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 struct mem_cgroup *root_mem_cgroup __read_mostly;
64 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_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_CGROUP_MEM_RES_CTLR_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_SWAPOUT, /* # of pages, swapped out */
91 MEM_CGROUP_STAT_DATA, /* end of data requires synchronization */
92 MEM_CGROUP_ON_MOVE, /* someone is moving account between groups */
93 MEM_CGROUP_STAT_NSTATS,
96 enum mem_cgroup_events_index {
97 MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */
98 MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */
99 MEM_CGROUP_EVENTS_COUNT, /* # of pages paged in/out */
100 MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */
101 MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */
102 MEM_CGROUP_EVENTS_NSTATS,
105 * Per memcg event counter is incremented at every pagein/pageout. With THP,
106 * it will be incremated by the number of pages. This counter is used for
107 * for trigger some periodic events. This is straightforward and better
108 * than using jiffies etc. to handle periodic memcg event.
110 enum mem_cgroup_events_target {
111 MEM_CGROUP_TARGET_THRESH,
112 MEM_CGROUP_TARGET_SOFTLIMIT,
113 MEM_CGROUP_TARGET_NUMAINFO,
116 #define THRESHOLDS_EVENTS_TARGET (128)
117 #define SOFTLIMIT_EVENTS_TARGET (1024)
118 #define NUMAINFO_EVENTS_TARGET (1024)
120 struct mem_cgroup_stat_cpu {
121 long count[MEM_CGROUP_STAT_NSTATS];
122 unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
123 unsigned long targets[MEM_CGROUP_NTARGETS];
126 struct mem_cgroup_reclaim_iter {
127 /* css_id of the last scanned hierarchy member */
129 /* scan generation, increased every round-trip */
130 unsigned int generation;
134 * per-zone information in memory controller.
136 struct mem_cgroup_per_zone {
137 struct lruvec lruvec;
138 unsigned long lru_size[NR_LRU_LISTS];
140 struct mem_cgroup_reclaim_iter reclaim_iter[DEF_PRIORITY + 1];
142 struct zone_reclaim_stat reclaim_stat;
143 struct rb_node tree_node; /* RB tree node */
144 unsigned long long usage_in_excess;/* Set to the value by which */
145 /* the soft limit is exceeded*/
147 struct mem_cgroup *memcg; /* Back pointer, we cannot */
148 /* use container_of */
151 struct mem_cgroup_per_node {
152 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
155 struct mem_cgroup_lru_info {
156 struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
160 * Cgroups above their limits are maintained in a RB-Tree, independent of
161 * their hierarchy representation
164 struct mem_cgroup_tree_per_zone {
165 struct rb_root rb_root;
169 struct mem_cgroup_tree_per_node {
170 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
173 struct mem_cgroup_tree {
174 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
177 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
179 struct mem_cgroup_threshold {
180 struct eventfd_ctx *eventfd;
185 struct mem_cgroup_threshold_ary {
186 /* An array index points to threshold just below usage. */
187 int current_threshold;
188 /* Size of entries[] */
190 /* Array of thresholds */
191 struct mem_cgroup_threshold entries[0];
194 struct mem_cgroup_thresholds {
195 /* Primary thresholds array */
196 struct mem_cgroup_threshold_ary *primary;
198 * Spare threshold array.
199 * This is needed to make mem_cgroup_unregister_event() "never fail".
200 * It must be able to store at least primary->size - 1 entries.
202 struct mem_cgroup_threshold_ary *spare;
206 struct mem_cgroup_eventfd_list {
207 struct list_head list;
208 struct eventfd_ctx *eventfd;
211 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
212 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
215 * The memory controller data structure. The memory controller controls both
216 * page cache and RSS per cgroup. We would eventually like to provide
217 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
218 * to help the administrator determine what knobs to tune.
220 * TODO: Add a water mark for the memory controller. Reclaim will begin when
221 * we hit the water mark. May be even add a low water mark, such that
222 * no reclaim occurs from a cgroup at it's low water mark, this is
223 * a feature that will be implemented much later in the future.
226 struct cgroup_subsys_state css;
228 * the counter to account for memory usage
230 struct res_counter res;
234 * the counter to account for mem+swap usage.
236 struct res_counter memsw;
239 * rcu_freeing is used only when freeing struct mem_cgroup,
240 * so put it into a union to avoid wasting more memory.
241 * It must be disjoint from the css field. It could be
242 * in a union with the res field, but res plays a much
243 * larger part in mem_cgroup life than memsw, and might
244 * be of interest, even at time of free, when debugging.
245 * So share rcu_head with the less interesting memsw.
247 struct rcu_head rcu_freeing;
249 * But when using vfree(), that cannot be done at
250 * interrupt time, so we must then queue the work.
252 struct work_struct work_freeing;
256 * Per cgroup active and inactive list, similar to the
257 * per zone LRU lists.
259 struct mem_cgroup_lru_info info;
260 int last_scanned_node;
262 nodemask_t scan_nodes;
263 atomic_t numainfo_events;
264 atomic_t numainfo_updating;
267 * Should the accounting and control be hierarchical, per subtree?
277 /* OOM-Killer disable */
278 int oom_kill_disable;
280 /* set when res.limit == memsw.limit */
281 bool memsw_is_minimum;
283 /* protect arrays of thresholds */
284 struct mutex thresholds_lock;
286 /* thresholds for memory usage. RCU-protected */
287 struct mem_cgroup_thresholds thresholds;
289 /* thresholds for mem+swap usage. RCU-protected */
290 struct mem_cgroup_thresholds memsw_thresholds;
292 /* For oom notifier event fd */
293 struct list_head oom_notify;
296 * Should we move charges of a task when a task is moved into this
297 * mem_cgroup ? And what type of charges should we move ?
299 unsigned long move_charge_at_immigrate;
303 struct mem_cgroup_stat_cpu *stat;
305 * used when a cpu is offlined or other synchronizations
306 * See mem_cgroup_read_stat().
308 struct mem_cgroup_stat_cpu nocpu_base;
309 spinlock_t pcp_counter_lock;
312 struct tcp_memcontrol tcp_mem;
316 /* Stuffs for move charges at task migration. */
318 * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
319 * left-shifted bitmap of these types.
322 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
323 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
327 /* "mc" and its members are protected by cgroup_mutex */
328 static struct move_charge_struct {
329 spinlock_t lock; /* for from, to */
330 struct mem_cgroup *from;
331 struct mem_cgroup *to;
332 unsigned long precharge;
333 unsigned long moved_charge;
334 unsigned long moved_swap;
335 struct task_struct *moving_task; /* a task moving charges */
336 wait_queue_head_t waitq; /* a waitq for other context */
338 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
339 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
342 static bool move_anon(void)
344 return test_bit(MOVE_CHARGE_TYPE_ANON,
345 &mc.to->move_charge_at_immigrate);
348 static bool move_file(void)
350 return test_bit(MOVE_CHARGE_TYPE_FILE,
351 &mc.to->move_charge_at_immigrate);
355 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
356 * limit reclaim to prevent infinite loops, if they ever occur.
358 #define MEM_CGROUP_MAX_RECLAIM_LOOPS (100)
359 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2)
362 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
363 MEM_CGROUP_CHARGE_TYPE_MAPPED,
364 MEM_CGROUP_CHARGE_TYPE_SHMEM, /* used by page migration of shmem */
365 MEM_CGROUP_CHARGE_TYPE_FORCE, /* used by force_empty */
366 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
367 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
371 /* for encoding cft->private value on file */
374 #define _OOM_TYPE (2)
375 #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
376 #define MEMFILE_TYPE(val) (((val) >> 16) & 0xffff)
377 #define MEMFILE_ATTR(val) ((val) & 0xffff)
378 /* Used for OOM nofiier */
379 #define OOM_CONTROL (0)
382 * Reclaim flags for mem_cgroup_hierarchical_reclaim
384 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
385 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
386 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
387 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
389 static void mem_cgroup_get(struct mem_cgroup *memcg);
390 static void mem_cgroup_put(struct mem_cgroup *memcg);
392 /* Writing them here to avoid exposing memcg's inner layout */
393 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_KMEM
394 #include <net/sock.h>
397 static bool mem_cgroup_is_root(struct mem_cgroup *memcg);
398 void sock_update_memcg(struct sock *sk)
400 if (mem_cgroup_sockets_enabled) {
401 struct mem_cgroup *memcg;
403 BUG_ON(!sk->sk_prot->proto_cgroup);
405 /* Socket cloning can throw us here with sk_cgrp already
406 * filled. It won't however, necessarily happen from
407 * process context. So the test for root memcg given
408 * the current task's memcg won't help us in this case.
410 * Respecting the original socket's memcg is a better
411 * decision in this case.
414 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
415 mem_cgroup_get(sk->sk_cgrp->memcg);
420 memcg = mem_cgroup_from_task(current);
421 if (!mem_cgroup_is_root(memcg)) {
422 mem_cgroup_get(memcg);
423 sk->sk_cgrp = sk->sk_prot->proto_cgroup(memcg);
428 EXPORT_SYMBOL(sock_update_memcg);
430 void sock_release_memcg(struct sock *sk)
432 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
433 struct mem_cgroup *memcg;
434 WARN_ON(!sk->sk_cgrp->memcg);
435 memcg = sk->sk_cgrp->memcg;
436 mem_cgroup_put(memcg);
441 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
443 if (!memcg || mem_cgroup_is_root(memcg))
446 return &memcg->tcp_mem.cg_proto;
448 EXPORT_SYMBOL(tcp_proto_cgroup);
449 #endif /* CONFIG_INET */
450 #endif /* CONFIG_CGROUP_MEM_RES_CTLR_KMEM */
452 static void drain_all_stock_async(struct mem_cgroup *memcg);
454 static struct mem_cgroup_per_zone *
455 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
457 return &memcg->info.nodeinfo[nid]->zoneinfo[zid];
460 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
465 static struct mem_cgroup_per_zone *
466 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
468 int nid = page_to_nid(page);
469 int zid = page_zonenum(page);
471 return mem_cgroup_zoneinfo(memcg, nid, zid);
474 static struct mem_cgroup_tree_per_zone *
475 soft_limit_tree_node_zone(int nid, int zid)
477 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
480 static struct mem_cgroup_tree_per_zone *
481 soft_limit_tree_from_page(struct page *page)
483 int nid = page_to_nid(page);
484 int zid = page_zonenum(page);
486 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
490 __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
491 struct mem_cgroup_per_zone *mz,
492 struct mem_cgroup_tree_per_zone *mctz,
493 unsigned long long new_usage_in_excess)
495 struct rb_node **p = &mctz->rb_root.rb_node;
496 struct rb_node *parent = NULL;
497 struct mem_cgroup_per_zone *mz_node;
502 mz->usage_in_excess = new_usage_in_excess;
503 if (!mz->usage_in_excess)
507 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
509 if (mz->usage_in_excess < mz_node->usage_in_excess)
512 * We can't avoid mem cgroups that are over their soft
513 * limit by the same amount
515 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
518 rb_link_node(&mz->tree_node, parent, p);
519 rb_insert_color(&mz->tree_node, &mctz->rb_root);
524 __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
525 struct mem_cgroup_per_zone *mz,
526 struct mem_cgroup_tree_per_zone *mctz)
530 rb_erase(&mz->tree_node, &mctz->rb_root);
535 mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
536 struct mem_cgroup_per_zone *mz,
537 struct mem_cgroup_tree_per_zone *mctz)
539 spin_lock(&mctz->lock);
540 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
541 spin_unlock(&mctz->lock);
545 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
547 unsigned long long excess;
548 struct mem_cgroup_per_zone *mz;
549 struct mem_cgroup_tree_per_zone *mctz;
550 int nid = page_to_nid(page);
551 int zid = page_zonenum(page);
552 mctz = soft_limit_tree_from_page(page);
555 * Necessary to update all ancestors when hierarchy is used.
556 * because their event counter is not touched.
558 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
559 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
560 excess = res_counter_soft_limit_excess(&memcg->res);
562 * We have to update the tree if mz is on RB-tree or
563 * mem is over its softlimit.
565 if (excess || mz->on_tree) {
566 spin_lock(&mctz->lock);
567 /* if on-tree, remove it */
569 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
571 * Insert again. mz->usage_in_excess will be updated.
572 * If excess is 0, no tree ops.
574 __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
575 spin_unlock(&mctz->lock);
580 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
583 struct mem_cgroup_per_zone *mz;
584 struct mem_cgroup_tree_per_zone *mctz;
586 for_each_node(node) {
587 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
588 mz = mem_cgroup_zoneinfo(memcg, node, zone);
589 mctz = soft_limit_tree_node_zone(node, zone);
590 mem_cgroup_remove_exceeded(memcg, mz, mctz);
595 static struct mem_cgroup_per_zone *
596 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
598 struct rb_node *rightmost = NULL;
599 struct mem_cgroup_per_zone *mz;
603 rightmost = rb_last(&mctz->rb_root);
605 goto done; /* Nothing to reclaim from */
607 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
609 * Remove the node now but someone else can add it back,
610 * we will to add it back at the end of reclaim to its correct
611 * position in the tree.
613 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
614 if (!res_counter_soft_limit_excess(&mz->memcg->res) ||
615 !css_tryget(&mz->memcg->css))
621 static struct mem_cgroup_per_zone *
622 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
624 struct mem_cgroup_per_zone *mz;
626 spin_lock(&mctz->lock);
627 mz = __mem_cgroup_largest_soft_limit_node(mctz);
628 spin_unlock(&mctz->lock);
633 * Implementation Note: reading percpu statistics for memcg.
635 * Both of vmstat[] and percpu_counter has threshold and do periodic
636 * synchronization to implement "quick" read. There are trade-off between
637 * reading cost and precision of value. Then, we may have a chance to implement
638 * a periodic synchronizion of counter in memcg's counter.
640 * But this _read() function is used for user interface now. The user accounts
641 * memory usage by memory cgroup and he _always_ requires exact value because
642 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
643 * have to visit all online cpus and make sum. So, for now, unnecessary
644 * synchronization is not implemented. (just implemented for cpu hotplug)
646 * If there are kernel internal actions which can make use of some not-exact
647 * value, and reading all cpu value can be performance bottleneck in some
648 * common workload, threashold and synchonization as vmstat[] should be
651 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
652 enum mem_cgroup_stat_index idx)
658 for_each_online_cpu(cpu)
659 val += per_cpu(memcg->stat->count[idx], cpu);
660 #ifdef CONFIG_HOTPLUG_CPU
661 spin_lock(&memcg->pcp_counter_lock);
662 val += memcg->nocpu_base.count[idx];
663 spin_unlock(&memcg->pcp_counter_lock);
669 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
672 int val = (charge) ? 1 : -1;
673 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAPOUT], val);
676 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
677 enum mem_cgroup_events_index idx)
679 unsigned long val = 0;
682 for_each_online_cpu(cpu)
683 val += per_cpu(memcg->stat->events[idx], cpu);
684 #ifdef CONFIG_HOTPLUG_CPU
685 spin_lock(&memcg->pcp_counter_lock);
686 val += memcg->nocpu_base.events[idx];
687 spin_unlock(&memcg->pcp_counter_lock);
692 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
693 bool file, int nr_pages)
698 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
701 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
704 /* pagein of a big page is an event. So, ignore page size */
706 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
708 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
709 nr_pages = -nr_pages; /* for event */
712 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT], nr_pages);
718 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
719 unsigned int lru_mask)
721 struct mem_cgroup_per_zone *mz;
723 unsigned long ret = 0;
725 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
728 if (BIT(lru) & lru_mask)
729 ret += mz->lru_size[lru];
735 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
736 int nid, unsigned int lru_mask)
741 for (zid = 0; zid < MAX_NR_ZONES; zid++)
742 total += mem_cgroup_zone_nr_lru_pages(memcg,
748 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
749 unsigned int lru_mask)
754 for_each_node_state(nid, N_HIGH_MEMORY)
755 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
759 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
760 enum mem_cgroup_events_target target)
762 unsigned long val, next;
764 val = __this_cpu_read(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT]);
765 next = __this_cpu_read(memcg->stat->targets[target]);
766 /* from time_after() in jiffies.h */
767 if ((long)next - (long)val < 0) {
769 case MEM_CGROUP_TARGET_THRESH:
770 next = val + THRESHOLDS_EVENTS_TARGET;
772 case MEM_CGROUP_TARGET_SOFTLIMIT:
773 next = val + SOFTLIMIT_EVENTS_TARGET;
775 case MEM_CGROUP_TARGET_NUMAINFO:
776 next = val + NUMAINFO_EVENTS_TARGET;
781 __this_cpu_write(memcg->stat->targets[target], next);
788 * Check events in order.
791 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
794 /* threshold event is triggered in finer grain than soft limit */
795 if (unlikely(mem_cgroup_event_ratelimit(memcg,
796 MEM_CGROUP_TARGET_THRESH))) {
798 bool do_numainfo __maybe_unused;
800 do_softlimit = mem_cgroup_event_ratelimit(memcg,
801 MEM_CGROUP_TARGET_SOFTLIMIT);
803 do_numainfo = mem_cgroup_event_ratelimit(memcg,
804 MEM_CGROUP_TARGET_NUMAINFO);
808 mem_cgroup_threshold(memcg);
809 if (unlikely(do_softlimit))
810 mem_cgroup_update_tree(memcg, page);
812 if (unlikely(do_numainfo))
813 atomic_inc(&memcg->numainfo_events);
819 struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
821 return container_of(cgroup_subsys_state(cont,
822 mem_cgroup_subsys_id), struct mem_cgroup,
826 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
829 * mm_update_next_owner() may clear mm->owner to NULL
830 * if it races with swapoff, page migration, etc.
831 * So this can be called with p == NULL.
836 return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
837 struct mem_cgroup, css);
840 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
842 struct mem_cgroup *memcg = NULL;
847 * Because we have no locks, mm->owner's may be being moved to other
848 * cgroup. We use css_tryget() here even if this looks
849 * pessimistic (rather than adding locks here).
853 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
854 if (unlikely(!memcg))
856 } while (!css_tryget(&memcg->css));
862 * mem_cgroup_iter - iterate over memory cgroup hierarchy
863 * @root: hierarchy root
864 * @prev: previously returned memcg, NULL on first invocation
865 * @reclaim: cookie for shared reclaim walks, NULL for full walks
867 * Returns references to children of the hierarchy below @root, or
868 * @root itself, or %NULL after a full round-trip.
870 * Caller must pass the return value in @prev on subsequent
871 * invocations for reference counting, or use mem_cgroup_iter_break()
872 * to cancel a hierarchy walk before the round-trip is complete.
874 * Reclaimers can specify a zone and a priority level in @reclaim to
875 * divide up the memcgs in the hierarchy among all concurrent
876 * reclaimers operating on the same zone and priority.
878 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
879 struct mem_cgroup *prev,
880 struct mem_cgroup_reclaim_cookie *reclaim)
882 struct mem_cgroup *memcg = NULL;
885 if (mem_cgroup_disabled())
889 root = root_mem_cgroup;
891 if (prev && !reclaim)
892 id = css_id(&prev->css);
894 if (prev && prev != root)
897 if (!root->use_hierarchy && root != root_mem_cgroup) {
904 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
905 struct cgroup_subsys_state *css;
908 int nid = zone_to_nid(reclaim->zone);
909 int zid = zone_idx(reclaim->zone);
910 struct mem_cgroup_per_zone *mz;
912 mz = mem_cgroup_zoneinfo(root, nid, zid);
913 iter = &mz->reclaim_iter[reclaim->priority];
914 if (prev && reclaim->generation != iter->generation)
920 css = css_get_next(&mem_cgroup_subsys, id + 1, &root->css, &id);
922 if (css == &root->css || css_tryget(css))
923 memcg = container_of(css,
924 struct mem_cgroup, css);
933 else if (!prev && memcg)
934 reclaim->generation = iter->generation;
944 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
945 * @root: hierarchy root
946 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
948 void mem_cgroup_iter_break(struct mem_cgroup *root,
949 struct mem_cgroup *prev)
952 root = root_mem_cgroup;
953 if (prev && prev != root)
958 * Iteration constructs for visiting all cgroups (under a tree). If
959 * loops are exited prematurely (break), mem_cgroup_iter_break() must
960 * be used for reference counting.
962 #define for_each_mem_cgroup_tree(iter, root) \
963 for (iter = mem_cgroup_iter(root, NULL, NULL); \
965 iter = mem_cgroup_iter(root, iter, NULL))
967 #define for_each_mem_cgroup(iter) \
968 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
970 iter = mem_cgroup_iter(NULL, iter, NULL))
972 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
974 return (memcg == root_mem_cgroup);
977 void mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
979 struct mem_cgroup *memcg;
985 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
986 if (unlikely(!memcg))
991 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
994 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1002 EXPORT_SYMBOL(mem_cgroup_count_vm_event);
1005 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1006 * @zone: zone of the wanted lruvec
1007 * @mem: memcg of the wanted lruvec
1009 * Returns the lru list vector holding pages for the given @zone and
1010 * @mem. This can be the global zone lruvec, if the memory controller
1013 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1014 struct mem_cgroup *memcg)
1016 struct mem_cgroup_per_zone *mz;
1018 if (mem_cgroup_disabled())
1019 return &zone->lruvec;
1021 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1026 * Following LRU functions are allowed to be used without PCG_LOCK.
1027 * Operations are called by routine of global LRU independently from memcg.
1028 * What we have to take care of here is validness of pc->mem_cgroup.
1030 * Changes to pc->mem_cgroup happens when
1033 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1034 * It is added to LRU before charge.
1035 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1036 * When moving account, the page is not on LRU. It's isolated.
1040 * mem_cgroup_lru_add_list - account for adding an lru page and return lruvec
1041 * @zone: zone of the page
1045 * This function accounts for @page being added to @lru, and returns
1046 * the lruvec for the given @zone and the memcg @page is charged to.
1048 * The callsite is then responsible for physically linking the page to
1049 * the returned lruvec->lists[@lru].
1051 struct lruvec *mem_cgroup_lru_add_list(struct zone *zone, struct page *page,
1054 struct mem_cgroup_per_zone *mz;
1055 struct mem_cgroup *memcg;
1056 struct page_cgroup *pc;
1058 if (mem_cgroup_disabled())
1059 return &zone->lruvec;
1061 pc = lookup_page_cgroup(page);
1062 memcg = pc->mem_cgroup;
1065 * Surreptitiously switch any uncharged page to root:
1066 * an uncharged page off lru does nothing to secure
1067 * its former mem_cgroup from sudden removal.
1069 * Our caller holds lru_lock, and PageCgroupUsed is updated
1070 * under page_cgroup lock: between them, they make all uses
1071 * of pc->mem_cgroup safe.
1073 if (!PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1074 pc->mem_cgroup = memcg = root_mem_cgroup;
1076 mz = page_cgroup_zoneinfo(memcg, page);
1077 /* compound_order() is stabilized through lru_lock */
1078 mz->lru_size[lru] += 1 << compound_order(page);
1083 * mem_cgroup_lru_del_list - account for removing an lru page
1087 * This function accounts for @page being removed from @lru.
1089 * The callsite is then responsible for physically unlinking
1092 void mem_cgroup_lru_del_list(struct page *page, enum lru_list lru)
1094 struct mem_cgroup_per_zone *mz;
1095 struct mem_cgroup *memcg;
1096 struct page_cgroup *pc;
1098 if (mem_cgroup_disabled())
1101 pc = lookup_page_cgroup(page);
1102 memcg = pc->mem_cgroup;
1104 mz = page_cgroup_zoneinfo(memcg, page);
1105 /* huge page split is done under lru_lock. so, we have no races. */
1106 VM_BUG_ON(mz->lru_size[lru] < (1 << compound_order(page)));
1107 mz->lru_size[lru] -= 1 << compound_order(page);
1110 void mem_cgroup_lru_del(struct page *page)
1112 mem_cgroup_lru_del_list(page, page_lru(page));
1116 * mem_cgroup_lru_move_lists - account for moving a page between lrus
1117 * @zone: zone of the page
1119 * @from: current lru
1122 * This function accounts for @page being moved between the lrus @from
1123 * and @to, and returns the lruvec for the given @zone and the memcg
1124 * @page is charged to.
1126 * The callsite is then responsible for physically relinking
1127 * @page->lru to the returned lruvec->lists[@to].
1129 struct lruvec *mem_cgroup_lru_move_lists(struct zone *zone,
1134 /* XXX: Optimize this, especially for @from == @to */
1135 mem_cgroup_lru_del_list(page, from);
1136 return mem_cgroup_lru_add_list(zone, page, to);
1140 * Checks whether given mem is same or in the root_mem_cgroup's
1143 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1144 struct mem_cgroup *memcg)
1146 if (root_memcg != memcg) {
1147 return (root_memcg->use_hierarchy &&
1148 css_is_ancestor(&memcg->css, &root_memcg->css));
1154 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *memcg)
1157 struct mem_cgroup *curr = NULL;
1158 struct task_struct *p;
1160 p = find_lock_task_mm(task);
1162 curr = try_get_mem_cgroup_from_mm(p->mm);
1166 * All threads may have already detached their mm's, but the oom
1167 * killer still needs to detect if they have already been oom
1168 * killed to prevent needlessly killing additional tasks.
1171 curr = mem_cgroup_from_task(task);
1173 css_get(&curr->css);
1179 * We should check use_hierarchy of "memcg" not "curr". Because checking
1180 * use_hierarchy of "curr" here make this function true if hierarchy is
1181 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1182 * hierarchy(even if use_hierarchy is disabled in "memcg").
1184 ret = mem_cgroup_same_or_subtree(memcg, curr);
1185 css_put(&curr->css);
1189 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg, struct zone *zone)
1191 unsigned long inactive_ratio;
1192 int nid = zone_to_nid(zone);
1193 int zid = zone_idx(zone);
1194 unsigned long inactive;
1195 unsigned long active;
1198 inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1199 BIT(LRU_INACTIVE_ANON));
1200 active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1201 BIT(LRU_ACTIVE_ANON));
1203 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1205 inactive_ratio = int_sqrt(10 * gb);
1209 return inactive * inactive_ratio < active;
1212 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg, struct zone *zone)
1214 unsigned long active;
1215 unsigned long inactive;
1216 int zid = zone_idx(zone);
1217 int nid = zone_to_nid(zone);
1219 inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1220 BIT(LRU_INACTIVE_FILE));
1221 active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1222 BIT(LRU_ACTIVE_FILE));
1224 return (active > inactive);
1227 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
1230 int nid = zone_to_nid(zone);
1231 int zid = zone_idx(zone);
1232 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
1234 return &mz->reclaim_stat;
1237 struct zone_reclaim_stat *
1238 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
1240 struct page_cgroup *pc;
1241 struct mem_cgroup_per_zone *mz;
1243 if (mem_cgroup_disabled())
1246 pc = lookup_page_cgroup(page);
1247 if (!PageCgroupUsed(pc))
1249 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1251 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1252 return &mz->reclaim_stat;
1255 #define mem_cgroup_from_res_counter(counter, member) \
1256 container_of(counter, struct mem_cgroup, member)
1259 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1260 * @mem: the memory cgroup
1262 * Returns the maximum amount of memory @mem can be charged with, in
1265 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1267 unsigned long long margin;
1269 margin = res_counter_margin(&memcg->res);
1270 if (do_swap_account)
1271 margin = min(margin, res_counter_margin(&memcg->memsw));
1272 return margin >> PAGE_SHIFT;
1275 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1277 struct cgroup *cgrp = memcg->css.cgroup;
1280 if (cgrp->parent == NULL)
1281 return vm_swappiness;
1283 return memcg->swappiness;
1286 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1291 spin_lock(&memcg->pcp_counter_lock);
1292 for_each_online_cpu(cpu)
1293 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) += 1;
1294 memcg->nocpu_base.count[MEM_CGROUP_ON_MOVE] += 1;
1295 spin_unlock(&memcg->pcp_counter_lock);
1301 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1308 spin_lock(&memcg->pcp_counter_lock);
1309 for_each_online_cpu(cpu)
1310 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) -= 1;
1311 memcg->nocpu_base.count[MEM_CGROUP_ON_MOVE] -= 1;
1312 spin_unlock(&memcg->pcp_counter_lock);
1316 * 2 routines for checking "mem" is under move_account() or not.
1318 * mem_cgroup_stealed() - checking a cgroup is mc.from or not. This is used
1319 * for avoiding race in accounting. If true,
1320 * pc->mem_cgroup may be overwritten.
1322 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1323 * under hierarchy of moving cgroups. This is for
1324 * waiting at hith-memory prressure caused by "move".
1327 static bool mem_cgroup_stealed(struct mem_cgroup *memcg)
1329 VM_BUG_ON(!rcu_read_lock_held());
1330 return this_cpu_read(memcg->stat->count[MEM_CGROUP_ON_MOVE]) > 0;
1333 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1335 struct mem_cgroup *from;
1336 struct mem_cgroup *to;
1339 * Unlike task_move routines, we access mc.to, mc.from not under
1340 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1342 spin_lock(&mc.lock);
1348 ret = mem_cgroup_same_or_subtree(memcg, from)
1349 || mem_cgroup_same_or_subtree(memcg, to);
1351 spin_unlock(&mc.lock);
1355 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1357 if (mc.moving_task && current != mc.moving_task) {
1358 if (mem_cgroup_under_move(memcg)) {
1360 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1361 /* moving charge context might have finished. */
1364 finish_wait(&mc.waitq, &wait);
1372 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1373 * @memcg: The memory cgroup that went over limit
1374 * @p: Task that is going to be killed
1376 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1379 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1381 struct cgroup *task_cgrp;
1382 struct cgroup *mem_cgrp;
1384 * Need a buffer in BSS, can't rely on allocations. The code relies
1385 * on the assumption that OOM is serialized for memory controller.
1386 * If this assumption is broken, revisit this code.
1388 static char memcg_name[PATH_MAX];
1396 mem_cgrp = memcg->css.cgroup;
1397 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1399 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1402 * Unfortunately, we are unable to convert to a useful name
1403 * But we'll still print out the usage information
1410 printk(KERN_INFO "Task in %s killed", memcg_name);
1413 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1421 * Continues from above, so we don't need an KERN_ level
1423 printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1426 printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1427 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1428 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1429 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1430 printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1432 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1433 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1434 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1438 * This function returns the number of memcg under hierarchy tree. Returns
1439 * 1(self count) if no children.
1441 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1444 struct mem_cgroup *iter;
1446 for_each_mem_cgroup_tree(iter, memcg)
1452 * Return the memory (and swap, if configured) limit for a memcg.
1454 u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1459 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1460 limit += total_swap_pages << PAGE_SHIFT;
1462 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1464 * If memsw is finite and limits the amount of swap space available
1465 * to this memcg, return that limit.
1467 return min(limit, memsw);
1470 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1472 unsigned long flags)
1474 unsigned long total = 0;
1475 bool noswap = false;
1478 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1480 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1483 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1485 drain_all_stock_async(memcg);
1486 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1488 * Allow limit shrinkers, which are triggered directly
1489 * by userspace, to catch signals and stop reclaim
1490 * after minimal progress, regardless of the margin.
1492 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1494 if (mem_cgroup_margin(memcg))
1497 * If nothing was reclaimed after two attempts, there
1498 * may be no reclaimable pages in this hierarchy.
1507 * test_mem_cgroup_node_reclaimable
1508 * @mem: the target memcg
1509 * @nid: the node ID to be checked.
1510 * @noswap : specify true here if the user wants flle only information.
1512 * This function returns whether the specified memcg contains any
1513 * reclaimable pages on a node. Returns true if there are any reclaimable
1514 * pages in the node.
1516 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1517 int nid, bool noswap)
1519 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1521 if (noswap || !total_swap_pages)
1523 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1528 #if MAX_NUMNODES > 1
1531 * Always updating the nodemask is not very good - even if we have an empty
1532 * list or the wrong list here, we can start from some node and traverse all
1533 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1536 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1540 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1541 * pagein/pageout changes since the last update.
1543 if (!atomic_read(&memcg->numainfo_events))
1545 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1548 /* make a nodemask where this memcg uses memory from */
1549 memcg->scan_nodes = node_states[N_HIGH_MEMORY];
1551 for_each_node_mask(nid, node_states[N_HIGH_MEMORY]) {
1553 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1554 node_clear(nid, memcg->scan_nodes);
1557 atomic_set(&memcg->numainfo_events, 0);
1558 atomic_set(&memcg->numainfo_updating, 0);
1562 * Selecting a node where we start reclaim from. Because what we need is just
1563 * reducing usage counter, start from anywhere is O,K. Considering
1564 * memory reclaim from current node, there are pros. and cons.
1566 * Freeing memory from current node means freeing memory from a node which
1567 * we'll use or we've used. So, it may make LRU bad. And if several threads
1568 * hit limits, it will see a contention on a node. But freeing from remote
1569 * node means more costs for memory reclaim because of memory latency.
1571 * Now, we use round-robin. Better algorithm is welcomed.
1573 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1577 mem_cgroup_may_update_nodemask(memcg);
1578 node = memcg->last_scanned_node;
1580 node = next_node(node, memcg->scan_nodes);
1581 if (node == MAX_NUMNODES)
1582 node = first_node(memcg->scan_nodes);
1584 * We call this when we hit limit, not when pages are added to LRU.
1585 * No LRU may hold pages because all pages are UNEVICTABLE or
1586 * memcg is too small and all pages are not on LRU. In that case,
1587 * we use curret node.
1589 if (unlikely(node == MAX_NUMNODES))
1590 node = numa_node_id();
1592 memcg->last_scanned_node = node;
1597 * Check all nodes whether it contains reclaimable pages or not.
1598 * For quick scan, we make use of scan_nodes. This will allow us to skip
1599 * unused nodes. But scan_nodes is lazily updated and may not cotain
1600 * enough new information. We need to do double check.
1602 bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1607 * quick check...making use of scan_node.
1608 * We can skip unused nodes.
1610 if (!nodes_empty(memcg->scan_nodes)) {
1611 for (nid = first_node(memcg->scan_nodes);
1613 nid = next_node(nid, memcg->scan_nodes)) {
1615 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1620 * Check rest of nodes.
1622 for_each_node_state(nid, N_HIGH_MEMORY) {
1623 if (node_isset(nid, memcg->scan_nodes))
1625 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1632 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1637 bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1639 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
1643 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1646 unsigned long *total_scanned)
1648 struct mem_cgroup *victim = NULL;
1651 unsigned long excess;
1652 unsigned long nr_scanned;
1653 struct mem_cgroup_reclaim_cookie reclaim = {
1658 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
1661 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1666 * If we have not been able to reclaim
1667 * anything, it might because there are
1668 * no reclaimable pages under this hierarchy
1673 * We want to do more targeted reclaim.
1674 * excess >> 2 is not to excessive so as to
1675 * reclaim too much, nor too less that we keep
1676 * coming back to reclaim from this cgroup
1678 if (total >= (excess >> 2) ||
1679 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1684 if (!mem_cgroup_reclaimable(victim, false))
1686 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1688 *total_scanned += nr_scanned;
1689 if (!res_counter_soft_limit_excess(&root_memcg->res))
1692 mem_cgroup_iter_break(root_memcg, victim);
1697 * Check OOM-Killer is already running under our hierarchy.
1698 * If someone is running, return false.
1699 * Has to be called with memcg_oom_lock
1701 static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg)
1703 struct mem_cgroup *iter, *failed = NULL;
1705 for_each_mem_cgroup_tree(iter, memcg) {
1706 if (iter->oom_lock) {
1708 * this subtree of our hierarchy is already locked
1709 * so we cannot give a lock.
1712 mem_cgroup_iter_break(memcg, iter);
1715 iter->oom_lock = true;
1722 * OK, we failed to lock the whole subtree so we have to clean up
1723 * what we set up to the failing subtree
1725 for_each_mem_cgroup_tree(iter, memcg) {
1726 if (iter == failed) {
1727 mem_cgroup_iter_break(memcg, iter);
1730 iter->oom_lock = false;
1736 * Has to be called with memcg_oom_lock
1738 static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1740 struct mem_cgroup *iter;
1742 for_each_mem_cgroup_tree(iter, memcg)
1743 iter->oom_lock = false;
1747 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1749 struct mem_cgroup *iter;
1751 for_each_mem_cgroup_tree(iter, memcg)
1752 atomic_inc(&iter->under_oom);
1755 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1757 struct mem_cgroup *iter;
1760 * When a new child is created while the hierarchy is under oom,
1761 * mem_cgroup_oom_lock() may not be called. We have to use
1762 * atomic_add_unless() here.
1764 for_each_mem_cgroup_tree(iter, memcg)
1765 atomic_add_unless(&iter->under_oom, -1, 0);
1768 static DEFINE_SPINLOCK(memcg_oom_lock);
1769 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1771 struct oom_wait_info {
1772 struct mem_cgroup *memcg;
1776 static int memcg_oom_wake_function(wait_queue_t *wait,
1777 unsigned mode, int sync, void *arg)
1779 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1780 struct mem_cgroup *oom_wait_memcg;
1781 struct oom_wait_info *oom_wait_info;
1783 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1784 oom_wait_memcg = oom_wait_info->memcg;
1787 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
1788 * Then we can use css_is_ancestor without taking care of RCU.
1790 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
1791 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
1793 return autoremove_wake_function(wait, mode, sync, arg);
1796 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
1798 /* for filtering, pass "memcg" as argument. */
1799 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1802 static void memcg_oom_recover(struct mem_cgroup *memcg)
1804 if (memcg && atomic_read(&memcg->under_oom))
1805 memcg_wakeup_oom(memcg);
1809 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1811 bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1813 struct oom_wait_info owait;
1814 bool locked, need_to_kill;
1816 owait.memcg = memcg;
1817 owait.wait.flags = 0;
1818 owait.wait.func = memcg_oom_wake_function;
1819 owait.wait.private = current;
1820 INIT_LIST_HEAD(&owait.wait.task_list);
1821 need_to_kill = true;
1822 mem_cgroup_mark_under_oom(memcg);
1824 /* At first, try to OOM lock hierarchy under memcg.*/
1825 spin_lock(&memcg_oom_lock);
1826 locked = mem_cgroup_oom_lock(memcg);
1828 * Even if signal_pending(), we can't quit charge() loop without
1829 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1830 * under OOM is always welcomed, use TASK_KILLABLE here.
1832 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1833 if (!locked || memcg->oom_kill_disable)
1834 need_to_kill = false;
1836 mem_cgroup_oom_notify(memcg);
1837 spin_unlock(&memcg_oom_lock);
1840 finish_wait(&memcg_oom_waitq, &owait.wait);
1841 mem_cgroup_out_of_memory(memcg, mask, order);
1844 finish_wait(&memcg_oom_waitq, &owait.wait);
1846 spin_lock(&memcg_oom_lock);
1848 mem_cgroup_oom_unlock(memcg);
1849 memcg_wakeup_oom(memcg);
1850 spin_unlock(&memcg_oom_lock);
1852 mem_cgroup_unmark_under_oom(memcg);
1854 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
1856 /* Give chance to dying process */
1857 schedule_timeout_uninterruptible(1);
1862 * Currently used to update mapped file statistics, but the routine can be
1863 * generalized to update other statistics as well.
1865 * Notes: Race condition
1867 * We usually use page_cgroup_lock() for accessing page_cgroup member but
1868 * it tends to be costly. But considering some conditions, we doesn't need
1869 * to do so _always_.
1871 * Considering "charge", lock_page_cgroup() is not required because all
1872 * file-stat operations happen after a page is attached to radix-tree. There
1873 * are no race with "charge".
1875 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
1876 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
1877 * if there are race with "uncharge". Statistics itself is properly handled
1880 * Considering "move", this is an only case we see a race. To make the race
1881 * small, we check MEM_CGROUP_ON_MOVE percpu value and detect there are
1882 * possibility of race condition. If there is, we take a lock.
1885 void mem_cgroup_update_page_stat(struct page *page,
1886 enum mem_cgroup_page_stat_item idx, int val)
1888 struct mem_cgroup *memcg;
1889 struct page_cgroup *pc = lookup_page_cgroup(page);
1890 bool need_unlock = false;
1891 unsigned long uninitialized_var(flags);
1893 if (mem_cgroup_disabled())
1897 memcg = pc->mem_cgroup;
1898 if (unlikely(!memcg || !PageCgroupUsed(pc)))
1900 /* pc->mem_cgroup is unstable ? */
1901 if (unlikely(mem_cgroup_stealed(memcg))) {
1902 /* take a lock against to access pc->mem_cgroup */
1903 move_lock_page_cgroup(pc, &flags);
1905 memcg = pc->mem_cgroup;
1906 if (!memcg || !PageCgroupUsed(pc))
1911 case MEMCG_NR_FILE_MAPPED:
1913 SetPageCgroupFileMapped(pc);
1914 else if (!page_mapped(page))
1915 ClearPageCgroupFileMapped(pc);
1916 idx = MEM_CGROUP_STAT_FILE_MAPPED;
1922 this_cpu_add(memcg->stat->count[idx], val);
1925 if (unlikely(need_unlock))
1926 move_unlock_page_cgroup(pc, &flags);
1929 EXPORT_SYMBOL(mem_cgroup_update_page_stat);
1932 * size of first charge trial. "32" comes from vmscan.c's magic value.
1933 * TODO: maybe necessary to use big numbers in big irons.
1935 #define CHARGE_BATCH 32U
1936 struct memcg_stock_pcp {
1937 struct mem_cgroup *cached; /* this never be root cgroup */
1938 unsigned int nr_pages;
1939 struct work_struct work;
1940 unsigned long flags;
1941 #define FLUSHING_CACHED_CHARGE (0)
1943 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1944 static DEFINE_MUTEX(percpu_charge_mutex);
1947 * Try to consume stocked charge on this cpu. If success, one page is consumed
1948 * from local stock and true is returned. If the stock is 0 or charges from a
1949 * cgroup which is not current target, returns false. This stock will be
1952 static bool consume_stock(struct mem_cgroup *memcg)
1954 struct memcg_stock_pcp *stock;
1957 stock = &get_cpu_var(memcg_stock);
1958 if (memcg == stock->cached && stock->nr_pages)
1960 else /* need to call res_counter_charge */
1962 put_cpu_var(memcg_stock);
1967 * Returns stocks cached in percpu to res_counter and reset cached information.
1969 static void drain_stock(struct memcg_stock_pcp *stock)
1971 struct mem_cgroup *old = stock->cached;
1973 if (stock->nr_pages) {
1974 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
1976 res_counter_uncharge(&old->res, bytes);
1977 if (do_swap_account)
1978 res_counter_uncharge(&old->memsw, bytes);
1979 stock->nr_pages = 0;
1981 stock->cached = NULL;
1985 * This must be called under preempt disabled or must be called by
1986 * a thread which is pinned to local cpu.
1988 static void drain_local_stock(struct work_struct *dummy)
1990 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
1992 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
1996 * Cache charges(val) which is from res_counter, to local per_cpu area.
1997 * This will be consumed by consume_stock() function, later.
1999 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2001 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2003 if (stock->cached != memcg) { /* reset if necessary */
2005 stock->cached = memcg;
2007 stock->nr_pages += nr_pages;
2008 put_cpu_var(memcg_stock);
2012 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2013 * of the hierarchy under it. sync flag says whether we should block
2014 * until the work is done.
2016 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2020 /* Notify other cpus that system-wide "drain" is running */
2023 for_each_online_cpu(cpu) {
2024 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2025 struct mem_cgroup *memcg;
2027 memcg = stock->cached;
2028 if (!memcg || !stock->nr_pages)
2030 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2032 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2034 drain_local_stock(&stock->work);
2036 schedule_work_on(cpu, &stock->work);
2044 for_each_online_cpu(cpu) {
2045 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2046 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2047 flush_work(&stock->work);
2054 * Tries to drain stocked charges in other cpus. This function is asynchronous
2055 * and just put a work per cpu for draining localy on each cpu. Caller can
2056 * expects some charges will be back to res_counter later but cannot wait for
2059 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2062 * If someone calls draining, avoid adding more kworker runs.
2064 if (!mutex_trylock(&percpu_charge_mutex))
2066 drain_all_stock(root_memcg, false);
2067 mutex_unlock(&percpu_charge_mutex);
2070 /* This is a synchronous drain interface. */
2071 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2073 /* called when force_empty is called */
2074 mutex_lock(&percpu_charge_mutex);
2075 drain_all_stock(root_memcg, true);
2076 mutex_unlock(&percpu_charge_mutex);
2080 * This function drains percpu counter value from DEAD cpu and
2081 * move it to local cpu. Note that this function can be preempted.
2083 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2087 spin_lock(&memcg->pcp_counter_lock);
2088 for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) {
2089 long x = per_cpu(memcg->stat->count[i], cpu);
2091 per_cpu(memcg->stat->count[i], cpu) = 0;
2092 memcg->nocpu_base.count[i] += x;
2094 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2095 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2097 per_cpu(memcg->stat->events[i], cpu) = 0;
2098 memcg->nocpu_base.events[i] += x;
2100 /* need to clear ON_MOVE value, works as a kind of lock. */
2101 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) = 0;
2102 spin_unlock(&memcg->pcp_counter_lock);
2105 static void synchronize_mem_cgroup_on_move(struct mem_cgroup *memcg, int cpu)
2107 int idx = MEM_CGROUP_ON_MOVE;
2109 spin_lock(&memcg->pcp_counter_lock);
2110 per_cpu(memcg->stat->count[idx], cpu) = memcg->nocpu_base.count[idx];
2111 spin_unlock(&memcg->pcp_counter_lock);
2114 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2115 unsigned long action,
2118 int cpu = (unsigned long)hcpu;
2119 struct memcg_stock_pcp *stock;
2120 struct mem_cgroup *iter;
2122 if ((action == CPU_ONLINE)) {
2123 for_each_mem_cgroup(iter)
2124 synchronize_mem_cgroup_on_move(iter, cpu);
2128 if ((action != CPU_DEAD) || action != CPU_DEAD_FROZEN)
2131 for_each_mem_cgroup(iter)
2132 mem_cgroup_drain_pcp_counter(iter, cpu);
2134 stock = &per_cpu(memcg_stock, cpu);
2140 /* See __mem_cgroup_try_charge() for details */
2142 CHARGE_OK, /* success */
2143 CHARGE_RETRY, /* need to retry but retry is not bad */
2144 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2145 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2146 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2149 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2150 unsigned int nr_pages, bool oom_check)
2152 unsigned long csize = nr_pages * PAGE_SIZE;
2153 struct mem_cgroup *mem_over_limit;
2154 struct res_counter *fail_res;
2155 unsigned long flags = 0;
2158 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2161 if (!do_swap_account)
2163 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2167 res_counter_uncharge(&memcg->res, csize);
2168 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2169 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2171 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2173 * nr_pages can be either a huge page (HPAGE_PMD_NR), a batch
2174 * of regular pages (CHARGE_BATCH), or a single regular page (1).
2176 * Never reclaim on behalf of optional batching, retry with a
2177 * single page instead.
2179 if (nr_pages == CHARGE_BATCH)
2180 return CHARGE_RETRY;
2182 if (!(gfp_mask & __GFP_WAIT))
2183 return CHARGE_WOULDBLOCK;
2185 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2186 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2187 return CHARGE_RETRY;
2189 * Even though the limit is exceeded at this point, reclaim
2190 * may have been able to free some pages. Retry the charge
2191 * before killing the task.
2193 * Only for regular pages, though: huge pages are rather
2194 * unlikely to succeed so close to the limit, and we fall back
2195 * to regular pages anyway in case of failure.
2197 if (nr_pages == 1 && ret)
2198 return CHARGE_RETRY;
2201 * At task move, charge accounts can be doubly counted. So, it's
2202 * better to wait until the end of task_move if something is going on.
2204 if (mem_cgroup_wait_acct_move(mem_over_limit))
2205 return CHARGE_RETRY;
2207 /* If we don't need to call oom-killer at el, return immediately */
2209 return CHARGE_NOMEM;
2211 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask, get_order(csize)))
2212 return CHARGE_OOM_DIE;
2214 return CHARGE_RETRY;
2218 * __mem_cgroup_try_charge() does
2219 * 1. detect memcg to be charged against from passed *mm and *ptr,
2220 * 2. update res_counter
2221 * 3. call memory reclaim if necessary.
2223 * In some special case, if the task is fatal, fatal_signal_pending() or
2224 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2225 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2226 * as possible without any hazards. 2: all pages should have a valid
2227 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2228 * pointer, that is treated as a charge to root_mem_cgroup.
2230 * So __mem_cgroup_try_charge() will return
2231 * 0 ... on success, filling *ptr with a valid memcg pointer.
2232 * -ENOMEM ... charge failure because of resource limits.
2233 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2235 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2236 * the oom-killer can be invoked.
2238 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2240 unsigned int nr_pages,
2241 struct mem_cgroup **ptr,
2244 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2245 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2246 struct mem_cgroup *memcg = NULL;
2250 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2251 * in system level. So, allow to go ahead dying process in addition to
2254 if (unlikely(test_thread_flag(TIF_MEMDIE)
2255 || fatal_signal_pending(current)))
2259 * We always charge the cgroup the mm_struct belongs to.
2260 * The mm_struct's mem_cgroup changes on task migration if the
2261 * thread group leader migrates. It's possible that mm is not
2262 * set, if so charge the init_mm (happens for pagecache usage).
2265 *ptr = root_mem_cgroup;
2267 if (*ptr) { /* css should be a valid one */
2269 VM_BUG_ON(css_is_removed(&memcg->css));
2270 if (mem_cgroup_is_root(memcg))
2272 if (nr_pages == 1 && consume_stock(memcg))
2274 css_get(&memcg->css);
2276 struct task_struct *p;
2279 p = rcu_dereference(mm->owner);
2281 * Because we don't have task_lock(), "p" can exit.
2282 * In that case, "memcg" can point to root or p can be NULL with
2283 * race with swapoff. Then, we have small risk of mis-accouning.
2284 * But such kind of mis-account by race always happens because
2285 * we don't have cgroup_mutex(). It's overkill and we allo that
2287 * (*) swapoff at el will charge against mm-struct not against
2288 * task-struct. So, mm->owner can be NULL.
2290 memcg = mem_cgroup_from_task(p);
2292 memcg = root_mem_cgroup;
2293 if (mem_cgroup_is_root(memcg)) {
2297 if (nr_pages == 1 && consume_stock(memcg)) {
2299 * It seems dagerous to access memcg without css_get().
2300 * But considering how consume_stok works, it's not
2301 * necessary. If consume_stock success, some charges
2302 * from this memcg are cached on this cpu. So, we
2303 * don't need to call css_get()/css_tryget() before
2304 * calling consume_stock().
2309 /* after here, we may be blocked. we need to get refcnt */
2310 if (!css_tryget(&memcg->css)) {
2320 /* If killed, bypass charge */
2321 if (fatal_signal_pending(current)) {
2322 css_put(&memcg->css);
2327 if (oom && !nr_oom_retries) {
2329 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2332 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, oom_check);
2336 case CHARGE_RETRY: /* not in OOM situation but retry */
2338 css_put(&memcg->css);
2341 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2342 css_put(&memcg->css);
2344 case CHARGE_NOMEM: /* OOM routine works */
2346 css_put(&memcg->css);
2349 /* If oom, we never return -ENOMEM */
2352 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2353 css_put(&memcg->css);
2356 } while (ret != CHARGE_OK);
2358 if (batch > nr_pages)
2359 refill_stock(memcg, batch - nr_pages);
2360 css_put(&memcg->css);
2368 *ptr = root_mem_cgroup;
2373 * Somemtimes we have to undo a charge we got by try_charge().
2374 * This function is for that and do uncharge, put css's refcnt.
2375 * gotten by try_charge().
2377 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2378 unsigned int nr_pages)
2380 if (!mem_cgroup_is_root(memcg)) {
2381 unsigned long bytes = nr_pages * PAGE_SIZE;
2383 res_counter_uncharge(&memcg->res, bytes);
2384 if (do_swap_account)
2385 res_counter_uncharge(&memcg->memsw, bytes);
2390 * A helper function to get mem_cgroup from ID. must be called under
2391 * rcu_read_lock(). The caller must check css_is_removed() or some if
2392 * it's concern. (dropping refcnt from swap can be called against removed
2395 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2397 struct cgroup_subsys_state *css;
2399 /* ID 0 is unused ID */
2402 css = css_lookup(&mem_cgroup_subsys, id);
2405 return container_of(css, struct mem_cgroup, css);
2408 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2410 struct mem_cgroup *memcg = NULL;
2411 struct page_cgroup *pc;
2415 VM_BUG_ON(!PageLocked(page));
2417 pc = lookup_page_cgroup(page);
2418 lock_page_cgroup(pc);
2419 if (PageCgroupUsed(pc)) {
2420 memcg = pc->mem_cgroup;
2421 if (memcg && !css_tryget(&memcg->css))
2423 } else if (PageSwapCache(page)) {
2424 ent.val = page_private(page);
2425 id = lookup_swap_cgroup_id(ent);
2427 memcg = mem_cgroup_lookup(id);
2428 if (memcg && !css_tryget(&memcg->css))
2432 unlock_page_cgroup(pc);
2436 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2438 unsigned int nr_pages,
2439 struct page_cgroup *pc,
2440 enum charge_type ctype,
2443 struct zone *uninitialized_var(zone);
2444 bool was_on_lru = false;
2446 lock_page_cgroup(pc);
2447 if (unlikely(PageCgroupUsed(pc))) {
2448 unlock_page_cgroup(pc);
2449 __mem_cgroup_cancel_charge(memcg, nr_pages);
2453 * we don't need page_cgroup_lock about tail pages, becase they are not
2454 * accessed by any other context at this point.
2458 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2459 * may already be on some other mem_cgroup's LRU. Take care of it.
2462 zone = page_zone(page);
2463 spin_lock_irq(&zone->lru_lock);
2464 if (PageLRU(page)) {
2466 del_page_from_lru_list(zone, page, page_lru(page));
2471 pc->mem_cgroup = memcg;
2473 * We access a page_cgroup asynchronously without lock_page_cgroup().
2474 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2475 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2476 * before USED bit, we need memory barrier here.
2477 * See mem_cgroup_add_lru_list(), etc.
2481 case MEM_CGROUP_CHARGE_TYPE_CACHE:
2482 case MEM_CGROUP_CHARGE_TYPE_SHMEM:
2483 SetPageCgroupCache(pc);
2484 SetPageCgroupUsed(pc);
2486 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2487 ClearPageCgroupCache(pc);
2488 SetPageCgroupUsed(pc);
2496 VM_BUG_ON(PageLRU(page));
2498 add_page_to_lru_list(zone, page, page_lru(page));
2500 spin_unlock_irq(&zone->lru_lock);
2503 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), nr_pages);
2504 unlock_page_cgroup(pc);
2507 * "charge_statistics" updated event counter. Then, check it.
2508 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2509 * if they exceeds softlimit.
2511 memcg_check_events(memcg, page);
2514 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2516 #define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MOVE_LOCK) |\
2517 (1 << PCG_MIGRATION))
2519 * Because tail pages are not marked as "used", set it. We're under
2520 * zone->lru_lock, 'splitting on pmd' and compound_lock.
2521 * charge/uncharge will be never happen and move_account() is done under
2522 * compound_lock(), so we don't have to take care of races.
2524 void mem_cgroup_split_huge_fixup(struct page *head)
2526 struct page_cgroup *head_pc = lookup_page_cgroup(head);
2527 struct page_cgroup *pc;
2530 if (mem_cgroup_disabled())
2532 for (i = 1; i < HPAGE_PMD_NR; i++) {
2534 pc->mem_cgroup = head_pc->mem_cgroup;
2535 smp_wmb();/* see __commit_charge() */
2536 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2539 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2542 * mem_cgroup_move_account - move account of the page
2544 * @nr_pages: number of regular pages (>1 for huge pages)
2545 * @pc: page_cgroup of the page.
2546 * @from: mem_cgroup which the page is moved from.
2547 * @to: mem_cgroup which the page is moved to. @from != @to.
2548 * @uncharge: whether we should call uncharge and css_put against @from.
2550 * The caller must confirm following.
2551 * - page is not on LRU (isolate_page() is useful.)
2552 * - compound_lock is held when nr_pages > 1
2554 * This function doesn't do "charge" nor css_get to new cgroup. It should be
2555 * done by a caller(__mem_cgroup_try_charge would be useful). If @uncharge is
2556 * true, this function does "uncharge" from old cgroup, but it doesn't if
2557 * @uncharge is false, so a caller should do "uncharge".
2559 static int mem_cgroup_move_account(struct page *page,
2560 unsigned int nr_pages,
2561 struct page_cgroup *pc,
2562 struct mem_cgroup *from,
2563 struct mem_cgroup *to,
2566 unsigned long flags;
2569 VM_BUG_ON(from == to);
2570 VM_BUG_ON(PageLRU(page));
2572 * The page is isolated from LRU. So, collapse function
2573 * will not handle this page. But page splitting can happen.
2574 * Do this check under compound_page_lock(). The caller should
2578 if (nr_pages > 1 && !PageTransHuge(page))
2581 lock_page_cgroup(pc);
2584 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
2587 move_lock_page_cgroup(pc, &flags);
2589 if (PageCgroupFileMapped(pc)) {
2590 /* Update mapped_file data for mem_cgroup */
2592 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2593 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2596 mem_cgroup_charge_statistics(from, PageCgroupCache(pc), -nr_pages);
2598 /* This is not "cancel", but cancel_charge does all we need. */
2599 __mem_cgroup_cancel_charge(from, nr_pages);
2601 /* caller should have done css_get */
2602 pc->mem_cgroup = to;
2603 mem_cgroup_charge_statistics(to, PageCgroupCache(pc), nr_pages);
2605 * We charges against "to" which may not have any tasks. Then, "to"
2606 * can be under rmdir(). But in current implementation, caller of
2607 * this function is just force_empty() and move charge, so it's
2608 * guaranteed that "to" is never removed. So, we don't check rmdir
2611 move_unlock_page_cgroup(pc, &flags);
2614 unlock_page_cgroup(pc);
2618 memcg_check_events(to, page);
2619 memcg_check_events(from, page);
2625 * move charges to its parent.
2628 static int mem_cgroup_move_parent(struct page *page,
2629 struct page_cgroup *pc,
2630 struct mem_cgroup *child,
2633 struct cgroup *cg = child->css.cgroup;
2634 struct cgroup *pcg = cg->parent;
2635 struct mem_cgroup *parent;
2636 unsigned int nr_pages;
2637 unsigned long uninitialized_var(flags);
2645 if (!get_page_unless_zero(page))
2647 if (isolate_lru_page(page))
2650 nr_pages = hpage_nr_pages(page);
2652 parent = mem_cgroup_from_cont(pcg);
2653 ret = __mem_cgroup_try_charge(NULL, gfp_mask, nr_pages, &parent, false);
2658 flags = compound_lock_irqsave(page);
2660 ret = mem_cgroup_move_account(page, nr_pages, pc, child, parent, true);
2662 __mem_cgroup_cancel_charge(parent, nr_pages);
2665 compound_unlock_irqrestore(page, flags);
2667 putback_lru_page(page);
2675 * Charge the memory controller for page usage.
2677 * 0 if the charge was successful
2678 * < 0 if the cgroup is over its limit
2680 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2681 gfp_t gfp_mask, enum charge_type ctype)
2683 struct mem_cgroup *memcg = NULL;
2684 unsigned int nr_pages = 1;
2685 struct page_cgroup *pc;
2689 if (PageTransHuge(page)) {
2690 nr_pages <<= compound_order(page);
2691 VM_BUG_ON(!PageTransHuge(page));
2693 * Never OOM-kill a process for a huge page. The
2694 * fault handler will fall back to regular pages.
2699 pc = lookup_page_cgroup(page);
2700 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
2703 __mem_cgroup_commit_charge(memcg, page, nr_pages, pc, ctype, false);
2707 int mem_cgroup_newpage_charge(struct page *page,
2708 struct mm_struct *mm, gfp_t gfp_mask)
2710 if (mem_cgroup_disabled())
2712 VM_BUG_ON(page_mapped(page));
2713 VM_BUG_ON(page->mapping && !PageAnon(page));
2715 return mem_cgroup_charge_common(page, mm, gfp_mask,
2716 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2720 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2721 enum charge_type ctype);
2723 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2726 struct mem_cgroup *memcg = NULL;
2727 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
2730 if (mem_cgroup_disabled())
2732 if (PageCompound(page))
2737 if (!page_is_file_cache(page))
2738 type = MEM_CGROUP_CHARGE_TYPE_SHMEM;
2740 if (!PageSwapCache(page))
2741 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
2742 else { /* page is swapcache/shmem */
2743 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &memcg);
2745 __mem_cgroup_commit_charge_swapin(page, memcg, type);
2751 * While swap-in, try_charge -> commit or cancel, the page is locked.
2752 * And when try_charge() successfully returns, one refcnt to memcg without
2753 * struct page_cgroup is acquired. This refcnt will be consumed by
2754 * "commit()" or removed by "cancel()"
2756 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2758 gfp_t mask, struct mem_cgroup **memcgp)
2760 struct mem_cgroup *memcg;
2765 if (mem_cgroup_disabled())
2768 if (!do_swap_account)
2771 * A racing thread's fault, or swapoff, may have already updated
2772 * the pte, and even removed page from swap cache: in those cases
2773 * do_swap_page()'s pte_same() test will fail; but there's also a
2774 * KSM case which does need to charge the page.
2776 if (!PageSwapCache(page))
2778 memcg = try_get_mem_cgroup_from_page(page);
2782 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
2783 css_put(&memcg->css);
2790 ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
2797 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
2798 enum charge_type ctype)
2800 struct page_cgroup *pc;
2802 if (mem_cgroup_disabled())
2806 cgroup_exclude_rmdir(&memcg->css);
2808 pc = lookup_page_cgroup(page);
2809 __mem_cgroup_commit_charge(memcg, page, 1, pc, ctype, true);
2811 * Now swap is on-memory. This means this page may be
2812 * counted both as mem and swap....double count.
2813 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2814 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2815 * may call delete_from_swap_cache() before reach here.
2817 if (do_swap_account && PageSwapCache(page)) {
2818 swp_entry_t ent = {.val = page_private(page)};
2819 struct mem_cgroup *swap_memcg;
2822 id = swap_cgroup_record(ent, 0);
2824 swap_memcg = mem_cgroup_lookup(id);
2827 * This recorded memcg can be obsolete one. So, avoid
2828 * calling css_tryget
2830 if (!mem_cgroup_is_root(swap_memcg))
2831 res_counter_uncharge(&swap_memcg->memsw,
2833 mem_cgroup_swap_statistics(swap_memcg, false);
2834 mem_cgroup_put(swap_memcg);
2839 * At swapin, we may charge account against cgroup which has no tasks.
2840 * So, rmdir()->pre_destroy() can be called while we do this charge.
2841 * In that case, we need to call pre_destroy() again. check it here.
2843 cgroup_release_and_wakeup_rmdir(&memcg->css);
2846 void mem_cgroup_commit_charge_swapin(struct page *page,
2847 struct mem_cgroup *memcg)
2849 __mem_cgroup_commit_charge_swapin(page, memcg,
2850 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2853 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
2855 if (mem_cgroup_disabled())
2859 __mem_cgroup_cancel_charge(memcg, 1);
2862 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
2863 unsigned int nr_pages,
2864 const enum charge_type ctype)
2866 struct memcg_batch_info *batch = NULL;
2867 bool uncharge_memsw = true;
2869 /* If swapout, usage of swap doesn't decrease */
2870 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2871 uncharge_memsw = false;
2873 batch = ¤t->memcg_batch;
2875 * In usual, we do css_get() when we remember memcg pointer.
2876 * But in this case, we keep res->usage until end of a series of
2877 * uncharges. Then, it's ok to ignore memcg's refcnt.
2880 batch->memcg = memcg;
2882 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2883 * In those cases, all pages freed continuously can be expected to be in
2884 * the same cgroup and we have chance to coalesce uncharges.
2885 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2886 * because we want to do uncharge as soon as possible.
2889 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
2890 goto direct_uncharge;
2893 goto direct_uncharge;
2896 * In typical case, batch->memcg == mem. This means we can
2897 * merge a series of uncharges to an uncharge of res_counter.
2898 * If not, we uncharge res_counter ony by one.
2900 if (batch->memcg != memcg)
2901 goto direct_uncharge;
2902 /* remember freed charge and uncharge it later */
2905 batch->memsw_nr_pages++;
2908 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
2910 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
2911 if (unlikely(batch->memcg != memcg))
2912 memcg_oom_recover(memcg);
2916 * uncharge if !page_mapped(page)
2918 static struct mem_cgroup *
2919 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
2921 struct mem_cgroup *memcg = NULL;
2922 unsigned int nr_pages = 1;
2923 struct page_cgroup *pc;
2925 if (mem_cgroup_disabled())
2928 if (PageSwapCache(page))
2931 if (PageTransHuge(page)) {
2932 nr_pages <<= compound_order(page);
2933 VM_BUG_ON(!PageTransHuge(page));
2936 * Check if our page_cgroup is valid
2938 pc = lookup_page_cgroup(page);
2939 if (unlikely(!PageCgroupUsed(pc)))
2942 lock_page_cgroup(pc);
2944 memcg = pc->mem_cgroup;
2946 if (!PageCgroupUsed(pc))
2950 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2951 case MEM_CGROUP_CHARGE_TYPE_DROP:
2952 /* See mem_cgroup_prepare_migration() */
2953 if (page_mapped(page) || PageCgroupMigration(pc))
2956 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
2957 if (!PageAnon(page)) { /* Shared memory */
2958 if (page->mapping && !page_is_file_cache(page))
2960 } else if (page_mapped(page)) /* Anon */
2967 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), -nr_pages);
2969 ClearPageCgroupUsed(pc);
2971 * pc->mem_cgroup is not cleared here. It will be accessed when it's
2972 * freed from LRU. This is safe because uncharged page is expected not
2973 * to be reused (freed soon). Exception is SwapCache, it's handled by
2974 * special functions.
2977 unlock_page_cgroup(pc);
2979 * even after unlock, we have memcg->res.usage here and this memcg
2980 * will never be freed.
2982 memcg_check_events(memcg, page);
2983 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
2984 mem_cgroup_swap_statistics(memcg, true);
2985 mem_cgroup_get(memcg);
2987 if (!mem_cgroup_is_root(memcg))
2988 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
2993 unlock_page_cgroup(pc);
2997 void mem_cgroup_uncharge_page(struct page *page)
3000 if (page_mapped(page))
3002 VM_BUG_ON(page->mapping && !PageAnon(page));
3003 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
3006 void mem_cgroup_uncharge_cache_page(struct page *page)
3008 VM_BUG_ON(page_mapped(page));
3009 VM_BUG_ON(page->mapping);
3010 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
3014 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
3015 * In that cases, pages are freed continuously and we can expect pages
3016 * are in the same memcg. All these calls itself limits the number of
3017 * pages freed at once, then uncharge_start/end() is called properly.
3018 * This may be called prural(2) times in a context,
3021 void mem_cgroup_uncharge_start(void)
3023 current->memcg_batch.do_batch++;
3024 /* We can do nest. */
3025 if (current->memcg_batch.do_batch == 1) {
3026 current->memcg_batch.memcg = NULL;
3027 current->memcg_batch.nr_pages = 0;
3028 current->memcg_batch.memsw_nr_pages = 0;
3032 void mem_cgroup_uncharge_end(void)
3034 struct memcg_batch_info *batch = ¤t->memcg_batch;
3036 if (!batch->do_batch)
3040 if (batch->do_batch) /* If stacked, do nothing. */
3046 * This "batch->memcg" is valid without any css_get/put etc...
3047 * bacause we hide charges behind us.
3049 if (batch->nr_pages)
3050 res_counter_uncharge(&batch->memcg->res,
3051 batch->nr_pages * PAGE_SIZE);
3052 if (batch->memsw_nr_pages)
3053 res_counter_uncharge(&batch->memcg->memsw,
3054 batch->memsw_nr_pages * PAGE_SIZE);
3055 memcg_oom_recover(batch->memcg);
3056 /* forget this pointer (for sanity check) */
3057 batch->memcg = NULL;
3062 * called after __delete_from_swap_cache() and drop "page" account.
3063 * memcg information is recorded to swap_cgroup of "ent"
3066 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
3068 struct mem_cgroup *memcg;
3069 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
3071 if (!swapout) /* this was a swap cache but the swap is unused ! */
3072 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
3074 memcg = __mem_cgroup_uncharge_common(page, ctype);
3077 * record memcg information, if swapout && memcg != NULL,
3078 * mem_cgroup_get() was called in uncharge().
3080 if (do_swap_account && swapout && memcg)
3081 swap_cgroup_record(ent, css_id(&memcg->css));
3085 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3087 * called from swap_entry_free(). remove record in swap_cgroup and
3088 * uncharge "memsw" account.
3090 void mem_cgroup_uncharge_swap(swp_entry_t ent)
3092 struct mem_cgroup *memcg;
3095 if (!do_swap_account)
3098 id = swap_cgroup_record(ent, 0);
3100 memcg = mem_cgroup_lookup(id);
3103 * We uncharge this because swap is freed.
3104 * This memcg can be obsolete one. We avoid calling css_tryget
3106 if (!mem_cgroup_is_root(memcg))
3107 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
3108 mem_cgroup_swap_statistics(memcg, false);
3109 mem_cgroup_put(memcg);
3115 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3116 * @entry: swap entry to be moved
3117 * @from: mem_cgroup which the entry is moved from
3118 * @to: mem_cgroup which the entry is moved to
3119 * @need_fixup: whether we should fixup res_counters and refcounts.
3121 * It succeeds only when the swap_cgroup's record for this entry is the same
3122 * as the mem_cgroup's id of @from.
3124 * Returns 0 on success, -EINVAL on failure.
3126 * The caller must have charged to @to, IOW, called res_counter_charge() about
3127 * both res and memsw, and called css_get().
3129 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3130 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3132 unsigned short old_id, new_id;
3134 old_id = css_id(&from->css);
3135 new_id = css_id(&to->css);
3137 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3138 mem_cgroup_swap_statistics(from, false);
3139 mem_cgroup_swap_statistics(to, true);
3141 * This function is only called from task migration context now.
3142 * It postpones res_counter and refcount handling till the end
3143 * of task migration(mem_cgroup_clear_mc()) for performance
3144 * improvement. But we cannot postpone mem_cgroup_get(to)
3145 * because if the process that has been moved to @to does
3146 * swap-in, the refcount of @to might be decreased to 0.
3150 if (!mem_cgroup_is_root(from))
3151 res_counter_uncharge(&from->memsw, PAGE_SIZE);
3152 mem_cgroup_put(from);
3154 * we charged both to->res and to->memsw, so we should
3157 if (!mem_cgroup_is_root(to))
3158 res_counter_uncharge(&to->res, PAGE_SIZE);
3165 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3166 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3173 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
3176 int mem_cgroup_prepare_migration(struct page *page,
3177 struct page *newpage, struct mem_cgroup **memcgp, gfp_t gfp_mask)
3179 struct mem_cgroup *memcg = NULL;
3180 struct page_cgroup *pc;
3181 enum charge_type ctype;
3186 VM_BUG_ON(PageTransHuge(page));
3187 if (mem_cgroup_disabled())
3190 pc = lookup_page_cgroup(page);
3191 lock_page_cgroup(pc);
3192 if (PageCgroupUsed(pc)) {
3193 memcg = pc->mem_cgroup;
3194 css_get(&memcg->css);
3196 * At migrating an anonymous page, its mapcount goes down
3197 * to 0 and uncharge() will be called. But, even if it's fully
3198 * unmapped, migration may fail and this page has to be
3199 * charged again. We set MIGRATION flag here and delay uncharge
3200 * until end_migration() is called
3202 * Corner Case Thinking
3204 * When the old page was mapped as Anon and it's unmap-and-freed
3205 * while migration was ongoing.
3206 * If unmap finds the old page, uncharge() of it will be delayed
3207 * until end_migration(). If unmap finds a new page, it's
3208 * uncharged when it make mapcount to be 1->0. If unmap code
3209 * finds swap_migration_entry, the new page will not be mapped
3210 * and end_migration() will find it(mapcount==0).
3213 * When the old page was mapped but migraion fails, the kernel
3214 * remaps it. A charge for it is kept by MIGRATION flag even
3215 * if mapcount goes down to 0. We can do remap successfully
3216 * without charging it again.
3219 * The "old" page is under lock_page() until the end of
3220 * migration, so, the old page itself will not be swapped-out.
3221 * If the new page is swapped out before end_migraton, our
3222 * hook to usual swap-out path will catch the event.
3225 SetPageCgroupMigration(pc);
3227 unlock_page_cgroup(pc);
3229 * If the page is not charged at this point,
3236 ret = __mem_cgroup_try_charge(NULL, gfp_mask, 1, memcgp, false);
3237 css_put(&memcg->css);/* drop extra refcnt */
3239 if (PageAnon(page)) {
3240 lock_page_cgroup(pc);
3241 ClearPageCgroupMigration(pc);
3242 unlock_page_cgroup(pc);
3244 * The old page may be fully unmapped while we kept it.
3246 mem_cgroup_uncharge_page(page);
3248 /* we'll need to revisit this error code (we have -EINTR) */
3252 * We charge new page before it's used/mapped. So, even if unlock_page()
3253 * is called before end_migration, we can catch all events on this new
3254 * page. In the case new page is migrated but not remapped, new page's
3255 * mapcount will be finally 0 and we call uncharge in end_migration().
3257 pc = lookup_page_cgroup(newpage);
3259 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
3260 else if (page_is_file_cache(page))
3261 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
3263 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3264 __mem_cgroup_commit_charge(memcg, newpage, 1, pc, ctype, false);
3268 /* remove redundant charge if migration failed*/
3269 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
3270 struct page *oldpage, struct page *newpage, bool migration_ok)
3272 struct page *used, *unused;
3273 struct page_cgroup *pc;
3277 /* blocks rmdir() */
3278 cgroup_exclude_rmdir(&memcg->css);
3279 if (!migration_ok) {
3287 * We disallowed uncharge of pages under migration because mapcount
3288 * of the page goes down to zero, temporarly.
3289 * Clear the flag and check the page should be charged.
3291 pc = lookup_page_cgroup(oldpage);
3292 lock_page_cgroup(pc);
3293 ClearPageCgroupMigration(pc);
3294 unlock_page_cgroup(pc);
3296 __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE);
3299 * If a page is a file cache, radix-tree replacement is very atomic
3300 * and we can skip this check. When it was an Anon page, its mapcount
3301 * goes down to 0. But because we added MIGRATION flage, it's not
3302 * uncharged yet. There are several case but page->mapcount check
3303 * and USED bit check in mem_cgroup_uncharge_page() will do enough
3304 * check. (see prepare_charge() also)
3307 mem_cgroup_uncharge_page(used);
3309 * At migration, we may charge account against cgroup which has no
3311 * So, rmdir()->pre_destroy() can be called while we do this charge.
3312 * In that case, we need to call pre_destroy() again. check it here.
3314 cgroup_release_and_wakeup_rmdir(&memcg->css);
3318 * At replace page cache, newpage is not under any memcg but it's on
3319 * LRU. So, this function doesn't touch res_counter but handles LRU
3320 * in correct way. Both pages are locked so we cannot race with uncharge.
3322 void mem_cgroup_replace_page_cache(struct page *oldpage,
3323 struct page *newpage)
3325 struct mem_cgroup *memcg;
3326 struct page_cgroup *pc;
3327 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
3329 if (mem_cgroup_disabled())
3332 pc = lookup_page_cgroup(oldpage);
3333 /* fix accounting on old pages */
3334 lock_page_cgroup(pc);
3335 memcg = pc->mem_cgroup;
3336 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), -1);
3337 ClearPageCgroupUsed(pc);
3338 unlock_page_cgroup(pc);
3340 if (PageSwapBacked(oldpage))
3341 type = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3344 * Even if newpage->mapping was NULL before starting replacement,
3345 * the newpage may be on LRU(or pagevec for LRU) already. We lock
3346 * LRU while we overwrite pc->mem_cgroup.
3348 __mem_cgroup_commit_charge(memcg, newpage, 1, pc, type, true);
3351 #ifdef CONFIG_DEBUG_VM
3352 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3354 struct page_cgroup *pc;
3356 pc = lookup_page_cgroup(page);
3358 * Can be NULL while feeding pages into the page allocator for
3359 * the first time, i.e. during boot or memory hotplug;
3360 * or when mem_cgroup_disabled().
3362 if (likely(pc) && PageCgroupUsed(pc))
3367 bool mem_cgroup_bad_page_check(struct page *page)
3369 if (mem_cgroup_disabled())
3372 return lookup_page_cgroup_used(page) != NULL;
3375 void mem_cgroup_print_bad_page(struct page *page)
3377 struct page_cgroup *pc;
3379 pc = lookup_page_cgroup_used(page);
3381 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
3382 pc, pc->flags, pc->mem_cgroup);
3387 static DEFINE_MUTEX(set_limit_mutex);
3389 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3390 unsigned long long val)
3393 u64 memswlimit, memlimit;
3395 int children = mem_cgroup_count_children(memcg);
3396 u64 curusage, oldusage;
3400 * For keeping hierarchical_reclaim simple, how long we should retry
3401 * is depends on callers. We set our retry-count to be function
3402 * of # of children which we should visit in this loop.
3404 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3406 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3409 while (retry_count) {
3410 if (signal_pending(current)) {
3415 * Rather than hide all in some function, I do this in
3416 * open coded manner. You see what this really does.
3417 * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3419 mutex_lock(&set_limit_mutex);
3420 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3421 if (memswlimit < val) {
3423 mutex_unlock(&set_limit_mutex);
3427 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3431 ret = res_counter_set_limit(&memcg->res, val);
3433 if (memswlimit == val)
3434 memcg->memsw_is_minimum = true;
3436 memcg->memsw_is_minimum = false;
3438 mutex_unlock(&set_limit_mutex);
3443 mem_cgroup_reclaim(memcg, GFP_KERNEL,
3444 MEM_CGROUP_RECLAIM_SHRINK);
3445 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3446 /* Usage is reduced ? */
3447 if (curusage >= oldusage)
3450 oldusage = curusage;
3452 if (!ret && enlarge)
3453 memcg_oom_recover(memcg);
3458 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3459 unsigned long long val)
3462 u64 memlimit, memswlimit, oldusage, curusage;
3463 int children = mem_cgroup_count_children(memcg);
3467 /* see mem_cgroup_resize_res_limit */
3468 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3469 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3470 while (retry_count) {
3471 if (signal_pending(current)) {
3476 * Rather than hide all in some function, I do this in
3477 * open coded manner. You see what this really does.
3478 * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3480 mutex_lock(&set_limit_mutex);
3481 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3482 if (memlimit > val) {
3484 mutex_unlock(&set_limit_mutex);
3487 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3488 if (memswlimit < val)
3490 ret = res_counter_set_limit(&memcg->memsw, val);
3492 if (memlimit == val)
3493 memcg->memsw_is_minimum = true;
3495 memcg->memsw_is_minimum = false;
3497 mutex_unlock(&set_limit_mutex);
3502 mem_cgroup_reclaim(memcg, GFP_KERNEL,
3503 MEM_CGROUP_RECLAIM_NOSWAP |
3504 MEM_CGROUP_RECLAIM_SHRINK);
3505 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3506 /* Usage is reduced ? */
3507 if (curusage >= oldusage)
3510 oldusage = curusage;
3512 if (!ret && enlarge)
3513 memcg_oom_recover(memcg);
3517 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3519 unsigned long *total_scanned)
3521 unsigned long nr_reclaimed = 0;
3522 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3523 unsigned long reclaimed;
3525 struct mem_cgroup_tree_per_zone *mctz;
3526 unsigned long long excess;
3527 unsigned long nr_scanned;
3532 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3534 * This loop can run a while, specially if mem_cgroup's continuously
3535 * keep exceeding their soft limit and putting the system under
3542 mz = mem_cgroup_largest_soft_limit_node(mctz);
3547 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
3548 gfp_mask, &nr_scanned);
3549 nr_reclaimed += reclaimed;
3550 *total_scanned += nr_scanned;
3551 spin_lock(&mctz->lock);
3554 * If we failed to reclaim anything from this memory cgroup
3555 * it is time to move on to the next cgroup
3561 * Loop until we find yet another one.
3563 * By the time we get the soft_limit lock
3564 * again, someone might have aded the
3565 * group back on the RB tree. Iterate to
3566 * make sure we get a different mem.
3567 * mem_cgroup_largest_soft_limit_node returns
3568 * NULL if no other cgroup is present on
3572 __mem_cgroup_largest_soft_limit_node(mctz);
3574 css_put(&next_mz->memcg->css);
3575 else /* next_mz == NULL or other memcg */
3579 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
3580 excess = res_counter_soft_limit_excess(&mz->memcg->res);
3582 * One school of thought says that we should not add
3583 * back the node to the tree if reclaim returns 0.
3584 * But our reclaim could return 0, simply because due
3585 * to priority we are exposing a smaller subset of
3586 * memory to reclaim from. Consider this as a longer
3589 /* If excess == 0, no tree ops */
3590 __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess);
3591 spin_unlock(&mctz->lock);
3592 css_put(&mz->memcg->css);
3595 * Could not reclaim anything and there are no more
3596 * mem cgroups to try or we seem to be looping without
3597 * reclaiming anything.
3599 if (!nr_reclaimed &&
3601 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3603 } while (!nr_reclaimed);
3605 css_put(&next_mz->memcg->css);
3606 return nr_reclaimed;
3610 * This routine traverse page_cgroup in given list and drop them all.
3611 * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3613 static int mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
3614 int node, int zid, enum lru_list lru)
3616 struct mem_cgroup_per_zone *mz;
3617 unsigned long flags, loop;
3618 struct list_head *list;
3623 zone = &NODE_DATA(node)->node_zones[zid];
3624 mz = mem_cgroup_zoneinfo(memcg, node, zid);
3625 list = &mz->lruvec.lists[lru];
3627 loop = mz->lru_size[lru];
3628 /* give some margin against EBUSY etc...*/
3632 struct page_cgroup *pc;
3636 spin_lock_irqsave(&zone->lru_lock, flags);
3637 if (list_empty(list)) {
3638 spin_unlock_irqrestore(&zone->lru_lock, flags);
3641 page = list_entry(list->prev, struct page, lru);
3643 list_move(&page->lru, list);
3645 spin_unlock_irqrestore(&zone->lru_lock, flags);
3648 spin_unlock_irqrestore(&zone->lru_lock, flags);
3650 pc = lookup_page_cgroup(page);
3652 ret = mem_cgroup_move_parent(page, pc, memcg, GFP_KERNEL);
3653 if (ret == -ENOMEM || ret == -EINTR)
3656 if (ret == -EBUSY || ret == -EINVAL) {
3657 /* found lock contention or "pc" is obsolete. */
3664 if (!ret && !list_empty(list))
3670 * make mem_cgroup's charge to be 0 if there is no task.
3671 * This enables deleting this mem_cgroup.
3673 static int mem_cgroup_force_empty(struct mem_cgroup *memcg, bool free_all)
3676 int node, zid, shrink;
3677 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3678 struct cgroup *cgrp = memcg->css.cgroup;
3680 css_get(&memcg->css);
3683 /* should free all ? */
3689 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3692 if (signal_pending(current))
3694 /* This is for making all *used* pages to be on LRU. */
3695 lru_add_drain_all();
3696 drain_all_stock_sync(memcg);
3698 mem_cgroup_start_move(memcg);
3699 for_each_node_state(node, N_HIGH_MEMORY) {
3700 for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3703 ret = mem_cgroup_force_empty_list(memcg,
3712 mem_cgroup_end_move(memcg);
3713 memcg_oom_recover(memcg);
3714 /* it seems parent cgroup doesn't have enough mem */
3718 /* "ret" should also be checked to ensure all lists are empty. */
3719 } while (memcg->res.usage > 0 || ret);
3721 css_put(&memcg->css);
3725 /* returns EBUSY if there is a task or if we come here twice. */
3726 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3730 /* we call try-to-free pages for make this cgroup empty */
3731 lru_add_drain_all();
3732 /* try to free all pages in this cgroup */
3734 while (nr_retries && memcg->res.usage > 0) {
3737 if (signal_pending(current)) {
3741 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
3745 /* maybe some writeback is necessary */
3746 congestion_wait(BLK_RW_ASYNC, HZ/10);
3751 /* try move_account...there may be some *locked* pages. */
3755 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3757 return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3761 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3763 return mem_cgroup_from_cont(cont)->use_hierarchy;
3766 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3770 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3771 struct cgroup *parent = cont->parent;
3772 struct mem_cgroup *parent_memcg = NULL;
3775 parent_memcg = mem_cgroup_from_cont(parent);
3779 * If parent's use_hierarchy is set, we can't make any modifications
3780 * in the child subtrees. If it is unset, then the change can
3781 * occur, provided the current cgroup has no children.
3783 * For the root cgroup, parent_mem is NULL, we allow value to be
3784 * set if there are no children.
3786 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3787 (val == 1 || val == 0)) {
3788 if (list_empty(&cont->children))
3789 memcg->use_hierarchy = val;
3800 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
3801 enum mem_cgroup_stat_index idx)
3803 struct mem_cgroup *iter;
3806 /* Per-cpu values can be negative, use a signed accumulator */
3807 for_each_mem_cgroup_tree(iter, memcg)
3808 val += mem_cgroup_read_stat(iter, idx);
3810 if (val < 0) /* race ? */
3815 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3819 if (!mem_cgroup_is_root(memcg)) {
3821 return res_counter_read_u64(&memcg->res, RES_USAGE);
3823 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
3826 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
3827 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
3830 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
3832 return val << PAGE_SHIFT;
3835 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
3837 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3841 type = MEMFILE_TYPE(cft->private);
3842 name = MEMFILE_ATTR(cft->private);
3845 if (name == RES_USAGE)
3846 val = mem_cgroup_usage(memcg, false);
3848 val = res_counter_read_u64(&memcg->res, name);
3851 if (name == RES_USAGE)
3852 val = mem_cgroup_usage(memcg, true);
3854 val = res_counter_read_u64(&memcg->memsw, name);
3863 * The user of this function is...
3866 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
3869 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3871 unsigned long long val;
3874 type = MEMFILE_TYPE(cft->private);
3875 name = MEMFILE_ATTR(cft->private);
3878 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3882 /* This function does all necessary parse...reuse it */
3883 ret = res_counter_memparse_write_strategy(buffer, &val);
3887 ret = mem_cgroup_resize_limit(memcg, val);
3889 ret = mem_cgroup_resize_memsw_limit(memcg, val);
3891 case RES_SOFT_LIMIT:
3892 ret = res_counter_memparse_write_strategy(buffer, &val);
3896 * For memsw, soft limits are hard to implement in terms
3897 * of semantics, for now, we support soft limits for
3898 * control without swap
3901 ret = res_counter_set_soft_limit(&memcg->res, val);
3906 ret = -EINVAL; /* should be BUG() ? */
3912 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
3913 unsigned long long *mem_limit, unsigned long long *memsw_limit)
3915 struct cgroup *cgroup;
3916 unsigned long long min_limit, min_memsw_limit, tmp;
3918 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3919 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3920 cgroup = memcg->css.cgroup;
3921 if (!memcg->use_hierarchy)
3924 while (cgroup->parent) {
3925 cgroup = cgroup->parent;
3926 memcg = mem_cgroup_from_cont(cgroup);
3927 if (!memcg->use_hierarchy)
3929 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
3930 min_limit = min(min_limit, tmp);
3931 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3932 min_memsw_limit = min(min_memsw_limit, tmp);
3935 *mem_limit = min_limit;
3936 *memsw_limit = min_memsw_limit;
3939 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
3941 struct mem_cgroup *memcg;
3944 memcg = mem_cgroup_from_cont(cont);
3945 type = MEMFILE_TYPE(event);
3946 name = MEMFILE_ATTR(event);
3950 res_counter_reset_max(&memcg->res);
3952 res_counter_reset_max(&memcg->memsw);
3956 res_counter_reset_failcnt(&memcg->res);
3958 res_counter_reset_failcnt(&memcg->memsw);
3965 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
3968 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
3972 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
3973 struct cftype *cft, u64 val)
3975 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3977 if (val >= (1 << NR_MOVE_TYPE))
3980 * We check this value several times in both in can_attach() and
3981 * attach(), so we need cgroup lock to prevent this value from being
3985 memcg->move_charge_at_immigrate = val;
3991 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
3992 struct cftype *cft, u64 val)
3999 /* For read statistics */
4017 struct mcs_total_stat {
4018 s64 stat[NR_MCS_STAT];
4024 } memcg_stat_strings[NR_MCS_STAT] = {
4025 {"cache", "total_cache"},
4026 {"rss", "total_rss"},
4027 {"mapped_file", "total_mapped_file"},
4028 {"pgpgin", "total_pgpgin"},
4029 {"pgpgout", "total_pgpgout"},
4030 {"swap", "total_swap"},
4031 {"pgfault", "total_pgfault"},
4032 {"pgmajfault", "total_pgmajfault"},
4033 {"inactive_anon", "total_inactive_anon"},
4034 {"active_anon", "total_active_anon"},
4035 {"inactive_file", "total_inactive_file"},
4036 {"active_file", "total_active_file"},
4037 {"unevictable", "total_unevictable"}
4042 mem_cgroup_get_local_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4047 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_CACHE);
4048 s->stat[MCS_CACHE] += val * PAGE_SIZE;
4049 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_RSS);
4050 s->stat[MCS_RSS] += val * PAGE_SIZE;
4051 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_FILE_MAPPED);
4052 s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
4053 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGIN);
4054 s->stat[MCS_PGPGIN] += val;
4055 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGOUT);
4056 s->stat[MCS_PGPGOUT] += val;
4057 if (do_swap_account) {
4058 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
4059 s->stat[MCS_SWAP] += val * PAGE_SIZE;
4061 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGFAULT);
4062 s->stat[MCS_PGFAULT] += val;
4063 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGMAJFAULT);
4064 s->stat[MCS_PGMAJFAULT] += val;
4067 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_ANON));
4068 s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
4069 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_ANON));
4070 s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
4071 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_FILE));
4072 s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
4073 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_FILE));
4074 s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
4075 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
4076 s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
4080 mem_cgroup_get_total_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4082 struct mem_cgroup *iter;
4084 for_each_mem_cgroup_tree(iter, memcg)
4085 mem_cgroup_get_local_stat(iter, s);
4089 static int mem_control_numa_stat_show(struct seq_file *m, void *arg)
4092 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
4093 unsigned long node_nr;
4094 struct cgroup *cont = m->private;
4095 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4097 total_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL);
4098 seq_printf(m, "total=%lu", total_nr);
4099 for_each_node_state(nid, N_HIGH_MEMORY) {
4100 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL);
4101 seq_printf(m, " N%d=%lu", nid, node_nr);
4105 file_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_FILE);
4106 seq_printf(m, "file=%lu", file_nr);
4107 for_each_node_state(nid, N_HIGH_MEMORY) {
4108 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4110 seq_printf(m, " N%d=%lu", nid, node_nr);
4114 anon_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_ANON);
4115 seq_printf(m, "anon=%lu", anon_nr);
4116 for_each_node_state(nid, N_HIGH_MEMORY) {
4117 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4119 seq_printf(m, " N%d=%lu", nid, node_nr);
4123 unevictable_nr = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
4124 seq_printf(m, "unevictable=%lu", unevictable_nr);
4125 for_each_node_state(nid, N_HIGH_MEMORY) {
4126 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4127 BIT(LRU_UNEVICTABLE));
4128 seq_printf(m, " N%d=%lu", nid, node_nr);
4133 #endif /* CONFIG_NUMA */
4135 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
4136 struct cgroup_map_cb *cb)
4138 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4139 struct mcs_total_stat mystat;
4142 memset(&mystat, 0, sizeof(mystat));
4143 mem_cgroup_get_local_stat(memcg, &mystat);
4146 for (i = 0; i < NR_MCS_STAT; i++) {
4147 if (i == MCS_SWAP && !do_swap_account)
4149 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
4152 /* Hierarchical information */
4154 unsigned long long limit, memsw_limit;
4155 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
4156 cb->fill(cb, "hierarchical_memory_limit", limit);
4157 if (do_swap_account)
4158 cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
4161 memset(&mystat, 0, sizeof(mystat));
4162 mem_cgroup_get_total_stat(memcg, &mystat);
4163 for (i = 0; i < NR_MCS_STAT; i++) {
4164 if (i == MCS_SWAP && !do_swap_account)
4166 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
4169 #ifdef CONFIG_DEBUG_VM
4172 struct mem_cgroup_per_zone *mz;
4173 unsigned long recent_rotated[2] = {0, 0};
4174 unsigned long recent_scanned[2] = {0, 0};
4176 for_each_online_node(nid)
4177 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4178 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
4180 recent_rotated[0] +=
4181 mz->reclaim_stat.recent_rotated[0];
4182 recent_rotated[1] +=
4183 mz->reclaim_stat.recent_rotated[1];
4184 recent_scanned[0] +=
4185 mz->reclaim_stat.recent_scanned[0];
4186 recent_scanned[1] +=
4187 mz->reclaim_stat.recent_scanned[1];
4189 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
4190 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
4191 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
4192 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
4199 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
4201 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4203 return mem_cgroup_swappiness(memcg);
4206 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
4209 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4210 struct mem_cgroup *parent;
4215 if (cgrp->parent == NULL)
4218 parent = mem_cgroup_from_cont(cgrp->parent);
4222 /* If under hierarchy, only empty-root can set this value */
4223 if ((parent->use_hierarchy) ||
4224 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4229 memcg->swappiness = val;
4236 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4238 struct mem_cgroup_threshold_ary *t;
4244 t = rcu_dereference(memcg->thresholds.primary);
4246 t = rcu_dereference(memcg->memsw_thresholds.primary);
4251 usage = mem_cgroup_usage(memcg, swap);
4254 * current_threshold points to threshold just below usage.
4255 * If it's not true, a threshold was crossed after last
4256 * call of __mem_cgroup_threshold().
4258 i = t->current_threshold;
4261 * Iterate backward over array of thresholds starting from
4262 * current_threshold and check if a threshold is crossed.
4263 * If none of thresholds below usage is crossed, we read
4264 * only one element of the array here.
4266 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4267 eventfd_signal(t->entries[i].eventfd, 1);
4269 /* i = current_threshold + 1 */
4273 * Iterate forward over array of thresholds starting from
4274 * current_threshold+1 and check if a threshold is crossed.
4275 * If none of thresholds above usage is crossed, we read
4276 * only one element of the array here.
4278 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4279 eventfd_signal(t->entries[i].eventfd, 1);
4281 /* Update current_threshold */
4282 t->current_threshold = i - 1;
4287 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4290 __mem_cgroup_threshold(memcg, false);
4291 if (do_swap_account)
4292 __mem_cgroup_threshold(memcg, true);
4294 memcg = parent_mem_cgroup(memcg);
4298 static int compare_thresholds(const void *a, const void *b)
4300 const struct mem_cgroup_threshold *_a = a;
4301 const struct mem_cgroup_threshold *_b = b;
4303 return _a->threshold - _b->threshold;
4306 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4308 struct mem_cgroup_eventfd_list *ev;
4310 list_for_each_entry(ev, &memcg->oom_notify, list)
4311 eventfd_signal(ev->eventfd, 1);
4315 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4317 struct mem_cgroup *iter;
4319 for_each_mem_cgroup_tree(iter, memcg)
4320 mem_cgroup_oom_notify_cb(iter);
4323 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
4324 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4326 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4327 struct mem_cgroup_thresholds *thresholds;
4328 struct mem_cgroup_threshold_ary *new;
4329 int type = MEMFILE_TYPE(cft->private);
4330 u64 threshold, usage;
4333 ret = res_counter_memparse_write_strategy(args, &threshold);
4337 mutex_lock(&memcg->thresholds_lock);
4340 thresholds = &memcg->thresholds;
4341 else if (type == _MEMSWAP)
4342 thresholds = &memcg->memsw_thresholds;
4346 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4348 /* Check if a threshold crossed before adding a new one */
4349 if (thresholds->primary)
4350 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4352 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4354 /* Allocate memory for new array of thresholds */
4355 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4363 /* Copy thresholds (if any) to new array */
4364 if (thresholds->primary) {
4365 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4366 sizeof(struct mem_cgroup_threshold));
4369 /* Add new threshold */
4370 new->entries[size - 1].eventfd = eventfd;
4371 new->entries[size - 1].threshold = threshold;
4373 /* Sort thresholds. Registering of new threshold isn't time-critical */
4374 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4375 compare_thresholds, NULL);
4377 /* Find current threshold */
4378 new->current_threshold = -1;
4379 for (i = 0; i < size; i++) {
4380 if (new->entries[i].threshold < usage) {
4382 * new->current_threshold will not be used until
4383 * rcu_assign_pointer(), so it's safe to increment
4386 ++new->current_threshold;
4390 /* Free old spare buffer and save old primary buffer as spare */
4391 kfree(thresholds->spare);
4392 thresholds->spare = thresholds->primary;
4394 rcu_assign_pointer(thresholds->primary, new);
4396 /* To be sure that nobody uses thresholds */
4400 mutex_unlock(&memcg->thresholds_lock);
4405 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
4406 struct cftype *cft, struct eventfd_ctx *eventfd)
4408 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4409 struct mem_cgroup_thresholds *thresholds;
4410 struct mem_cgroup_threshold_ary *new;
4411 int type = MEMFILE_TYPE(cft->private);
4415 mutex_lock(&memcg->thresholds_lock);
4417 thresholds = &memcg->thresholds;
4418 else if (type == _MEMSWAP)
4419 thresholds = &memcg->memsw_thresholds;
4424 * Something went wrong if we trying to unregister a threshold
4425 * if we don't have thresholds
4427 BUG_ON(!thresholds);
4429 if (!thresholds->primary)
4432 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4434 /* Check if a threshold crossed before removing */
4435 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4437 /* Calculate new number of threshold */
4439 for (i = 0; i < thresholds->primary->size; i++) {
4440 if (thresholds->primary->entries[i].eventfd != eventfd)
4444 new = thresholds->spare;
4446 /* Set thresholds array to NULL if we don't have thresholds */
4455 /* Copy thresholds and find current threshold */
4456 new->current_threshold = -1;
4457 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4458 if (thresholds->primary->entries[i].eventfd == eventfd)
4461 new->entries[j] = thresholds->primary->entries[i];
4462 if (new->entries[j].threshold < usage) {
4464 * new->current_threshold will not be used
4465 * until rcu_assign_pointer(), so it's safe to increment
4468 ++new->current_threshold;
4474 /* Swap primary and spare array */
4475 thresholds->spare = thresholds->primary;
4476 rcu_assign_pointer(thresholds->primary, new);
4478 /* To be sure that nobody uses thresholds */
4481 mutex_unlock(&memcg->thresholds_lock);
4484 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4485 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4487 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4488 struct mem_cgroup_eventfd_list *event;
4489 int type = MEMFILE_TYPE(cft->private);
4491 BUG_ON(type != _OOM_TYPE);
4492 event = kmalloc(sizeof(*event), GFP_KERNEL);
4496 spin_lock(&memcg_oom_lock);
4498 event->eventfd = eventfd;
4499 list_add(&event->list, &memcg->oom_notify);
4501 /* already in OOM ? */
4502 if (atomic_read(&memcg->under_oom))
4503 eventfd_signal(eventfd, 1);
4504 spin_unlock(&memcg_oom_lock);
4509 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4510 struct cftype *cft, struct eventfd_ctx *eventfd)
4512 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4513 struct mem_cgroup_eventfd_list *ev, *tmp;
4514 int type = MEMFILE_TYPE(cft->private);
4516 BUG_ON(type != _OOM_TYPE);
4518 spin_lock(&memcg_oom_lock);
4520 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4521 if (ev->eventfd == eventfd) {
4522 list_del(&ev->list);
4527 spin_unlock(&memcg_oom_lock);
4530 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4531 struct cftype *cft, struct cgroup_map_cb *cb)
4533 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4535 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
4537 if (atomic_read(&memcg->under_oom))
4538 cb->fill(cb, "under_oom", 1);
4540 cb->fill(cb, "under_oom", 0);
4544 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4545 struct cftype *cft, u64 val)
4547 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4548 struct mem_cgroup *parent;
4550 /* cannot set to root cgroup and only 0 and 1 are allowed */
4551 if (!cgrp->parent || !((val == 0) || (val == 1)))
4554 parent = mem_cgroup_from_cont(cgrp->parent);
4557 /* oom-kill-disable is a flag for subhierarchy. */
4558 if ((parent->use_hierarchy) ||
4559 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4563 memcg->oom_kill_disable = val;
4565 memcg_oom_recover(memcg);
4571 static const struct file_operations mem_control_numa_stat_file_operations = {
4573 .llseek = seq_lseek,
4574 .release = single_release,
4577 static int mem_control_numa_stat_open(struct inode *unused, struct file *file)
4579 struct cgroup *cont = file->f_dentry->d_parent->d_fsdata;
4581 file->f_op = &mem_control_numa_stat_file_operations;
4582 return single_open(file, mem_control_numa_stat_show, cont);
4584 #endif /* CONFIG_NUMA */
4586 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_KMEM
4587 static int register_kmem_files(struct cgroup *cont, struct cgroup_subsys *ss)
4590 * Part of this would be better living in a separate allocation
4591 * function, leaving us with just the cgroup tree population work.
4592 * We, however, depend on state such as network's proto_list that
4593 * is only initialized after cgroup creation. I found the less
4594 * cumbersome way to deal with it to defer it all to populate time
4596 return mem_cgroup_sockets_init(cont, ss);
4599 static void kmem_cgroup_destroy(struct cgroup *cont)
4601 mem_cgroup_sockets_destroy(cont);
4604 static int register_kmem_files(struct cgroup *cont, struct cgroup_subsys *ss)
4609 static void kmem_cgroup_destroy(struct cgroup *cont)
4614 static struct cftype mem_cgroup_files[] = {
4616 .name = "usage_in_bytes",
4617 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4618 .read_u64 = mem_cgroup_read,
4619 .register_event = mem_cgroup_usage_register_event,
4620 .unregister_event = mem_cgroup_usage_unregister_event,
4623 .name = "max_usage_in_bytes",
4624 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4625 .trigger = mem_cgroup_reset,
4626 .read_u64 = mem_cgroup_read,
4629 .name = "limit_in_bytes",
4630 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4631 .write_string = mem_cgroup_write,
4632 .read_u64 = mem_cgroup_read,
4635 .name = "soft_limit_in_bytes",
4636 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4637 .write_string = mem_cgroup_write,
4638 .read_u64 = mem_cgroup_read,
4642 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4643 .trigger = mem_cgroup_reset,
4644 .read_u64 = mem_cgroup_read,
4648 .read_map = mem_control_stat_show,
4651 .name = "force_empty",
4652 .trigger = mem_cgroup_force_empty_write,
4655 .name = "use_hierarchy",
4656 .write_u64 = mem_cgroup_hierarchy_write,
4657 .read_u64 = mem_cgroup_hierarchy_read,
4660 .name = "swappiness",
4661 .read_u64 = mem_cgroup_swappiness_read,
4662 .write_u64 = mem_cgroup_swappiness_write,
4665 .name = "move_charge_at_immigrate",
4666 .read_u64 = mem_cgroup_move_charge_read,
4667 .write_u64 = mem_cgroup_move_charge_write,
4670 .name = "oom_control",
4671 .read_map = mem_cgroup_oom_control_read,
4672 .write_u64 = mem_cgroup_oom_control_write,
4673 .register_event = mem_cgroup_oom_register_event,
4674 .unregister_event = mem_cgroup_oom_unregister_event,
4675 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4679 .name = "numa_stat",
4680 .open = mem_control_numa_stat_open,
4686 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4687 static struct cftype memsw_cgroup_files[] = {
4689 .name = "memsw.usage_in_bytes",
4690 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4691 .read_u64 = mem_cgroup_read,
4692 .register_event = mem_cgroup_usage_register_event,
4693 .unregister_event = mem_cgroup_usage_unregister_event,
4696 .name = "memsw.max_usage_in_bytes",
4697 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4698 .trigger = mem_cgroup_reset,
4699 .read_u64 = mem_cgroup_read,
4702 .name = "memsw.limit_in_bytes",
4703 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4704 .write_string = mem_cgroup_write,
4705 .read_u64 = mem_cgroup_read,
4708 .name = "memsw.failcnt",
4709 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4710 .trigger = mem_cgroup_reset,
4711 .read_u64 = mem_cgroup_read,
4715 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4717 if (!do_swap_account)
4719 return cgroup_add_files(cont, ss, memsw_cgroup_files,
4720 ARRAY_SIZE(memsw_cgroup_files));
4723 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4729 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4731 struct mem_cgroup_per_node *pn;
4732 struct mem_cgroup_per_zone *mz;
4734 int zone, tmp = node;
4736 * This routine is called against possible nodes.
4737 * But it's BUG to call kmalloc() against offline node.
4739 * TODO: this routine can waste much memory for nodes which will
4740 * never be onlined. It's better to use memory hotplug callback
4743 if (!node_state(node, N_NORMAL_MEMORY))
4745 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4749 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4750 mz = &pn->zoneinfo[zone];
4752 INIT_LIST_HEAD(&mz->lruvec.lists[lru]);
4753 mz->usage_in_excess = 0;
4754 mz->on_tree = false;
4757 memcg->info.nodeinfo[node] = pn;
4761 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4763 kfree(memcg->info.nodeinfo[node]);
4766 static struct mem_cgroup *mem_cgroup_alloc(void)
4768 struct mem_cgroup *memcg;
4769 int size = sizeof(struct mem_cgroup);
4771 /* Can be very big if MAX_NUMNODES is very big */
4772 if (size < PAGE_SIZE)
4773 memcg = kzalloc(size, GFP_KERNEL);
4775 memcg = vzalloc(size);
4780 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4783 spin_lock_init(&memcg->pcp_counter_lock);
4787 if (size < PAGE_SIZE)
4795 * Helpers for freeing a vzalloc()ed mem_cgroup by RCU,
4796 * but in process context. The work_freeing structure is overlaid
4797 * on the rcu_freeing structure, which itself is overlaid on memsw.
4799 static void vfree_work(struct work_struct *work)
4801 struct mem_cgroup *memcg;
4803 memcg = container_of(work, struct mem_cgroup, work_freeing);
4806 static void vfree_rcu(struct rcu_head *rcu_head)
4808 struct mem_cgroup *memcg;
4810 memcg = container_of(rcu_head, struct mem_cgroup, rcu_freeing);
4811 INIT_WORK(&memcg->work_freeing, vfree_work);
4812 schedule_work(&memcg->work_freeing);
4816 * At destroying mem_cgroup, references from swap_cgroup can remain.
4817 * (scanning all at force_empty is too costly...)
4819 * Instead of clearing all references at force_empty, we remember
4820 * the number of reference from swap_cgroup and free mem_cgroup when
4821 * it goes down to 0.
4823 * Removal of cgroup itself succeeds regardless of refs from swap.
4826 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4830 mem_cgroup_remove_from_trees(memcg);
4831 free_css_id(&mem_cgroup_subsys, &memcg->css);
4834 free_mem_cgroup_per_zone_info(memcg, node);
4836 free_percpu(memcg->stat);
4837 if (sizeof(struct mem_cgroup) < PAGE_SIZE)
4838 kfree_rcu(memcg, rcu_freeing);
4840 call_rcu(&memcg->rcu_freeing, vfree_rcu);
4843 static void mem_cgroup_get(struct mem_cgroup *memcg)
4845 atomic_inc(&memcg->refcnt);
4848 static void __mem_cgroup_put(struct mem_cgroup *memcg, int count)
4850 if (atomic_sub_and_test(count, &memcg->refcnt)) {
4851 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4852 __mem_cgroup_free(memcg);
4854 mem_cgroup_put(parent);
4858 static void mem_cgroup_put(struct mem_cgroup *memcg)
4860 __mem_cgroup_put(memcg, 1);
4864 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4866 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4868 if (!memcg->res.parent)
4870 return mem_cgroup_from_res_counter(memcg->res.parent, res);
4872 EXPORT_SYMBOL(parent_mem_cgroup);
4874 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4875 static void __init enable_swap_cgroup(void)
4877 if (!mem_cgroup_disabled() && really_do_swap_account)
4878 do_swap_account = 1;
4881 static void __init enable_swap_cgroup(void)
4886 static int mem_cgroup_soft_limit_tree_init(void)
4888 struct mem_cgroup_tree_per_node *rtpn;
4889 struct mem_cgroup_tree_per_zone *rtpz;
4890 int tmp, node, zone;
4892 for_each_node(node) {
4894 if (!node_state(node, N_NORMAL_MEMORY))
4896 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
4900 soft_limit_tree.rb_tree_per_node[node] = rtpn;
4902 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4903 rtpz = &rtpn->rb_tree_per_zone[zone];
4904 rtpz->rb_root = RB_ROOT;
4905 spin_lock_init(&rtpz->lock);
4911 for_each_node(node) {
4912 if (!soft_limit_tree.rb_tree_per_node[node])
4914 kfree(soft_limit_tree.rb_tree_per_node[node]);
4915 soft_limit_tree.rb_tree_per_node[node] = NULL;
4921 static struct cgroup_subsys_state * __ref
4922 mem_cgroup_create(struct cgroup *cont)
4924 struct mem_cgroup *memcg, *parent;
4925 long error = -ENOMEM;
4928 memcg = mem_cgroup_alloc();
4930 return ERR_PTR(error);
4933 if (alloc_mem_cgroup_per_zone_info(memcg, node))
4937 if (cont->parent == NULL) {
4939 enable_swap_cgroup();
4941 if (mem_cgroup_soft_limit_tree_init())
4943 root_mem_cgroup = memcg;
4944 for_each_possible_cpu(cpu) {
4945 struct memcg_stock_pcp *stock =
4946 &per_cpu(memcg_stock, cpu);
4947 INIT_WORK(&stock->work, drain_local_stock);
4949 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
4951 parent = mem_cgroup_from_cont(cont->parent);
4952 memcg->use_hierarchy = parent->use_hierarchy;
4953 memcg->oom_kill_disable = parent->oom_kill_disable;
4956 if (parent && parent->use_hierarchy) {
4957 res_counter_init(&memcg->res, &parent->res);
4958 res_counter_init(&memcg->memsw, &parent->memsw);
4960 * We increment refcnt of the parent to ensure that we can
4961 * safely access it on res_counter_charge/uncharge.
4962 * This refcnt will be decremented when freeing this
4963 * mem_cgroup(see mem_cgroup_put).
4965 mem_cgroup_get(parent);
4967 res_counter_init(&memcg->res, NULL);
4968 res_counter_init(&memcg->memsw, NULL);
4970 memcg->last_scanned_node = MAX_NUMNODES;
4971 INIT_LIST_HEAD(&memcg->oom_notify);
4974 memcg->swappiness = mem_cgroup_swappiness(parent);
4975 atomic_set(&memcg->refcnt, 1);
4976 memcg->move_charge_at_immigrate = 0;
4977 mutex_init(&memcg->thresholds_lock);
4980 __mem_cgroup_free(memcg);
4981 return ERR_PTR(error);
4984 static int mem_cgroup_pre_destroy(struct cgroup *cont)
4986 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4988 return mem_cgroup_force_empty(memcg, false);
4991 static void mem_cgroup_destroy(struct cgroup *cont)
4993 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4995 kmem_cgroup_destroy(cont);
4997 mem_cgroup_put(memcg);
5000 static int mem_cgroup_populate(struct cgroup_subsys *ss,
5001 struct cgroup *cont)
5005 ret = cgroup_add_files(cont, ss, mem_cgroup_files,
5006 ARRAY_SIZE(mem_cgroup_files));
5009 ret = register_memsw_files(cont, ss);
5012 ret = register_kmem_files(cont, ss);
5018 /* Handlers for move charge at task migration. */
5019 #define PRECHARGE_COUNT_AT_ONCE 256
5020 static int mem_cgroup_do_precharge(unsigned long count)
5023 int batch_count = PRECHARGE_COUNT_AT_ONCE;
5024 struct mem_cgroup *memcg = mc.to;
5026 if (mem_cgroup_is_root(memcg)) {
5027 mc.precharge += count;
5028 /* we don't need css_get for root */
5031 /* try to charge at once */
5033 struct res_counter *dummy;
5035 * "memcg" cannot be under rmdir() because we've already checked
5036 * by cgroup_lock_live_cgroup() that it is not removed and we
5037 * are still under the same cgroup_mutex. So we can postpone
5040 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
5042 if (do_swap_account && res_counter_charge(&memcg->memsw,
5043 PAGE_SIZE * count, &dummy)) {
5044 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
5047 mc.precharge += count;
5051 /* fall back to one by one charge */
5053 if (signal_pending(current)) {
5057 if (!batch_count--) {
5058 batch_count = PRECHARGE_COUNT_AT_ONCE;
5061 ret = __mem_cgroup_try_charge(NULL,
5062 GFP_KERNEL, 1, &memcg, false);
5064 /* mem_cgroup_clear_mc() will do uncharge later */
5072 * is_target_pte_for_mc - check a pte whether it is valid for move charge
5073 * @vma: the vma the pte to be checked belongs
5074 * @addr: the address corresponding to the pte to be checked
5075 * @ptent: the pte to be checked
5076 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5079 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5080 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5081 * move charge. if @target is not NULL, the page is stored in target->page
5082 * with extra refcnt got(Callers should handle it).
5083 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5084 * target for charge migration. if @target is not NULL, the entry is stored
5087 * Called with pte lock held.
5094 enum mc_target_type {
5095 MC_TARGET_NONE, /* not used */
5100 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5101 unsigned long addr, pte_t ptent)
5103 struct page *page = vm_normal_page(vma, addr, ptent);
5105 if (!page || !page_mapped(page))
5107 if (PageAnon(page)) {
5108 /* we don't move shared anon */
5109 if (!move_anon() || page_mapcount(page) > 2)
5111 } else if (!move_file())
5112 /* we ignore mapcount for file pages */
5114 if (!get_page_unless_zero(page))
5120 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5121 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5124 struct page *page = NULL;
5125 swp_entry_t ent = pte_to_swp_entry(ptent);
5127 if (!move_anon() || non_swap_entry(ent))
5129 usage_count = mem_cgroup_count_swap_user(ent, &page);
5130 if (usage_count > 1) { /* we don't move shared anon */
5135 if (do_swap_account)
5136 entry->val = ent.val;
5141 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5142 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5144 struct page *page = NULL;
5145 struct inode *inode;
5146 struct address_space *mapping;
5149 if (!vma->vm_file) /* anonymous vma */
5154 inode = vma->vm_file->f_path.dentry->d_inode;
5155 mapping = vma->vm_file->f_mapping;
5156 if (pte_none(ptent))
5157 pgoff = linear_page_index(vma, addr);
5158 else /* pte_file(ptent) is true */
5159 pgoff = pte_to_pgoff(ptent);
5161 /* page is moved even if it's not RSS of this task(page-faulted). */
5162 page = find_get_page(mapping, pgoff);
5165 /* shmem/tmpfs may report page out on swap: account for that too. */
5166 if (radix_tree_exceptional_entry(page)) {
5167 swp_entry_t swap = radix_to_swp_entry(page);
5168 if (do_swap_account)
5170 page = find_get_page(&swapper_space, swap.val);
5176 static int is_target_pte_for_mc(struct vm_area_struct *vma,
5177 unsigned long addr, pte_t ptent, union mc_target *target)
5179 struct page *page = NULL;
5180 struct page_cgroup *pc;
5182 swp_entry_t ent = { .val = 0 };
5184 if (pte_present(ptent))
5185 page = mc_handle_present_pte(vma, addr, ptent);
5186 else if (is_swap_pte(ptent))
5187 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5188 else if (pte_none(ptent) || pte_file(ptent))
5189 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5191 if (!page && !ent.val)
5194 pc = lookup_page_cgroup(page);
5196 * Do only loose check w/o page_cgroup lock.
5197 * mem_cgroup_move_account() checks the pc is valid or not under
5200 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5201 ret = MC_TARGET_PAGE;
5203 target->page = page;
5205 if (!ret || !target)
5208 /* There is a swap entry and a page doesn't exist or isn't charged */
5209 if (ent.val && !ret &&
5210 css_id(&mc.from->css) == lookup_swap_cgroup_id(ent)) {
5211 ret = MC_TARGET_SWAP;
5218 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5219 unsigned long addr, unsigned long end,
5220 struct mm_walk *walk)
5222 struct vm_area_struct *vma = walk->private;
5226 split_huge_page_pmd(walk->mm, pmd);
5227 if (pmd_trans_unstable(pmd))
5230 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5231 for (; addr != end; pte++, addr += PAGE_SIZE)
5232 if (is_target_pte_for_mc(vma, addr, *pte, NULL))
5233 mc.precharge++; /* increment precharge temporarily */
5234 pte_unmap_unlock(pte - 1, ptl);
5240 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5242 unsigned long precharge;
5243 struct vm_area_struct *vma;
5245 down_read(&mm->mmap_sem);
5246 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5247 struct mm_walk mem_cgroup_count_precharge_walk = {
5248 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5252 if (is_vm_hugetlb_page(vma))
5254 walk_page_range(vma->vm_start, vma->vm_end,
5255 &mem_cgroup_count_precharge_walk);
5257 up_read(&mm->mmap_sem);
5259 precharge = mc.precharge;
5265 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5267 unsigned long precharge = mem_cgroup_count_precharge(mm);
5269 VM_BUG_ON(mc.moving_task);
5270 mc.moving_task = current;
5271 return mem_cgroup_do_precharge(precharge);
5274 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5275 static void __mem_cgroup_clear_mc(void)
5277 struct mem_cgroup *from = mc.from;
5278 struct mem_cgroup *to = mc.to;
5280 /* we must uncharge all the leftover precharges from mc.to */
5282 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
5286 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5287 * we must uncharge here.
5289 if (mc.moved_charge) {
5290 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
5291 mc.moved_charge = 0;
5293 /* we must fixup refcnts and charges */
5294 if (mc.moved_swap) {
5295 /* uncharge swap account from the old cgroup */
5296 if (!mem_cgroup_is_root(mc.from))
5297 res_counter_uncharge(&mc.from->memsw,
5298 PAGE_SIZE * mc.moved_swap);
5299 __mem_cgroup_put(mc.from, mc.moved_swap);
5301 if (!mem_cgroup_is_root(mc.to)) {
5303 * we charged both to->res and to->memsw, so we should
5306 res_counter_uncharge(&mc.to->res,
5307 PAGE_SIZE * mc.moved_swap);
5309 /* we've already done mem_cgroup_get(mc.to) */
5312 memcg_oom_recover(from);
5313 memcg_oom_recover(to);
5314 wake_up_all(&mc.waitq);
5317 static void mem_cgroup_clear_mc(void)
5319 struct mem_cgroup *from = mc.from;
5322 * we must clear moving_task before waking up waiters at the end of
5325 mc.moving_task = NULL;
5326 __mem_cgroup_clear_mc();
5327 spin_lock(&mc.lock);
5330 spin_unlock(&mc.lock);
5331 mem_cgroup_end_move(from);
5334 static int mem_cgroup_can_attach(struct cgroup *cgroup,
5335 struct cgroup_taskset *tset)
5337 struct task_struct *p = cgroup_taskset_first(tset);
5339 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup);
5341 if (memcg->move_charge_at_immigrate) {
5342 struct mm_struct *mm;
5343 struct mem_cgroup *from = mem_cgroup_from_task(p);
5345 VM_BUG_ON(from == memcg);
5347 mm = get_task_mm(p);
5350 /* We move charges only when we move a owner of the mm */
5351 if (mm->owner == p) {
5354 VM_BUG_ON(mc.precharge);
5355 VM_BUG_ON(mc.moved_charge);
5356 VM_BUG_ON(mc.moved_swap);
5357 mem_cgroup_start_move(from);
5358 spin_lock(&mc.lock);
5361 spin_unlock(&mc.lock);
5362 /* We set mc.moving_task later */
5364 ret = mem_cgroup_precharge_mc(mm);
5366 mem_cgroup_clear_mc();
5373 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
5374 struct cgroup_taskset *tset)
5376 mem_cgroup_clear_mc();
5379 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5380 unsigned long addr, unsigned long end,
5381 struct mm_walk *walk)
5384 struct vm_area_struct *vma = walk->private;
5388 split_huge_page_pmd(walk->mm, pmd);
5389 if (pmd_trans_unstable(pmd))
5392 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5393 for (; addr != end; addr += PAGE_SIZE) {
5394 pte_t ptent = *(pte++);
5395 union mc_target target;
5398 struct page_cgroup *pc;
5404 type = is_target_pte_for_mc(vma, addr, ptent, &target);
5406 case MC_TARGET_PAGE:
5408 if (isolate_lru_page(page))
5410 pc = lookup_page_cgroup(page);
5411 if (!mem_cgroup_move_account(page, 1, pc,
5412 mc.from, mc.to, false)) {
5414 /* we uncharge from mc.from later. */
5417 putback_lru_page(page);
5418 put: /* is_target_pte_for_mc() gets the page */
5421 case MC_TARGET_SWAP:
5423 if (!mem_cgroup_move_swap_account(ent,
5424 mc.from, mc.to, false)) {
5426 /* we fixup refcnts and charges later. */
5434 pte_unmap_unlock(pte - 1, ptl);
5439 * We have consumed all precharges we got in can_attach().
5440 * We try charge one by one, but don't do any additional
5441 * charges to mc.to if we have failed in charge once in attach()
5444 ret = mem_cgroup_do_precharge(1);
5452 static void mem_cgroup_move_charge(struct mm_struct *mm)
5454 struct vm_area_struct *vma;
5456 lru_add_drain_all();
5458 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5460 * Someone who are holding the mmap_sem might be waiting in
5461 * waitq. So we cancel all extra charges, wake up all waiters,
5462 * and retry. Because we cancel precharges, we might not be able
5463 * to move enough charges, but moving charge is a best-effort
5464 * feature anyway, so it wouldn't be a big problem.
5466 __mem_cgroup_clear_mc();
5470 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5472 struct mm_walk mem_cgroup_move_charge_walk = {
5473 .pmd_entry = mem_cgroup_move_charge_pte_range,
5477 if (is_vm_hugetlb_page(vma))
5479 ret = walk_page_range(vma->vm_start, vma->vm_end,
5480 &mem_cgroup_move_charge_walk);
5483 * means we have consumed all precharges and failed in
5484 * doing additional charge. Just abandon here.
5488 up_read(&mm->mmap_sem);
5491 static void mem_cgroup_move_task(struct cgroup *cont,
5492 struct cgroup_taskset *tset)
5494 struct task_struct *p = cgroup_taskset_first(tset);
5495 struct mm_struct *mm = get_task_mm(p);
5499 mem_cgroup_move_charge(mm);
5504 mem_cgroup_clear_mc();
5506 #else /* !CONFIG_MMU */
5507 static int mem_cgroup_can_attach(struct cgroup *cgroup,
5508 struct cgroup_taskset *tset)
5512 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
5513 struct cgroup_taskset *tset)
5516 static void mem_cgroup_move_task(struct cgroup *cont,
5517 struct cgroup_taskset *tset)
5522 struct cgroup_subsys mem_cgroup_subsys = {
5524 .subsys_id = mem_cgroup_subsys_id,
5525 .create = mem_cgroup_create,
5526 .pre_destroy = mem_cgroup_pre_destroy,
5527 .destroy = mem_cgroup_destroy,
5528 .populate = mem_cgroup_populate,
5529 .can_attach = mem_cgroup_can_attach,
5530 .cancel_attach = mem_cgroup_cancel_attach,
5531 .attach = mem_cgroup_move_task,
5536 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5537 static int __init enable_swap_account(char *s)
5539 /* consider enabled if no parameter or 1 is given */
5540 if (!strcmp(s, "1"))
5541 really_do_swap_account = 1;
5542 else if (!strcmp(s, "0"))
5543 really_do_swap_account = 0;
5546 __setup("swapaccount=", enable_swap_account);