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 #define MEM_CGROUP_RECLAIM_RETRIES 5
63 static struct mem_cgroup *root_mem_cgroup __read_mostly;
65 #ifdef CONFIG_MEMCG_SWAP
66 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
67 int do_swap_account __read_mostly;
69 /* for remember boot option*/
70 #ifdef CONFIG_MEMCG_SWAP_ENABLED
71 static int really_do_swap_account __initdata = 1;
73 static int really_do_swap_account __initdata = 0;
77 #define do_swap_account 0
82 * Statistics for memory cgroup.
84 enum mem_cgroup_stat_index {
86 * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
88 MEM_CGROUP_STAT_CACHE, /* # of pages charged as cache */
89 MEM_CGROUP_STAT_RSS, /* # of pages charged as anon rss */
90 MEM_CGROUP_STAT_FILE_MAPPED, /* # of pages charged as file rss */
91 MEM_CGROUP_STAT_SWAP, /* # of pages, swapped out */
92 MEM_CGROUP_STAT_NSTATS,
95 static const char * const mem_cgroup_stat_names[] = {
102 enum mem_cgroup_events_index {
103 MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */
104 MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */
105 MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */
106 MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */
107 MEM_CGROUP_EVENTS_NSTATS,
110 static const char * const mem_cgroup_events_names[] = {
118 * Per memcg event counter is incremented at every pagein/pageout. With THP,
119 * it will be incremated by the number of pages. This counter is used for
120 * for trigger some periodic events. This is straightforward and better
121 * than using jiffies etc. to handle periodic memcg event.
123 enum mem_cgroup_events_target {
124 MEM_CGROUP_TARGET_THRESH,
125 MEM_CGROUP_TARGET_SOFTLIMIT,
126 MEM_CGROUP_TARGET_NUMAINFO,
129 #define THRESHOLDS_EVENTS_TARGET 128
130 #define SOFTLIMIT_EVENTS_TARGET 1024
131 #define NUMAINFO_EVENTS_TARGET 1024
133 struct mem_cgroup_stat_cpu {
134 long count[MEM_CGROUP_STAT_NSTATS];
135 unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
136 unsigned long nr_page_events;
137 unsigned long targets[MEM_CGROUP_NTARGETS];
140 struct mem_cgroup_reclaim_iter {
141 /* css_id of the last scanned hierarchy member */
143 /* scan generation, increased every round-trip */
144 unsigned int generation;
148 * per-zone information in memory controller.
150 struct mem_cgroup_per_zone {
151 struct lruvec lruvec;
152 unsigned long lru_size[NR_LRU_LISTS];
154 struct mem_cgroup_reclaim_iter reclaim_iter[DEF_PRIORITY + 1];
156 struct rb_node tree_node; /* RB tree node */
157 unsigned long long usage_in_excess;/* Set to the value by which */
158 /* the soft limit is exceeded*/
160 struct mem_cgroup *memcg; /* Back pointer, we cannot */
161 /* use container_of */
164 struct mem_cgroup_per_node {
165 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
168 struct mem_cgroup_lru_info {
169 struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
173 * Cgroups above their limits are maintained in a RB-Tree, independent of
174 * their hierarchy representation
177 struct mem_cgroup_tree_per_zone {
178 struct rb_root rb_root;
182 struct mem_cgroup_tree_per_node {
183 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
186 struct mem_cgroup_tree {
187 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
190 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
192 struct mem_cgroup_threshold {
193 struct eventfd_ctx *eventfd;
198 struct mem_cgroup_threshold_ary {
199 /* An array index points to threshold just below or equal to usage. */
200 int current_threshold;
201 /* Size of entries[] */
203 /* Array of thresholds */
204 struct mem_cgroup_threshold entries[0];
207 struct mem_cgroup_thresholds {
208 /* Primary thresholds array */
209 struct mem_cgroup_threshold_ary *primary;
211 * Spare threshold array.
212 * This is needed to make mem_cgroup_unregister_event() "never fail".
213 * It must be able to store at least primary->size - 1 entries.
215 struct mem_cgroup_threshold_ary *spare;
219 struct mem_cgroup_eventfd_list {
220 struct list_head list;
221 struct eventfd_ctx *eventfd;
224 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
225 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
228 * The memory controller data structure. The memory controller controls both
229 * page cache and RSS per cgroup. We would eventually like to provide
230 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
231 * to help the administrator determine what knobs to tune.
233 * TODO: Add a water mark for the memory controller. Reclaim will begin when
234 * we hit the water mark. May be even add a low water mark, such that
235 * no reclaim occurs from a cgroup at it's low water mark, this is
236 * a feature that will be implemented much later in the future.
239 struct cgroup_subsys_state css;
241 * the counter to account for memory usage
243 struct res_counter res;
247 * the counter to account for mem+swap usage.
249 struct res_counter memsw;
252 * rcu_freeing is used only when freeing struct mem_cgroup,
253 * so put it into a union to avoid wasting more memory.
254 * It must be disjoint from the css field. It could be
255 * in a union with the res field, but res plays a much
256 * larger part in mem_cgroup life than memsw, and might
257 * be of interest, even at time of free, when debugging.
258 * So share rcu_head with the less interesting memsw.
260 struct rcu_head rcu_freeing;
262 * We also need some space for a worker in deferred freeing.
263 * By the time we call it, rcu_freeing is no longer in use.
265 struct work_struct work_freeing;
269 * Per cgroup active and inactive list, similar to the
270 * per zone LRU lists.
272 struct mem_cgroup_lru_info info;
273 int last_scanned_node;
275 nodemask_t scan_nodes;
276 atomic_t numainfo_events;
277 atomic_t numainfo_updating;
280 * Should the accounting and control be hierarchical, per subtree?
290 /* OOM-Killer disable */
291 int oom_kill_disable;
293 /* set when res.limit == memsw.limit */
294 bool memsw_is_minimum;
296 /* protect arrays of thresholds */
297 struct mutex thresholds_lock;
299 /* thresholds for memory usage. RCU-protected */
300 struct mem_cgroup_thresholds thresholds;
302 /* thresholds for mem+swap usage. RCU-protected */
303 struct mem_cgroup_thresholds memsw_thresholds;
305 /* For oom notifier event fd */
306 struct list_head oom_notify;
309 * Should we move charges of a task when a task is moved into this
310 * mem_cgroup ? And what type of charges should we move ?
312 unsigned long move_charge_at_immigrate;
314 * set > 0 if pages under this cgroup are moving to other cgroup.
316 atomic_t moving_account;
317 /* taken only while moving_account > 0 */
318 spinlock_t move_lock;
322 struct mem_cgroup_stat_cpu __percpu *stat;
324 * used when a cpu is offlined or other synchronizations
325 * See mem_cgroup_read_stat().
327 struct mem_cgroup_stat_cpu nocpu_base;
328 spinlock_t pcp_counter_lock;
330 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
331 struct tcp_memcontrol tcp_mem;
335 /* Stuffs for move charges at task migration. */
337 * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
338 * left-shifted bitmap of these types.
341 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
342 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
346 /* "mc" and its members are protected by cgroup_mutex */
347 static struct move_charge_struct {
348 spinlock_t lock; /* for from, to */
349 struct mem_cgroup *from;
350 struct mem_cgroup *to;
351 unsigned long precharge;
352 unsigned long moved_charge;
353 unsigned long moved_swap;
354 struct task_struct *moving_task; /* a task moving charges */
355 wait_queue_head_t waitq; /* a waitq for other context */
357 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
358 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
361 static bool move_anon(void)
363 return test_bit(MOVE_CHARGE_TYPE_ANON,
364 &mc.to->move_charge_at_immigrate);
367 static bool move_file(void)
369 return test_bit(MOVE_CHARGE_TYPE_FILE,
370 &mc.to->move_charge_at_immigrate);
374 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
375 * limit reclaim to prevent infinite loops, if they ever occur.
377 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
378 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
381 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
382 MEM_CGROUP_CHARGE_TYPE_ANON,
383 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
384 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
388 /* for encoding cft->private value on file */
391 #define _OOM_TYPE (2)
392 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
393 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
394 #define MEMFILE_ATTR(val) ((val) & 0xffff)
395 /* Used for OOM nofiier */
396 #define OOM_CONTROL (0)
399 * Reclaim flags for mem_cgroup_hierarchical_reclaim
401 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
402 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
403 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
404 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
406 static void mem_cgroup_get(struct mem_cgroup *memcg);
407 static void mem_cgroup_put(struct mem_cgroup *memcg);
410 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
412 return container_of(s, struct mem_cgroup, css);
415 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
417 return (memcg == root_mem_cgroup);
420 /* Writing them here to avoid exposing memcg's inner layout */
421 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
423 void sock_update_memcg(struct sock *sk)
425 if (mem_cgroup_sockets_enabled) {
426 struct mem_cgroup *memcg;
427 struct cg_proto *cg_proto;
429 BUG_ON(!sk->sk_prot->proto_cgroup);
431 /* Socket cloning can throw us here with sk_cgrp already
432 * filled. It won't however, necessarily happen from
433 * process context. So the test for root memcg given
434 * the current task's memcg won't help us in this case.
436 * Respecting the original socket's memcg is a better
437 * decision in this case.
440 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
441 mem_cgroup_get(sk->sk_cgrp->memcg);
446 memcg = mem_cgroup_from_task(current);
447 cg_proto = sk->sk_prot->proto_cgroup(memcg);
448 if (!mem_cgroup_is_root(memcg) && memcg_proto_active(cg_proto)) {
449 mem_cgroup_get(memcg);
450 sk->sk_cgrp = cg_proto;
455 EXPORT_SYMBOL(sock_update_memcg);
457 void sock_release_memcg(struct sock *sk)
459 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
460 struct mem_cgroup *memcg;
461 WARN_ON(!sk->sk_cgrp->memcg);
462 memcg = sk->sk_cgrp->memcg;
463 mem_cgroup_put(memcg);
467 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
469 if (!memcg || mem_cgroup_is_root(memcg))
472 return &memcg->tcp_mem.cg_proto;
474 EXPORT_SYMBOL(tcp_proto_cgroup);
476 static void disarm_sock_keys(struct mem_cgroup *memcg)
478 if (!memcg_proto_activated(&memcg->tcp_mem.cg_proto))
480 static_key_slow_dec(&memcg_socket_limit_enabled);
483 static void disarm_sock_keys(struct mem_cgroup *memcg)
488 static void drain_all_stock_async(struct mem_cgroup *memcg);
490 static struct mem_cgroup_per_zone *
491 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
493 return &memcg->info.nodeinfo[nid]->zoneinfo[zid];
496 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
501 static struct mem_cgroup_per_zone *
502 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
504 int nid = page_to_nid(page);
505 int zid = page_zonenum(page);
507 return mem_cgroup_zoneinfo(memcg, nid, zid);
510 static struct mem_cgroup_tree_per_zone *
511 soft_limit_tree_node_zone(int nid, int zid)
513 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
516 static struct mem_cgroup_tree_per_zone *
517 soft_limit_tree_from_page(struct page *page)
519 int nid = page_to_nid(page);
520 int zid = page_zonenum(page);
522 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
526 __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
527 struct mem_cgroup_per_zone *mz,
528 struct mem_cgroup_tree_per_zone *mctz,
529 unsigned long long new_usage_in_excess)
531 struct rb_node **p = &mctz->rb_root.rb_node;
532 struct rb_node *parent = NULL;
533 struct mem_cgroup_per_zone *mz_node;
538 mz->usage_in_excess = new_usage_in_excess;
539 if (!mz->usage_in_excess)
543 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
545 if (mz->usage_in_excess < mz_node->usage_in_excess)
548 * We can't avoid mem cgroups that are over their soft
549 * limit by the same amount
551 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
554 rb_link_node(&mz->tree_node, parent, p);
555 rb_insert_color(&mz->tree_node, &mctz->rb_root);
560 __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
561 struct mem_cgroup_per_zone *mz,
562 struct mem_cgroup_tree_per_zone *mctz)
566 rb_erase(&mz->tree_node, &mctz->rb_root);
571 mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
572 struct mem_cgroup_per_zone *mz,
573 struct mem_cgroup_tree_per_zone *mctz)
575 spin_lock(&mctz->lock);
576 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
577 spin_unlock(&mctz->lock);
581 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
583 unsigned long long excess;
584 struct mem_cgroup_per_zone *mz;
585 struct mem_cgroup_tree_per_zone *mctz;
586 int nid = page_to_nid(page);
587 int zid = page_zonenum(page);
588 mctz = soft_limit_tree_from_page(page);
591 * Necessary to update all ancestors when hierarchy is used.
592 * because their event counter is not touched.
594 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
595 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
596 excess = res_counter_soft_limit_excess(&memcg->res);
598 * We have to update the tree if mz is on RB-tree or
599 * mem is over its softlimit.
601 if (excess || mz->on_tree) {
602 spin_lock(&mctz->lock);
603 /* if on-tree, remove it */
605 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
607 * Insert again. mz->usage_in_excess will be updated.
608 * If excess is 0, no tree ops.
610 __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
611 spin_unlock(&mctz->lock);
616 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
619 struct mem_cgroup_per_zone *mz;
620 struct mem_cgroup_tree_per_zone *mctz;
622 for_each_node(node) {
623 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
624 mz = mem_cgroup_zoneinfo(memcg, node, zone);
625 mctz = soft_limit_tree_node_zone(node, zone);
626 mem_cgroup_remove_exceeded(memcg, mz, mctz);
631 static struct mem_cgroup_per_zone *
632 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
634 struct rb_node *rightmost = NULL;
635 struct mem_cgroup_per_zone *mz;
639 rightmost = rb_last(&mctz->rb_root);
641 goto done; /* Nothing to reclaim from */
643 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
645 * Remove the node now but someone else can add it back,
646 * we will to add it back at the end of reclaim to its correct
647 * position in the tree.
649 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
650 if (!res_counter_soft_limit_excess(&mz->memcg->res) ||
651 !css_tryget(&mz->memcg->css))
657 static struct mem_cgroup_per_zone *
658 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
660 struct mem_cgroup_per_zone *mz;
662 spin_lock(&mctz->lock);
663 mz = __mem_cgroup_largest_soft_limit_node(mctz);
664 spin_unlock(&mctz->lock);
669 * Implementation Note: reading percpu statistics for memcg.
671 * Both of vmstat[] and percpu_counter has threshold and do periodic
672 * synchronization to implement "quick" read. There are trade-off between
673 * reading cost and precision of value. Then, we may have a chance to implement
674 * a periodic synchronizion of counter in memcg's counter.
676 * But this _read() function is used for user interface now. The user accounts
677 * memory usage by memory cgroup and he _always_ requires exact value because
678 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
679 * have to visit all online cpus and make sum. So, for now, unnecessary
680 * synchronization is not implemented. (just implemented for cpu hotplug)
682 * If there are kernel internal actions which can make use of some not-exact
683 * value, and reading all cpu value can be performance bottleneck in some
684 * common workload, threashold and synchonization as vmstat[] should be
687 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
688 enum mem_cgroup_stat_index idx)
694 for_each_online_cpu(cpu)
695 val += per_cpu(memcg->stat->count[idx], cpu);
696 #ifdef CONFIG_HOTPLUG_CPU
697 spin_lock(&memcg->pcp_counter_lock);
698 val += memcg->nocpu_base.count[idx];
699 spin_unlock(&memcg->pcp_counter_lock);
705 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
708 int val = (charge) ? 1 : -1;
709 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
712 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
713 enum mem_cgroup_events_index idx)
715 unsigned long val = 0;
718 for_each_online_cpu(cpu)
719 val += per_cpu(memcg->stat->events[idx], cpu);
720 #ifdef CONFIG_HOTPLUG_CPU
721 spin_lock(&memcg->pcp_counter_lock);
722 val += memcg->nocpu_base.events[idx];
723 spin_unlock(&memcg->pcp_counter_lock);
728 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
729 bool anon, int nr_pages)
734 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
735 * counted as CACHE even if it's on ANON LRU.
738 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
741 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
744 /* pagein of a big page is an event. So, ignore page size */
746 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
748 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
749 nr_pages = -nr_pages; /* for event */
752 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
758 mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
760 struct mem_cgroup_per_zone *mz;
762 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
763 return mz->lru_size[lru];
767 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
768 unsigned int lru_mask)
770 struct mem_cgroup_per_zone *mz;
772 unsigned long ret = 0;
774 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
777 if (BIT(lru) & lru_mask)
778 ret += mz->lru_size[lru];
784 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
785 int nid, unsigned int lru_mask)
790 for (zid = 0; zid < MAX_NR_ZONES; zid++)
791 total += mem_cgroup_zone_nr_lru_pages(memcg,
797 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
798 unsigned int lru_mask)
803 for_each_node_state(nid, N_HIGH_MEMORY)
804 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
808 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
809 enum mem_cgroup_events_target target)
811 unsigned long val, next;
813 val = __this_cpu_read(memcg->stat->nr_page_events);
814 next = __this_cpu_read(memcg->stat->targets[target]);
815 /* from time_after() in jiffies.h */
816 if ((long)next - (long)val < 0) {
818 case MEM_CGROUP_TARGET_THRESH:
819 next = val + THRESHOLDS_EVENTS_TARGET;
821 case MEM_CGROUP_TARGET_SOFTLIMIT:
822 next = val + SOFTLIMIT_EVENTS_TARGET;
824 case MEM_CGROUP_TARGET_NUMAINFO:
825 next = val + NUMAINFO_EVENTS_TARGET;
830 __this_cpu_write(memcg->stat->targets[target], next);
837 * Check events in order.
840 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
843 /* threshold event is triggered in finer grain than soft limit */
844 if (unlikely(mem_cgroup_event_ratelimit(memcg,
845 MEM_CGROUP_TARGET_THRESH))) {
847 bool do_numainfo __maybe_unused;
849 do_softlimit = mem_cgroup_event_ratelimit(memcg,
850 MEM_CGROUP_TARGET_SOFTLIMIT);
852 do_numainfo = mem_cgroup_event_ratelimit(memcg,
853 MEM_CGROUP_TARGET_NUMAINFO);
857 mem_cgroup_threshold(memcg);
858 if (unlikely(do_softlimit))
859 mem_cgroup_update_tree(memcg, page);
861 if (unlikely(do_numainfo))
862 atomic_inc(&memcg->numainfo_events);
868 struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
870 return mem_cgroup_from_css(
871 cgroup_subsys_state(cont, mem_cgroup_subsys_id));
874 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
877 * mm_update_next_owner() may clear mm->owner to NULL
878 * if it races with swapoff, page migration, etc.
879 * So this can be called with p == NULL.
884 return mem_cgroup_from_css(task_subsys_state(p, mem_cgroup_subsys_id));
887 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
889 struct mem_cgroup *memcg = NULL;
894 * Because we have no locks, mm->owner's may be being moved to other
895 * cgroup. We use css_tryget() here even if this looks
896 * pessimistic (rather than adding locks here).
900 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
901 if (unlikely(!memcg))
903 } while (!css_tryget(&memcg->css));
909 * mem_cgroup_iter - iterate over memory cgroup hierarchy
910 * @root: hierarchy root
911 * @prev: previously returned memcg, NULL on first invocation
912 * @reclaim: cookie for shared reclaim walks, NULL for full walks
914 * Returns references to children of the hierarchy below @root, or
915 * @root itself, or %NULL after a full round-trip.
917 * Caller must pass the return value in @prev on subsequent
918 * invocations for reference counting, or use mem_cgroup_iter_break()
919 * to cancel a hierarchy walk before the round-trip is complete.
921 * Reclaimers can specify a zone and a priority level in @reclaim to
922 * divide up the memcgs in the hierarchy among all concurrent
923 * reclaimers operating on the same zone and priority.
925 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
926 struct mem_cgroup *prev,
927 struct mem_cgroup_reclaim_cookie *reclaim)
929 struct mem_cgroup *memcg = NULL;
932 if (mem_cgroup_disabled())
936 root = root_mem_cgroup;
938 if (prev && !reclaim)
939 id = css_id(&prev->css);
941 if (prev && prev != root)
944 if (!root->use_hierarchy && root != root_mem_cgroup) {
951 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
952 struct cgroup_subsys_state *css;
955 int nid = zone_to_nid(reclaim->zone);
956 int zid = zone_idx(reclaim->zone);
957 struct mem_cgroup_per_zone *mz;
959 mz = mem_cgroup_zoneinfo(root, nid, zid);
960 iter = &mz->reclaim_iter[reclaim->priority];
961 if (prev && reclaim->generation != iter->generation)
967 css = css_get_next(&mem_cgroup_subsys, id + 1, &root->css, &id);
969 if (css == &root->css || css_tryget(css))
970 memcg = mem_cgroup_from_css(css);
979 else if (!prev && memcg)
980 reclaim->generation = iter->generation;
990 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
991 * @root: hierarchy root
992 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
994 void mem_cgroup_iter_break(struct mem_cgroup *root,
995 struct mem_cgroup *prev)
998 root = root_mem_cgroup;
999 if (prev && prev != root)
1000 css_put(&prev->css);
1004 * Iteration constructs for visiting all cgroups (under a tree). If
1005 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1006 * be used for reference counting.
1008 #define for_each_mem_cgroup_tree(iter, root) \
1009 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1011 iter = mem_cgroup_iter(root, iter, NULL))
1013 #define for_each_mem_cgroup(iter) \
1014 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1016 iter = mem_cgroup_iter(NULL, iter, NULL))
1018 void mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1020 struct mem_cgroup *memcg;
1026 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1027 if (unlikely(!memcg))
1032 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1035 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1043 EXPORT_SYMBOL(mem_cgroup_count_vm_event);
1046 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1047 * @zone: zone of the wanted lruvec
1048 * @memcg: memcg of the wanted lruvec
1050 * Returns the lru list vector holding pages for the given @zone and
1051 * @mem. This can be the global zone lruvec, if the memory controller
1054 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1055 struct mem_cgroup *memcg)
1057 struct mem_cgroup_per_zone *mz;
1058 struct lruvec *lruvec;
1060 if (mem_cgroup_disabled()) {
1061 lruvec = &zone->lruvec;
1065 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1066 lruvec = &mz->lruvec;
1069 * Since a node can be onlined after the mem_cgroup was created,
1070 * we have to be prepared to initialize lruvec->zone here;
1071 * and if offlined then reonlined, we need to reinitialize it.
1073 if (unlikely(lruvec->zone != zone))
1074 lruvec->zone = zone;
1079 * Following LRU functions are allowed to be used without PCG_LOCK.
1080 * Operations are called by routine of global LRU independently from memcg.
1081 * What we have to take care of here is validness of pc->mem_cgroup.
1083 * Changes to pc->mem_cgroup happens when
1086 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1087 * It is added to LRU before charge.
1088 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1089 * When moving account, the page is not on LRU. It's isolated.
1093 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1095 * @zone: zone of the page
1097 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1099 struct mem_cgroup_per_zone *mz;
1100 struct mem_cgroup *memcg;
1101 struct page_cgroup *pc;
1102 struct lruvec *lruvec;
1104 if (mem_cgroup_disabled()) {
1105 lruvec = &zone->lruvec;
1109 pc = lookup_page_cgroup(page);
1110 memcg = pc->mem_cgroup;
1113 * Surreptitiously switch any uncharged offlist page to root:
1114 * an uncharged page off lru does nothing to secure
1115 * its former mem_cgroup from sudden removal.
1117 * Our caller holds lru_lock, and PageCgroupUsed is updated
1118 * under page_cgroup lock: between them, they make all uses
1119 * of pc->mem_cgroup safe.
1121 if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1122 pc->mem_cgroup = memcg = root_mem_cgroup;
1124 mz = page_cgroup_zoneinfo(memcg, page);
1125 lruvec = &mz->lruvec;
1128 * Since a node can be onlined after the mem_cgroup was created,
1129 * we have to be prepared to initialize lruvec->zone here;
1130 * and if offlined then reonlined, we need to reinitialize it.
1132 if (unlikely(lruvec->zone != zone))
1133 lruvec->zone = zone;
1138 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1139 * @lruvec: mem_cgroup per zone lru vector
1140 * @lru: index of lru list the page is sitting on
1141 * @nr_pages: positive when adding or negative when removing
1143 * This function must be called when a page is added to or removed from an
1146 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1149 struct mem_cgroup_per_zone *mz;
1150 unsigned long *lru_size;
1152 if (mem_cgroup_disabled())
1155 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1156 lru_size = mz->lru_size + lru;
1157 *lru_size += nr_pages;
1158 VM_BUG_ON((long)(*lru_size) < 0);
1162 * Checks whether given mem is same or in the root_mem_cgroup's
1165 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1166 struct mem_cgroup *memcg)
1168 if (root_memcg == memcg)
1170 if (!root_memcg->use_hierarchy || !memcg)
1172 return css_is_ancestor(&memcg->css, &root_memcg->css);
1175 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1176 struct mem_cgroup *memcg)
1181 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1186 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *memcg)
1189 struct mem_cgroup *curr = NULL;
1190 struct task_struct *p;
1192 p = find_lock_task_mm(task);
1194 curr = try_get_mem_cgroup_from_mm(p->mm);
1198 * All threads may have already detached their mm's, but the oom
1199 * killer still needs to detect if they have already been oom
1200 * killed to prevent needlessly killing additional tasks.
1203 curr = mem_cgroup_from_task(task);
1205 css_get(&curr->css);
1211 * We should check use_hierarchy of "memcg" not "curr". Because checking
1212 * use_hierarchy of "curr" here make this function true if hierarchy is
1213 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1214 * hierarchy(even if use_hierarchy is disabled in "memcg").
1216 ret = mem_cgroup_same_or_subtree(memcg, curr);
1217 css_put(&curr->css);
1221 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1223 unsigned long inactive_ratio;
1224 unsigned long inactive;
1225 unsigned long active;
1228 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1229 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1231 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1233 inactive_ratio = int_sqrt(10 * gb);
1237 return inactive * inactive_ratio < active;
1240 int mem_cgroup_inactive_file_is_low(struct lruvec *lruvec)
1242 unsigned long active;
1243 unsigned long inactive;
1245 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_FILE);
1246 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_FILE);
1248 return (active > inactive);
1251 #define mem_cgroup_from_res_counter(counter, member) \
1252 container_of(counter, struct mem_cgroup, member)
1255 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1256 * @memcg: the memory cgroup
1258 * Returns the maximum amount of memory @mem can be charged with, in
1261 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1263 unsigned long long margin;
1265 margin = res_counter_margin(&memcg->res);
1266 if (do_swap_account)
1267 margin = min(margin, res_counter_margin(&memcg->memsw));
1268 return margin >> PAGE_SHIFT;
1271 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1273 struct cgroup *cgrp = memcg->css.cgroup;
1276 if (cgrp->parent == NULL)
1277 return vm_swappiness;
1279 return memcg->swappiness;
1283 * memcg->moving_account is used for checking possibility that some thread is
1284 * calling move_account(). When a thread on CPU-A starts moving pages under
1285 * a memcg, other threads should check memcg->moving_account under
1286 * rcu_read_lock(), like this:
1290 * memcg->moving_account+1 if (memcg->mocing_account)
1292 * synchronize_rcu() update something.
1297 /* for quick checking without looking up memcg */
1298 atomic_t memcg_moving __read_mostly;
1300 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1302 atomic_inc(&memcg_moving);
1303 atomic_inc(&memcg->moving_account);
1307 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1310 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1311 * We check NULL in callee rather than caller.
1314 atomic_dec(&memcg_moving);
1315 atomic_dec(&memcg->moving_account);
1320 * 2 routines for checking "mem" is under move_account() or not.
1322 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1323 * is used for avoiding races in accounting. If true,
1324 * pc->mem_cgroup may be overwritten.
1326 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1327 * under hierarchy of moving cgroups. This is for
1328 * waiting at hith-memory prressure caused by "move".
1331 static bool mem_cgroup_stolen(struct mem_cgroup *memcg)
1333 VM_BUG_ON(!rcu_read_lock_held());
1334 return atomic_read(&memcg->moving_account) > 0;
1337 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1339 struct mem_cgroup *from;
1340 struct mem_cgroup *to;
1343 * Unlike task_move routines, we access mc.to, mc.from not under
1344 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1346 spin_lock(&mc.lock);
1352 ret = mem_cgroup_same_or_subtree(memcg, from)
1353 || mem_cgroup_same_or_subtree(memcg, to);
1355 spin_unlock(&mc.lock);
1359 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1361 if (mc.moving_task && current != mc.moving_task) {
1362 if (mem_cgroup_under_move(memcg)) {
1364 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1365 /* moving charge context might have finished. */
1368 finish_wait(&mc.waitq, &wait);
1376 * Take this lock when
1377 * - a code tries to modify page's memcg while it's USED.
1378 * - a code tries to modify page state accounting in a memcg.
1379 * see mem_cgroup_stolen(), too.
1381 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1382 unsigned long *flags)
1384 spin_lock_irqsave(&memcg->move_lock, *flags);
1387 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1388 unsigned long *flags)
1390 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1394 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1395 * @memcg: The memory cgroup that went over limit
1396 * @p: Task that is going to be killed
1398 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1401 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1403 struct cgroup *task_cgrp;
1404 struct cgroup *mem_cgrp;
1406 * Need a buffer in BSS, can't rely on allocations. The code relies
1407 * on the assumption that OOM is serialized for memory controller.
1408 * If this assumption is broken, revisit this code.
1410 static char memcg_name[PATH_MAX];
1418 mem_cgrp = memcg->css.cgroup;
1419 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1421 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1424 * Unfortunately, we are unable to convert to a useful name
1425 * But we'll still print out the usage information
1432 printk(KERN_INFO "Task in %s killed", memcg_name);
1435 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1443 * Continues from above, so we don't need an KERN_ level
1445 printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1448 printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1449 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1450 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1451 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1452 printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1454 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1455 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1456 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1460 * This function returns the number of memcg under hierarchy tree. Returns
1461 * 1(self count) if no children.
1463 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1466 struct mem_cgroup *iter;
1468 for_each_mem_cgroup_tree(iter, memcg)
1474 * Return the memory (and swap, if configured) limit for a memcg.
1476 static u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1480 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1483 * Do not consider swap space if we cannot swap due to swappiness
1485 if (mem_cgroup_swappiness(memcg)) {
1488 limit += total_swap_pages << PAGE_SHIFT;
1489 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1492 * If memsw is finite and limits the amount of swap space
1493 * available to this memcg, return that limit.
1495 limit = min(limit, memsw);
1501 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1504 struct mem_cgroup *iter;
1505 unsigned long chosen_points = 0;
1506 unsigned long totalpages;
1507 unsigned int points = 0;
1508 struct task_struct *chosen = NULL;
1511 * If current has a pending SIGKILL, then automatically select it. The
1512 * goal is to allow it to allocate so that it may quickly exit and free
1515 if (fatal_signal_pending(current)) {
1516 set_thread_flag(TIF_MEMDIE);
1520 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1521 totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
1522 for_each_mem_cgroup_tree(iter, memcg) {
1523 struct cgroup *cgroup = iter->css.cgroup;
1524 struct cgroup_iter it;
1525 struct task_struct *task;
1527 cgroup_iter_start(cgroup, &it);
1528 while ((task = cgroup_iter_next(cgroup, &it))) {
1529 switch (oom_scan_process_thread(task, totalpages, NULL,
1531 case OOM_SCAN_SELECT:
1533 put_task_struct(chosen);
1535 chosen_points = ULONG_MAX;
1536 get_task_struct(chosen);
1538 case OOM_SCAN_CONTINUE:
1540 case OOM_SCAN_ABORT:
1541 cgroup_iter_end(cgroup, &it);
1542 mem_cgroup_iter_break(memcg, iter);
1544 put_task_struct(chosen);
1549 points = oom_badness(task, memcg, NULL, totalpages);
1550 if (points > chosen_points) {
1552 put_task_struct(chosen);
1554 chosen_points = points;
1555 get_task_struct(chosen);
1558 cgroup_iter_end(cgroup, &it);
1563 points = chosen_points * 1000 / totalpages;
1564 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1565 NULL, "Memory cgroup out of memory");
1568 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1570 unsigned long flags)
1572 unsigned long total = 0;
1573 bool noswap = false;
1576 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1578 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1581 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1583 drain_all_stock_async(memcg);
1584 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1586 * Allow limit shrinkers, which are triggered directly
1587 * by userspace, to catch signals and stop reclaim
1588 * after minimal progress, regardless of the margin.
1590 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1592 if (mem_cgroup_margin(memcg))
1595 * If nothing was reclaimed after two attempts, there
1596 * may be no reclaimable pages in this hierarchy.
1605 * test_mem_cgroup_node_reclaimable
1606 * @memcg: the target memcg
1607 * @nid: the node ID to be checked.
1608 * @noswap : specify true here if the user wants flle only information.
1610 * This function returns whether the specified memcg contains any
1611 * reclaimable pages on a node. Returns true if there are any reclaimable
1612 * pages in the node.
1614 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1615 int nid, bool noswap)
1617 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1619 if (noswap || !total_swap_pages)
1621 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1626 #if MAX_NUMNODES > 1
1629 * Always updating the nodemask is not very good - even if we have an empty
1630 * list or the wrong list here, we can start from some node and traverse all
1631 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1634 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1638 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1639 * pagein/pageout changes since the last update.
1641 if (!atomic_read(&memcg->numainfo_events))
1643 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1646 /* make a nodemask where this memcg uses memory from */
1647 memcg->scan_nodes = node_states[N_HIGH_MEMORY];
1649 for_each_node_mask(nid, node_states[N_HIGH_MEMORY]) {
1651 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1652 node_clear(nid, memcg->scan_nodes);
1655 atomic_set(&memcg->numainfo_events, 0);
1656 atomic_set(&memcg->numainfo_updating, 0);
1660 * Selecting a node where we start reclaim from. Because what we need is just
1661 * reducing usage counter, start from anywhere is O,K. Considering
1662 * memory reclaim from current node, there are pros. and cons.
1664 * Freeing memory from current node means freeing memory from a node which
1665 * we'll use or we've used. So, it may make LRU bad. And if several threads
1666 * hit limits, it will see a contention on a node. But freeing from remote
1667 * node means more costs for memory reclaim because of memory latency.
1669 * Now, we use round-robin. Better algorithm is welcomed.
1671 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1675 mem_cgroup_may_update_nodemask(memcg);
1676 node = memcg->last_scanned_node;
1678 node = next_node(node, memcg->scan_nodes);
1679 if (node == MAX_NUMNODES)
1680 node = first_node(memcg->scan_nodes);
1682 * We call this when we hit limit, not when pages are added to LRU.
1683 * No LRU may hold pages because all pages are UNEVICTABLE or
1684 * memcg is too small and all pages are not on LRU. In that case,
1685 * we use curret node.
1687 if (unlikely(node == MAX_NUMNODES))
1688 node = numa_node_id();
1690 memcg->last_scanned_node = node;
1695 * Check all nodes whether it contains reclaimable pages or not.
1696 * For quick scan, we make use of scan_nodes. This will allow us to skip
1697 * unused nodes. But scan_nodes is lazily updated and may not cotain
1698 * enough new information. We need to do double check.
1700 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1705 * quick check...making use of scan_node.
1706 * We can skip unused nodes.
1708 if (!nodes_empty(memcg->scan_nodes)) {
1709 for (nid = first_node(memcg->scan_nodes);
1711 nid = next_node(nid, memcg->scan_nodes)) {
1713 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1718 * Check rest of nodes.
1720 for_each_node_state(nid, N_HIGH_MEMORY) {
1721 if (node_isset(nid, memcg->scan_nodes))
1723 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1730 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1735 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1737 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
1741 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1744 unsigned long *total_scanned)
1746 struct mem_cgroup *victim = NULL;
1749 unsigned long excess;
1750 unsigned long nr_scanned;
1751 struct mem_cgroup_reclaim_cookie reclaim = {
1756 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
1759 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1764 * If we have not been able to reclaim
1765 * anything, it might because there are
1766 * no reclaimable pages under this hierarchy
1771 * We want to do more targeted reclaim.
1772 * excess >> 2 is not to excessive so as to
1773 * reclaim too much, nor too less that we keep
1774 * coming back to reclaim from this cgroup
1776 if (total >= (excess >> 2) ||
1777 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1782 if (!mem_cgroup_reclaimable(victim, false))
1784 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1786 *total_scanned += nr_scanned;
1787 if (!res_counter_soft_limit_excess(&root_memcg->res))
1790 mem_cgroup_iter_break(root_memcg, victim);
1795 * Check OOM-Killer is already running under our hierarchy.
1796 * If someone is running, return false.
1797 * Has to be called with memcg_oom_lock
1799 static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg)
1801 struct mem_cgroup *iter, *failed = NULL;
1803 for_each_mem_cgroup_tree(iter, memcg) {
1804 if (iter->oom_lock) {
1806 * this subtree of our hierarchy is already locked
1807 * so we cannot give a lock.
1810 mem_cgroup_iter_break(memcg, iter);
1813 iter->oom_lock = true;
1820 * OK, we failed to lock the whole subtree so we have to clean up
1821 * what we set up to the failing subtree
1823 for_each_mem_cgroup_tree(iter, memcg) {
1824 if (iter == failed) {
1825 mem_cgroup_iter_break(memcg, iter);
1828 iter->oom_lock = false;
1834 * Has to be called with memcg_oom_lock
1836 static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1838 struct mem_cgroup *iter;
1840 for_each_mem_cgroup_tree(iter, memcg)
1841 iter->oom_lock = false;
1845 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1847 struct mem_cgroup *iter;
1849 for_each_mem_cgroup_tree(iter, memcg)
1850 atomic_inc(&iter->under_oom);
1853 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1855 struct mem_cgroup *iter;
1858 * When a new child is created while the hierarchy is under oom,
1859 * mem_cgroup_oom_lock() may not be called. We have to use
1860 * atomic_add_unless() here.
1862 for_each_mem_cgroup_tree(iter, memcg)
1863 atomic_add_unless(&iter->under_oom, -1, 0);
1866 static DEFINE_SPINLOCK(memcg_oom_lock);
1867 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1869 struct oom_wait_info {
1870 struct mem_cgroup *memcg;
1874 static int memcg_oom_wake_function(wait_queue_t *wait,
1875 unsigned mode, int sync, void *arg)
1877 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1878 struct mem_cgroup *oom_wait_memcg;
1879 struct oom_wait_info *oom_wait_info;
1881 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1882 oom_wait_memcg = oom_wait_info->memcg;
1885 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
1886 * Then we can use css_is_ancestor without taking care of RCU.
1888 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
1889 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
1891 return autoremove_wake_function(wait, mode, sync, arg);
1894 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
1896 /* for filtering, pass "memcg" as argument. */
1897 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1900 static void memcg_oom_recover(struct mem_cgroup *memcg)
1902 if (memcg && atomic_read(&memcg->under_oom))
1903 memcg_wakeup_oom(memcg);
1907 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1909 static bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask,
1912 struct oom_wait_info owait;
1913 bool locked, need_to_kill;
1915 owait.memcg = memcg;
1916 owait.wait.flags = 0;
1917 owait.wait.func = memcg_oom_wake_function;
1918 owait.wait.private = current;
1919 INIT_LIST_HEAD(&owait.wait.task_list);
1920 need_to_kill = true;
1921 mem_cgroup_mark_under_oom(memcg);
1923 /* At first, try to OOM lock hierarchy under memcg.*/
1924 spin_lock(&memcg_oom_lock);
1925 locked = mem_cgroup_oom_lock(memcg);
1927 * Even if signal_pending(), we can't quit charge() loop without
1928 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1929 * under OOM is always welcomed, use TASK_KILLABLE here.
1931 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1932 if (!locked || memcg->oom_kill_disable)
1933 need_to_kill = false;
1935 mem_cgroup_oom_notify(memcg);
1936 spin_unlock(&memcg_oom_lock);
1939 finish_wait(&memcg_oom_waitq, &owait.wait);
1940 mem_cgroup_out_of_memory(memcg, mask, order);
1943 finish_wait(&memcg_oom_waitq, &owait.wait);
1945 spin_lock(&memcg_oom_lock);
1947 mem_cgroup_oom_unlock(memcg);
1948 memcg_wakeup_oom(memcg);
1949 spin_unlock(&memcg_oom_lock);
1951 mem_cgroup_unmark_under_oom(memcg);
1953 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
1955 /* Give chance to dying process */
1956 schedule_timeout_uninterruptible(1);
1961 * Currently used to update mapped file statistics, but the routine can be
1962 * generalized to update other statistics as well.
1964 * Notes: Race condition
1966 * We usually use page_cgroup_lock() for accessing page_cgroup member but
1967 * it tends to be costly. But considering some conditions, we doesn't need
1968 * to do so _always_.
1970 * Considering "charge", lock_page_cgroup() is not required because all
1971 * file-stat operations happen after a page is attached to radix-tree. There
1972 * are no race with "charge".
1974 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
1975 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
1976 * if there are race with "uncharge". Statistics itself is properly handled
1979 * Considering "move", this is an only case we see a race. To make the race
1980 * small, we check mm->moving_account and detect there are possibility of race
1981 * If there is, we take a lock.
1984 void __mem_cgroup_begin_update_page_stat(struct page *page,
1985 bool *locked, unsigned long *flags)
1987 struct mem_cgroup *memcg;
1988 struct page_cgroup *pc;
1990 pc = lookup_page_cgroup(page);
1992 memcg = pc->mem_cgroup;
1993 if (unlikely(!memcg || !PageCgroupUsed(pc)))
1996 * If this memory cgroup is not under account moving, we don't
1997 * need to take move_lock_mem_cgroup(). Because we already hold
1998 * rcu_read_lock(), any calls to move_account will be delayed until
1999 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2001 if (!mem_cgroup_stolen(memcg))
2004 move_lock_mem_cgroup(memcg, flags);
2005 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2006 move_unlock_mem_cgroup(memcg, flags);
2012 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
2014 struct page_cgroup *pc = lookup_page_cgroup(page);
2017 * It's guaranteed that pc->mem_cgroup never changes while
2018 * lock is held because a routine modifies pc->mem_cgroup
2019 * should take move_lock_mem_cgroup().
2021 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
2024 void mem_cgroup_update_page_stat(struct page *page,
2025 enum mem_cgroup_page_stat_item idx, int val)
2027 struct mem_cgroup *memcg;
2028 struct page_cgroup *pc = lookup_page_cgroup(page);
2029 unsigned long uninitialized_var(flags);
2031 if (mem_cgroup_disabled())
2034 memcg = pc->mem_cgroup;
2035 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2039 case MEMCG_NR_FILE_MAPPED:
2040 idx = MEM_CGROUP_STAT_FILE_MAPPED;
2046 this_cpu_add(memcg->stat->count[idx], val);
2050 * size of first charge trial. "32" comes from vmscan.c's magic value.
2051 * TODO: maybe necessary to use big numbers in big irons.
2053 #define CHARGE_BATCH 32U
2054 struct memcg_stock_pcp {
2055 struct mem_cgroup *cached; /* this never be root cgroup */
2056 unsigned int nr_pages;
2057 struct work_struct work;
2058 unsigned long flags;
2059 #define FLUSHING_CACHED_CHARGE 0
2061 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2062 static DEFINE_MUTEX(percpu_charge_mutex);
2065 * Try to consume stocked charge on this cpu. If success, one page is consumed
2066 * from local stock and true is returned. If the stock is 0 or charges from a
2067 * cgroup which is not current target, returns false. This stock will be
2070 static bool consume_stock(struct mem_cgroup *memcg)
2072 struct memcg_stock_pcp *stock;
2075 stock = &get_cpu_var(memcg_stock);
2076 if (memcg == stock->cached && stock->nr_pages)
2078 else /* need to call res_counter_charge */
2080 put_cpu_var(memcg_stock);
2085 * Returns stocks cached in percpu to res_counter and reset cached information.
2087 static void drain_stock(struct memcg_stock_pcp *stock)
2089 struct mem_cgroup *old = stock->cached;
2091 if (stock->nr_pages) {
2092 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2094 res_counter_uncharge(&old->res, bytes);
2095 if (do_swap_account)
2096 res_counter_uncharge(&old->memsw, bytes);
2097 stock->nr_pages = 0;
2099 stock->cached = NULL;
2103 * This must be called under preempt disabled or must be called by
2104 * a thread which is pinned to local cpu.
2106 static void drain_local_stock(struct work_struct *dummy)
2108 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2110 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2114 * Cache charges(val) which is from res_counter, to local per_cpu area.
2115 * This will be consumed by consume_stock() function, later.
2117 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2119 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2121 if (stock->cached != memcg) { /* reset if necessary */
2123 stock->cached = memcg;
2125 stock->nr_pages += nr_pages;
2126 put_cpu_var(memcg_stock);
2130 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2131 * of the hierarchy under it. sync flag says whether we should block
2132 * until the work is done.
2134 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2138 /* Notify other cpus that system-wide "drain" is running */
2141 for_each_online_cpu(cpu) {
2142 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2143 struct mem_cgroup *memcg;
2145 memcg = stock->cached;
2146 if (!memcg || !stock->nr_pages)
2148 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2150 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2152 drain_local_stock(&stock->work);
2154 schedule_work_on(cpu, &stock->work);
2162 for_each_online_cpu(cpu) {
2163 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2164 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2165 flush_work(&stock->work);
2172 * Tries to drain stocked charges in other cpus. This function is asynchronous
2173 * and just put a work per cpu for draining localy on each cpu. Caller can
2174 * expects some charges will be back to res_counter later but cannot wait for
2177 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2180 * If someone calls draining, avoid adding more kworker runs.
2182 if (!mutex_trylock(&percpu_charge_mutex))
2184 drain_all_stock(root_memcg, false);
2185 mutex_unlock(&percpu_charge_mutex);
2188 /* This is a synchronous drain interface. */
2189 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2191 /* called when force_empty is called */
2192 mutex_lock(&percpu_charge_mutex);
2193 drain_all_stock(root_memcg, true);
2194 mutex_unlock(&percpu_charge_mutex);
2198 * This function drains percpu counter value from DEAD cpu and
2199 * move it to local cpu. Note that this function can be preempted.
2201 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2205 spin_lock(&memcg->pcp_counter_lock);
2206 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2207 long x = per_cpu(memcg->stat->count[i], cpu);
2209 per_cpu(memcg->stat->count[i], cpu) = 0;
2210 memcg->nocpu_base.count[i] += x;
2212 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2213 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2215 per_cpu(memcg->stat->events[i], cpu) = 0;
2216 memcg->nocpu_base.events[i] += x;
2218 spin_unlock(&memcg->pcp_counter_lock);
2221 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2222 unsigned long action,
2225 int cpu = (unsigned long)hcpu;
2226 struct memcg_stock_pcp *stock;
2227 struct mem_cgroup *iter;
2229 if (action == CPU_ONLINE)
2232 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2235 for_each_mem_cgroup(iter)
2236 mem_cgroup_drain_pcp_counter(iter, cpu);
2238 stock = &per_cpu(memcg_stock, cpu);
2244 /* See __mem_cgroup_try_charge() for details */
2246 CHARGE_OK, /* success */
2247 CHARGE_RETRY, /* need to retry but retry is not bad */
2248 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2249 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2250 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2253 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2254 unsigned int nr_pages, bool oom_check)
2256 unsigned long csize = nr_pages * PAGE_SIZE;
2257 struct mem_cgroup *mem_over_limit;
2258 struct res_counter *fail_res;
2259 unsigned long flags = 0;
2262 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2265 if (!do_swap_account)
2267 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2271 res_counter_uncharge(&memcg->res, csize);
2272 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2273 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2275 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2277 * nr_pages can be either a huge page (HPAGE_PMD_NR), a batch
2278 * of regular pages (CHARGE_BATCH), or a single regular page (1).
2280 * Never reclaim on behalf of optional batching, retry with a
2281 * single page instead.
2283 if (nr_pages == CHARGE_BATCH)
2284 return CHARGE_RETRY;
2286 if (!(gfp_mask & __GFP_WAIT))
2287 return CHARGE_WOULDBLOCK;
2289 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2290 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2291 return CHARGE_RETRY;
2293 * Even though the limit is exceeded at this point, reclaim
2294 * may have been able to free some pages. Retry the charge
2295 * before killing the task.
2297 * Only for regular pages, though: huge pages are rather
2298 * unlikely to succeed so close to the limit, and we fall back
2299 * to regular pages anyway in case of failure.
2301 if (nr_pages == 1 && ret)
2302 return CHARGE_RETRY;
2305 * At task move, charge accounts can be doubly counted. So, it's
2306 * better to wait until the end of task_move if something is going on.
2308 if (mem_cgroup_wait_acct_move(mem_over_limit))
2309 return CHARGE_RETRY;
2311 /* If we don't need to call oom-killer at el, return immediately */
2313 return CHARGE_NOMEM;
2315 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask, get_order(csize)))
2316 return CHARGE_OOM_DIE;
2318 return CHARGE_RETRY;
2322 * __mem_cgroup_try_charge() does
2323 * 1. detect memcg to be charged against from passed *mm and *ptr,
2324 * 2. update res_counter
2325 * 3. call memory reclaim if necessary.
2327 * In some special case, if the task is fatal, fatal_signal_pending() or
2328 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2329 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2330 * as possible without any hazards. 2: all pages should have a valid
2331 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2332 * pointer, that is treated as a charge to root_mem_cgroup.
2334 * So __mem_cgroup_try_charge() will return
2335 * 0 ... on success, filling *ptr with a valid memcg pointer.
2336 * -ENOMEM ... charge failure because of resource limits.
2337 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2339 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2340 * the oom-killer can be invoked.
2342 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2344 unsigned int nr_pages,
2345 struct mem_cgroup **ptr,
2348 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2349 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2350 struct mem_cgroup *memcg = NULL;
2354 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2355 * in system level. So, allow to go ahead dying process in addition to
2358 if (unlikely(test_thread_flag(TIF_MEMDIE)
2359 || fatal_signal_pending(current)))
2363 * We always charge the cgroup the mm_struct belongs to.
2364 * The mm_struct's mem_cgroup changes on task migration if the
2365 * thread group leader migrates. It's possible that mm is not
2366 * set, if so charge the root memcg (happens for pagecache usage).
2369 *ptr = root_mem_cgroup;
2371 if (*ptr) { /* css should be a valid one */
2373 if (mem_cgroup_is_root(memcg))
2375 if (nr_pages == 1 && consume_stock(memcg))
2377 css_get(&memcg->css);
2379 struct task_struct *p;
2382 p = rcu_dereference(mm->owner);
2384 * Because we don't have task_lock(), "p" can exit.
2385 * In that case, "memcg" can point to root or p can be NULL with
2386 * race with swapoff. Then, we have small risk of mis-accouning.
2387 * But such kind of mis-account by race always happens because
2388 * we don't have cgroup_mutex(). It's overkill and we allo that
2390 * (*) swapoff at el will charge against mm-struct not against
2391 * task-struct. So, mm->owner can be NULL.
2393 memcg = mem_cgroup_from_task(p);
2395 memcg = root_mem_cgroup;
2396 if (mem_cgroup_is_root(memcg)) {
2400 if (nr_pages == 1 && consume_stock(memcg)) {
2402 * It seems dagerous to access memcg without css_get().
2403 * But considering how consume_stok works, it's not
2404 * necessary. If consume_stock success, some charges
2405 * from this memcg are cached on this cpu. So, we
2406 * don't need to call css_get()/css_tryget() before
2407 * calling consume_stock().
2412 /* after here, we may be blocked. we need to get refcnt */
2413 if (!css_tryget(&memcg->css)) {
2423 /* If killed, bypass charge */
2424 if (fatal_signal_pending(current)) {
2425 css_put(&memcg->css);
2430 if (oom && !nr_oom_retries) {
2432 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2435 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, oom_check);
2439 case CHARGE_RETRY: /* not in OOM situation but retry */
2441 css_put(&memcg->css);
2444 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2445 css_put(&memcg->css);
2447 case CHARGE_NOMEM: /* OOM routine works */
2449 css_put(&memcg->css);
2452 /* If oom, we never return -ENOMEM */
2455 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2456 css_put(&memcg->css);
2459 } while (ret != CHARGE_OK);
2461 if (batch > nr_pages)
2462 refill_stock(memcg, batch - nr_pages);
2463 css_put(&memcg->css);
2471 *ptr = root_mem_cgroup;
2476 * Somemtimes we have to undo a charge we got by try_charge().
2477 * This function is for that and do uncharge, put css's refcnt.
2478 * gotten by try_charge().
2480 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2481 unsigned int nr_pages)
2483 if (!mem_cgroup_is_root(memcg)) {
2484 unsigned long bytes = nr_pages * PAGE_SIZE;
2486 res_counter_uncharge(&memcg->res, bytes);
2487 if (do_swap_account)
2488 res_counter_uncharge(&memcg->memsw, bytes);
2493 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2494 * This is useful when moving usage to parent cgroup.
2496 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2497 unsigned int nr_pages)
2499 unsigned long bytes = nr_pages * PAGE_SIZE;
2501 if (mem_cgroup_is_root(memcg))
2504 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2505 if (do_swap_account)
2506 res_counter_uncharge_until(&memcg->memsw,
2507 memcg->memsw.parent, bytes);
2511 * A helper function to get mem_cgroup from ID. must be called under
2512 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2513 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2514 * called against removed memcg.)
2516 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2518 struct cgroup_subsys_state *css;
2520 /* ID 0 is unused ID */
2523 css = css_lookup(&mem_cgroup_subsys, id);
2526 return mem_cgroup_from_css(css);
2529 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2531 struct mem_cgroup *memcg = NULL;
2532 struct page_cgroup *pc;
2536 VM_BUG_ON(!PageLocked(page));
2538 pc = lookup_page_cgroup(page);
2539 lock_page_cgroup(pc);
2540 if (PageCgroupUsed(pc)) {
2541 memcg = pc->mem_cgroup;
2542 if (memcg && !css_tryget(&memcg->css))
2544 } else if (PageSwapCache(page)) {
2545 ent.val = page_private(page);
2546 id = lookup_swap_cgroup_id(ent);
2548 memcg = mem_cgroup_lookup(id);
2549 if (memcg && !css_tryget(&memcg->css))
2553 unlock_page_cgroup(pc);
2557 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2559 unsigned int nr_pages,
2560 enum charge_type ctype,
2563 struct page_cgroup *pc = lookup_page_cgroup(page);
2564 struct zone *uninitialized_var(zone);
2565 struct lruvec *lruvec;
2566 bool was_on_lru = false;
2569 lock_page_cgroup(pc);
2570 VM_BUG_ON(PageCgroupUsed(pc));
2572 * we don't need page_cgroup_lock about tail pages, becase they are not
2573 * accessed by any other context at this point.
2577 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2578 * may already be on some other mem_cgroup's LRU. Take care of it.
2581 zone = page_zone(page);
2582 spin_lock_irq(&zone->lru_lock);
2583 if (PageLRU(page)) {
2584 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2586 del_page_from_lru_list(page, lruvec, page_lru(page));
2591 pc->mem_cgroup = memcg;
2593 * We access a page_cgroup asynchronously without lock_page_cgroup().
2594 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2595 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2596 * before USED bit, we need memory barrier here.
2597 * See mem_cgroup_add_lru_list(), etc.
2600 SetPageCgroupUsed(pc);
2604 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2605 VM_BUG_ON(PageLRU(page));
2607 add_page_to_lru_list(page, lruvec, page_lru(page));
2609 spin_unlock_irq(&zone->lru_lock);
2612 if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
2617 mem_cgroup_charge_statistics(memcg, anon, nr_pages);
2618 unlock_page_cgroup(pc);
2621 * "charge_statistics" updated event counter. Then, check it.
2622 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2623 * if they exceeds softlimit.
2625 memcg_check_events(memcg, page);
2628 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2630 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
2632 * Because tail pages are not marked as "used", set it. We're under
2633 * zone->lru_lock, 'splitting on pmd' and compound_lock.
2634 * charge/uncharge will be never happen and move_account() is done under
2635 * compound_lock(), so we don't have to take care of races.
2637 void mem_cgroup_split_huge_fixup(struct page *head)
2639 struct page_cgroup *head_pc = lookup_page_cgroup(head);
2640 struct page_cgroup *pc;
2643 if (mem_cgroup_disabled())
2645 for (i = 1; i < HPAGE_PMD_NR; i++) {
2647 pc->mem_cgroup = head_pc->mem_cgroup;
2648 smp_wmb();/* see __commit_charge() */
2649 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2652 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2655 * mem_cgroup_move_account - move account of the page
2657 * @nr_pages: number of regular pages (>1 for huge pages)
2658 * @pc: page_cgroup of the page.
2659 * @from: mem_cgroup which the page is moved from.
2660 * @to: mem_cgroup which the page is moved to. @from != @to.
2662 * The caller must confirm following.
2663 * - page is not on LRU (isolate_page() is useful.)
2664 * - compound_lock is held when nr_pages > 1
2666 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
2669 static int mem_cgroup_move_account(struct page *page,
2670 unsigned int nr_pages,
2671 struct page_cgroup *pc,
2672 struct mem_cgroup *from,
2673 struct mem_cgroup *to)
2675 unsigned long flags;
2677 bool anon = PageAnon(page);
2679 VM_BUG_ON(from == to);
2680 VM_BUG_ON(PageLRU(page));
2682 * The page is isolated from LRU. So, collapse function
2683 * will not handle this page. But page splitting can happen.
2684 * Do this check under compound_page_lock(). The caller should
2688 if (nr_pages > 1 && !PageTransHuge(page))
2691 lock_page_cgroup(pc);
2694 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
2697 move_lock_mem_cgroup(from, &flags);
2699 if (!anon && page_mapped(page)) {
2700 /* Update mapped_file data for mem_cgroup */
2702 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2703 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2706 mem_cgroup_charge_statistics(from, anon, -nr_pages);
2708 /* caller should have done css_get */
2709 pc->mem_cgroup = to;
2710 mem_cgroup_charge_statistics(to, anon, nr_pages);
2711 move_unlock_mem_cgroup(from, &flags);
2714 unlock_page_cgroup(pc);
2718 memcg_check_events(to, page);
2719 memcg_check_events(from, page);
2725 * mem_cgroup_move_parent - moves page to the parent group
2726 * @page: the page to move
2727 * @pc: page_cgroup of the page
2728 * @child: page's cgroup
2730 * move charges to its parent or the root cgroup if the group has no
2731 * parent (aka use_hierarchy==0).
2732 * Although this might fail (get_page_unless_zero, isolate_lru_page or
2733 * mem_cgroup_move_account fails) the failure is always temporary and
2734 * it signals a race with a page removal/uncharge or migration. In the
2735 * first case the page is on the way out and it will vanish from the LRU
2736 * on the next attempt and the call should be retried later.
2737 * Isolation from the LRU fails only if page has been isolated from
2738 * the LRU since we looked at it and that usually means either global
2739 * reclaim or migration going on. The page will either get back to the
2741 * Finaly mem_cgroup_move_account fails only if the page got uncharged
2742 * (!PageCgroupUsed) or moved to a different group. The page will
2743 * disappear in the next attempt.
2745 static int mem_cgroup_move_parent(struct page *page,
2746 struct page_cgroup *pc,
2747 struct mem_cgroup *child)
2749 struct mem_cgroup *parent;
2750 unsigned int nr_pages;
2751 unsigned long uninitialized_var(flags);
2754 VM_BUG_ON(mem_cgroup_is_root(child));
2757 if (!get_page_unless_zero(page))
2759 if (isolate_lru_page(page))
2762 nr_pages = hpage_nr_pages(page);
2764 parent = parent_mem_cgroup(child);
2766 * If no parent, move charges to root cgroup.
2769 parent = root_mem_cgroup;
2772 VM_BUG_ON(!PageTransHuge(page));
2773 flags = compound_lock_irqsave(page);
2776 ret = mem_cgroup_move_account(page, nr_pages,
2779 __mem_cgroup_cancel_local_charge(child, nr_pages);
2782 compound_unlock_irqrestore(page, flags);
2783 putback_lru_page(page);
2791 * Charge the memory controller for page usage.
2793 * 0 if the charge was successful
2794 * < 0 if the cgroup is over its limit
2796 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2797 gfp_t gfp_mask, enum charge_type ctype)
2799 struct mem_cgroup *memcg = NULL;
2800 unsigned int nr_pages = 1;
2804 if (PageTransHuge(page)) {
2805 nr_pages <<= compound_order(page);
2806 VM_BUG_ON(!PageTransHuge(page));
2808 * Never OOM-kill a process for a huge page. The
2809 * fault handler will fall back to regular pages.
2814 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
2817 __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false);
2821 int mem_cgroup_newpage_charge(struct page *page,
2822 struct mm_struct *mm, gfp_t gfp_mask)
2824 if (mem_cgroup_disabled())
2826 VM_BUG_ON(page_mapped(page));
2827 VM_BUG_ON(page->mapping && !PageAnon(page));
2829 return mem_cgroup_charge_common(page, mm, gfp_mask,
2830 MEM_CGROUP_CHARGE_TYPE_ANON);
2834 * While swap-in, try_charge -> commit or cancel, the page is locked.
2835 * And when try_charge() successfully returns, one refcnt to memcg without
2836 * struct page_cgroup is acquired. This refcnt will be consumed by
2837 * "commit()" or removed by "cancel()"
2839 static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2842 struct mem_cgroup **memcgp)
2844 struct mem_cgroup *memcg;
2845 struct page_cgroup *pc;
2848 pc = lookup_page_cgroup(page);
2850 * Every swap fault against a single page tries to charge the
2851 * page, bail as early as possible. shmem_unuse() encounters
2852 * already charged pages, too. The USED bit is protected by
2853 * the page lock, which serializes swap cache removal, which
2854 * in turn serializes uncharging.
2856 if (PageCgroupUsed(pc))
2858 if (!do_swap_account)
2860 memcg = try_get_mem_cgroup_from_page(page);
2864 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
2865 css_put(&memcg->css);
2870 ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
2876 int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page,
2877 gfp_t gfp_mask, struct mem_cgroup **memcgp)
2880 if (mem_cgroup_disabled())
2883 * A racing thread's fault, or swapoff, may have already
2884 * updated the pte, and even removed page from swap cache: in
2885 * those cases unuse_pte()'s pte_same() test will fail; but
2886 * there's also a KSM case which does need to charge the page.
2888 if (!PageSwapCache(page)) {
2891 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, memcgp, true);
2896 return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp);
2899 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
2901 if (mem_cgroup_disabled())
2905 __mem_cgroup_cancel_charge(memcg, 1);
2909 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
2910 enum charge_type ctype)
2912 if (mem_cgroup_disabled())
2917 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
2919 * Now swap is on-memory. This means this page may be
2920 * counted both as mem and swap....double count.
2921 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2922 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2923 * may call delete_from_swap_cache() before reach here.
2925 if (do_swap_account && PageSwapCache(page)) {
2926 swp_entry_t ent = {.val = page_private(page)};
2927 mem_cgroup_uncharge_swap(ent);
2931 void mem_cgroup_commit_charge_swapin(struct page *page,
2932 struct mem_cgroup *memcg)
2934 __mem_cgroup_commit_charge_swapin(page, memcg,
2935 MEM_CGROUP_CHARGE_TYPE_ANON);
2938 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2941 struct mem_cgroup *memcg = NULL;
2942 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
2945 if (mem_cgroup_disabled())
2947 if (PageCompound(page))
2950 if (!PageSwapCache(page))
2951 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
2952 else { /* page is swapcache/shmem */
2953 ret = __mem_cgroup_try_charge_swapin(mm, page,
2956 __mem_cgroup_commit_charge_swapin(page, memcg, type);
2961 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
2962 unsigned int nr_pages,
2963 const enum charge_type ctype)
2965 struct memcg_batch_info *batch = NULL;
2966 bool uncharge_memsw = true;
2968 /* If swapout, usage of swap doesn't decrease */
2969 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2970 uncharge_memsw = false;
2972 batch = ¤t->memcg_batch;
2974 * In usual, we do css_get() when we remember memcg pointer.
2975 * But in this case, we keep res->usage until end of a series of
2976 * uncharges. Then, it's ok to ignore memcg's refcnt.
2979 batch->memcg = memcg;
2981 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2982 * In those cases, all pages freed continuously can be expected to be in
2983 * the same cgroup and we have chance to coalesce uncharges.
2984 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2985 * because we want to do uncharge as soon as possible.
2988 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
2989 goto direct_uncharge;
2992 goto direct_uncharge;
2995 * In typical case, batch->memcg == mem. This means we can
2996 * merge a series of uncharges to an uncharge of res_counter.
2997 * If not, we uncharge res_counter ony by one.
2999 if (batch->memcg != memcg)
3000 goto direct_uncharge;
3001 /* remember freed charge and uncharge it later */
3004 batch->memsw_nr_pages++;
3007 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
3009 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
3010 if (unlikely(batch->memcg != memcg))
3011 memcg_oom_recover(memcg);
3015 * uncharge if !page_mapped(page)
3017 static struct mem_cgroup *
3018 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
3021 struct mem_cgroup *memcg = NULL;
3022 unsigned int nr_pages = 1;
3023 struct page_cgroup *pc;
3026 if (mem_cgroup_disabled())
3029 VM_BUG_ON(PageSwapCache(page));
3031 if (PageTransHuge(page)) {
3032 nr_pages <<= compound_order(page);
3033 VM_BUG_ON(!PageTransHuge(page));
3036 * Check if our page_cgroup is valid
3038 pc = lookup_page_cgroup(page);
3039 if (unlikely(!PageCgroupUsed(pc)))
3042 lock_page_cgroup(pc);
3044 memcg = pc->mem_cgroup;
3046 if (!PageCgroupUsed(pc))
3049 anon = PageAnon(page);
3052 case MEM_CGROUP_CHARGE_TYPE_ANON:
3054 * Generally PageAnon tells if it's the anon statistics to be
3055 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
3056 * used before page reached the stage of being marked PageAnon.
3060 case MEM_CGROUP_CHARGE_TYPE_DROP:
3061 /* See mem_cgroup_prepare_migration() */
3062 if (page_mapped(page))
3065 * Pages under migration may not be uncharged. But
3066 * end_migration() /must/ be the one uncharging the
3067 * unused post-migration page and so it has to call
3068 * here with the migration bit still set. See the
3069 * res_counter handling below.
3071 if (!end_migration && PageCgroupMigration(pc))
3074 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
3075 if (!PageAnon(page)) { /* Shared memory */
3076 if (page->mapping && !page_is_file_cache(page))
3078 } else if (page_mapped(page)) /* Anon */
3085 mem_cgroup_charge_statistics(memcg, anon, -nr_pages);
3087 ClearPageCgroupUsed(pc);
3089 * pc->mem_cgroup is not cleared here. It will be accessed when it's
3090 * freed from LRU. This is safe because uncharged page is expected not
3091 * to be reused (freed soon). Exception is SwapCache, it's handled by
3092 * special functions.
3095 unlock_page_cgroup(pc);
3097 * even after unlock, we have memcg->res.usage here and this memcg
3098 * will never be freed.
3100 memcg_check_events(memcg, page);
3101 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
3102 mem_cgroup_swap_statistics(memcg, true);
3103 mem_cgroup_get(memcg);
3106 * Migration does not charge the res_counter for the
3107 * replacement page, so leave it alone when phasing out the
3108 * page that is unused after the migration.
3110 if (!end_migration && !mem_cgroup_is_root(memcg))
3111 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
3116 unlock_page_cgroup(pc);
3120 void mem_cgroup_uncharge_page(struct page *page)
3123 if (page_mapped(page))
3125 VM_BUG_ON(page->mapping && !PageAnon(page));
3126 if (PageSwapCache(page))
3128 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
3131 void mem_cgroup_uncharge_cache_page(struct page *page)
3133 VM_BUG_ON(page_mapped(page));
3134 VM_BUG_ON(page->mapping);
3135 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
3139 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
3140 * In that cases, pages are freed continuously and we can expect pages
3141 * are in the same memcg. All these calls itself limits the number of
3142 * pages freed at once, then uncharge_start/end() is called properly.
3143 * This may be called prural(2) times in a context,
3146 void mem_cgroup_uncharge_start(void)
3148 current->memcg_batch.do_batch++;
3149 /* We can do nest. */
3150 if (current->memcg_batch.do_batch == 1) {
3151 current->memcg_batch.memcg = NULL;
3152 current->memcg_batch.nr_pages = 0;
3153 current->memcg_batch.memsw_nr_pages = 0;
3157 void mem_cgroup_uncharge_end(void)
3159 struct memcg_batch_info *batch = ¤t->memcg_batch;
3161 if (!batch->do_batch)
3165 if (batch->do_batch) /* If stacked, do nothing. */
3171 * This "batch->memcg" is valid without any css_get/put etc...
3172 * bacause we hide charges behind us.
3174 if (batch->nr_pages)
3175 res_counter_uncharge(&batch->memcg->res,
3176 batch->nr_pages * PAGE_SIZE);
3177 if (batch->memsw_nr_pages)
3178 res_counter_uncharge(&batch->memcg->memsw,
3179 batch->memsw_nr_pages * PAGE_SIZE);
3180 memcg_oom_recover(batch->memcg);
3181 /* forget this pointer (for sanity check) */
3182 batch->memcg = NULL;
3187 * called after __delete_from_swap_cache() and drop "page" account.
3188 * memcg information is recorded to swap_cgroup of "ent"
3191 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
3193 struct mem_cgroup *memcg;
3194 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
3196 if (!swapout) /* this was a swap cache but the swap is unused ! */
3197 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
3199 memcg = __mem_cgroup_uncharge_common(page, ctype, false);
3202 * record memcg information, if swapout && memcg != NULL,
3203 * mem_cgroup_get() was called in uncharge().
3205 if (do_swap_account && swapout && memcg)
3206 swap_cgroup_record(ent, css_id(&memcg->css));
3210 #ifdef CONFIG_MEMCG_SWAP
3212 * called from swap_entry_free(). remove record in swap_cgroup and
3213 * uncharge "memsw" account.
3215 void mem_cgroup_uncharge_swap(swp_entry_t ent)
3217 struct mem_cgroup *memcg;
3220 if (!do_swap_account)
3223 id = swap_cgroup_record(ent, 0);
3225 memcg = mem_cgroup_lookup(id);
3228 * We uncharge this because swap is freed.
3229 * This memcg can be obsolete one. We avoid calling css_tryget
3231 if (!mem_cgroup_is_root(memcg))
3232 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
3233 mem_cgroup_swap_statistics(memcg, false);
3234 mem_cgroup_put(memcg);
3240 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3241 * @entry: swap entry to be moved
3242 * @from: mem_cgroup which the entry is moved from
3243 * @to: mem_cgroup which the entry is moved to
3245 * It succeeds only when the swap_cgroup's record for this entry is the same
3246 * as the mem_cgroup's id of @from.
3248 * Returns 0 on success, -EINVAL on failure.
3250 * The caller must have charged to @to, IOW, called res_counter_charge() about
3251 * both res and memsw, and called css_get().
3253 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3254 struct mem_cgroup *from, struct mem_cgroup *to)
3256 unsigned short old_id, new_id;
3258 old_id = css_id(&from->css);
3259 new_id = css_id(&to->css);
3261 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3262 mem_cgroup_swap_statistics(from, false);
3263 mem_cgroup_swap_statistics(to, true);
3265 * This function is only called from task migration context now.
3266 * It postpones res_counter and refcount handling till the end
3267 * of task migration(mem_cgroup_clear_mc()) for performance
3268 * improvement. But we cannot postpone mem_cgroup_get(to)
3269 * because if the process that has been moved to @to does
3270 * swap-in, the refcount of @to might be decreased to 0.
3278 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3279 struct mem_cgroup *from, struct mem_cgroup *to)
3286 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
3289 void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
3290 struct mem_cgroup **memcgp)
3292 struct mem_cgroup *memcg = NULL;
3293 struct page_cgroup *pc;
3294 enum charge_type ctype;
3298 VM_BUG_ON(PageTransHuge(page));
3299 if (mem_cgroup_disabled())
3302 pc = lookup_page_cgroup(page);
3303 lock_page_cgroup(pc);
3304 if (PageCgroupUsed(pc)) {
3305 memcg = pc->mem_cgroup;
3306 css_get(&memcg->css);
3308 * At migrating an anonymous page, its mapcount goes down
3309 * to 0 and uncharge() will be called. But, even if it's fully
3310 * unmapped, migration may fail and this page has to be
3311 * charged again. We set MIGRATION flag here and delay uncharge
3312 * until end_migration() is called
3314 * Corner Case Thinking
3316 * When the old page was mapped as Anon and it's unmap-and-freed
3317 * while migration was ongoing.
3318 * If unmap finds the old page, uncharge() of it will be delayed
3319 * until end_migration(). If unmap finds a new page, it's
3320 * uncharged when it make mapcount to be 1->0. If unmap code
3321 * finds swap_migration_entry, the new page will not be mapped
3322 * and end_migration() will find it(mapcount==0).
3325 * When the old page was mapped but migraion fails, the kernel
3326 * remaps it. A charge for it is kept by MIGRATION flag even
3327 * if mapcount goes down to 0. We can do remap successfully
3328 * without charging it again.
3331 * The "old" page is under lock_page() until the end of
3332 * migration, so, the old page itself will not be swapped-out.
3333 * If the new page is swapped out before end_migraton, our
3334 * hook to usual swap-out path will catch the event.
3337 SetPageCgroupMigration(pc);
3339 unlock_page_cgroup(pc);
3341 * If the page is not charged at this point,
3349 * We charge new page before it's used/mapped. So, even if unlock_page()
3350 * is called before end_migration, we can catch all events on this new
3351 * page. In the case new page is migrated but not remapped, new page's
3352 * mapcount will be finally 0 and we call uncharge in end_migration().
3355 ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
3357 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
3359 * The page is committed to the memcg, but it's not actually
3360 * charged to the res_counter since we plan on replacing the
3361 * old one and only one page is going to be left afterwards.
3363 __mem_cgroup_commit_charge(memcg, newpage, 1, ctype, false);
3366 /* remove redundant charge if migration failed*/
3367 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
3368 struct page *oldpage, struct page *newpage, bool migration_ok)
3370 struct page *used, *unused;
3371 struct page_cgroup *pc;
3377 if (!migration_ok) {
3384 anon = PageAnon(used);
3385 __mem_cgroup_uncharge_common(unused,
3386 anon ? MEM_CGROUP_CHARGE_TYPE_ANON
3387 : MEM_CGROUP_CHARGE_TYPE_CACHE,
3389 css_put(&memcg->css);
3391 * We disallowed uncharge of pages under migration because mapcount
3392 * of the page goes down to zero, temporarly.
3393 * Clear the flag and check the page should be charged.
3395 pc = lookup_page_cgroup(oldpage);
3396 lock_page_cgroup(pc);
3397 ClearPageCgroupMigration(pc);
3398 unlock_page_cgroup(pc);
3401 * If a page is a file cache, radix-tree replacement is very atomic
3402 * and we can skip this check. When it was an Anon page, its mapcount
3403 * goes down to 0. But because we added MIGRATION flage, it's not
3404 * uncharged yet. There are several case but page->mapcount check
3405 * and USED bit check in mem_cgroup_uncharge_page() will do enough
3406 * check. (see prepare_charge() also)
3409 mem_cgroup_uncharge_page(used);
3413 * At replace page cache, newpage is not under any memcg but it's on
3414 * LRU. So, this function doesn't touch res_counter but handles LRU
3415 * in correct way. Both pages are locked so we cannot race with uncharge.
3417 void mem_cgroup_replace_page_cache(struct page *oldpage,
3418 struct page *newpage)
3420 struct mem_cgroup *memcg = NULL;
3421 struct page_cgroup *pc;
3422 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
3424 if (mem_cgroup_disabled())
3427 pc = lookup_page_cgroup(oldpage);
3428 /* fix accounting on old pages */
3429 lock_page_cgroup(pc);
3430 if (PageCgroupUsed(pc)) {
3431 memcg = pc->mem_cgroup;
3432 mem_cgroup_charge_statistics(memcg, false, -1);
3433 ClearPageCgroupUsed(pc);
3435 unlock_page_cgroup(pc);
3438 * When called from shmem_replace_page(), in some cases the
3439 * oldpage has already been charged, and in some cases not.
3444 * Even if newpage->mapping was NULL before starting replacement,
3445 * the newpage may be on LRU(or pagevec for LRU) already. We lock
3446 * LRU while we overwrite pc->mem_cgroup.
3448 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
3451 #ifdef CONFIG_DEBUG_VM
3452 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3454 struct page_cgroup *pc;
3456 pc = lookup_page_cgroup(page);
3458 * Can be NULL while feeding pages into the page allocator for
3459 * the first time, i.e. during boot or memory hotplug;
3460 * or when mem_cgroup_disabled().
3462 if (likely(pc) && PageCgroupUsed(pc))
3467 bool mem_cgroup_bad_page_check(struct page *page)
3469 if (mem_cgroup_disabled())
3472 return lookup_page_cgroup_used(page) != NULL;
3475 void mem_cgroup_print_bad_page(struct page *page)
3477 struct page_cgroup *pc;
3479 pc = lookup_page_cgroup_used(page);
3481 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
3482 pc, pc->flags, pc->mem_cgroup);
3487 static DEFINE_MUTEX(set_limit_mutex);
3489 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3490 unsigned long long val)
3493 u64 memswlimit, memlimit;
3495 int children = mem_cgroup_count_children(memcg);
3496 u64 curusage, oldusage;
3500 * For keeping hierarchical_reclaim simple, how long we should retry
3501 * is depends on callers. We set our retry-count to be function
3502 * of # of children which we should visit in this loop.
3504 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3506 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3509 while (retry_count) {
3510 if (signal_pending(current)) {
3515 * Rather than hide all in some function, I do this in
3516 * open coded manner. You see what this really does.
3517 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
3519 mutex_lock(&set_limit_mutex);
3520 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3521 if (memswlimit < val) {
3523 mutex_unlock(&set_limit_mutex);
3527 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3531 ret = res_counter_set_limit(&memcg->res, val);
3533 if (memswlimit == val)
3534 memcg->memsw_is_minimum = true;
3536 memcg->memsw_is_minimum = false;
3538 mutex_unlock(&set_limit_mutex);
3543 mem_cgroup_reclaim(memcg, GFP_KERNEL,
3544 MEM_CGROUP_RECLAIM_SHRINK);
3545 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3546 /* Usage is reduced ? */
3547 if (curusage >= oldusage)
3550 oldusage = curusage;
3552 if (!ret && enlarge)
3553 memcg_oom_recover(memcg);
3558 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3559 unsigned long long val)
3562 u64 memlimit, memswlimit, oldusage, curusage;
3563 int children = mem_cgroup_count_children(memcg);
3567 /* see mem_cgroup_resize_res_limit */
3568 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3569 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3570 while (retry_count) {
3571 if (signal_pending(current)) {
3576 * Rather than hide all in some function, I do this in
3577 * open coded manner. You see what this really does.
3578 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
3580 mutex_lock(&set_limit_mutex);
3581 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3582 if (memlimit > val) {
3584 mutex_unlock(&set_limit_mutex);
3587 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3588 if (memswlimit < val)
3590 ret = res_counter_set_limit(&memcg->memsw, val);
3592 if (memlimit == val)
3593 memcg->memsw_is_minimum = true;
3595 memcg->memsw_is_minimum = false;
3597 mutex_unlock(&set_limit_mutex);
3602 mem_cgroup_reclaim(memcg, GFP_KERNEL,
3603 MEM_CGROUP_RECLAIM_NOSWAP |
3604 MEM_CGROUP_RECLAIM_SHRINK);
3605 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3606 /* Usage is reduced ? */
3607 if (curusage >= oldusage)
3610 oldusage = curusage;
3612 if (!ret && enlarge)
3613 memcg_oom_recover(memcg);
3617 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3619 unsigned long *total_scanned)
3621 unsigned long nr_reclaimed = 0;
3622 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3623 unsigned long reclaimed;
3625 struct mem_cgroup_tree_per_zone *mctz;
3626 unsigned long long excess;
3627 unsigned long nr_scanned;
3632 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3634 * This loop can run a while, specially if mem_cgroup's continuously
3635 * keep exceeding their soft limit and putting the system under
3642 mz = mem_cgroup_largest_soft_limit_node(mctz);
3647 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
3648 gfp_mask, &nr_scanned);
3649 nr_reclaimed += reclaimed;
3650 *total_scanned += nr_scanned;
3651 spin_lock(&mctz->lock);
3654 * If we failed to reclaim anything from this memory cgroup
3655 * it is time to move on to the next cgroup
3661 * Loop until we find yet another one.
3663 * By the time we get the soft_limit lock
3664 * again, someone might have aded the
3665 * group back on the RB tree. Iterate to
3666 * make sure we get a different mem.
3667 * mem_cgroup_largest_soft_limit_node returns
3668 * NULL if no other cgroup is present on
3672 __mem_cgroup_largest_soft_limit_node(mctz);
3674 css_put(&next_mz->memcg->css);
3675 else /* next_mz == NULL or other memcg */
3679 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
3680 excess = res_counter_soft_limit_excess(&mz->memcg->res);
3682 * One school of thought says that we should not add
3683 * back the node to the tree if reclaim returns 0.
3684 * But our reclaim could return 0, simply because due
3685 * to priority we are exposing a smaller subset of
3686 * memory to reclaim from. Consider this as a longer
3689 /* If excess == 0, no tree ops */
3690 __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess);
3691 spin_unlock(&mctz->lock);
3692 css_put(&mz->memcg->css);
3695 * Could not reclaim anything and there are no more
3696 * mem cgroups to try or we seem to be looping without
3697 * reclaiming anything.
3699 if (!nr_reclaimed &&
3701 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3703 } while (!nr_reclaimed);
3705 css_put(&next_mz->memcg->css);
3706 return nr_reclaimed;
3710 * mem_cgroup_force_empty_list - clears LRU of a group
3711 * @memcg: group to clear
3714 * @lru: lru to to clear
3716 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
3717 * reclaim the pages page themselves - pages are moved to the parent (or root)
3720 static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
3721 int node, int zid, enum lru_list lru)
3723 struct lruvec *lruvec;
3724 unsigned long flags;
3725 struct list_head *list;
3729 zone = &NODE_DATA(node)->node_zones[zid];
3730 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
3731 list = &lruvec->lists[lru];
3735 struct page_cgroup *pc;
3738 spin_lock_irqsave(&zone->lru_lock, flags);
3739 if (list_empty(list)) {
3740 spin_unlock_irqrestore(&zone->lru_lock, flags);
3743 page = list_entry(list->prev, struct page, lru);
3745 list_move(&page->lru, list);
3747 spin_unlock_irqrestore(&zone->lru_lock, flags);
3750 spin_unlock_irqrestore(&zone->lru_lock, flags);
3752 pc = lookup_page_cgroup(page);
3754 if (mem_cgroup_move_parent(page, pc, memcg)) {
3755 /* found lock contention or "pc" is obsolete. */
3760 } while (!list_empty(list));
3764 * make mem_cgroup's charge to be 0 if there is no task by moving
3765 * all the charges and pages to the parent.
3766 * This enables deleting this mem_cgroup.
3768 * Caller is responsible for holding css reference on the memcg.
3770 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
3775 /* This is for making all *used* pages to be on LRU. */
3776 lru_add_drain_all();
3777 drain_all_stock_sync(memcg);
3778 mem_cgroup_start_move(memcg);
3779 for_each_node_state(node, N_HIGH_MEMORY) {
3780 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3783 mem_cgroup_force_empty_list(memcg,
3788 mem_cgroup_end_move(memcg);
3789 memcg_oom_recover(memcg);
3793 * This is a safety check because mem_cgroup_force_empty_list
3794 * could have raced with mem_cgroup_replace_page_cache callers
3795 * so the lru seemed empty but the page could have been added
3796 * right after the check. RES_USAGE should be safe as we always
3797 * charge before adding to the LRU.
3799 } while (res_counter_read_u64(&memcg->res, RES_USAGE) > 0);
3803 * Reclaims as many pages from the given memcg as possible and moves
3804 * the rest to the parent.
3806 * Caller is responsible for holding css reference for memcg.
3808 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3810 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3811 struct cgroup *cgrp = memcg->css.cgroup;
3813 /* returns EBUSY if there is a task or if we come here twice. */
3814 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3817 /* we call try-to-free pages for make this cgroup empty */
3818 lru_add_drain_all();
3819 /* try to free all pages in this cgroup */
3820 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
3823 if (signal_pending(current))
3826 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
3830 /* maybe some writeback is necessary */
3831 congestion_wait(BLK_RW_ASYNC, HZ/10);
3836 mem_cgroup_reparent_charges(memcg);
3841 static int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3843 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3846 if (mem_cgroup_is_root(memcg))
3848 css_get(&memcg->css);
3849 ret = mem_cgroup_force_empty(memcg);
3850 css_put(&memcg->css);
3856 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3858 return mem_cgroup_from_cont(cont)->use_hierarchy;
3861 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3865 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3866 struct cgroup *parent = cont->parent;
3867 struct mem_cgroup *parent_memcg = NULL;
3870 parent_memcg = mem_cgroup_from_cont(parent);
3874 if (memcg->use_hierarchy == val)
3878 * If parent's use_hierarchy is set, we can't make any modifications
3879 * in the child subtrees. If it is unset, then the change can
3880 * occur, provided the current cgroup has no children.
3882 * For the root cgroup, parent_mem is NULL, we allow value to be
3883 * set if there are no children.
3885 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3886 (val == 1 || val == 0)) {
3887 if (list_empty(&cont->children))
3888 memcg->use_hierarchy = val;
3901 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
3902 enum mem_cgroup_stat_index idx)
3904 struct mem_cgroup *iter;
3907 /* Per-cpu values can be negative, use a signed accumulator */
3908 for_each_mem_cgroup_tree(iter, memcg)
3909 val += mem_cgroup_read_stat(iter, idx);
3911 if (val < 0) /* race ? */
3916 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3920 if (!mem_cgroup_is_root(memcg)) {
3922 return res_counter_read_u64(&memcg->res, RES_USAGE);
3924 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
3927 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
3928 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
3931 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
3933 return val << PAGE_SHIFT;
3936 static ssize_t mem_cgroup_read(struct cgroup *cont, struct cftype *cft,
3937 struct file *file, char __user *buf,
3938 size_t nbytes, loff_t *ppos)
3940 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3943 int type, name, len;
3945 type = MEMFILE_TYPE(cft->private);
3946 name = MEMFILE_ATTR(cft->private);
3948 if (!do_swap_account && type == _MEMSWAP)
3953 if (name == RES_USAGE)
3954 val = mem_cgroup_usage(memcg, false);
3956 val = res_counter_read_u64(&memcg->res, name);
3959 if (name == RES_USAGE)
3960 val = mem_cgroup_usage(memcg, true);
3962 val = res_counter_read_u64(&memcg->memsw, name);
3968 len = scnprintf(str, sizeof(str), "%llu\n", (unsigned long long)val);
3969 return simple_read_from_buffer(buf, nbytes, ppos, str, len);
3972 * The user of this function is...
3975 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
3978 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3980 unsigned long long val;
3983 type = MEMFILE_TYPE(cft->private);
3984 name = MEMFILE_ATTR(cft->private);
3986 if (!do_swap_account && type == _MEMSWAP)
3991 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3995 /* This function does all necessary parse...reuse it */
3996 ret = res_counter_memparse_write_strategy(buffer, &val);
4000 ret = mem_cgroup_resize_limit(memcg, val);
4002 ret = mem_cgroup_resize_memsw_limit(memcg, val);
4004 case RES_SOFT_LIMIT:
4005 ret = res_counter_memparse_write_strategy(buffer, &val);
4009 * For memsw, soft limits are hard to implement in terms
4010 * of semantics, for now, we support soft limits for
4011 * control without swap
4014 ret = res_counter_set_soft_limit(&memcg->res, val);
4019 ret = -EINVAL; /* should be BUG() ? */
4025 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
4026 unsigned long long *mem_limit, unsigned long long *memsw_limit)
4028 struct cgroup *cgroup;
4029 unsigned long long min_limit, min_memsw_limit, tmp;
4031 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4032 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4033 cgroup = memcg->css.cgroup;
4034 if (!memcg->use_hierarchy)
4037 while (cgroup->parent) {
4038 cgroup = cgroup->parent;
4039 memcg = mem_cgroup_from_cont(cgroup);
4040 if (!memcg->use_hierarchy)
4042 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
4043 min_limit = min(min_limit, tmp);
4044 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4045 min_memsw_limit = min(min_memsw_limit, tmp);
4048 *mem_limit = min_limit;
4049 *memsw_limit = min_memsw_limit;
4052 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
4054 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4057 type = MEMFILE_TYPE(event);
4058 name = MEMFILE_ATTR(event);
4060 if (!do_swap_account && type == _MEMSWAP)
4066 res_counter_reset_max(&memcg->res);
4068 res_counter_reset_max(&memcg->memsw);
4072 res_counter_reset_failcnt(&memcg->res);
4074 res_counter_reset_failcnt(&memcg->memsw);
4081 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
4084 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
4088 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4089 struct cftype *cft, u64 val)
4091 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4093 if (val >= (1 << NR_MOVE_TYPE))
4096 * We check this value several times in both in can_attach() and
4097 * attach(), so we need cgroup lock to prevent this value from being
4101 memcg->move_charge_at_immigrate = val;
4107 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4108 struct cftype *cft, u64 val)
4115 static int memcg_numa_stat_show(struct cgroup *cont, struct cftype *cft,
4119 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
4120 unsigned long node_nr;
4121 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4123 total_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL);
4124 seq_printf(m, "total=%lu", total_nr);
4125 for_each_node_state(nid, N_HIGH_MEMORY) {
4126 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL);
4127 seq_printf(m, " N%d=%lu", nid, node_nr);
4131 file_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_FILE);
4132 seq_printf(m, "file=%lu", file_nr);
4133 for_each_node_state(nid, N_HIGH_MEMORY) {
4134 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4136 seq_printf(m, " N%d=%lu", nid, node_nr);
4140 anon_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_ANON);
4141 seq_printf(m, "anon=%lu", anon_nr);
4142 for_each_node_state(nid, N_HIGH_MEMORY) {
4143 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4145 seq_printf(m, " N%d=%lu", nid, node_nr);
4149 unevictable_nr = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
4150 seq_printf(m, "unevictable=%lu", unevictable_nr);
4151 for_each_node_state(nid, N_HIGH_MEMORY) {
4152 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4153 BIT(LRU_UNEVICTABLE));
4154 seq_printf(m, " N%d=%lu", nid, node_nr);
4159 #endif /* CONFIG_NUMA */
4161 static const char * const mem_cgroup_lru_names[] = {
4169 static inline void mem_cgroup_lru_names_not_uptodate(void)
4171 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
4174 static int memcg_stat_show(struct cgroup *cont, struct cftype *cft,
4177 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4178 struct mem_cgroup *mi;
4181 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
4182 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
4184 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
4185 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
4188 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
4189 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
4190 mem_cgroup_read_events(memcg, i));
4192 for (i = 0; i < NR_LRU_LISTS; i++)
4193 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
4194 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
4196 /* Hierarchical information */
4198 unsigned long long limit, memsw_limit;
4199 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
4200 seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
4201 if (do_swap_account)
4202 seq_printf(m, "hierarchical_memsw_limit %llu\n",
4206 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
4209 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
4211 for_each_mem_cgroup_tree(mi, memcg)
4212 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
4213 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
4216 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
4217 unsigned long long val = 0;
4219 for_each_mem_cgroup_tree(mi, memcg)
4220 val += mem_cgroup_read_events(mi, i);
4221 seq_printf(m, "total_%s %llu\n",
4222 mem_cgroup_events_names[i], val);
4225 for (i = 0; i < NR_LRU_LISTS; i++) {
4226 unsigned long long val = 0;
4228 for_each_mem_cgroup_tree(mi, memcg)
4229 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
4230 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
4233 #ifdef CONFIG_DEBUG_VM
4236 struct mem_cgroup_per_zone *mz;
4237 struct zone_reclaim_stat *rstat;
4238 unsigned long recent_rotated[2] = {0, 0};
4239 unsigned long recent_scanned[2] = {0, 0};
4241 for_each_online_node(nid)
4242 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4243 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
4244 rstat = &mz->lruvec.reclaim_stat;
4246 recent_rotated[0] += rstat->recent_rotated[0];
4247 recent_rotated[1] += rstat->recent_rotated[1];
4248 recent_scanned[0] += rstat->recent_scanned[0];
4249 recent_scanned[1] += rstat->recent_scanned[1];
4251 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
4252 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
4253 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
4254 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
4261 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
4263 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4265 return mem_cgroup_swappiness(memcg);
4268 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
4271 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4272 struct mem_cgroup *parent;
4277 if (cgrp->parent == NULL)
4280 parent = mem_cgroup_from_cont(cgrp->parent);
4284 /* If under hierarchy, only empty-root can set this value */
4285 if ((parent->use_hierarchy) ||
4286 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4291 memcg->swappiness = val;
4298 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4300 struct mem_cgroup_threshold_ary *t;
4306 t = rcu_dereference(memcg->thresholds.primary);
4308 t = rcu_dereference(memcg->memsw_thresholds.primary);
4313 usage = mem_cgroup_usage(memcg, swap);
4316 * current_threshold points to threshold just below or equal to usage.
4317 * If it's not true, a threshold was crossed after last
4318 * call of __mem_cgroup_threshold().
4320 i = t->current_threshold;
4323 * Iterate backward over array of thresholds starting from
4324 * current_threshold and check if a threshold is crossed.
4325 * If none of thresholds below usage is crossed, we read
4326 * only one element of the array here.
4328 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4329 eventfd_signal(t->entries[i].eventfd, 1);
4331 /* i = current_threshold + 1 */
4335 * Iterate forward over array of thresholds starting from
4336 * current_threshold+1 and check if a threshold is crossed.
4337 * If none of thresholds above usage is crossed, we read
4338 * only one element of the array here.
4340 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4341 eventfd_signal(t->entries[i].eventfd, 1);
4343 /* Update current_threshold */
4344 t->current_threshold = i - 1;
4349 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4352 __mem_cgroup_threshold(memcg, false);
4353 if (do_swap_account)
4354 __mem_cgroup_threshold(memcg, true);
4356 memcg = parent_mem_cgroup(memcg);
4360 static int compare_thresholds(const void *a, const void *b)
4362 const struct mem_cgroup_threshold *_a = a;
4363 const struct mem_cgroup_threshold *_b = b;
4365 return _a->threshold - _b->threshold;
4368 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4370 struct mem_cgroup_eventfd_list *ev;
4372 list_for_each_entry(ev, &memcg->oom_notify, list)
4373 eventfd_signal(ev->eventfd, 1);
4377 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4379 struct mem_cgroup *iter;
4381 for_each_mem_cgroup_tree(iter, memcg)
4382 mem_cgroup_oom_notify_cb(iter);
4385 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
4386 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4388 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4389 struct mem_cgroup_thresholds *thresholds;
4390 struct mem_cgroup_threshold_ary *new;
4391 int type = MEMFILE_TYPE(cft->private);
4392 u64 threshold, usage;
4395 ret = res_counter_memparse_write_strategy(args, &threshold);
4399 mutex_lock(&memcg->thresholds_lock);
4402 thresholds = &memcg->thresholds;
4403 else if (type == _MEMSWAP)
4404 thresholds = &memcg->memsw_thresholds;
4408 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4410 /* Check if a threshold crossed before adding a new one */
4411 if (thresholds->primary)
4412 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4414 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4416 /* Allocate memory for new array of thresholds */
4417 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4425 /* Copy thresholds (if any) to new array */
4426 if (thresholds->primary) {
4427 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4428 sizeof(struct mem_cgroup_threshold));
4431 /* Add new threshold */
4432 new->entries[size - 1].eventfd = eventfd;
4433 new->entries[size - 1].threshold = threshold;
4435 /* Sort thresholds. Registering of new threshold isn't time-critical */
4436 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4437 compare_thresholds, NULL);
4439 /* Find current threshold */
4440 new->current_threshold = -1;
4441 for (i = 0; i < size; i++) {
4442 if (new->entries[i].threshold <= usage) {
4444 * new->current_threshold will not be used until
4445 * rcu_assign_pointer(), so it's safe to increment
4448 ++new->current_threshold;
4453 /* Free old spare buffer and save old primary buffer as spare */
4454 kfree(thresholds->spare);
4455 thresholds->spare = thresholds->primary;
4457 rcu_assign_pointer(thresholds->primary, new);
4459 /* To be sure that nobody uses thresholds */
4463 mutex_unlock(&memcg->thresholds_lock);
4468 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
4469 struct cftype *cft, struct eventfd_ctx *eventfd)
4471 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4472 struct mem_cgroup_thresholds *thresholds;
4473 struct mem_cgroup_threshold_ary *new;
4474 int type = MEMFILE_TYPE(cft->private);
4478 mutex_lock(&memcg->thresholds_lock);
4480 thresholds = &memcg->thresholds;
4481 else if (type == _MEMSWAP)
4482 thresholds = &memcg->memsw_thresholds;
4486 if (!thresholds->primary)
4489 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4491 /* Check if a threshold crossed before removing */
4492 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4494 /* Calculate new number of threshold */
4496 for (i = 0; i < thresholds->primary->size; i++) {
4497 if (thresholds->primary->entries[i].eventfd != eventfd)
4501 new = thresholds->spare;
4503 /* Set thresholds array to NULL if we don't have thresholds */
4512 /* Copy thresholds and find current threshold */
4513 new->current_threshold = -1;
4514 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4515 if (thresholds->primary->entries[i].eventfd == eventfd)
4518 new->entries[j] = thresholds->primary->entries[i];
4519 if (new->entries[j].threshold <= usage) {
4521 * new->current_threshold will not be used
4522 * until rcu_assign_pointer(), so it's safe to increment
4525 ++new->current_threshold;
4531 /* Swap primary and spare array */
4532 thresholds->spare = thresholds->primary;
4533 /* If all events are unregistered, free the spare array */
4535 kfree(thresholds->spare);
4536 thresholds->spare = NULL;
4539 rcu_assign_pointer(thresholds->primary, new);
4541 /* To be sure that nobody uses thresholds */
4544 mutex_unlock(&memcg->thresholds_lock);
4547 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4548 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4550 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4551 struct mem_cgroup_eventfd_list *event;
4552 int type = MEMFILE_TYPE(cft->private);
4554 BUG_ON(type != _OOM_TYPE);
4555 event = kmalloc(sizeof(*event), GFP_KERNEL);
4559 spin_lock(&memcg_oom_lock);
4561 event->eventfd = eventfd;
4562 list_add(&event->list, &memcg->oom_notify);
4564 /* already in OOM ? */
4565 if (atomic_read(&memcg->under_oom))
4566 eventfd_signal(eventfd, 1);
4567 spin_unlock(&memcg_oom_lock);
4572 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4573 struct cftype *cft, struct eventfd_ctx *eventfd)
4575 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4576 struct mem_cgroup_eventfd_list *ev, *tmp;
4577 int type = MEMFILE_TYPE(cft->private);
4579 BUG_ON(type != _OOM_TYPE);
4581 spin_lock(&memcg_oom_lock);
4583 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4584 if (ev->eventfd == eventfd) {
4585 list_del(&ev->list);
4590 spin_unlock(&memcg_oom_lock);
4593 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4594 struct cftype *cft, struct cgroup_map_cb *cb)
4596 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4598 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
4600 if (atomic_read(&memcg->under_oom))
4601 cb->fill(cb, "under_oom", 1);
4603 cb->fill(cb, "under_oom", 0);
4607 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4608 struct cftype *cft, u64 val)
4610 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4611 struct mem_cgroup *parent;
4613 /* cannot set to root cgroup and only 0 and 1 are allowed */
4614 if (!cgrp->parent || !((val == 0) || (val == 1)))
4617 parent = mem_cgroup_from_cont(cgrp->parent);
4620 /* oom-kill-disable is a flag for subhierarchy. */
4621 if ((parent->use_hierarchy) ||
4622 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4626 memcg->oom_kill_disable = val;
4628 memcg_oom_recover(memcg);
4633 #ifdef CONFIG_MEMCG_KMEM
4634 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
4636 return mem_cgroup_sockets_init(memcg, ss);
4639 static void kmem_cgroup_destroy(struct mem_cgroup *memcg)
4641 mem_cgroup_sockets_destroy(memcg);
4644 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
4649 static void kmem_cgroup_destroy(struct mem_cgroup *memcg)
4654 static struct cftype mem_cgroup_files[] = {
4656 .name = "usage_in_bytes",
4657 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4658 .read = mem_cgroup_read,
4659 .register_event = mem_cgroup_usage_register_event,
4660 .unregister_event = mem_cgroup_usage_unregister_event,
4663 .name = "max_usage_in_bytes",
4664 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4665 .trigger = mem_cgroup_reset,
4666 .read = mem_cgroup_read,
4669 .name = "limit_in_bytes",
4670 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4671 .write_string = mem_cgroup_write,
4672 .read = mem_cgroup_read,
4675 .name = "soft_limit_in_bytes",
4676 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4677 .write_string = mem_cgroup_write,
4678 .read = mem_cgroup_read,
4682 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4683 .trigger = mem_cgroup_reset,
4684 .read = mem_cgroup_read,
4688 .read_seq_string = memcg_stat_show,
4691 .name = "force_empty",
4692 .trigger = mem_cgroup_force_empty_write,
4695 .name = "use_hierarchy",
4696 .write_u64 = mem_cgroup_hierarchy_write,
4697 .read_u64 = mem_cgroup_hierarchy_read,
4700 .name = "swappiness",
4701 .read_u64 = mem_cgroup_swappiness_read,
4702 .write_u64 = mem_cgroup_swappiness_write,
4705 .name = "move_charge_at_immigrate",
4706 .read_u64 = mem_cgroup_move_charge_read,
4707 .write_u64 = mem_cgroup_move_charge_write,
4710 .name = "oom_control",
4711 .read_map = mem_cgroup_oom_control_read,
4712 .write_u64 = mem_cgroup_oom_control_write,
4713 .register_event = mem_cgroup_oom_register_event,
4714 .unregister_event = mem_cgroup_oom_unregister_event,
4715 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4719 .name = "numa_stat",
4720 .read_seq_string = memcg_numa_stat_show,
4723 #ifdef CONFIG_MEMCG_SWAP
4725 .name = "memsw.usage_in_bytes",
4726 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4727 .read = mem_cgroup_read,
4728 .register_event = mem_cgroup_usage_register_event,
4729 .unregister_event = mem_cgroup_usage_unregister_event,
4732 .name = "memsw.max_usage_in_bytes",
4733 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4734 .trigger = mem_cgroup_reset,
4735 .read = mem_cgroup_read,
4738 .name = "memsw.limit_in_bytes",
4739 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4740 .write_string = mem_cgroup_write,
4741 .read = mem_cgroup_read,
4744 .name = "memsw.failcnt",
4745 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4746 .trigger = mem_cgroup_reset,
4747 .read = mem_cgroup_read,
4750 { }, /* terminate */
4753 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4755 struct mem_cgroup_per_node *pn;
4756 struct mem_cgroup_per_zone *mz;
4757 int zone, tmp = node;
4759 * This routine is called against possible nodes.
4760 * But it's BUG to call kmalloc() against offline node.
4762 * TODO: this routine can waste much memory for nodes which will
4763 * never be onlined. It's better to use memory hotplug callback
4766 if (!node_state(node, N_NORMAL_MEMORY))
4768 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4772 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4773 mz = &pn->zoneinfo[zone];
4774 lruvec_init(&mz->lruvec);
4775 mz->usage_in_excess = 0;
4776 mz->on_tree = false;
4779 memcg->info.nodeinfo[node] = pn;
4783 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4785 kfree(memcg->info.nodeinfo[node]);
4788 static struct mem_cgroup *mem_cgroup_alloc(void)
4790 struct mem_cgroup *memcg;
4791 int size = sizeof(struct mem_cgroup);
4793 /* Can be very big if MAX_NUMNODES is very big */
4794 if (size < PAGE_SIZE)
4795 memcg = kzalloc(size, GFP_KERNEL);
4797 memcg = vzalloc(size);
4802 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4805 spin_lock_init(&memcg->pcp_counter_lock);
4809 if (size < PAGE_SIZE)
4817 * Helpers for freeing a kmalloc()ed/vzalloc()ed mem_cgroup by RCU,
4818 * but in process context. The work_freeing structure is overlaid
4819 * on the rcu_freeing structure, which itself is overlaid on memsw.
4821 static void free_work(struct work_struct *work)
4823 struct mem_cgroup *memcg;
4824 int size = sizeof(struct mem_cgroup);
4826 memcg = container_of(work, struct mem_cgroup, work_freeing);
4828 * We need to make sure that (at least for now), the jump label
4829 * destruction code runs outside of the cgroup lock. This is because
4830 * get_online_cpus(), which is called from the static_branch update,
4831 * can't be called inside the cgroup_lock. cpusets are the ones
4832 * enforcing this dependency, so if they ever change, we might as well.
4834 * schedule_work() will guarantee this happens. Be careful if you need
4835 * to move this code around, and make sure it is outside
4838 disarm_sock_keys(memcg);
4839 if (size < PAGE_SIZE)
4845 static void free_rcu(struct rcu_head *rcu_head)
4847 struct mem_cgroup *memcg;
4849 memcg = container_of(rcu_head, struct mem_cgroup, rcu_freeing);
4850 INIT_WORK(&memcg->work_freeing, free_work);
4851 schedule_work(&memcg->work_freeing);
4855 * At destroying mem_cgroup, references from swap_cgroup can remain.
4856 * (scanning all at force_empty is too costly...)
4858 * Instead of clearing all references at force_empty, we remember
4859 * the number of reference from swap_cgroup and free mem_cgroup when
4860 * it goes down to 0.
4862 * Removal of cgroup itself succeeds regardless of refs from swap.
4865 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4869 mem_cgroup_remove_from_trees(memcg);
4870 free_css_id(&mem_cgroup_subsys, &memcg->css);
4873 free_mem_cgroup_per_zone_info(memcg, node);
4875 free_percpu(memcg->stat);
4876 call_rcu(&memcg->rcu_freeing, free_rcu);
4879 static void mem_cgroup_get(struct mem_cgroup *memcg)
4881 atomic_inc(&memcg->refcnt);
4884 static void __mem_cgroup_put(struct mem_cgroup *memcg, int count)
4886 if (atomic_sub_and_test(count, &memcg->refcnt)) {
4887 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4888 __mem_cgroup_free(memcg);
4890 mem_cgroup_put(parent);
4894 static void mem_cgroup_put(struct mem_cgroup *memcg)
4896 __mem_cgroup_put(memcg, 1);
4900 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4902 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4904 if (!memcg->res.parent)
4906 return mem_cgroup_from_res_counter(memcg->res.parent, res);
4908 EXPORT_SYMBOL(parent_mem_cgroup);
4910 #ifdef CONFIG_MEMCG_SWAP
4911 static void __init enable_swap_cgroup(void)
4913 if (!mem_cgroup_disabled() && really_do_swap_account)
4914 do_swap_account = 1;
4917 static void __init enable_swap_cgroup(void)
4922 static int mem_cgroup_soft_limit_tree_init(void)
4924 struct mem_cgroup_tree_per_node *rtpn;
4925 struct mem_cgroup_tree_per_zone *rtpz;
4926 int tmp, node, zone;
4928 for_each_node(node) {
4930 if (!node_state(node, N_NORMAL_MEMORY))
4932 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
4936 soft_limit_tree.rb_tree_per_node[node] = rtpn;
4938 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4939 rtpz = &rtpn->rb_tree_per_zone[zone];
4940 rtpz->rb_root = RB_ROOT;
4941 spin_lock_init(&rtpz->lock);
4947 for_each_node(node) {
4948 if (!soft_limit_tree.rb_tree_per_node[node])
4950 kfree(soft_limit_tree.rb_tree_per_node[node]);
4951 soft_limit_tree.rb_tree_per_node[node] = NULL;
4957 static struct cgroup_subsys_state * __ref
4958 mem_cgroup_css_alloc(struct cgroup *cont)
4960 struct mem_cgroup *memcg, *parent;
4961 long error = -ENOMEM;
4964 memcg = mem_cgroup_alloc();
4966 return ERR_PTR(error);
4969 if (alloc_mem_cgroup_per_zone_info(memcg, node))
4973 if (cont->parent == NULL) {
4975 enable_swap_cgroup();
4977 if (mem_cgroup_soft_limit_tree_init())
4979 root_mem_cgroup = memcg;
4980 for_each_possible_cpu(cpu) {
4981 struct memcg_stock_pcp *stock =
4982 &per_cpu(memcg_stock, cpu);
4983 INIT_WORK(&stock->work, drain_local_stock);
4985 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
4987 parent = mem_cgroup_from_cont(cont->parent);
4988 memcg->use_hierarchy = parent->use_hierarchy;
4989 memcg->oom_kill_disable = parent->oom_kill_disable;
4992 if (parent && parent->use_hierarchy) {
4993 res_counter_init(&memcg->res, &parent->res);
4994 res_counter_init(&memcg->memsw, &parent->memsw);
4996 * We increment refcnt of the parent to ensure that we can
4997 * safely access it on res_counter_charge/uncharge.
4998 * This refcnt will be decremented when freeing this
4999 * mem_cgroup(see mem_cgroup_put).
5001 mem_cgroup_get(parent);
5003 res_counter_init(&memcg->res, NULL);
5004 res_counter_init(&memcg->memsw, NULL);
5006 * Deeper hierachy with use_hierarchy == false doesn't make
5007 * much sense so let cgroup subsystem know about this
5008 * unfortunate state in our controller.
5010 if (parent && parent != root_mem_cgroup)
5011 mem_cgroup_subsys.broken_hierarchy = true;
5013 memcg->last_scanned_node = MAX_NUMNODES;
5014 INIT_LIST_HEAD(&memcg->oom_notify);
5017 memcg->swappiness = mem_cgroup_swappiness(parent);
5018 atomic_set(&memcg->refcnt, 1);
5019 memcg->move_charge_at_immigrate = 0;
5020 mutex_init(&memcg->thresholds_lock);
5021 spin_lock_init(&memcg->move_lock);
5023 error = memcg_init_kmem(memcg, &mem_cgroup_subsys);
5026 * We call put now because our (and parent's) refcnts
5027 * are already in place. mem_cgroup_put() will internally
5028 * call __mem_cgroup_free, so return directly
5030 mem_cgroup_put(memcg);
5031 return ERR_PTR(error);
5035 __mem_cgroup_free(memcg);
5036 return ERR_PTR(error);
5039 static void mem_cgroup_css_offline(struct cgroup *cont)
5041 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5043 mem_cgroup_reparent_charges(memcg);
5046 static void mem_cgroup_css_free(struct cgroup *cont)
5048 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5050 kmem_cgroup_destroy(memcg);
5052 mem_cgroup_put(memcg);
5056 /* Handlers for move charge at task migration. */
5057 #define PRECHARGE_COUNT_AT_ONCE 256
5058 static int mem_cgroup_do_precharge(unsigned long count)
5061 int batch_count = PRECHARGE_COUNT_AT_ONCE;
5062 struct mem_cgroup *memcg = mc.to;
5064 if (mem_cgroup_is_root(memcg)) {
5065 mc.precharge += count;
5066 /* we don't need css_get for root */
5069 /* try to charge at once */
5071 struct res_counter *dummy;
5073 * "memcg" cannot be under rmdir() because we've already checked
5074 * by cgroup_lock_live_cgroup() that it is not removed and we
5075 * are still under the same cgroup_mutex. So we can postpone
5078 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
5080 if (do_swap_account && res_counter_charge(&memcg->memsw,
5081 PAGE_SIZE * count, &dummy)) {
5082 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
5085 mc.precharge += count;
5089 /* fall back to one by one charge */
5091 if (signal_pending(current)) {
5095 if (!batch_count--) {
5096 batch_count = PRECHARGE_COUNT_AT_ONCE;
5099 ret = __mem_cgroup_try_charge(NULL,
5100 GFP_KERNEL, 1, &memcg, false);
5102 /* mem_cgroup_clear_mc() will do uncharge later */
5110 * get_mctgt_type - get target type of moving charge
5111 * @vma: the vma the pte to be checked belongs
5112 * @addr: the address corresponding to the pte to be checked
5113 * @ptent: the pte to be checked
5114 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5117 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5118 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5119 * move charge. if @target is not NULL, the page is stored in target->page
5120 * with extra refcnt got(Callers should handle it).
5121 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5122 * target for charge migration. if @target is not NULL, the entry is stored
5125 * Called with pte lock held.
5132 enum mc_target_type {
5138 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5139 unsigned long addr, pte_t ptent)
5141 struct page *page = vm_normal_page(vma, addr, ptent);
5143 if (!page || !page_mapped(page))
5145 if (PageAnon(page)) {
5146 /* we don't move shared anon */
5149 } else if (!move_file())
5150 /* we ignore mapcount for file pages */
5152 if (!get_page_unless_zero(page))
5159 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5160 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5162 struct page *page = NULL;
5163 swp_entry_t ent = pte_to_swp_entry(ptent);
5165 if (!move_anon() || non_swap_entry(ent))
5168 * Because lookup_swap_cache() updates some statistics counter,
5169 * we call find_get_page() with swapper_space directly.
5171 page = find_get_page(&swapper_space, ent.val);
5172 if (do_swap_account)
5173 entry->val = ent.val;
5178 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5179 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5185 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5186 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5188 struct page *page = NULL;
5189 struct address_space *mapping;
5192 if (!vma->vm_file) /* anonymous vma */
5197 mapping = vma->vm_file->f_mapping;
5198 if (pte_none(ptent))
5199 pgoff = linear_page_index(vma, addr);
5200 else /* pte_file(ptent) is true */
5201 pgoff = pte_to_pgoff(ptent);
5203 /* page is moved even if it's not RSS of this task(page-faulted). */
5204 page = find_get_page(mapping, pgoff);
5207 /* shmem/tmpfs may report page out on swap: account for that too. */
5208 if (radix_tree_exceptional_entry(page)) {
5209 swp_entry_t swap = radix_to_swp_entry(page);
5210 if (do_swap_account)
5212 page = find_get_page(&swapper_space, swap.val);
5218 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5219 unsigned long addr, pte_t ptent, union mc_target *target)
5221 struct page *page = NULL;
5222 struct page_cgroup *pc;
5223 enum mc_target_type ret = MC_TARGET_NONE;
5224 swp_entry_t ent = { .val = 0 };
5226 if (pte_present(ptent))
5227 page = mc_handle_present_pte(vma, addr, ptent);
5228 else if (is_swap_pte(ptent))
5229 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5230 else if (pte_none(ptent) || pte_file(ptent))
5231 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5233 if (!page && !ent.val)
5236 pc = lookup_page_cgroup(page);
5238 * Do only loose check w/o page_cgroup lock.
5239 * mem_cgroup_move_account() checks the pc is valid or not under
5242 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5243 ret = MC_TARGET_PAGE;
5245 target->page = page;
5247 if (!ret || !target)
5250 /* There is a swap entry and a page doesn't exist or isn't charged */
5251 if (ent.val && !ret &&
5252 css_id(&mc.from->css) == lookup_swap_cgroup_id(ent)) {
5253 ret = MC_TARGET_SWAP;
5260 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5262 * We don't consider swapping or file mapped pages because THP does not
5263 * support them for now.
5264 * Caller should make sure that pmd_trans_huge(pmd) is true.
5266 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5267 unsigned long addr, pmd_t pmd, union mc_target *target)
5269 struct page *page = NULL;
5270 struct page_cgroup *pc;
5271 enum mc_target_type ret = MC_TARGET_NONE;
5273 page = pmd_page(pmd);
5274 VM_BUG_ON(!page || !PageHead(page));
5277 pc = lookup_page_cgroup(page);
5278 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5279 ret = MC_TARGET_PAGE;
5282 target->page = page;
5288 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5289 unsigned long addr, pmd_t pmd, union mc_target *target)
5291 return MC_TARGET_NONE;
5295 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5296 unsigned long addr, unsigned long end,
5297 struct mm_walk *walk)
5299 struct vm_area_struct *vma = walk->private;
5303 if (pmd_trans_huge_lock(pmd, vma) == 1) {
5304 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5305 mc.precharge += HPAGE_PMD_NR;
5306 spin_unlock(&vma->vm_mm->page_table_lock);
5310 if (pmd_trans_unstable(pmd))
5312 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5313 for (; addr != end; pte++, addr += PAGE_SIZE)
5314 if (get_mctgt_type(vma, addr, *pte, NULL))
5315 mc.precharge++; /* increment precharge temporarily */
5316 pte_unmap_unlock(pte - 1, ptl);
5322 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5324 unsigned long precharge;
5325 struct vm_area_struct *vma;
5327 down_read(&mm->mmap_sem);
5328 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5329 struct mm_walk mem_cgroup_count_precharge_walk = {
5330 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5334 if (is_vm_hugetlb_page(vma))
5336 walk_page_range(vma->vm_start, vma->vm_end,
5337 &mem_cgroup_count_precharge_walk);
5339 up_read(&mm->mmap_sem);
5341 precharge = mc.precharge;
5347 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5349 unsigned long precharge = mem_cgroup_count_precharge(mm);
5351 VM_BUG_ON(mc.moving_task);
5352 mc.moving_task = current;
5353 return mem_cgroup_do_precharge(precharge);
5356 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5357 static void __mem_cgroup_clear_mc(void)
5359 struct mem_cgroup *from = mc.from;
5360 struct mem_cgroup *to = mc.to;
5362 /* we must uncharge all the leftover precharges from mc.to */
5364 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
5368 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5369 * we must uncharge here.
5371 if (mc.moved_charge) {
5372 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
5373 mc.moved_charge = 0;
5375 /* we must fixup refcnts and charges */
5376 if (mc.moved_swap) {
5377 /* uncharge swap account from the old cgroup */
5378 if (!mem_cgroup_is_root(mc.from))
5379 res_counter_uncharge(&mc.from->memsw,
5380 PAGE_SIZE * mc.moved_swap);
5381 __mem_cgroup_put(mc.from, mc.moved_swap);
5383 if (!mem_cgroup_is_root(mc.to)) {
5385 * we charged both to->res and to->memsw, so we should
5388 res_counter_uncharge(&mc.to->res,
5389 PAGE_SIZE * mc.moved_swap);
5391 /* we've already done mem_cgroup_get(mc.to) */
5394 memcg_oom_recover(from);
5395 memcg_oom_recover(to);
5396 wake_up_all(&mc.waitq);
5399 static void mem_cgroup_clear_mc(void)
5401 struct mem_cgroup *from = mc.from;
5404 * we must clear moving_task before waking up waiters at the end of
5407 mc.moving_task = NULL;
5408 __mem_cgroup_clear_mc();
5409 spin_lock(&mc.lock);
5412 spin_unlock(&mc.lock);
5413 mem_cgroup_end_move(from);
5416 static int mem_cgroup_can_attach(struct cgroup *cgroup,
5417 struct cgroup_taskset *tset)
5419 struct task_struct *p = cgroup_taskset_first(tset);
5421 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup);
5423 if (memcg->move_charge_at_immigrate) {
5424 struct mm_struct *mm;
5425 struct mem_cgroup *from = mem_cgroup_from_task(p);
5427 VM_BUG_ON(from == memcg);
5429 mm = get_task_mm(p);
5432 /* We move charges only when we move a owner of the mm */
5433 if (mm->owner == p) {
5436 VM_BUG_ON(mc.precharge);
5437 VM_BUG_ON(mc.moved_charge);
5438 VM_BUG_ON(mc.moved_swap);
5439 mem_cgroup_start_move(from);
5440 spin_lock(&mc.lock);
5443 spin_unlock(&mc.lock);
5444 /* We set mc.moving_task later */
5446 ret = mem_cgroup_precharge_mc(mm);
5448 mem_cgroup_clear_mc();
5455 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
5456 struct cgroup_taskset *tset)
5458 mem_cgroup_clear_mc();
5461 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5462 unsigned long addr, unsigned long end,
5463 struct mm_walk *walk)
5466 struct vm_area_struct *vma = walk->private;
5469 enum mc_target_type target_type;
5470 union mc_target target;
5472 struct page_cgroup *pc;
5475 * We don't take compound_lock() here but no race with splitting thp
5477 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
5478 * under splitting, which means there's no concurrent thp split,
5479 * - if another thread runs into split_huge_page() just after we
5480 * entered this if-block, the thread must wait for page table lock
5481 * to be unlocked in __split_huge_page_splitting(), where the main
5482 * part of thp split is not executed yet.
5484 if (pmd_trans_huge_lock(pmd, vma) == 1) {
5485 if (mc.precharge < HPAGE_PMD_NR) {
5486 spin_unlock(&vma->vm_mm->page_table_lock);
5489 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5490 if (target_type == MC_TARGET_PAGE) {
5492 if (!isolate_lru_page(page)) {
5493 pc = lookup_page_cgroup(page);
5494 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
5495 pc, mc.from, mc.to)) {
5496 mc.precharge -= HPAGE_PMD_NR;
5497 mc.moved_charge += HPAGE_PMD_NR;
5499 putback_lru_page(page);
5503 spin_unlock(&vma->vm_mm->page_table_lock);
5507 if (pmd_trans_unstable(pmd))
5510 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5511 for (; addr != end; addr += PAGE_SIZE) {
5512 pte_t ptent = *(pte++);
5518 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5519 case MC_TARGET_PAGE:
5521 if (isolate_lru_page(page))
5523 pc = lookup_page_cgroup(page);
5524 if (!mem_cgroup_move_account(page, 1, pc,
5527 /* we uncharge from mc.from later. */
5530 putback_lru_page(page);
5531 put: /* get_mctgt_type() gets the page */
5534 case MC_TARGET_SWAP:
5536 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5538 /* we fixup refcnts and charges later. */
5546 pte_unmap_unlock(pte - 1, ptl);
5551 * We have consumed all precharges we got in can_attach().
5552 * We try charge one by one, but don't do any additional
5553 * charges to mc.to if we have failed in charge once in attach()
5556 ret = mem_cgroup_do_precharge(1);
5564 static void mem_cgroup_move_charge(struct mm_struct *mm)
5566 struct vm_area_struct *vma;
5568 lru_add_drain_all();
5570 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5572 * Someone who are holding the mmap_sem might be waiting in
5573 * waitq. So we cancel all extra charges, wake up all waiters,
5574 * and retry. Because we cancel precharges, we might not be able
5575 * to move enough charges, but moving charge is a best-effort
5576 * feature anyway, so it wouldn't be a big problem.
5578 __mem_cgroup_clear_mc();
5582 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5584 struct mm_walk mem_cgroup_move_charge_walk = {
5585 .pmd_entry = mem_cgroup_move_charge_pte_range,
5589 if (is_vm_hugetlb_page(vma))
5591 ret = walk_page_range(vma->vm_start, vma->vm_end,
5592 &mem_cgroup_move_charge_walk);
5595 * means we have consumed all precharges and failed in
5596 * doing additional charge. Just abandon here.
5600 up_read(&mm->mmap_sem);
5603 static void mem_cgroup_move_task(struct cgroup *cont,
5604 struct cgroup_taskset *tset)
5606 struct task_struct *p = cgroup_taskset_first(tset);
5607 struct mm_struct *mm = get_task_mm(p);
5611 mem_cgroup_move_charge(mm);
5615 mem_cgroup_clear_mc();
5617 #else /* !CONFIG_MMU */
5618 static int mem_cgroup_can_attach(struct cgroup *cgroup,
5619 struct cgroup_taskset *tset)
5623 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
5624 struct cgroup_taskset *tset)
5627 static void mem_cgroup_move_task(struct cgroup *cont,
5628 struct cgroup_taskset *tset)
5633 struct cgroup_subsys mem_cgroup_subsys = {
5635 .subsys_id = mem_cgroup_subsys_id,
5636 .css_alloc = mem_cgroup_css_alloc,
5637 .css_offline = mem_cgroup_css_offline,
5638 .css_free = mem_cgroup_css_free,
5639 .can_attach = mem_cgroup_can_attach,
5640 .cancel_attach = mem_cgroup_cancel_attach,
5641 .attach = mem_cgroup_move_task,
5642 .base_cftypes = mem_cgroup_files,
5647 #ifdef CONFIG_MEMCG_SWAP
5648 static int __init enable_swap_account(char *s)
5650 /* consider enabled if no parameter or 1 is given */
5651 if (!strcmp(s, "1"))
5652 really_do_swap_account = 1;
5653 else if (!strcmp(s, "0"))
5654 really_do_swap_account = 0;
5657 __setup("swapaccount=", enable_swap_account);