&memcg->move_lock, *flags);
2112 if (memcg != pc->mem_cgroup) {
2113 spin_unlock_irqrestore(&memcg->move_lock, *flags);
2122 * mem_cgroup_end_page_stat - finish a page state statistics transaction
2123 * @memcg: the memcg that was accounted against
2124 * @locked: value received from mem_cgroup_begin_page_stat()
2125 * @flags: value received from mem_cgroup_begin_page_stat()
2127 void mem_cgroup_end_page_stat(struct mem_cgroup *memcg, bool locked,
2128 unsigned long flags)
2130 if (memcg && locked)
2131 spin_unlock_irqrestore(&memcg->move_lock, flags);
2137 * mem_cgroup_update_page_stat - update page state statistics
2138 * @memcg: memcg to account against
2139 * @idx: page state item to account
2140 * @val: number of pages (positive or negative)
2142 * See mem_cgroup_begin_page_stat() for locking requirements.
2144 void mem_cgroup_update_page_stat(struct mem_cgroup *memcg,
2145 enum mem_cgroup_stat_index idx, int val)
2147 VM_BUG_ON(!rcu_read_lock_held());
2150 this_cpu_add(memcg->stat->count[idx], val);
2154 * size of first charge trial. "32" comes from vmscan.c's magic value.
2155 * TODO: maybe necessary to use big numbers in big irons.
2157 #define CHARGE_BATCH 32U
2158 struct memcg_stock_pcp {
2159 struct mem_cgroup *cached; /* this never be root cgroup */
2160 unsigned int nr_pages;
2161 struct work_struct work;
2162 unsigned long flags;
2163 #define FLUSHING_CACHED_CHARGE 0
2165 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2166 static DEFINE_MUTEX(percpu_charge_mutex);
2169 * consume_stock: Try to consume stocked charge on this cpu.
2170 * @memcg: memcg to consume from.
2171 * @nr_pages: how many pages to charge.
2173 * The charges will only happen if @memcg matches the current cpu's memcg
2174 * stock, and at least @nr_pages are available in that stock. Failure to
2175 * service an allocation will refill the stock.
2177 * returns true if successful, false otherwise.
2179 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2181 struct memcg_stock_pcp *stock;
2184 if (nr_pages > CHARGE_BATCH)
2187 stock = &get_cpu_var(memcg_stock);
2188 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2189 stock->nr_pages -= nr_pages;
2192 put_cpu_var(memcg_stock);
2197 * Returns stocks cached in percpu and reset cached information.
2199 static void drain_stock(struct memcg_stock_pcp *stock)
2201 struct mem_cgroup *old = stock->cached;
2203 if (stock->nr_pages) {
2204 page_counter_uncharge(&old->memory, stock->nr_pages);
2205 if (do_swap_account)
2206 page_counter_uncharge(&old->memsw, stock->nr_pages);
2207 css_put_many(&old->css, stock->nr_pages);
2208 stock->nr_pages = 0;
2210 stock->cached = NULL;
2214 * This must be called under preempt disabled or must be called by
2215 * a thread which is pinned to local cpu.
2217 static void drain_local_stock(struct work_struct *dummy)
2219 struct memcg_stock_pcp *stock = this_cpu_ptr(&memcg_stock);
2221 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2224 static void __init memcg_stock_init(void)
2228 for_each_possible_cpu(cpu) {
2229 struct memcg_stock_pcp *stock =
2230 &per_cpu(memcg_stock, cpu);
2231 INIT_WORK(&stock->work, drain_local_stock);
2236 * Cache charges(val) to local per_cpu area.
2237 * This will be consumed by consume_stock() function, later.
2239 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2241 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2243 if (stock->cached != memcg) { /* reset if necessary */
2245 stock->cached = memcg;
2247 stock->nr_pages += nr_pages;
2248 put_cpu_var(memcg_stock);
2252 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2253 * of the hierarchy under it.
2255 static void drain_all_stock(struct mem_cgroup *root_memcg)
2259 /* If someone's already draining, avoid adding running more workers. */
2260 if (!mutex_trylock(&percpu_charge_mutex))
2262 /* Notify other cpus that system-wide "drain" is running */
2265 for_each_online_cpu(cpu) {
2266 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2267 struct mem_cgroup *memcg;
2269 memcg = stock->cached;
2270 if (!memcg || !stock->nr_pages)
2272 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2274 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2276 drain_local_stock(&stock->work);
2278 schedule_work_on(cpu, &stock->work);
2283 mutex_unlock(&percpu_charge_mutex);
2287 * This function drains percpu counter value from DEAD cpu and
2288 * move it to local cpu. Note that this function can be preempted.
2290 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2294 spin_lock(&memcg->pcp_counter_lock);
2295 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2296 long x = per_cpu(memcg->stat->count[i], cpu);
2298 per_cpu(memcg->stat->count[i], cpu) = 0;
2299 memcg->nocpu_base.count[i] += x;
2301 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2302 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2304 per_cpu(memcg->stat->events[i], cpu) = 0;
2305 memcg->nocpu_base.events[i] += x;
2307 spin_unlock(&memcg->pcp_counter_lock);
2310 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
2311 unsigned long action,
2314 int cpu = (unsigned long)hcpu;
2315 struct memcg_stock_pcp *stock;
2316 struct mem_cgroup *iter;
2318 if (action == CPU_ONLINE)
2321 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2324 for_each_mem_cgroup(iter)
2325 mem_cgroup_drain_pcp_counter(iter, cpu);
2327 stock = &per_cpu(memcg_stock, cpu);
2332 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2333 unsigned int nr_pages)
2335 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2336 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2337 struct mem_cgroup *mem_over_limit;
2338 struct page_counter *counter;
2339 unsigned long nr_reclaimed;
2340 bool may_swap = true;
2341 bool drained = false;
2344 if (mem_cgroup_is_root(memcg))
2347 if (consume_stock(memcg, nr_pages))
2350 if (!do_swap_account ||
2351 !page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2352 if (!page_counter_try_charge(&memcg->memory, batch, &counter))
2354 if (do_swap_account)
2355 page_counter_uncharge(&memcg->memsw, batch);
2356 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2358 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2362 if (batch > nr_pages) {
2368 * Unlike in global OOM situations, memcg is not in a physical
2369 * memory shortage. Allow dying and OOM-killed tasks to
2370 * bypass the last charges so that they can exit quickly and
2371 * free their memory.
2373 if (unlikely(test_thread_flag(TIF_MEMDIE) ||
2374 fatal_signal_pending(current) ||
2375 current->flags & PF_EXITING))
2378 if (unlikely(task_in_memcg_oom(current)))
2381 if (!(gfp_mask & __GFP_WAIT))
2384 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2385 gfp_mask, may_swap);
2387 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2391 drain_all_stock(mem_over_limit);
2396 if (gfp_mask & __GFP_NORETRY)
2399 * Even though the limit is exceeded at this point, reclaim
2400 * may have been able to free some pages. Retry the charge
2401 * before killing the task.
2403 * Only for regular pages, though: huge pages are rather
2404 * unlikely to succeed so close to the limit, and we fall back
2405 * to regular pages anyway in case of failure.
2407 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2410 * At task move, charge accounts can be doubly counted. So, it's
2411 * better to wait until the end of task_move if something is going on.
2413 if (mem_cgroup_wait_acct_move(mem_over_limit))
2419 if (gfp_mask & __GFP_NOFAIL)
2422 if (fatal_signal_pending(current))
2425 mem_cgroup_oom(mem_over_limit, gfp_mask, get_order(nr_pages));
2427 if (!(gfp_mask & __GFP_NOFAIL))
2433 css_get_many(&memcg->css, batch);
2434 if (batch > nr_pages)
2435 refill_stock(memcg, batch - nr_pages);
2440 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2442 if (mem_cgroup_is_root(memcg))
2445 page_counter_uncharge(&memcg->memory, nr_pages);
2446 if (do_swap_account)
2447 page_counter_uncharge(&memcg->memsw, nr_pages);
2449 css_put_many(&memcg->css, nr_pages);
2453 * A helper function to get mem_cgroup from ID. must be called under
2454 * rcu_read_lock(). The caller is responsible for calling
2455 * css_tryget_online() if the mem_cgroup is used for charging. (dropping
2456 * refcnt from swap can be called against removed memcg.)
2458 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2460 /* ID 0 is unused ID */
2463 return mem_cgroup_from_id(id);
2467 * try_get_mem_cgroup_from_page - look up page's memcg association
2470 * Look up, get a css reference, and return the memcg that owns @page.
2472 * The page must be locked to prevent racing with swap-in and page
2473 * cache charges. If coming from an unlocked page table, the caller
2474 * must ensure the page is on the LRU or this can race with charging.
2476 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2478 struct mem_cgroup *memcg;
2479 struct page_cgroup *pc;
2483 VM_BUG_ON_PAGE(!PageLocked(page), page);
2485 pc = lookup_page_cgroup(page);
2486 memcg = pc->mem_cgroup;
2489 if (!css_tryget_online(&memcg->css))
2491 } else if (PageSwapCache(page)) {
2492 ent.val = page_private(page);
2493 id = lookup_swap_cgroup_id(ent);
2495 memcg = mem_cgroup_lookup(id);
2496 if (memcg && !css_tryget_online(&memcg->css))
2503 static void lock_page_lru(struct page *page, int *isolated)
2505 struct zone *zone = page_zone(page);
2507 spin_lock_irq(&zone->lru_lock);
2508 if (PageLRU(page)) {
2509 struct lruvec *lruvec;
2511 lruvec = mem_cgroup_page_lruvec(page, zone);
2513 del_page_from_lru_list(page, lruvec, page_lru(page));
2519 static void unlock_page_lru(struct page *page, int isolated)
2521 struct zone *zone = page_zone(page);
2524 struct lruvec *lruvec;
2526 lruvec = mem_cgroup_page_lruvec(page, zone);
2527 VM_BUG_ON_PAGE(PageLRU(page), page);
2529 add_page_to_lru_list(page, lruvec, page_lru(page));
2531 spin_unlock_irq(&zone->lru_lock);
2534 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2537 struct page_cgroup *pc = lookup_page_cgroup(page);
2540 VM_BUG_ON_PAGE(pc->mem_cgroup, page);
2542 * we don't need page_cgroup_lock about tail pages, becase they are not
2543 * accessed by any other context at this point.
2547 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2548 * may already be on some other mem_cgroup's LRU. Take care of it.
2551 lock_page_lru(page, &isolated);
2554 * Nobody should be changing or seriously looking at
2555 * pc->mem_cgroup at this point:
2557 * - the page is uncharged
2559 * - the page is off-LRU
2561 * - an anonymous fault has exclusive page access, except for
2562 * a locked page table
2564 * - a page cache insertion, a swapin fault, or a migration
2565 * have the page locked
2567 pc->mem_cgroup = memcg;
2570 unlock_page_lru(page, isolated);
2573 #ifdef CONFIG_MEMCG_KMEM
2575 * The memcg_slab_mutex is held whenever a per memcg kmem cache is created or
2576 * destroyed. It protects memcg_caches arrays and memcg_slab_caches lists.
2578 static DEFINE_MUTEX(memcg_slab_mutex);
2581 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2582 * in the memcg_cache_params struct.
2584 static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
2586 struct kmem_cache *cachep;
2588 VM_BUG_ON(p->is_root_cache);
2589 cachep = p->root_cache;
2590 return cache_from_memcg_idx(cachep, memcg_cache_id(p->memcg));
2593 #ifdef CONFIG_SLABINFO
2594 static int mem_cgroup_slabinfo_read(struct seq_file *m, void *v)
2596 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
2597 struct memcg_cache_params *params;
2599 if (!memcg_kmem_is_active(memcg))
2602 print_slabinfo_header(m);
2604 mutex_lock(&memcg_slab_mutex);
2605 list_for_each_entry(params, &memcg->memcg_slab_caches, list)
2606 cache_show(memcg_params_to_cache(params), m);
2607 mutex_unlock(&memcg_slab_mutex);
2613 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp,
2614 unsigned long nr_pages)
2616 struct page_counter *counter;
2619 ret = page_counter_try_charge(&memcg->kmem, nr_pages, &counter);
2623 ret = try_charge(memcg, gfp, nr_pages);
2624 if (ret == -EINTR) {
2626 * try_charge() chose to bypass to root due to OOM kill or
2627 * fatal signal. Since our only options are to either fail
2628 * the allocation or charge it to this cgroup, do it as a
2629 * temporary condition. But we can't fail. From a kmem/slab
2630 * perspective, the cache has already been selected, by
2631 * mem_cgroup_kmem_get_cache(), so it is too late to change
2634 * This condition will only trigger if the task entered
2635 * memcg_charge_kmem in a sane state, but was OOM-killed
2636 * during try_charge() above. Tasks that were already dying
2637 * when the allocation triggers should have been already
2638 * directed to the root cgroup in memcontrol.h
2640 page_counter_charge(&memcg->memory, nr_pages);
2641 if (do_swap_account)
2642 page_counter_charge(&memcg->memsw, nr_pages);
2643 css_get_many(&memcg->css, nr_pages);
2646 page_counter_uncharge(&memcg->kmem, nr_pages);
2651 static void memcg_uncharge_kmem(struct mem_cgroup *memcg,
2652 unsigned long nr_pages)
2654 page_counter_uncharge(&memcg->memory, nr_pages);
2655 if (do_swap_account)
2656 page_counter_uncharge(&memcg->memsw, nr_pages);
2658 page_counter_uncharge(&memcg->kmem, nr_pages);
2660 css_put_many(&memcg->css, nr_pages);
2664 * helper for acessing a memcg's index. It will be used as an index in the
2665 * child cache array in kmem_cache, and also to derive its name. This function
2666 * will return -1 when this is not a kmem-limited memcg.
2668 int memcg_cache_id(struct mem_cgroup *memcg)
2670 return memcg ? memcg->kmemcg_id : -1;
2673 static int memcg_alloc_cache_id(void)
2678 id = ida_simple_get(&kmem_limited_groups,
2679 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2683 if (id < memcg_limited_groups_array_size)
2687 * There's no space for the new id in memcg_caches arrays,
2688 * so we have to grow them.
2691 size = 2 * (id + 1);
2692 if (size < MEMCG_CACHES_MIN_SIZE)
2693 size = MEMCG_CACHES_MIN_SIZE;
2694 else if (size > MEMCG_CACHES_MAX_SIZE)
2695 size = MEMCG_CACHES_MAX_SIZE;
2697 mutex_lock(&memcg_slab_mutex);
2698 err = memcg_update_all_caches(size);
2699 mutex_unlock(&memcg_slab_mutex);
2702 ida_simple_remove(&kmem_limited_groups, id);
2708 static void memcg_free_cache_id(int id)
2710 ida_simple_remove(&kmem_limited_groups, id);
2714 * We should update the current array size iff all caches updates succeed. This
2715 * can only be done from the slab side. The slab mutex needs to be held when
2718 void memcg_update_array_size(int num)
2720 memcg_limited_groups_array_size = num;
2723 static void memcg_register_cache(struct mem_cgroup *memcg,
2724 struct kmem_cache *root_cache)
2726 static char memcg_name_buf[NAME_MAX + 1]; /* protected by
2728 struct kmem_cache *cachep;
2731 lockdep_assert_held(&memcg_slab_mutex);
2733 id = memcg_cache_id(memcg);
2736 * Since per-memcg caches are created asynchronously on first
2737 * allocation (see memcg_kmem_get_cache()), several threads can try to
2738 * create the same cache, but only one of them may succeed.
2740 if (cache_from_memcg_idx(root_cache, id))
2743 cgroup_name(memcg->css.cgroup, memcg_name_buf, NAME_MAX + 1);
2744 cachep = memcg_create_kmem_cache(memcg, root_cache, memcg_name_buf);
2746 * If we could not create a memcg cache, do not complain, because
2747 * that's not critical at all as we can always proceed with the root
2753 css_get(&memcg->css);
2754 list_add(&cachep->memcg_params->list, &memcg->memcg_slab_caches);
2757 * Since readers won't lock (see cache_from_memcg_idx()), we need a
2758 * barrier here to ensure nobody will see the kmem_cache partially
2763 BUG_ON(root_cache->memcg_params->memcg_caches[id]);
2764 root_cache->memcg_params->memcg_caches[id] = cachep;
2767 static void memcg_unregister_cache(struct kmem_cache *cachep)
2769 struct kmem_cache *root_cache;
2770 struct mem_cgroup *memcg;
2773 lockdep_assert_held(&memcg_slab_mutex);
2775 BUG_ON(is_root_cache(cachep));
2777 root_cache = cachep->memcg_params->root_cache;
2778 memcg = cachep->memcg_params->memcg;
2779 id = memcg_cache_id(memcg);
2781 BUG_ON(root_cache->memcg_params->memcg_caches[id] != cachep);
2782 root_cache->memcg_params->memcg_caches[id] = NULL;
2784 list_del(&cachep->memcg_params->list);
2786 kmem_cache_destroy(cachep);
2788 /* drop the reference taken in memcg_register_cache */
2789 css_put(&memcg->css);
2793 * During the creation a new cache, we need to disable our accounting mechanism
2794 * altogether. This is true even if we are not creating, but rather just
2795 * enqueing new caches to be created.
2797 * This is because that process will trigger allocations; some visible, like
2798 * explicit kmallocs to auxiliary data structures, name strings and internal
2799 * cache structures; some well concealed, like INIT_WORK() that can allocate
2800 * objects during debug.
2802 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
2803 * to it. This may not be a bounded recursion: since the first cache creation
2804 * failed to complete (waiting on the allocation), we'll just try to create the
2805 * cache again, failing at the same point.
2807 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
2808 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
2809 * inside the following two functions.
2811 static inline void memcg_stop_kmem_account(void)
2813 VM_BUG_ON(!current->mm);
2814 current->memcg_kmem_skip_account++;
2817 static inline void memcg_resume_kmem_account(void)
2819 VM_BUG_ON(!current->mm);
2820 current->memcg_kmem_skip_account--;
2823 int __memcg_cleanup_cache_params(struct kmem_cache *s)
2825 struct kmem_cache *c;
2828 mutex_lock(&memcg_slab_mutex);
2829 for_each_memcg_cache_index(i) {
2830 c = cache_from_memcg_idx(s, i);
2834 memcg_unregister_cache(c);
2836 if (cache_from_memcg_idx(s, i))
2839 mutex_unlock(&memcg_slab_mutex);
2843 static void memcg_unregister_all_caches(struct mem_cgroup *memcg)
2845 struct kmem_cache *cachep;
2846 struct memcg_cache_params *params, *tmp;
2848 if (!memcg_kmem_is_active(memcg))
2851 mutex_lock(&memcg_slab_mutex);
2852 list_for_each_entry_safe(params, tmp, &memcg->memcg_slab_caches, list) {
2853 cachep = memcg_params_to_cache(params);
2854 kmem_cache_shrink(cachep);
2855 if (atomic_read(&cachep->memcg_params->nr_pages) == 0)
2856 memcg_unregister_cache(cachep);
2858 mutex_unlock(&memcg_slab_mutex);
2861 struct memcg_register_cache_work {
2862 struct mem_cgroup *memcg;
2863 struct kmem_cache *cachep;
2864 struct work_struct work;
2867 static void memcg_register_cache_func(struct work_struct *w)
2869 struct memcg_register_cache_work *cw =
2870 container_of(w, struct memcg_register_cache_work, work);
2871 struct mem_cgroup *memcg = cw->memcg;
2872 struct kmem_cache *cachep = cw->cachep;
2874 mutex_lock(&memcg_slab_mutex);
2875 memcg_register_cache(memcg, cachep);
2876 mutex_unlock(&memcg_slab_mutex);
2878 css_put(&memcg->css);
2883 * Enqueue the creation of a per-memcg kmem_cache.
2885 static void __memcg_schedule_register_cache(struct mem_cgroup *memcg,
2886 struct kmem_cache *cachep)
2888 struct memcg_register_cache_work *cw;
2890 cw = kmalloc(sizeof(*cw), GFP_NOWAIT);
2892 css_put(&memcg->css);
2897 cw->cachep = cachep;
2899 INIT_WORK(&cw->work, memcg_register_cache_func);
2900 schedule_work(&cw->work);
2903 static void memcg_schedule_register_cache(struct mem_cgroup *memcg,
2904 struct kmem_cache *cachep)
2907 * We need to stop accounting when we kmalloc, because if the
2908 * corresponding kmalloc cache is not yet created, the first allocation
2909 * in __memcg_schedule_register_cache will recurse.
2911 * However, it is better to enclose the whole function. Depending on
2912 * the debugging options enabled, INIT_WORK(), for instance, can
2913 * trigger an allocation. This too, will make us recurse. Because at
2914 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2915 * the safest choice is to do it like this, wrapping the whole function.
2917 memcg_stop_kmem_account();
2918 __memcg_schedule_register_cache(memcg, cachep);
2919 memcg_resume_kmem_account();
2922 int __memcg_charge_slab(struct kmem_cache *cachep, gfp_t gfp, int order)
2924 unsigned int nr_pages = 1 << order;
2927 res = memcg_charge_kmem(cachep->memcg_params->memcg, gfp, nr_pages);
2929 atomic_add(nr_pages, &cachep->memcg_params->nr_pages);
2933 void __memcg_uncharge_slab(struct kmem_cache *cachep, int order)
2935 unsigned int nr_pages = 1 << order;
2937 memcg_uncharge_kmem(cachep->memcg_params->memcg, nr_pages);
2938 atomic_sub(nr_pages, &cachep->memcg_params->nr_pages);
2942 * Return the kmem_cache we're supposed to use for a slab allocation.
2943 * We try to use the current memcg's version of the cache.
2945 * If the cache does not exist yet, if we are the first user of it,
2946 * we either create it immediately, if possible, or create it asynchronously
2948 * In the latter case, we will let the current allocation go through with
2949 * the original cache.
2951 * Can't be called in interrupt context or from kernel threads.
2952 * This function needs to be called with rcu_read_lock() held.
2954 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
2957 struct mem_cgroup *memcg;
2958 struct kmem_cache *memcg_cachep;
2960 VM_BUG_ON(!cachep->memcg_params);
2961 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
2963 if (!current->mm || current->memcg_kmem_skip_account)
2967 memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
2969 if (!memcg_kmem_is_active(memcg))
2972 memcg_cachep = cache_from_memcg_idx(cachep, memcg_cache_id(memcg));
2973 if (likely(memcg_cachep)) {
2974 cachep = memcg_cachep;
2978 /* The corresponding put will be done in the workqueue. */
2979 if (!css_tryget_online(&memcg->css))
2984 * If we are in a safe context (can wait, and not in interrupt
2985 * context), we could be be predictable and return right away.
2986 * This would guarantee that the allocation being performed
2987 * already belongs in the new cache.
2989 * However, there are some clashes that can arrive from locking.
2990 * For instance, because we acquire the slab_mutex while doing
2991 * memcg_create_kmem_cache, this means no further allocation
2992 * could happen with the slab_mutex held. So it's better to
2995 memcg_schedule_register_cache(memcg, cachep);
3003 * We need to verify if the allocation against current->mm->owner's memcg is
3004 * possible for the given order. But the page is not allocated yet, so we'll
3005 * need a further commit step to do the final arrangements.
3007 * It is possible for the task to switch cgroups in this mean time, so at
3008 * commit time, we can't rely on task conversion any longer. We'll then use
3009 * the handle argument to return to the caller which cgroup we should commit
3010 * against. We could also return the memcg directly and avoid the pointer
3011 * passing, but a boolean return value gives better semantics considering
3012 * the compiled-out case as well.
3014 * Returning true means the allocation is possible.
3017 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3019 struct mem_cgroup *memcg;
3025 * Disabling accounting is only relevant for some specific memcg
3026 * internal allocations. Therefore we would initially not have such
3027 * check here, since direct calls to the page allocator that are
3028 * accounted to kmemcg (alloc_kmem_pages and friends) only happen
3029 * outside memcg core. We are mostly concerned with cache allocations,
3030 * and by having this test at memcg_kmem_get_cache, we are already able
3031 * to relay the allocation to the root cache and bypass the memcg cache
3034 * There is one exception, though: the SLUB allocator does not create
3035 * large order caches, but rather service large kmallocs directly from
3036 * the page allocator. Therefore, the following sequence when backed by
3037 * the SLUB allocator:
3039 * memcg_stop_kmem_account();
3040 * kmalloc(<large_number>)
3041 * memcg_resume_kmem_account();
3043 * would effectively ignore the fact that we should skip accounting,
3044 * since it will drive us directly to this function without passing
3045 * through the cache selector memcg_kmem_get_cache. Such large
3046 * allocations are extremely rare but can happen, for instance, for the
3047 * cache arrays. We bring this test here.
3049 if (!current->mm || current->memcg_kmem_skip_account)
3052 memcg = get_mem_cgroup_from_mm(current->mm);
3054 if (!memcg_kmem_is_active(memcg)) {
3055 css_put(&memcg->css);
3059 ret = memcg_charge_kmem(memcg, gfp, 1 << order);
3063 css_put(&memcg->css);
3067 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3070 struct page_cgroup *pc;
3072 VM_BUG_ON(mem_cgroup_is_root(memcg));
3074 /* The page allocation failed. Revert */
3076 memcg_uncharge_kmem(memcg, 1 << order);
3079 pc = lookup_page_cgroup(page);
3080 pc->mem_cgroup = memcg;
3083 void __memcg_kmem_uncharge_pages(struct page *page, int order)
3085 struct page_cgroup *pc = lookup_page_cgroup(page);
3086 struct mem_cgroup *memcg = pc->mem_cgroup;
3091 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
3093 memcg_uncharge_kmem(memcg, 1 << order);
3094 pc->mem_cgroup = NULL;
3097 static inline void memcg_unregister_all_caches(struct mem_cgroup *memcg)
3100 #endif /* CONFIG_MEMCG_KMEM */
3102 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3105 * Because tail pages are not marked as "used", set it. We're under
3106 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3107 * charge/uncharge will be never happen and move_account() is done under
3108 * compound_lock(), so we don't have to take care of races.
3110 void mem_cgroup_split_huge_fixup(struct page *head)
3112 struct page_cgroup *pc = lookup_page_cgroup(head);
3115 if (mem_cgroup_disabled())
3118 for (i = 1; i < HPAGE_PMD_NR; i++)
3119 pc[i].mem_cgroup = pc[0].mem_cgroup;
3121 __this_cpu_sub(pc[0].mem_cgroup->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
3124 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3127 * mem_cgroup_move_account - move account of the page
3129 * @nr_pages: number of regular pages (>1 for huge pages)
3130 * @pc: page_cgroup of the page.
3131 * @from: mem_cgroup which the page is moved from.
3132 * @to: mem_cgroup which the page is moved to. @from != @to.
3134 * The caller must confirm following.
3135 * - page is not on LRU (isolate_page() is useful.)
3136 * - compound_lock is held when nr_pages > 1
3138 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3141 static int mem_cgroup_move_account(struct page *page,
3142 unsigned int nr_pages,
3143 struct page_cgroup *pc,
3144 struct mem_cgroup *from,
3145 struct mem_cgroup *to)
3147 unsigned long flags;
3150 VM_BUG_ON(from == to);
3151 VM_BUG_ON_PAGE(PageLRU(page), page);
3153 * The page is isolated from LRU. So, collapse function
3154 * will not handle this page. But page splitting can happen.
3155 * Do this check under compound_page_lock(). The caller should
3159 if (nr_pages > 1 && !PageTransHuge(page))
3163 * Prevent mem_cgroup_migrate() from looking at pc->mem_cgroup
3164 * of its source page while we change it: page migration takes
3165 * both pages off the LRU, but page cache replacement doesn't.
3167 if (!trylock_page(page))
3171 if (pc->mem_cgroup != from)
3174 spin_lock_irqsave(&from->move_lock, flags);
3176 if (!PageAnon(page) && page_mapped(page)) {
3177 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
3179 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
3183 if (PageWriteback(page)) {
3184 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK],
3186 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK],
3191 * It is safe to change pc->mem_cgroup here because the page
3192 * is referenced, charged, and isolated - we can't race with
3193 * uncharging, charging, migration, or LRU putback.
3196 /* caller should have done css_get */
3197 pc->mem_cgroup = to;
3198 spin_unlock_irqrestore(&from->move_lock, flags);
3202 local_irq_disable();
3203 mem_cgroup_charge_statistics(to, page, nr_pages);
3204 memcg_check_events(to, page);
3205 mem_cgroup_charge_statistics(from, page, -nr_pages);
3206 memcg_check_events(from, page);
3214 #ifdef CONFIG_MEMCG_SWAP
3215 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
3218 int val = (charge) ? 1 : -1;
3219 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
3223 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3224 * @entry: swap entry to be moved
3225 * @from: mem_cgroup which the entry is moved from
3226 * @to: mem_cgroup which the entry is moved to
3228 * It succeeds only when the swap_cgroup's record for this entry is the same
3229 * as the mem_cgroup's id of @from.
3231 * Returns 0 on success, -EINVAL on failure.
3233 * The caller must have charged to @to, IOW, called page_counter_charge() about
3234 * both res and memsw, and called css_get().
3236 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3237 struct mem_cgroup *from, struct mem_cgroup *to)
3239 unsigned short old_id, new_id;
3241 old_id = mem_cgroup_id(from);
3242 new_id = mem_cgroup_id(to);
3244 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3245 mem_cgroup_swap_statistics(from, false);
3246 mem_cgroup_swap_statistics(to, true);
3248 * This function is only called from task migration context now.
3249 * It postpones page_counter and refcount handling till the end
3250 * of task migration(mem_cgroup_clear_mc()) for performance
3251 * improvement. But we cannot postpone css_get(to) because if
3252 * the process that has been moved to @to does swap-in, the
3253 * refcount of @to might be decreased to 0.
3255 * We are in attach() phase, so the cgroup is guaranteed to be
3256 * alive, so we can just call css_get().
3264 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3265 struct mem_cgroup *from, struct mem_cgroup *to)
3271 #ifdef CONFIG_DEBUG_VM
3272 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3274 struct page_cgroup *pc;
3276 pc = lookup_page_cgroup(page);
3278 * Can be NULL while feeding pages into the page allocator for
3279 * the first time, i.e. during boot or memory hotplug;
3280 * or when mem_cgroup_disabled().
3282 if (likely(pc) && pc->mem_cgroup)
3287 bool mem_cgroup_bad_page_check(struct page *page)
3289 if (mem_cgroup_disabled())
3292 return lookup_page_cgroup_used(page) != NULL;
3295 void mem_cgroup_print_bad_page(struct page *page)
3297 struct page_cgroup *pc;
3299 pc = lookup_page_cgroup_used(page);
3301 pr_alert("pc:%p pc->mem_cgroup:%p\n", pc, pc->mem_cgroup);
3305 static DEFINE_MUTEX(memcg_limit_mutex);
3307 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3308 unsigned long limit)
3310 unsigned long curusage;
3311 unsigned long oldusage;
3312 bool enlarge = false;
3317 * For keeping hierarchical_reclaim simple, how long we should retry
3318 * is depends on callers. We set our retry-count to be function
3319 * of # of children which we should visit in this loop.
3321 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
3322 mem_cgroup_count_children(memcg);
3324 oldusage = page_counter_read(&memcg->memory);
3327 if (signal_pending(current)) {
3332 mutex_lock(&memcg_limit_mutex);
3333 if (limit > memcg->memsw.limit) {
3334 mutex_unlock(&memcg_limit_mutex);
3338 if (limit > memcg->memory.limit)
3340 ret = page_counter_limit(&memcg->memory, limit);
3341 mutex_unlock(&memcg_limit_mutex);
3346 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, true);
3348 curusage = page_counter_read(&memcg->memory);
3349 /* Usage is reduced ? */
3350 if (curusage >= oldusage)
3353 oldusage = curusage;
3354 } while (retry_count);
3356 if (!ret && enlarge)
3357 memcg_oom_recover(memcg);
3362 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3363 unsigned long limit)
3365 unsigned long curusage;
3366 unsigned long oldusage;
3367 bool enlarge = false;
3371 /* see mem_cgroup_resize_res_limit */
3372 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
3373 mem_cgroup_count_children(memcg);
3375 oldusage = page_counter_read(&memcg->memsw);
3378 if (signal_pending(current)) {
3383 mutex_lock(&memcg_limit_mutex);
3384 if (limit < memcg->memory.limit) {
3385 mutex_unlock(&memcg_limit_mutex);
3389 if (limit > memcg->memsw.limit)
3391 ret = page_counter_limit(&memcg->memsw, limit);
3392 mutex_unlock(&memcg_limit_mutex);
3397 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, false);
3399 curusage = page_counter_read(&memcg->memsw);
3400 /* Usage is reduced ? */
3401 if (curusage >= oldusage)
3404 oldusage = curusage;
3405 } while (retry_count);
3407 if (!ret && enlarge)
3408 memcg_oom_recover(memcg);
3413 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3415 unsigned long *total_scanned)
3417 unsigned long nr_reclaimed = 0;
3418 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3419 unsigned long reclaimed;
3421 struct mem_cgroup_tree_per_zone *mctz;
3422 unsigned long excess;
3423 unsigned long nr_scanned;
3428 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3430 * This loop can run a while, specially if mem_cgroup's continuously
3431 * keep exceeding their soft limit and putting the system under
3438 mz = mem_cgroup_largest_soft_limit_node(mctz);
3443 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
3444 gfp_mask, &nr_scanned);
3445 nr_reclaimed += reclaimed;
3446 *total_scanned += nr_scanned;
3447 spin_lock_irq(&mctz->lock);
3448 __mem_cgroup_remove_exceeded(mz, mctz);
3451 * If we failed to reclaim anything from this memory cgroup
3452 * it is time to move on to the next cgroup
3456 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3458 excess = soft_limit_excess(mz->memcg);
3460 * One school of thought says that we should not add
3461 * back the node to the tree if reclaim returns 0.
3462 * But our reclaim could return 0, simply because due
3463 * to priority we are exposing a smaller subset of
3464 * memory to reclaim from. Consider this as a longer
3467 /* If excess == 0, no tree ops */
3468 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3469 spin_unlock_irq(&mctz->lock);
3470 css_put(&mz->memcg->css);
3473 * Could not reclaim anything and there are no more
3474 * mem cgroups to try or we seem to be looping without
3475 * reclaiming anything.
3477 if (!nr_reclaimed &&
3479 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3481 } while (!nr_reclaimed);
3483 css_put(&next_mz->memcg->css);
3484 return nr_reclaimed;
3488 * Test whether @memcg has children, dead or alive. Note that this
3489 * function doesn't care whether @memcg has use_hierarchy enabled and
3490 * returns %true if there are child csses according to the cgroup
3491 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
3493 static inline bool memcg_has_children(struct mem_cgroup *memcg)
3498 * The lock does not prevent addition or deletion of children, but
3499 * it prevents a new child from being initialized based on this
3500 * parent in css_online(), so it's enough to decide whether
3501 * hierarchically inherited attributes can still be changed or not.
3503 lockdep_assert_held(&memcg_create_mutex);
3506 ret = css_next_child(NULL, &memcg->css);
3512 * Reclaims as many pages from the given memcg as possible and moves
3513 * the rest to the parent.
3515 * Caller is responsible for holding css reference for memcg.
3517 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3519 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3521 /* we call try-to-free pages for make this cgroup empty */
3522 lru_add_drain_all();
3523 /* try to free all pages in this cgroup */
3524 while (nr_retries && page_counter_read(&memcg->memory)) {
3527 if (signal_pending(current))
3530 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3534 /* maybe some writeback is necessary */
3535 congestion_wait(BLK_RW_ASYNC, HZ/10);
3543 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3544 char *buf, size_t nbytes,
3547 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3549 if (mem_cgroup_is_root(memcg))
3551 return mem_cgroup_force_empty(memcg) ?: nbytes;
3554 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3557 return mem_cgroup_from_css(css)->use_hierarchy;
3560 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3561 struct cftype *cft, u64 val)
3564 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3565 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
3567 mutex_lock(&memcg_create_mutex);
3569 if (memcg->use_hierarchy == val)
3573 * If parent's use_hierarchy is set, we can't make any modifications
3574 * in the child subtrees. If it is unset, then the change can
3575 * occur, provided the current cgroup has no children.
3577 * For the root cgroup, parent_mem is NULL, we allow value to be
3578 * set if there are no children.
3580 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3581 (val == 1 || val == 0)) {
3582 if (!memcg_has_children(memcg))
3583 memcg->use_hierarchy = val;
3590 mutex_unlock(&memcg_create_mutex);
3595 static unsigned long tree_stat(struct mem_cgroup *memcg,
3596 enum mem_cgroup_stat_index idx)
3598 struct mem_cgroup *iter;
3601 /* Per-cpu values can be negative, use a signed accumulator */
3602 for_each_mem_cgroup_tree(iter, memcg)
3603 val += mem_cgroup_read_stat(iter, idx);
3605 if (val < 0) /* race ? */
3610 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3614 if (mem_cgroup_is_root(memcg)) {
3615 val = tree_stat(memcg, MEM_CGROUP_STAT_CACHE);
3616 val += tree_stat(memcg, MEM_CGROUP_STAT_RSS);
3618 val += tree_stat(memcg, MEM_CGROUP_STAT_SWAP);
3621 val = page_counter_read(&memcg->memory);
3623 val = page_counter_read(&memcg->memsw);
3625 return val << PAGE_SHIFT;
3636 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3639 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3640 struct page_counter *counter;
3642 switch (MEMFILE_TYPE(cft->private)) {
3644 counter = &memcg->memory;
3647 counter = &memcg->memsw;
3650 counter = &memcg->kmem;
3656 switch (MEMFILE_ATTR(cft->private)) {
3658 if (counter == &memcg->memory)
3659 return mem_cgroup_usage(memcg, false);
3660 if (counter == &memcg->memsw)
3661 return mem_cgroup_usage(memcg, true);
3662 return (u64)page_counter_read(counter) * PAGE_SIZE;
3664 return (u64)counter->limit * PAGE_SIZE;
3666 return (u64)counter->watermark * PAGE_SIZE;
3668 return counter->failcnt;
3669 case RES_SOFT_LIMIT:
3670 return (u64)memcg->soft_limit * PAGE_SIZE;
3676 #ifdef CONFIG_MEMCG_KMEM
3677 static int memcg_activate_kmem(struct mem_cgroup *memcg,
3678 unsigned long nr_pages)
3683 if (memcg_kmem_is_active(memcg))
3687 * We are going to allocate memory for data shared by all memory
3688 * cgroups so let's stop accounting here.
3690 memcg_stop_kmem_account();
3693 * For simplicity, we won't allow this to be disabled. It also can't
3694 * be changed if the cgroup has children already, or if tasks had
3697 * If tasks join before we set the limit, a person looking at
3698 * kmem.usage_in_bytes will have no way to determine when it took
3699 * place, which makes the value quite meaningless.
3701 * After it first became limited, changes in the value of the limit are
3702 * of course permitted.
3704 mutex_lock(&memcg_create_mutex);
3705 if (cgroup_has_tasks(memcg->css.cgroup) ||
3706 (memcg->use_hierarchy && memcg_has_children(memcg)))
3708 mutex_unlock(&memcg_create_mutex);
3712 memcg_id = memcg_alloc_cache_id();
3718 memcg->kmemcg_id = memcg_id;
3719 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
3722 * We couldn't have accounted to this cgroup, because it hasn't got the
3723 * active bit set yet, so this should succeed.
3725 err = page_counter_limit(&memcg->kmem, nr_pages);
3728 static_key_slow_inc(&memcg_kmem_enabled_key);
3730 * Setting the active bit after enabling static branching will
3731 * guarantee no one starts accounting before all call sites are
3734 memcg_kmem_set_active(memcg);
3736 memcg_resume_kmem_account();
3740 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
3741 unsigned long limit)
3745 mutex_lock(&memcg_limit_mutex);
3746 if (!memcg_kmem_is_active(memcg))
3747 ret = memcg_activate_kmem(memcg, limit);
3749 ret = page_counter_limit(&memcg->kmem, limit);
3750 mutex_unlock(&memcg_limit_mutex);
3754 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
3757 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
3762 mutex_lock(&memcg_limit_mutex);
3764 * If the parent cgroup is not kmem-active now, it cannot be activated
3765 * after this point, because it has at least one child already.
3767 if (memcg_kmem_is_active(parent))
3768 ret = memcg_activate_kmem(memcg, PAGE_COUNTER_MAX);
3769 mutex_unlock(&memcg_limit_mutex);
3773 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
3774 unsigned long limit)
3778 #endif /* CONFIG_MEMCG_KMEM */
3781 * The user of this function is...
3784 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3785 char *buf, size_t nbytes, loff_t off)
3787 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3788 unsigned long nr_pages;
3791 buf = strstrip(buf);
3792 ret = page_counter_memparse(buf, &nr_pages);
3796 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3798 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3802 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3804 ret = mem_cgroup_resize_limit(memcg, nr_pages);
3807 ret = mem_cgroup_resize_memsw_limit(memcg, nr_pages);
3810 ret = memcg_update_kmem_limit(memcg, nr_pages);
3814 case RES_SOFT_LIMIT:
3815 memcg->soft_limit = nr_pages;
3819 return ret ?: nbytes;
3822 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3823 size_t nbytes, loff_t off)
3825 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3826 struct page_counter *counter;
3828 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3830 counter = &memcg->memory;
3833 counter = &memcg->memsw;
3836 counter = &memcg->kmem;
3842 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3844 page_counter_reset_watermark(counter);
3847 counter->failcnt = 0;
3856 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3859 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3863 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3864 struct cftype *cft, u64 val)
3866 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3868 if (val >= (1 << NR_MOVE_TYPE))
3872 * No kind of locking is needed in here, because ->can_attach() will
3873 * check this value once in the beginning of the process, and then carry
3874 * on with stale data. This means that changes to this value will only
3875 * affect task migrations starting after the change.
3877 memcg->move_charge_at_immigrate = val;
3881 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3882 struct cftype *cft, u64 val)
3889 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3893 unsigned int lru_mask;
3896 static const struct numa_stat stats[] = {
3897 { "total", LRU_ALL },
3898 { "file", LRU_ALL_FILE },
3899 { "anon", LRU_ALL_ANON },
3900 { "unevictable", BIT(LRU_UNEVICTABLE) },
3902 const struct numa_stat *stat;
3905 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3907 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3908 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3909 seq_printf(m, "%s=%lu", stat->name, nr);
3910 for_each_node_state(nid, N_MEMORY) {
3911 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3913 seq_printf(m, " N%d=%lu", nid, nr);
3918 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3919 struct mem_cgroup *iter;
3922 for_each_mem_cgroup_tree(iter, memcg)
3923 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3924 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3925 for_each_node_state(nid, N_MEMORY) {
3927 for_each_mem_cgroup_tree(iter, memcg)
3928 nr += mem_cgroup_node_nr_lru_pages(
3929 iter, nid, stat->lru_mask);
3930 seq_printf(m, " N%d=%lu", nid, nr);
3937 #endif /* CONFIG_NUMA */
3939 static inline void mem_cgroup_lru_names_not_uptodate(void)
3941 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3944 static int memcg_stat_show(struct seq_file *m, void *v)
3946 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3947 unsigned long memory, memsw;
3948 struct mem_cgroup *mi;
3951 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3952 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
3954 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
3955 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
3958 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
3959 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
3960 mem_cgroup_read_events(memcg, i));
3962 for (i = 0; i < NR_LRU_LISTS; i++)
3963 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3964 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
3966 /* Hierarchical information */
3967 memory = memsw = PAGE_COUNTER_MAX;
3968 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3969 memory = min(memory, mi->memory.limit);
3970 memsw = min(memsw, mi->memsw.limit);
3972 seq_printf(m, "hierarchical_memory_limit %llu\n",
3973 (u64)memory * PAGE_SIZE);
3974 if (do_swap_account)
3975 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3976 (u64)memsw * PAGE_SIZE);
3978 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3981 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
3983 for_each_mem_cgroup_tree(mi, memcg)
3984 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
3985 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
3988 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
3989 unsigned long long val = 0;
3991 for_each_mem_cgroup_tree(mi, memcg)
3992 val += mem_cgroup_read_events(mi, i);
3993 seq_printf(m, "total_%s %llu\n",
3994 mem_cgroup_events_names[i], val);
3997 for (i = 0; i < NR_LRU_LISTS; i++) {
3998 unsigned long long val = 0;
4000 for_each_mem_cgroup_tree(mi, memcg)
4001 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
4002 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
4005 #ifdef CONFIG_DEBUG_VM
4008 struct mem_cgroup_per_zone *mz;
4009 struct zone_reclaim_stat *rstat;
4010 unsigned long recent_rotated[2] = {0, 0};
4011 unsigned long recent_scanned[2] = {0, 0};
4013 for_each_online_node(nid)
4014 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4015 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
4016 rstat = &mz->lruvec.reclaim_stat;
4018 recent_rotated[0] += rstat->recent_rotated[0];
4019 recent_rotated[1] += rstat->recent_rotated[1];
4020 recent_scanned[0] += rstat->recent_scanned[0];
4021 recent_scanned[1] += rstat->recent_scanned[1];
4023 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
4024 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
4025 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
4026 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
4033 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
4036 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4038 return mem_cgroup_swappiness(memcg);
4041 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
4042 struct cftype *cft, u64 val)
4044 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4050 memcg->swappiness = val;
4052 vm_swappiness = val;
4057 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4059 struct mem_cgroup_threshold_ary *t;
4060 unsigned long usage;
4065 t = rcu_dereference(memcg->thresholds.primary);
4067 t = rcu_dereference(memcg->memsw_thresholds.primary);
4072 usage = mem_cgroup_usage(memcg, swap);
4075 * current_threshold points to threshold just below or equal to usage.
4076 * If it's not true, a threshold was crossed after last
4077 * call of __mem_cgroup_threshold().
4079 i = t->current_threshold;
4082 * Iterate backward over array of thresholds starting from
4083 * current_threshold and check if a threshold is crossed.
4084 * If none of thresholds below usage is crossed, we read
4085 * only one element of the array here.
4087 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4088 eventfd_signal(t->entries[i].eventfd, 1);
4090 /* i = current_threshold + 1 */
4094 * Iterate forward over array of thresholds starting from
4095 * current_threshold+1 and check if a threshold is crossed.
4096 * If none of thresholds above usage is crossed, we read
4097 * only one element of the array here.
4099 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4100 eventfd_signal(t->entries[i].eventfd, 1);
4102 /* Update current_threshold */
4103 t->current_threshold = i - 1;
4108 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4111 __mem_cgroup_threshold(memcg, false);
4112 if (do_swap_account)
4113 __mem_cgroup_threshold(memcg, true);
4115 memcg = parent_mem_cgroup(memcg);
4119 static int compare_thresholds(const void *a, const void *b)
4121 const struct mem_cgroup_threshold *_a = a;
4122 const struct mem_cgroup_threshold *_b = b;
4124 if (_a->threshold > _b->threshold)
4127 if (_a->threshold < _b->threshold)
4133 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4135 struct mem_cgroup_eventfd_list *ev;
4137 spin_lock(&memcg_oom_lock);
4139 list_for_each_entry(ev, &memcg->oom_notify, list)
4140 eventfd_signal(ev->eventfd, 1);
4142 spin_unlock(&memcg_oom_lock);
4146 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4148 struct mem_cgroup *iter;
4150 for_each_mem_cgroup_tree(iter, memcg)
4151 mem_cgroup_oom_notify_cb(iter);
4154 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4155 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4157 struct mem_cgroup_thresholds *thresholds;
4158 struct mem_cgroup_threshold_ary *new;
4159 unsigned long threshold;
4160 unsigned long usage;
4163 ret = page_counter_memparse(args, &threshold);
4167 mutex_lock(&memcg->thresholds_lock);
4170 thresholds = &memcg->thresholds;
4171 usage = mem_cgroup_usage(memcg, false);
4172 } else if (type == _MEMSWAP) {
4173 thresholds = &memcg->memsw_thresholds;
4174 usage = mem_cgroup_usage(memcg, true);
4178 /* Check if a threshold crossed before adding a new one */
4179 if (thresholds->primary)
4180 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4182 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4184 /* Allocate memory for new array of thresholds */
4185 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4193 /* Copy thresholds (if any) to new array */
4194 if (thresholds->primary) {
4195 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4196 sizeof(struct mem_cgroup_threshold));
4199 /* Add new threshold */
4200 new->entries[size - 1].eventfd = eventfd;
4201 new->entries[size - 1].threshold = threshold;
4203 /* Sort thresholds. Registering of new threshold isn't time-critical */
4204 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4205 compare_thresholds, NULL);
4207 /* Find current threshold */
4208 new->current_threshold = -1;
4209 for (i = 0; i < size; i++) {
4210 if (new->entries[i].threshold <= usage) {
4212 * new->current_threshold will not be used until
4213 * rcu_assign_pointer(), so it's safe to increment
4216 ++new->current_threshold;
4221 /* Free old spare buffer and save old primary buffer as spare */
4222 kfree(thresholds->spare);
4223 thresholds->spare = thresholds->primary;
4225 rcu_assign_pointer(thresholds->primary, new);
4227 /* To be sure that nobody uses thresholds */
4231 mutex_unlock(&memcg->thresholds_lock);
4236 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4237 struct eventfd_ctx *eventfd, const char *args)
4239 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4242 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4243 struct eventfd_ctx *eventfd, const char *args)
4245 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4248 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4249 struct eventfd_ctx *eventfd, enum res_type type)
4251 struct mem_cgroup_thresholds *thresholds;
4252 struct mem_cgroup_threshold_ary *new;
4253 unsigned long usage;
4256 mutex_lock(&memcg->thresholds_lock);
4259 thresholds = &memcg->thresholds;
4260 usage = mem_cgroup_usage(memcg, false);
4261 } else if (type == _MEMSWAP) {
4262 thresholds = &memcg->memsw_thresholds;
4263 usage = mem_cgroup_usage(memcg, true);
4267 if (!thresholds->primary)
4270 /* Check if a threshold crossed before removing */
4271 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4273 /* Calculate new number of threshold */
4275 for (i = 0; i < thresholds->primary->size; i++) {
4276 if (thresholds->primary->entries[i].eventfd != eventfd)
4280 new = thresholds->spare;
4282 /* Set thresholds array to NULL if we don't have thresholds */
4291 /* Copy thresholds and find current threshold */
4292 new->current_threshold = -1;
4293 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4294 if (thresholds->primary->entries[i].eventfd == eventfd)
4297 new->entries[j] = thresholds->primary->entries[i];
4298 if (new->entries[j].threshold <= usage) {
4300 * new->current_threshold will not be used
4301 * until rcu_assign_pointer(), so it's safe to increment
4304 ++new->current_threshold;
4310 /* Swap primary and spare array */
4311 thresholds->spare = thresholds->primary;
4312 /* If all events are unregistered, free the spare array */
4314 kfree(thresholds->spare);
4315 thresholds->spare = NULL;
4318 rcu_assign_pointer(thresholds->primary, new);
4320 /* To be sure that nobody uses thresholds */
4323 mutex_unlock(&memcg->thresholds_lock);
4326 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4327 struct eventfd_ctx *eventfd)
4329 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4332 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4333 struct eventfd_ctx *eventfd)
4335 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4338 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4339 struct eventfd_ctx *eventfd, const char *args)
4341 struct mem_cgroup_eventfd_list *event;
4343 event = kmalloc(sizeof(*event), GFP_KERNEL);
4347 spin_lock(&memcg_oom_lock);
4349 event->eventfd = eventfd;
4350 list_add(&event->list, &memcg->oom_notify);
4352 /* already in OOM ? */
4353 if (atomic_read(&memcg->under_oom))
4354 eventfd_signal(eventfd, 1);
4355 spin_unlock(&memcg_oom_lock);
4360 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4361 struct eventfd_ctx *eventfd)
4363 struct mem_cgroup_eventfd_list *ev, *tmp;
4365 spin_lock(&memcg_oom_lock);
4367 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4368 if (ev->eventfd == eventfd) {
4369 list_del(&ev->list);
4374 spin_unlock(&memcg_oom_lock);
4377 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4379 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
4381 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4382 seq_printf(sf, "under_oom %d\n", (bool)atomic_read(&memcg->under_oom));
4386 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4387 struct cftype *cft, u64 val)
4389 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4391 /* cannot set to root cgroup and only 0 and 1 are allowed */
4392 if (!css->parent || !((val == 0) || (val == 1)))
4395 memcg->oom_kill_disable = val;
4397 memcg_oom_recover(memcg);
4402 #ifdef CONFIG_MEMCG_KMEM
4403 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
4407 memcg->kmemcg_id = -1;
4408 ret = memcg_propagate_kmem(memcg);
4412 return mem_cgroup_sockets_init(memcg, ss);
4415 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
4417 mem_cgroup_sockets_destroy(memcg);
4420 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
4425 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
4431 * DO NOT USE IN NEW FILES.
4433 * "cgroup.event_control" implementation.
4435 * This is way over-engineered. It tries to support fully configurable
4436 * events for each user. Such level of flexibility is completely
4437 * unnecessary especially in the light of the planned unified hierarchy.
4439 * Please deprecate this and replace with something simpler if at all
4444 * Unregister event and free resources.
4446 * Gets called from workqueue.
4448 static void memcg_event_remove(struct work_struct *work)
4450 struct mem_cgroup_event *event =
4451 container_of(work, struct mem_cgroup_event, remove);
4452 struct mem_cgroup *memcg = event->memcg;
4454 remove_wait_queue(event->wqh, &event->wait);
4456 event->unregister_event(memcg, event->eventfd);
4458 /* Notify userspace the event is going away. */
4459 eventfd_signal(event->eventfd, 1);
4461 eventfd_ctx_put(event->eventfd);
4463 css_put(&memcg->css);
4467 * Gets called on POLLHUP on eventfd when user closes it.
4469 * Called with wqh->lock held and interrupts disabled.
4471 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
4472 int sync, void *key)
4474 struct mem_cgroup_event *event =
4475 container_of(wait, struct mem_cgroup_event, wait);
4476 struct mem_cgroup *memcg = event->memcg;
4477 unsigned long flags = (unsigned long)key;
4479 if (flags & POLLHUP) {
4481 * If the event has been detached at cgroup removal, we
4482 * can simply return knowing the other side will cleanup
4485 * We can't race against event freeing since the other
4486 * side will require wqh->lock via remove_wait_queue(),
4489 spin_lock(&memcg->event_list_lock);
4490 if (!list_empty(&event->list)) {
4491 list_del_init(&event->list);
4493 * We are in atomic context, but cgroup_event_remove()
4494 * may sleep, so we have to call it in workqueue.
4496 schedule_work(&event->remove);
4498 spin_unlock(&memcg->event_list_lock);
4504 static void memcg_event_ptable_queue_proc(struct file *file,
4505 wait_queue_head_t *wqh, poll_table *pt)
4507 struct mem_cgroup_event *event =
4508 container_of(pt, struct mem_cgroup_event, pt);
4511 add_wait_queue(wqh, &event->wait);
4515 * DO NOT USE IN NEW FILES.
4517 * Parse input and register new cgroup event handler.
4519 * Input must be in format '<event_fd> <control_fd> <args>'.
4520 * Interpretation of args is defined by control file implementation.
4522 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4523 char *buf, size_t nbytes, loff_t off)
4525 struct cgroup_subsys_state *css = of_css(of);
4526 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4527 struct mem_cgroup_event *event;
4528 struct cgroup_subsys_state *cfile_css;
4529 unsigned int efd, cfd;
4536 buf = strstrip(buf);
4538 efd = simple_strtoul(buf, &endp, 10);
4543 cfd = simple_strtoul(buf, &endp, 10);
4544 if ((*endp != ' ') && (*endp != '\0'))
4548 event = kzalloc(sizeof(*event), GFP_KERNEL);
4552 event->memcg = memcg;
4553 INIT_LIST_HEAD(&event->list);
4554 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4555 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4556 INIT_WORK(&event->remove, memcg_event_remove);
4564 event->eventfd = eventfd_ctx_fileget(efile.file);
4565 if (IS_ERR(event->eventfd)) {
4566 ret = PTR_ERR(event->eventfd);
4573 goto out_put_eventfd;
4576 /* the process need read permission on control file */
4577 /* AV: shouldn't we check that it's been opened for read instead? */
4578 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4583 * Determine the event callbacks and set them in @event. This used
4584 * to be done via struct cftype but cgroup core no longer knows
4585 * about these events. The following is crude but the whole thing
4586 * is for compatibility anyway.
4588 * DO NOT ADD NEW FILES.
4590 name = cfile.file->f_dentry->d_name.name;
4592 if (!strcmp(name, "memory.usage_in_bytes")) {
4593 event->register_event = mem_cgroup_usage_register_event;
4594 event->unregister_event = mem_cgroup_usage_unregister_event;
4595 } else if (!strcmp(name, "memory.oom_control")) {
4596 event->register_event = mem_cgroup_oom_register_event;
4597 event->unregister_event = mem_cgroup_oom_unregister_event;
4598 } else if (!strcmp(name, "memory.pressure_level")) {
4599 event->register_event = vmpressure_register_event;
4600 event->unregister_event = vmpressure_unregister_event;
4601 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4602 event->register_event = memsw_cgroup_usage_register_event;
4603 event->unregister_event = memsw_cgroup_usage_unregister_event;
4610 * Verify @cfile should belong to @css. Also, remaining events are
4611 * automatically removed on cgroup destruction but the removal is
4612 * asynchronous, so take an extra ref on @css.
4614 cfile_css = css_tryget_online_from_dir(cfile.file->f_dentry->d_parent,
4615 &memory_cgrp_subsys);
4617 if (IS_ERR(cfile_css))
4619 if (cfile_css != css) {
4624 ret = event->register_event(memcg, event->eventfd, buf);
4628 efile.file->f_op->poll(efile.file, &event->pt);
4630 spin_lock(&memcg->event_list_lock);
4631 list_add(&event->list, &memcg->event_list);
4632 spin_unlock(&memcg->event_list_lock);
4644 eventfd_ctx_put(event->eventfd);
4653 static struct cftype mem_cgroup_files[] = {
4655 .name = "usage_in_bytes",
4656 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4657 .read_u64 = mem_cgroup_read_u64,
4660 .name = "max_usage_in_bytes",
4661 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4662 .write = mem_cgroup_reset,
4663 .read_u64 = mem_cgroup_read_u64,
4666 .name = "limit_in_bytes",
4667 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4668 .write = mem_cgroup_write,
4669 .read_u64 = mem_cgroup_read_u64,
4672 .name = "soft_limit_in_bytes",
4673 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4674 .write = mem_cgroup_write,
4675 .read_u64 = mem_cgroup_read_u64,
4679 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4680 .write = mem_cgroup_reset,
4681 .read_u64 = mem_cgroup_read_u64,
4685 .seq_show = memcg_stat_show,
4688 .name = "force_empty",
4689 .write = mem_cgroup_force_empty_write,
4692 .name = "use_hierarchy",
4693 .write_u64 = mem_cgroup_hierarchy_write,
4694 .read_u64 = mem_cgroup_hierarchy_read,
4697 .name = "cgroup.event_control", /* XXX: for compat */
4698 .write = memcg_write_event_control,
4699 .flags = CFTYPE_NO_PREFIX,
4703 .name = "swappiness",
4704 .read_u64 = mem_cgroup_swappiness_read,
4705 .write_u64 = mem_cgroup_swappiness_write,
4708 .name = "move_charge_at_immigrate",
4709 .read_u64 = mem_cgroup_move_charge_read,
4710 .write_u64 = mem_cgroup_move_charge_write,
4713 .name = "oom_control",
4714 .seq_show = mem_cgroup_oom_control_read,
4715 .write_u64 = mem_cgroup_oom_control_write,
4716 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4719 .name = "pressure_level",
4723 .name = "numa_stat",
4724 .seq_show = memcg_numa_stat_show,
4727 #ifdef CONFIG_MEMCG_KMEM
4729 .name = "kmem.limit_in_bytes",
4730 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4731 .write = mem_cgroup_write,
4732 .read_u64 = mem_cgroup_read_u64,
4735 .name = "kmem.usage_in_bytes",
4736 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4737 .read_u64 = mem_cgroup_read_u64,
4740 .name = "kmem.failcnt",
4741 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4742 .write = mem_cgroup_reset,
4743 .read_u64 = mem_cgroup_read_u64,
4746 .name = "kmem.max_usage_in_bytes",
4747 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4748 .write = mem_cgroup_reset,
4749 .read_u64 = mem_cgroup_read_u64,
4751 #ifdef CONFIG_SLABINFO
4753 .name = "kmem.slabinfo",
4754 .seq_show = mem_cgroup_slabinfo_read,
4758 { }, /* terminate */
4761 #ifdef CONFIG_MEMCG_SWAP
4762 static struct cftype memsw_cgroup_files[] = {
4764 .name = "memsw.usage_in_bytes",
4765 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4766 .read_u64 = mem_cgroup_read_u64,
4769 .name = "memsw.max_usage_in_bytes",
4770 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4771 .write = mem_cgroup_reset,
4772 .read_u64 = mem_cgroup_read_u64,
4775 .name = "memsw.limit_in_bytes",
4776 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4777 .write = mem_cgroup_write,
4778 .read_u64 = mem_cgroup_read_u64,
4781 .name = "memsw.failcnt",
4782 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4783 .write = mem_cgroup_reset,
4784 .read_u64 = mem_cgroup_read_u64,
4786 { }, /* terminate */
4789 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4791 struct mem_cgroup_per_node *pn;
4792 struct mem_cgroup_per_zone *mz;
4793 int zone, tmp = node;
4795 * This routine is called against possible nodes.
4796 * But it's BUG to call kmalloc() against offline node.
4798 * TODO: this routine can waste much memory for nodes which will
4799 * never be onlined. It's better to use memory hotplug callback
4802 if (!node_state(node, N_NORMAL_MEMORY))
4804 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4808 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4809 mz = &pn->zoneinfo[zone];
4810 lruvec_init(&mz->lruvec);
4811 mz->usage_in_excess = 0;
4812 mz->on_tree = false;
4815 memcg->nodeinfo[node] = pn;
4819 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4821 kfree(memcg->nodeinfo[node]);
4824 static struct mem_cgroup *mem_cgroup_alloc(void)
4826 struct mem_cgroup *memcg;
4829 size = sizeof(struct mem_cgroup);
4830 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4832 memcg = kzalloc(size, GFP_KERNEL);
4836 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4839 spin_lock_init(&memcg->pcp_counter_lock);
4848 * At destroying mem_cgroup, references from swap_cgroup can remain.
4849 * (scanning all at force_empty is too costly...)
4851 * Instead of clearing all references at force_empty, we remember
4852 * the number of reference from swap_cgroup and free mem_cgroup when
4853 * it goes down to 0.
4855 * Removal of cgroup itself succeeds regardless of refs from swap.
4858 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4862 mem_cgroup_remove_from_trees(memcg);
4865 free_mem_cgroup_per_zone_info(memcg, node);
4867 free_percpu(memcg->stat);
4870 * We need to make sure that (at least for now), the jump label
4871 * destruction code runs outside of the cgroup lock. This is because
4872 * get_online_cpus(), which is called from the static_branch update,
4873 * can't be called inside the cgroup_lock. cpusets are the ones
4874 * enforcing this dependency, so if they ever change, we might as well.
4876 * schedule_work() will guarantee this happens. Be careful if you need
4877 * to move this code around, and make sure it is outside
4880 disarm_static_keys(memcg);
4885 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4887 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4889 if (!memcg->memory.parent)
4891 return mem_cgroup_from_counter(memcg->memory.parent, memory);
4893 EXPORT_SYMBOL(parent_mem_cgroup);
4895 static void __init mem_cgroup_soft_limit_tree_init(void)
4897 struct mem_cgroup_tree_per_node *rtpn;
4898 struct mem_cgroup_tree_per_zone *rtpz;
4899 int tmp, node, zone;
4901 for_each_node(node) {
4903 if (!node_state(node, N_NORMAL_MEMORY))
4905 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
4908 soft_limit_tree.rb_tree_per_node[node] = rtpn;
4910 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4911 rtpz = &rtpn->rb_tree_per_zone[zone];
4912 rtpz->rb_root = RB_ROOT;
4913 spin_lock_init(&rtpz->lock);
4918 static struct cgroup_subsys_state * __ref
4919 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4921 struct mem_cgroup *memcg;
4922 long error = -ENOMEM;
4925 memcg = mem_cgroup_alloc();
4927 return ERR_PTR(error);
4930 if (alloc_mem_cgroup_per_zone_info(memcg, node))
4934 if (parent_css == NULL) {
4935 root_mem_cgroup = memcg;
4936 page_counter_init(&memcg->memory, NULL);
4937 page_counter_init(&memcg->memsw, NULL);
4938 page_counter_init(&memcg->kmem, NULL);
4941 memcg->last_scanned_node = MAX_NUMNODES;
4942 INIT_LIST_HEAD(&memcg->oom_notify);
4943 memcg->move_charge_at_immigrate = 0;
4944 mutex_init(&memcg->thresholds_lock);
4945 spin_lock_init(&memcg->move_lock);
4946 vmpressure_init(&memcg->vmpressure);
4947 INIT_LIST_HEAD(&memcg->event_list);
4948 spin_lock_init(&memcg->event_list_lock);
4953 __mem_cgroup_free(memcg);
4954 return ERR_PTR(error);
4958 mem_cgroup_css_online(struct cgroup_subsys_state *css)
4960 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4961 struct mem_cgroup *parent = mem_cgroup_from_css(css->parent);
4964 if (css->id > MEM_CGROUP_ID_MAX)
4970 mutex_lock(&memcg_create_mutex);
4972 memcg->use_hierarchy = parent->use_hierarchy;
4973 memcg->oom_kill_disable = parent->oom_kill_disable;
4974 memcg->swappiness = mem_cgroup_swappiness(parent);
4976 if (parent->use_hierarchy) {
4977 page_counter_init(&memcg->memory, &parent->memory);
4978 page_counter_init(&memcg->memsw, &parent->memsw);
4979 page_counter_init(&memcg->kmem, &parent->kmem);
4982 * No need to take a reference to the parent because cgroup
4983 * core guarantees its existence.
4986 page_counter_init(&memcg->memory, NULL);
4987 page_counter_init(&memcg->memsw, NULL);
4988 page_counter_init(&memcg->kmem, NULL);
4990 * Deeper hierachy with use_hierarchy == false doesn't make
4991 * much sense so let cgroup subsystem know about this
4992 * unfortunate state in our controller.
4994 if (parent != root_mem_cgroup)
4995 memory_cgrp_subsys.broken_hierarchy = true;
4997 mutex_unlock(&memcg_create_mutex);
4999 ret = memcg_init_kmem(memcg, &memory_cgrp_subsys);
5004 * Make sure the memcg is initialized: mem_cgroup_iter()
5005 * orders reading memcg->initialized against its callers
5006 * reading the memcg members.
5008 smp_store_release(&memcg->initialized, 1);
5013 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5015 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5016 struct mem_cgroup_event *event, *tmp;
5019 * Unregister events and notify userspace.
5020 * Notify userspace about cgroup removing only after rmdir of cgroup
5021 * directory to avoid race between userspace and kernelspace.
5023 spin_lock(&memcg->event_list_lock);
5024 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5025 list_del_init(&event->list);
5026 schedule_work(&event->remove);
5028 spin_unlock(&memcg->event_list_lock);
5030 memcg_unregister_all_caches(memcg);
5031 vmpressure_cleanup(&memcg->vmpressure);
5034 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5036 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5038 memcg_destroy_kmem(memcg);
5039 __mem_cgroup_free(memcg);
5043 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5044 * @css: the target css
5046 * Reset the states of the mem_cgroup associated with @css. This is
5047 * invoked when the userland requests disabling on the default hierarchy
5048 * but the memcg is pinned through dependency. The memcg should stop
5049 * applying policies and should revert to the vanilla state as it may be
5050 * made visible again.
5052 * The current implementation only resets the essential configurations.
5053 * This needs to be expanded to cover all the visible parts.
5055 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5057 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5059 mem_cgroup_resize_limit(memcg, PAGE_COUNTER_MAX);
5060 mem_cgroup_resize_memsw_limit(memcg, PAGE_COUNTER_MAX);
5061 memcg_update_kmem_limit(memcg, PAGE_COUNTER_MAX);
5062 memcg->soft_limit = 0;
5066 /* Handlers for move charge at task migration. */
5067 static int mem_cgroup_do_precharge(unsigned long count)
5071 /* Try a single bulk charge without reclaim first */
5072 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_WAIT, count);
5074 mc.precharge += count;
5077 if (ret == -EINTR) {
5078 cancel_charge(root_mem_cgroup, count);
5082 /* Try charges one by one with reclaim */
5084 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_NORETRY, 1);
5086 * In case of failure, any residual charges against
5087 * mc.to will be dropped by mem_cgroup_clear_mc()
5088 * later on. However, cancel any charges that are
5089 * bypassed to root right away or they'll be lost.
5092 cancel_charge(root_mem_cgroup, 1);
5102 * get_mctgt_type - get target type of moving charge
5103 * @vma: the vma the pte to be checked belongs
5104 * @addr: the address corresponding to the pte to be checked
5105 * @ptent: the pte to be checked
5106 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5109 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5110 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5111 * move charge. if @target is not NULL, the page is stored in target->page
5112 * with extra refcnt got(Callers should handle it).
5113 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5114 * target for charge migration. if @target is not NULL, the entry is stored
5117 * Called with pte lock held.
5124 enum mc_target_type {
5130 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5131 unsigned long addr, pte_t ptent)
5133 struct page *page = vm_normal_page(vma, addr, ptent);
5135 if (!page || !page_mapped(page))
5137 if (PageAnon(page)) {
5138 /* we don't move shared anon */
5141 } else if (!move_file())
5142 /* we ignore mapcount for file pages */
5144 if (!get_page_unless_zero(page))
5151 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5152 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5154 struct page *page = NULL;
5155 swp_entry_t ent = pte_to_swp_entry(ptent);
5157 if (!move_anon() || non_swap_entry(ent))
5160 * Because lookup_swap_cache() updates some statistics counter,
5161 * we call find_get_page() with swapper_space directly.
5163 page = find_get_page(swap_address_space(ent), ent.val);
5164 if (do_swap_account)
5165 entry->val = ent.val;
5170 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5171 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5177 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5178 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5180 struct page *page = NULL;
5181 struct address_space *mapping;
5184 if (!vma->vm_file) /* anonymous vma */
5189 mapping = vma->vm_file->f_mapping;
5190 if (pte_none(ptent))
5191 pgoff = linear_page_index(vma, addr);
5192 else /* pte_file(ptent) is true */
5193 pgoff = pte_to_pgoff(ptent);
5195 /* page is moved even if it's not RSS of this task(page-faulted). */
5197 /* shmem/tmpfs may report page out on swap: account for that too. */
5198 if (shmem_mapping(mapping)) {
5199 page = find_get_entry(mapping, pgoff);
5200 if (radix_tree_exceptional_entry(page)) {
5201 swp_entry_t swp = radix_to_swp_entry(page);
5202 if (do_swap_account)
5204 page = find_get_page(swap_address_space(swp), swp.val);
5207 page = find_get_page(mapping, pgoff);
5209 page = find_get_page(mapping, pgoff);
5214 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5215 unsigned long addr, pte_t ptent, union mc_target *target)
5217 struct page *page = NULL;
5218 struct page_cgroup *pc;
5219 enum mc_target_type ret = MC_TARGET_NONE;
5220 swp_entry_t ent = { .val = 0 };
5222 if (pte_present(ptent))
5223 page = mc_handle_present_pte(vma, addr, ptent);
5224 else if (is_swap_pte(ptent))
5225 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5226 else if (pte_none(ptent) || pte_file(ptent))
5227 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5229 if (!page && !ent.val)
5232 pc = lookup_page_cgroup(page);
5234 * Do only loose check w/o serialization.
5235 * mem_cgroup_move_account() checks the pc is valid or
5236 * not under LRU exclusion.
5238 if (pc->mem_cgroup == mc.from) {
5239 ret = MC_TARGET_PAGE;
5241 target->page = page;
5243 if (!ret || !target)
5246 /* There is a swap entry and a page doesn't exist or isn't charged */
5247 if (ent.val && !ret &&
5248 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5249 ret = MC_TARGET_SWAP;
5256 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5258 * We don't consider swapping or file mapped pages because THP does not
5259 * support them for now.
5260 * Caller should make sure that pmd_trans_huge(pmd) is true.
5262 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5263 unsigned long addr, pmd_t pmd, union mc_target *target)
5265 struct page *page = NULL;
5266 struct page_cgroup *pc;
5267 enum mc_target_type ret = MC_TARGET_NONE;
5269 page = pmd_page(pmd);
5270 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5273 pc = lookup_page_cgroup(page);
5274 if (pc->mem_cgroup == mc.from) {
5275 ret = MC_TARGET_PAGE;
5278 target->page = page;
5284 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5285 unsigned long addr, pmd_t pmd, union mc_target *target)
5287 return MC_TARGET_NONE;
5291 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5292 unsigned long addr, unsigned long end,
5293 struct mm_walk *walk)
5295 struct vm_area_struct *vma = walk->private;
5299 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
5300 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5301 mc.precharge += HPAGE_PMD_NR;
5306 if (pmd_trans_unstable(pmd))
5308 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5309 for (; addr != end; pte++, addr += PAGE_SIZE)
5310 if (get_mctgt_type(vma, addr, *pte, NULL))
5311 mc.precharge++; /* increment precharge temporarily */
5312 pte_unmap_unlock(pte - 1, ptl);
5318 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5320 unsigned long precharge;
5321 struct vm_area_struct *vma;
5323 down_read(&mm->mmap_sem);
5324 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5325 struct mm_walk mem_cgroup_count_precharge_walk = {
5326 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5330 if (is_vm_hugetlb_page(vma))
5332 walk_page_range(vma->vm_start, vma->vm_end,
5333 &mem_cgroup_count_precharge_walk);
5335 up_read(&mm->mmap_sem);
5337 precharge = mc.precharge;
5343 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5345 unsigned long precharge = mem_cgroup_count_precharge(mm);
5347 VM_BUG_ON(mc.moving_task);
5348 mc.moving_task = current;
5349 return mem_cgroup_do_precharge(precharge);
5352 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5353 static void __mem_cgroup_clear_mc(void)
5355 struct mem_cgroup *from = mc.from;
5356 struct mem_cgroup *to = mc.to;
5358 /* we must uncharge all the leftover precharges from mc.to */
5360 cancel_charge(mc.to, mc.precharge);
5364 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5365 * we must uncharge here.
5367 if (mc.moved_charge) {
5368 cancel_charge(mc.from, mc.moved_charge);
5369 mc.moved_charge = 0;
5371 /* we must fixup refcnts and charges */
5372 if (mc.moved_swap) {
5373 /* uncharge swap account from the old cgroup */
5374 if (!mem_cgroup_is_root(mc.from))
5375 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5378 * we charged both to->memory and to->memsw, so we
5379 * should uncharge to->memory.
5381 if (!mem_cgroup_is_root(mc.to))
5382 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5384 css_put_many(&mc.from->css, mc.moved_swap);
5386 /* we've already done css_get(mc.to) */
5389 memcg_oom_recover(from);
5390 memcg_oom_recover(to);
5391 wake_up_all(&mc.waitq);
5394 static void mem_cgroup_clear_mc(void)
5396 struct mem_cgroup *from = mc.from;
5399 * we must clear moving_task before waking up waiters at the end of
5402 mc.moving_task = NULL;
5403 __mem_cgroup_clear_mc();
5404 spin_lock(&mc.lock);
5407 spin_unlock(&mc.lock);
5409 atomic_dec(&from->moving_account);
5412 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
5413 struct cgroup_taskset *tset)
5415 struct task_struct *p = cgroup_taskset_first(tset);
5417 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5418 unsigned long move_charge_at_immigrate;
5421 * We are now commited to this value whatever it is. Changes in this
5422 * tunable will only affect upcoming migrations, not the current one.
5423 * So we need to save it, and keep it going.
5425 move_charge_at_immigrate = memcg->move_charge_at_immigrate;
5426 if (move_charge_at_immigrate) {
5427 struct mm_struct *mm;
5428 struct mem_cgroup *from = mem_cgroup_from_task(p);
5430 VM_BUG_ON(from == memcg);
5432 mm = get_task_mm(p);
5435 /* We move charges only when we move a owner of the mm */
5436 if (mm->owner == p) {
5439 VM_BUG_ON(mc.precharge);
5440 VM_BUG_ON(mc.moved_charge);
5441 VM_BUG_ON(mc.moved_swap);
5444 * Signal mem_cgroup_begin_page_stat() to take
5445 * the memcg's move_lock while we're moving
5446 * its pages to another memcg. Then wait for
5447 * already started RCU-only updates to finish.
5449 atomic_inc(&from->moving_account);
5452 spin_lock(&mc.lock);
5455 mc.immigrate_flags = move_charge_at_immigrate;
5456 spin_unlock(&mc.lock);
5457 /* We set mc.moving_task later */
5459 ret = mem_cgroup_precharge_mc(mm);
5461 mem_cgroup_clear_mc();
5468 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
5469 struct cgroup_taskset *tset)
5472 mem_cgroup_clear_mc();
5475 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5476 unsigned long addr, unsigned long end,
5477 struct mm_walk *walk)
5480 struct vm_area_struct *vma = walk->private;
5483 enum mc_target_type target_type;
5484 union mc_target target;
5486 struct page_cgroup *pc;
5489 * We don't take compound_lock() here but no race with splitting thp
5491 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
5492 * under splitting, which means there's no concurrent thp split,
5493 * - if another thread runs into split_huge_page() just after we
5494 * entered this if-block, the thread must wait for page table lock
5495 * to be unlocked in __split_huge_page_splitting(), where the main
5496 * part of thp split is not executed yet.
5498 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
5499 if (mc.precharge < HPAGE_PMD_NR) {
5503 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5504 if (target_type == MC_TARGET_PAGE) {
5506 if (!isolate_lru_page(page)) {
5507 pc = lookup_page_cgroup(page);
5508 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
5509 pc, mc.from, mc.to)) {
5510 mc.precharge -= HPAGE_PMD_NR;
5511 mc.moved_charge += HPAGE_PMD_NR;
5513 putback_lru_page(page);
5521 if (pmd_trans_unstable(pmd))
5524 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5525 for (; addr != end; addr += PAGE_SIZE) {
5526 pte_t ptent = *(pte++);
5532 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5533 case MC_TARGET_PAGE:
5535 if (isolate_lru_page(page))
5537 pc = lookup_page_cgroup(page);
5538 if (!mem_cgroup_move_account(page, 1, pc,
5541 /* we uncharge from mc.from later. */
5544 putback_lru_page(page);
5545 put: /* get_mctgt_type() gets the page */
5548 case MC_TARGET_SWAP:
5550 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5552 /* we fixup refcnts and charges later. */
5560 pte_unmap_unlock(pte - 1, ptl);
5565 * We have consumed all precharges we got in can_attach().
5566 * We try charge one by one, but don't do any additional
5567 * charges to mc.to if we have failed in charge once in attach()
5570 ret = mem_cgroup_do_precharge(1);
5578 static void mem_cgroup_move_charge(struct mm_struct *mm)
5580 struct vm_area_struct *vma;
5582 lru_add_drain_all();
5584 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5586 * Someone who are holding the mmap_sem might be waiting in
5587 * waitq. So we cancel all extra charges, wake up all waiters,
5588 * and retry. Because we cancel precharges, we might not be able
5589 * to move enough charges, but moving charge is a best-effort
5590 * feature anyway, so it wouldn't be a big problem.
5592 __mem_cgroup_clear_mc();
5596 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5598 struct mm_walk mem_cgroup_move_charge_walk = {
5599 .pmd_entry = mem_cgroup_move_charge_pte_range,
5603 if (is_vm_hugetlb_page(vma))
5605 ret = walk_page_range(vma->vm_start, vma->vm_end,
5606 &mem_cgroup_move_charge_walk);
5609 * means we have consumed all precharges and failed in
5610 * doing additional charge. Just abandon here.
5614 up_read(&mm->mmap_sem);
5617 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
5618 struct cgroup_taskset *tset)
5620 struct task_struct *p = cgroup_taskset_first(tset);
5621 struct mm_struct *mm = get_task_mm(p);
5625 mem_cgroup_move_charge(mm);
5629 mem_cgroup_clear_mc();
5631 #else /* !CONFIG_MMU */
5632 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
5633 struct cgroup_taskset *tset)
5637 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
5638 struct cgroup_taskset *tset)
5641 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
5642 struct cgroup_taskset *tset)
5648 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5649 * to verify whether we're attached to the default hierarchy on each mount
5652 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5655 * use_hierarchy is forced on the default hierarchy. cgroup core
5656 * guarantees that @root doesn't have any children, so turning it
5657 * on for the root memcg is enough.
5659 if (cgroup_on_dfl(root_css->cgroup))
5660 mem_cgroup_from_css(root_css)->use_hierarchy = true;
5663 struct cgroup_subsys memory_cgrp_subsys = {
5664 .css_alloc = mem_cgroup_css_alloc,
5665 .css_online = mem_cgroup_css_online,
5666 .css_offline = mem_cgroup_css_offline,
5667 .css_free = mem_cgroup_css_free,
5668 .css_reset = mem_cgroup_css_reset,
5669 .can_attach = mem_cgroup_can_attach,
5670 .cancel_attach = mem_cgroup_cancel_attach,
5671 .attach = mem_cgroup_move_task,
5672 .bind = mem_cgroup_bind,
5673 .legacy_cftypes = mem_cgroup_files,
5677 #ifdef CONFIG_MEMCG_SWAP
5678 static int __init enable_swap_account(char *s)
5680 if (!strcmp(s, "1"))
5681 really_do_swap_account = 1;
5682 else if (!strcmp(s, "0"))
5683 really_do_swap_account = 0;
5686 __setup("swapaccount=", enable_swap_account);
5688 static void __init memsw_file_init(void)
5690 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
5691 memsw_cgroup_files));
5694 static void __init enable_swap_cgroup(void)
5696 if (!mem_cgroup_disabled() && really_do_swap_account) {
5697 do_swap_account = 1;
5703 static void __init enable_swap_cgroup(void)
5708 #ifdef CONFIG_MEMCG_SWAP
5710 * mem_cgroup_swapout - transfer a memsw charge to swap
5711 * @page: page whose memsw charge to transfer
5712 * @entry: swap entry to move the charge to
5714 * Transfer the memsw charge of @page to @entry.
5716 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
5718 struct mem_cgroup *memcg;
5719 struct page_cgroup *pc;
5720 unsigned short oldid;
5722 VM_BUG_ON_PAGE(PageLRU(page), page);
5723 VM_BUG_ON_PAGE(page_count(page), page);
5725 if (!do_swap_account)
5728 pc = lookup_page_cgroup(page);
5729 memcg = pc->mem_cgroup;
5731 /* Readahead page, never charged */
5735 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg));
5736 VM_BUG_ON_PAGE(oldid, page);
5737 mem_cgroup_swap_statistics(memcg, true);
5739 pc->mem_cgroup = NULL;
5741 if (!mem_cgroup_is_root(memcg))
5742 page_counter_uncharge(&memcg->memory, 1);
5744 /* XXX: caller holds IRQ-safe mapping->tree_lock */
5745 VM_BUG_ON(!irqs_disabled());
5747 mem_cgroup_charge_statistics(memcg, page, -1);
5748 memcg_check_events(memcg, page);
5752 * mem_cgroup_uncharge_swap - uncharge a swap entry
5753 * @entry: swap entry to uncharge
5755 * Drop the memsw charge associated with @entry.
5757 void mem_cgroup_uncharge_swap(swp_entry_t entry)
5759 struct mem_cgroup *memcg;
5762 if (!do_swap_account)
5765 id = swap_cgroup_record(entry, 0);
5767 memcg = mem_cgroup_lookup(id);
5769 if (!mem_cgroup_is_root(memcg))
5770 page_counter_uncharge(&memcg->memsw, 1);
5771 mem_cgroup_swap_statistics(memcg, false);
5772 css_put(&memcg->css);
5779 * mem_cgroup_try_charge - try charging a page
5780 * @page: page to charge
5781 * @mm: mm context of the victim
5782 * @gfp_mask: reclaim mode
5783 * @memcgp: charged memcg return
5785 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5786 * pages according to @gfp_mask if necessary.
5788 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5789 * Otherwise, an error code is returned.
5791 * After page->mapping has been set up, the caller must finalize the
5792 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5793 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5795 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
5796 gfp_t gfp_mask, struct mem_cgroup **memcgp)
5798 struct mem_cgroup *memcg = NULL;
5799 unsigned int nr_pages = 1;
5802 if (mem_cgroup_disabled())
5805 if (PageSwapCache(page)) {
5806 struct page_cgroup *pc = lookup_page_cgroup(page);
5808 * Every swap fault against a single page tries to charge the
5809 * page, bail as early as possible. shmem_unuse() encounters
5810 * already charged pages, too. The USED bit is protected by
5811 * the page lock, which serializes swap cache removal, which
5812 * in turn serializes uncharging.
5818 if (PageTransHuge(page)) {
5819 nr_pages <<= compound_order(page);
5820 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5823 if (do_swap_account && PageSwapCache(page))
5824 memcg = try_get_mem_cgroup_from_page(page);
5826 memcg = get_mem_cgroup_from_mm(mm);
5828 ret = try_charge(memcg, gfp_mask, nr_pages);
5830 css_put(&memcg->css);
5832 if (ret == -EINTR) {
5833 memcg = root_mem_cgroup;
5842 * mem_cgroup_commit_charge - commit a page charge
5843 * @page: page to charge
5844 * @memcg: memcg to charge the page to
5845 * @lrucare: page might be on LRU already
5847 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5848 * after page->mapping has been set up. This must happen atomically
5849 * as part of the page instantiation, i.e. under the page table lock
5850 * for anonymous pages, under the page lock for page and swap cache.
5852 * In addition, the page must not be on the LRU during the commit, to
5853 * prevent racing with task migration. If it might be, use @lrucare.
5855 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5857 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
5860 unsigned int nr_pages = 1;
5862 VM_BUG_ON_PAGE(!page->mapping, page);
5863 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
5865 if (mem_cgroup_disabled())
5868 * Swap faults will attempt to charge the same page multiple
5869 * times. But reuse_swap_page() might have removed the page
5870 * from swapcache already, so we can't check PageSwapCache().
5875 commit_charge(page, memcg, lrucare);
5877 if (PageTransHuge(page)) {
5878 nr_pages <<= compound_order(page);
5879 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5882 local_irq_disable();
5883 mem_cgroup_charge_statistics(memcg, page, nr_pages);
5884 memcg_check_events(memcg, page);
5887 if (do_swap_account && PageSwapCache(page)) {
5888 swp_entry_t entry = { .val = page_private(page) };
5890 * The swap entry might not get freed for a long time,
5891 * let's not wait for it. The page already received a
5892 * memory+swap charge, drop the swap entry duplicate.
5894 mem_cgroup_uncharge_swap(entry);
5899 * mem_cgroup_cancel_charge - cancel a page charge
5900 * @page: page to charge
5901 * @memcg: memcg to charge the page to
5903 * Cancel a charge transaction started by mem_cgroup_try_charge().
5905 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg)
5907 unsigned int nr_pages = 1;
5909 if (mem_cgroup_disabled())
5912 * Swap faults will attempt to charge the same page multiple
5913 * times. But reuse_swap_page() might have removed the page
5914 * from swapcache already, so we can't check PageSwapCache().
5919 if (PageTransHuge(page)) {
5920 nr_pages <<= compound_order(page);
5921 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5924 cancel_charge(memcg, nr_pages);
5927 static void uncharge_batch(struct mem_cgroup *memcg, unsigned long pgpgout,
5928 unsigned long nr_anon, unsigned long nr_file,
5929 unsigned long nr_huge, struct page *dummy_page)
5931 unsigned long nr_pages = nr_anon + nr_file;
5932 unsigned long flags;
5934 if (!mem_cgroup_is_root(memcg)) {
5935 page_counter_uncharge(&memcg->memory, nr_pages);
5936 if (do_swap_account)
5937 page_counter_uncharge(&memcg->memsw, nr_pages);
5938 memcg_oom_recover(memcg);
5941 local_irq_save(flags);
5942 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS], nr_anon);
5943 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_CACHE], nr_file);
5944 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE], nr_huge);
5945 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT], pgpgout);
5946 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
5947 memcg_check_events(memcg, dummy_page);
5948 local_irq_restore(flags);
5950 if (!mem_cgroup_is_root(memcg))
5951 css_put_many(&memcg->css, nr_pages);
5954 static void uncharge_list(struct list_head *page_list)
5956 struct mem_cgroup *memcg = NULL;
5957 unsigned long nr_anon = 0;
5958 unsigned long nr_file = 0;
5959 unsigned long nr_huge = 0;
5960 unsigned long pgpgout = 0;
5961 struct list_head *next;
5964 next = page_list->next;
5966 unsigned int nr_pages = 1;
5967 struct page_cgroup *pc;
5969 page = list_entry(next, struct page, lru);
5970 next = page->lru.next;
5972 VM_BUG_ON_PAGE(PageLRU(page), page);
5973 VM_BUG_ON_PAGE(page_count(page), page);
5975 pc = lookup_page_cgroup(page);
5976 if (!pc->mem_cgroup)
5980 * Nobody should be changing or seriously looking at
5981 * pc->mem_cgroup at this point, we have fully
5982 * exclusive access to the page.
5985 if (memcg != pc->mem_cgroup) {
5987 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5989 pgpgout = nr_anon = nr_file = nr_huge = 0;
5991 memcg = pc->mem_cgroup;
5994 if (PageTransHuge(page)) {
5995 nr_pages <<= compound_order(page);
5996 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5997 nr_huge += nr_pages;
6001 nr_anon += nr_pages;
6003 nr_file += nr_pages;
6005 pc->mem_cgroup = NULL;
6008 } while (next != page_list);
6011 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
6016 * mem_cgroup_uncharge - uncharge a page
6017 * @page: page to uncharge
6019 * Uncharge a page previously charged with mem_cgroup_try_charge() and
6020 * mem_cgroup_commit_charge().
6022 void mem_cgroup_uncharge(struct page *page)
6024 struct page_cgroup *pc;
6026 if (mem_cgroup_disabled())
6029 /* Don't touch page->lru of any random page, pre-check: */
6030 pc = lookup_page_cgroup(page);
6031 if (!pc->mem_cgroup)
6034 INIT_LIST_HEAD(&page->lru);
6035 uncharge_list(&page->lru);
6039 * mem_cgroup_uncharge_list - uncharge a list of page
6040 * @page_list: list of pages to uncharge
6042 * Uncharge a list of pages previously charged with
6043 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
6045 void mem_cgroup_uncharge_list(struct list_head *page_list)
6047 if (mem_cgroup_disabled())
6050 if (!list_empty(page_list))
6051 uncharge_list(page_list);
6055 * mem_cgroup_migrate - migrate a charge to another page
6056 * @oldpage: currently charged page
6057 * @newpage: page to transfer the charge to
6058 * @lrucare: both pages might be on the LRU already
6060 * Migrate the charge from @oldpage to @newpage.
6062 * Both pages must be locked, @newpage->mapping must be set up.
6064 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage,
6067 struct mem_cgroup *memcg;
6068 struct page_cgroup *pc;
6071 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6072 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6073 VM_BUG_ON_PAGE(!lrucare && PageLRU(oldpage), oldpage);
6074 VM_BUG_ON_PAGE(!lrucare && PageLRU(newpage), newpage);
6075 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6076 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6079 if (mem_cgroup_disabled())
6082 /* Page cache replacement: new page already charged? */
6083 pc = lookup_page_cgroup(newpage);
6088 * Swapcache readahead pages can get migrated before being
6089 * charged, and migration from compaction can happen to an
6090 * uncharged page when the PFN walker finds a page that
6091 * reclaim just put back on the LRU but has not released yet.
6093 pc = lookup_page_cgroup(oldpage);
6094 memcg = pc->mem_cgroup;
6099 lock_page_lru(oldpage, &isolated);
6101 pc->mem_cgroup = NULL;
6104 unlock_page_lru(oldpage, isolated);
6106 commit_charge(newpage, memcg, lrucare);
6110 * subsys_initcall() for memory controller.
6112 * Some parts like hotcpu_notifier() have to be initialized from this context
6113 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
6114 * everything that doesn't depend on a specific mem_cgroup structure should
6115 * be initialized from here.
6117 static int __init mem_cgroup_init(void)
6119 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
6120 enable_swap_cgroup();
6121 mem_cgroup_soft_limit_tree_init();
6125 subsys_initcall(mem_cgroup_init);