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 * This program is free software; you can redistribute it and/or modify
14 * it under the terms of the GNU General Public License as published by
15 * the Free Software Foundation; either version 2 of the License, or
16 * (at your option) any later version.
18 * This program is distributed in the hope that it will be useful,
19 * but WITHOUT ANY WARRANTY; without even the implied warranty of
20 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
21 * GNU General Public License for more details.
24 #include <linux/res_counter.h>
25 #include <linux/memcontrol.h>
26 #include <linux/cgroup.h>
28 #include <linux/hugetlb.h>
29 #include <linux/pagemap.h>
30 #include <linux/smp.h>
31 #include <linux/page-flags.h>
32 #include <linux/backing-dev.h>
33 #include <linux/bit_spinlock.h>
34 #include <linux/rcupdate.h>
35 #include <linux/limits.h>
36 #include <linux/export.h>
37 #include <linux/mutex.h>
38 #include <linux/rbtree.h>
39 #include <linux/slab.h>
40 #include <linux/swap.h>
41 #include <linux/swapops.h>
42 #include <linux/spinlock.h>
43 #include <linux/eventfd.h>
44 #include <linux/sort.h>
46 #include <linux/seq_file.h>
47 #include <linux/vmalloc.h>
48 #include <linux/mm_inline.h>
49 #include <linux/page_cgroup.h>
50 #include <linux/cpu.h>
51 #include <linux/oom.h>
55 #include <net/tcp_memcontrol.h>
57 #include <asm/uaccess.h>
59 #include <trace/events/vmscan.h>
61 struct cgroup_subsys mem_cgroup_subsys __read_mostly;
62 EXPORT_SYMBOL(mem_cgroup_subsys);
64 #define MEM_CGROUP_RECLAIM_RETRIES 5
65 static struct mem_cgroup *root_mem_cgroup __read_mostly;
67 #ifdef CONFIG_MEMCG_SWAP
68 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
69 int do_swap_account __read_mostly;
71 /* for remember boot option*/
72 #ifdef CONFIG_MEMCG_SWAP_ENABLED
73 static int really_do_swap_account __initdata = 1;
75 static int really_do_swap_account __initdata = 0;
79 #define do_swap_account 0
84 * Statistics for memory cgroup.
86 enum mem_cgroup_stat_index {
88 * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
90 MEM_CGROUP_STAT_CACHE, /* # of pages charged as cache */
91 MEM_CGROUP_STAT_RSS, /* # of pages charged as anon rss */
92 MEM_CGROUP_STAT_FILE_MAPPED, /* # of pages charged as file rss */
93 MEM_CGROUP_STAT_SWAP, /* # of pages, swapped out */
94 MEM_CGROUP_STAT_NSTATS,
97 static const char * const mem_cgroup_stat_names[] = {
104 enum mem_cgroup_events_index {
105 MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */
106 MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */
107 MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */
108 MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */
109 MEM_CGROUP_EVENTS_NSTATS,
112 static const char * const mem_cgroup_events_names[] = {
120 * Per memcg event counter is incremented at every pagein/pageout. With THP,
121 * it will be incremated by the number of pages. This counter is used for
122 * for trigger some periodic events. This is straightforward and better
123 * than using jiffies etc. to handle periodic memcg event.
125 enum mem_cgroup_events_target {
126 MEM_CGROUP_TARGET_THRESH,
127 MEM_CGROUP_TARGET_SOFTLIMIT,
128 MEM_CGROUP_TARGET_NUMAINFO,
131 #define THRESHOLDS_EVENTS_TARGET 128
132 #define SOFTLIMIT_EVENTS_TARGET 1024
133 #define NUMAINFO_EVENTS_TARGET 1024
135 struct mem_cgroup_stat_cpu {
136 long count[MEM_CGROUP_STAT_NSTATS];
137 unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
138 unsigned long nr_page_events;
139 unsigned long targets[MEM_CGROUP_NTARGETS];
142 struct mem_cgroup_reclaim_iter {
143 /* css_id of the last scanned hierarchy member */
145 /* scan generation, increased every round-trip */
146 unsigned int generation;
150 * per-zone information in memory controller.
152 struct mem_cgroup_per_zone {
153 struct lruvec lruvec;
154 unsigned long lru_size[NR_LRU_LISTS];
156 struct mem_cgroup_reclaim_iter reclaim_iter[DEF_PRIORITY + 1];
158 struct rb_node tree_node; /* RB tree node */
159 unsigned long long usage_in_excess;/* Set to the value by which */
160 /* the soft limit is exceeded*/
162 struct mem_cgroup *memcg; /* Back pointer, we cannot */
163 /* use container_of */
166 struct mem_cgroup_per_node {
167 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
170 struct mem_cgroup_lru_info {
171 struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
175 * Cgroups above their limits are maintained in a RB-Tree, independent of
176 * their hierarchy representation
179 struct mem_cgroup_tree_per_zone {
180 struct rb_root rb_root;
184 struct mem_cgroup_tree_per_node {
185 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
188 struct mem_cgroup_tree {
189 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
192 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
194 struct mem_cgroup_threshold {
195 struct eventfd_ctx *eventfd;
200 struct mem_cgroup_threshold_ary {
201 /* An array index points to threshold just below or equal to usage. */
202 int current_threshold;
203 /* Size of entries[] */
205 /* Array of thresholds */
206 struct mem_cgroup_threshold entries[0];
209 struct mem_cgroup_thresholds {
210 /* Primary thresholds array */
211 struct mem_cgroup_threshold_ary *primary;
213 * Spare threshold array.
214 * This is needed to make mem_cgroup_unregister_event() "never fail".
215 * It must be able to store at least primary->size - 1 entries.
217 struct mem_cgroup_threshold_ary *spare;
221 struct mem_cgroup_eventfd_list {
222 struct list_head list;
223 struct eventfd_ctx *eventfd;
226 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
227 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
230 * The memory controller data structure. The memory controller controls both
231 * page cache and RSS per cgroup. We would eventually like to provide
232 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
233 * to help the administrator determine what knobs to tune.
235 * TODO: Add a water mark for the memory controller. Reclaim will begin when
236 * we hit the water mark. May be even add a low water mark, such that
237 * no reclaim occurs from a cgroup at it's low water mark, this is
238 * a feature that will be implemented much later in the future.
241 struct cgroup_subsys_state css;
243 * the counter to account for memory usage
245 struct res_counter res;
249 * the counter to account for mem+swap usage.
251 struct res_counter memsw;
254 * rcu_freeing is used only when freeing struct mem_cgroup,
255 * so put it into a union to avoid wasting more memory.
256 * It must be disjoint from the css field. It could be
257 * in a union with the res field, but res plays a much
258 * larger part in mem_cgroup life than memsw, and might
259 * be of interest, even at time of free, when debugging.
260 * So share rcu_head with the less interesting memsw.
262 struct rcu_head rcu_freeing;
264 * We also need some space for a worker in deferred freeing.
265 * By the time we call it, rcu_freeing is no longer in use.
267 struct work_struct work_freeing;
271 * the counter to account for kernel memory usage.
273 struct res_counter kmem;
275 * Per cgroup active and inactive list, similar to the
276 * per zone LRU lists.
278 struct mem_cgroup_lru_info info;
279 int last_scanned_node;
281 nodemask_t scan_nodes;
282 atomic_t numainfo_events;
283 atomic_t numainfo_updating;
286 * Should the accounting and control be hierarchical, per subtree?
289 unsigned long kmem_account_flags; /* See KMEM_ACCOUNTED_*, below */
297 /* OOM-Killer disable */
298 int oom_kill_disable;
300 /* set when res.limit == memsw.limit */
301 bool memsw_is_minimum;
303 /* protect arrays of thresholds */
304 struct mutex thresholds_lock;
306 /* thresholds for memory usage. RCU-protected */
307 struct mem_cgroup_thresholds thresholds;
309 /* thresholds for mem+swap usage. RCU-protected */
310 struct mem_cgroup_thresholds memsw_thresholds;
312 /* For oom notifier event fd */
313 struct list_head oom_notify;
316 * Should we move charges of a task when a task is moved into this
317 * mem_cgroup ? And what type of charges should we move ?
319 unsigned long move_charge_at_immigrate;
321 * set > 0 if pages under this cgroup are moving to other cgroup.
323 atomic_t moving_account;
324 /* taken only while moving_account > 0 */
325 spinlock_t move_lock;
329 struct mem_cgroup_stat_cpu __percpu *stat;
331 * used when a cpu is offlined or other synchronizations
332 * See mem_cgroup_read_stat().
334 struct mem_cgroup_stat_cpu nocpu_base;
335 spinlock_t pcp_counter_lock;
337 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
338 struct tcp_memcontrol tcp_mem;
342 /* internal only representation about the status of kmem accounting. */
344 KMEM_ACCOUNTED_ACTIVE = 0, /* accounted by this cgroup itself */
347 #define KMEM_ACCOUNTED_MASK (1 << KMEM_ACCOUNTED_ACTIVE)
349 #ifdef CONFIG_MEMCG_KMEM
350 static inline void memcg_kmem_set_active(struct mem_cgroup *memcg)
352 set_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
356 /* Stuffs for move charges at task migration. */
358 * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
359 * left-shifted bitmap of these types.
362 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
363 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
367 /* "mc" and its members are protected by cgroup_mutex */
368 static struct move_charge_struct {
369 spinlock_t lock; /* for from, to */
370 struct mem_cgroup *from;
371 struct mem_cgroup *to;
372 unsigned long precharge;
373 unsigned long moved_charge;
374 unsigned long moved_swap;
375 struct task_struct *moving_task; /* a task moving charges */
376 wait_queue_head_t waitq; /* a waitq for other context */
378 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
379 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
382 static bool move_anon(void)
384 return test_bit(MOVE_CHARGE_TYPE_ANON,
385 &mc.to->move_charge_at_immigrate);
388 static bool move_file(void)
390 return test_bit(MOVE_CHARGE_TYPE_FILE,
391 &mc.to->move_charge_at_immigrate);
395 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
396 * limit reclaim to prevent infinite loops, if they ever occur.
398 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
399 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
402 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
403 MEM_CGROUP_CHARGE_TYPE_ANON,
404 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
405 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
409 /* for encoding cft->private value on file */
417 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
418 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
419 #define MEMFILE_ATTR(val) ((val) & 0xffff)
420 /* Used for OOM nofiier */
421 #define OOM_CONTROL (0)
424 * Reclaim flags for mem_cgroup_hierarchical_reclaim
426 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
427 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
428 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
429 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
431 static void mem_cgroup_get(struct mem_cgroup *memcg);
432 static void mem_cgroup_put(struct mem_cgroup *memcg);
435 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
437 return container_of(s, struct mem_cgroup, css);
440 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
442 return (memcg == root_mem_cgroup);
445 /* Writing them here to avoid exposing memcg's inner layout */
446 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
448 void sock_update_memcg(struct sock *sk)
450 if (mem_cgroup_sockets_enabled) {
451 struct mem_cgroup *memcg;
452 struct cg_proto *cg_proto;
454 BUG_ON(!sk->sk_prot->proto_cgroup);
456 /* Socket cloning can throw us here with sk_cgrp already
457 * filled. It won't however, necessarily happen from
458 * process context. So the test for root memcg given
459 * the current task's memcg won't help us in this case.
461 * Respecting the original socket's memcg is a better
462 * decision in this case.
465 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
466 mem_cgroup_get(sk->sk_cgrp->memcg);
471 memcg = mem_cgroup_from_task(current);
472 cg_proto = sk->sk_prot->proto_cgroup(memcg);
473 if (!mem_cgroup_is_root(memcg) && memcg_proto_active(cg_proto)) {
474 mem_cgroup_get(memcg);
475 sk->sk_cgrp = cg_proto;
480 EXPORT_SYMBOL(sock_update_memcg);
482 void sock_release_memcg(struct sock *sk)
484 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
485 struct mem_cgroup *memcg;
486 WARN_ON(!sk->sk_cgrp->memcg);
487 memcg = sk->sk_cgrp->memcg;
488 mem_cgroup_put(memcg);
492 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
494 if (!memcg || mem_cgroup_is_root(memcg))
497 return &memcg->tcp_mem.cg_proto;
499 EXPORT_SYMBOL(tcp_proto_cgroup);
501 static void disarm_sock_keys(struct mem_cgroup *memcg)
503 if (!memcg_proto_activated(&memcg->tcp_mem.cg_proto))
505 static_key_slow_dec(&memcg_socket_limit_enabled);
508 static void disarm_sock_keys(struct mem_cgroup *memcg)
513 static void drain_all_stock_async(struct mem_cgroup *memcg);
515 static struct mem_cgroup_per_zone *
516 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
518 return &memcg->info.nodeinfo[nid]->zoneinfo[zid];
521 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
526 static struct mem_cgroup_per_zone *
527 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
529 int nid = page_to_nid(page);
530 int zid = page_zonenum(page);
532 return mem_cgroup_zoneinfo(memcg, nid, zid);
535 static struct mem_cgroup_tree_per_zone *
536 soft_limit_tree_node_zone(int nid, int zid)
538 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
541 static struct mem_cgroup_tree_per_zone *
542 soft_limit_tree_from_page(struct page *page)
544 int nid = page_to_nid(page);
545 int zid = page_zonenum(page);
547 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
551 __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
552 struct mem_cgroup_per_zone *mz,
553 struct mem_cgroup_tree_per_zone *mctz,
554 unsigned long long new_usage_in_excess)
556 struct rb_node **p = &mctz->rb_root.rb_node;
557 struct rb_node *parent = NULL;
558 struct mem_cgroup_per_zone *mz_node;
563 mz->usage_in_excess = new_usage_in_excess;
564 if (!mz->usage_in_excess)
568 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
570 if (mz->usage_in_excess < mz_node->usage_in_excess)
573 * We can't avoid mem cgroups that are over their soft
574 * limit by the same amount
576 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
579 rb_link_node(&mz->tree_node, parent, p);
580 rb_insert_color(&mz->tree_node, &mctz->rb_root);
585 __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
586 struct mem_cgroup_per_zone *mz,
587 struct mem_cgroup_tree_per_zone *mctz)
591 rb_erase(&mz->tree_node, &mctz->rb_root);
596 mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
597 struct mem_cgroup_per_zone *mz,
598 struct mem_cgroup_tree_per_zone *mctz)
600 spin_lock(&mctz->lock);
601 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
602 spin_unlock(&mctz->lock);
606 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
608 unsigned long long excess;
609 struct mem_cgroup_per_zone *mz;
610 struct mem_cgroup_tree_per_zone *mctz;
611 int nid = page_to_nid(page);
612 int zid = page_zonenum(page);
613 mctz = soft_limit_tree_from_page(page);
616 * Necessary to update all ancestors when hierarchy is used.
617 * because their event counter is not touched.
619 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
620 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
621 excess = res_counter_soft_limit_excess(&memcg->res);
623 * We have to update the tree if mz is on RB-tree or
624 * mem is over its softlimit.
626 if (excess || mz->on_tree) {
627 spin_lock(&mctz->lock);
628 /* if on-tree, remove it */
630 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
632 * Insert again. mz->usage_in_excess will be updated.
633 * If excess is 0, no tree ops.
635 __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
636 spin_unlock(&mctz->lock);
641 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
644 struct mem_cgroup_per_zone *mz;
645 struct mem_cgroup_tree_per_zone *mctz;
647 for_each_node(node) {
648 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
649 mz = mem_cgroup_zoneinfo(memcg, node, zone);
650 mctz = soft_limit_tree_node_zone(node, zone);
651 mem_cgroup_remove_exceeded(memcg, mz, mctz);
656 static struct mem_cgroup_per_zone *
657 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
659 struct rb_node *rightmost = NULL;
660 struct mem_cgroup_per_zone *mz;
664 rightmost = rb_last(&mctz->rb_root);
666 goto done; /* Nothing to reclaim from */
668 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
670 * Remove the node now but someone else can add it back,
671 * we will to add it back at the end of reclaim to its correct
672 * position in the tree.
674 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
675 if (!res_counter_soft_limit_excess(&mz->memcg->res) ||
676 !css_tryget(&mz->memcg->css))
682 static struct mem_cgroup_per_zone *
683 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
685 struct mem_cgroup_per_zone *mz;
687 spin_lock(&mctz->lock);
688 mz = __mem_cgroup_largest_soft_limit_node(mctz);
689 spin_unlock(&mctz->lock);
694 * Implementation Note: reading percpu statistics for memcg.
696 * Both of vmstat[] and percpu_counter has threshold and do periodic
697 * synchronization to implement "quick" read. There are trade-off between
698 * reading cost and precision of value. Then, we may have a chance to implement
699 * a periodic synchronizion of counter in memcg's counter.
701 * But this _read() function is used for user interface now. The user accounts
702 * memory usage by memory cgroup and he _always_ requires exact value because
703 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
704 * have to visit all online cpus and make sum. So, for now, unnecessary
705 * synchronization is not implemented. (just implemented for cpu hotplug)
707 * If there are kernel internal actions which can make use of some not-exact
708 * value, and reading all cpu value can be performance bottleneck in some
709 * common workload, threashold and synchonization as vmstat[] should be
712 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
713 enum mem_cgroup_stat_index idx)
719 for_each_online_cpu(cpu)
720 val += per_cpu(memcg->stat->count[idx], cpu);
721 #ifdef CONFIG_HOTPLUG_CPU
722 spin_lock(&memcg->pcp_counter_lock);
723 val += memcg->nocpu_base.count[idx];
724 spin_unlock(&memcg->pcp_counter_lock);
730 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
733 int val = (charge) ? 1 : -1;
734 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
737 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
738 enum mem_cgroup_events_index idx)
740 unsigned long val = 0;
743 for_each_online_cpu(cpu)
744 val += per_cpu(memcg->stat->events[idx], cpu);
745 #ifdef CONFIG_HOTPLUG_CPU
746 spin_lock(&memcg->pcp_counter_lock);
747 val += memcg->nocpu_base.events[idx];
748 spin_unlock(&memcg->pcp_counter_lock);
753 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
754 bool anon, int nr_pages)
759 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
760 * counted as CACHE even if it's on ANON LRU.
763 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
766 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
769 /* pagein of a big page is an event. So, ignore page size */
771 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
773 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
774 nr_pages = -nr_pages; /* for event */
777 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
783 mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
785 struct mem_cgroup_per_zone *mz;
787 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
788 return mz->lru_size[lru];
792 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
793 unsigned int lru_mask)
795 struct mem_cgroup_per_zone *mz;
797 unsigned long ret = 0;
799 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
802 if (BIT(lru) & lru_mask)
803 ret += mz->lru_size[lru];
809 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
810 int nid, unsigned int lru_mask)
815 for (zid = 0; zid < MAX_NR_ZONES; zid++)
816 total += mem_cgroup_zone_nr_lru_pages(memcg,
822 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
823 unsigned int lru_mask)
828 for_each_node_state(nid, N_MEMORY)
829 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
833 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
834 enum mem_cgroup_events_target target)
836 unsigned long val, next;
838 val = __this_cpu_read(memcg->stat->nr_page_events);
839 next = __this_cpu_read(memcg->stat->targets[target]);
840 /* from time_after() in jiffies.h */
841 if ((long)next - (long)val < 0) {
843 case MEM_CGROUP_TARGET_THRESH:
844 next = val + THRESHOLDS_EVENTS_TARGET;
846 case MEM_CGROUP_TARGET_SOFTLIMIT:
847 next = val + SOFTLIMIT_EVENTS_TARGET;
849 case MEM_CGROUP_TARGET_NUMAINFO:
850 next = val + NUMAINFO_EVENTS_TARGET;
855 __this_cpu_write(memcg->stat->targets[target], next);
862 * Check events in order.
865 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
868 /* threshold event is triggered in finer grain than soft limit */
869 if (unlikely(mem_cgroup_event_ratelimit(memcg,
870 MEM_CGROUP_TARGET_THRESH))) {
872 bool do_numainfo __maybe_unused;
874 do_softlimit = mem_cgroup_event_ratelimit(memcg,
875 MEM_CGROUP_TARGET_SOFTLIMIT);
877 do_numainfo = mem_cgroup_event_ratelimit(memcg,
878 MEM_CGROUP_TARGET_NUMAINFO);
882 mem_cgroup_threshold(memcg);
883 if (unlikely(do_softlimit))
884 mem_cgroup_update_tree(memcg, page);
886 if (unlikely(do_numainfo))
887 atomic_inc(&memcg->numainfo_events);
893 struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
895 return mem_cgroup_from_css(
896 cgroup_subsys_state(cont, mem_cgroup_subsys_id));
899 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
902 * mm_update_next_owner() may clear mm->owner to NULL
903 * if it races with swapoff, page migration, etc.
904 * So this can be called with p == NULL.
909 return mem_cgroup_from_css(task_subsys_state(p, mem_cgroup_subsys_id));
912 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
914 struct mem_cgroup *memcg = NULL;
919 * Because we have no locks, mm->owner's may be being moved to other
920 * cgroup. We use css_tryget() here even if this looks
921 * pessimistic (rather than adding locks here).
925 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
926 if (unlikely(!memcg))
928 } while (!css_tryget(&memcg->css));
934 * mem_cgroup_iter - iterate over memory cgroup hierarchy
935 * @root: hierarchy root
936 * @prev: previously returned memcg, NULL on first invocation
937 * @reclaim: cookie for shared reclaim walks, NULL for full walks
939 * Returns references to children of the hierarchy below @root, or
940 * @root itself, or %NULL after a full round-trip.
942 * Caller must pass the return value in @prev on subsequent
943 * invocations for reference counting, or use mem_cgroup_iter_break()
944 * to cancel a hierarchy walk before the round-trip is complete.
946 * Reclaimers can specify a zone and a priority level in @reclaim to
947 * divide up the memcgs in the hierarchy among all concurrent
948 * reclaimers operating on the same zone and priority.
950 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
951 struct mem_cgroup *prev,
952 struct mem_cgroup_reclaim_cookie *reclaim)
954 struct mem_cgroup *memcg = NULL;
957 if (mem_cgroup_disabled())
961 root = root_mem_cgroup;
963 if (prev && !reclaim)
964 id = css_id(&prev->css);
966 if (prev && prev != root)
969 if (!root->use_hierarchy && root != root_mem_cgroup) {
976 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
977 struct cgroup_subsys_state *css;
980 int nid = zone_to_nid(reclaim->zone);
981 int zid = zone_idx(reclaim->zone);
982 struct mem_cgroup_per_zone *mz;
984 mz = mem_cgroup_zoneinfo(root, nid, zid);
985 iter = &mz->reclaim_iter[reclaim->priority];
986 if (prev && reclaim->generation != iter->generation)
992 css = css_get_next(&mem_cgroup_subsys, id + 1, &root->css, &id);
994 if (css == &root->css || css_tryget(css))
995 memcg = mem_cgroup_from_css(css);
1001 iter->position = id;
1004 else if (!prev && memcg)
1005 reclaim->generation = iter->generation;
1015 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1016 * @root: hierarchy root
1017 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1019 void mem_cgroup_iter_break(struct mem_cgroup *root,
1020 struct mem_cgroup *prev)
1023 root = root_mem_cgroup;
1024 if (prev && prev != root)
1025 css_put(&prev->css);
1029 * Iteration constructs for visiting all cgroups (under a tree). If
1030 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1031 * be used for reference counting.
1033 #define for_each_mem_cgroup_tree(iter, root) \
1034 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1036 iter = mem_cgroup_iter(root, iter, NULL))
1038 #define for_each_mem_cgroup(iter) \
1039 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1041 iter = mem_cgroup_iter(NULL, iter, NULL))
1043 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1045 struct mem_cgroup *memcg;
1048 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1049 if (unlikely(!memcg))
1054 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1057 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1065 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1068 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1069 * @zone: zone of the wanted lruvec
1070 * @memcg: memcg of the wanted lruvec
1072 * Returns the lru list vector holding pages for the given @zone and
1073 * @mem. This can be the global zone lruvec, if the memory controller
1076 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1077 struct mem_cgroup *memcg)
1079 struct mem_cgroup_per_zone *mz;
1080 struct lruvec *lruvec;
1082 if (mem_cgroup_disabled()) {
1083 lruvec = &zone->lruvec;
1087 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1088 lruvec = &mz->lruvec;
1091 * Since a node can be onlined after the mem_cgroup was created,
1092 * we have to be prepared to initialize lruvec->zone here;
1093 * and if offlined then reonlined, we need to reinitialize it.
1095 if (unlikely(lruvec->zone != zone))
1096 lruvec->zone = zone;
1101 * Following LRU functions are allowed to be used without PCG_LOCK.
1102 * Operations are called by routine of global LRU independently from memcg.
1103 * What we have to take care of here is validness of pc->mem_cgroup.
1105 * Changes to pc->mem_cgroup happens when
1108 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1109 * It is added to LRU before charge.
1110 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1111 * When moving account, the page is not on LRU. It's isolated.
1115 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1117 * @zone: zone of the page
1119 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1121 struct mem_cgroup_per_zone *mz;
1122 struct mem_cgroup *memcg;
1123 struct page_cgroup *pc;
1124 struct lruvec *lruvec;
1126 if (mem_cgroup_disabled()) {
1127 lruvec = &zone->lruvec;
1131 pc = lookup_page_cgroup(page);
1132 memcg = pc->mem_cgroup;
1135 * Surreptitiously switch any uncharged offlist page to root:
1136 * an uncharged page off lru does nothing to secure
1137 * its former mem_cgroup from sudden removal.
1139 * Our caller holds lru_lock, and PageCgroupUsed is updated
1140 * under page_cgroup lock: between them, they make all uses
1141 * of pc->mem_cgroup safe.
1143 if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1144 pc->mem_cgroup = memcg = root_mem_cgroup;
1146 mz = page_cgroup_zoneinfo(memcg, page);
1147 lruvec = &mz->lruvec;
1150 * Since a node can be onlined after the mem_cgroup was created,
1151 * we have to be prepared to initialize lruvec->zone here;
1152 * and if offlined then reonlined, we need to reinitialize it.
1154 if (unlikely(lruvec->zone != zone))
1155 lruvec->zone = zone;
1160 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1161 * @lruvec: mem_cgroup per zone lru vector
1162 * @lru: index of lru list the page is sitting on
1163 * @nr_pages: positive when adding or negative when removing
1165 * This function must be called when a page is added to or removed from an
1168 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1171 struct mem_cgroup_per_zone *mz;
1172 unsigned long *lru_size;
1174 if (mem_cgroup_disabled())
1177 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1178 lru_size = mz->lru_size + lru;
1179 *lru_size += nr_pages;
1180 VM_BUG_ON((long)(*lru_size) < 0);
1184 * Checks whether given mem is same or in the root_mem_cgroup's
1187 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1188 struct mem_cgroup *memcg)
1190 if (root_memcg == memcg)
1192 if (!root_memcg->use_hierarchy || !memcg)
1194 return css_is_ancestor(&memcg->css, &root_memcg->css);
1197 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1198 struct mem_cgroup *memcg)
1203 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1208 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *memcg)
1211 struct mem_cgroup *curr = NULL;
1212 struct task_struct *p;
1214 p = find_lock_task_mm(task);
1216 curr = try_get_mem_cgroup_from_mm(p->mm);
1220 * All threads may have already detached their mm's, but the oom
1221 * killer still needs to detect if they have already been oom
1222 * killed to prevent needlessly killing additional tasks.
1225 curr = mem_cgroup_from_task(task);
1227 css_get(&curr->css);
1233 * We should check use_hierarchy of "memcg" not "curr". Because checking
1234 * use_hierarchy of "curr" here make this function true if hierarchy is
1235 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1236 * hierarchy(even if use_hierarchy is disabled in "memcg").
1238 ret = mem_cgroup_same_or_subtree(memcg, curr);
1239 css_put(&curr->css);
1243 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1245 unsigned long inactive_ratio;
1246 unsigned long inactive;
1247 unsigned long active;
1250 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1251 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1253 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1255 inactive_ratio = int_sqrt(10 * gb);
1259 return inactive * inactive_ratio < active;
1262 int mem_cgroup_inactive_file_is_low(struct lruvec *lruvec)
1264 unsigned long active;
1265 unsigned long inactive;
1267 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_FILE);
1268 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_FILE);
1270 return (active > inactive);
1273 #define mem_cgroup_from_res_counter(counter, member) \
1274 container_of(counter, struct mem_cgroup, member)
1277 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1278 * @memcg: the memory cgroup
1280 * Returns the maximum amount of memory @mem can be charged with, in
1283 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1285 unsigned long long margin;
1287 margin = res_counter_margin(&memcg->res);
1288 if (do_swap_account)
1289 margin = min(margin, res_counter_margin(&memcg->memsw));
1290 return margin >> PAGE_SHIFT;
1293 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1295 struct cgroup *cgrp = memcg->css.cgroup;
1298 if (cgrp->parent == NULL)
1299 return vm_swappiness;
1301 return memcg->swappiness;
1305 * memcg->moving_account is used for checking possibility that some thread is
1306 * calling move_account(). When a thread on CPU-A starts moving pages under
1307 * a memcg, other threads should check memcg->moving_account under
1308 * rcu_read_lock(), like this:
1312 * memcg->moving_account+1 if (memcg->mocing_account)
1314 * synchronize_rcu() update something.
1319 /* for quick checking without looking up memcg */
1320 atomic_t memcg_moving __read_mostly;
1322 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1324 atomic_inc(&memcg_moving);
1325 atomic_inc(&memcg->moving_account);
1329 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1332 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1333 * We check NULL in callee rather than caller.
1336 atomic_dec(&memcg_moving);
1337 atomic_dec(&memcg->moving_account);
1342 * 2 routines for checking "mem" is under move_account() or not.
1344 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1345 * is used for avoiding races in accounting. If true,
1346 * pc->mem_cgroup may be overwritten.
1348 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1349 * under hierarchy of moving cgroups. This is for
1350 * waiting at hith-memory prressure caused by "move".
1353 static bool mem_cgroup_stolen(struct mem_cgroup *memcg)
1355 VM_BUG_ON(!rcu_read_lock_held());
1356 return atomic_read(&memcg->moving_account) > 0;
1359 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1361 struct mem_cgroup *from;
1362 struct mem_cgroup *to;
1365 * Unlike task_move routines, we access mc.to, mc.from not under
1366 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1368 spin_lock(&mc.lock);
1374 ret = mem_cgroup_same_or_subtree(memcg, from)
1375 || mem_cgroup_same_or_subtree(memcg, to);
1377 spin_unlock(&mc.lock);
1381 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1383 if (mc.moving_task && current != mc.moving_task) {
1384 if (mem_cgroup_under_move(memcg)) {
1386 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1387 /* moving charge context might have finished. */
1390 finish_wait(&mc.waitq, &wait);
1398 * Take this lock when
1399 * - a code tries to modify page's memcg while it's USED.
1400 * - a code tries to modify page state accounting in a memcg.
1401 * see mem_cgroup_stolen(), too.
1403 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1404 unsigned long *flags)
1406 spin_lock_irqsave(&memcg->move_lock, *flags);
1409 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1410 unsigned long *flags)
1412 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1416 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1417 * @memcg: The memory cgroup that went over limit
1418 * @p: Task that is going to be killed
1420 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1423 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1425 struct cgroup *task_cgrp;
1426 struct cgroup *mem_cgrp;
1428 * Need a buffer in BSS, can't rely on allocations. The code relies
1429 * on the assumption that OOM is serialized for memory controller.
1430 * If this assumption is broken, revisit this code.
1432 static char memcg_name[PATH_MAX];
1440 mem_cgrp = memcg->css.cgroup;
1441 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1443 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1446 * Unfortunately, we are unable to convert to a useful name
1447 * But we'll still print out the usage information
1454 printk(KERN_INFO "Task in %s killed", memcg_name);
1457 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1465 * Continues from above, so we don't need an KERN_ level
1467 printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1470 printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1471 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1472 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1473 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1474 printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1476 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1477 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1478 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1479 printk(KERN_INFO "kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1480 res_counter_read_u64(&memcg->kmem, RES_USAGE) >> 10,
1481 res_counter_read_u64(&memcg->kmem, RES_LIMIT) >> 10,
1482 res_counter_read_u64(&memcg->kmem, RES_FAILCNT));
1486 * This function returns the number of memcg under hierarchy tree. Returns
1487 * 1(self count) if no children.
1489 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1492 struct mem_cgroup *iter;
1494 for_each_mem_cgroup_tree(iter, memcg)
1500 * Return the memory (and swap, if configured) limit for a memcg.
1502 static u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1506 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1509 * Do not consider swap space if we cannot swap due to swappiness
1511 if (mem_cgroup_swappiness(memcg)) {
1514 limit += total_swap_pages << PAGE_SHIFT;
1515 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1518 * If memsw is finite and limits the amount of swap space
1519 * available to this memcg, return that limit.
1521 limit = min(limit, memsw);
1527 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1530 struct mem_cgroup *iter;
1531 unsigned long chosen_points = 0;
1532 unsigned long totalpages;
1533 unsigned int points = 0;
1534 struct task_struct *chosen = NULL;
1537 * If current has a pending SIGKILL, then automatically select it. The
1538 * goal is to allow it to allocate so that it may quickly exit and free
1541 if (fatal_signal_pending(current)) {
1542 set_thread_flag(TIF_MEMDIE);
1546 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1547 totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
1548 for_each_mem_cgroup_tree(iter, memcg) {
1549 struct cgroup *cgroup = iter->css.cgroup;
1550 struct cgroup_iter it;
1551 struct task_struct *task;
1553 cgroup_iter_start(cgroup, &it);
1554 while ((task = cgroup_iter_next(cgroup, &it))) {
1555 switch (oom_scan_process_thread(task, totalpages, NULL,
1557 case OOM_SCAN_SELECT:
1559 put_task_struct(chosen);
1561 chosen_points = ULONG_MAX;
1562 get_task_struct(chosen);
1564 case OOM_SCAN_CONTINUE:
1566 case OOM_SCAN_ABORT:
1567 cgroup_iter_end(cgroup, &it);
1568 mem_cgroup_iter_break(memcg, iter);
1570 put_task_struct(chosen);
1575 points = oom_badness(task, memcg, NULL, totalpages);
1576 if (points > chosen_points) {
1578 put_task_struct(chosen);
1580 chosen_points = points;
1581 get_task_struct(chosen);
1584 cgroup_iter_end(cgroup, &it);
1589 points = chosen_points * 1000 / totalpages;
1590 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1591 NULL, "Memory cgroup out of memory");
1594 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1596 unsigned long flags)
1598 unsigned long total = 0;
1599 bool noswap = false;
1602 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1604 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1607 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1609 drain_all_stock_async(memcg);
1610 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1612 * Allow limit shrinkers, which are triggered directly
1613 * by userspace, to catch signals and stop reclaim
1614 * after minimal progress, regardless of the margin.
1616 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1618 if (mem_cgroup_margin(memcg))
1621 * If nothing was reclaimed after two attempts, there
1622 * may be no reclaimable pages in this hierarchy.
1631 * test_mem_cgroup_node_reclaimable
1632 * @memcg: the target memcg
1633 * @nid: the node ID to be checked.
1634 * @noswap : specify true here if the user wants flle only information.
1636 * This function returns whether the specified memcg contains any
1637 * reclaimable pages on a node. Returns true if there are any reclaimable
1638 * pages in the node.
1640 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1641 int nid, bool noswap)
1643 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1645 if (noswap || !total_swap_pages)
1647 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1652 #if MAX_NUMNODES > 1
1655 * Always updating the nodemask is not very good - even if we have an empty
1656 * list or the wrong list here, we can start from some node and traverse all
1657 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1660 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1664 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1665 * pagein/pageout changes since the last update.
1667 if (!atomic_read(&memcg->numainfo_events))
1669 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1672 /* make a nodemask where this memcg uses memory from */
1673 memcg->scan_nodes = node_states[N_MEMORY];
1675 for_each_node_mask(nid, node_states[N_MEMORY]) {
1677 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1678 node_clear(nid, memcg->scan_nodes);
1681 atomic_set(&memcg->numainfo_events, 0);
1682 atomic_set(&memcg->numainfo_updating, 0);
1686 * Selecting a node where we start reclaim from. Because what we need is just
1687 * reducing usage counter, start from anywhere is O,K. Considering
1688 * memory reclaim from current node, there are pros. and cons.
1690 * Freeing memory from current node means freeing memory from a node which
1691 * we'll use or we've used. So, it may make LRU bad. And if several threads
1692 * hit limits, it will see a contention on a node. But freeing from remote
1693 * node means more costs for memory reclaim because of memory latency.
1695 * Now, we use round-robin. Better algorithm is welcomed.
1697 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1701 mem_cgroup_may_update_nodemask(memcg);
1702 node = memcg->last_scanned_node;
1704 node = next_node(node, memcg->scan_nodes);
1705 if (node == MAX_NUMNODES)
1706 node = first_node(memcg->scan_nodes);
1708 * We call this when we hit limit, not when pages are added to LRU.
1709 * No LRU may hold pages because all pages are UNEVICTABLE or
1710 * memcg is too small and all pages are not on LRU. In that case,
1711 * we use curret node.
1713 if (unlikely(node == MAX_NUMNODES))
1714 node = numa_node_id();
1716 memcg->last_scanned_node = node;
1721 * Check all nodes whether it contains reclaimable pages or not.
1722 * For quick scan, we make use of scan_nodes. This will allow us to skip
1723 * unused nodes. But scan_nodes is lazily updated and may not cotain
1724 * enough new information. We need to do double check.
1726 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1731 * quick check...making use of scan_node.
1732 * We can skip unused nodes.
1734 if (!nodes_empty(memcg->scan_nodes)) {
1735 for (nid = first_node(memcg->scan_nodes);
1737 nid = next_node(nid, memcg->scan_nodes)) {
1739 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1744 * Check rest of nodes.
1746 for_each_node_state(nid, N_MEMORY) {
1747 if (node_isset(nid, memcg->scan_nodes))
1749 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1756 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1761 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1763 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
1767 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1770 unsigned long *total_scanned)
1772 struct mem_cgroup *victim = NULL;
1775 unsigned long excess;
1776 unsigned long nr_scanned;
1777 struct mem_cgroup_reclaim_cookie reclaim = {
1782 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
1785 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1790 * If we have not been able to reclaim
1791 * anything, it might because there are
1792 * no reclaimable pages under this hierarchy
1797 * We want to do more targeted reclaim.
1798 * excess >> 2 is not to excessive so as to
1799 * reclaim too much, nor too less that we keep
1800 * coming back to reclaim from this cgroup
1802 if (total >= (excess >> 2) ||
1803 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1808 if (!mem_cgroup_reclaimable(victim, false))
1810 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1812 *total_scanned += nr_scanned;
1813 if (!res_counter_soft_limit_excess(&root_memcg->res))
1816 mem_cgroup_iter_break(root_memcg, victim);
1821 * Check OOM-Killer is already running under our hierarchy.
1822 * If someone is running, return false.
1823 * Has to be called with memcg_oom_lock
1825 static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg)
1827 struct mem_cgroup *iter, *failed = NULL;
1829 for_each_mem_cgroup_tree(iter, memcg) {
1830 if (iter->oom_lock) {
1832 * this subtree of our hierarchy is already locked
1833 * so we cannot give a lock.
1836 mem_cgroup_iter_break(memcg, iter);
1839 iter->oom_lock = true;
1846 * OK, we failed to lock the whole subtree so we have to clean up
1847 * what we set up to the failing subtree
1849 for_each_mem_cgroup_tree(iter, memcg) {
1850 if (iter == failed) {
1851 mem_cgroup_iter_break(memcg, iter);
1854 iter->oom_lock = false;
1860 * Has to be called with memcg_oom_lock
1862 static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1864 struct mem_cgroup *iter;
1866 for_each_mem_cgroup_tree(iter, memcg)
1867 iter->oom_lock = false;
1871 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1873 struct mem_cgroup *iter;
1875 for_each_mem_cgroup_tree(iter, memcg)
1876 atomic_inc(&iter->under_oom);
1879 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1881 struct mem_cgroup *iter;
1884 * When a new child is created while the hierarchy is under oom,
1885 * mem_cgroup_oom_lock() may not be called. We have to use
1886 * atomic_add_unless() here.
1888 for_each_mem_cgroup_tree(iter, memcg)
1889 atomic_add_unless(&iter->under_oom, -1, 0);
1892 static DEFINE_SPINLOCK(memcg_oom_lock);
1893 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1895 struct oom_wait_info {
1896 struct mem_cgroup *memcg;
1900 static int memcg_oom_wake_function(wait_queue_t *wait,
1901 unsigned mode, int sync, void *arg)
1903 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1904 struct mem_cgroup *oom_wait_memcg;
1905 struct oom_wait_info *oom_wait_info;
1907 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1908 oom_wait_memcg = oom_wait_info->memcg;
1911 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
1912 * Then we can use css_is_ancestor without taking care of RCU.
1914 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
1915 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
1917 return autoremove_wake_function(wait, mode, sync, arg);
1920 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
1922 /* for filtering, pass "memcg" as argument. */
1923 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1926 static void memcg_oom_recover(struct mem_cgroup *memcg)
1928 if (memcg && atomic_read(&memcg->under_oom))
1929 memcg_wakeup_oom(memcg);
1933 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1935 static bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask,
1938 struct oom_wait_info owait;
1939 bool locked, need_to_kill;
1941 owait.memcg = memcg;
1942 owait.wait.flags = 0;
1943 owait.wait.func = memcg_oom_wake_function;
1944 owait.wait.private = current;
1945 INIT_LIST_HEAD(&owait.wait.task_list);
1946 need_to_kill = true;
1947 mem_cgroup_mark_under_oom(memcg);
1949 /* At first, try to OOM lock hierarchy under memcg.*/
1950 spin_lock(&memcg_oom_lock);
1951 locked = mem_cgroup_oom_lock(memcg);
1953 * Even if signal_pending(), we can't quit charge() loop without
1954 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1955 * under OOM is always welcomed, use TASK_KILLABLE here.
1957 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1958 if (!locked || memcg->oom_kill_disable)
1959 need_to_kill = false;
1961 mem_cgroup_oom_notify(memcg);
1962 spin_unlock(&memcg_oom_lock);
1965 finish_wait(&memcg_oom_waitq, &owait.wait);
1966 mem_cgroup_out_of_memory(memcg, mask, order);
1969 finish_wait(&memcg_oom_waitq, &owait.wait);
1971 spin_lock(&memcg_oom_lock);
1973 mem_cgroup_oom_unlock(memcg);
1974 memcg_wakeup_oom(memcg);
1975 spin_unlock(&memcg_oom_lock);
1977 mem_cgroup_unmark_under_oom(memcg);
1979 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
1981 /* Give chance to dying process */
1982 schedule_timeout_uninterruptible(1);
1987 * Currently used to update mapped file statistics, but the routine can be
1988 * generalized to update other statistics as well.
1990 * Notes: Race condition
1992 * We usually use page_cgroup_lock() for accessing page_cgroup member but
1993 * it tends to be costly. But considering some conditions, we doesn't need
1994 * to do so _always_.
1996 * Considering "charge", lock_page_cgroup() is not required because all
1997 * file-stat operations happen after a page is attached to radix-tree. There
1998 * are no race with "charge".
2000 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2001 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2002 * if there are race with "uncharge". Statistics itself is properly handled
2005 * Considering "move", this is an only case we see a race. To make the race
2006 * small, we check mm->moving_account and detect there are possibility of race
2007 * If there is, we take a lock.
2010 void __mem_cgroup_begin_update_page_stat(struct page *page,
2011 bool *locked, unsigned long *flags)
2013 struct mem_cgroup *memcg;
2014 struct page_cgroup *pc;
2016 pc = lookup_page_cgroup(page);
2018 memcg = pc->mem_cgroup;
2019 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2022 * If this memory cgroup is not under account moving, we don't
2023 * need to take move_lock_mem_cgroup(). Because we already hold
2024 * rcu_read_lock(), any calls to move_account will be delayed until
2025 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2027 if (!mem_cgroup_stolen(memcg))
2030 move_lock_mem_cgroup(memcg, flags);
2031 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2032 move_unlock_mem_cgroup(memcg, flags);
2038 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
2040 struct page_cgroup *pc = lookup_page_cgroup(page);
2043 * It's guaranteed that pc->mem_cgroup never changes while
2044 * lock is held because a routine modifies pc->mem_cgroup
2045 * should take move_lock_mem_cgroup().
2047 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
2050 void mem_cgroup_update_page_stat(struct page *page,
2051 enum mem_cgroup_page_stat_item idx, int val)
2053 struct mem_cgroup *memcg;
2054 struct page_cgroup *pc = lookup_page_cgroup(page);
2055 unsigned long uninitialized_var(flags);
2057 if (mem_cgroup_disabled())
2060 memcg = pc->mem_cgroup;
2061 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2065 case MEMCG_NR_FILE_MAPPED:
2066 idx = MEM_CGROUP_STAT_FILE_MAPPED;
2072 this_cpu_add(memcg->stat->count[idx], val);
2076 * size of first charge trial. "32" comes from vmscan.c's magic value.
2077 * TODO: maybe necessary to use big numbers in big irons.
2079 #define CHARGE_BATCH 32U
2080 struct memcg_stock_pcp {
2081 struct mem_cgroup *cached; /* this never be root cgroup */
2082 unsigned int nr_pages;
2083 struct work_struct work;
2084 unsigned long flags;
2085 #define FLUSHING_CACHED_CHARGE 0
2087 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2088 static DEFINE_MUTEX(percpu_charge_mutex);
2091 * consume_stock: Try to consume stocked charge on this cpu.
2092 * @memcg: memcg to consume from.
2093 * @nr_pages: how many pages to charge.
2095 * The charges will only happen if @memcg matches the current cpu's memcg
2096 * stock, and at least @nr_pages are available in that stock. Failure to
2097 * service an allocation will refill the stock.
2099 * returns true if successful, false otherwise.
2101 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2103 struct memcg_stock_pcp *stock;
2106 if (nr_pages > CHARGE_BATCH)
2109 stock = &get_cpu_var(memcg_stock);
2110 if (memcg == stock->cached && stock->nr_pages >= nr_pages)
2111 stock->nr_pages -= nr_pages;
2112 else /* need to call res_counter_charge */
2114 put_cpu_var(memcg_stock);
2119 * Returns stocks cached in percpu to res_counter and reset cached information.
2121 static void drain_stock(struct memcg_stock_pcp *stock)
2123 struct mem_cgroup *old = stock->cached;
2125 if (stock->nr_pages) {
2126 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2128 res_counter_uncharge(&old->res, bytes);
2129 if (do_swap_account)
2130 res_counter_uncharge(&old->memsw, bytes);
2131 stock->nr_pages = 0;
2133 stock->cached = NULL;
2137 * This must be called under preempt disabled or must be called by
2138 * a thread which is pinned to local cpu.
2140 static void drain_local_stock(struct work_struct *dummy)
2142 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2144 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2148 * Cache charges(val) which is from res_counter, to local per_cpu area.
2149 * This will be consumed by consume_stock() function, later.
2151 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2153 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2155 if (stock->cached != memcg) { /* reset if necessary */
2157 stock->cached = memcg;
2159 stock->nr_pages += nr_pages;
2160 put_cpu_var(memcg_stock);
2164 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2165 * of the hierarchy under it. sync flag says whether we should block
2166 * until the work is done.
2168 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2172 /* Notify other cpus that system-wide "drain" is running */
2175 for_each_online_cpu(cpu) {
2176 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2177 struct mem_cgroup *memcg;
2179 memcg = stock->cached;
2180 if (!memcg || !stock->nr_pages)
2182 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2184 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2186 drain_local_stock(&stock->work);
2188 schedule_work_on(cpu, &stock->work);
2196 for_each_online_cpu(cpu) {
2197 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2198 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2199 flush_work(&stock->work);
2206 * Tries to drain stocked charges in other cpus. This function is asynchronous
2207 * and just put a work per cpu for draining localy on each cpu. Caller can
2208 * expects some charges will be back to res_counter later but cannot wait for
2211 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2214 * If someone calls draining, avoid adding more kworker runs.
2216 if (!mutex_trylock(&percpu_charge_mutex))
2218 drain_all_stock(root_memcg, false);
2219 mutex_unlock(&percpu_charge_mutex);
2222 /* This is a synchronous drain interface. */
2223 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2225 /* called when force_empty is called */
2226 mutex_lock(&percpu_charge_mutex);
2227 drain_all_stock(root_memcg, true);
2228 mutex_unlock(&percpu_charge_mutex);
2232 * This function drains percpu counter value from DEAD cpu and
2233 * move it to local cpu. Note that this function can be preempted.
2235 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2239 spin_lock(&memcg->pcp_counter_lock);
2240 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2241 long x = per_cpu(memcg->stat->count[i], cpu);
2243 per_cpu(memcg->stat->count[i], cpu) = 0;
2244 memcg->nocpu_base.count[i] += x;
2246 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2247 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2249 per_cpu(memcg->stat->events[i], cpu) = 0;
2250 memcg->nocpu_base.events[i] += x;
2252 spin_unlock(&memcg->pcp_counter_lock);
2255 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2256 unsigned long action,
2259 int cpu = (unsigned long)hcpu;
2260 struct memcg_stock_pcp *stock;
2261 struct mem_cgroup *iter;
2263 if (action == CPU_ONLINE)
2266 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2269 for_each_mem_cgroup(iter)
2270 mem_cgroup_drain_pcp_counter(iter, cpu);
2272 stock = &per_cpu(memcg_stock, cpu);
2278 /* See __mem_cgroup_try_charge() for details */
2280 CHARGE_OK, /* success */
2281 CHARGE_RETRY, /* need to retry but retry is not bad */
2282 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2283 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2284 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2287 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2288 unsigned int nr_pages, unsigned int min_pages,
2291 unsigned long csize = nr_pages * PAGE_SIZE;
2292 struct mem_cgroup *mem_over_limit;
2293 struct res_counter *fail_res;
2294 unsigned long flags = 0;
2297 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2300 if (!do_swap_account)
2302 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2306 res_counter_uncharge(&memcg->res, csize);
2307 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2308 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2310 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2312 * Never reclaim on behalf of optional batching, retry with a
2313 * single page instead.
2315 if (nr_pages > min_pages)
2316 return CHARGE_RETRY;
2318 if (!(gfp_mask & __GFP_WAIT))
2319 return CHARGE_WOULDBLOCK;
2321 if (gfp_mask & __GFP_NORETRY)
2322 return CHARGE_NOMEM;
2324 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2325 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2326 return CHARGE_RETRY;
2328 * Even though the limit is exceeded at this point, reclaim
2329 * may have been able to free some pages. Retry the charge
2330 * before killing the task.
2332 * Only for regular pages, though: huge pages are rather
2333 * unlikely to succeed so close to the limit, and we fall back
2334 * to regular pages anyway in case of failure.
2336 if (nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER) && ret)
2337 return CHARGE_RETRY;
2340 * At task move, charge accounts can be doubly counted. So, it's
2341 * better to wait until the end of task_move if something is going on.
2343 if (mem_cgroup_wait_acct_move(mem_over_limit))
2344 return CHARGE_RETRY;
2346 /* If we don't need to call oom-killer at el, return immediately */
2348 return CHARGE_NOMEM;
2350 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask, get_order(csize)))
2351 return CHARGE_OOM_DIE;
2353 return CHARGE_RETRY;
2357 * __mem_cgroup_try_charge() does
2358 * 1. detect memcg to be charged against from passed *mm and *ptr,
2359 * 2. update res_counter
2360 * 3. call memory reclaim if necessary.
2362 * In some special case, if the task is fatal, fatal_signal_pending() or
2363 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2364 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2365 * as possible without any hazards. 2: all pages should have a valid
2366 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2367 * pointer, that is treated as a charge to root_mem_cgroup.
2369 * So __mem_cgroup_try_charge() will return
2370 * 0 ... on success, filling *ptr with a valid memcg pointer.
2371 * -ENOMEM ... charge failure because of resource limits.
2372 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2374 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2375 * the oom-killer can be invoked.
2377 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2379 unsigned int nr_pages,
2380 struct mem_cgroup **ptr,
2383 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2384 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2385 struct mem_cgroup *memcg = NULL;
2389 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2390 * in system level. So, allow to go ahead dying process in addition to
2393 if (unlikely(test_thread_flag(TIF_MEMDIE)
2394 || fatal_signal_pending(current)))
2398 * We always charge the cgroup the mm_struct belongs to.
2399 * The mm_struct's mem_cgroup changes on task migration if the
2400 * thread group leader migrates. It's possible that mm is not
2401 * set, if so charge the root memcg (happens for pagecache usage).
2404 *ptr = root_mem_cgroup;
2406 if (*ptr) { /* css should be a valid one */
2408 if (mem_cgroup_is_root(memcg))
2410 if (consume_stock(memcg, nr_pages))
2412 css_get(&memcg->css);
2414 struct task_struct *p;
2417 p = rcu_dereference(mm->owner);
2419 * Because we don't have task_lock(), "p" can exit.
2420 * In that case, "memcg" can point to root or p can be NULL with
2421 * race with swapoff. Then, we have small risk of mis-accouning.
2422 * But such kind of mis-account by race always happens because
2423 * we don't have cgroup_mutex(). It's overkill and we allo that
2425 * (*) swapoff at el will charge against mm-struct not against
2426 * task-struct. So, mm->owner can be NULL.
2428 memcg = mem_cgroup_from_task(p);
2430 memcg = root_mem_cgroup;
2431 if (mem_cgroup_is_root(memcg)) {
2435 if (consume_stock(memcg, nr_pages)) {
2437 * It seems dagerous to access memcg without css_get().
2438 * But considering how consume_stok works, it's not
2439 * necessary. If consume_stock success, some charges
2440 * from this memcg are cached on this cpu. So, we
2441 * don't need to call css_get()/css_tryget() before
2442 * calling consume_stock().
2447 /* after here, we may be blocked. we need to get refcnt */
2448 if (!css_tryget(&memcg->css)) {
2458 /* If killed, bypass charge */
2459 if (fatal_signal_pending(current)) {
2460 css_put(&memcg->css);
2465 if (oom && !nr_oom_retries) {
2467 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2470 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, nr_pages,
2475 case CHARGE_RETRY: /* not in OOM situation but retry */
2477 css_put(&memcg->css);
2480 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2481 css_put(&memcg->css);
2483 case CHARGE_NOMEM: /* OOM routine works */
2485 css_put(&memcg->css);
2488 /* If oom, we never return -ENOMEM */
2491 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2492 css_put(&memcg->css);
2495 } while (ret != CHARGE_OK);
2497 if (batch > nr_pages)
2498 refill_stock(memcg, batch - nr_pages);
2499 css_put(&memcg->css);
2507 *ptr = root_mem_cgroup;
2512 * Somemtimes we have to undo a charge we got by try_charge().
2513 * This function is for that and do uncharge, put css's refcnt.
2514 * gotten by try_charge().
2516 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2517 unsigned int nr_pages)
2519 if (!mem_cgroup_is_root(memcg)) {
2520 unsigned long bytes = nr_pages * PAGE_SIZE;
2522 res_counter_uncharge(&memcg->res, bytes);
2523 if (do_swap_account)
2524 res_counter_uncharge(&memcg->memsw, bytes);
2529 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2530 * This is useful when moving usage to parent cgroup.
2532 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2533 unsigned int nr_pages)
2535 unsigned long bytes = nr_pages * PAGE_SIZE;
2537 if (mem_cgroup_is_root(memcg))
2540 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2541 if (do_swap_account)
2542 res_counter_uncharge_until(&memcg->memsw,
2543 memcg->memsw.parent, bytes);
2547 * A helper function to get mem_cgroup from ID. must be called under
2548 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2549 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2550 * called against removed memcg.)
2552 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2554 struct cgroup_subsys_state *css;
2556 /* ID 0 is unused ID */
2559 css = css_lookup(&mem_cgroup_subsys, id);
2562 return mem_cgroup_from_css(css);
2565 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2567 struct mem_cgroup *memcg = NULL;
2568 struct page_cgroup *pc;
2572 VM_BUG_ON(!PageLocked(page));
2574 pc = lookup_page_cgroup(page);
2575 lock_page_cgroup(pc);
2576 if (PageCgroupUsed(pc)) {
2577 memcg = pc->mem_cgroup;
2578 if (memcg && !css_tryget(&memcg->css))
2580 } else if (PageSwapCache(page)) {
2581 ent.val = page_private(page);
2582 id = lookup_swap_cgroup_id(ent);
2584 memcg = mem_cgroup_lookup(id);
2585 if (memcg && !css_tryget(&memcg->css))
2589 unlock_page_cgroup(pc);
2593 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2595 unsigned int nr_pages,
2596 enum charge_type ctype,
2599 struct page_cgroup *pc = lookup_page_cgroup(page);
2600 struct zone *uninitialized_var(zone);
2601 struct lruvec *lruvec;
2602 bool was_on_lru = false;
2605 lock_page_cgroup(pc);
2606 VM_BUG_ON(PageCgroupUsed(pc));
2608 * we don't need page_cgroup_lock about tail pages, becase they are not
2609 * accessed by any other context at this point.
2613 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2614 * may already be on some other mem_cgroup's LRU. Take care of it.
2617 zone = page_zone(page);
2618 spin_lock_irq(&zone->lru_lock);
2619 if (PageLRU(page)) {
2620 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2622 del_page_from_lru_list(page, lruvec, page_lru(page));
2627 pc->mem_cgroup = memcg;
2629 * We access a page_cgroup asynchronously without lock_page_cgroup().
2630 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2631 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2632 * before USED bit, we need memory barrier here.
2633 * See mem_cgroup_add_lru_list(), etc.
2636 SetPageCgroupUsed(pc);
2640 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2641 VM_BUG_ON(PageLRU(page));
2643 add_page_to_lru_list(page, lruvec, page_lru(page));
2645 spin_unlock_irq(&zone->lru_lock);
2648 if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
2653 mem_cgroup_charge_statistics(memcg, anon, nr_pages);
2654 unlock_page_cgroup(pc);
2657 * "charge_statistics" updated event counter. Then, check it.
2658 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2659 * if they exceeds softlimit.
2661 memcg_check_events(memcg, page);
2664 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2666 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
2668 * Because tail pages are not marked as "used", set it. We're under
2669 * zone->lru_lock, 'splitting on pmd' and compound_lock.
2670 * charge/uncharge will be never happen and move_account() is done under
2671 * compound_lock(), so we don't have to take care of races.
2673 void mem_cgroup_split_huge_fixup(struct page *head)
2675 struct page_cgroup *head_pc = lookup_page_cgroup(head);
2676 struct page_cgroup *pc;
2679 if (mem_cgroup_disabled())
2681 for (i = 1; i < HPAGE_PMD_NR; i++) {
2683 pc->mem_cgroup = head_pc->mem_cgroup;
2684 smp_wmb();/* see __commit_charge() */
2685 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2688 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2691 * mem_cgroup_move_account - move account of the page
2693 * @nr_pages: number of regular pages (>1 for huge pages)
2694 * @pc: page_cgroup of the page.
2695 * @from: mem_cgroup which the page is moved from.
2696 * @to: mem_cgroup which the page is moved to. @from != @to.
2698 * The caller must confirm following.
2699 * - page is not on LRU (isolate_page() is useful.)
2700 * - compound_lock is held when nr_pages > 1
2702 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
2705 static int mem_cgroup_move_account(struct page *page,
2706 unsigned int nr_pages,
2707 struct page_cgroup *pc,
2708 struct mem_cgroup *from,
2709 struct mem_cgroup *to)
2711 unsigned long flags;
2713 bool anon = PageAnon(page);
2715 VM_BUG_ON(from == to);
2716 VM_BUG_ON(PageLRU(page));
2718 * The page is isolated from LRU. So, collapse function
2719 * will not handle this page. But page splitting can happen.
2720 * Do this check under compound_page_lock(). The caller should
2724 if (nr_pages > 1 && !PageTransHuge(page))
2727 lock_page_cgroup(pc);
2730 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
2733 move_lock_mem_cgroup(from, &flags);
2735 if (!anon && page_mapped(page)) {
2736 /* Update mapped_file data for mem_cgroup */
2738 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2739 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2742 mem_cgroup_charge_statistics(from, anon, -nr_pages);
2744 /* caller should have done css_get */
2745 pc->mem_cgroup = to;
2746 mem_cgroup_charge_statistics(to, anon, nr_pages);
2747 move_unlock_mem_cgroup(from, &flags);
2750 unlock_page_cgroup(pc);
2754 memcg_check_events(to, page);
2755 memcg_check_events(from, page);
2761 * mem_cgroup_move_parent - moves page to the parent group
2762 * @page: the page to move
2763 * @pc: page_cgroup of the page
2764 * @child: page's cgroup
2766 * move charges to its parent or the root cgroup if the group has no
2767 * parent (aka use_hierarchy==0).
2768 * Although this might fail (get_page_unless_zero, isolate_lru_page or
2769 * mem_cgroup_move_account fails) the failure is always temporary and
2770 * it signals a race with a page removal/uncharge or migration. In the
2771 * first case the page is on the way out and it will vanish from the LRU
2772 * on the next attempt and the call should be retried later.
2773 * Isolation from the LRU fails only if page has been isolated from
2774 * the LRU since we looked at it and that usually means either global
2775 * reclaim or migration going on. The page will either get back to the
2777 * Finaly mem_cgroup_move_account fails only if the page got uncharged
2778 * (!PageCgroupUsed) or moved to a different group. The page will
2779 * disappear in the next attempt.
2781 static int mem_cgroup_move_parent(struct page *page,
2782 struct page_cgroup *pc,
2783 struct mem_cgroup *child)
2785 struct mem_cgroup *parent;
2786 unsigned int nr_pages;
2787 unsigned long uninitialized_var(flags);
2790 VM_BUG_ON(mem_cgroup_is_root(child));
2793 if (!get_page_unless_zero(page))
2795 if (isolate_lru_page(page))
2798 nr_pages = hpage_nr_pages(page);
2800 parent = parent_mem_cgroup(child);
2802 * If no parent, move charges to root cgroup.
2805 parent = root_mem_cgroup;
2808 VM_BUG_ON(!PageTransHuge(page));
2809 flags = compound_lock_irqsave(page);
2812 ret = mem_cgroup_move_account(page, nr_pages,
2815 __mem_cgroup_cancel_local_charge(child, nr_pages);
2818 compound_unlock_irqrestore(page, flags);
2819 putback_lru_page(page);
2827 * Charge the memory controller for page usage.
2829 * 0 if the charge was successful
2830 * < 0 if the cgroup is over its limit
2832 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2833 gfp_t gfp_mask, enum charge_type ctype)
2835 struct mem_cgroup *memcg = NULL;
2836 unsigned int nr_pages = 1;
2840 if (PageTransHuge(page)) {
2841 nr_pages <<= compound_order(page);
2842 VM_BUG_ON(!PageTransHuge(page));
2844 * Never OOM-kill a process for a huge page. The
2845 * fault handler will fall back to regular pages.
2850 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
2853 __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false);
2857 int mem_cgroup_newpage_charge(struct page *page,
2858 struct mm_struct *mm, gfp_t gfp_mask)
2860 if (mem_cgroup_disabled())
2862 VM_BUG_ON(page_mapped(page));
2863 VM_BUG_ON(page->mapping && !PageAnon(page));
2865 return mem_cgroup_charge_common(page, mm, gfp_mask,
2866 MEM_CGROUP_CHARGE_TYPE_ANON);
2870 * While swap-in, try_charge -> commit or cancel, the page is locked.
2871 * And when try_charge() successfully returns, one refcnt to memcg without
2872 * struct page_cgroup is acquired. This refcnt will be consumed by
2873 * "commit()" or removed by "cancel()"
2875 static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2878 struct mem_cgroup **memcgp)
2880 struct mem_cgroup *memcg;
2881 struct page_cgroup *pc;
2884 pc = lookup_page_cgroup(page);
2886 * Every swap fault against a single page tries to charge the
2887 * page, bail as early as possible. shmem_unuse() encounters
2888 * already charged pages, too. The USED bit is protected by
2889 * the page lock, which serializes swap cache removal, which
2890 * in turn serializes uncharging.
2892 if (PageCgroupUsed(pc))
2894 if (!do_swap_account)
2896 memcg = try_get_mem_cgroup_from_page(page);
2900 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
2901 css_put(&memcg->css);
2906 ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
2912 int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page,
2913 gfp_t gfp_mask, struct mem_cgroup **memcgp)
2916 if (mem_cgroup_disabled())
2919 * A racing thread's fault, or swapoff, may have already
2920 * updated the pte, and even removed page from swap cache: in
2921 * those cases unuse_pte()'s pte_same() test will fail; but
2922 * there's also a KSM case which does need to charge the page.
2924 if (!PageSwapCache(page)) {
2927 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, memcgp, true);
2932 return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp);
2935 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
2937 if (mem_cgroup_disabled())
2941 __mem_cgroup_cancel_charge(memcg, 1);
2945 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
2946 enum charge_type ctype)
2948 if (mem_cgroup_disabled())
2953 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
2955 * Now swap is on-memory. This means this page may be
2956 * counted both as mem and swap....double count.
2957 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2958 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2959 * may call delete_from_swap_cache() before reach here.
2961 if (do_swap_account && PageSwapCache(page)) {
2962 swp_entry_t ent = {.val = page_private(page)};
2963 mem_cgroup_uncharge_swap(ent);
2967 void mem_cgroup_commit_charge_swapin(struct page *page,
2968 struct mem_cgroup *memcg)
2970 __mem_cgroup_commit_charge_swapin(page, memcg,
2971 MEM_CGROUP_CHARGE_TYPE_ANON);
2974 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2977 struct mem_cgroup *memcg = NULL;
2978 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
2981 if (mem_cgroup_disabled())
2983 if (PageCompound(page))
2986 if (!PageSwapCache(page))
2987 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
2988 else { /* page is swapcache/shmem */
2989 ret = __mem_cgroup_try_charge_swapin(mm, page,
2992 __mem_cgroup_commit_charge_swapin(page, memcg, type);
2997 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
2998 unsigned int nr_pages,
2999 const enum charge_type ctype)
3001 struct memcg_batch_info *batch = NULL;
3002 bool uncharge_memsw = true;
3004 /* If swapout, usage of swap doesn't decrease */
3005 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
3006 uncharge_memsw = false;
3008 batch = ¤t->memcg_batch;
3010 * In usual, we do css_get() when we remember memcg pointer.
3011 * But in this case, we keep res->usage until end of a series of
3012 * uncharges. Then, it's ok to ignore memcg's refcnt.
3015 batch->memcg = memcg;
3017 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
3018 * In those cases, all pages freed continuously can be expected to be in
3019 * the same cgroup and we have chance to coalesce uncharges.
3020 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
3021 * because we want to do uncharge as soon as possible.
3024 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
3025 goto direct_uncharge;
3028 goto direct_uncharge;
3031 * In typical case, batch->memcg == mem. This means we can
3032 * merge a series of uncharges to an uncharge of res_counter.
3033 * If not, we uncharge res_counter ony by one.
3035 if (batch->memcg != memcg)
3036 goto direct_uncharge;
3037 /* remember freed charge and uncharge it later */
3040 batch->memsw_nr_pages++;
3043 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
3045 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
3046 if (unlikely(batch->memcg != memcg))
3047 memcg_oom_recover(memcg);
3051 * uncharge if !page_mapped(page)
3053 static struct mem_cgroup *
3054 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
3057 struct mem_cgroup *memcg = NULL;
3058 unsigned int nr_pages = 1;
3059 struct page_cgroup *pc;
3062 if (mem_cgroup_disabled())
3065 VM_BUG_ON(PageSwapCache(page));
3067 if (PageTransHuge(page)) {
3068 nr_pages <<= compound_order(page);
3069 VM_BUG_ON(!PageTransHuge(page));
3072 * Check if our page_cgroup is valid
3074 pc = lookup_page_cgroup(page);
3075 if (unlikely(!PageCgroupUsed(pc)))
3078 lock_page_cgroup(pc);
3080 memcg = pc->mem_cgroup;
3082 if (!PageCgroupUsed(pc))
3085 anon = PageAnon(page);
3088 case MEM_CGROUP_CHARGE_TYPE_ANON:
3090 * Generally PageAnon tells if it's the anon statistics to be
3091 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
3092 * used before page reached the stage of being marked PageAnon.
3096 case MEM_CGROUP_CHARGE_TYPE_DROP:
3097 /* See mem_cgroup_prepare_migration() */
3098 if (page_mapped(page))
3101 * Pages under migration may not be uncharged. But
3102 * end_migration() /must/ be the one uncharging the
3103 * unused post-migration page and so it has to call
3104 * here with the migration bit still set. See the
3105 * res_counter handling below.
3107 if (!end_migration && PageCgroupMigration(pc))
3110 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
3111 if (!PageAnon(page)) { /* Shared memory */
3112 if (page->mapping && !page_is_file_cache(page))
3114 } else if (page_mapped(page)) /* Anon */
3121 mem_cgroup_charge_statistics(memcg, anon, -nr_pages);
3123 ClearPageCgroupUsed(pc);
3125 * pc->mem_cgroup is not cleared here. It will be accessed when it's
3126 * freed from LRU. This is safe because uncharged page is expected not
3127 * to be reused (freed soon). Exception is SwapCache, it's handled by
3128 * special functions.
3131 unlock_page_cgroup(pc);
3133 * even after unlock, we have memcg->res.usage here and this memcg
3134 * will never be freed.
3136 memcg_check_events(memcg, page);
3137 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
3138 mem_cgroup_swap_statistics(memcg, true);
3139 mem_cgroup_get(memcg);
3142 * Migration does not charge the res_counter for the
3143 * replacement page, so leave it alone when phasing out the
3144 * page that is unused after the migration.
3146 if (!end_migration && !mem_cgroup_is_root(memcg))
3147 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
3152 unlock_page_cgroup(pc);
3156 void mem_cgroup_uncharge_page(struct page *page)
3159 if (page_mapped(page))
3161 VM_BUG_ON(page->mapping && !PageAnon(page));
3162 if (PageSwapCache(page))
3164 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
3167 void mem_cgroup_uncharge_cache_page(struct page *page)
3169 VM_BUG_ON(page_mapped(page));
3170 VM_BUG_ON(page->mapping);
3171 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
3175 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
3176 * In that cases, pages are freed continuously and we can expect pages
3177 * are in the same memcg. All these calls itself limits the number of
3178 * pages freed at once, then uncharge_start/end() is called properly.
3179 * This may be called prural(2) times in a context,
3182 void mem_cgroup_uncharge_start(void)
3184 current->memcg_batch.do_batch++;
3185 /* We can do nest. */
3186 if (current->memcg_batch.do_batch == 1) {
3187 current->memcg_batch.memcg = NULL;
3188 current->memcg_batch.nr_pages = 0;
3189 current->memcg_batch.memsw_nr_pages = 0;
3193 void mem_cgroup_uncharge_end(void)
3195 struct memcg_batch_info *batch = ¤t->memcg_batch;
3197 if (!batch->do_batch)
3201 if (batch->do_batch) /* If stacked, do nothing. */
3207 * This "batch->memcg" is valid without any css_get/put etc...
3208 * bacause we hide charges behind us.
3210 if (batch->nr_pages)
3211 res_counter_uncharge(&batch->memcg->res,
3212 batch->nr_pages * PAGE_SIZE);
3213 if (batch->memsw_nr_pages)
3214 res_counter_uncharge(&batch->memcg->memsw,
3215 batch->memsw_nr_pages * PAGE_SIZE);
3216 memcg_oom_recover(batch->memcg);
3217 /* forget this pointer (for sanity check) */
3218 batch->memcg = NULL;
3223 * called after __delete_from_swap_cache() and drop "page" account.
3224 * memcg information is recorded to swap_cgroup of "ent"
3227 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
3229 struct mem_cgroup *memcg;
3230 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
3232 if (!swapout) /* this was a swap cache but the swap is unused ! */
3233 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
3235 memcg = __mem_cgroup_uncharge_common(page, ctype, false);
3238 * record memcg information, if swapout && memcg != NULL,
3239 * mem_cgroup_get() was called in uncharge().
3241 if (do_swap_account && swapout && memcg)
3242 swap_cgroup_record(ent, css_id(&memcg->css));
3246 #ifdef CONFIG_MEMCG_SWAP
3248 * called from swap_entry_free(). remove record in swap_cgroup and
3249 * uncharge "memsw" account.
3251 void mem_cgroup_uncharge_swap(swp_entry_t ent)
3253 struct mem_cgroup *memcg;
3256 if (!do_swap_account)
3259 id = swap_cgroup_record(ent, 0);
3261 memcg = mem_cgroup_lookup(id);
3264 * We uncharge this because swap is freed.
3265 * This memcg can be obsolete one. We avoid calling css_tryget
3267 if (!mem_cgroup_is_root(memcg))
3268 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
3269 mem_cgroup_swap_statistics(memcg, false);
3270 mem_cgroup_put(memcg);
3276 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3277 * @entry: swap entry to be moved
3278 * @from: mem_cgroup which the entry is moved from
3279 * @to: mem_cgroup which the entry is moved to
3281 * It succeeds only when the swap_cgroup's record for this entry is the same
3282 * as the mem_cgroup's id of @from.
3284 * Returns 0 on success, -EINVAL on failure.
3286 * The caller must have charged to @to, IOW, called res_counter_charge() about
3287 * both res and memsw, and called css_get().
3289 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3290 struct mem_cgroup *from, struct mem_cgroup *to)
3292 unsigned short old_id, new_id;
3294 old_id = css_id(&from->css);
3295 new_id = css_id(&to->css);
3297 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3298 mem_cgroup_swap_statistics(from, false);
3299 mem_cgroup_swap_statistics(to, true);
3301 * This function is only called from task migration context now.
3302 * It postpones res_counter and refcount handling till the end
3303 * of task migration(mem_cgroup_clear_mc()) for performance
3304 * improvement. But we cannot postpone mem_cgroup_get(to)
3305 * because if the process that has been moved to @to does
3306 * swap-in, the refcount of @to might be decreased to 0.
3314 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3315 struct mem_cgroup *from, struct mem_cgroup *to)
3322 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
3325 void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
3326 struct mem_cgroup **memcgp)
3328 struct mem_cgroup *memcg = NULL;
3329 unsigned int nr_pages = 1;
3330 struct page_cgroup *pc;
3331 enum charge_type ctype;
3335 if (mem_cgroup_disabled())
3338 if (PageTransHuge(page))
3339 nr_pages <<= compound_order(page);
3341 pc = lookup_page_cgroup(page);
3342 lock_page_cgroup(pc);
3343 if (PageCgroupUsed(pc)) {
3344 memcg = pc->mem_cgroup;
3345 css_get(&memcg->css);
3347 * At migrating an anonymous page, its mapcount goes down
3348 * to 0 and uncharge() will be called. But, even if it's fully
3349 * unmapped, migration may fail and this page has to be
3350 * charged again. We set MIGRATION flag here and delay uncharge
3351 * until end_migration() is called
3353 * Corner Case Thinking
3355 * When the old page was mapped as Anon and it's unmap-and-freed
3356 * while migration was ongoing.
3357 * If unmap finds the old page, uncharge() of it will be delayed
3358 * until end_migration(). If unmap finds a new page, it's
3359 * uncharged when it make mapcount to be 1->0. If unmap code
3360 * finds swap_migration_entry, the new page will not be mapped
3361 * and end_migration() will find it(mapcount==0).
3364 * When the old page was mapped but migraion fails, the kernel
3365 * remaps it. A charge for it is kept by MIGRATION flag even
3366 * if mapcount goes down to 0. We can do remap successfully
3367 * without charging it again.
3370 * The "old" page is under lock_page() until the end of
3371 * migration, so, the old page itself will not be swapped-out.
3372 * If the new page is swapped out before end_migraton, our
3373 * hook to usual swap-out path will catch the event.
3376 SetPageCgroupMigration(pc);
3378 unlock_page_cgroup(pc);
3380 * If the page is not charged at this point,
3388 * We charge new page before it's used/mapped. So, even if unlock_page()
3389 * is called before end_migration, we can catch all events on this new
3390 * page. In the case new page is migrated but not remapped, new page's
3391 * mapcount will be finally 0 and we call uncharge in end_migration().
3394 ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
3396 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
3398 * The page is committed to the memcg, but it's not actually
3399 * charged to the res_counter since we plan on replacing the
3400 * old one and only one page is going to be left afterwards.
3402 __mem_cgroup_commit_charge(memcg, newpage, nr_pages, ctype, false);
3405 /* remove redundant charge if migration failed*/
3406 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
3407 struct page *oldpage, struct page *newpage, bool migration_ok)
3409 struct page *used, *unused;
3410 struct page_cgroup *pc;
3416 if (!migration_ok) {
3423 anon = PageAnon(used);
3424 __mem_cgroup_uncharge_common(unused,
3425 anon ? MEM_CGROUP_CHARGE_TYPE_ANON
3426 : MEM_CGROUP_CHARGE_TYPE_CACHE,
3428 css_put(&memcg->css);
3430 * We disallowed uncharge of pages under migration because mapcount
3431 * of the page goes down to zero, temporarly.
3432 * Clear the flag and check the page should be charged.
3434 pc = lookup_page_cgroup(oldpage);
3435 lock_page_cgroup(pc);
3436 ClearPageCgroupMigration(pc);
3437 unlock_page_cgroup(pc);
3440 * If a page is a file cache, radix-tree replacement is very atomic
3441 * and we can skip this check. When it was an Anon page, its mapcount
3442 * goes down to 0. But because we added MIGRATION flage, it's not
3443 * uncharged yet. There are several case but page->mapcount check
3444 * and USED bit check in mem_cgroup_uncharge_page() will do enough
3445 * check. (see prepare_charge() also)
3448 mem_cgroup_uncharge_page(used);
3452 * At replace page cache, newpage is not under any memcg but it's on
3453 * LRU. So, this function doesn't touch res_counter but handles LRU
3454 * in correct way. Both pages are locked so we cannot race with uncharge.
3456 void mem_cgroup_replace_page_cache(struct page *oldpage,
3457 struct page *newpage)
3459 struct mem_cgroup *memcg = NULL;
3460 struct page_cgroup *pc;
3461 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
3463 if (mem_cgroup_disabled())
3466 pc = lookup_page_cgroup(oldpage);
3467 /* fix accounting on old pages */
3468 lock_page_cgroup(pc);
3469 if (PageCgroupUsed(pc)) {
3470 memcg = pc->mem_cgroup;
3471 mem_cgroup_charge_statistics(memcg, false, -1);
3472 ClearPageCgroupUsed(pc);
3474 unlock_page_cgroup(pc);
3477 * When called from shmem_replace_page(), in some cases the
3478 * oldpage has already been charged, and in some cases not.
3483 * Even if newpage->mapping was NULL before starting replacement,
3484 * the newpage may be on LRU(or pagevec for LRU) already. We lock
3485 * LRU while we overwrite pc->mem_cgroup.
3487 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
3490 #ifdef CONFIG_DEBUG_VM
3491 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3493 struct page_cgroup *pc;
3495 pc = lookup_page_cgroup(page);
3497 * Can be NULL while feeding pages into the page allocator for
3498 * the first time, i.e. during boot or memory hotplug;
3499 * or when mem_cgroup_disabled().
3501 if (likely(pc) && PageCgroupUsed(pc))
3506 bool mem_cgroup_bad_page_check(struct page *page)
3508 if (mem_cgroup_disabled())
3511 return lookup_page_cgroup_used(page) != NULL;
3514 void mem_cgroup_print_bad_page(struct page *page)
3516 struct page_cgroup *pc;
3518 pc = lookup_page_cgroup_used(page);
3520 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
3521 pc, pc->flags, pc->mem_cgroup);
3526 static DEFINE_MUTEX(set_limit_mutex);
3528 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3529 unsigned long long val)
3532 u64 memswlimit, memlimit;
3534 int children = mem_cgroup_count_children(memcg);
3535 u64 curusage, oldusage;
3539 * For keeping hierarchical_reclaim simple, how long we should retry
3540 * is depends on callers. We set our retry-count to be function
3541 * of # of children which we should visit in this loop.
3543 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3545 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3548 while (retry_count) {
3549 if (signal_pending(current)) {
3554 * Rather than hide all in some function, I do this in
3555 * open coded manner. You see what this really does.
3556 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
3558 mutex_lock(&set_limit_mutex);
3559 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3560 if (memswlimit < val) {
3562 mutex_unlock(&set_limit_mutex);
3566 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3570 ret = res_counter_set_limit(&memcg->res, val);
3572 if (memswlimit == val)
3573 memcg->memsw_is_minimum = true;
3575 memcg->memsw_is_minimum = false;
3577 mutex_unlock(&set_limit_mutex);
3582 mem_cgroup_reclaim(memcg, GFP_KERNEL,
3583 MEM_CGROUP_RECLAIM_SHRINK);
3584 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3585 /* Usage is reduced ? */
3586 if (curusage >= oldusage)
3589 oldusage = curusage;
3591 if (!ret && enlarge)
3592 memcg_oom_recover(memcg);
3597 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3598 unsigned long long val)
3601 u64 memlimit, memswlimit, oldusage, curusage;
3602 int children = mem_cgroup_count_children(memcg);
3606 /* see mem_cgroup_resize_res_limit */
3607 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3608 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3609 while (retry_count) {
3610 if (signal_pending(current)) {
3615 * Rather than hide all in some function, I do this in
3616 * open coded manner. You see what this really does.
3617 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
3619 mutex_lock(&set_limit_mutex);
3620 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3621 if (memlimit > val) {
3623 mutex_unlock(&set_limit_mutex);
3626 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3627 if (memswlimit < val)
3629 ret = res_counter_set_limit(&memcg->memsw, val);
3631 if (memlimit == val)
3632 memcg->memsw_is_minimum = true;
3634 memcg->memsw_is_minimum = false;
3636 mutex_unlock(&set_limit_mutex);
3641 mem_cgroup_reclaim(memcg, GFP_KERNEL,
3642 MEM_CGROUP_RECLAIM_NOSWAP |
3643 MEM_CGROUP_RECLAIM_SHRINK);
3644 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3645 /* Usage is reduced ? */
3646 if (curusage >= oldusage)
3649 oldusage = curusage;
3651 if (!ret && enlarge)
3652 memcg_oom_recover(memcg);
3656 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3658 unsigned long *total_scanned)
3660 unsigned long nr_reclaimed = 0;
3661 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3662 unsigned long reclaimed;
3664 struct mem_cgroup_tree_per_zone *mctz;
3665 unsigned long long excess;
3666 unsigned long nr_scanned;
3671 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3673 * This loop can run a while, specially if mem_cgroup's continuously
3674 * keep exceeding their soft limit and putting the system under
3681 mz = mem_cgroup_largest_soft_limit_node(mctz);
3686 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
3687 gfp_mask, &nr_scanned);
3688 nr_reclaimed += reclaimed;
3689 *total_scanned += nr_scanned;
3690 spin_lock(&mctz->lock);
3693 * If we failed to reclaim anything from this memory cgroup
3694 * it is time to move on to the next cgroup
3700 * Loop until we find yet another one.
3702 * By the time we get the soft_limit lock
3703 * again, someone might have aded the
3704 * group back on the RB tree. Iterate to
3705 * make sure we get a different mem.
3706 * mem_cgroup_largest_soft_limit_node returns
3707 * NULL if no other cgroup is present on
3711 __mem_cgroup_largest_soft_limit_node(mctz);
3713 css_put(&next_mz->memcg->css);
3714 else /* next_mz == NULL or other memcg */
3718 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
3719 excess = res_counter_soft_limit_excess(&mz->memcg->res);
3721 * One school of thought says that we should not add
3722 * back the node to the tree if reclaim returns 0.
3723 * But our reclaim could return 0, simply because due
3724 * to priority we are exposing a smaller subset of
3725 * memory to reclaim from. Consider this as a longer
3728 /* If excess == 0, no tree ops */
3729 __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess);
3730 spin_unlock(&mctz->lock);
3731 css_put(&mz->memcg->css);
3734 * Could not reclaim anything and there are no more
3735 * mem cgroups to try or we seem to be looping without
3736 * reclaiming anything.
3738 if (!nr_reclaimed &&
3740 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3742 } while (!nr_reclaimed);
3744 css_put(&next_mz->memcg->css);
3745 return nr_reclaimed;
3749 * mem_cgroup_force_empty_list - clears LRU of a group
3750 * @memcg: group to clear
3753 * @lru: lru to to clear
3755 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
3756 * reclaim the pages page themselves - pages are moved to the parent (or root)
3759 static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
3760 int node, int zid, enum lru_list lru)
3762 struct lruvec *lruvec;
3763 unsigned long flags;
3764 struct list_head *list;
3768 zone = &NODE_DATA(node)->node_zones[zid];
3769 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
3770 list = &lruvec->lists[lru];
3774 struct page_cgroup *pc;
3777 spin_lock_irqsave(&zone->lru_lock, flags);
3778 if (list_empty(list)) {
3779 spin_unlock_irqrestore(&zone->lru_lock, flags);
3782 page = list_entry(list->prev, struct page, lru);
3784 list_move(&page->lru, list);
3786 spin_unlock_irqrestore(&zone->lru_lock, flags);
3789 spin_unlock_irqrestore(&zone->lru_lock, flags);
3791 pc = lookup_page_cgroup(page);
3793 if (mem_cgroup_move_parent(page, pc, memcg)) {
3794 /* found lock contention or "pc" is obsolete. */
3799 } while (!list_empty(list));
3803 * make mem_cgroup's charge to be 0 if there is no task by moving
3804 * all the charges and pages to the parent.
3805 * This enables deleting this mem_cgroup.
3807 * Caller is responsible for holding css reference on the memcg.
3809 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
3814 /* This is for making all *used* pages to be on LRU. */
3815 lru_add_drain_all();
3816 drain_all_stock_sync(memcg);
3817 mem_cgroup_start_move(memcg);
3818 for_each_node_state(node, N_MEMORY) {
3819 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3822 mem_cgroup_force_empty_list(memcg,
3827 mem_cgroup_end_move(memcg);
3828 memcg_oom_recover(memcg);
3832 * This is a safety check because mem_cgroup_force_empty_list
3833 * could have raced with mem_cgroup_replace_page_cache callers
3834 * so the lru seemed empty but the page could have been added
3835 * right after the check. RES_USAGE should be safe as we always
3836 * charge before adding to the LRU.
3838 } while (res_counter_read_u64(&memcg->res, RES_USAGE) > 0);
3842 * Reclaims as many pages from the given memcg as possible and moves
3843 * the rest to the parent.
3845 * Caller is responsible for holding css reference for memcg.
3847 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3849 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3850 struct cgroup *cgrp = memcg->css.cgroup;
3852 /* returns EBUSY if there is a task or if we come here twice. */
3853 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3856 /* we call try-to-free pages for make this cgroup empty */
3857 lru_add_drain_all();
3858 /* try to free all pages in this cgroup */
3859 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
3862 if (signal_pending(current))
3865 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
3869 /* maybe some writeback is necessary */
3870 congestion_wait(BLK_RW_ASYNC, HZ/10);
3875 mem_cgroup_reparent_charges(memcg);
3880 static int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3882 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3885 if (mem_cgroup_is_root(memcg))
3887 css_get(&memcg->css);
3888 ret = mem_cgroup_force_empty(memcg);
3889 css_put(&memcg->css);
3895 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3897 return mem_cgroup_from_cont(cont)->use_hierarchy;
3900 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3904 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3905 struct cgroup *parent = cont->parent;
3906 struct mem_cgroup *parent_memcg = NULL;
3909 parent_memcg = mem_cgroup_from_cont(parent);
3913 if (memcg->use_hierarchy == val)
3917 * If parent's use_hierarchy is set, we can't make any modifications
3918 * in the child subtrees. If it is unset, then the change can
3919 * occur, provided the current cgroup has no children.
3921 * For the root cgroup, parent_mem is NULL, we allow value to be
3922 * set if there are no children.
3924 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3925 (val == 1 || val == 0)) {
3926 if (list_empty(&cont->children))
3927 memcg->use_hierarchy = val;
3940 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
3941 enum mem_cgroup_stat_index idx)
3943 struct mem_cgroup *iter;
3946 /* Per-cpu values can be negative, use a signed accumulator */
3947 for_each_mem_cgroup_tree(iter, memcg)
3948 val += mem_cgroup_read_stat(iter, idx);
3950 if (val < 0) /* race ? */
3955 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3959 if (!mem_cgroup_is_root(memcg)) {
3961 return res_counter_read_u64(&memcg->res, RES_USAGE);
3963 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
3966 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
3967 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
3970 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
3972 return val << PAGE_SHIFT;
3975 static ssize_t mem_cgroup_read(struct cgroup *cont, struct cftype *cft,
3976 struct file *file, char __user *buf,
3977 size_t nbytes, loff_t *ppos)
3979 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3985 type = MEMFILE_TYPE(cft->private);
3986 name = MEMFILE_ATTR(cft->private);
3988 if (!do_swap_account && type == _MEMSWAP)
3993 if (name == RES_USAGE)
3994 val = mem_cgroup_usage(memcg, false);
3996 val = res_counter_read_u64(&memcg->res, name);
3999 if (name == RES_USAGE)
4000 val = mem_cgroup_usage(memcg, true);
4002 val = res_counter_read_u64(&memcg->memsw, name);
4005 val = res_counter_read_u64(&memcg->kmem, name);
4011 len = scnprintf(str, sizeof(str), "%llu\n", (unsigned long long)val);
4012 return simple_read_from_buffer(buf, nbytes, ppos, str, len);
4015 static int memcg_update_kmem_limit(struct cgroup *cont, u64 val)
4018 #ifdef CONFIG_MEMCG_KMEM
4019 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4021 * For simplicity, we won't allow this to be disabled. It also can't
4022 * be changed if the cgroup has children already, or if tasks had
4025 * If tasks join before we set the limit, a person looking at
4026 * kmem.usage_in_bytes will have no way to determine when it took
4027 * place, which makes the value quite meaningless.
4029 * After it first became limited, changes in the value of the limit are
4030 * of course permitted.
4032 * Taking the cgroup_lock is really offensive, but it is so far the only
4033 * way to guarantee that no children will appear. There are plenty of
4034 * other offenders, and they should all go away. Fine grained locking
4035 * is probably the way to go here. When we are fully hierarchical, we
4036 * can also get rid of the use_hierarchy check.
4039 mutex_lock(&set_limit_mutex);
4040 if (!memcg->kmem_account_flags && val != RESOURCE_MAX) {
4041 if (cgroup_task_count(cont) || (memcg->use_hierarchy &&
4042 !list_empty(&cont->children))) {
4046 ret = res_counter_set_limit(&memcg->kmem, val);
4049 memcg_kmem_set_active(memcg);
4051 ret = res_counter_set_limit(&memcg->kmem, val);
4053 mutex_unlock(&set_limit_mutex);
4059 static void memcg_propagate_kmem(struct mem_cgroup *memcg)
4061 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4064 memcg->kmem_account_flags = parent->kmem_account_flags;
4068 * The user of this function is...
4071 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
4074 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4077 unsigned long long val;
4080 type = MEMFILE_TYPE(cft->private);
4081 name = MEMFILE_ATTR(cft->private);
4083 if (!do_swap_account && type == _MEMSWAP)
4088 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
4092 /* This function does all necessary parse...reuse it */
4093 ret = res_counter_memparse_write_strategy(buffer, &val);
4097 ret = mem_cgroup_resize_limit(memcg, val);
4098 else if (type == _MEMSWAP)
4099 ret = mem_cgroup_resize_memsw_limit(memcg, val);
4100 else if (type == _KMEM)
4101 ret = memcg_update_kmem_limit(cont, val);
4105 case RES_SOFT_LIMIT:
4106 ret = res_counter_memparse_write_strategy(buffer, &val);
4110 * For memsw, soft limits are hard to implement in terms
4111 * of semantics, for now, we support soft limits for
4112 * control without swap
4115 ret = res_counter_set_soft_limit(&memcg->res, val);
4120 ret = -EINVAL; /* should be BUG() ? */
4126 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
4127 unsigned long long *mem_limit, unsigned long long *memsw_limit)
4129 struct cgroup *cgroup;
4130 unsigned long long min_limit, min_memsw_limit, tmp;
4132 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4133 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4134 cgroup = memcg->css.cgroup;
4135 if (!memcg->use_hierarchy)
4138 while (cgroup->parent) {
4139 cgroup = cgroup->parent;
4140 memcg = mem_cgroup_from_cont(cgroup);
4141 if (!memcg->use_hierarchy)
4143 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
4144 min_limit = min(min_limit, tmp);
4145 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4146 min_memsw_limit = min(min_memsw_limit, tmp);
4149 *mem_limit = min_limit;
4150 *memsw_limit = min_memsw_limit;
4153 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
4155 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4159 type = MEMFILE_TYPE(event);
4160 name = MEMFILE_ATTR(event);
4162 if (!do_swap_account && type == _MEMSWAP)
4168 res_counter_reset_max(&memcg->res);
4169 else if (type == _MEMSWAP)
4170 res_counter_reset_max(&memcg->memsw);
4171 else if (type == _KMEM)
4172 res_counter_reset_max(&memcg->kmem);
4178 res_counter_reset_failcnt(&memcg->res);
4179 else if (type == _MEMSWAP)
4180 res_counter_reset_failcnt(&memcg->memsw);
4181 else if (type == _KMEM)
4182 res_counter_reset_failcnt(&memcg->kmem);
4191 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
4194 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
4198 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4199 struct cftype *cft, u64 val)
4201 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4203 if (val >= (1 << NR_MOVE_TYPE))
4206 * We check this value several times in both in can_attach() and
4207 * attach(), so we need cgroup lock to prevent this value from being
4211 memcg->move_charge_at_immigrate = val;
4217 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4218 struct cftype *cft, u64 val)
4225 static int memcg_numa_stat_show(struct cgroup *cont, struct cftype *cft,
4229 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
4230 unsigned long node_nr;
4231 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4233 total_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL);
4234 seq_printf(m, "total=%lu", total_nr);
4235 for_each_node_state(nid, N_MEMORY) {
4236 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL);
4237 seq_printf(m, " N%d=%lu", nid, node_nr);
4241 file_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_FILE);
4242 seq_printf(m, "file=%lu", file_nr);
4243 for_each_node_state(nid, N_MEMORY) {
4244 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4246 seq_printf(m, " N%d=%lu", nid, node_nr);
4250 anon_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_ANON);
4251 seq_printf(m, "anon=%lu", anon_nr);
4252 for_each_node_state(nid, N_MEMORY) {
4253 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4255 seq_printf(m, " N%d=%lu", nid, node_nr);
4259 unevictable_nr = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
4260 seq_printf(m, "unevictable=%lu", unevictable_nr);
4261 for_each_node_state(nid, N_MEMORY) {
4262 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4263 BIT(LRU_UNEVICTABLE));
4264 seq_printf(m, " N%d=%lu", nid, node_nr);
4269 #endif /* CONFIG_NUMA */
4271 static const char * const mem_cgroup_lru_names[] = {
4279 static inline void mem_cgroup_lru_names_not_uptodate(void)
4281 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
4284 static int memcg_stat_show(struct cgroup *cont, struct cftype *cft,
4287 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4288 struct mem_cgroup *mi;
4291 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
4292 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
4294 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
4295 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
4298 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
4299 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
4300 mem_cgroup_read_events(memcg, i));
4302 for (i = 0; i < NR_LRU_LISTS; i++)
4303 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
4304 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
4306 /* Hierarchical information */
4308 unsigned long long limit, memsw_limit;
4309 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
4310 seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
4311 if (do_swap_account)
4312 seq_printf(m, "hierarchical_memsw_limit %llu\n",
4316 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
4319 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
4321 for_each_mem_cgroup_tree(mi, memcg)
4322 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
4323 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
4326 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
4327 unsigned long long val = 0;
4329 for_each_mem_cgroup_tree(mi, memcg)
4330 val += mem_cgroup_read_events(mi, i);
4331 seq_printf(m, "total_%s %llu\n",
4332 mem_cgroup_events_names[i], val);
4335 for (i = 0; i < NR_LRU_LISTS; i++) {
4336 unsigned long long val = 0;
4338 for_each_mem_cgroup_tree(mi, memcg)
4339 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
4340 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
4343 #ifdef CONFIG_DEBUG_VM
4346 struct mem_cgroup_per_zone *mz;
4347 struct zone_reclaim_stat *rstat;
4348 unsigned long recent_rotated[2] = {0, 0};
4349 unsigned long recent_scanned[2] = {0, 0};
4351 for_each_online_node(nid)
4352 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4353 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
4354 rstat = &mz->lruvec.reclaim_stat;
4356 recent_rotated[0] += rstat->recent_rotated[0];
4357 recent_rotated[1] += rstat->recent_rotated[1];
4358 recent_scanned[0] += rstat->recent_scanned[0];
4359 recent_scanned[1] += rstat->recent_scanned[1];
4361 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
4362 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
4363 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
4364 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
4371 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
4373 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4375 return mem_cgroup_swappiness(memcg);
4378 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
4381 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4382 struct mem_cgroup *parent;
4387 if (cgrp->parent == NULL)
4390 parent = mem_cgroup_from_cont(cgrp->parent);
4394 /* If under hierarchy, only empty-root can set this value */
4395 if ((parent->use_hierarchy) ||
4396 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4401 memcg->swappiness = val;
4408 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4410 struct mem_cgroup_threshold_ary *t;
4416 t = rcu_dereference(memcg->thresholds.primary);
4418 t = rcu_dereference(memcg->memsw_thresholds.primary);
4423 usage = mem_cgroup_usage(memcg, swap);
4426 * current_threshold points to threshold just below or equal to usage.
4427 * If it's not true, a threshold was crossed after last
4428 * call of __mem_cgroup_threshold().
4430 i = t->current_threshold;
4433 * Iterate backward over array of thresholds starting from
4434 * current_threshold and check if a threshold is crossed.
4435 * If none of thresholds below usage is crossed, we read
4436 * only one element of the array here.
4438 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4439 eventfd_signal(t->entries[i].eventfd, 1);
4441 /* i = current_threshold + 1 */
4445 * Iterate forward over array of thresholds starting from
4446 * current_threshold+1 and check if a threshold is crossed.
4447 * If none of thresholds above usage is crossed, we read
4448 * only one element of the array here.
4450 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4451 eventfd_signal(t->entries[i].eventfd, 1);
4453 /* Update current_threshold */
4454 t->current_threshold = i - 1;
4459 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4462 __mem_cgroup_threshold(memcg, false);
4463 if (do_swap_account)
4464 __mem_cgroup_threshold(memcg, true);
4466 memcg = parent_mem_cgroup(memcg);
4470 static int compare_thresholds(const void *a, const void *b)
4472 const struct mem_cgroup_threshold *_a = a;
4473 const struct mem_cgroup_threshold *_b = b;
4475 return _a->threshold - _b->threshold;
4478 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4480 struct mem_cgroup_eventfd_list *ev;
4482 list_for_each_entry(ev, &memcg->oom_notify, list)
4483 eventfd_signal(ev->eventfd, 1);
4487 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4489 struct mem_cgroup *iter;
4491 for_each_mem_cgroup_tree(iter, memcg)
4492 mem_cgroup_oom_notify_cb(iter);
4495 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
4496 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4498 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4499 struct mem_cgroup_thresholds *thresholds;
4500 struct mem_cgroup_threshold_ary *new;
4501 enum res_type type = MEMFILE_TYPE(cft->private);
4502 u64 threshold, usage;
4505 ret = res_counter_memparse_write_strategy(args, &threshold);
4509 mutex_lock(&memcg->thresholds_lock);
4512 thresholds = &memcg->thresholds;
4513 else if (type == _MEMSWAP)
4514 thresholds = &memcg->memsw_thresholds;
4518 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4520 /* Check if a threshold crossed before adding a new one */
4521 if (thresholds->primary)
4522 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4524 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4526 /* Allocate memory for new array of thresholds */
4527 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4535 /* Copy thresholds (if any) to new array */
4536 if (thresholds->primary) {
4537 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4538 sizeof(struct mem_cgroup_threshold));
4541 /* Add new threshold */
4542 new->entries[size - 1].eventfd = eventfd;
4543 new->entries[size - 1].threshold = threshold;
4545 /* Sort thresholds. Registering of new threshold isn't time-critical */
4546 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4547 compare_thresholds, NULL);
4549 /* Find current threshold */
4550 new->current_threshold = -1;
4551 for (i = 0; i < size; i++) {
4552 if (new->entries[i].threshold <= usage) {
4554 * new->current_threshold will not be used until
4555 * rcu_assign_pointer(), so it's safe to increment
4558 ++new->current_threshold;
4563 /* Free old spare buffer and save old primary buffer as spare */
4564 kfree(thresholds->spare);
4565 thresholds->spare = thresholds->primary;
4567 rcu_assign_pointer(thresholds->primary, new);
4569 /* To be sure that nobody uses thresholds */
4573 mutex_unlock(&memcg->thresholds_lock);
4578 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
4579 struct cftype *cft, struct eventfd_ctx *eventfd)
4581 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4582 struct mem_cgroup_thresholds *thresholds;
4583 struct mem_cgroup_threshold_ary *new;
4584 enum res_type type = MEMFILE_TYPE(cft->private);
4588 mutex_lock(&memcg->thresholds_lock);
4590 thresholds = &memcg->thresholds;
4591 else if (type == _MEMSWAP)
4592 thresholds = &memcg->memsw_thresholds;
4596 if (!thresholds->primary)
4599 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4601 /* Check if a threshold crossed before removing */
4602 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4604 /* Calculate new number of threshold */
4606 for (i = 0; i < thresholds->primary->size; i++) {
4607 if (thresholds->primary->entries[i].eventfd != eventfd)
4611 new = thresholds->spare;
4613 /* Set thresholds array to NULL if we don't have thresholds */
4622 /* Copy thresholds and find current threshold */
4623 new->current_threshold = -1;
4624 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4625 if (thresholds->primary->entries[i].eventfd == eventfd)
4628 new->entries[j] = thresholds->primary->entries[i];
4629 if (new->entries[j].threshold <= usage) {
4631 * new->current_threshold will not be used
4632 * until rcu_assign_pointer(), so it's safe to increment
4635 ++new->current_threshold;
4641 /* Swap primary and spare array */
4642 thresholds->spare = thresholds->primary;
4643 /* If all events are unregistered, free the spare array */
4645 kfree(thresholds->spare);
4646 thresholds->spare = NULL;
4649 rcu_assign_pointer(thresholds->primary, new);
4651 /* To be sure that nobody uses thresholds */
4654 mutex_unlock(&memcg->thresholds_lock);
4657 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4658 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4660 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4661 struct mem_cgroup_eventfd_list *event;
4662 enum res_type type = MEMFILE_TYPE(cft->private);
4664 BUG_ON(type != _OOM_TYPE);
4665 event = kmalloc(sizeof(*event), GFP_KERNEL);
4669 spin_lock(&memcg_oom_lock);
4671 event->eventfd = eventfd;
4672 list_add(&event->list, &memcg->oom_notify);
4674 /* already in OOM ? */
4675 if (atomic_read(&memcg->under_oom))
4676 eventfd_signal(eventfd, 1);
4677 spin_unlock(&memcg_oom_lock);
4682 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4683 struct cftype *cft, struct eventfd_ctx *eventfd)
4685 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4686 struct mem_cgroup_eventfd_list *ev, *tmp;
4687 enum res_type type = MEMFILE_TYPE(cft->private);
4689 BUG_ON(type != _OOM_TYPE);
4691 spin_lock(&memcg_oom_lock);
4693 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4694 if (ev->eventfd == eventfd) {
4695 list_del(&ev->list);
4700 spin_unlock(&memcg_oom_lock);
4703 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4704 struct cftype *cft, struct cgroup_map_cb *cb)
4706 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4708 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
4710 if (atomic_read(&memcg->under_oom))
4711 cb->fill(cb, "under_oom", 1);
4713 cb->fill(cb, "under_oom", 0);
4717 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4718 struct cftype *cft, u64 val)
4720 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4721 struct mem_cgroup *parent;
4723 /* cannot set to root cgroup and only 0 and 1 are allowed */
4724 if (!cgrp->parent || !((val == 0) || (val == 1)))
4727 parent = mem_cgroup_from_cont(cgrp->parent);
4730 /* oom-kill-disable is a flag for subhierarchy. */
4731 if ((parent->use_hierarchy) ||
4732 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4736 memcg->oom_kill_disable = val;
4738 memcg_oom_recover(memcg);
4743 #ifdef CONFIG_MEMCG_KMEM
4744 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
4746 memcg_propagate_kmem(memcg);
4747 return mem_cgroup_sockets_init(memcg, ss);
4750 static void kmem_cgroup_destroy(struct mem_cgroup *memcg)
4752 mem_cgroup_sockets_destroy(memcg);
4755 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
4760 static void kmem_cgroup_destroy(struct mem_cgroup *memcg)
4765 static struct cftype mem_cgroup_files[] = {
4767 .name = "usage_in_bytes",
4768 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4769 .read = mem_cgroup_read,
4770 .register_event = mem_cgroup_usage_register_event,
4771 .unregister_event = mem_cgroup_usage_unregister_event,
4774 .name = "max_usage_in_bytes",
4775 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4776 .trigger = mem_cgroup_reset,
4777 .read = mem_cgroup_read,
4780 .name = "limit_in_bytes",
4781 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4782 .write_string = mem_cgroup_write,
4783 .read = mem_cgroup_read,
4786 .name = "soft_limit_in_bytes",
4787 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4788 .write_string = mem_cgroup_write,
4789 .read = mem_cgroup_read,
4793 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4794 .trigger = mem_cgroup_reset,
4795 .read = mem_cgroup_read,
4799 .read_seq_string = memcg_stat_show,
4802 .name = "force_empty",
4803 .trigger = mem_cgroup_force_empty_write,
4806 .name = "use_hierarchy",
4807 .write_u64 = mem_cgroup_hierarchy_write,
4808 .read_u64 = mem_cgroup_hierarchy_read,
4811 .name = "swappiness",
4812 .read_u64 = mem_cgroup_swappiness_read,
4813 .write_u64 = mem_cgroup_swappiness_write,
4816 .name = "move_charge_at_immigrate",
4817 .read_u64 = mem_cgroup_move_charge_read,
4818 .write_u64 = mem_cgroup_move_charge_write,
4821 .name = "oom_control",
4822 .read_map = mem_cgroup_oom_control_read,
4823 .write_u64 = mem_cgroup_oom_control_write,
4824 .register_event = mem_cgroup_oom_register_event,
4825 .unregister_event = mem_cgroup_oom_unregister_event,
4826 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4830 .name = "numa_stat",
4831 .read_seq_string = memcg_numa_stat_show,
4834 #ifdef CONFIG_MEMCG_SWAP
4836 .name = "memsw.usage_in_bytes",
4837 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4838 .read = mem_cgroup_read,
4839 .register_event = mem_cgroup_usage_register_event,
4840 .unregister_event = mem_cgroup_usage_unregister_event,
4843 .name = "memsw.max_usage_in_bytes",
4844 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4845 .trigger = mem_cgroup_reset,
4846 .read = mem_cgroup_read,
4849 .name = "memsw.limit_in_bytes",
4850 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4851 .write_string = mem_cgroup_write,
4852 .read = mem_cgroup_read,
4855 .name = "memsw.failcnt",
4856 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4857 .trigger = mem_cgroup_reset,
4858 .read = mem_cgroup_read,
4861 #ifdef CONFIG_MEMCG_KMEM
4863 .name = "kmem.limit_in_bytes",
4864 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4865 .write_string = mem_cgroup_write,
4866 .read = mem_cgroup_read,
4869 .name = "kmem.usage_in_bytes",
4870 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4871 .read = mem_cgroup_read,
4874 .name = "kmem.failcnt",
4875 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4876 .trigger = mem_cgroup_reset,
4877 .read = mem_cgroup_read,
4880 .name = "kmem.max_usage_in_bytes",
4881 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4882 .trigger = mem_cgroup_reset,
4883 .read = mem_cgroup_read,
4886 { }, /* terminate */
4889 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4891 struct mem_cgroup_per_node *pn;
4892 struct mem_cgroup_per_zone *mz;
4893 int zone, tmp = node;
4895 * This routine is called against possible nodes.
4896 * But it's BUG to call kmalloc() against offline node.
4898 * TODO: this routine can waste much memory for nodes which will
4899 * never be onlined. It's better to use memory hotplug callback
4902 if (!node_state(node, N_NORMAL_MEMORY))
4904 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4908 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4909 mz = &pn->zoneinfo[zone];
4910 lruvec_init(&mz->lruvec);
4911 mz->usage_in_excess = 0;
4912 mz->on_tree = false;
4915 memcg->info.nodeinfo[node] = pn;
4919 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4921 kfree(memcg->info.nodeinfo[node]);
4924 static struct mem_cgroup *mem_cgroup_alloc(void)
4926 struct mem_cgroup *memcg;
4927 int size = sizeof(struct mem_cgroup);
4929 /* Can be very big if MAX_NUMNODES is very big */
4930 if (size < PAGE_SIZE)
4931 memcg = kzalloc(size, GFP_KERNEL);
4933 memcg = vzalloc(size);
4938 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4941 spin_lock_init(&memcg->pcp_counter_lock);
4945 if (size < PAGE_SIZE)
4953 * Helpers for freeing a kmalloc()ed/vzalloc()ed mem_cgroup by RCU,
4954 * but in process context. The work_freeing structure is overlaid
4955 * on the rcu_freeing structure, which itself is overlaid on memsw.
4957 static void free_work(struct work_struct *work)
4959 struct mem_cgroup *memcg;
4960 int size = sizeof(struct mem_cgroup);
4962 memcg = container_of(work, struct mem_cgroup, work_freeing);
4964 * We need to make sure that (at least for now), the jump label
4965 * destruction code runs outside of the cgroup lock. This is because
4966 * get_online_cpus(), which is called from the static_branch update,
4967 * can't be called inside the cgroup_lock. cpusets are the ones
4968 * enforcing this dependency, so if they ever change, we might as well.
4970 * schedule_work() will guarantee this happens. Be careful if you need
4971 * to move this code around, and make sure it is outside
4974 disarm_sock_keys(memcg);
4975 if (size < PAGE_SIZE)
4981 static void free_rcu(struct rcu_head *rcu_head)
4983 struct mem_cgroup *memcg;
4985 memcg = container_of(rcu_head, struct mem_cgroup, rcu_freeing);
4986 INIT_WORK(&memcg->work_freeing, free_work);
4987 schedule_work(&memcg->work_freeing);
4991 * At destroying mem_cgroup, references from swap_cgroup can remain.
4992 * (scanning all at force_empty is too costly...)
4994 * Instead of clearing all references at force_empty, we remember
4995 * the number of reference from swap_cgroup and free mem_cgroup when
4996 * it goes down to 0.
4998 * Removal of cgroup itself succeeds regardless of refs from swap.
5001 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5005 mem_cgroup_remove_from_trees(memcg);
5006 free_css_id(&mem_cgroup_subsys, &memcg->css);
5009 free_mem_cgroup_per_zone_info(memcg, node);
5011 free_percpu(memcg->stat);
5012 call_rcu(&memcg->rcu_freeing, free_rcu);
5015 static void mem_cgroup_get(struct mem_cgroup *memcg)
5017 atomic_inc(&memcg->refcnt);
5020 static void __mem_cgroup_put(struct mem_cgroup *memcg, int count)
5022 if (atomic_sub_and_test(count, &memcg->refcnt)) {
5023 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5024 __mem_cgroup_free(memcg);
5026 mem_cgroup_put(parent);
5030 static void mem_cgroup_put(struct mem_cgroup *memcg)
5032 __mem_cgroup_put(memcg, 1);
5036 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
5038 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
5040 if (!memcg->res.parent)
5042 return mem_cgroup_from_res_counter(memcg->res.parent, res);
5044 EXPORT_SYMBOL(parent_mem_cgroup);
5046 #ifdef CONFIG_MEMCG_SWAP
5047 static void __init enable_swap_cgroup(void)
5049 if (!mem_cgroup_disabled() && really_do_swap_account)
5050 do_swap_account = 1;
5053 static void __init enable_swap_cgroup(void)
5058 static int mem_cgroup_soft_limit_tree_init(void)
5060 struct mem_cgroup_tree_per_node *rtpn;
5061 struct mem_cgroup_tree_per_zone *rtpz;
5062 int tmp, node, zone;
5064 for_each_node(node) {
5066 if (!node_state(node, N_NORMAL_MEMORY))
5068 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
5072 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5074 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5075 rtpz = &rtpn->rb_tree_per_zone[zone];
5076 rtpz->rb_root = RB_ROOT;
5077 spin_lock_init(&rtpz->lock);
5083 for_each_node(node) {
5084 if (!soft_limit_tree.rb_tree_per_node[node])
5086 kfree(soft_limit_tree.rb_tree_per_node[node]);
5087 soft_limit_tree.rb_tree_per_node[node] = NULL;
5093 static struct cgroup_subsys_state * __ref
5094 mem_cgroup_css_alloc(struct cgroup *cont)
5096 struct mem_cgroup *memcg, *parent;
5097 long error = -ENOMEM;
5100 memcg = mem_cgroup_alloc();
5102 return ERR_PTR(error);
5105 if (alloc_mem_cgroup_per_zone_info(memcg, node))
5109 if (cont->parent == NULL) {
5111 enable_swap_cgroup();
5113 if (mem_cgroup_soft_limit_tree_init())
5115 root_mem_cgroup = memcg;
5116 for_each_possible_cpu(cpu) {
5117 struct memcg_stock_pcp *stock =
5118 &per_cpu(memcg_stock, cpu);
5119 INIT_WORK(&stock->work, drain_local_stock);
5121 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5123 parent = mem_cgroup_from_cont(cont->parent);
5124 memcg->use_hierarchy = parent->use_hierarchy;
5125 memcg->oom_kill_disable = parent->oom_kill_disable;
5128 if (parent && parent->use_hierarchy) {
5129 res_counter_init(&memcg->res, &parent->res);
5130 res_counter_init(&memcg->memsw, &parent->memsw);
5131 res_counter_init(&memcg->kmem, &parent->kmem);
5133 * We increment refcnt of the parent to ensure that we can
5134 * safely access it on res_counter_charge/uncharge.
5135 * This refcnt will be decremented when freeing this
5136 * mem_cgroup(see mem_cgroup_put).
5138 mem_cgroup_get(parent);
5140 res_counter_init(&memcg->res, NULL);
5141 res_counter_init(&memcg->memsw, NULL);
5142 res_counter_init(&memcg->kmem, NULL);
5144 * Deeper hierachy with use_hierarchy == false doesn't make
5145 * much sense so let cgroup subsystem know about this
5146 * unfortunate state in our controller.
5148 if (parent && parent != root_mem_cgroup)
5149 mem_cgroup_subsys.broken_hierarchy = true;
5151 memcg->last_scanned_node = MAX_NUMNODES;
5152 INIT_LIST_HEAD(&memcg->oom_notify);
5155 memcg->swappiness = mem_cgroup_swappiness(parent);
5156 atomic_set(&memcg->refcnt, 1);
5157 memcg->move_charge_at_immigrate = 0;
5158 mutex_init(&memcg->thresholds_lock);
5159 spin_lock_init(&memcg->move_lock);
5161 error = memcg_init_kmem(memcg, &mem_cgroup_subsys);
5164 * We call put now because our (and parent's) refcnts
5165 * are already in place. mem_cgroup_put() will internally
5166 * call __mem_cgroup_free, so return directly
5168 mem_cgroup_put(memcg);
5169 return ERR_PTR(error);
5173 __mem_cgroup_free(memcg);
5174 return ERR_PTR(error);
5177 static void mem_cgroup_css_offline(struct cgroup *cont)
5179 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5181 mem_cgroup_reparent_charges(memcg);
5184 static void mem_cgroup_css_free(struct cgroup *cont)
5186 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5188 kmem_cgroup_destroy(memcg);
5190 mem_cgroup_put(memcg);
5194 /* Handlers for move charge at task migration. */
5195 #define PRECHARGE_COUNT_AT_ONCE 256
5196 static int mem_cgroup_do_precharge(unsigned long count)
5199 int batch_count = PRECHARGE_COUNT_AT_ONCE;
5200 struct mem_cgroup *memcg = mc.to;
5202 if (mem_cgroup_is_root(memcg)) {
5203 mc.precharge += count;
5204 /* we don't need css_get for root */
5207 /* try to charge at once */
5209 struct res_counter *dummy;
5211 * "memcg" cannot be under rmdir() because we've already checked
5212 * by cgroup_lock_live_cgroup() that it is not removed and we
5213 * are still under the same cgroup_mutex. So we can postpone
5216 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
5218 if (do_swap_account && res_counter_charge(&memcg->memsw,
5219 PAGE_SIZE * count, &dummy)) {
5220 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
5223 mc.precharge += count;
5227 /* fall back to one by one charge */
5229 if (signal_pending(current)) {
5233 if (!batch_count--) {
5234 batch_count = PRECHARGE_COUNT_AT_ONCE;
5237 ret = __mem_cgroup_try_charge(NULL,
5238 GFP_KERNEL, 1, &memcg, false);
5240 /* mem_cgroup_clear_mc() will do uncharge later */
5248 * get_mctgt_type - get target type of moving charge
5249 * @vma: the vma the pte to be checked belongs
5250 * @addr: the address corresponding to the pte to be checked
5251 * @ptent: the pte to be checked
5252 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5255 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5256 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5257 * move charge. if @target is not NULL, the page is stored in target->page
5258 * with extra refcnt got(Callers should handle it).
5259 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5260 * target for charge migration. if @target is not NULL, the entry is stored
5263 * Called with pte lock held.
5270 enum mc_target_type {
5276 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5277 unsigned long addr, pte_t ptent)
5279 struct page *page = vm_normal_page(vma, addr, ptent);
5281 if (!page || !page_mapped(page))
5283 if (PageAnon(page)) {
5284 /* we don't move shared anon */
5287 } else if (!move_file())
5288 /* we ignore mapcount for file pages */
5290 if (!get_page_unless_zero(page))
5297 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5298 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5300 struct page *page = NULL;
5301 swp_entry_t ent = pte_to_swp_entry(ptent);
5303 if (!move_anon() || non_swap_entry(ent))
5306 * Because lookup_swap_cache() updates some statistics counter,
5307 * we call find_get_page() with swapper_space directly.
5309 page = find_get_page(&swapper_space, ent.val);
5310 if (do_swap_account)
5311 entry->val = ent.val;
5316 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5317 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5323 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5324 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5326 struct page *page = NULL;
5327 struct address_space *mapping;
5330 if (!vma->vm_file) /* anonymous vma */
5335 mapping = vma->vm_file->f_mapping;
5336 if (pte_none(ptent))
5337 pgoff = linear_page_index(vma, addr);
5338 else /* pte_file(ptent) is true */
5339 pgoff = pte_to_pgoff(ptent);
5341 /* page is moved even if it's not RSS of this task(page-faulted). */
5342 page = find_get_page(mapping, pgoff);
5345 /* shmem/tmpfs may report page out on swap: account for that too. */
5346 if (radix_tree_exceptional_entry(page)) {
5347 swp_entry_t swap = radix_to_swp_entry(page);
5348 if (do_swap_account)
5350 page = find_get_page(&swapper_space, swap.val);
5356 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5357 unsigned long addr, pte_t ptent, union mc_target *target)
5359 struct page *page = NULL;
5360 struct page_cgroup *pc;
5361 enum mc_target_type ret = MC_TARGET_NONE;
5362 swp_entry_t ent = { .val = 0 };
5364 if (pte_present(ptent))
5365 page = mc_handle_present_pte(vma, addr, ptent);
5366 else if (is_swap_pte(ptent))
5367 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5368 else if (pte_none(ptent) || pte_file(ptent))
5369 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5371 if (!page && !ent.val)
5374 pc = lookup_page_cgroup(page);
5376 * Do only loose check w/o page_cgroup lock.
5377 * mem_cgroup_move_account() checks the pc is valid or not under
5380 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5381 ret = MC_TARGET_PAGE;
5383 target->page = page;
5385 if (!ret || !target)
5388 /* There is a swap entry and a page doesn't exist or isn't charged */
5389 if (ent.val && !ret &&
5390 css_id(&mc.from->css) == lookup_swap_cgroup_id(ent)) {
5391 ret = MC_TARGET_SWAP;
5398 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5400 * We don't consider swapping or file mapped pages because THP does not
5401 * support them for now.
5402 * Caller should make sure that pmd_trans_huge(pmd) is true.
5404 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5405 unsigned long addr, pmd_t pmd, union mc_target *target)
5407 struct page *page = NULL;
5408 struct page_cgroup *pc;
5409 enum mc_target_type ret = MC_TARGET_NONE;
5411 page = pmd_page(pmd);
5412 VM_BUG_ON(!page || !PageHead(page));
5415 pc = lookup_page_cgroup(page);
5416 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5417 ret = MC_TARGET_PAGE;
5420 target->page = page;
5426 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5427 unsigned long addr, pmd_t pmd, union mc_target *target)
5429 return MC_TARGET_NONE;
5433 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5434 unsigned long addr, unsigned long end,
5435 struct mm_walk *walk)
5437 struct vm_area_struct *vma = walk->private;
5441 if (pmd_trans_huge_lock(pmd, vma) == 1) {
5442 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5443 mc.precharge += HPAGE_PMD_NR;
5444 spin_unlock(&vma->vm_mm->page_table_lock);
5448 if (pmd_trans_unstable(pmd))
5450 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5451 for (; addr != end; pte++, addr += PAGE_SIZE)
5452 if (get_mctgt_type(vma, addr, *pte, NULL))
5453 mc.precharge++; /* increment precharge temporarily */
5454 pte_unmap_unlock(pte - 1, ptl);
5460 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5462 unsigned long precharge;
5463 struct vm_area_struct *vma;
5465 down_read(&mm->mmap_sem);
5466 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5467 struct mm_walk mem_cgroup_count_precharge_walk = {
5468 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5472 if (is_vm_hugetlb_page(vma))
5474 walk_page_range(vma->vm_start, vma->vm_end,
5475 &mem_cgroup_count_precharge_walk);
5477 up_read(&mm->mmap_sem);
5479 precharge = mc.precharge;
5485 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5487 unsigned long precharge = mem_cgroup_count_precharge(mm);
5489 VM_BUG_ON(mc.moving_task);
5490 mc.moving_task = current;
5491 return mem_cgroup_do_precharge(precharge);
5494 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5495 static void __mem_cgroup_clear_mc(void)
5497 struct mem_cgroup *from = mc.from;
5498 struct mem_cgroup *to = mc.to;
5500 /* we must uncharge all the leftover precharges from mc.to */
5502 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
5506 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5507 * we must uncharge here.
5509 if (mc.moved_charge) {
5510 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
5511 mc.moved_charge = 0;
5513 /* we must fixup refcnts and charges */
5514 if (mc.moved_swap) {
5515 /* uncharge swap account from the old cgroup */
5516 if (!mem_cgroup_is_root(mc.from))
5517 res_counter_uncharge(&mc.from->memsw,
5518 PAGE_SIZE * mc.moved_swap);
5519 __mem_cgroup_put(mc.from, mc.moved_swap);
5521 if (!mem_cgroup_is_root(mc.to)) {
5523 * we charged both to->res and to->memsw, so we should
5526 res_counter_uncharge(&mc.to->res,
5527 PAGE_SIZE * mc.moved_swap);
5529 /* we've already done mem_cgroup_get(mc.to) */
5532 memcg_oom_recover(from);
5533 memcg_oom_recover(to);
5534 wake_up_all(&mc.waitq);
5537 static void mem_cgroup_clear_mc(void)
5539 struct mem_cgroup *from = mc.from;
5542 * we must clear moving_task before waking up waiters at the end of
5545 mc.moving_task = NULL;
5546 __mem_cgroup_clear_mc();
5547 spin_lock(&mc.lock);
5550 spin_unlock(&mc.lock);
5551 mem_cgroup_end_move(from);
5554 static int mem_cgroup_can_attach(struct cgroup *cgroup,
5555 struct cgroup_taskset *tset)
5557 struct task_struct *p = cgroup_taskset_first(tset);
5559 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup);
5561 if (memcg->move_charge_at_immigrate) {
5562 struct mm_struct *mm;
5563 struct mem_cgroup *from = mem_cgroup_from_task(p);
5565 VM_BUG_ON(from == memcg);
5567 mm = get_task_mm(p);
5570 /* We move charges only when we move a owner of the mm */
5571 if (mm->owner == p) {
5574 VM_BUG_ON(mc.precharge);
5575 VM_BUG_ON(mc.moved_charge);
5576 VM_BUG_ON(mc.moved_swap);
5577 mem_cgroup_start_move(from);
5578 spin_lock(&mc.lock);
5581 spin_unlock(&mc.lock);
5582 /* We set mc.moving_task later */
5584 ret = mem_cgroup_precharge_mc(mm);
5586 mem_cgroup_clear_mc();
5593 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
5594 struct cgroup_taskset *tset)
5596 mem_cgroup_clear_mc();
5599 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5600 unsigned long addr, unsigned long end,
5601 struct mm_walk *walk)
5604 struct vm_area_struct *vma = walk->private;
5607 enum mc_target_type target_type;
5608 union mc_target target;
5610 struct page_cgroup *pc;
5613 * We don't take compound_lock() here but no race with splitting thp
5615 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
5616 * under splitting, which means there's no concurrent thp split,
5617 * - if another thread runs into split_huge_page() just after we
5618 * entered this if-block, the thread must wait for page table lock
5619 * to be unlocked in __split_huge_page_splitting(), where the main
5620 * part of thp split is not executed yet.
5622 if (pmd_trans_huge_lock(pmd, vma) == 1) {
5623 if (mc.precharge < HPAGE_PMD_NR) {
5624 spin_unlock(&vma->vm_mm->page_table_lock);
5627 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5628 if (target_type == MC_TARGET_PAGE) {
5630 if (!isolate_lru_page(page)) {
5631 pc = lookup_page_cgroup(page);
5632 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
5633 pc, mc.from, mc.to)) {
5634 mc.precharge -= HPAGE_PMD_NR;
5635 mc.moved_charge += HPAGE_PMD_NR;
5637 putback_lru_page(page);
5641 spin_unlock(&vma->vm_mm->page_table_lock);
5645 if (pmd_trans_unstable(pmd))
5648 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5649 for (; addr != end; addr += PAGE_SIZE) {
5650 pte_t ptent = *(pte++);
5656 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5657 case MC_TARGET_PAGE:
5659 if (isolate_lru_page(page))
5661 pc = lookup_page_cgroup(page);
5662 if (!mem_cgroup_move_account(page, 1, pc,
5665 /* we uncharge from mc.from later. */
5668 putback_lru_page(page);
5669 put: /* get_mctgt_type() gets the page */
5672 case MC_TARGET_SWAP:
5674 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5676 /* we fixup refcnts and charges later. */
5684 pte_unmap_unlock(pte - 1, ptl);
5689 * We have consumed all precharges we got in can_attach().
5690 * We try charge one by one, but don't do any additional
5691 * charges to mc.to if we have failed in charge once in attach()
5694 ret = mem_cgroup_do_precharge(1);
5702 static void mem_cgroup_move_charge(struct mm_struct *mm)
5704 struct vm_area_struct *vma;
5706 lru_add_drain_all();
5708 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5710 * Someone who are holding the mmap_sem might be waiting in
5711 * waitq. So we cancel all extra charges, wake up all waiters,
5712 * and retry. Because we cancel precharges, we might not be able
5713 * to move enough charges, but moving charge is a best-effort
5714 * feature anyway, so it wouldn't be a big problem.
5716 __mem_cgroup_clear_mc();
5720 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5722 struct mm_walk mem_cgroup_move_charge_walk = {
5723 .pmd_entry = mem_cgroup_move_charge_pte_range,
5727 if (is_vm_hugetlb_page(vma))
5729 ret = walk_page_range(vma->vm_start, vma->vm_end,
5730 &mem_cgroup_move_charge_walk);
5733 * means we have consumed all precharges and failed in
5734 * doing additional charge. Just abandon here.
5738 up_read(&mm->mmap_sem);
5741 static void mem_cgroup_move_task(struct cgroup *cont,
5742 struct cgroup_taskset *tset)
5744 struct task_struct *p = cgroup_taskset_first(tset);
5745 struct mm_struct *mm = get_task_mm(p);
5749 mem_cgroup_move_charge(mm);
5753 mem_cgroup_clear_mc();
5755 #else /* !CONFIG_MMU */
5756 static int mem_cgroup_can_attach(struct cgroup *cgroup,
5757 struct cgroup_taskset *tset)
5761 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
5762 struct cgroup_taskset *tset)
5765 static void mem_cgroup_move_task(struct cgroup *cont,
5766 struct cgroup_taskset *tset)
5771 struct cgroup_subsys mem_cgroup_subsys = {
5773 .subsys_id = mem_cgroup_subsys_id,
5774 .css_alloc = mem_cgroup_css_alloc,
5775 .css_offline = mem_cgroup_css_offline,
5776 .css_free = mem_cgroup_css_free,
5777 .can_attach = mem_cgroup_can_attach,
5778 .cancel_attach = mem_cgroup_cancel_attach,
5779 .attach = mem_cgroup_move_task,
5780 .base_cftypes = mem_cgroup_files,
5785 #ifdef CONFIG_MEMCG_SWAP
5786 static int __init enable_swap_account(char *s)
5788 /* consider enabled if no parameter or 1 is given */
5789 if (!strcmp(s, "1"))
5790 really_do_swap_account = 1;
5791 else if (!strcmp(s, "0"))
5792 really_do_swap_account = 0;
5795 __setup("swapaccount=", enable_swap_account);