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>
54 #include <net/tcp_memcontrol.h>
56 #include <asm/uaccess.h>
58 #include <trace/events/vmscan.h>
60 struct cgroup_subsys mem_cgroup_subsys __read_mostly;
61 #define MEM_CGROUP_RECLAIM_RETRIES 5
62 static struct mem_cgroup *root_mem_cgroup __read_mostly;
64 #ifdef CONFIG_MEMCG_SWAP
65 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
66 int do_swap_account __read_mostly;
68 /* for remember boot option*/
69 #ifdef CONFIG_MEMCG_SWAP_ENABLED
70 static int really_do_swap_account __initdata = 1;
72 static int really_do_swap_account __initdata = 0;
76 #define do_swap_account 0
81 * Statistics for memory cgroup.
83 enum mem_cgroup_stat_index {
85 * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
87 MEM_CGROUP_STAT_CACHE, /* # of pages charged as cache */
88 MEM_CGROUP_STAT_RSS, /* # of pages charged as anon rss */
89 MEM_CGROUP_STAT_FILE_MAPPED, /* # of pages charged as file rss */
90 MEM_CGROUP_STAT_SWAP, /* # of pages, swapped out */
91 MEM_CGROUP_STAT_NSTATS,
94 static const char * const mem_cgroup_stat_names[] = {
101 enum mem_cgroup_events_index {
102 MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */
103 MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */
104 MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */
105 MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */
106 MEM_CGROUP_EVENTS_NSTATS,
109 static const char * const mem_cgroup_events_names[] = {
117 * Per memcg event counter is incremented at every pagein/pageout. With THP,
118 * it will be incremated by the number of pages. This counter is used for
119 * for trigger some periodic events. This is straightforward and better
120 * than using jiffies etc. to handle periodic memcg event.
122 enum mem_cgroup_events_target {
123 MEM_CGROUP_TARGET_THRESH,
124 MEM_CGROUP_TARGET_SOFTLIMIT,
125 MEM_CGROUP_TARGET_NUMAINFO,
128 #define THRESHOLDS_EVENTS_TARGET 128
129 #define SOFTLIMIT_EVENTS_TARGET 1024
130 #define NUMAINFO_EVENTS_TARGET 1024
132 struct mem_cgroup_stat_cpu {
133 long count[MEM_CGROUP_STAT_NSTATS];
134 unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
135 unsigned long nr_page_events;
136 unsigned long targets[MEM_CGROUP_NTARGETS];
139 struct mem_cgroup_reclaim_iter {
140 /* css_id of the last scanned hierarchy member */
142 /* scan generation, increased every round-trip */
143 unsigned int generation;
147 * per-zone information in memory controller.
149 struct mem_cgroup_per_zone {
150 struct lruvec lruvec;
151 unsigned long lru_size[NR_LRU_LISTS];
153 struct mem_cgroup_reclaim_iter reclaim_iter[DEF_PRIORITY + 1];
155 struct rb_node tree_node; /* RB tree node */
156 unsigned long long usage_in_excess;/* Set to the value by which */
157 /* the soft limit is exceeded*/
159 struct mem_cgroup *memcg; /* Back pointer, we cannot */
160 /* use container_of */
163 struct mem_cgroup_per_node {
164 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
167 struct mem_cgroup_lru_info {
168 struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
172 * Cgroups above their limits are maintained in a RB-Tree, independent of
173 * their hierarchy representation
176 struct mem_cgroup_tree_per_zone {
177 struct rb_root rb_root;
181 struct mem_cgroup_tree_per_node {
182 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
185 struct mem_cgroup_tree {
186 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
189 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
191 struct mem_cgroup_threshold {
192 struct eventfd_ctx *eventfd;
197 struct mem_cgroup_threshold_ary {
198 /* An array index points to threshold just below or equal to usage. */
199 int current_threshold;
200 /* Size of entries[] */
202 /* Array of thresholds */
203 struct mem_cgroup_threshold entries[0];
206 struct mem_cgroup_thresholds {
207 /* Primary thresholds array */
208 struct mem_cgroup_threshold_ary *primary;
210 * Spare threshold array.
211 * This is needed to make mem_cgroup_unregister_event() "never fail".
212 * It must be able to store at least primary->size - 1 entries.
214 struct mem_cgroup_threshold_ary *spare;
218 struct mem_cgroup_eventfd_list {
219 struct list_head list;
220 struct eventfd_ctx *eventfd;
223 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
224 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
227 * The memory controller data structure. The memory controller controls both
228 * page cache and RSS per cgroup. We would eventually like to provide
229 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
230 * to help the administrator determine what knobs to tune.
232 * TODO: Add a water mark for the memory controller. Reclaim will begin when
233 * we hit the water mark. May be even add a low water mark, such that
234 * no reclaim occurs from a cgroup at it's low water mark, this is
235 * a feature that will be implemented much later in the future.
238 struct cgroup_subsys_state css;
240 * the counter to account for memory usage
242 struct res_counter res;
246 * the counter to account for mem+swap usage.
248 struct res_counter memsw;
251 * rcu_freeing is used only when freeing struct mem_cgroup,
252 * so put it into a union to avoid wasting more memory.
253 * It must be disjoint from the css field. It could be
254 * in a union with the res field, but res plays a much
255 * larger part in mem_cgroup life than memsw, and might
256 * be of interest, even at time of free, when debugging.
257 * So share rcu_head with the less interesting memsw.
259 struct rcu_head rcu_freeing;
261 * We also need some space for a worker in deferred freeing.
262 * By the time we call it, rcu_freeing is no longer in use.
264 struct work_struct work_freeing;
268 * Per cgroup active and inactive list, similar to the
269 * per zone LRU lists.
271 struct mem_cgroup_lru_info info;
272 int last_scanned_node;
274 nodemask_t scan_nodes;
275 atomic_t numainfo_events;
276 atomic_t numainfo_updating;
279 * Should the accounting and control be hierarchical, per subtree?
289 /* OOM-Killer disable */
290 int oom_kill_disable;
292 /* set when res.limit == memsw.limit */
293 bool memsw_is_minimum;
295 /* protect arrays of thresholds */
296 struct mutex thresholds_lock;
298 /* thresholds for memory usage. RCU-protected */
299 struct mem_cgroup_thresholds thresholds;
301 /* thresholds for mem+swap usage. RCU-protected */
302 struct mem_cgroup_thresholds memsw_thresholds;
304 /* For oom notifier event fd */
305 struct list_head oom_notify;
308 * Should we move charges of a task when a task is moved into this
309 * mem_cgroup ? And what type of charges should we move ?
311 unsigned long move_charge_at_immigrate;
313 * set > 0 if pages under this cgroup are moving to other cgroup.
315 atomic_t moving_account;
316 /* taken only while moving_account > 0 */
317 spinlock_t move_lock;
321 struct mem_cgroup_stat_cpu __percpu *stat;
323 * used when a cpu is offlined or other synchronizations
324 * See mem_cgroup_read_stat().
326 struct mem_cgroup_stat_cpu nocpu_base;
327 spinlock_t pcp_counter_lock;
330 struct tcp_memcontrol tcp_mem;
334 /* Stuffs for move charges at task migration. */
336 * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
337 * left-shifted bitmap of these types.
340 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
341 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
345 /* "mc" and its members are protected by cgroup_mutex */
346 static struct move_charge_struct {
347 spinlock_t lock; /* for from, to */
348 struct mem_cgroup *from;
349 struct mem_cgroup *to;
350 unsigned long precharge;
351 unsigned long moved_charge;
352 unsigned long moved_swap;
353 struct task_struct *moving_task; /* a task moving charges */
354 wait_queue_head_t waitq; /* a waitq for other context */
356 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
357 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
360 static bool move_anon(void)
362 return test_bit(MOVE_CHARGE_TYPE_ANON,
363 &mc.to->move_charge_at_immigrate);
366 static bool move_file(void)
368 return test_bit(MOVE_CHARGE_TYPE_FILE,
369 &mc.to->move_charge_at_immigrate);
373 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
374 * limit reclaim to prevent infinite loops, if they ever occur.
376 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
377 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
380 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
381 MEM_CGROUP_CHARGE_TYPE_ANON,
382 MEM_CGROUP_CHARGE_TYPE_SHMEM, /* used by page migration of shmem */
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);
409 /* Writing them here to avoid exposing memcg's inner layout */
410 #ifdef CONFIG_MEMCG_KMEM
411 #include <net/sock.h>
414 static bool mem_cgroup_is_root(struct mem_cgroup *memcg);
415 void sock_update_memcg(struct sock *sk)
417 if (mem_cgroup_sockets_enabled) {
418 struct mem_cgroup *memcg;
419 struct cg_proto *cg_proto;
421 BUG_ON(!sk->sk_prot->proto_cgroup);
423 /* Socket cloning can throw us here with sk_cgrp already
424 * filled. It won't however, necessarily happen from
425 * process context. So the test for root memcg given
426 * the current task's memcg won't help us in this case.
428 * Respecting the original socket's memcg is a better
429 * decision in this case.
432 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
433 mem_cgroup_get(sk->sk_cgrp->memcg);
438 memcg = mem_cgroup_from_task(current);
439 cg_proto = sk->sk_prot->proto_cgroup(memcg);
440 if (!mem_cgroup_is_root(memcg) && memcg_proto_active(cg_proto)) {
441 mem_cgroup_get(memcg);
442 sk->sk_cgrp = cg_proto;
447 EXPORT_SYMBOL(sock_update_memcg);
449 void sock_release_memcg(struct sock *sk)
451 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
452 struct mem_cgroup *memcg;
453 WARN_ON(!sk->sk_cgrp->memcg);
454 memcg = sk->sk_cgrp->memcg;
455 mem_cgroup_put(memcg);
460 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
462 if (!memcg || mem_cgroup_is_root(memcg))
465 return &memcg->tcp_mem.cg_proto;
467 EXPORT_SYMBOL(tcp_proto_cgroup);
468 #endif /* CONFIG_INET */
469 #endif /* CONFIG_MEMCG_KMEM */
471 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
472 static void disarm_sock_keys(struct mem_cgroup *memcg)
474 if (!memcg_proto_activated(&memcg->tcp_mem.cg_proto))
476 static_key_slow_dec(&memcg_socket_limit_enabled);
479 static void disarm_sock_keys(struct mem_cgroup *memcg)
484 static void drain_all_stock_async(struct mem_cgroup *memcg);
486 static struct mem_cgroup_per_zone *
487 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
489 return &memcg->info.nodeinfo[nid]->zoneinfo[zid];
492 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
497 static struct mem_cgroup_per_zone *
498 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
500 int nid = page_to_nid(page);
501 int zid = page_zonenum(page);
503 return mem_cgroup_zoneinfo(memcg, nid, zid);
506 static struct mem_cgroup_tree_per_zone *
507 soft_limit_tree_node_zone(int nid, int zid)
509 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
512 static struct mem_cgroup_tree_per_zone *
513 soft_limit_tree_from_page(struct page *page)
515 int nid = page_to_nid(page);
516 int zid = page_zonenum(page);
518 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
522 __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
523 struct mem_cgroup_per_zone *mz,
524 struct mem_cgroup_tree_per_zone *mctz,
525 unsigned long long new_usage_in_excess)
527 struct rb_node **p = &mctz->rb_root.rb_node;
528 struct rb_node *parent = NULL;
529 struct mem_cgroup_per_zone *mz_node;
534 mz->usage_in_excess = new_usage_in_excess;
535 if (!mz->usage_in_excess)
539 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
541 if (mz->usage_in_excess < mz_node->usage_in_excess)
544 * We can't avoid mem cgroups that are over their soft
545 * limit by the same amount
547 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
550 rb_link_node(&mz->tree_node, parent, p);
551 rb_insert_color(&mz->tree_node, &mctz->rb_root);
556 __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
557 struct mem_cgroup_per_zone *mz,
558 struct mem_cgroup_tree_per_zone *mctz)
562 rb_erase(&mz->tree_node, &mctz->rb_root);
567 mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
568 struct mem_cgroup_per_zone *mz,
569 struct mem_cgroup_tree_per_zone *mctz)
571 spin_lock(&mctz->lock);
572 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
573 spin_unlock(&mctz->lock);
577 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
579 unsigned long long excess;
580 struct mem_cgroup_per_zone *mz;
581 struct mem_cgroup_tree_per_zone *mctz;
582 int nid = page_to_nid(page);
583 int zid = page_zonenum(page);
584 mctz = soft_limit_tree_from_page(page);
587 * Necessary to update all ancestors when hierarchy is used.
588 * because their event counter is not touched.
590 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
591 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
592 excess = res_counter_soft_limit_excess(&memcg->res);
594 * We have to update the tree if mz is on RB-tree or
595 * mem is over its softlimit.
597 if (excess || mz->on_tree) {
598 spin_lock(&mctz->lock);
599 /* if on-tree, remove it */
601 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
603 * Insert again. mz->usage_in_excess will be updated.
604 * If excess is 0, no tree ops.
606 __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
607 spin_unlock(&mctz->lock);
612 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
615 struct mem_cgroup_per_zone *mz;
616 struct mem_cgroup_tree_per_zone *mctz;
618 for_each_node(node) {
619 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
620 mz = mem_cgroup_zoneinfo(memcg, node, zone);
621 mctz = soft_limit_tree_node_zone(node, zone);
622 mem_cgroup_remove_exceeded(memcg, mz, mctz);
627 static struct mem_cgroup_per_zone *
628 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
630 struct rb_node *rightmost = NULL;
631 struct mem_cgroup_per_zone *mz;
635 rightmost = rb_last(&mctz->rb_root);
637 goto done; /* Nothing to reclaim from */
639 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
641 * Remove the node now but someone else can add it back,
642 * we will to add it back at the end of reclaim to its correct
643 * position in the tree.
645 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
646 if (!res_counter_soft_limit_excess(&mz->memcg->res) ||
647 !css_tryget(&mz->memcg->css))
653 static struct mem_cgroup_per_zone *
654 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
656 struct mem_cgroup_per_zone *mz;
658 spin_lock(&mctz->lock);
659 mz = __mem_cgroup_largest_soft_limit_node(mctz);
660 spin_unlock(&mctz->lock);
665 * Implementation Note: reading percpu statistics for memcg.
667 * Both of vmstat[] and percpu_counter has threshold and do periodic
668 * synchronization to implement "quick" read. There are trade-off between
669 * reading cost and precision of value. Then, we may have a chance to implement
670 * a periodic synchronizion of counter in memcg's counter.
672 * But this _read() function is used for user interface now. The user accounts
673 * memory usage by memory cgroup and he _always_ requires exact value because
674 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
675 * have to visit all online cpus and make sum. So, for now, unnecessary
676 * synchronization is not implemented. (just implemented for cpu hotplug)
678 * If there are kernel internal actions which can make use of some not-exact
679 * value, and reading all cpu value can be performance bottleneck in some
680 * common workload, threashold and synchonization as vmstat[] should be
683 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
684 enum mem_cgroup_stat_index idx)
690 for_each_online_cpu(cpu)
691 val += per_cpu(memcg->stat->count[idx], cpu);
692 #ifdef CONFIG_HOTPLUG_CPU
693 spin_lock(&memcg->pcp_counter_lock);
694 val += memcg->nocpu_base.count[idx];
695 spin_unlock(&memcg->pcp_counter_lock);
701 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
704 int val = (charge) ? 1 : -1;
705 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
708 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
709 enum mem_cgroup_events_index idx)
711 unsigned long val = 0;
714 for_each_online_cpu(cpu)
715 val += per_cpu(memcg->stat->events[idx], cpu);
716 #ifdef CONFIG_HOTPLUG_CPU
717 spin_lock(&memcg->pcp_counter_lock);
718 val += memcg->nocpu_base.events[idx];
719 spin_unlock(&memcg->pcp_counter_lock);
724 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
725 bool anon, int nr_pages)
730 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
731 * counted as CACHE even if it's on ANON LRU.
734 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
737 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
740 /* pagein of a big page is an event. So, ignore page size */
742 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
744 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
745 nr_pages = -nr_pages; /* for event */
748 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
754 mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
756 struct mem_cgroup_per_zone *mz;
758 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
759 return mz->lru_size[lru];
763 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
764 unsigned int lru_mask)
766 struct mem_cgroup_per_zone *mz;
768 unsigned long ret = 0;
770 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
773 if (BIT(lru) & lru_mask)
774 ret += mz->lru_size[lru];
780 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
781 int nid, unsigned int lru_mask)
786 for (zid = 0; zid < MAX_NR_ZONES; zid++)
787 total += mem_cgroup_zone_nr_lru_pages(memcg,
793 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
794 unsigned int lru_mask)
799 for_each_node_state(nid, N_HIGH_MEMORY)
800 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
804 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
805 enum mem_cgroup_events_target target)
807 unsigned long val, next;
809 val = __this_cpu_read(memcg->stat->nr_page_events);
810 next = __this_cpu_read(memcg->stat->targets[target]);
811 /* from time_after() in jiffies.h */
812 if ((long)next - (long)val < 0) {
814 case MEM_CGROUP_TARGET_THRESH:
815 next = val + THRESHOLDS_EVENTS_TARGET;
817 case MEM_CGROUP_TARGET_SOFTLIMIT:
818 next = val + SOFTLIMIT_EVENTS_TARGET;
820 case MEM_CGROUP_TARGET_NUMAINFO:
821 next = val + NUMAINFO_EVENTS_TARGET;
826 __this_cpu_write(memcg->stat->targets[target], next);
833 * Check events in order.
836 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
839 /* threshold event is triggered in finer grain than soft limit */
840 if (unlikely(mem_cgroup_event_ratelimit(memcg,
841 MEM_CGROUP_TARGET_THRESH))) {
843 bool do_numainfo __maybe_unused;
845 do_softlimit = mem_cgroup_event_ratelimit(memcg,
846 MEM_CGROUP_TARGET_SOFTLIMIT);
848 do_numainfo = mem_cgroup_event_ratelimit(memcg,
849 MEM_CGROUP_TARGET_NUMAINFO);
853 mem_cgroup_threshold(memcg);
854 if (unlikely(do_softlimit))
855 mem_cgroup_update_tree(memcg, page);
857 if (unlikely(do_numainfo))
858 atomic_inc(&memcg->numainfo_events);
864 struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
866 return container_of(cgroup_subsys_state(cont,
867 mem_cgroup_subsys_id), struct mem_cgroup,
871 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
874 * mm_update_next_owner() may clear mm->owner to NULL
875 * if it races with swapoff, page migration, etc.
876 * So this can be called with p == NULL.
881 return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
882 struct mem_cgroup, css);
885 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
887 struct mem_cgroup *memcg = NULL;
892 * Because we have no locks, mm->owner's may be being moved to other
893 * cgroup. We use css_tryget() here even if this looks
894 * pessimistic (rather than adding locks here).
898 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
899 if (unlikely(!memcg))
901 } while (!css_tryget(&memcg->css));
907 * mem_cgroup_iter - iterate over memory cgroup hierarchy
908 * @root: hierarchy root
909 * @prev: previously returned memcg, NULL on first invocation
910 * @reclaim: cookie for shared reclaim walks, NULL for full walks
912 * Returns references to children of the hierarchy below @root, or
913 * @root itself, or %NULL after a full round-trip.
915 * Caller must pass the return value in @prev on subsequent
916 * invocations for reference counting, or use mem_cgroup_iter_break()
917 * to cancel a hierarchy walk before the round-trip is complete.
919 * Reclaimers can specify a zone and a priority level in @reclaim to
920 * divide up the memcgs in the hierarchy among all concurrent
921 * reclaimers operating on the same zone and priority.
923 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
924 struct mem_cgroup *prev,
925 struct mem_cgroup_reclaim_cookie *reclaim)
927 struct mem_cgroup *memcg = NULL;
930 if (mem_cgroup_disabled())
934 root = root_mem_cgroup;
936 if (prev && !reclaim)
937 id = css_id(&prev->css);
939 if (prev && prev != root)
942 if (!root->use_hierarchy && root != root_mem_cgroup) {
949 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
950 struct cgroup_subsys_state *css;
953 int nid = zone_to_nid(reclaim->zone);
954 int zid = zone_idx(reclaim->zone);
955 struct mem_cgroup_per_zone *mz;
957 mz = mem_cgroup_zoneinfo(root, nid, zid);
958 iter = &mz->reclaim_iter[reclaim->priority];
959 if (prev && reclaim->generation != iter->generation)
965 css = css_get_next(&mem_cgroup_subsys, id + 1, &root->css, &id);
967 if (css == &root->css || css_tryget(css))
968 memcg = container_of(css,
969 struct mem_cgroup, css);
978 else if (!prev && memcg)
979 reclaim->generation = iter->generation;
989 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
990 * @root: hierarchy root
991 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
993 void mem_cgroup_iter_break(struct mem_cgroup *root,
994 struct mem_cgroup *prev)
997 root = root_mem_cgroup;
998 if (prev && prev != root)
1003 * Iteration constructs for visiting all cgroups (under a tree). If
1004 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1005 * be used for reference counting.
1007 #define for_each_mem_cgroup_tree(iter, root) \
1008 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1010 iter = mem_cgroup_iter(root, iter, NULL))
1012 #define for_each_mem_cgroup(iter) \
1013 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1015 iter = mem_cgroup_iter(NULL, iter, NULL))
1017 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
1019 return (memcg == root_mem_cgroup);
1022 void mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1024 struct mem_cgroup *memcg;
1030 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1031 if (unlikely(!memcg))
1036 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1039 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1047 EXPORT_SYMBOL(mem_cgroup_count_vm_event);
1050 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1051 * @zone: zone of the wanted lruvec
1052 * @memcg: memcg of the wanted lruvec
1054 * Returns the lru list vector holding pages for the given @zone and
1055 * @mem. This can be the global zone lruvec, if the memory controller
1058 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1059 struct mem_cgroup *memcg)
1061 struct mem_cgroup_per_zone *mz;
1063 if (mem_cgroup_disabled())
1064 return &zone->lruvec;
1066 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1071 * Following LRU functions are allowed to be used without PCG_LOCK.
1072 * Operations are called by routine of global LRU independently from memcg.
1073 * What we have to take care of here is validness of pc->mem_cgroup.
1075 * Changes to pc->mem_cgroup happens when
1078 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1079 * It is added to LRU before charge.
1080 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1081 * When moving account, the page is not on LRU. It's isolated.
1085 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1087 * @zone: zone of the page
1089 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1091 struct mem_cgroup_per_zone *mz;
1092 struct mem_cgroup *memcg;
1093 struct page_cgroup *pc;
1095 if (mem_cgroup_disabled())
1096 return &zone->lruvec;
1098 pc = lookup_page_cgroup(page);
1099 memcg = pc->mem_cgroup;
1102 * Surreptitiously switch any uncharged offlist page to root:
1103 * an uncharged page off lru does nothing to secure
1104 * its former mem_cgroup from sudden removal.
1106 * Our caller holds lru_lock, and PageCgroupUsed is updated
1107 * under page_cgroup lock: between them, they make all uses
1108 * of pc->mem_cgroup safe.
1110 if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1111 pc->mem_cgroup = memcg = root_mem_cgroup;
1113 mz = page_cgroup_zoneinfo(memcg, page);
1118 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1119 * @lruvec: mem_cgroup per zone lru vector
1120 * @lru: index of lru list the page is sitting on
1121 * @nr_pages: positive when adding or negative when removing
1123 * This function must be called when a page is added to or removed from an
1126 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1129 struct mem_cgroup_per_zone *mz;
1130 unsigned long *lru_size;
1132 if (mem_cgroup_disabled())
1135 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1136 lru_size = mz->lru_size + lru;
1137 *lru_size += nr_pages;
1138 VM_BUG_ON((long)(*lru_size) < 0);
1142 * Checks whether given mem is same or in the root_mem_cgroup's
1145 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1146 struct mem_cgroup *memcg)
1148 if (root_memcg == memcg)
1150 if (!root_memcg->use_hierarchy || !memcg)
1152 return css_is_ancestor(&memcg->css, &root_memcg->css);
1155 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1156 struct mem_cgroup *memcg)
1161 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1166 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *memcg)
1169 struct mem_cgroup *curr = NULL;
1170 struct task_struct *p;
1172 p = find_lock_task_mm(task);
1174 curr = try_get_mem_cgroup_from_mm(p->mm);
1178 * All threads may have already detached their mm's, but the oom
1179 * killer still needs to detect if they have already been oom
1180 * killed to prevent needlessly killing additional tasks.
1183 curr = mem_cgroup_from_task(task);
1185 css_get(&curr->css);
1191 * We should check use_hierarchy of "memcg" not "curr". Because checking
1192 * use_hierarchy of "curr" here make this function true if hierarchy is
1193 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1194 * hierarchy(even if use_hierarchy is disabled in "memcg").
1196 ret = mem_cgroup_same_or_subtree(memcg, curr);
1197 css_put(&curr->css);
1201 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1203 unsigned long inactive_ratio;
1204 unsigned long inactive;
1205 unsigned long active;
1208 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1209 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1211 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1213 inactive_ratio = int_sqrt(10 * gb);
1217 return inactive * inactive_ratio < active;
1220 int mem_cgroup_inactive_file_is_low(struct lruvec *lruvec)
1222 unsigned long active;
1223 unsigned long inactive;
1225 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_FILE);
1226 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_FILE);
1228 return (active > inactive);
1231 #define mem_cgroup_from_res_counter(counter, member) \
1232 container_of(counter, struct mem_cgroup, member)
1235 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1236 * @memcg: the memory cgroup
1238 * Returns the maximum amount of memory @mem can be charged with, in
1241 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1243 unsigned long long margin;
1245 margin = res_counter_margin(&memcg->res);
1246 if (do_swap_account)
1247 margin = min(margin, res_counter_margin(&memcg->memsw));
1248 return margin >> PAGE_SHIFT;
1251 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1253 struct cgroup *cgrp = memcg->css.cgroup;
1256 if (cgrp->parent == NULL)
1257 return vm_swappiness;
1259 return memcg->swappiness;
1263 * memcg->moving_account is used for checking possibility that some thread is
1264 * calling move_account(). When a thread on CPU-A starts moving pages under
1265 * a memcg, other threads should check memcg->moving_account under
1266 * rcu_read_lock(), like this:
1270 * memcg->moving_account+1 if (memcg->mocing_account)
1272 * synchronize_rcu() update something.
1277 /* for quick checking without looking up memcg */
1278 atomic_t memcg_moving __read_mostly;
1280 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1282 atomic_inc(&memcg_moving);
1283 atomic_inc(&memcg->moving_account);
1287 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1290 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1291 * We check NULL in callee rather than caller.
1294 atomic_dec(&memcg_moving);
1295 atomic_dec(&memcg->moving_account);
1300 * 2 routines for checking "mem" is under move_account() or not.
1302 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1303 * is used for avoiding races in accounting. If true,
1304 * pc->mem_cgroup may be overwritten.
1306 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1307 * under hierarchy of moving cgroups. This is for
1308 * waiting at hith-memory prressure caused by "move".
1311 static bool mem_cgroup_stolen(struct mem_cgroup *memcg)
1313 VM_BUG_ON(!rcu_read_lock_held());
1314 return atomic_read(&memcg->moving_account) > 0;
1317 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1319 struct mem_cgroup *from;
1320 struct mem_cgroup *to;
1323 * Unlike task_move routines, we access mc.to, mc.from not under
1324 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1326 spin_lock(&mc.lock);
1332 ret = mem_cgroup_same_or_subtree(memcg, from)
1333 || mem_cgroup_same_or_subtree(memcg, to);
1335 spin_unlock(&mc.lock);
1339 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1341 if (mc.moving_task && current != mc.moving_task) {
1342 if (mem_cgroup_under_move(memcg)) {
1344 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1345 /* moving charge context might have finished. */
1348 finish_wait(&mc.waitq, &wait);
1356 * Take this lock when
1357 * - a code tries to modify page's memcg while it's USED.
1358 * - a code tries to modify page state accounting in a memcg.
1359 * see mem_cgroup_stolen(), too.
1361 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1362 unsigned long *flags)
1364 spin_lock_irqsave(&memcg->move_lock, *flags);
1367 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1368 unsigned long *flags)
1370 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1374 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1375 * @memcg: The memory cgroup that went over limit
1376 * @p: Task that is going to be killed
1378 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1381 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1383 struct cgroup *task_cgrp;
1384 struct cgroup *mem_cgrp;
1386 * Need a buffer in BSS, can't rely on allocations. The code relies
1387 * on the assumption that OOM is serialized for memory controller.
1388 * If this assumption is broken, revisit this code.
1390 static char memcg_name[PATH_MAX];
1398 mem_cgrp = memcg->css.cgroup;
1399 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1401 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1404 * Unfortunately, we are unable to convert to a useful name
1405 * But we'll still print out the usage information
1412 printk(KERN_INFO "Task in %s killed", memcg_name);
1415 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1423 * Continues from above, so we don't need an KERN_ level
1425 printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1428 printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1429 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1430 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1431 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1432 printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1434 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1435 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1436 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1440 * This function returns the number of memcg under hierarchy tree. Returns
1441 * 1(self count) if no children.
1443 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1446 struct mem_cgroup *iter;
1448 for_each_mem_cgroup_tree(iter, memcg)
1454 * Return the memory (and swap, if configured) limit for a memcg.
1456 static u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1461 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1462 limit += total_swap_pages << PAGE_SHIFT;
1464 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1466 * If memsw is finite and limits the amount of swap space available
1467 * to this memcg, return that limit.
1469 return min(limit, memsw);
1472 void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1475 struct mem_cgroup *iter;
1476 unsigned long chosen_points = 0;
1477 unsigned long totalpages;
1478 unsigned int points = 0;
1479 struct task_struct *chosen = NULL;
1482 * If current has a pending SIGKILL, then automatically select it. The
1483 * goal is to allow it to allocate so that it may quickly exit and free
1486 if (fatal_signal_pending(current)) {
1487 set_thread_flag(TIF_MEMDIE);
1491 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1492 totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
1493 for_each_mem_cgroup_tree(iter, memcg) {
1494 struct cgroup *cgroup = iter->css.cgroup;
1495 struct cgroup_iter it;
1496 struct task_struct *task;
1498 cgroup_iter_start(cgroup, &it);
1499 while ((task = cgroup_iter_next(cgroup, &it))) {
1500 switch (oom_scan_process_thread(task, totalpages, NULL,
1502 case OOM_SCAN_SELECT:
1504 put_task_struct(chosen);
1506 chosen_points = ULONG_MAX;
1507 get_task_struct(chosen);
1509 case OOM_SCAN_CONTINUE:
1511 case OOM_SCAN_ABORT:
1512 cgroup_iter_end(cgroup, &it);
1513 mem_cgroup_iter_break(memcg, iter);
1515 put_task_struct(chosen);
1520 points = oom_badness(task, memcg, NULL, totalpages);
1521 if (points > chosen_points) {
1523 put_task_struct(chosen);
1525 chosen_points = points;
1526 get_task_struct(chosen);
1529 cgroup_iter_end(cgroup, &it);
1534 points = chosen_points * 1000 / totalpages;
1535 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1536 NULL, "Memory cgroup out of memory");
1539 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1541 unsigned long flags)
1543 unsigned long total = 0;
1544 bool noswap = false;
1547 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1549 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1552 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1554 drain_all_stock_async(memcg);
1555 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1557 * Allow limit shrinkers, which are triggered directly
1558 * by userspace, to catch signals and stop reclaim
1559 * after minimal progress, regardless of the margin.
1561 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1563 if (mem_cgroup_margin(memcg))
1566 * If nothing was reclaimed after two attempts, there
1567 * may be no reclaimable pages in this hierarchy.
1576 * test_mem_cgroup_node_reclaimable
1577 * @memcg: the target memcg
1578 * @nid: the node ID to be checked.
1579 * @noswap : specify true here if the user wants flle only information.
1581 * This function returns whether the specified memcg contains any
1582 * reclaimable pages on a node. Returns true if there are any reclaimable
1583 * pages in the node.
1585 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1586 int nid, bool noswap)
1588 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1590 if (noswap || !total_swap_pages)
1592 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1597 #if MAX_NUMNODES > 1
1600 * Always updating the nodemask is not very good - even if we have an empty
1601 * list or the wrong list here, we can start from some node and traverse all
1602 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1605 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1609 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1610 * pagein/pageout changes since the last update.
1612 if (!atomic_read(&memcg->numainfo_events))
1614 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1617 /* make a nodemask where this memcg uses memory from */
1618 memcg->scan_nodes = node_states[N_HIGH_MEMORY];
1620 for_each_node_mask(nid, node_states[N_HIGH_MEMORY]) {
1622 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1623 node_clear(nid, memcg->scan_nodes);
1626 atomic_set(&memcg->numainfo_events, 0);
1627 atomic_set(&memcg->numainfo_updating, 0);
1631 * Selecting a node where we start reclaim from. Because what we need is just
1632 * reducing usage counter, start from anywhere is O,K. Considering
1633 * memory reclaim from current node, there are pros. and cons.
1635 * Freeing memory from current node means freeing memory from a node which
1636 * we'll use or we've used. So, it may make LRU bad. And if several threads
1637 * hit limits, it will see a contention on a node. But freeing from remote
1638 * node means more costs for memory reclaim because of memory latency.
1640 * Now, we use round-robin. Better algorithm is welcomed.
1642 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1646 mem_cgroup_may_update_nodemask(memcg);
1647 node = memcg->last_scanned_node;
1649 node = next_node(node, memcg->scan_nodes);
1650 if (node == MAX_NUMNODES)
1651 node = first_node(memcg->scan_nodes);
1653 * We call this when we hit limit, not when pages are added to LRU.
1654 * No LRU may hold pages because all pages are UNEVICTABLE or
1655 * memcg is too small and all pages are not on LRU. In that case,
1656 * we use curret node.
1658 if (unlikely(node == MAX_NUMNODES))
1659 node = numa_node_id();
1661 memcg->last_scanned_node = node;
1666 * Check all nodes whether it contains reclaimable pages or not.
1667 * For quick scan, we make use of scan_nodes. This will allow us to skip
1668 * unused nodes. But scan_nodes is lazily updated and may not cotain
1669 * enough new information. We need to do double check.
1671 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1676 * quick check...making use of scan_node.
1677 * We can skip unused nodes.
1679 if (!nodes_empty(memcg->scan_nodes)) {
1680 for (nid = first_node(memcg->scan_nodes);
1682 nid = next_node(nid, memcg->scan_nodes)) {
1684 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1689 * Check rest of nodes.
1691 for_each_node_state(nid, N_HIGH_MEMORY) {
1692 if (node_isset(nid, memcg->scan_nodes))
1694 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1701 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1706 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1708 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
1712 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1715 unsigned long *total_scanned)
1717 struct mem_cgroup *victim = NULL;
1720 unsigned long excess;
1721 unsigned long nr_scanned;
1722 struct mem_cgroup_reclaim_cookie reclaim = {
1727 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
1730 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1735 * If we have not been able to reclaim
1736 * anything, it might because there are
1737 * no reclaimable pages under this hierarchy
1742 * We want to do more targeted reclaim.
1743 * excess >> 2 is not to excessive so as to
1744 * reclaim too much, nor too less that we keep
1745 * coming back to reclaim from this cgroup
1747 if (total >= (excess >> 2) ||
1748 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1753 if (!mem_cgroup_reclaimable(victim, false))
1755 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1757 *total_scanned += nr_scanned;
1758 if (!res_counter_soft_limit_excess(&root_memcg->res))
1761 mem_cgroup_iter_break(root_memcg, victim);
1766 * Check OOM-Killer is already running under our hierarchy.
1767 * If someone is running, return false.
1768 * Has to be called with memcg_oom_lock
1770 static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg)
1772 struct mem_cgroup *iter, *failed = NULL;
1774 for_each_mem_cgroup_tree(iter, memcg) {
1775 if (iter->oom_lock) {
1777 * this subtree of our hierarchy is already locked
1778 * so we cannot give a lock.
1781 mem_cgroup_iter_break(memcg, iter);
1784 iter->oom_lock = true;
1791 * OK, we failed to lock the whole subtree so we have to clean up
1792 * what we set up to the failing subtree
1794 for_each_mem_cgroup_tree(iter, memcg) {
1795 if (iter == failed) {
1796 mem_cgroup_iter_break(memcg, iter);
1799 iter->oom_lock = false;
1805 * Has to be called with memcg_oom_lock
1807 static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1809 struct mem_cgroup *iter;
1811 for_each_mem_cgroup_tree(iter, memcg)
1812 iter->oom_lock = false;
1816 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1818 struct mem_cgroup *iter;
1820 for_each_mem_cgroup_tree(iter, memcg)
1821 atomic_inc(&iter->under_oom);
1824 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1826 struct mem_cgroup *iter;
1829 * When a new child is created while the hierarchy is under oom,
1830 * mem_cgroup_oom_lock() may not be called. We have to use
1831 * atomic_add_unless() here.
1833 for_each_mem_cgroup_tree(iter, memcg)
1834 atomic_add_unless(&iter->under_oom, -1, 0);
1837 static DEFINE_SPINLOCK(memcg_oom_lock);
1838 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1840 struct oom_wait_info {
1841 struct mem_cgroup *memcg;
1845 static int memcg_oom_wake_function(wait_queue_t *wait,
1846 unsigned mode, int sync, void *arg)
1848 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1849 struct mem_cgroup *oom_wait_memcg;
1850 struct oom_wait_info *oom_wait_info;
1852 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1853 oom_wait_memcg = oom_wait_info->memcg;
1856 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
1857 * Then we can use css_is_ancestor without taking care of RCU.
1859 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
1860 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
1862 return autoremove_wake_function(wait, mode, sync, arg);
1865 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
1867 /* for filtering, pass "memcg" as argument. */
1868 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1871 static void memcg_oom_recover(struct mem_cgroup *memcg)
1873 if (memcg && atomic_read(&memcg->under_oom))
1874 memcg_wakeup_oom(memcg);
1878 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1880 static bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask,
1883 struct oom_wait_info owait;
1884 bool locked, need_to_kill;
1886 owait.memcg = memcg;
1887 owait.wait.flags = 0;
1888 owait.wait.func = memcg_oom_wake_function;
1889 owait.wait.private = current;
1890 INIT_LIST_HEAD(&owait.wait.task_list);
1891 need_to_kill = true;
1892 mem_cgroup_mark_under_oom(memcg);
1894 /* At first, try to OOM lock hierarchy under memcg.*/
1895 spin_lock(&memcg_oom_lock);
1896 locked = mem_cgroup_oom_lock(memcg);
1898 * Even if signal_pending(), we can't quit charge() loop without
1899 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1900 * under OOM is always welcomed, use TASK_KILLABLE here.
1902 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1903 if (!locked || memcg->oom_kill_disable)
1904 need_to_kill = false;
1906 mem_cgroup_oom_notify(memcg);
1907 spin_unlock(&memcg_oom_lock);
1910 finish_wait(&memcg_oom_waitq, &owait.wait);
1911 mem_cgroup_out_of_memory(memcg, mask, order);
1914 finish_wait(&memcg_oom_waitq, &owait.wait);
1916 spin_lock(&memcg_oom_lock);
1918 mem_cgroup_oom_unlock(memcg);
1919 memcg_wakeup_oom(memcg);
1920 spin_unlock(&memcg_oom_lock);
1922 mem_cgroup_unmark_under_oom(memcg);
1924 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
1926 /* Give chance to dying process */
1927 schedule_timeout_uninterruptible(1);
1932 * Currently used to update mapped file statistics, but the routine can be
1933 * generalized to update other statistics as well.
1935 * Notes: Race condition
1937 * We usually use page_cgroup_lock() for accessing page_cgroup member but
1938 * it tends to be costly. But considering some conditions, we doesn't need
1939 * to do so _always_.
1941 * Considering "charge", lock_page_cgroup() is not required because all
1942 * file-stat operations happen after a page is attached to radix-tree. There
1943 * are no race with "charge".
1945 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
1946 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
1947 * if there are race with "uncharge". Statistics itself is properly handled
1950 * Considering "move", this is an only case we see a race. To make the race
1951 * small, we check mm->moving_account and detect there are possibility of race
1952 * If there is, we take a lock.
1955 void __mem_cgroup_begin_update_page_stat(struct page *page,
1956 bool *locked, unsigned long *flags)
1958 struct mem_cgroup *memcg;
1959 struct page_cgroup *pc;
1961 pc = lookup_page_cgroup(page);
1963 memcg = pc->mem_cgroup;
1964 if (unlikely(!memcg || !PageCgroupUsed(pc)))
1967 * If this memory cgroup is not under account moving, we don't
1968 * need to take move_lock_mem_cgroup(). Because we already hold
1969 * rcu_read_lock(), any calls to move_account will be delayed until
1970 * rcu_read_unlock() if mem_cgroup_stolen() == true.
1972 if (!mem_cgroup_stolen(memcg))
1975 move_lock_mem_cgroup(memcg, flags);
1976 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
1977 move_unlock_mem_cgroup(memcg, flags);
1983 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
1985 struct page_cgroup *pc = lookup_page_cgroup(page);
1988 * It's guaranteed that pc->mem_cgroup never changes while
1989 * lock is held because a routine modifies pc->mem_cgroup
1990 * should take move_lock_mem_cgroup().
1992 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
1995 void mem_cgroup_update_page_stat(struct page *page,
1996 enum mem_cgroup_page_stat_item idx, int val)
1998 struct mem_cgroup *memcg;
1999 struct page_cgroup *pc = lookup_page_cgroup(page);
2000 unsigned long uninitialized_var(flags);
2002 if (mem_cgroup_disabled())
2005 memcg = pc->mem_cgroup;
2006 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2010 case MEMCG_NR_FILE_MAPPED:
2011 idx = MEM_CGROUP_STAT_FILE_MAPPED;
2017 this_cpu_add(memcg->stat->count[idx], val);
2021 * size of first charge trial. "32" comes from vmscan.c's magic value.
2022 * TODO: maybe necessary to use big numbers in big irons.
2024 #define CHARGE_BATCH 32U
2025 struct memcg_stock_pcp {
2026 struct mem_cgroup *cached; /* this never be root cgroup */
2027 unsigned int nr_pages;
2028 struct work_struct work;
2029 unsigned long flags;
2030 #define FLUSHING_CACHED_CHARGE 0
2032 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2033 static DEFINE_MUTEX(percpu_charge_mutex);
2036 * Try to consume stocked charge on this cpu. If success, one page is consumed
2037 * from local stock and true is returned. If the stock is 0 or charges from a
2038 * cgroup which is not current target, returns false. This stock will be
2041 static bool consume_stock(struct mem_cgroup *memcg)
2043 struct memcg_stock_pcp *stock;
2046 stock = &get_cpu_var(memcg_stock);
2047 if (memcg == stock->cached && stock->nr_pages)
2049 else /* need to call res_counter_charge */
2051 put_cpu_var(memcg_stock);
2056 * Returns stocks cached in percpu to res_counter and reset cached information.
2058 static void drain_stock(struct memcg_stock_pcp *stock)
2060 struct mem_cgroup *old = stock->cached;
2062 if (stock->nr_pages) {
2063 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2065 res_counter_uncharge(&old->res, bytes);
2066 if (do_swap_account)
2067 res_counter_uncharge(&old->memsw, bytes);
2068 stock->nr_pages = 0;
2070 stock->cached = NULL;
2074 * This must be called under preempt disabled or must be called by
2075 * a thread which is pinned to local cpu.
2077 static void drain_local_stock(struct work_struct *dummy)
2079 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2081 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2085 * Cache charges(val) which is from res_counter, to local per_cpu area.
2086 * This will be consumed by consume_stock() function, later.
2088 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2090 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2092 if (stock->cached != memcg) { /* reset if necessary */
2094 stock->cached = memcg;
2096 stock->nr_pages += nr_pages;
2097 put_cpu_var(memcg_stock);
2101 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2102 * of the hierarchy under it. sync flag says whether we should block
2103 * until the work is done.
2105 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2109 /* Notify other cpus that system-wide "drain" is running */
2112 for_each_online_cpu(cpu) {
2113 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2114 struct mem_cgroup *memcg;
2116 memcg = stock->cached;
2117 if (!memcg || !stock->nr_pages)
2119 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2121 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2123 drain_local_stock(&stock->work);
2125 schedule_work_on(cpu, &stock->work);
2133 for_each_online_cpu(cpu) {
2134 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2135 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2136 flush_work(&stock->work);
2143 * Tries to drain stocked charges in other cpus. This function is asynchronous
2144 * and just put a work per cpu for draining localy on each cpu. Caller can
2145 * expects some charges will be back to res_counter later but cannot wait for
2148 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2151 * If someone calls draining, avoid adding more kworker runs.
2153 if (!mutex_trylock(&percpu_charge_mutex))
2155 drain_all_stock(root_memcg, false);
2156 mutex_unlock(&percpu_charge_mutex);
2159 /* This is a synchronous drain interface. */
2160 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2162 /* called when force_empty is called */
2163 mutex_lock(&percpu_charge_mutex);
2164 drain_all_stock(root_memcg, true);
2165 mutex_unlock(&percpu_charge_mutex);
2169 * This function drains percpu counter value from DEAD cpu and
2170 * move it to local cpu. Note that this function can be preempted.
2172 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2176 spin_lock(&memcg->pcp_counter_lock);
2177 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2178 long x = per_cpu(memcg->stat->count[i], cpu);
2180 per_cpu(memcg->stat->count[i], cpu) = 0;
2181 memcg->nocpu_base.count[i] += x;
2183 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2184 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2186 per_cpu(memcg->stat->events[i], cpu) = 0;
2187 memcg->nocpu_base.events[i] += x;
2189 spin_unlock(&memcg->pcp_counter_lock);
2192 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2193 unsigned long action,
2196 int cpu = (unsigned long)hcpu;
2197 struct memcg_stock_pcp *stock;
2198 struct mem_cgroup *iter;
2200 if (action == CPU_ONLINE)
2203 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2206 for_each_mem_cgroup(iter)
2207 mem_cgroup_drain_pcp_counter(iter, cpu);
2209 stock = &per_cpu(memcg_stock, cpu);
2215 /* See __mem_cgroup_try_charge() for details */
2217 CHARGE_OK, /* success */
2218 CHARGE_RETRY, /* need to retry but retry is not bad */
2219 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2220 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2221 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2224 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2225 unsigned int nr_pages, bool oom_check)
2227 unsigned long csize = nr_pages * PAGE_SIZE;
2228 struct mem_cgroup *mem_over_limit;
2229 struct res_counter *fail_res;
2230 unsigned long flags = 0;
2233 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2236 if (!do_swap_account)
2238 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2242 res_counter_uncharge(&memcg->res, csize);
2243 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2244 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2246 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2248 * nr_pages can be either a huge page (HPAGE_PMD_NR), a batch
2249 * of regular pages (CHARGE_BATCH), or a single regular page (1).
2251 * Never reclaim on behalf of optional batching, retry with a
2252 * single page instead.
2254 if (nr_pages == CHARGE_BATCH)
2255 return CHARGE_RETRY;
2257 if (!(gfp_mask & __GFP_WAIT))
2258 return CHARGE_WOULDBLOCK;
2260 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2261 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2262 return CHARGE_RETRY;
2264 * Even though the limit is exceeded at this point, reclaim
2265 * may have been able to free some pages. Retry the charge
2266 * before killing the task.
2268 * Only for regular pages, though: huge pages are rather
2269 * unlikely to succeed so close to the limit, and we fall back
2270 * to regular pages anyway in case of failure.
2272 if (nr_pages == 1 && ret)
2273 return CHARGE_RETRY;
2276 * At task move, charge accounts can be doubly counted. So, it's
2277 * better to wait until the end of task_move if something is going on.
2279 if (mem_cgroup_wait_acct_move(mem_over_limit))
2280 return CHARGE_RETRY;
2282 /* If we don't need to call oom-killer at el, return immediately */
2284 return CHARGE_NOMEM;
2286 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask, get_order(csize)))
2287 return CHARGE_OOM_DIE;
2289 return CHARGE_RETRY;
2293 * __mem_cgroup_try_charge() does
2294 * 1. detect memcg to be charged against from passed *mm and *ptr,
2295 * 2. update res_counter
2296 * 3. call memory reclaim if necessary.
2298 * In some special case, if the task is fatal, fatal_signal_pending() or
2299 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2300 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2301 * as possible without any hazards. 2: all pages should have a valid
2302 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2303 * pointer, that is treated as a charge to root_mem_cgroup.
2305 * So __mem_cgroup_try_charge() will return
2306 * 0 ... on success, filling *ptr with a valid memcg pointer.
2307 * -ENOMEM ... charge failure because of resource limits.
2308 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2310 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2311 * the oom-killer can be invoked.
2313 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2315 unsigned int nr_pages,
2316 struct mem_cgroup **ptr,
2319 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2320 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2321 struct mem_cgroup *memcg = NULL;
2325 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2326 * in system level. So, allow to go ahead dying process in addition to
2329 if (unlikely(test_thread_flag(TIF_MEMDIE)
2330 || fatal_signal_pending(current)))
2334 * We always charge the cgroup the mm_struct belongs to.
2335 * The mm_struct's mem_cgroup changes on task migration if the
2336 * thread group leader migrates. It's possible that mm is not
2337 * set, if so charge the init_mm (happens for pagecache usage).
2340 *ptr = root_mem_cgroup;
2342 if (*ptr) { /* css should be a valid one */
2344 VM_BUG_ON(css_is_removed(&memcg->css));
2345 if (mem_cgroup_is_root(memcg))
2347 if (nr_pages == 1 && consume_stock(memcg))
2349 css_get(&memcg->css);
2351 struct task_struct *p;
2354 p = rcu_dereference(mm->owner);
2356 * Because we don't have task_lock(), "p" can exit.
2357 * In that case, "memcg" can point to root or p can be NULL with
2358 * race with swapoff. Then, we have small risk of mis-accouning.
2359 * But such kind of mis-account by race always happens because
2360 * we don't have cgroup_mutex(). It's overkill and we allo that
2362 * (*) swapoff at el will charge against mm-struct not against
2363 * task-struct. So, mm->owner can be NULL.
2365 memcg = mem_cgroup_from_task(p);
2367 memcg = root_mem_cgroup;
2368 if (mem_cgroup_is_root(memcg)) {
2372 if (nr_pages == 1 && consume_stock(memcg)) {
2374 * It seems dagerous to access memcg without css_get().
2375 * But considering how consume_stok works, it's not
2376 * necessary. If consume_stock success, some charges
2377 * from this memcg are cached on this cpu. So, we
2378 * don't need to call css_get()/css_tryget() before
2379 * calling consume_stock().
2384 /* after here, we may be blocked. we need to get refcnt */
2385 if (!css_tryget(&memcg->css)) {
2395 /* If killed, bypass charge */
2396 if (fatal_signal_pending(current)) {
2397 css_put(&memcg->css);
2402 if (oom && !nr_oom_retries) {
2404 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2407 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, oom_check);
2411 case CHARGE_RETRY: /* not in OOM situation but retry */
2413 css_put(&memcg->css);
2416 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2417 css_put(&memcg->css);
2419 case CHARGE_NOMEM: /* OOM routine works */
2421 css_put(&memcg->css);
2424 /* If oom, we never return -ENOMEM */
2427 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2428 css_put(&memcg->css);
2431 } while (ret != CHARGE_OK);
2433 if (batch > nr_pages)
2434 refill_stock(memcg, batch - nr_pages);
2435 css_put(&memcg->css);
2443 *ptr = root_mem_cgroup;
2448 * Somemtimes we have to undo a charge we got by try_charge().
2449 * This function is for that and do uncharge, put css's refcnt.
2450 * gotten by try_charge().
2452 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2453 unsigned int nr_pages)
2455 if (!mem_cgroup_is_root(memcg)) {
2456 unsigned long bytes = nr_pages * PAGE_SIZE;
2458 res_counter_uncharge(&memcg->res, bytes);
2459 if (do_swap_account)
2460 res_counter_uncharge(&memcg->memsw, bytes);
2465 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2466 * This is useful when moving usage to parent cgroup.
2468 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2469 unsigned int nr_pages)
2471 unsigned long bytes = nr_pages * PAGE_SIZE;
2473 if (mem_cgroup_is_root(memcg))
2476 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2477 if (do_swap_account)
2478 res_counter_uncharge_until(&memcg->memsw,
2479 memcg->memsw.parent, bytes);
2483 * A helper function to get mem_cgroup from ID. must be called under
2484 * rcu_read_lock(). The caller must check css_is_removed() or some if
2485 * it's concern. (dropping refcnt from swap can be called against removed
2488 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2490 struct cgroup_subsys_state *css;
2492 /* ID 0 is unused ID */
2495 css = css_lookup(&mem_cgroup_subsys, id);
2498 return container_of(css, struct mem_cgroup, css);
2501 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2503 struct mem_cgroup *memcg = NULL;
2504 struct page_cgroup *pc;
2508 VM_BUG_ON(!PageLocked(page));
2510 pc = lookup_page_cgroup(page);
2511 lock_page_cgroup(pc);
2512 if (PageCgroupUsed(pc)) {
2513 memcg = pc->mem_cgroup;
2514 if (memcg && !css_tryget(&memcg->css))
2516 } else if (PageSwapCache(page)) {
2517 ent.val = page_private(page);
2518 id = lookup_swap_cgroup_id(ent);
2520 memcg = mem_cgroup_lookup(id);
2521 if (memcg && !css_tryget(&memcg->css))
2525 unlock_page_cgroup(pc);
2529 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2531 unsigned int nr_pages,
2532 enum charge_type ctype,
2535 struct page_cgroup *pc = lookup_page_cgroup(page);
2536 struct zone *uninitialized_var(zone);
2537 struct lruvec *lruvec;
2538 bool was_on_lru = false;
2541 lock_page_cgroup(pc);
2542 if (unlikely(PageCgroupUsed(pc))) {
2543 unlock_page_cgroup(pc);
2544 __mem_cgroup_cancel_charge(memcg, nr_pages);
2548 * we don't need page_cgroup_lock about tail pages, becase they are not
2549 * accessed by any other context at this point.
2553 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2554 * may already be on some other mem_cgroup's LRU. Take care of it.
2557 zone = page_zone(page);
2558 spin_lock_irq(&zone->lru_lock);
2559 if (PageLRU(page)) {
2560 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2562 del_page_from_lru_list(page, lruvec, page_lru(page));
2567 pc->mem_cgroup = memcg;
2569 * We access a page_cgroup asynchronously without lock_page_cgroup().
2570 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2571 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2572 * before USED bit, we need memory barrier here.
2573 * See mem_cgroup_add_lru_list(), etc.
2576 SetPageCgroupUsed(pc);
2580 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2581 VM_BUG_ON(PageLRU(page));
2583 add_page_to_lru_list(page, lruvec, page_lru(page));
2585 spin_unlock_irq(&zone->lru_lock);
2588 if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
2593 mem_cgroup_charge_statistics(memcg, anon, nr_pages);
2594 unlock_page_cgroup(pc);
2597 * "charge_statistics" updated event counter. Then, check it.
2598 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2599 * if they exceeds softlimit.
2601 memcg_check_events(memcg, page);
2604 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2606 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
2608 * Because tail pages are not marked as "used", set it. We're under
2609 * zone->lru_lock, 'splitting on pmd' and compound_lock.
2610 * charge/uncharge will be never happen and move_account() is done under
2611 * compound_lock(), so we don't have to take care of races.
2613 void mem_cgroup_split_huge_fixup(struct page *head)
2615 struct page_cgroup *head_pc = lookup_page_cgroup(head);
2616 struct page_cgroup *pc;
2619 if (mem_cgroup_disabled())
2621 for (i = 1; i < HPAGE_PMD_NR; i++) {
2623 pc->mem_cgroup = head_pc->mem_cgroup;
2624 smp_wmb();/* see __commit_charge() */
2625 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2628 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2631 * mem_cgroup_move_account - move account of the page
2633 * @nr_pages: number of regular pages (>1 for huge pages)
2634 * @pc: page_cgroup of the page.
2635 * @from: mem_cgroup which the page is moved from.
2636 * @to: mem_cgroup which the page is moved to. @from != @to.
2638 * The caller must confirm following.
2639 * - page is not on LRU (isolate_page() is useful.)
2640 * - compound_lock is held when nr_pages > 1
2642 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
2645 static int mem_cgroup_move_account(struct page *page,
2646 unsigned int nr_pages,
2647 struct page_cgroup *pc,
2648 struct mem_cgroup *from,
2649 struct mem_cgroup *to)
2651 unsigned long flags;
2653 bool anon = PageAnon(page);
2655 VM_BUG_ON(from == to);
2656 VM_BUG_ON(PageLRU(page));
2658 * The page is isolated from LRU. So, collapse function
2659 * will not handle this page. But page splitting can happen.
2660 * Do this check under compound_page_lock(). The caller should
2664 if (nr_pages > 1 && !PageTransHuge(page))
2667 lock_page_cgroup(pc);
2670 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
2673 move_lock_mem_cgroup(from, &flags);
2675 if (!anon && page_mapped(page)) {
2676 /* Update mapped_file data for mem_cgroup */
2678 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2679 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2682 mem_cgroup_charge_statistics(from, anon, -nr_pages);
2684 /* caller should have done css_get */
2685 pc->mem_cgroup = to;
2686 mem_cgroup_charge_statistics(to, anon, nr_pages);
2688 * We charges against "to" which may not have any tasks. Then, "to"
2689 * can be under rmdir(). But in current implementation, caller of
2690 * this function is just force_empty() and move charge, so it's
2691 * guaranteed that "to" is never removed. So, we don't check rmdir
2694 move_unlock_mem_cgroup(from, &flags);
2697 unlock_page_cgroup(pc);
2701 memcg_check_events(to, page);
2702 memcg_check_events(from, page);
2708 * move charges to its parent.
2711 static int mem_cgroup_move_parent(struct page *page,
2712 struct page_cgroup *pc,
2713 struct mem_cgroup *child)
2715 struct mem_cgroup *parent;
2716 unsigned int nr_pages;
2717 unsigned long uninitialized_var(flags);
2721 if (mem_cgroup_is_root(child))
2725 if (!get_page_unless_zero(page))
2727 if (isolate_lru_page(page))
2730 nr_pages = hpage_nr_pages(page);
2732 parent = parent_mem_cgroup(child);
2734 * If no parent, move charges to root cgroup.
2737 parent = root_mem_cgroup;
2740 flags = compound_lock_irqsave(page);
2742 ret = mem_cgroup_move_account(page, nr_pages,
2745 __mem_cgroup_cancel_local_charge(child, nr_pages);
2748 compound_unlock_irqrestore(page, flags);
2749 putback_lru_page(page);
2757 * Charge the memory controller for page usage.
2759 * 0 if the charge was successful
2760 * < 0 if the cgroup is over its limit
2762 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2763 gfp_t gfp_mask, enum charge_type ctype)
2765 struct mem_cgroup *memcg = NULL;
2766 unsigned int nr_pages = 1;
2770 if (PageTransHuge(page)) {
2771 nr_pages <<= compound_order(page);
2772 VM_BUG_ON(!PageTransHuge(page));
2774 * Never OOM-kill a process for a huge page. The
2775 * fault handler will fall back to regular pages.
2780 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
2783 __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false);
2787 int mem_cgroup_newpage_charge(struct page *page,
2788 struct mm_struct *mm, gfp_t gfp_mask)
2790 if (mem_cgroup_disabled())
2792 VM_BUG_ON(page_mapped(page));
2793 VM_BUG_ON(page->mapping && !PageAnon(page));
2795 return mem_cgroup_charge_common(page, mm, gfp_mask,
2796 MEM_CGROUP_CHARGE_TYPE_ANON);
2800 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2801 enum charge_type ctype);
2803 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2806 struct mem_cgroup *memcg = NULL;
2807 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
2810 if (mem_cgroup_disabled())
2812 if (PageCompound(page))
2817 if (!page_is_file_cache(page))
2818 type = MEM_CGROUP_CHARGE_TYPE_SHMEM;
2820 if (!PageSwapCache(page))
2821 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
2822 else { /* page is swapcache/shmem */
2823 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &memcg);
2825 __mem_cgroup_commit_charge_swapin(page, memcg, type);
2831 * While swap-in, try_charge -> commit or cancel, the page is locked.
2832 * And when try_charge() successfully returns, one refcnt to memcg without
2833 * struct page_cgroup is acquired. This refcnt will be consumed by
2834 * "commit()" or removed by "cancel()"
2836 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2838 gfp_t mask, struct mem_cgroup **memcgp)
2840 struct mem_cgroup *memcg;
2845 if (mem_cgroup_disabled())
2848 if (!do_swap_account)
2851 * A racing thread's fault, or swapoff, may have already updated
2852 * the pte, and even removed page from swap cache: in those cases
2853 * do_swap_page()'s pte_same() test will fail; but there's also a
2854 * KSM case which does need to charge the page.
2856 if (!PageSwapCache(page))
2858 memcg = try_get_mem_cgroup_from_page(page);
2862 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
2863 css_put(&memcg->css);
2870 ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
2877 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
2878 enum charge_type ctype)
2880 if (mem_cgroup_disabled())
2884 cgroup_exclude_rmdir(&memcg->css);
2886 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
2888 * Now swap is on-memory. This means this page may be
2889 * counted both as mem and swap....double count.
2890 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2891 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2892 * may call delete_from_swap_cache() before reach here.
2894 if (do_swap_account && PageSwapCache(page)) {
2895 swp_entry_t ent = {.val = page_private(page)};
2896 mem_cgroup_uncharge_swap(ent);
2899 * At swapin, we may charge account against cgroup which has no tasks.
2900 * So, rmdir()->pre_destroy() can be called while we do this charge.
2901 * In that case, we need to call pre_destroy() again. check it here.
2903 cgroup_release_and_wakeup_rmdir(&memcg->css);
2906 void mem_cgroup_commit_charge_swapin(struct page *page,
2907 struct mem_cgroup *memcg)
2909 __mem_cgroup_commit_charge_swapin(page, memcg,
2910 MEM_CGROUP_CHARGE_TYPE_ANON);
2913 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
2915 if (mem_cgroup_disabled())
2919 __mem_cgroup_cancel_charge(memcg, 1);
2922 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
2923 unsigned int nr_pages,
2924 const enum charge_type ctype)
2926 struct memcg_batch_info *batch = NULL;
2927 bool uncharge_memsw = true;
2929 /* If swapout, usage of swap doesn't decrease */
2930 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2931 uncharge_memsw = false;
2933 batch = ¤t->memcg_batch;
2935 * In usual, we do css_get() when we remember memcg pointer.
2936 * But in this case, we keep res->usage until end of a series of
2937 * uncharges. Then, it's ok to ignore memcg's refcnt.
2940 batch->memcg = memcg;
2942 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2943 * In those cases, all pages freed continuously can be expected to be in
2944 * the same cgroup and we have chance to coalesce uncharges.
2945 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2946 * because we want to do uncharge as soon as possible.
2949 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
2950 goto direct_uncharge;
2953 goto direct_uncharge;
2956 * In typical case, batch->memcg == mem. This means we can
2957 * merge a series of uncharges to an uncharge of res_counter.
2958 * If not, we uncharge res_counter ony by one.
2960 if (batch->memcg != memcg)
2961 goto direct_uncharge;
2962 /* remember freed charge and uncharge it later */
2965 batch->memsw_nr_pages++;
2968 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
2970 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
2971 if (unlikely(batch->memcg != memcg))
2972 memcg_oom_recover(memcg);
2976 * uncharge if !page_mapped(page)
2978 static struct mem_cgroup *
2979 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
2982 struct mem_cgroup *memcg = NULL;
2983 unsigned int nr_pages = 1;
2984 struct page_cgroup *pc;
2987 if (mem_cgroup_disabled())
2990 VM_BUG_ON(PageSwapCache(page));
2992 if (PageTransHuge(page)) {
2993 nr_pages <<= compound_order(page);
2994 VM_BUG_ON(!PageTransHuge(page));
2997 * Check if our page_cgroup is valid
2999 pc = lookup_page_cgroup(page);
3000 if (unlikely(!PageCgroupUsed(pc)))
3003 lock_page_cgroup(pc);
3005 memcg = pc->mem_cgroup;
3007 if (!PageCgroupUsed(pc))
3010 anon = PageAnon(page);
3013 case MEM_CGROUP_CHARGE_TYPE_ANON:
3015 * Generally PageAnon tells if it's the anon statistics to be
3016 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
3017 * used before page reached the stage of being marked PageAnon.
3021 case MEM_CGROUP_CHARGE_TYPE_DROP:
3022 /* See mem_cgroup_prepare_migration() */
3023 if (page_mapped(page))
3026 * Pages under migration may not be uncharged. But
3027 * end_migration() /must/ be the one uncharging the
3028 * unused post-migration page and so it has to call
3029 * here with the migration bit still set. See the
3030 * res_counter handling below.
3032 if (!end_migration && PageCgroupMigration(pc))
3035 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
3036 if (!PageAnon(page)) { /* Shared memory */
3037 if (page->mapping && !page_is_file_cache(page))
3039 } else if (page_mapped(page)) /* Anon */
3046 mem_cgroup_charge_statistics(memcg, anon, -nr_pages);
3048 ClearPageCgroupUsed(pc);
3050 * pc->mem_cgroup is not cleared here. It will be accessed when it's
3051 * freed from LRU. This is safe because uncharged page is expected not
3052 * to be reused (freed soon). Exception is SwapCache, it's handled by
3053 * special functions.
3056 unlock_page_cgroup(pc);
3058 * even after unlock, we have memcg->res.usage here and this memcg
3059 * will never be freed.
3061 memcg_check_events(memcg, page);
3062 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
3063 mem_cgroup_swap_statistics(memcg, true);
3064 mem_cgroup_get(memcg);
3067 * Migration does not charge the res_counter for the
3068 * replacement page, so leave it alone when phasing out the
3069 * page that is unused after the migration.
3071 if (!end_migration && !mem_cgroup_is_root(memcg))
3072 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
3077 unlock_page_cgroup(pc);
3081 void mem_cgroup_uncharge_page(struct page *page)
3084 if (page_mapped(page))
3086 VM_BUG_ON(page->mapping && !PageAnon(page));
3087 if (PageSwapCache(page))
3089 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
3092 void mem_cgroup_uncharge_cache_page(struct page *page)
3094 VM_BUG_ON(page_mapped(page));
3095 VM_BUG_ON(page->mapping);
3096 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
3100 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
3101 * In that cases, pages are freed continuously and we can expect pages
3102 * are in the same memcg. All these calls itself limits the number of
3103 * pages freed at once, then uncharge_start/end() is called properly.
3104 * This may be called prural(2) times in a context,
3107 void mem_cgroup_uncharge_start(void)
3109 current->memcg_batch.do_batch++;
3110 /* We can do nest. */
3111 if (current->memcg_batch.do_batch == 1) {
3112 current->memcg_batch.memcg = NULL;
3113 current->memcg_batch.nr_pages = 0;
3114 current->memcg_batch.memsw_nr_pages = 0;
3118 void mem_cgroup_uncharge_end(void)
3120 struct memcg_batch_info *batch = ¤t->memcg_batch;
3122 if (!batch->do_batch)
3126 if (batch->do_batch) /* If stacked, do nothing. */
3132 * This "batch->memcg" is valid without any css_get/put etc...
3133 * bacause we hide charges behind us.
3135 if (batch->nr_pages)
3136 res_counter_uncharge(&batch->memcg->res,
3137 batch->nr_pages * PAGE_SIZE);
3138 if (batch->memsw_nr_pages)
3139 res_counter_uncharge(&batch->memcg->memsw,
3140 batch->memsw_nr_pages * PAGE_SIZE);
3141 memcg_oom_recover(batch->memcg);
3142 /* forget this pointer (for sanity check) */
3143 batch->memcg = NULL;
3148 * called after __delete_from_swap_cache() and drop "page" account.
3149 * memcg information is recorded to swap_cgroup of "ent"
3152 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
3154 struct mem_cgroup *memcg;
3155 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
3157 if (!swapout) /* this was a swap cache but the swap is unused ! */
3158 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
3160 memcg = __mem_cgroup_uncharge_common(page, ctype, false);
3163 * record memcg information, if swapout && memcg != NULL,
3164 * mem_cgroup_get() was called in uncharge().
3166 if (do_swap_account && swapout && memcg)
3167 swap_cgroup_record(ent, css_id(&memcg->css));
3171 #ifdef CONFIG_MEMCG_SWAP
3173 * called from swap_entry_free(). remove record in swap_cgroup and
3174 * uncharge "memsw" account.
3176 void mem_cgroup_uncharge_swap(swp_entry_t ent)
3178 struct mem_cgroup *memcg;
3181 if (!do_swap_account)
3184 id = swap_cgroup_record(ent, 0);
3186 memcg = mem_cgroup_lookup(id);
3189 * We uncharge this because swap is freed.
3190 * This memcg can be obsolete one. We avoid calling css_tryget
3192 if (!mem_cgroup_is_root(memcg))
3193 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
3194 mem_cgroup_swap_statistics(memcg, false);
3195 mem_cgroup_put(memcg);
3201 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3202 * @entry: swap entry to be moved
3203 * @from: mem_cgroup which the entry is moved from
3204 * @to: mem_cgroup which the entry is moved to
3206 * It succeeds only when the swap_cgroup's record for this entry is the same
3207 * as the mem_cgroup's id of @from.
3209 * Returns 0 on success, -EINVAL on failure.
3211 * The caller must have charged to @to, IOW, called res_counter_charge() about
3212 * both res and memsw, and called css_get().
3214 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3215 struct mem_cgroup *from, struct mem_cgroup *to)
3217 unsigned short old_id, new_id;
3219 old_id = css_id(&from->css);
3220 new_id = css_id(&to->css);
3222 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3223 mem_cgroup_swap_statistics(from, false);
3224 mem_cgroup_swap_statistics(to, true);
3226 * This function is only called from task migration context now.
3227 * It postpones res_counter and refcount handling till the end
3228 * of task migration(mem_cgroup_clear_mc()) for performance
3229 * improvement. But we cannot postpone mem_cgroup_get(to)
3230 * because if the process that has been moved to @to does
3231 * swap-in, the refcount of @to might be decreased to 0.
3239 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3240 struct mem_cgroup *from, struct mem_cgroup *to)
3247 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
3250 void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
3251 struct mem_cgroup **memcgp)
3253 struct mem_cgroup *memcg = NULL;
3254 struct page_cgroup *pc;
3255 enum charge_type ctype;
3259 VM_BUG_ON(PageTransHuge(page));
3260 if (mem_cgroup_disabled())
3263 pc = lookup_page_cgroup(page);
3264 lock_page_cgroup(pc);
3265 if (PageCgroupUsed(pc)) {
3266 memcg = pc->mem_cgroup;
3267 css_get(&memcg->css);
3269 * At migrating an anonymous page, its mapcount goes down
3270 * to 0 and uncharge() will be called. But, even if it's fully
3271 * unmapped, migration may fail and this page has to be
3272 * charged again. We set MIGRATION flag here and delay uncharge
3273 * until end_migration() is called
3275 * Corner Case Thinking
3277 * When the old page was mapped as Anon and it's unmap-and-freed
3278 * while migration was ongoing.
3279 * If unmap finds the old page, uncharge() of it will be delayed
3280 * until end_migration(). If unmap finds a new page, it's
3281 * uncharged when it make mapcount to be 1->0. If unmap code
3282 * finds swap_migration_entry, the new page will not be mapped
3283 * and end_migration() will find it(mapcount==0).
3286 * When the old page was mapped but migraion fails, the kernel
3287 * remaps it. A charge for it is kept by MIGRATION flag even
3288 * if mapcount goes down to 0. We can do remap successfully
3289 * without charging it again.
3292 * The "old" page is under lock_page() until the end of
3293 * migration, so, the old page itself will not be swapped-out.
3294 * If the new page is swapped out before end_migraton, our
3295 * hook to usual swap-out path will catch the event.
3298 SetPageCgroupMigration(pc);
3300 unlock_page_cgroup(pc);
3302 * If the page is not charged at this point,
3310 * We charge new page before it's used/mapped. So, even if unlock_page()
3311 * is called before end_migration, we can catch all events on this new
3312 * page. In the case new page is migrated but not remapped, new page's
3313 * mapcount will be finally 0 and we call uncharge in end_migration().
3316 ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
3317 else if (page_is_file_cache(page))
3318 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
3320 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3322 * The page is committed to the memcg, but it's not actually
3323 * charged to the res_counter since we plan on replacing the
3324 * old one and only one page is going to be left afterwards.
3326 __mem_cgroup_commit_charge(memcg, newpage, 1, ctype, false);
3329 /* remove redundant charge if migration failed*/
3330 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
3331 struct page *oldpage, struct page *newpage, bool migration_ok)
3333 struct page *used, *unused;
3334 struct page_cgroup *pc;
3339 /* blocks rmdir() */
3340 cgroup_exclude_rmdir(&memcg->css);
3341 if (!migration_ok) {
3348 anon = PageAnon(used);
3349 __mem_cgroup_uncharge_common(unused,
3350 anon ? MEM_CGROUP_CHARGE_TYPE_ANON
3351 : MEM_CGROUP_CHARGE_TYPE_CACHE,
3353 css_put(&memcg->css);
3355 * We disallowed uncharge of pages under migration because mapcount
3356 * of the page goes down to zero, temporarly.
3357 * Clear the flag and check the page should be charged.
3359 pc = lookup_page_cgroup(oldpage);
3360 lock_page_cgroup(pc);
3361 ClearPageCgroupMigration(pc);
3362 unlock_page_cgroup(pc);
3365 * If a page is a file cache, radix-tree replacement is very atomic
3366 * and we can skip this check. When it was an Anon page, its mapcount
3367 * goes down to 0. But because we added MIGRATION flage, it's not
3368 * uncharged yet. There are several case but page->mapcount check
3369 * and USED bit check in mem_cgroup_uncharge_page() will do enough
3370 * check. (see prepare_charge() also)
3373 mem_cgroup_uncharge_page(used);
3375 * At migration, we may charge account against cgroup which has no
3377 * So, rmdir()->pre_destroy() can be called while we do this charge.
3378 * In that case, we need to call pre_destroy() again. check it here.
3380 cgroup_release_and_wakeup_rmdir(&memcg->css);
3384 * At replace page cache, newpage is not under any memcg but it's on
3385 * LRU. So, this function doesn't touch res_counter but handles LRU
3386 * in correct way. Both pages are locked so we cannot race with uncharge.
3388 void mem_cgroup_replace_page_cache(struct page *oldpage,
3389 struct page *newpage)
3391 struct mem_cgroup *memcg = NULL;
3392 struct page_cgroup *pc;
3393 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
3395 if (mem_cgroup_disabled())
3398 pc = lookup_page_cgroup(oldpage);
3399 /* fix accounting on old pages */
3400 lock_page_cgroup(pc);
3401 if (PageCgroupUsed(pc)) {
3402 memcg = pc->mem_cgroup;
3403 mem_cgroup_charge_statistics(memcg, false, -1);
3404 ClearPageCgroupUsed(pc);
3406 unlock_page_cgroup(pc);
3409 * When called from shmem_replace_page(), in some cases the
3410 * oldpage has already been charged, and in some cases not.
3415 if (PageSwapBacked(oldpage))
3416 type = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3419 * Even if newpage->mapping was NULL before starting replacement,
3420 * the newpage may be on LRU(or pagevec for LRU) already. We lock
3421 * LRU while we overwrite pc->mem_cgroup.
3423 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
3426 #ifdef CONFIG_DEBUG_VM
3427 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3429 struct page_cgroup *pc;
3431 pc = lookup_page_cgroup(page);
3433 * Can be NULL while feeding pages into the page allocator for
3434 * the first time, i.e. during boot or memory hotplug;
3435 * or when mem_cgroup_disabled().
3437 if (likely(pc) && PageCgroupUsed(pc))
3442 bool mem_cgroup_bad_page_check(struct page *page)
3444 if (mem_cgroup_disabled())
3447 return lookup_page_cgroup_used(page) != NULL;
3450 void mem_cgroup_print_bad_page(struct page *page)
3452 struct page_cgroup *pc;
3454 pc = lookup_page_cgroup_used(page);
3456 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
3457 pc, pc->flags, pc->mem_cgroup);
3462 static DEFINE_MUTEX(set_limit_mutex);
3464 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3465 unsigned long long val)
3468 u64 memswlimit, memlimit;
3470 int children = mem_cgroup_count_children(memcg);
3471 u64 curusage, oldusage;
3475 * For keeping hierarchical_reclaim simple, how long we should retry
3476 * is depends on callers. We set our retry-count to be function
3477 * of # of children which we should visit in this loop.
3479 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3481 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3484 while (retry_count) {
3485 if (signal_pending(current)) {
3490 * Rather than hide all in some function, I do this in
3491 * open coded manner. You see what this really does.
3492 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
3494 mutex_lock(&set_limit_mutex);
3495 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3496 if (memswlimit < val) {
3498 mutex_unlock(&set_limit_mutex);
3502 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3506 ret = res_counter_set_limit(&memcg->res, val);
3508 if (memswlimit == val)
3509 memcg->memsw_is_minimum = true;
3511 memcg->memsw_is_minimum = false;
3513 mutex_unlock(&set_limit_mutex);
3518 mem_cgroup_reclaim(memcg, GFP_KERNEL,
3519 MEM_CGROUP_RECLAIM_SHRINK);
3520 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3521 /* Usage is reduced ? */
3522 if (curusage >= oldusage)
3525 oldusage = curusage;
3527 if (!ret && enlarge)
3528 memcg_oom_recover(memcg);
3533 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3534 unsigned long long val)
3537 u64 memlimit, memswlimit, oldusage, curusage;
3538 int children = mem_cgroup_count_children(memcg);
3542 /* see mem_cgroup_resize_res_limit */
3543 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3544 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3545 while (retry_count) {
3546 if (signal_pending(current)) {
3551 * Rather than hide all in some function, I do this in
3552 * open coded manner. You see what this really does.
3553 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
3555 mutex_lock(&set_limit_mutex);
3556 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3557 if (memlimit > val) {
3559 mutex_unlock(&set_limit_mutex);
3562 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3563 if (memswlimit < val)
3565 ret = res_counter_set_limit(&memcg->memsw, val);
3567 if (memlimit == val)
3568 memcg->memsw_is_minimum = true;
3570 memcg->memsw_is_minimum = false;
3572 mutex_unlock(&set_limit_mutex);
3577 mem_cgroup_reclaim(memcg, GFP_KERNEL,
3578 MEM_CGROUP_RECLAIM_NOSWAP |
3579 MEM_CGROUP_RECLAIM_SHRINK);
3580 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3581 /* Usage is reduced ? */
3582 if (curusage >= oldusage)
3585 oldusage = curusage;
3587 if (!ret && enlarge)
3588 memcg_oom_recover(memcg);
3592 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3594 unsigned long *total_scanned)
3596 unsigned long nr_reclaimed = 0;
3597 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3598 unsigned long reclaimed;
3600 struct mem_cgroup_tree_per_zone *mctz;
3601 unsigned long long excess;
3602 unsigned long nr_scanned;
3607 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3609 * This loop can run a while, specially if mem_cgroup's continuously
3610 * keep exceeding their soft limit and putting the system under
3617 mz = mem_cgroup_largest_soft_limit_node(mctz);
3622 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
3623 gfp_mask, &nr_scanned);
3624 nr_reclaimed += reclaimed;
3625 *total_scanned += nr_scanned;
3626 spin_lock(&mctz->lock);
3629 * If we failed to reclaim anything from this memory cgroup
3630 * it is time to move on to the next cgroup
3636 * Loop until we find yet another one.
3638 * By the time we get the soft_limit lock
3639 * again, someone might have aded the
3640 * group back on the RB tree. Iterate to
3641 * make sure we get a different mem.
3642 * mem_cgroup_largest_soft_limit_node returns
3643 * NULL if no other cgroup is present on
3647 __mem_cgroup_largest_soft_limit_node(mctz);
3649 css_put(&next_mz->memcg->css);
3650 else /* next_mz == NULL or other memcg */
3654 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
3655 excess = res_counter_soft_limit_excess(&mz->memcg->res);
3657 * One school of thought says that we should not add
3658 * back the node to the tree if reclaim returns 0.
3659 * But our reclaim could return 0, simply because due
3660 * to priority we are exposing a smaller subset of
3661 * memory to reclaim from. Consider this as a longer
3664 /* If excess == 0, no tree ops */
3665 __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess);
3666 spin_unlock(&mctz->lock);
3667 css_put(&mz->memcg->css);
3670 * Could not reclaim anything and there are no more
3671 * mem cgroups to try or we seem to be looping without
3672 * reclaiming anything.
3674 if (!nr_reclaimed &&
3676 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3678 } while (!nr_reclaimed);
3680 css_put(&next_mz->memcg->css);
3681 return nr_reclaimed;
3685 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
3686 * reclaim the pages page themselves - it just removes the page_cgroups.
3687 * Returns true if some page_cgroups were not freed, indicating that the caller
3688 * must retry this operation.
3690 static bool mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
3691 int node, int zid, enum lru_list lru)
3693 struct mem_cgroup_per_zone *mz;
3694 unsigned long flags, loop;
3695 struct list_head *list;
3699 zone = &NODE_DATA(node)->node_zones[zid];
3700 mz = mem_cgroup_zoneinfo(memcg, node, zid);
3701 list = &mz->lruvec.lists[lru];
3703 loop = mz->lru_size[lru];
3704 /* give some margin against EBUSY etc...*/
3708 struct page_cgroup *pc;
3711 spin_lock_irqsave(&zone->lru_lock, flags);
3712 if (list_empty(list)) {
3713 spin_unlock_irqrestore(&zone->lru_lock, flags);
3716 page = list_entry(list->prev, struct page, lru);
3718 list_move(&page->lru, list);
3720 spin_unlock_irqrestore(&zone->lru_lock, flags);
3723 spin_unlock_irqrestore(&zone->lru_lock, flags);
3725 pc = lookup_page_cgroup(page);
3727 if (mem_cgroup_move_parent(page, pc, memcg)) {
3728 /* found lock contention or "pc" is obsolete. */
3734 return !list_empty(list);
3738 * make mem_cgroup's charge to be 0 if there is no task.
3739 * This enables deleting this mem_cgroup.
3741 static int mem_cgroup_force_empty(struct mem_cgroup *memcg, bool free_all)
3744 int node, zid, shrink;
3745 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3746 struct cgroup *cgrp = memcg->css.cgroup;
3748 css_get(&memcg->css);
3751 /* should free all ? */
3757 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3759 /* This is for making all *used* pages to be on LRU. */
3760 lru_add_drain_all();
3761 drain_all_stock_sync(memcg);
3763 mem_cgroup_start_move(memcg);
3764 for_each_node_state(node, N_HIGH_MEMORY) {
3765 for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3768 ret = mem_cgroup_force_empty_list(memcg,
3777 mem_cgroup_end_move(memcg);
3778 memcg_oom_recover(memcg);
3780 /* "ret" should also be checked to ensure all lists are empty. */
3781 } while (res_counter_read_u64(&memcg->res, RES_USAGE) > 0 || ret);
3783 css_put(&memcg->css);
3787 /* returns EBUSY if there is a task or if we come here twice. */
3788 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3792 /* we call try-to-free pages for make this cgroup empty */
3793 lru_add_drain_all();
3794 /* try to free all pages in this cgroup */
3796 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
3799 if (signal_pending(current)) {
3803 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
3807 /* maybe some writeback is necessary */
3808 congestion_wait(BLK_RW_ASYNC, HZ/10);
3813 /* try move_account...there may be some *locked* pages. */
3817 static int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3819 return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3823 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3825 return mem_cgroup_from_cont(cont)->use_hierarchy;
3828 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3832 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3833 struct cgroup *parent = cont->parent;
3834 struct mem_cgroup *parent_memcg = NULL;
3837 parent_memcg = mem_cgroup_from_cont(parent);
3841 if (memcg->use_hierarchy == val)
3845 * If parent's use_hierarchy is set, we can't make any modifications
3846 * in the child subtrees. If it is unset, then the change can
3847 * occur, provided the current cgroup has no children.
3849 * For the root cgroup, parent_mem is NULL, we allow value to be
3850 * set if there are no children.
3852 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3853 (val == 1 || val == 0)) {
3854 if (list_empty(&cont->children))
3855 memcg->use_hierarchy = val;
3868 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
3869 enum mem_cgroup_stat_index idx)
3871 struct mem_cgroup *iter;
3874 /* Per-cpu values can be negative, use a signed accumulator */
3875 for_each_mem_cgroup_tree(iter, memcg)
3876 val += mem_cgroup_read_stat(iter, idx);
3878 if (val < 0) /* race ? */
3883 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3887 if (!mem_cgroup_is_root(memcg)) {
3889 return res_counter_read_u64(&memcg->res, RES_USAGE);
3891 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
3894 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
3895 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
3898 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
3900 return val << PAGE_SHIFT;
3903 static ssize_t mem_cgroup_read(struct cgroup *cont, struct cftype *cft,
3904 struct file *file, char __user *buf,
3905 size_t nbytes, loff_t *ppos)
3907 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3910 int type, name, len;
3912 type = MEMFILE_TYPE(cft->private);
3913 name = MEMFILE_ATTR(cft->private);
3915 if (!do_swap_account && type == _MEMSWAP)
3920 if (name == RES_USAGE)
3921 val = mem_cgroup_usage(memcg, false);
3923 val = res_counter_read_u64(&memcg->res, name);
3926 if (name == RES_USAGE)
3927 val = mem_cgroup_usage(memcg, true);
3929 val = res_counter_read_u64(&memcg->memsw, name);
3935 len = scnprintf(str, sizeof(str), "%llu\n", (unsigned long long)val);
3936 return simple_read_from_buffer(buf, nbytes, ppos, str, len);
3939 * The user of this function is...
3942 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
3945 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3947 unsigned long long val;
3950 type = MEMFILE_TYPE(cft->private);
3951 name = MEMFILE_ATTR(cft->private);
3953 if (!do_swap_account && type == _MEMSWAP)
3958 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3962 /* This function does all necessary parse...reuse it */
3963 ret = res_counter_memparse_write_strategy(buffer, &val);
3967 ret = mem_cgroup_resize_limit(memcg, val);
3969 ret = mem_cgroup_resize_memsw_limit(memcg, val);
3971 case RES_SOFT_LIMIT:
3972 ret = res_counter_memparse_write_strategy(buffer, &val);
3976 * For memsw, soft limits are hard to implement in terms
3977 * of semantics, for now, we support soft limits for
3978 * control without swap
3981 ret = res_counter_set_soft_limit(&memcg->res, val);
3986 ret = -EINVAL; /* should be BUG() ? */
3992 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
3993 unsigned long long *mem_limit, unsigned long long *memsw_limit)
3995 struct cgroup *cgroup;
3996 unsigned long long min_limit, min_memsw_limit, tmp;
3998 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3999 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4000 cgroup = memcg->css.cgroup;
4001 if (!memcg->use_hierarchy)
4004 while (cgroup->parent) {
4005 cgroup = cgroup->parent;
4006 memcg = mem_cgroup_from_cont(cgroup);
4007 if (!memcg->use_hierarchy)
4009 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
4010 min_limit = min(min_limit, tmp);
4011 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4012 min_memsw_limit = min(min_memsw_limit, tmp);
4015 *mem_limit = min_limit;
4016 *memsw_limit = min_memsw_limit;
4019 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
4021 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4024 type = MEMFILE_TYPE(event);
4025 name = MEMFILE_ATTR(event);
4027 if (!do_swap_account && type == _MEMSWAP)
4033 res_counter_reset_max(&memcg->res);
4035 res_counter_reset_max(&memcg->memsw);
4039 res_counter_reset_failcnt(&memcg->res);
4041 res_counter_reset_failcnt(&memcg->memsw);
4048 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
4051 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
4055 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4056 struct cftype *cft, u64 val)
4058 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4060 if (val >= (1 << NR_MOVE_TYPE))
4063 * We check this value several times in both in can_attach() and
4064 * attach(), so we need cgroup lock to prevent this value from being
4068 memcg->move_charge_at_immigrate = val;
4074 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4075 struct cftype *cft, u64 val)
4082 static int memcg_numa_stat_show(struct cgroup *cont, struct cftype *cft,
4086 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
4087 unsigned long node_nr;
4088 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4090 total_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL);
4091 seq_printf(m, "total=%lu", total_nr);
4092 for_each_node_state(nid, N_HIGH_MEMORY) {
4093 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL);
4094 seq_printf(m, " N%d=%lu", nid, node_nr);
4098 file_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_FILE);
4099 seq_printf(m, "file=%lu", file_nr);
4100 for_each_node_state(nid, N_HIGH_MEMORY) {
4101 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4103 seq_printf(m, " N%d=%lu", nid, node_nr);
4107 anon_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_ANON);
4108 seq_printf(m, "anon=%lu", anon_nr);
4109 for_each_node_state(nid, N_HIGH_MEMORY) {
4110 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4112 seq_printf(m, " N%d=%lu", nid, node_nr);
4116 unevictable_nr = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
4117 seq_printf(m, "unevictable=%lu", unevictable_nr);
4118 for_each_node_state(nid, N_HIGH_MEMORY) {
4119 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4120 BIT(LRU_UNEVICTABLE));
4121 seq_printf(m, " N%d=%lu", nid, node_nr);
4126 #endif /* CONFIG_NUMA */
4128 static const char * const mem_cgroup_lru_names[] = {
4136 static inline void mem_cgroup_lru_names_not_uptodate(void)
4138 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
4141 static int memcg_stat_show(struct cgroup *cont, struct cftype *cft,
4144 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4145 struct mem_cgroup *mi;
4148 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
4149 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
4151 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
4152 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
4155 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
4156 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
4157 mem_cgroup_read_events(memcg, i));
4159 for (i = 0; i < NR_LRU_LISTS; i++)
4160 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
4161 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
4163 /* Hierarchical information */
4165 unsigned long long limit, memsw_limit;
4166 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
4167 seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
4168 if (do_swap_account)
4169 seq_printf(m, "hierarchical_memsw_limit %llu\n",
4173 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
4176 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
4178 for_each_mem_cgroup_tree(mi, memcg)
4179 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
4180 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
4183 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
4184 unsigned long long val = 0;
4186 for_each_mem_cgroup_tree(mi, memcg)
4187 val += mem_cgroup_read_events(mi, i);
4188 seq_printf(m, "total_%s %llu\n",
4189 mem_cgroup_events_names[i], val);
4192 for (i = 0; i < NR_LRU_LISTS; i++) {
4193 unsigned long long val = 0;
4195 for_each_mem_cgroup_tree(mi, memcg)
4196 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
4197 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
4200 #ifdef CONFIG_DEBUG_VM
4203 struct mem_cgroup_per_zone *mz;
4204 struct zone_reclaim_stat *rstat;
4205 unsigned long recent_rotated[2] = {0, 0};
4206 unsigned long recent_scanned[2] = {0, 0};
4208 for_each_online_node(nid)
4209 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4210 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
4211 rstat = &mz->lruvec.reclaim_stat;
4213 recent_rotated[0] += rstat->recent_rotated[0];
4214 recent_rotated[1] += rstat->recent_rotated[1];
4215 recent_scanned[0] += rstat->recent_scanned[0];
4216 recent_scanned[1] += rstat->recent_scanned[1];
4218 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
4219 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
4220 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
4221 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
4228 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
4230 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4232 return mem_cgroup_swappiness(memcg);
4235 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
4238 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4239 struct mem_cgroup *parent;
4244 if (cgrp->parent == NULL)
4247 parent = mem_cgroup_from_cont(cgrp->parent);
4251 /* If under hierarchy, only empty-root can set this value */
4252 if ((parent->use_hierarchy) ||
4253 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4258 memcg->swappiness = val;
4265 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4267 struct mem_cgroup_threshold_ary *t;
4273 t = rcu_dereference(memcg->thresholds.primary);
4275 t = rcu_dereference(memcg->memsw_thresholds.primary);
4280 usage = mem_cgroup_usage(memcg, swap);
4283 * current_threshold points to threshold just below or equal to usage.
4284 * If it's not true, a threshold was crossed after last
4285 * call of __mem_cgroup_threshold().
4287 i = t->current_threshold;
4290 * Iterate backward over array of thresholds starting from
4291 * current_threshold and check if a threshold is crossed.
4292 * If none of thresholds below usage is crossed, we read
4293 * only one element of the array here.
4295 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4296 eventfd_signal(t->entries[i].eventfd, 1);
4298 /* i = current_threshold + 1 */
4302 * Iterate forward over array of thresholds starting from
4303 * current_threshold+1 and check if a threshold is crossed.
4304 * If none of thresholds above usage is crossed, we read
4305 * only one element of the array here.
4307 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4308 eventfd_signal(t->entries[i].eventfd, 1);
4310 /* Update current_threshold */
4311 t->current_threshold = i - 1;
4316 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4319 __mem_cgroup_threshold(memcg, false);
4320 if (do_swap_account)
4321 __mem_cgroup_threshold(memcg, true);
4323 memcg = parent_mem_cgroup(memcg);
4327 static int compare_thresholds(const void *a, const void *b)
4329 const struct mem_cgroup_threshold *_a = a;
4330 const struct mem_cgroup_threshold *_b = b;
4332 return _a->threshold - _b->threshold;
4335 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4337 struct mem_cgroup_eventfd_list *ev;
4339 list_for_each_entry(ev, &memcg->oom_notify, list)
4340 eventfd_signal(ev->eventfd, 1);
4344 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4346 struct mem_cgroup *iter;
4348 for_each_mem_cgroup_tree(iter, memcg)
4349 mem_cgroup_oom_notify_cb(iter);
4352 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
4353 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4355 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4356 struct mem_cgroup_thresholds *thresholds;
4357 struct mem_cgroup_threshold_ary *new;
4358 int type = MEMFILE_TYPE(cft->private);
4359 u64 threshold, usage;
4362 ret = res_counter_memparse_write_strategy(args, &threshold);
4366 mutex_lock(&memcg->thresholds_lock);
4369 thresholds = &memcg->thresholds;
4370 else if (type == _MEMSWAP)
4371 thresholds = &memcg->memsw_thresholds;
4375 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4377 /* Check if a threshold crossed before adding a new one */
4378 if (thresholds->primary)
4379 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4381 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4383 /* Allocate memory for new array of thresholds */
4384 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4392 /* Copy thresholds (if any) to new array */
4393 if (thresholds->primary) {
4394 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4395 sizeof(struct mem_cgroup_threshold));
4398 /* Add new threshold */
4399 new->entries[size - 1].eventfd = eventfd;
4400 new->entries[size - 1].threshold = threshold;
4402 /* Sort thresholds. Registering of new threshold isn't time-critical */
4403 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4404 compare_thresholds, NULL);
4406 /* Find current threshold */
4407 new->current_threshold = -1;
4408 for (i = 0; i < size; i++) {
4409 if (new->entries[i].threshold <= usage) {
4411 * new->current_threshold will not be used until
4412 * rcu_assign_pointer(), so it's safe to increment
4415 ++new->current_threshold;
4420 /* Free old spare buffer and save old primary buffer as spare */
4421 kfree(thresholds->spare);
4422 thresholds->spare = thresholds->primary;
4424 rcu_assign_pointer(thresholds->primary, new);
4426 /* To be sure that nobody uses thresholds */
4430 mutex_unlock(&memcg->thresholds_lock);
4435 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
4436 struct cftype *cft, struct eventfd_ctx *eventfd)
4438 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4439 struct mem_cgroup_thresholds *thresholds;
4440 struct mem_cgroup_threshold_ary *new;
4441 int type = MEMFILE_TYPE(cft->private);
4445 mutex_lock(&memcg->thresholds_lock);
4447 thresholds = &memcg->thresholds;
4448 else if (type == _MEMSWAP)
4449 thresholds = &memcg->memsw_thresholds;
4453 if (!thresholds->primary)
4456 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4458 /* Check if a threshold crossed before removing */
4459 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4461 /* Calculate new number of threshold */
4463 for (i = 0; i < thresholds->primary->size; i++) {
4464 if (thresholds->primary->entries[i].eventfd != eventfd)
4468 new = thresholds->spare;
4470 /* Set thresholds array to NULL if we don't have thresholds */
4479 /* Copy thresholds and find current threshold */
4480 new->current_threshold = -1;
4481 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4482 if (thresholds->primary->entries[i].eventfd == eventfd)
4485 new->entries[j] = thresholds->primary->entries[i];
4486 if (new->entries[j].threshold <= usage) {
4488 * new->current_threshold will not be used
4489 * until rcu_assign_pointer(), so it's safe to increment
4492 ++new->current_threshold;
4498 /* Swap primary and spare array */
4499 thresholds->spare = thresholds->primary;
4500 /* If all events are unregistered, free the spare array */
4502 kfree(thresholds->spare);
4503 thresholds->spare = NULL;
4506 rcu_assign_pointer(thresholds->primary, new);
4508 /* To be sure that nobody uses thresholds */
4511 mutex_unlock(&memcg->thresholds_lock);
4514 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4515 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4517 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4518 struct mem_cgroup_eventfd_list *event;
4519 int type = MEMFILE_TYPE(cft->private);
4521 BUG_ON(type != _OOM_TYPE);
4522 event = kmalloc(sizeof(*event), GFP_KERNEL);
4526 spin_lock(&memcg_oom_lock);
4528 event->eventfd = eventfd;
4529 list_add(&event->list, &memcg->oom_notify);
4531 /* already in OOM ? */
4532 if (atomic_read(&memcg->under_oom))
4533 eventfd_signal(eventfd, 1);
4534 spin_unlock(&memcg_oom_lock);
4539 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4540 struct cftype *cft, struct eventfd_ctx *eventfd)
4542 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4543 struct mem_cgroup_eventfd_list *ev, *tmp;
4544 int type = MEMFILE_TYPE(cft->private);
4546 BUG_ON(type != _OOM_TYPE);
4548 spin_lock(&memcg_oom_lock);
4550 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4551 if (ev->eventfd == eventfd) {
4552 list_del(&ev->list);
4557 spin_unlock(&memcg_oom_lock);
4560 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4561 struct cftype *cft, struct cgroup_map_cb *cb)
4563 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4565 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
4567 if (atomic_read(&memcg->under_oom))
4568 cb->fill(cb, "under_oom", 1);
4570 cb->fill(cb, "under_oom", 0);
4574 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4575 struct cftype *cft, u64 val)
4577 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4578 struct mem_cgroup *parent;
4580 /* cannot set to root cgroup and only 0 and 1 are allowed */
4581 if (!cgrp->parent || !((val == 0) || (val == 1)))
4584 parent = mem_cgroup_from_cont(cgrp->parent);
4587 /* oom-kill-disable is a flag for subhierarchy. */
4588 if ((parent->use_hierarchy) ||
4589 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4593 memcg->oom_kill_disable = val;
4595 memcg_oom_recover(memcg);
4600 #ifdef CONFIG_MEMCG_KMEM
4601 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
4603 return mem_cgroup_sockets_init(memcg, ss);
4606 static void kmem_cgroup_destroy(struct mem_cgroup *memcg)
4608 mem_cgroup_sockets_destroy(memcg);
4611 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
4616 static void kmem_cgroup_destroy(struct mem_cgroup *memcg)
4621 static struct cftype mem_cgroup_files[] = {
4623 .name = "usage_in_bytes",
4624 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4625 .read = mem_cgroup_read,
4626 .register_event = mem_cgroup_usage_register_event,
4627 .unregister_event = mem_cgroup_usage_unregister_event,
4630 .name = "max_usage_in_bytes",
4631 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4632 .trigger = mem_cgroup_reset,
4633 .read = mem_cgroup_read,
4636 .name = "limit_in_bytes",
4637 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4638 .write_string = mem_cgroup_write,
4639 .read = mem_cgroup_read,
4642 .name = "soft_limit_in_bytes",
4643 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4644 .write_string = mem_cgroup_write,
4645 .read = mem_cgroup_read,
4649 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4650 .trigger = mem_cgroup_reset,
4651 .read = mem_cgroup_read,
4655 .read_seq_string = memcg_stat_show,
4658 .name = "force_empty",
4659 .trigger = mem_cgroup_force_empty_write,
4662 .name = "use_hierarchy",
4663 .write_u64 = mem_cgroup_hierarchy_write,
4664 .read_u64 = mem_cgroup_hierarchy_read,
4667 .name = "swappiness",
4668 .read_u64 = mem_cgroup_swappiness_read,
4669 .write_u64 = mem_cgroup_swappiness_write,
4672 .name = "move_charge_at_immigrate",
4673 .read_u64 = mem_cgroup_move_charge_read,
4674 .write_u64 = mem_cgroup_move_charge_write,
4677 .name = "oom_control",
4678 .read_map = mem_cgroup_oom_control_read,
4679 .write_u64 = mem_cgroup_oom_control_write,
4680 .register_event = mem_cgroup_oom_register_event,
4681 .unregister_event = mem_cgroup_oom_unregister_event,
4682 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4686 .name = "numa_stat",
4687 .read_seq_string = memcg_numa_stat_show,
4690 #ifdef CONFIG_MEMCG_SWAP
4692 .name = "memsw.usage_in_bytes",
4693 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4694 .read = mem_cgroup_read,
4695 .register_event = mem_cgroup_usage_register_event,
4696 .unregister_event = mem_cgroup_usage_unregister_event,
4699 .name = "memsw.max_usage_in_bytes",
4700 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4701 .trigger = mem_cgroup_reset,
4702 .read = mem_cgroup_read,
4705 .name = "memsw.limit_in_bytes",
4706 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4707 .write_string = mem_cgroup_write,
4708 .read = mem_cgroup_read,
4711 .name = "memsw.failcnt",
4712 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4713 .trigger = mem_cgroup_reset,
4714 .read = mem_cgroup_read,
4717 { }, /* terminate */
4720 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4722 struct mem_cgroup_per_node *pn;
4723 struct mem_cgroup_per_zone *mz;
4724 int zone, tmp = node;
4726 * This routine is called against possible nodes.
4727 * But it's BUG to call kmalloc() against offline node.
4729 * TODO: this routine can waste much memory for nodes which will
4730 * never be onlined. It's better to use memory hotplug callback
4733 if (!node_state(node, N_NORMAL_MEMORY))
4735 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4739 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4740 mz = &pn->zoneinfo[zone];
4741 lruvec_init(&mz->lruvec, &NODE_DATA(node)->node_zones[zone]);
4742 mz->usage_in_excess = 0;
4743 mz->on_tree = false;
4746 memcg->info.nodeinfo[node] = pn;
4750 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4752 kfree(memcg->info.nodeinfo[node]);
4755 static struct mem_cgroup *mem_cgroup_alloc(void)
4757 struct mem_cgroup *memcg;
4758 int size = sizeof(struct mem_cgroup);
4760 /* Can be very big if MAX_NUMNODES is very big */
4761 if (size < PAGE_SIZE)
4762 memcg = kzalloc(size, GFP_KERNEL);
4764 memcg = vzalloc(size);
4769 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4772 spin_lock_init(&memcg->pcp_counter_lock);
4776 if (size < PAGE_SIZE)
4784 * Helpers for freeing a kmalloc()ed/vzalloc()ed mem_cgroup by RCU,
4785 * but in process context. The work_freeing structure is overlaid
4786 * on the rcu_freeing structure, which itself is overlaid on memsw.
4788 static void free_work(struct work_struct *work)
4790 struct mem_cgroup *memcg;
4791 int size = sizeof(struct mem_cgroup);
4793 memcg = container_of(work, struct mem_cgroup, work_freeing);
4795 * We need to make sure that (at least for now), the jump label
4796 * destruction code runs outside of the cgroup lock. This is because
4797 * get_online_cpus(), which is called from the static_branch update,
4798 * can't be called inside the cgroup_lock. cpusets are the ones
4799 * enforcing this dependency, so if they ever change, we might as well.
4801 * schedule_work() will guarantee this happens. Be careful if you need
4802 * to move this code around, and make sure it is outside
4805 disarm_sock_keys(memcg);
4806 if (size < PAGE_SIZE)
4812 static void free_rcu(struct rcu_head *rcu_head)
4814 struct mem_cgroup *memcg;
4816 memcg = container_of(rcu_head, struct mem_cgroup, rcu_freeing);
4817 INIT_WORK(&memcg->work_freeing, free_work);
4818 schedule_work(&memcg->work_freeing);
4822 * At destroying mem_cgroup, references from swap_cgroup can remain.
4823 * (scanning all at force_empty is too costly...)
4825 * Instead of clearing all references at force_empty, we remember
4826 * the number of reference from swap_cgroup and free mem_cgroup when
4827 * it goes down to 0.
4829 * Removal of cgroup itself succeeds regardless of refs from swap.
4832 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4836 mem_cgroup_remove_from_trees(memcg);
4837 free_css_id(&mem_cgroup_subsys, &memcg->css);
4840 free_mem_cgroup_per_zone_info(memcg, node);
4842 free_percpu(memcg->stat);
4843 call_rcu(&memcg->rcu_freeing, free_rcu);
4846 static void mem_cgroup_get(struct mem_cgroup *memcg)
4848 atomic_inc(&memcg->refcnt);
4851 static void __mem_cgroup_put(struct mem_cgroup *memcg, int count)
4853 if (atomic_sub_and_test(count, &memcg->refcnt)) {
4854 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4855 __mem_cgroup_free(memcg);
4857 mem_cgroup_put(parent);
4861 static void mem_cgroup_put(struct mem_cgroup *memcg)
4863 __mem_cgroup_put(memcg, 1);
4867 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4869 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4871 if (!memcg->res.parent)
4873 return mem_cgroup_from_res_counter(memcg->res.parent, res);
4875 EXPORT_SYMBOL(parent_mem_cgroup);
4877 #ifdef CONFIG_MEMCG_SWAP
4878 static void __init enable_swap_cgroup(void)
4880 if (!mem_cgroup_disabled() && really_do_swap_account)
4881 do_swap_account = 1;
4884 static void __init enable_swap_cgroup(void)
4889 static int mem_cgroup_soft_limit_tree_init(void)
4891 struct mem_cgroup_tree_per_node *rtpn;
4892 struct mem_cgroup_tree_per_zone *rtpz;
4893 int tmp, node, zone;
4895 for_each_node(node) {
4897 if (!node_state(node, N_NORMAL_MEMORY))
4899 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
4903 soft_limit_tree.rb_tree_per_node[node] = rtpn;
4905 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4906 rtpz = &rtpn->rb_tree_per_zone[zone];
4907 rtpz->rb_root = RB_ROOT;
4908 spin_lock_init(&rtpz->lock);
4914 for_each_node(node) {
4915 if (!soft_limit_tree.rb_tree_per_node[node])
4917 kfree(soft_limit_tree.rb_tree_per_node[node]);
4918 soft_limit_tree.rb_tree_per_node[node] = NULL;
4924 static struct cgroup_subsys_state * __ref
4925 mem_cgroup_create(struct cgroup *cont)
4927 struct mem_cgroup *memcg, *parent;
4928 long error = -ENOMEM;
4931 memcg = mem_cgroup_alloc();
4933 return ERR_PTR(error);
4936 if (alloc_mem_cgroup_per_zone_info(memcg, node))
4940 if (cont->parent == NULL) {
4942 enable_swap_cgroup();
4944 if (mem_cgroup_soft_limit_tree_init())
4946 root_mem_cgroup = memcg;
4947 for_each_possible_cpu(cpu) {
4948 struct memcg_stock_pcp *stock =
4949 &per_cpu(memcg_stock, cpu);
4950 INIT_WORK(&stock->work, drain_local_stock);
4952 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
4954 parent = mem_cgroup_from_cont(cont->parent);
4955 memcg->use_hierarchy = parent->use_hierarchy;
4956 memcg->oom_kill_disable = parent->oom_kill_disable;
4959 if (parent && parent->use_hierarchy) {
4960 res_counter_init(&memcg->res, &parent->res);
4961 res_counter_init(&memcg->memsw, &parent->memsw);
4963 * We increment refcnt of the parent to ensure that we can
4964 * safely access it on res_counter_charge/uncharge.
4965 * This refcnt will be decremented when freeing this
4966 * mem_cgroup(see mem_cgroup_put).
4968 mem_cgroup_get(parent);
4970 res_counter_init(&memcg->res, NULL);
4971 res_counter_init(&memcg->memsw, NULL);
4973 memcg->last_scanned_node = MAX_NUMNODES;
4974 INIT_LIST_HEAD(&memcg->oom_notify);
4977 memcg->swappiness = mem_cgroup_swappiness(parent);
4978 atomic_set(&memcg->refcnt, 1);
4979 memcg->move_charge_at_immigrate = 0;
4980 mutex_init(&memcg->thresholds_lock);
4981 spin_lock_init(&memcg->move_lock);
4983 error = memcg_init_kmem(memcg, &mem_cgroup_subsys);
4986 * We call put now because our (and parent's) refcnts
4987 * are already in place. mem_cgroup_put() will internally
4988 * call __mem_cgroup_free, so return directly
4990 mem_cgroup_put(memcg);
4991 return ERR_PTR(error);
4995 __mem_cgroup_free(memcg);
4996 return ERR_PTR(error);
4999 static int mem_cgroup_pre_destroy(struct cgroup *cont)
5001 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5003 return mem_cgroup_force_empty(memcg, false);
5006 static void mem_cgroup_destroy(struct cgroup *cont)
5008 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5010 kmem_cgroup_destroy(memcg);
5012 mem_cgroup_put(memcg);
5016 /* Handlers for move charge at task migration. */
5017 #define PRECHARGE_COUNT_AT_ONCE 256
5018 static int mem_cgroup_do_precharge(unsigned long count)
5021 int batch_count = PRECHARGE_COUNT_AT_ONCE;
5022 struct mem_cgroup *memcg = mc.to;
5024 if (mem_cgroup_is_root(memcg)) {
5025 mc.precharge += count;
5026 /* we don't need css_get for root */
5029 /* try to charge at once */
5031 struct res_counter *dummy;
5033 * "memcg" cannot be under rmdir() because we've already checked
5034 * by cgroup_lock_live_cgroup() that it is not removed and we
5035 * are still under the same cgroup_mutex. So we can postpone
5038 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
5040 if (do_swap_account && res_counter_charge(&memcg->memsw,
5041 PAGE_SIZE * count, &dummy)) {
5042 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
5045 mc.precharge += count;
5049 /* fall back to one by one charge */
5051 if (signal_pending(current)) {
5055 if (!batch_count--) {
5056 batch_count = PRECHARGE_COUNT_AT_ONCE;
5059 ret = __mem_cgroup_try_charge(NULL,
5060 GFP_KERNEL, 1, &memcg, false);
5062 /* mem_cgroup_clear_mc() will do uncharge later */
5070 * get_mctgt_type - get target type of moving charge
5071 * @vma: the vma the pte to be checked belongs
5072 * @addr: the address corresponding to the pte to be checked
5073 * @ptent: the pte to be checked
5074 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5077 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5078 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5079 * move charge. if @target is not NULL, the page is stored in target->page
5080 * with extra refcnt got(Callers should handle it).
5081 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5082 * target for charge migration. if @target is not NULL, the entry is stored
5085 * Called with pte lock held.
5092 enum mc_target_type {
5098 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5099 unsigned long addr, pte_t ptent)
5101 struct page *page = vm_normal_page(vma, addr, ptent);
5103 if (!page || !page_mapped(page))
5105 if (PageAnon(page)) {
5106 /* we don't move shared anon */
5109 } else if (!move_file())
5110 /* we ignore mapcount for file pages */
5112 if (!get_page_unless_zero(page))
5119 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5120 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5122 struct page *page = NULL;
5123 swp_entry_t ent = pte_to_swp_entry(ptent);
5125 if (!move_anon() || non_swap_entry(ent))
5128 * Because lookup_swap_cache() updates some statistics counter,
5129 * we call find_get_page() with swapper_space directly.
5131 page = find_get_page(&swapper_space, ent.val);
5132 if (do_swap_account)
5133 entry->val = ent.val;
5138 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5139 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5145 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5146 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5148 struct page *page = NULL;
5149 struct address_space *mapping;
5152 if (!vma->vm_file) /* anonymous vma */
5157 mapping = vma->vm_file->f_mapping;
5158 if (pte_none(ptent))
5159 pgoff = linear_page_index(vma, addr);
5160 else /* pte_file(ptent) is true */
5161 pgoff = pte_to_pgoff(ptent);
5163 /* page is moved even if it's not RSS of this task(page-faulted). */
5164 page = find_get_page(mapping, pgoff);
5167 /* shmem/tmpfs may report page out on swap: account for that too. */
5168 if (radix_tree_exceptional_entry(page)) {
5169 swp_entry_t swap = radix_to_swp_entry(page);
5170 if (do_swap_account)
5172 page = find_get_page(&swapper_space, swap.val);
5178 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5179 unsigned long addr, pte_t ptent, union mc_target *target)
5181 struct page *page = NULL;
5182 struct page_cgroup *pc;
5183 enum mc_target_type ret = MC_TARGET_NONE;
5184 swp_entry_t ent = { .val = 0 };
5186 if (pte_present(ptent))
5187 page = mc_handle_present_pte(vma, addr, ptent);
5188 else if (is_swap_pte(ptent))
5189 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5190 else if (pte_none(ptent) || pte_file(ptent))
5191 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5193 if (!page && !ent.val)
5196 pc = lookup_page_cgroup(page);
5198 * Do only loose check w/o page_cgroup lock.
5199 * mem_cgroup_move_account() checks the pc is valid or not under
5202 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5203 ret = MC_TARGET_PAGE;
5205 target->page = page;
5207 if (!ret || !target)
5210 /* There is a swap entry and a page doesn't exist or isn't charged */
5211 if (ent.val && !ret &&
5212 css_id(&mc.from->css) == lookup_swap_cgroup_id(ent)) {
5213 ret = MC_TARGET_SWAP;
5220 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5222 * We don't consider swapping or file mapped pages because THP does not
5223 * support them for now.
5224 * Caller should make sure that pmd_trans_huge(pmd) is true.
5226 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5227 unsigned long addr, pmd_t pmd, union mc_target *target)
5229 struct page *page = NULL;
5230 struct page_cgroup *pc;
5231 enum mc_target_type ret = MC_TARGET_NONE;
5233 page = pmd_page(pmd);
5234 VM_BUG_ON(!page || !PageHead(page));
5237 pc = lookup_page_cgroup(page);
5238 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5239 ret = MC_TARGET_PAGE;
5242 target->page = page;
5248 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5249 unsigned long addr, pmd_t pmd, union mc_target *target)
5251 return MC_TARGET_NONE;
5255 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5256 unsigned long addr, unsigned long end,
5257 struct mm_walk *walk)
5259 struct vm_area_struct *vma = walk->private;
5263 if (pmd_trans_huge_lock(pmd, vma) == 1) {
5264 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5265 mc.precharge += HPAGE_PMD_NR;
5266 spin_unlock(&vma->vm_mm->page_table_lock);
5270 if (pmd_trans_unstable(pmd))
5272 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5273 for (; addr != end; pte++, addr += PAGE_SIZE)
5274 if (get_mctgt_type(vma, addr, *pte, NULL))
5275 mc.precharge++; /* increment precharge temporarily */
5276 pte_unmap_unlock(pte - 1, ptl);
5282 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5284 unsigned long precharge;
5285 struct vm_area_struct *vma;
5287 down_read(&mm->mmap_sem);
5288 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5289 struct mm_walk mem_cgroup_count_precharge_walk = {
5290 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5294 if (is_vm_hugetlb_page(vma))
5296 walk_page_range(vma->vm_start, vma->vm_end,
5297 &mem_cgroup_count_precharge_walk);
5299 up_read(&mm->mmap_sem);
5301 precharge = mc.precharge;
5307 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5309 unsigned long precharge = mem_cgroup_count_precharge(mm);
5311 VM_BUG_ON(mc.moving_task);
5312 mc.moving_task = current;
5313 return mem_cgroup_do_precharge(precharge);
5316 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5317 static void __mem_cgroup_clear_mc(void)
5319 struct mem_cgroup *from = mc.from;
5320 struct mem_cgroup *to = mc.to;
5322 /* we must uncharge all the leftover precharges from mc.to */
5324 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
5328 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5329 * we must uncharge here.
5331 if (mc.moved_charge) {
5332 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
5333 mc.moved_charge = 0;
5335 /* we must fixup refcnts and charges */
5336 if (mc.moved_swap) {
5337 /* uncharge swap account from the old cgroup */
5338 if (!mem_cgroup_is_root(mc.from))
5339 res_counter_uncharge(&mc.from->memsw,
5340 PAGE_SIZE * mc.moved_swap);
5341 __mem_cgroup_put(mc.from, mc.moved_swap);
5343 if (!mem_cgroup_is_root(mc.to)) {
5345 * we charged both to->res and to->memsw, so we should
5348 res_counter_uncharge(&mc.to->res,
5349 PAGE_SIZE * mc.moved_swap);
5351 /* we've already done mem_cgroup_get(mc.to) */
5354 memcg_oom_recover(from);
5355 memcg_oom_recover(to);
5356 wake_up_all(&mc.waitq);
5359 static void mem_cgroup_clear_mc(void)
5361 struct mem_cgroup *from = mc.from;
5364 * we must clear moving_task before waking up waiters at the end of
5367 mc.moving_task = NULL;
5368 __mem_cgroup_clear_mc();
5369 spin_lock(&mc.lock);
5372 spin_unlock(&mc.lock);
5373 mem_cgroup_end_move(from);
5376 static int mem_cgroup_can_attach(struct cgroup *cgroup,
5377 struct cgroup_taskset *tset)
5379 struct task_struct *p = cgroup_taskset_first(tset);
5381 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup);
5383 if (memcg->move_charge_at_immigrate) {
5384 struct mm_struct *mm;
5385 struct mem_cgroup *from = mem_cgroup_from_task(p);
5387 VM_BUG_ON(from == memcg);
5389 mm = get_task_mm(p);
5392 /* We move charges only when we move a owner of the mm */
5393 if (mm->owner == p) {
5396 VM_BUG_ON(mc.precharge);
5397 VM_BUG_ON(mc.moved_charge);
5398 VM_BUG_ON(mc.moved_swap);
5399 mem_cgroup_start_move(from);
5400 spin_lock(&mc.lock);
5403 spin_unlock(&mc.lock);
5404 /* We set mc.moving_task later */
5406 ret = mem_cgroup_precharge_mc(mm);
5408 mem_cgroup_clear_mc();
5415 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
5416 struct cgroup_taskset *tset)
5418 mem_cgroup_clear_mc();
5421 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5422 unsigned long addr, unsigned long end,
5423 struct mm_walk *walk)
5426 struct vm_area_struct *vma = walk->private;
5429 enum mc_target_type target_type;
5430 union mc_target target;
5432 struct page_cgroup *pc;
5435 * We don't take compound_lock() here but no race with splitting thp
5437 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
5438 * under splitting, which means there's no concurrent thp split,
5439 * - if another thread runs into split_huge_page() just after we
5440 * entered this if-block, the thread must wait for page table lock
5441 * to be unlocked in __split_huge_page_splitting(), where the main
5442 * part of thp split is not executed yet.
5444 if (pmd_trans_huge_lock(pmd, vma) == 1) {
5445 if (mc.precharge < HPAGE_PMD_NR) {
5446 spin_unlock(&vma->vm_mm->page_table_lock);
5449 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5450 if (target_type == MC_TARGET_PAGE) {
5452 if (!isolate_lru_page(page)) {
5453 pc = lookup_page_cgroup(page);
5454 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
5455 pc, mc.from, mc.to)) {
5456 mc.precharge -= HPAGE_PMD_NR;
5457 mc.moved_charge += HPAGE_PMD_NR;
5459 putback_lru_page(page);
5463 spin_unlock(&vma->vm_mm->page_table_lock);
5467 if (pmd_trans_unstable(pmd))
5470 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5471 for (; addr != end; addr += PAGE_SIZE) {
5472 pte_t ptent = *(pte++);
5478 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5479 case MC_TARGET_PAGE:
5481 if (isolate_lru_page(page))
5483 pc = lookup_page_cgroup(page);
5484 if (!mem_cgroup_move_account(page, 1, pc,
5487 /* we uncharge from mc.from later. */
5490 putback_lru_page(page);
5491 put: /* get_mctgt_type() gets the page */
5494 case MC_TARGET_SWAP:
5496 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5498 /* we fixup refcnts and charges later. */
5506 pte_unmap_unlock(pte - 1, ptl);
5511 * We have consumed all precharges we got in can_attach().
5512 * We try charge one by one, but don't do any additional
5513 * charges to mc.to if we have failed in charge once in attach()
5516 ret = mem_cgroup_do_precharge(1);
5524 static void mem_cgroup_move_charge(struct mm_struct *mm)
5526 struct vm_area_struct *vma;
5528 lru_add_drain_all();
5530 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5532 * Someone who are holding the mmap_sem might be waiting in
5533 * waitq. So we cancel all extra charges, wake up all waiters,
5534 * and retry. Because we cancel precharges, we might not be able
5535 * to move enough charges, but moving charge is a best-effort
5536 * feature anyway, so it wouldn't be a big problem.
5538 __mem_cgroup_clear_mc();
5542 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5544 struct mm_walk mem_cgroup_move_charge_walk = {
5545 .pmd_entry = mem_cgroup_move_charge_pte_range,
5549 if (is_vm_hugetlb_page(vma))
5551 ret = walk_page_range(vma->vm_start, vma->vm_end,
5552 &mem_cgroup_move_charge_walk);
5555 * means we have consumed all precharges and failed in
5556 * doing additional charge. Just abandon here.
5560 up_read(&mm->mmap_sem);
5563 static void mem_cgroup_move_task(struct cgroup *cont,
5564 struct cgroup_taskset *tset)
5566 struct task_struct *p = cgroup_taskset_first(tset);
5567 struct mm_struct *mm = get_task_mm(p);
5571 mem_cgroup_move_charge(mm);
5575 mem_cgroup_clear_mc();
5577 #else /* !CONFIG_MMU */
5578 static int mem_cgroup_can_attach(struct cgroup *cgroup,
5579 struct cgroup_taskset *tset)
5583 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
5584 struct cgroup_taskset *tset)
5587 static void mem_cgroup_move_task(struct cgroup *cont,
5588 struct cgroup_taskset *tset)
5593 struct cgroup_subsys mem_cgroup_subsys = {
5595 .subsys_id = mem_cgroup_subsys_id,
5596 .create = mem_cgroup_create,
5597 .pre_destroy = mem_cgroup_pre_destroy,
5598 .destroy = mem_cgroup_destroy,
5599 .can_attach = mem_cgroup_can_attach,
5600 .cancel_attach = mem_cgroup_cancel_attach,
5601 .attach = mem_cgroup_move_task,
5602 .base_cftypes = mem_cgroup_files,
5605 .__DEPRECATED_clear_css_refs = true,
5608 #ifdef CONFIG_MEMCG_SWAP
5609 static int __init enable_swap_account(char *s)
5611 /* consider enabled if no parameter or 1 is given */
5612 if (!strcmp(s, "1"))
5613 really_do_swap_account = 1;
5614 else if (!strcmp(s, "0"))
5615 really_do_swap_account = 0;
5618 __setup("swapaccount=", enable_swap_account);