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 anon, int nr_pages)
698 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
699 * counted as CACHE even if it's on ANON LRU.
702 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
705 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
708 /* pagein of a big page is an event. So, ignore page size */
710 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
712 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
713 nr_pages = -nr_pages; /* for event */
716 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT], nr_pages);
722 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
723 unsigned int lru_mask)
725 struct mem_cgroup_per_zone *mz;
727 unsigned long ret = 0;
729 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
732 if (BIT(lru) & lru_mask)
733 ret += mz->lru_size[lru];
739 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
740 int nid, unsigned int lru_mask)
745 for (zid = 0; zid < MAX_NR_ZONES; zid++)
746 total += mem_cgroup_zone_nr_lru_pages(memcg,
752 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
753 unsigned int lru_mask)
758 for_each_node_state(nid, N_HIGH_MEMORY)
759 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
763 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
764 enum mem_cgroup_events_target target)
766 unsigned long val, next;
768 val = __this_cpu_read(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT]);
769 next = __this_cpu_read(memcg->stat->targets[target]);
770 /* from time_after() in jiffies.h */
771 if ((long)next - (long)val < 0) {
773 case MEM_CGROUP_TARGET_THRESH:
774 next = val + THRESHOLDS_EVENTS_TARGET;
776 case MEM_CGROUP_TARGET_SOFTLIMIT:
777 next = val + SOFTLIMIT_EVENTS_TARGET;
779 case MEM_CGROUP_TARGET_NUMAINFO:
780 next = val + NUMAINFO_EVENTS_TARGET;
785 __this_cpu_write(memcg->stat->targets[target], next);
792 * Check events in order.
795 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
798 /* threshold event is triggered in finer grain than soft limit */
799 if (unlikely(mem_cgroup_event_ratelimit(memcg,
800 MEM_CGROUP_TARGET_THRESH))) {
802 bool do_numainfo __maybe_unused;
804 do_softlimit = mem_cgroup_event_ratelimit(memcg,
805 MEM_CGROUP_TARGET_SOFTLIMIT);
807 do_numainfo = mem_cgroup_event_ratelimit(memcg,
808 MEM_CGROUP_TARGET_NUMAINFO);
812 mem_cgroup_threshold(memcg);
813 if (unlikely(do_softlimit))
814 mem_cgroup_update_tree(memcg, page);
816 if (unlikely(do_numainfo))
817 atomic_inc(&memcg->numainfo_events);
823 struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
825 return container_of(cgroup_subsys_state(cont,
826 mem_cgroup_subsys_id), struct mem_cgroup,
830 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
833 * mm_update_next_owner() may clear mm->owner to NULL
834 * if it races with swapoff, page migration, etc.
835 * So this can be called with p == NULL.
840 return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
841 struct mem_cgroup, css);
844 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
846 struct mem_cgroup *memcg = NULL;
851 * Because we have no locks, mm->owner's may be being moved to other
852 * cgroup. We use css_tryget() here even if this looks
853 * pessimistic (rather than adding locks here).
857 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
858 if (unlikely(!memcg))
860 } while (!css_tryget(&memcg->css));
866 * mem_cgroup_iter - iterate over memory cgroup hierarchy
867 * @root: hierarchy root
868 * @prev: previously returned memcg, NULL on first invocation
869 * @reclaim: cookie for shared reclaim walks, NULL for full walks
871 * Returns references to children of the hierarchy below @root, or
872 * @root itself, or %NULL after a full round-trip.
874 * Caller must pass the return value in @prev on subsequent
875 * invocations for reference counting, or use mem_cgroup_iter_break()
876 * to cancel a hierarchy walk before the round-trip is complete.
878 * Reclaimers can specify a zone and a priority level in @reclaim to
879 * divide up the memcgs in the hierarchy among all concurrent
880 * reclaimers operating on the same zone and priority.
882 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
883 struct mem_cgroup *prev,
884 struct mem_cgroup_reclaim_cookie *reclaim)
886 struct mem_cgroup *memcg = NULL;
889 if (mem_cgroup_disabled())
893 root = root_mem_cgroup;
895 if (prev && !reclaim)
896 id = css_id(&prev->css);
898 if (prev && prev != root)
901 if (!root->use_hierarchy && root != root_mem_cgroup) {
908 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
909 struct cgroup_subsys_state *css;
912 int nid = zone_to_nid(reclaim->zone);
913 int zid = zone_idx(reclaim->zone);
914 struct mem_cgroup_per_zone *mz;
916 mz = mem_cgroup_zoneinfo(root, nid, zid);
917 iter = &mz->reclaim_iter[reclaim->priority];
918 if (prev && reclaim->generation != iter->generation)
924 css = css_get_next(&mem_cgroup_subsys, id + 1, &root->css, &id);
926 if (css == &root->css || css_tryget(css))
927 memcg = container_of(css,
928 struct mem_cgroup, css);
937 else if (!prev && memcg)
938 reclaim->generation = iter->generation;
948 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
949 * @root: hierarchy root
950 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
952 void mem_cgroup_iter_break(struct mem_cgroup *root,
953 struct mem_cgroup *prev)
956 root = root_mem_cgroup;
957 if (prev && prev != root)
962 * Iteration constructs for visiting all cgroups (under a tree). If
963 * loops are exited prematurely (break), mem_cgroup_iter_break() must
964 * be used for reference counting.
966 #define for_each_mem_cgroup_tree(iter, root) \
967 for (iter = mem_cgroup_iter(root, NULL, NULL); \
969 iter = mem_cgroup_iter(root, iter, NULL))
971 #define for_each_mem_cgroup(iter) \
972 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
974 iter = mem_cgroup_iter(NULL, iter, NULL))
976 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
978 return (memcg == root_mem_cgroup);
981 void mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
983 struct mem_cgroup *memcg;
989 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
990 if (unlikely(!memcg))
995 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
998 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1006 EXPORT_SYMBOL(mem_cgroup_count_vm_event);
1009 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1010 * @zone: zone of the wanted lruvec
1011 * @mem: memcg of the wanted lruvec
1013 * Returns the lru list vector holding pages for the given @zone and
1014 * @mem. This can be the global zone lruvec, if the memory controller
1017 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1018 struct mem_cgroup *memcg)
1020 struct mem_cgroup_per_zone *mz;
1022 if (mem_cgroup_disabled())
1023 return &zone->lruvec;
1025 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1030 * Following LRU functions are allowed to be used without PCG_LOCK.
1031 * Operations are called by routine of global LRU independently from memcg.
1032 * What we have to take care of here is validness of pc->mem_cgroup.
1034 * Changes to pc->mem_cgroup happens when
1037 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1038 * It is added to LRU before charge.
1039 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1040 * When moving account, the page is not on LRU. It's isolated.
1044 * mem_cgroup_lru_add_list - account for adding an lru page and return lruvec
1045 * @zone: zone of the page
1049 * This function accounts for @page being added to @lru, and returns
1050 * the lruvec for the given @zone and the memcg @page is charged to.
1052 * The callsite is then responsible for physically linking the page to
1053 * the returned lruvec->lists[@lru].
1055 struct lruvec *mem_cgroup_lru_add_list(struct zone *zone, struct page *page,
1058 struct mem_cgroup_per_zone *mz;
1059 struct mem_cgroup *memcg;
1060 struct page_cgroup *pc;
1062 if (mem_cgroup_disabled())
1063 return &zone->lruvec;
1065 pc = lookup_page_cgroup(page);
1066 memcg = pc->mem_cgroup;
1069 * Surreptitiously switch any uncharged page to root:
1070 * an uncharged page off lru does nothing to secure
1071 * its former mem_cgroup from sudden removal.
1073 * Our caller holds lru_lock, and PageCgroupUsed is updated
1074 * under page_cgroup lock: between them, they make all uses
1075 * of pc->mem_cgroup safe.
1077 if (!PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1078 pc->mem_cgroup = memcg = root_mem_cgroup;
1080 mz = page_cgroup_zoneinfo(memcg, page);
1081 /* compound_order() is stabilized through lru_lock */
1082 mz->lru_size[lru] += 1 << compound_order(page);
1087 * mem_cgroup_lru_del_list - account for removing an lru page
1091 * This function accounts for @page being removed from @lru.
1093 * The callsite is then responsible for physically unlinking
1096 void mem_cgroup_lru_del_list(struct page *page, enum lru_list lru)
1098 struct mem_cgroup_per_zone *mz;
1099 struct mem_cgroup *memcg;
1100 struct page_cgroup *pc;
1102 if (mem_cgroup_disabled())
1105 pc = lookup_page_cgroup(page);
1106 memcg = pc->mem_cgroup;
1108 mz = page_cgroup_zoneinfo(memcg, page);
1109 /* huge page split is done under lru_lock. so, we have no races. */
1110 VM_BUG_ON(mz->lru_size[lru] < (1 << compound_order(page)));
1111 mz->lru_size[lru] -= 1 << compound_order(page);
1114 void mem_cgroup_lru_del(struct page *page)
1116 mem_cgroup_lru_del_list(page, page_lru(page));
1120 * mem_cgroup_lru_move_lists - account for moving a page between lrus
1121 * @zone: zone of the page
1123 * @from: current lru
1126 * This function accounts for @page being moved between the lrus @from
1127 * and @to, and returns the lruvec for the given @zone and the memcg
1128 * @page is charged to.
1130 * The callsite is then responsible for physically relinking
1131 * @page->lru to the returned lruvec->lists[@to].
1133 struct lruvec *mem_cgroup_lru_move_lists(struct zone *zone,
1138 /* XXX: Optimize this, especially for @from == @to */
1139 mem_cgroup_lru_del_list(page, from);
1140 return mem_cgroup_lru_add_list(zone, page, to);
1144 * Checks whether given mem is same or in the root_mem_cgroup's
1147 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1148 struct mem_cgroup *memcg)
1150 if (root_memcg != memcg) {
1151 return (root_memcg->use_hierarchy &&
1152 css_is_ancestor(&memcg->css, &root_memcg->css));
1158 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *memcg)
1161 struct mem_cgroup *curr = NULL;
1162 struct task_struct *p;
1164 p = find_lock_task_mm(task);
1166 curr = try_get_mem_cgroup_from_mm(p->mm);
1170 * All threads may have already detached their mm's, but the oom
1171 * killer still needs to detect if they have already been oom
1172 * killed to prevent needlessly killing additional tasks.
1175 curr = mem_cgroup_from_task(task);
1177 css_get(&curr->css);
1183 * We should check use_hierarchy of "memcg" not "curr". Because checking
1184 * use_hierarchy of "curr" here make this function true if hierarchy is
1185 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1186 * hierarchy(even if use_hierarchy is disabled in "memcg").
1188 ret = mem_cgroup_same_or_subtree(memcg, curr);
1189 css_put(&curr->css);
1193 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg, struct zone *zone)
1195 unsigned long inactive_ratio;
1196 int nid = zone_to_nid(zone);
1197 int zid = zone_idx(zone);
1198 unsigned long inactive;
1199 unsigned long active;
1202 inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1203 BIT(LRU_INACTIVE_ANON));
1204 active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1205 BIT(LRU_ACTIVE_ANON));
1207 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1209 inactive_ratio = int_sqrt(10 * gb);
1213 return inactive * inactive_ratio < active;
1216 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg, struct zone *zone)
1218 unsigned long active;
1219 unsigned long inactive;
1220 int zid = zone_idx(zone);
1221 int nid = zone_to_nid(zone);
1223 inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1224 BIT(LRU_INACTIVE_FILE));
1225 active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1226 BIT(LRU_ACTIVE_FILE));
1228 return (active > inactive);
1231 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
1234 int nid = zone_to_nid(zone);
1235 int zid = zone_idx(zone);
1236 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
1238 return &mz->reclaim_stat;
1241 struct zone_reclaim_stat *
1242 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
1244 struct page_cgroup *pc;
1245 struct mem_cgroup_per_zone *mz;
1247 if (mem_cgroup_disabled())
1250 pc = lookup_page_cgroup(page);
1251 if (!PageCgroupUsed(pc))
1253 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1255 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1256 return &mz->reclaim_stat;
1259 #define mem_cgroup_from_res_counter(counter, member) \
1260 container_of(counter, struct mem_cgroup, member)
1263 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1264 * @mem: the memory cgroup
1266 * Returns the maximum amount of memory @mem can be charged with, in
1269 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1271 unsigned long long margin;
1273 margin = res_counter_margin(&memcg->res);
1274 if (do_swap_account)
1275 margin = min(margin, res_counter_margin(&memcg->memsw));
1276 return margin >> PAGE_SHIFT;
1279 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1281 struct cgroup *cgrp = memcg->css.cgroup;
1284 if (cgrp->parent == NULL)
1285 return vm_swappiness;
1287 return memcg->swappiness;
1290 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1295 spin_lock(&memcg->pcp_counter_lock);
1296 for_each_online_cpu(cpu)
1297 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) += 1;
1298 memcg->nocpu_base.count[MEM_CGROUP_ON_MOVE] += 1;
1299 spin_unlock(&memcg->pcp_counter_lock);
1305 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1312 spin_lock(&memcg->pcp_counter_lock);
1313 for_each_online_cpu(cpu)
1314 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) -= 1;
1315 memcg->nocpu_base.count[MEM_CGROUP_ON_MOVE] -= 1;
1316 spin_unlock(&memcg->pcp_counter_lock);
1320 * 2 routines for checking "mem" is under move_account() or not.
1322 * mem_cgroup_stealed() - checking a cgroup is mc.from or not. This is used
1323 * for avoiding race in accounting. If true,
1324 * pc->mem_cgroup may be overwritten.
1326 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1327 * under hierarchy of moving cgroups. This is for
1328 * waiting at hith-memory prressure caused by "move".
1331 static bool mem_cgroup_stealed(struct mem_cgroup *memcg)
1333 VM_BUG_ON(!rcu_read_lock_held());
1334 return this_cpu_read(memcg->stat->count[MEM_CGROUP_ON_MOVE]) > 0;
1337 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1339 struct mem_cgroup *from;
1340 struct mem_cgroup *to;
1343 * Unlike task_move routines, we access mc.to, mc.from not under
1344 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1346 spin_lock(&mc.lock);
1352 ret = mem_cgroup_same_or_subtree(memcg, from)
1353 || mem_cgroup_same_or_subtree(memcg, to);
1355 spin_unlock(&mc.lock);
1359 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1361 if (mc.moving_task && current != mc.moving_task) {
1362 if (mem_cgroup_under_move(memcg)) {
1364 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1365 /* moving charge context might have finished. */
1368 finish_wait(&mc.waitq, &wait);
1376 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1377 * @memcg: The memory cgroup that went over limit
1378 * @p: Task that is going to be killed
1380 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1383 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1385 struct cgroup *task_cgrp;
1386 struct cgroup *mem_cgrp;
1388 * Need a buffer in BSS, can't rely on allocations. The code relies
1389 * on the assumption that OOM is serialized for memory controller.
1390 * If this assumption is broken, revisit this code.
1392 static char memcg_name[PATH_MAX];
1400 mem_cgrp = memcg->css.cgroup;
1401 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1403 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1406 * Unfortunately, we are unable to convert to a useful name
1407 * But we'll still print out the usage information
1414 printk(KERN_INFO "Task in %s killed", memcg_name);
1417 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1425 * Continues from above, so we don't need an KERN_ level
1427 printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1430 printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1431 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1432 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1433 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1434 printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1436 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1437 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1438 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1442 * This function returns the number of memcg under hierarchy tree. Returns
1443 * 1(self count) if no children.
1445 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1448 struct mem_cgroup *iter;
1450 for_each_mem_cgroup_tree(iter, memcg)
1456 * Return the memory (and swap, if configured) limit for a memcg.
1458 u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1463 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1464 limit += total_swap_pages << PAGE_SHIFT;
1466 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1468 * If memsw is finite and limits the amount of swap space available
1469 * to this memcg, return that limit.
1471 return min(limit, memsw);
1474 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1476 unsigned long flags)
1478 unsigned long total = 0;
1479 bool noswap = false;
1482 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1484 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1487 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1489 drain_all_stock_async(memcg);
1490 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1492 * Allow limit shrinkers, which are triggered directly
1493 * by userspace, to catch signals and stop reclaim
1494 * after minimal progress, regardless of the margin.
1496 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1498 if (mem_cgroup_margin(memcg))
1501 * If nothing was reclaimed after two attempts, there
1502 * may be no reclaimable pages in this hierarchy.
1511 * test_mem_cgroup_node_reclaimable
1512 * @mem: the target memcg
1513 * @nid: the node ID to be checked.
1514 * @noswap : specify true here if the user wants flle only information.
1516 * This function returns whether the specified memcg contains any
1517 * reclaimable pages on a node. Returns true if there are any reclaimable
1518 * pages in the node.
1520 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1521 int nid, bool noswap)
1523 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1525 if (noswap || !total_swap_pages)
1527 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1532 #if MAX_NUMNODES > 1
1535 * Always updating the nodemask is not very good - even if we have an empty
1536 * list or the wrong list here, we can start from some node and traverse all
1537 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1540 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1544 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1545 * pagein/pageout changes since the last update.
1547 if (!atomic_read(&memcg->numainfo_events))
1549 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1552 /* make a nodemask where this memcg uses memory from */
1553 memcg->scan_nodes = node_states[N_HIGH_MEMORY];
1555 for_each_node_mask(nid, node_states[N_HIGH_MEMORY]) {
1557 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1558 node_clear(nid, memcg->scan_nodes);
1561 atomic_set(&memcg->numainfo_events, 0);
1562 atomic_set(&memcg->numainfo_updating, 0);
1566 * Selecting a node where we start reclaim from. Because what we need is just
1567 * reducing usage counter, start from anywhere is O,K. Considering
1568 * memory reclaim from current node, there are pros. and cons.
1570 * Freeing memory from current node means freeing memory from a node which
1571 * we'll use or we've used. So, it may make LRU bad. And if several threads
1572 * hit limits, it will see a contention on a node. But freeing from remote
1573 * node means more costs for memory reclaim because of memory latency.
1575 * Now, we use round-robin. Better algorithm is welcomed.
1577 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1581 mem_cgroup_may_update_nodemask(memcg);
1582 node = memcg->last_scanned_node;
1584 node = next_node(node, memcg->scan_nodes);
1585 if (node == MAX_NUMNODES)
1586 node = first_node(memcg->scan_nodes);
1588 * We call this when we hit limit, not when pages are added to LRU.
1589 * No LRU may hold pages because all pages are UNEVICTABLE or
1590 * memcg is too small and all pages are not on LRU. In that case,
1591 * we use curret node.
1593 if (unlikely(node == MAX_NUMNODES))
1594 node = numa_node_id();
1596 memcg->last_scanned_node = node;
1601 * Check all nodes whether it contains reclaimable pages or not.
1602 * For quick scan, we make use of scan_nodes. This will allow us to skip
1603 * unused nodes. But scan_nodes is lazily updated and may not cotain
1604 * enough new information. We need to do double check.
1606 bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1611 * quick check...making use of scan_node.
1612 * We can skip unused nodes.
1614 if (!nodes_empty(memcg->scan_nodes)) {
1615 for (nid = first_node(memcg->scan_nodes);
1617 nid = next_node(nid, memcg->scan_nodes)) {
1619 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1624 * Check rest of nodes.
1626 for_each_node_state(nid, N_HIGH_MEMORY) {
1627 if (node_isset(nid, memcg->scan_nodes))
1629 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1636 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1641 bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1643 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
1647 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1650 unsigned long *total_scanned)
1652 struct mem_cgroup *victim = NULL;
1655 unsigned long excess;
1656 unsigned long nr_scanned;
1657 struct mem_cgroup_reclaim_cookie reclaim = {
1662 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
1665 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1670 * If we have not been able to reclaim
1671 * anything, it might because there are
1672 * no reclaimable pages under this hierarchy
1677 * We want to do more targeted reclaim.
1678 * excess >> 2 is not to excessive so as to
1679 * reclaim too much, nor too less that we keep
1680 * coming back to reclaim from this cgroup
1682 if (total >= (excess >> 2) ||
1683 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1688 if (!mem_cgroup_reclaimable(victim, false))
1690 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1692 *total_scanned += nr_scanned;
1693 if (!res_counter_soft_limit_excess(&root_memcg->res))
1696 mem_cgroup_iter_break(root_memcg, victim);
1701 * Check OOM-Killer is already running under our hierarchy.
1702 * If someone is running, return false.
1703 * Has to be called with memcg_oom_lock
1705 static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg)
1707 struct mem_cgroup *iter, *failed = NULL;
1709 for_each_mem_cgroup_tree(iter, memcg) {
1710 if (iter->oom_lock) {
1712 * this subtree of our hierarchy is already locked
1713 * so we cannot give a lock.
1716 mem_cgroup_iter_break(memcg, iter);
1719 iter->oom_lock = true;
1726 * OK, we failed to lock the whole subtree so we have to clean up
1727 * what we set up to the failing subtree
1729 for_each_mem_cgroup_tree(iter, memcg) {
1730 if (iter == failed) {
1731 mem_cgroup_iter_break(memcg, iter);
1734 iter->oom_lock = false;
1740 * Has to be called with memcg_oom_lock
1742 static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1744 struct mem_cgroup *iter;
1746 for_each_mem_cgroup_tree(iter, memcg)
1747 iter->oom_lock = false;
1751 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1753 struct mem_cgroup *iter;
1755 for_each_mem_cgroup_tree(iter, memcg)
1756 atomic_inc(&iter->under_oom);
1759 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1761 struct mem_cgroup *iter;
1764 * When a new child is created while the hierarchy is under oom,
1765 * mem_cgroup_oom_lock() may not be called. We have to use
1766 * atomic_add_unless() here.
1768 for_each_mem_cgroup_tree(iter, memcg)
1769 atomic_add_unless(&iter->under_oom, -1, 0);
1772 static DEFINE_SPINLOCK(memcg_oom_lock);
1773 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1775 struct oom_wait_info {
1776 struct mem_cgroup *memcg;
1780 static int memcg_oom_wake_function(wait_queue_t *wait,
1781 unsigned mode, int sync, void *arg)
1783 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1784 struct mem_cgroup *oom_wait_memcg;
1785 struct oom_wait_info *oom_wait_info;
1787 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1788 oom_wait_memcg = oom_wait_info->memcg;
1791 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
1792 * Then we can use css_is_ancestor without taking care of RCU.
1794 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
1795 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
1797 return autoremove_wake_function(wait, mode, sync, arg);
1800 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
1802 /* for filtering, pass "memcg" as argument. */
1803 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1806 static void memcg_oom_recover(struct mem_cgroup *memcg)
1808 if (memcg && atomic_read(&memcg->under_oom))
1809 memcg_wakeup_oom(memcg);
1813 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1815 bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1817 struct oom_wait_info owait;
1818 bool locked, need_to_kill;
1820 owait.memcg = memcg;
1821 owait.wait.flags = 0;
1822 owait.wait.func = memcg_oom_wake_function;
1823 owait.wait.private = current;
1824 INIT_LIST_HEAD(&owait.wait.task_list);
1825 need_to_kill = true;
1826 mem_cgroup_mark_under_oom(memcg);
1828 /* At first, try to OOM lock hierarchy under memcg.*/
1829 spin_lock(&memcg_oom_lock);
1830 locked = mem_cgroup_oom_lock(memcg);
1832 * Even if signal_pending(), we can't quit charge() loop without
1833 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1834 * under OOM is always welcomed, use TASK_KILLABLE here.
1836 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1837 if (!locked || memcg->oom_kill_disable)
1838 need_to_kill = false;
1840 mem_cgroup_oom_notify(memcg);
1841 spin_unlock(&memcg_oom_lock);
1844 finish_wait(&memcg_oom_waitq, &owait.wait);
1845 mem_cgroup_out_of_memory(memcg, mask, order);
1848 finish_wait(&memcg_oom_waitq, &owait.wait);
1850 spin_lock(&memcg_oom_lock);
1852 mem_cgroup_oom_unlock(memcg);
1853 memcg_wakeup_oom(memcg);
1854 spin_unlock(&memcg_oom_lock);
1856 mem_cgroup_unmark_under_oom(memcg);
1858 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
1860 /* Give chance to dying process */
1861 schedule_timeout_uninterruptible(1);
1866 * Currently used to update mapped file statistics, but the routine can be
1867 * generalized to update other statistics as well.
1869 * Notes: Race condition
1871 * We usually use page_cgroup_lock() for accessing page_cgroup member but
1872 * it tends to be costly. But considering some conditions, we doesn't need
1873 * to do so _always_.
1875 * Considering "charge", lock_page_cgroup() is not required because all
1876 * file-stat operations happen after a page is attached to radix-tree. There
1877 * are no race with "charge".
1879 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
1880 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
1881 * if there are race with "uncharge". Statistics itself is properly handled
1884 * Considering "move", this is an only case we see a race. To make the race
1885 * small, we check MEM_CGROUP_ON_MOVE percpu value and detect there are
1886 * possibility of race condition. If there is, we take a lock.
1889 void mem_cgroup_update_page_stat(struct page *page,
1890 enum mem_cgroup_page_stat_item idx, int val)
1892 struct mem_cgroup *memcg;
1893 struct page_cgroup *pc = lookup_page_cgroup(page);
1894 bool need_unlock = false;
1895 unsigned long uninitialized_var(flags);
1897 if (mem_cgroup_disabled())
1901 memcg = pc->mem_cgroup;
1902 if (unlikely(!memcg || !PageCgroupUsed(pc)))
1904 /* pc->mem_cgroup is unstable ? */
1905 if (unlikely(mem_cgroup_stealed(memcg))) {
1906 /* take a lock against to access pc->mem_cgroup */
1907 move_lock_page_cgroup(pc, &flags);
1909 memcg = pc->mem_cgroup;
1910 if (!memcg || !PageCgroupUsed(pc))
1915 case MEMCG_NR_FILE_MAPPED:
1917 SetPageCgroupFileMapped(pc);
1918 else if (!page_mapped(page))
1919 ClearPageCgroupFileMapped(pc);
1920 idx = MEM_CGROUP_STAT_FILE_MAPPED;
1926 this_cpu_add(memcg->stat->count[idx], val);
1929 if (unlikely(need_unlock))
1930 move_unlock_page_cgroup(pc, &flags);
1933 EXPORT_SYMBOL(mem_cgroup_update_page_stat);
1936 * size of first charge trial. "32" comes from vmscan.c's magic value.
1937 * TODO: maybe necessary to use big numbers in big irons.
1939 #define CHARGE_BATCH 32U
1940 struct memcg_stock_pcp {
1941 struct mem_cgroup *cached; /* this never be root cgroup */
1942 unsigned int nr_pages;
1943 struct work_struct work;
1944 unsigned long flags;
1945 #define FLUSHING_CACHED_CHARGE (0)
1947 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1948 static DEFINE_MUTEX(percpu_charge_mutex);
1951 * Try to consume stocked charge on this cpu. If success, one page is consumed
1952 * from local stock and true is returned. If the stock is 0 or charges from a
1953 * cgroup which is not current target, returns false. This stock will be
1956 static bool consume_stock(struct mem_cgroup *memcg)
1958 struct memcg_stock_pcp *stock;
1961 stock = &get_cpu_var(memcg_stock);
1962 if (memcg == stock->cached && stock->nr_pages)
1964 else /* need to call res_counter_charge */
1966 put_cpu_var(memcg_stock);
1971 * Returns stocks cached in percpu to res_counter and reset cached information.
1973 static void drain_stock(struct memcg_stock_pcp *stock)
1975 struct mem_cgroup *old = stock->cached;
1977 if (stock->nr_pages) {
1978 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
1980 res_counter_uncharge(&old->res, bytes);
1981 if (do_swap_account)
1982 res_counter_uncharge(&old->memsw, bytes);
1983 stock->nr_pages = 0;
1985 stock->cached = NULL;
1989 * This must be called under preempt disabled or must be called by
1990 * a thread which is pinned to local cpu.
1992 static void drain_local_stock(struct work_struct *dummy)
1994 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
1996 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2000 * Cache charges(val) which is from res_counter, to local per_cpu area.
2001 * This will be consumed by consume_stock() function, later.
2003 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2005 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2007 if (stock->cached != memcg) { /* reset if necessary */
2009 stock->cached = memcg;
2011 stock->nr_pages += nr_pages;
2012 put_cpu_var(memcg_stock);
2016 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2017 * of the hierarchy under it. sync flag says whether we should block
2018 * until the work is done.
2020 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2024 /* Notify other cpus that system-wide "drain" is running */
2027 for_each_online_cpu(cpu) {
2028 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2029 struct mem_cgroup *memcg;
2031 memcg = stock->cached;
2032 if (!memcg || !stock->nr_pages)
2034 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2036 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2038 drain_local_stock(&stock->work);
2040 schedule_work_on(cpu, &stock->work);
2048 for_each_online_cpu(cpu) {
2049 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2050 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2051 flush_work(&stock->work);
2058 * Tries to drain stocked charges in other cpus. This function is asynchronous
2059 * and just put a work per cpu for draining localy on each cpu. Caller can
2060 * expects some charges will be back to res_counter later but cannot wait for
2063 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2066 * If someone calls draining, avoid adding more kworker runs.
2068 if (!mutex_trylock(&percpu_charge_mutex))
2070 drain_all_stock(root_memcg, false);
2071 mutex_unlock(&percpu_charge_mutex);
2074 /* This is a synchronous drain interface. */
2075 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2077 /* called when force_empty is called */
2078 mutex_lock(&percpu_charge_mutex);
2079 drain_all_stock(root_memcg, true);
2080 mutex_unlock(&percpu_charge_mutex);
2084 * This function drains percpu counter value from DEAD cpu and
2085 * move it to local cpu. Note that this function can be preempted.
2087 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2091 spin_lock(&memcg->pcp_counter_lock);
2092 for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) {
2093 long x = per_cpu(memcg->stat->count[i], cpu);
2095 per_cpu(memcg->stat->count[i], cpu) = 0;
2096 memcg->nocpu_base.count[i] += x;
2098 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2099 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2101 per_cpu(memcg->stat->events[i], cpu) = 0;
2102 memcg->nocpu_base.events[i] += x;
2104 /* need to clear ON_MOVE value, works as a kind of lock. */
2105 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) = 0;
2106 spin_unlock(&memcg->pcp_counter_lock);
2109 static void synchronize_mem_cgroup_on_move(struct mem_cgroup *memcg, int cpu)
2111 int idx = MEM_CGROUP_ON_MOVE;
2113 spin_lock(&memcg->pcp_counter_lock);
2114 per_cpu(memcg->stat->count[idx], cpu) = memcg->nocpu_base.count[idx];
2115 spin_unlock(&memcg->pcp_counter_lock);
2118 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2119 unsigned long action,
2122 int cpu = (unsigned long)hcpu;
2123 struct memcg_stock_pcp *stock;
2124 struct mem_cgroup *iter;
2126 if ((action == CPU_ONLINE)) {
2127 for_each_mem_cgroup(iter)
2128 synchronize_mem_cgroup_on_move(iter, cpu);
2132 if ((action != CPU_DEAD) || action != CPU_DEAD_FROZEN)
2135 for_each_mem_cgroup(iter)
2136 mem_cgroup_drain_pcp_counter(iter, cpu);
2138 stock = &per_cpu(memcg_stock, cpu);
2144 /* See __mem_cgroup_try_charge() for details */
2146 CHARGE_OK, /* success */
2147 CHARGE_RETRY, /* need to retry but retry is not bad */
2148 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2149 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2150 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2153 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2154 unsigned int nr_pages, bool oom_check)
2156 unsigned long csize = nr_pages * PAGE_SIZE;
2157 struct mem_cgroup *mem_over_limit;
2158 struct res_counter *fail_res;
2159 unsigned long flags = 0;
2162 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2165 if (!do_swap_account)
2167 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2171 res_counter_uncharge(&memcg->res, csize);
2172 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2173 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2175 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2177 * nr_pages can be either a huge page (HPAGE_PMD_NR), a batch
2178 * of regular pages (CHARGE_BATCH), or a single regular page (1).
2180 * Never reclaim on behalf of optional batching, retry with a
2181 * single page instead.
2183 if (nr_pages == CHARGE_BATCH)
2184 return CHARGE_RETRY;
2186 if (!(gfp_mask & __GFP_WAIT))
2187 return CHARGE_WOULDBLOCK;
2189 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2190 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2191 return CHARGE_RETRY;
2193 * Even though the limit is exceeded at this point, reclaim
2194 * may have been able to free some pages. Retry the charge
2195 * before killing the task.
2197 * Only for regular pages, though: huge pages are rather
2198 * unlikely to succeed so close to the limit, and we fall back
2199 * to regular pages anyway in case of failure.
2201 if (nr_pages == 1 && ret)
2202 return CHARGE_RETRY;
2205 * At task move, charge accounts can be doubly counted. So, it's
2206 * better to wait until the end of task_move if something is going on.
2208 if (mem_cgroup_wait_acct_move(mem_over_limit))
2209 return CHARGE_RETRY;
2211 /* If we don't need to call oom-killer at el, return immediately */
2213 return CHARGE_NOMEM;
2215 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask, get_order(csize)))
2216 return CHARGE_OOM_DIE;
2218 return CHARGE_RETRY;
2222 * __mem_cgroup_try_charge() does
2223 * 1. detect memcg to be charged against from passed *mm and *ptr,
2224 * 2. update res_counter
2225 * 3. call memory reclaim if necessary.
2227 * In some special case, if the task is fatal, fatal_signal_pending() or
2228 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2229 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2230 * as possible without any hazards. 2: all pages should have a valid
2231 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2232 * pointer, that is treated as a charge to root_mem_cgroup.
2234 * So __mem_cgroup_try_charge() will return
2235 * 0 ... on success, filling *ptr with a valid memcg pointer.
2236 * -ENOMEM ... charge failure because of resource limits.
2237 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2239 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2240 * the oom-killer can be invoked.
2242 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2244 unsigned int nr_pages,
2245 struct mem_cgroup **ptr,
2248 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2249 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2250 struct mem_cgroup *memcg = NULL;
2254 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2255 * in system level. So, allow to go ahead dying process in addition to
2258 if (unlikely(test_thread_flag(TIF_MEMDIE)
2259 || fatal_signal_pending(current)))
2263 * We always charge the cgroup the mm_struct belongs to.
2264 * The mm_struct's mem_cgroup changes on task migration if the
2265 * thread group leader migrates. It's possible that mm is not
2266 * set, if so charge the init_mm (happens for pagecache usage).
2269 *ptr = root_mem_cgroup;
2271 if (*ptr) { /* css should be a valid one */
2273 VM_BUG_ON(css_is_removed(&memcg->css));
2274 if (mem_cgroup_is_root(memcg))
2276 if (nr_pages == 1 && consume_stock(memcg))
2278 css_get(&memcg->css);
2280 struct task_struct *p;
2283 p = rcu_dereference(mm->owner);
2285 * Because we don't have task_lock(), "p" can exit.
2286 * In that case, "memcg" can point to root or p can be NULL with
2287 * race with swapoff. Then, we have small risk of mis-accouning.
2288 * But such kind of mis-account by race always happens because
2289 * we don't have cgroup_mutex(). It's overkill and we allo that
2291 * (*) swapoff at el will charge against mm-struct not against
2292 * task-struct. So, mm->owner can be NULL.
2294 memcg = mem_cgroup_from_task(p);
2296 memcg = root_mem_cgroup;
2297 if (mem_cgroup_is_root(memcg)) {
2301 if (nr_pages == 1 && consume_stock(memcg)) {
2303 * It seems dagerous to access memcg without css_get().
2304 * But considering how consume_stok works, it's not
2305 * necessary. If consume_stock success, some charges
2306 * from this memcg are cached on this cpu. So, we
2307 * don't need to call css_get()/css_tryget() before
2308 * calling consume_stock().
2313 /* after here, we may be blocked. we need to get refcnt */
2314 if (!css_tryget(&memcg->css)) {
2324 /* If killed, bypass charge */
2325 if (fatal_signal_pending(current)) {
2326 css_put(&memcg->css);
2331 if (oom && !nr_oom_retries) {
2333 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2336 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, oom_check);
2340 case CHARGE_RETRY: /* not in OOM situation but retry */
2342 css_put(&memcg->css);
2345 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2346 css_put(&memcg->css);
2348 case CHARGE_NOMEM: /* OOM routine works */
2350 css_put(&memcg->css);
2353 /* If oom, we never return -ENOMEM */
2356 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2357 css_put(&memcg->css);
2360 } while (ret != CHARGE_OK);
2362 if (batch > nr_pages)
2363 refill_stock(memcg, batch - nr_pages);
2364 css_put(&memcg->css);
2372 *ptr = root_mem_cgroup;
2377 * Somemtimes we have to undo a charge we got by try_charge().
2378 * This function is for that and do uncharge, put css's refcnt.
2379 * gotten by try_charge().
2381 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2382 unsigned int nr_pages)
2384 if (!mem_cgroup_is_root(memcg)) {
2385 unsigned long bytes = nr_pages * PAGE_SIZE;
2387 res_counter_uncharge(&memcg->res, bytes);
2388 if (do_swap_account)
2389 res_counter_uncharge(&memcg->memsw, bytes);
2394 * A helper function to get mem_cgroup from ID. must be called under
2395 * rcu_read_lock(). The caller must check css_is_removed() or some if
2396 * it's concern. (dropping refcnt from swap can be called against removed
2399 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2401 struct cgroup_subsys_state *css;
2403 /* ID 0 is unused ID */
2406 css = css_lookup(&mem_cgroup_subsys, id);
2409 return container_of(css, struct mem_cgroup, css);
2412 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2414 struct mem_cgroup *memcg = NULL;
2415 struct page_cgroup *pc;
2419 VM_BUG_ON(!PageLocked(page));
2421 pc = lookup_page_cgroup(page);
2422 lock_page_cgroup(pc);
2423 if (PageCgroupUsed(pc)) {
2424 memcg = pc->mem_cgroup;
2425 if (memcg && !css_tryget(&memcg->css))
2427 } else if (PageSwapCache(page)) {
2428 ent.val = page_private(page);
2429 id = lookup_swap_cgroup_id(ent);
2431 memcg = mem_cgroup_lookup(id);
2432 if (memcg && !css_tryget(&memcg->css))
2436 unlock_page_cgroup(pc);
2440 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2442 unsigned int nr_pages,
2443 struct page_cgroup *pc,
2444 enum charge_type ctype,
2447 struct zone *uninitialized_var(zone);
2448 bool was_on_lru = false;
2451 lock_page_cgroup(pc);
2452 if (unlikely(PageCgroupUsed(pc))) {
2453 unlock_page_cgroup(pc);
2454 __mem_cgroup_cancel_charge(memcg, nr_pages);
2458 * we don't need page_cgroup_lock about tail pages, becase they are not
2459 * accessed by any other context at this point.
2463 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2464 * may already be on some other mem_cgroup's LRU. Take care of it.
2467 zone = page_zone(page);
2468 spin_lock_irq(&zone->lru_lock);
2469 if (PageLRU(page)) {
2471 del_page_from_lru_list(zone, page, page_lru(page));
2476 pc->mem_cgroup = memcg;
2478 * We access a page_cgroup asynchronously without lock_page_cgroup().
2479 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2480 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2481 * before USED bit, we need memory barrier here.
2482 * See mem_cgroup_add_lru_list(), etc.
2485 SetPageCgroupUsed(pc);
2489 VM_BUG_ON(PageLRU(page));
2491 add_page_to_lru_list(zone, page, page_lru(page));
2493 spin_unlock_irq(&zone->lru_lock);
2496 if (ctype == MEM_CGROUP_CHARGE_TYPE_MAPPED)
2501 mem_cgroup_charge_statistics(memcg, anon, nr_pages);
2502 unlock_page_cgroup(pc);
2505 * "charge_statistics" updated event counter. Then, check it.
2506 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2507 * if they exceeds softlimit.
2509 memcg_check_events(memcg, page);
2512 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2514 #define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MOVE_LOCK) |\
2515 (1 << PCG_MIGRATION))
2517 * Because tail pages are not marked as "used", set it. We're under
2518 * zone->lru_lock, 'splitting on pmd' and compound_lock.
2519 * charge/uncharge will be never happen and move_account() is done under
2520 * compound_lock(), so we don't have to take care of races.
2522 void mem_cgroup_split_huge_fixup(struct page *head)
2524 struct page_cgroup *head_pc = lookup_page_cgroup(head);
2525 struct page_cgroup *pc;
2528 if (mem_cgroup_disabled())
2530 for (i = 1; i < HPAGE_PMD_NR; i++) {
2532 pc->mem_cgroup = head_pc->mem_cgroup;
2533 smp_wmb();/* see __commit_charge() */
2534 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2537 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2540 * mem_cgroup_move_account - move account of the page
2542 * @nr_pages: number of regular pages (>1 for huge pages)
2543 * @pc: page_cgroup of the page.
2544 * @from: mem_cgroup which the page is moved from.
2545 * @to: mem_cgroup which the page is moved to. @from != @to.
2546 * @uncharge: whether we should call uncharge and css_put against @from.
2548 * The caller must confirm following.
2549 * - page is not on LRU (isolate_page() is useful.)
2550 * - compound_lock is held when nr_pages > 1
2552 * This function doesn't do "charge" nor css_get to new cgroup. It should be
2553 * done by a caller(__mem_cgroup_try_charge would be useful). If @uncharge is
2554 * true, this function does "uncharge" from old cgroup, but it doesn't if
2555 * @uncharge is false, so a caller should do "uncharge".
2557 static int mem_cgroup_move_account(struct page *page,
2558 unsigned int nr_pages,
2559 struct page_cgroup *pc,
2560 struct mem_cgroup *from,
2561 struct mem_cgroup *to,
2564 unsigned long flags;
2566 bool anon = PageAnon(page);
2568 VM_BUG_ON(from == to);
2569 VM_BUG_ON(PageLRU(page));
2571 * The page is isolated from LRU. So, collapse function
2572 * will not handle this page. But page splitting can happen.
2573 * Do this check under compound_page_lock(). The caller should
2577 if (nr_pages > 1 && !PageTransHuge(page))
2580 lock_page_cgroup(pc);
2583 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
2586 move_lock_page_cgroup(pc, &flags);
2588 if (PageCgroupFileMapped(pc)) {
2589 /* Update mapped_file data for mem_cgroup */
2591 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2592 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2595 mem_cgroup_charge_statistics(from, anon, -nr_pages);
2597 /* This is not "cancel", but cancel_charge does all we need. */
2598 __mem_cgroup_cancel_charge(from, nr_pages);
2600 /* caller should have done css_get */
2601 pc->mem_cgroup = to;
2602 mem_cgroup_charge_statistics(to, anon, nr_pages);
2604 * We charges against "to" which may not have any tasks. Then, "to"
2605 * can be under rmdir(). But in current implementation, caller of
2606 * this function is just force_empty() and move charge, so it's
2607 * guaranteed that "to" is never removed. So, we don't check rmdir
2610 move_unlock_page_cgroup(pc, &flags);
2613 unlock_page_cgroup(pc);
2617 memcg_check_events(to, page);
2618 memcg_check_events(from, page);
2624 * move charges to its parent.
2627 static int mem_cgroup_move_parent(struct page *page,
2628 struct page_cgroup *pc,
2629 struct mem_cgroup *child,
2632 struct cgroup *cg = child->css.cgroup;
2633 struct cgroup *pcg = cg->parent;
2634 struct mem_cgroup *parent;
2635 unsigned int nr_pages;
2636 unsigned long uninitialized_var(flags);
2644 if (!get_page_unless_zero(page))
2646 if (isolate_lru_page(page))
2649 nr_pages = hpage_nr_pages(page);
2651 parent = mem_cgroup_from_cont(pcg);
2652 ret = __mem_cgroup_try_charge(NULL, gfp_mask, nr_pages, &parent, false);
2657 flags = compound_lock_irqsave(page);
2659 ret = mem_cgroup_move_account(page, nr_pages, pc, child, parent, true);
2661 __mem_cgroup_cancel_charge(parent, nr_pages);
2664 compound_unlock_irqrestore(page, flags);
2666 putback_lru_page(page);
2674 * Charge the memory controller for page usage.
2676 * 0 if the charge was successful
2677 * < 0 if the cgroup is over its limit
2679 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2680 gfp_t gfp_mask, enum charge_type ctype)
2682 struct mem_cgroup *memcg = NULL;
2683 unsigned int nr_pages = 1;
2684 struct page_cgroup *pc;
2688 if (PageTransHuge(page)) {
2689 nr_pages <<= compound_order(page);
2690 VM_BUG_ON(!PageTransHuge(page));
2692 * Never OOM-kill a process for a huge page. The
2693 * fault handler will fall back to regular pages.
2698 pc = lookup_page_cgroup(page);
2699 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
2702 __mem_cgroup_commit_charge(memcg, page, nr_pages, pc, ctype, false);
2706 int mem_cgroup_newpage_charge(struct page *page,
2707 struct mm_struct *mm, gfp_t gfp_mask)
2709 if (mem_cgroup_disabled())
2711 VM_BUG_ON(page_mapped(page));
2712 VM_BUG_ON(page->mapping && !PageAnon(page));
2714 return mem_cgroup_charge_common(page, mm, gfp_mask,
2715 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2719 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2720 enum charge_type ctype);
2722 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2725 struct mem_cgroup *memcg = NULL;
2726 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
2729 if (mem_cgroup_disabled())
2731 if (PageCompound(page))
2736 if (!page_is_file_cache(page))
2737 type = MEM_CGROUP_CHARGE_TYPE_SHMEM;
2739 if (!PageSwapCache(page))
2740 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
2741 else { /* page is swapcache/shmem */
2742 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &memcg);
2744 __mem_cgroup_commit_charge_swapin(page, memcg, type);
2750 * While swap-in, try_charge -> commit or cancel, the page is locked.
2751 * And when try_charge() successfully returns, one refcnt to memcg without
2752 * struct page_cgroup is acquired. This refcnt will be consumed by
2753 * "commit()" or removed by "cancel()"
2755 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2757 gfp_t mask, struct mem_cgroup **memcgp)
2759 struct mem_cgroup *memcg;
2764 if (mem_cgroup_disabled())
2767 if (!do_swap_account)
2770 * A racing thread's fault, or swapoff, may have already updated
2771 * the pte, and even removed page from swap cache: in those cases
2772 * do_swap_page()'s pte_same() test will fail; but there's also a
2773 * KSM case which does need to charge the page.
2775 if (!PageSwapCache(page))
2777 memcg = try_get_mem_cgroup_from_page(page);
2781 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
2782 css_put(&memcg->css);
2789 ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
2796 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
2797 enum charge_type ctype)
2799 struct page_cgroup *pc;
2801 if (mem_cgroup_disabled())
2805 cgroup_exclude_rmdir(&memcg->css);
2807 pc = lookup_page_cgroup(page);
2808 __mem_cgroup_commit_charge(memcg, page, 1, pc, ctype, true);
2810 * Now swap is on-memory. This means this page may be
2811 * counted both as mem and swap....double count.
2812 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2813 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2814 * may call delete_from_swap_cache() before reach here.
2816 if (do_swap_account && PageSwapCache(page)) {
2817 swp_entry_t ent = {.val = page_private(page)};
2818 struct mem_cgroup *swap_memcg;
2821 id = swap_cgroup_record(ent, 0);
2823 swap_memcg = mem_cgroup_lookup(id);
2826 * This recorded memcg can be obsolete one. So, avoid
2827 * calling css_tryget
2829 if (!mem_cgroup_is_root(swap_memcg))
2830 res_counter_uncharge(&swap_memcg->memsw,
2832 mem_cgroup_swap_statistics(swap_memcg, false);
2833 mem_cgroup_put(swap_memcg);
2838 * At swapin, we may charge account against cgroup which has no tasks.
2839 * So, rmdir()->pre_destroy() can be called while we do this charge.
2840 * In that case, we need to call pre_destroy() again. check it here.
2842 cgroup_release_and_wakeup_rmdir(&memcg->css);
2845 void mem_cgroup_commit_charge_swapin(struct page *page,
2846 struct mem_cgroup *memcg)
2848 __mem_cgroup_commit_charge_swapin(page, memcg,
2849 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2852 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
2854 if (mem_cgroup_disabled())
2858 __mem_cgroup_cancel_charge(memcg, 1);
2861 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
2862 unsigned int nr_pages,
2863 const enum charge_type ctype)
2865 struct memcg_batch_info *batch = NULL;
2866 bool uncharge_memsw = true;
2868 /* If swapout, usage of swap doesn't decrease */
2869 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2870 uncharge_memsw = false;
2872 batch = ¤t->memcg_batch;
2874 * In usual, we do css_get() when we remember memcg pointer.
2875 * But in this case, we keep res->usage until end of a series of
2876 * uncharges. Then, it's ok to ignore memcg's refcnt.
2879 batch->memcg = memcg;
2881 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2882 * In those cases, all pages freed continuously can be expected to be in
2883 * the same cgroup and we have chance to coalesce uncharges.
2884 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2885 * because we want to do uncharge as soon as possible.
2888 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
2889 goto direct_uncharge;
2892 goto direct_uncharge;
2895 * In typical case, batch->memcg == mem. This means we can
2896 * merge a series of uncharges to an uncharge of res_counter.
2897 * If not, we uncharge res_counter ony by one.
2899 if (batch->memcg != memcg)
2900 goto direct_uncharge;
2901 /* remember freed charge and uncharge it later */
2904 batch->memsw_nr_pages++;
2907 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
2909 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
2910 if (unlikely(batch->memcg != memcg))
2911 memcg_oom_recover(memcg);
2915 * uncharge if !page_mapped(page)
2917 static struct mem_cgroup *
2918 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
2920 struct mem_cgroup *memcg = NULL;
2921 unsigned int nr_pages = 1;
2922 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))
2949 anon = PageAnon(page);
2952 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2955 case MEM_CGROUP_CHARGE_TYPE_DROP:
2956 /* See mem_cgroup_prepare_migration() */
2957 if (page_mapped(page) || PageCgroupMigration(pc))
2960 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
2961 if (!PageAnon(page)) { /* Shared memory */
2962 if (page->mapping && !page_is_file_cache(page))
2964 } else if (page_mapped(page)) /* Anon */
2971 mem_cgroup_charge_statistics(memcg, anon, -nr_pages);
2973 ClearPageCgroupUsed(pc);
2975 * pc->mem_cgroup is not cleared here. It will be accessed when it's
2976 * freed from LRU. This is safe because uncharged page is expected not
2977 * to be reused (freed soon). Exception is SwapCache, it's handled by
2978 * special functions.
2981 unlock_page_cgroup(pc);
2983 * even after unlock, we have memcg->res.usage here and this memcg
2984 * will never be freed.
2986 memcg_check_events(memcg, page);
2987 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
2988 mem_cgroup_swap_statistics(memcg, true);
2989 mem_cgroup_get(memcg);
2991 if (!mem_cgroup_is_root(memcg))
2992 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
2997 unlock_page_cgroup(pc);
3001 void mem_cgroup_uncharge_page(struct page *page)
3004 if (page_mapped(page))
3006 VM_BUG_ON(page->mapping && !PageAnon(page));
3007 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
3010 void mem_cgroup_uncharge_cache_page(struct page *page)
3012 VM_BUG_ON(page_mapped(page));
3013 VM_BUG_ON(page->mapping);
3014 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
3018 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
3019 * In that cases, pages are freed continuously and we can expect pages
3020 * are in the same memcg. All these calls itself limits the number of
3021 * pages freed at once, then uncharge_start/end() is called properly.
3022 * This may be called prural(2) times in a context,
3025 void mem_cgroup_uncharge_start(void)
3027 current->memcg_batch.do_batch++;
3028 /* We can do nest. */
3029 if (current->memcg_batch.do_batch == 1) {
3030 current->memcg_batch.memcg = NULL;
3031 current->memcg_batch.nr_pages = 0;
3032 current->memcg_batch.memsw_nr_pages = 0;
3036 void mem_cgroup_uncharge_end(void)
3038 struct memcg_batch_info *batch = ¤t->memcg_batch;
3040 if (!batch->do_batch)
3044 if (batch->do_batch) /* If stacked, do nothing. */
3050 * This "batch->memcg" is valid without any css_get/put etc...
3051 * bacause we hide charges behind us.
3053 if (batch->nr_pages)
3054 res_counter_uncharge(&batch->memcg->res,
3055 batch->nr_pages * PAGE_SIZE);
3056 if (batch->memsw_nr_pages)
3057 res_counter_uncharge(&batch->memcg->memsw,
3058 batch->memsw_nr_pages * PAGE_SIZE);
3059 memcg_oom_recover(batch->memcg);
3060 /* forget this pointer (for sanity check) */
3061 batch->memcg = NULL;
3066 * called after __delete_from_swap_cache() and drop "page" account.
3067 * memcg information is recorded to swap_cgroup of "ent"
3070 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
3072 struct mem_cgroup *memcg;
3073 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
3075 if (!swapout) /* this was a swap cache but the swap is unused ! */
3076 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
3078 memcg = __mem_cgroup_uncharge_common(page, ctype);
3081 * record memcg information, if swapout && memcg != NULL,
3082 * mem_cgroup_get() was called in uncharge().
3084 if (do_swap_account && swapout && memcg)
3085 swap_cgroup_record(ent, css_id(&memcg->css));
3089 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3091 * called from swap_entry_free(). remove record in swap_cgroup and
3092 * uncharge "memsw" account.
3094 void mem_cgroup_uncharge_swap(swp_entry_t ent)
3096 struct mem_cgroup *memcg;
3099 if (!do_swap_account)
3102 id = swap_cgroup_record(ent, 0);
3104 memcg = mem_cgroup_lookup(id);
3107 * We uncharge this because swap is freed.
3108 * This memcg can be obsolete one. We avoid calling css_tryget
3110 if (!mem_cgroup_is_root(memcg))
3111 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
3112 mem_cgroup_swap_statistics(memcg, false);
3113 mem_cgroup_put(memcg);
3119 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3120 * @entry: swap entry to be moved
3121 * @from: mem_cgroup which the entry is moved from
3122 * @to: mem_cgroup which the entry is moved to
3123 * @need_fixup: whether we should fixup res_counters and refcounts.
3125 * It succeeds only when the swap_cgroup's record for this entry is the same
3126 * as the mem_cgroup's id of @from.
3128 * Returns 0 on success, -EINVAL on failure.
3130 * The caller must have charged to @to, IOW, called res_counter_charge() about
3131 * both res and memsw, and called css_get().
3133 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3134 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3136 unsigned short old_id, new_id;
3138 old_id = css_id(&from->css);
3139 new_id = css_id(&to->css);
3141 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3142 mem_cgroup_swap_statistics(from, false);
3143 mem_cgroup_swap_statistics(to, true);
3145 * This function is only called from task migration context now.
3146 * It postpones res_counter and refcount handling till the end
3147 * of task migration(mem_cgroup_clear_mc()) for performance
3148 * improvement. But we cannot postpone mem_cgroup_get(to)
3149 * because if the process that has been moved to @to does
3150 * swap-in, the refcount of @to might be decreased to 0.
3154 if (!mem_cgroup_is_root(from))
3155 res_counter_uncharge(&from->memsw, PAGE_SIZE);
3156 mem_cgroup_put(from);
3158 * we charged both to->res and to->memsw, so we should
3161 if (!mem_cgroup_is_root(to))
3162 res_counter_uncharge(&to->res, PAGE_SIZE);
3169 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3170 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3177 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
3180 int mem_cgroup_prepare_migration(struct page *page,
3181 struct page *newpage, struct mem_cgroup **memcgp, gfp_t gfp_mask)
3183 struct mem_cgroup *memcg = NULL;
3184 struct page_cgroup *pc;
3185 enum charge_type ctype;
3190 VM_BUG_ON(PageTransHuge(page));
3191 if (mem_cgroup_disabled())
3194 pc = lookup_page_cgroup(page);
3195 lock_page_cgroup(pc);
3196 if (PageCgroupUsed(pc)) {
3197 memcg = pc->mem_cgroup;
3198 css_get(&memcg->css);
3200 * At migrating an anonymous page, its mapcount goes down
3201 * to 0 and uncharge() will be called. But, even if it's fully
3202 * unmapped, migration may fail and this page has to be
3203 * charged again. We set MIGRATION flag here and delay uncharge
3204 * until end_migration() is called
3206 * Corner Case Thinking
3208 * When the old page was mapped as Anon and it's unmap-and-freed
3209 * while migration was ongoing.
3210 * If unmap finds the old page, uncharge() of it will be delayed
3211 * until end_migration(). If unmap finds a new page, it's
3212 * uncharged when it make mapcount to be 1->0. If unmap code
3213 * finds swap_migration_entry, the new page will not be mapped
3214 * and end_migration() will find it(mapcount==0).
3217 * When the old page was mapped but migraion fails, the kernel
3218 * remaps it. A charge for it is kept by MIGRATION flag even
3219 * if mapcount goes down to 0. We can do remap successfully
3220 * without charging it again.
3223 * The "old" page is under lock_page() until the end of
3224 * migration, so, the old page itself will not be swapped-out.
3225 * If the new page is swapped out before end_migraton, our
3226 * hook to usual swap-out path will catch the event.
3229 SetPageCgroupMigration(pc);
3231 unlock_page_cgroup(pc);
3233 * If the page is not charged at this point,
3240 ret = __mem_cgroup_try_charge(NULL, gfp_mask, 1, memcgp, false);
3241 css_put(&memcg->css);/* drop extra refcnt */
3243 if (PageAnon(page)) {
3244 lock_page_cgroup(pc);
3245 ClearPageCgroupMigration(pc);
3246 unlock_page_cgroup(pc);
3248 * The old page may be fully unmapped while we kept it.
3250 mem_cgroup_uncharge_page(page);
3252 /* we'll need to revisit this error code (we have -EINTR) */
3256 * We charge new page before it's used/mapped. So, even if unlock_page()
3257 * is called before end_migration, we can catch all events on this new
3258 * page. In the case new page is migrated but not remapped, new page's
3259 * mapcount will be finally 0 and we call uncharge in end_migration().
3261 pc = lookup_page_cgroup(newpage);
3263 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
3264 else if (page_is_file_cache(page))
3265 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
3267 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3268 __mem_cgroup_commit_charge(memcg, newpage, 1, pc, ctype, false);
3272 /* remove redundant charge if migration failed*/
3273 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
3274 struct page *oldpage, struct page *newpage, bool migration_ok)
3276 struct page *used, *unused;
3277 struct page_cgroup *pc;
3282 /* blocks rmdir() */
3283 cgroup_exclude_rmdir(&memcg->css);
3284 if (!migration_ok) {
3292 * We disallowed uncharge of pages under migration because mapcount
3293 * of the page goes down to zero, temporarly.
3294 * Clear the flag and check the page should be charged.
3296 pc = lookup_page_cgroup(oldpage);
3297 lock_page_cgroup(pc);
3298 ClearPageCgroupMigration(pc);
3299 unlock_page_cgroup(pc);
3300 anon = PageAnon(used);
3301 __mem_cgroup_uncharge_common(unused,
3302 anon ? MEM_CGROUP_CHARGE_TYPE_MAPPED
3303 : MEM_CGROUP_CHARGE_TYPE_CACHE);
3306 * If a page is a file cache, radix-tree replacement is very atomic
3307 * and we can skip this check. When it was an Anon page, its mapcount
3308 * goes down to 0. But because we added MIGRATION flage, it's not
3309 * uncharged yet. There are several case but page->mapcount check
3310 * and USED bit check in mem_cgroup_uncharge_page() will do enough
3311 * check. (see prepare_charge() also)
3314 mem_cgroup_uncharge_page(used);
3316 * At migration, we may charge account against cgroup which has no
3318 * So, rmdir()->pre_destroy() can be called while we do this charge.
3319 * In that case, we need to call pre_destroy() again. check it here.
3321 cgroup_release_and_wakeup_rmdir(&memcg->css);
3325 * At replace page cache, newpage is not under any memcg but it's on
3326 * LRU. So, this function doesn't touch res_counter but handles LRU
3327 * in correct way. Both pages are locked so we cannot race with uncharge.
3329 void mem_cgroup_replace_page_cache(struct page *oldpage,
3330 struct page *newpage)
3332 struct mem_cgroup *memcg;
3333 struct page_cgroup *pc;
3334 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
3336 if (mem_cgroup_disabled())
3339 pc = lookup_page_cgroup(oldpage);
3340 /* fix accounting on old pages */
3341 lock_page_cgroup(pc);
3342 memcg = pc->mem_cgroup;
3343 mem_cgroup_charge_statistics(memcg, false, -1);
3344 ClearPageCgroupUsed(pc);
3345 unlock_page_cgroup(pc);
3347 if (PageSwapBacked(oldpage))
3348 type = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3351 * Even if newpage->mapping was NULL before starting replacement,
3352 * the newpage may be on LRU(or pagevec for LRU) already. We lock
3353 * LRU while we overwrite pc->mem_cgroup.
3355 __mem_cgroup_commit_charge(memcg, newpage, 1, pc, type, true);
3358 #ifdef CONFIG_DEBUG_VM
3359 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3361 struct page_cgroup *pc;
3363 pc = lookup_page_cgroup(page);
3365 * Can be NULL while feeding pages into the page allocator for
3366 * the first time, i.e. during boot or memory hotplug;
3367 * or when mem_cgroup_disabled().
3369 if (likely(pc) && PageCgroupUsed(pc))
3374 bool mem_cgroup_bad_page_check(struct page *page)
3376 if (mem_cgroup_disabled())
3379 return lookup_page_cgroup_used(page) != NULL;
3382 void mem_cgroup_print_bad_page(struct page *page)
3384 struct page_cgroup *pc;
3386 pc = lookup_page_cgroup_used(page);
3388 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
3389 pc, pc->flags, pc->mem_cgroup);
3394 static DEFINE_MUTEX(set_limit_mutex);
3396 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3397 unsigned long long val)
3400 u64 memswlimit, memlimit;
3402 int children = mem_cgroup_count_children(memcg);
3403 u64 curusage, oldusage;
3407 * For keeping hierarchical_reclaim simple, how long we should retry
3408 * is depends on callers. We set our retry-count to be function
3409 * of # of children which we should visit in this loop.
3411 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3413 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3416 while (retry_count) {
3417 if (signal_pending(current)) {
3422 * Rather than hide all in some function, I do this in
3423 * open coded manner. You see what this really does.
3424 * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3426 mutex_lock(&set_limit_mutex);
3427 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3428 if (memswlimit < val) {
3430 mutex_unlock(&set_limit_mutex);
3434 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3438 ret = res_counter_set_limit(&memcg->res, val);
3440 if (memswlimit == val)
3441 memcg->memsw_is_minimum = true;
3443 memcg->memsw_is_minimum = false;
3445 mutex_unlock(&set_limit_mutex);
3450 mem_cgroup_reclaim(memcg, GFP_KERNEL,
3451 MEM_CGROUP_RECLAIM_SHRINK);
3452 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3453 /* Usage is reduced ? */
3454 if (curusage >= oldusage)
3457 oldusage = curusage;
3459 if (!ret && enlarge)
3460 memcg_oom_recover(memcg);
3465 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3466 unsigned long long val)
3469 u64 memlimit, memswlimit, oldusage, curusage;
3470 int children = mem_cgroup_count_children(memcg);
3474 /* see mem_cgroup_resize_res_limit */
3475 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3476 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3477 while (retry_count) {
3478 if (signal_pending(current)) {
3483 * Rather than hide all in some function, I do this in
3484 * open coded manner. You see what this really does.
3485 * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3487 mutex_lock(&set_limit_mutex);
3488 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3489 if (memlimit > val) {
3491 mutex_unlock(&set_limit_mutex);
3494 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3495 if (memswlimit < val)
3497 ret = res_counter_set_limit(&memcg->memsw, val);
3499 if (memlimit == val)
3500 memcg->memsw_is_minimum = true;
3502 memcg->memsw_is_minimum = false;
3504 mutex_unlock(&set_limit_mutex);
3509 mem_cgroup_reclaim(memcg, GFP_KERNEL,
3510 MEM_CGROUP_RECLAIM_NOSWAP |
3511 MEM_CGROUP_RECLAIM_SHRINK);
3512 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3513 /* Usage is reduced ? */
3514 if (curusage >= oldusage)
3517 oldusage = curusage;
3519 if (!ret && enlarge)
3520 memcg_oom_recover(memcg);
3524 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3526 unsigned long *total_scanned)
3528 unsigned long nr_reclaimed = 0;
3529 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3530 unsigned long reclaimed;
3532 struct mem_cgroup_tree_per_zone *mctz;
3533 unsigned long long excess;
3534 unsigned long nr_scanned;
3539 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3541 * This loop can run a while, specially if mem_cgroup's continuously
3542 * keep exceeding their soft limit and putting the system under
3549 mz = mem_cgroup_largest_soft_limit_node(mctz);
3554 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
3555 gfp_mask, &nr_scanned);
3556 nr_reclaimed += reclaimed;
3557 *total_scanned += nr_scanned;
3558 spin_lock(&mctz->lock);
3561 * If we failed to reclaim anything from this memory cgroup
3562 * it is time to move on to the next cgroup
3568 * Loop until we find yet another one.
3570 * By the time we get the soft_limit lock
3571 * again, someone might have aded the
3572 * group back on the RB tree. Iterate to
3573 * make sure we get a different mem.
3574 * mem_cgroup_largest_soft_limit_node returns
3575 * NULL if no other cgroup is present on
3579 __mem_cgroup_largest_soft_limit_node(mctz);
3581 css_put(&next_mz->memcg->css);
3582 else /* next_mz == NULL or other memcg */
3586 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
3587 excess = res_counter_soft_limit_excess(&mz->memcg->res);
3589 * One school of thought says that we should not add
3590 * back the node to the tree if reclaim returns 0.
3591 * But our reclaim could return 0, simply because due
3592 * to priority we are exposing a smaller subset of
3593 * memory to reclaim from. Consider this as a longer
3596 /* If excess == 0, no tree ops */
3597 __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess);
3598 spin_unlock(&mctz->lock);
3599 css_put(&mz->memcg->css);
3602 * Could not reclaim anything and there are no more
3603 * mem cgroups to try or we seem to be looping without
3604 * reclaiming anything.
3606 if (!nr_reclaimed &&
3608 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3610 } while (!nr_reclaimed);
3612 css_put(&next_mz->memcg->css);
3613 return nr_reclaimed;
3617 * This routine traverse page_cgroup in given list and drop them all.
3618 * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3620 static int mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
3621 int node, int zid, enum lru_list lru)
3623 struct mem_cgroup_per_zone *mz;
3624 unsigned long flags, loop;
3625 struct list_head *list;
3630 zone = &NODE_DATA(node)->node_zones[zid];
3631 mz = mem_cgroup_zoneinfo(memcg, node, zid);
3632 list = &mz->lruvec.lists[lru];
3634 loop = mz->lru_size[lru];
3635 /* give some margin against EBUSY etc...*/
3639 struct page_cgroup *pc;
3643 spin_lock_irqsave(&zone->lru_lock, flags);
3644 if (list_empty(list)) {
3645 spin_unlock_irqrestore(&zone->lru_lock, flags);
3648 page = list_entry(list->prev, struct page, lru);
3650 list_move(&page->lru, list);
3652 spin_unlock_irqrestore(&zone->lru_lock, flags);
3655 spin_unlock_irqrestore(&zone->lru_lock, flags);
3657 pc = lookup_page_cgroup(page);
3659 ret = mem_cgroup_move_parent(page, pc, memcg, GFP_KERNEL);
3660 if (ret == -ENOMEM || ret == -EINTR)
3663 if (ret == -EBUSY || ret == -EINVAL) {
3664 /* found lock contention or "pc" is obsolete. */
3671 if (!ret && !list_empty(list))
3677 * make mem_cgroup's charge to be 0 if there is no task.
3678 * This enables deleting this mem_cgroup.
3680 static int mem_cgroup_force_empty(struct mem_cgroup *memcg, bool free_all)
3683 int node, zid, shrink;
3684 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3685 struct cgroup *cgrp = memcg->css.cgroup;
3687 css_get(&memcg->css);
3690 /* should free all ? */
3696 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3699 if (signal_pending(current))
3701 /* This is for making all *used* pages to be on LRU. */
3702 lru_add_drain_all();
3703 drain_all_stock_sync(memcg);
3705 mem_cgroup_start_move(memcg);
3706 for_each_node_state(node, N_HIGH_MEMORY) {
3707 for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3710 ret = mem_cgroup_force_empty_list(memcg,
3719 mem_cgroup_end_move(memcg);
3720 memcg_oom_recover(memcg);
3721 /* it seems parent cgroup doesn't have enough mem */
3725 /* "ret" should also be checked to ensure all lists are empty. */
3726 } while (memcg->res.usage > 0 || ret);
3728 css_put(&memcg->css);
3732 /* returns EBUSY if there is a task or if we come here twice. */
3733 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3737 /* we call try-to-free pages for make this cgroup empty */
3738 lru_add_drain_all();
3739 /* try to free all pages in this cgroup */
3741 while (nr_retries && memcg->res.usage > 0) {
3744 if (signal_pending(current)) {
3748 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
3752 /* maybe some writeback is necessary */
3753 congestion_wait(BLK_RW_ASYNC, HZ/10);
3758 /* try move_account...there may be some *locked* pages. */
3762 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3764 return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3768 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3770 return mem_cgroup_from_cont(cont)->use_hierarchy;
3773 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3777 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3778 struct cgroup *parent = cont->parent;
3779 struct mem_cgroup *parent_memcg = NULL;
3782 parent_memcg = mem_cgroup_from_cont(parent);
3786 * If parent's use_hierarchy is set, we can't make any modifications
3787 * in the child subtrees. If it is unset, then the change can
3788 * occur, provided the current cgroup has no children.
3790 * For the root cgroup, parent_mem is NULL, we allow value to be
3791 * set if there are no children.
3793 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3794 (val == 1 || val == 0)) {
3795 if (list_empty(&cont->children))
3796 memcg->use_hierarchy = val;
3807 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
3808 enum mem_cgroup_stat_index idx)
3810 struct mem_cgroup *iter;
3813 /* Per-cpu values can be negative, use a signed accumulator */
3814 for_each_mem_cgroup_tree(iter, memcg)
3815 val += mem_cgroup_read_stat(iter, idx);
3817 if (val < 0) /* race ? */
3822 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3826 if (!mem_cgroup_is_root(memcg)) {
3828 return res_counter_read_u64(&memcg->res, RES_USAGE);
3830 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
3833 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
3834 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
3837 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
3839 return val << PAGE_SHIFT;
3842 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
3844 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3848 type = MEMFILE_TYPE(cft->private);
3849 name = MEMFILE_ATTR(cft->private);
3852 if (name == RES_USAGE)
3853 val = mem_cgroup_usage(memcg, false);
3855 val = res_counter_read_u64(&memcg->res, name);
3858 if (name == RES_USAGE)
3859 val = mem_cgroup_usage(memcg, true);
3861 val = res_counter_read_u64(&memcg->memsw, name);
3870 * The user of this function is...
3873 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
3876 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3878 unsigned long long val;
3881 type = MEMFILE_TYPE(cft->private);
3882 name = MEMFILE_ATTR(cft->private);
3885 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3889 /* This function does all necessary parse...reuse it */
3890 ret = res_counter_memparse_write_strategy(buffer, &val);
3894 ret = mem_cgroup_resize_limit(memcg, val);
3896 ret = mem_cgroup_resize_memsw_limit(memcg, val);
3898 case RES_SOFT_LIMIT:
3899 ret = res_counter_memparse_write_strategy(buffer, &val);
3903 * For memsw, soft limits are hard to implement in terms
3904 * of semantics, for now, we support soft limits for
3905 * control without swap
3908 ret = res_counter_set_soft_limit(&memcg->res, val);
3913 ret = -EINVAL; /* should be BUG() ? */
3919 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
3920 unsigned long long *mem_limit, unsigned long long *memsw_limit)
3922 struct cgroup *cgroup;
3923 unsigned long long min_limit, min_memsw_limit, tmp;
3925 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3926 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3927 cgroup = memcg->css.cgroup;
3928 if (!memcg->use_hierarchy)
3931 while (cgroup->parent) {
3932 cgroup = cgroup->parent;
3933 memcg = mem_cgroup_from_cont(cgroup);
3934 if (!memcg->use_hierarchy)
3936 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
3937 min_limit = min(min_limit, tmp);
3938 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3939 min_memsw_limit = min(min_memsw_limit, tmp);
3942 *mem_limit = min_limit;
3943 *memsw_limit = min_memsw_limit;
3946 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
3948 struct mem_cgroup *memcg;
3951 memcg = mem_cgroup_from_cont(cont);
3952 type = MEMFILE_TYPE(event);
3953 name = MEMFILE_ATTR(event);
3957 res_counter_reset_max(&memcg->res);
3959 res_counter_reset_max(&memcg->memsw);
3963 res_counter_reset_failcnt(&memcg->res);
3965 res_counter_reset_failcnt(&memcg->memsw);
3972 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
3975 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
3979 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
3980 struct cftype *cft, u64 val)
3982 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3984 if (val >= (1 << NR_MOVE_TYPE))
3987 * We check this value several times in both in can_attach() and
3988 * attach(), so we need cgroup lock to prevent this value from being
3992 memcg->move_charge_at_immigrate = val;
3998 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
3999 struct cftype *cft, u64 val)
4006 /* For read statistics */
4024 struct mcs_total_stat {
4025 s64 stat[NR_MCS_STAT];
4031 } memcg_stat_strings[NR_MCS_STAT] = {
4032 {"cache", "total_cache"},
4033 {"rss", "total_rss"},
4034 {"mapped_file", "total_mapped_file"},
4035 {"pgpgin", "total_pgpgin"},
4036 {"pgpgout", "total_pgpgout"},
4037 {"swap", "total_swap"},
4038 {"pgfault", "total_pgfault"},
4039 {"pgmajfault", "total_pgmajfault"},
4040 {"inactive_anon", "total_inactive_anon"},
4041 {"active_anon", "total_active_anon"},
4042 {"inactive_file", "total_inactive_file"},
4043 {"active_file", "total_active_file"},
4044 {"unevictable", "total_unevictable"}
4049 mem_cgroup_get_local_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4054 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_CACHE);
4055 s->stat[MCS_CACHE] += val * PAGE_SIZE;
4056 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_RSS);
4057 s->stat[MCS_RSS] += val * PAGE_SIZE;
4058 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_FILE_MAPPED);
4059 s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
4060 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGIN);
4061 s->stat[MCS_PGPGIN] += val;
4062 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGOUT);
4063 s->stat[MCS_PGPGOUT] += val;
4064 if (do_swap_account) {
4065 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
4066 s->stat[MCS_SWAP] += val * PAGE_SIZE;
4068 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGFAULT);
4069 s->stat[MCS_PGFAULT] += val;
4070 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGMAJFAULT);
4071 s->stat[MCS_PGMAJFAULT] += val;
4074 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_ANON));
4075 s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
4076 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_ANON));
4077 s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
4078 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_FILE));
4079 s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
4080 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_FILE));
4081 s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
4082 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
4083 s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
4087 mem_cgroup_get_total_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4089 struct mem_cgroup *iter;
4091 for_each_mem_cgroup_tree(iter, memcg)
4092 mem_cgroup_get_local_stat(iter, s);
4096 static int mem_control_numa_stat_show(struct seq_file *m, void *arg)
4099 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
4100 unsigned long node_nr;
4101 struct cgroup *cont = m->private;
4102 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4104 total_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL);
4105 seq_printf(m, "total=%lu", total_nr);
4106 for_each_node_state(nid, N_HIGH_MEMORY) {
4107 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL);
4108 seq_printf(m, " N%d=%lu", nid, node_nr);
4112 file_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_FILE);
4113 seq_printf(m, "file=%lu", file_nr);
4114 for_each_node_state(nid, N_HIGH_MEMORY) {
4115 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4117 seq_printf(m, " N%d=%lu", nid, node_nr);
4121 anon_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_ANON);
4122 seq_printf(m, "anon=%lu", anon_nr);
4123 for_each_node_state(nid, N_HIGH_MEMORY) {
4124 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4126 seq_printf(m, " N%d=%lu", nid, node_nr);
4130 unevictable_nr = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
4131 seq_printf(m, "unevictable=%lu", unevictable_nr);
4132 for_each_node_state(nid, N_HIGH_MEMORY) {
4133 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4134 BIT(LRU_UNEVICTABLE));
4135 seq_printf(m, " N%d=%lu", nid, node_nr);
4140 #endif /* CONFIG_NUMA */
4142 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
4143 struct cgroup_map_cb *cb)
4145 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4146 struct mcs_total_stat mystat;
4149 memset(&mystat, 0, sizeof(mystat));
4150 mem_cgroup_get_local_stat(memcg, &mystat);
4153 for (i = 0; i < NR_MCS_STAT; i++) {
4154 if (i == MCS_SWAP && !do_swap_account)
4156 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
4159 /* Hierarchical information */
4161 unsigned long long limit, memsw_limit;
4162 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
4163 cb->fill(cb, "hierarchical_memory_limit", limit);
4164 if (do_swap_account)
4165 cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
4168 memset(&mystat, 0, sizeof(mystat));
4169 mem_cgroup_get_total_stat(memcg, &mystat);
4170 for (i = 0; i < NR_MCS_STAT; i++) {
4171 if (i == MCS_SWAP && !do_swap_account)
4173 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
4176 #ifdef CONFIG_DEBUG_VM
4179 struct mem_cgroup_per_zone *mz;
4180 unsigned long recent_rotated[2] = {0, 0};
4181 unsigned long recent_scanned[2] = {0, 0};
4183 for_each_online_node(nid)
4184 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4185 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
4187 recent_rotated[0] +=
4188 mz->reclaim_stat.recent_rotated[0];
4189 recent_rotated[1] +=
4190 mz->reclaim_stat.recent_rotated[1];
4191 recent_scanned[0] +=
4192 mz->reclaim_stat.recent_scanned[0];
4193 recent_scanned[1] +=
4194 mz->reclaim_stat.recent_scanned[1];
4196 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
4197 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
4198 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
4199 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
4206 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
4208 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4210 return mem_cgroup_swappiness(memcg);
4213 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
4216 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4217 struct mem_cgroup *parent;
4222 if (cgrp->parent == NULL)
4225 parent = mem_cgroup_from_cont(cgrp->parent);
4229 /* If under hierarchy, only empty-root can set this value */
4230 if ((parent->use_hierarchy) ||
4231 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4236 memcg->swappiness = val;
4243 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4245 struct mem_cgroup_threshold_ary *t;
4251 t = rcu_dereference(memcg->thresholds.primary);
4253 t = rcu_dereference(memcg->memsw_thresholds.primary);
4258 usage = mem_cgroup_usage(memcg, swap);
4261 * current_threshold points to threshold just below usage.
4262 * If it's not true, a threshold was crossed after last
4263 * call of __mem_cgroup_threshold().
4265 i = t->current_threshold;
4268 * Iterate backward over array of thresholds starting from
4269 * current_threshold and check if a threshold is crossed.
4270 * If none of thresholds below usage is crossed, we read
4271 * only one element of the array here.
4273 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4274 eventfd_signal(t->entries[i].eventfd, 1);
4276 /* i = current_threshold + 1 */
4280 * Iterate forward over array of thresholds starting from
4281 * current_threshold+1 and check if a threshold is crossed.
4282 * If none of thresholds above usage is crossed, we read
4283 * only one element of the array here.
4285 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4286 eventfd_signal(t->entries[i].eventfd, 1);
4288 /* Update current_threshold */
4289 t->current_threshold = i - 1;
4294 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4297 __mem_cgroup_threshold(memcg, false);
4298 if (do_swap_account)
4299 __mem_cgroup_threshold(memcg, true);
4301 memcg = parent_mem_cgroup(memcg);
4305 static int compare_thresholds(const void *a, const void *b)
4307 const struct mem_cgroup_threshold *_a = a;
4308 const struct mem_cgroup_threshold *_b = b;
4310 return _a->threshold - _b->threshold;
4313 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4315 struct mem_cgroup_eventfd_list *ev;
4317 list_for_each_entry(ev, &memcg->oom_notify, list)
4318 eventfd_signal(ev->eventfd, 1);
4322 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4324 struct mem_cgroup *iter;
4326 for_each_mem_cgroup_tree(iter, memcg)
4327 mem_cgroup_oom_notify_cb(iter);
4330 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
4331 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4333 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4334 struct mem_cgroup_thresholds *thresholds;
4335 struct mem_cgroup_threshold_ary *new;
4336 int type = MEMFILE_TYPE(cft->private);
4337 u64 threshold, usage;
4340 ret = res_counter_memparse_write_strategy(args, &threshold);
4344 mutex_lock(&memcg->thresholds_lock);
4347 thresholds = &memcg->thresholds;
4348 else if (type == _MEMSWAP)
4349 thresholds = &memcg->memsw_thresholds;
4353 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4355 /* Check if a threshold crossed before adding a new one */
4356 if (thresholds->primary)
4357 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4359 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4361 /* Allocate memory for new array of thresholds */
4362 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4370 /* Copy thresholds (if any) to new array */
4371 if (thresholds->primary) {
4372 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4373 sizeof(struct mem_cgroup_threshold));
4376 /* Add new threshold */
4377 new->entries[size - 1].eventfd = eventfd;
4378 new->entries[size - 1].threshold = threshold;
4380 /* Sort thresholds. Registering of new threshold isn't time-critical */
4381 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4382 compare_thresholds, NULL);
4384 /* Find current threshold */
4385 new->current_threshold = -1;
4386 for (i = 0; i < size; i++) {
4387 if (new->entries[i].threshold < usage) {
4389 * new->current_threshold will not be used until
4390 * rcu_assign_pointer(), so it's safe to increment
4393 ++new->current_threshold;
4397 /* Free old spare buffer and save old primary buffer as spare */
4398 kfree(thresholds->spare);
4399 thresholds->spare = thresholds->primary;
4401 rcu_assign_pointer(thresholds->primary, new);
4403 /* To be sure that nobody uses thresholds */
4407 mutex_unlock(&memcg->thresholds_lock);
4412 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
4413 struct cftype *cft, struct eventfd_ctx *eventfd)
4415 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4416 struct mem_cgroup_thresholds *thresholds;
4417 struct mem_cgroup_threshold_ary *new;
4418 int type = MEMFILE_TYPE(cft->private);
4422 mutex_lock(&memcg->thresholds_lock);
4424 thresholds = &memcg->thresholds;
4425 else if (type == _MEMSWAP)
4426 thresholds = &memcg->memsw_thresholds;
4431 * Something went wrong if we trying to unregister a threshold
4432 * if we don't have thresholds
4434 BUG_ON(!thresholds);
4436 if (!thresholds->primary)
4439 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4441 /* Check if a threshold crossed before removing */
4442 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4444 /* Calculate new number of threshold */
4446 for (i = 0; i < thresholds->primary->size; i++) {
4447 if (thresholds->primary->entries[i].eventfd != eventfd)
4451 new = thresholds->spare;
4453 /* Set thresholds array to NULL if we don't have thresholds */
4462 /* Copy thresholds and find current threshold */
4463 new->current_threshold = -1;
4464 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4465 if (thresholds->primary->entries[i].eventfd == eventfd)
4468 new->entries[j] = thresholds->primary->entries[i];
4469 if (new->entries[j].threshold < usage) {
4471 * new->current_threshold will not be used
4472 * until rcu_assign_pointer(), so it's safe to increment
4475 ++new->current_threshold;
4481 /* Swap primary and spare array */
4482 thresholds->spare = thresholds->primary;
4483 rcu_assign_pointer(thresholds->primary, new);
4485 /* To be sure that nobody uses thresholds */
4488 mutex_unlock(&memcg->thresholds_lock);
4491 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4492 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4494 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4495 struct mem_cgroup_eventfd_list *event;
4496 int type = MEMFILE_TYPE(cft->private);
4498 BUG_ON(type != _OOM_TYPE);
4499 event = kmalloc(sizeof(*event), GFP_KERNEL);
4503 spin_lock(&memcg_oom_lock);
4505 event->eventfd = eventfd;
4506 list_add(&event->list, &memcg->oom_notify);
4508 /* already in OOM ? */
4509 if (atomic_read(&memcg->under_oom))
4510 eventfd_signal(eventfd, 1);
4511 spin_unlock(&memcg_oom_lock);
4516 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4517 struct cftype *cft, struct eventfd_ctx *eventfd)
4519 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4520 struct mem_cgroup_eventfd_list *ev, *tmp;
4521 int type = MEMFILE_TYPE(cft->private);
4523 BUG_ON(type != _OOM_TYPE);
4525 spin_lock(&memcg_oom_lock);
4527 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4528 if (ev->eventfd == eventfd) {
4529 list_del(&ev->list);
4534 spin_unlock(&memcg_oom_lock);
4537 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4538 struct cftype *cft, struct cgroup_map_cb *cb)
4540 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4542 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
4544 if (atomic_read(&memcg->under_oom))
4545 cb->fill(cb, "under_oom", 1);
4547 cb->fill(cb, "under_oom", 0);
4551 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4552 struct cftype *cft, u64 val)
4554 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4555 struct mem_cgroup *parent;
4557 /* cannot set to root cgroup and only 0 and 1 are allowed */
4558 if (!cgrp->parent || !((val == 0) || (val == 1)))
4561 parent = mem_cgroup_from_cont(cgrp->parent);
4564 /* oom-kill-disable is a flag for subhierarchy. */
4565 if ((parent->use_hierarchy) ||
4566 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4570 memcg->oom_kill_disable = val;
4572 memcg_oom_recover(memcg);
4578 static const struct file_operations mem_control_numa_stat_file_operations = {
4580 .llseek = seq_lseek,
4581 .release = single_release,
4584 static int mem_control_numa_stat_open(struct inode *unused, struct file *file)
4586 struct cgroup *cont = file->f_dentry->d_parent->d_fsdata;
4588 file->f_op = &mem_control_numa_stat_file_operations;
4589 return single_open(file, mem_control_numa_stat_show, cont);
4591 #endif /* CONFIG_NUMA */
4593 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_KMEM
4594 static int register_kmem_files(struct cgroup *cont, struct cgroup_subsys *ss)
4597 * Part of this would be better living in a separate allocation
4598 * function, leaving us with just the cgroup tree population work.
4599 * We, however, depend on state such as network's proto_list that
4600 * is only initialized after cgroup creation. I found the less
4601 * cumbersome way to deal with it to defer it all to populate time
4603 return mem_cgroup_sockets_init(cont, ss);
4606 static void kmem_cgroup_destroy(struct cgroup *cont)
4608 mem_cgroup_sockets_destroy(cont);
4611 static int register_kmem_files(struct cgroup *cont, struct cgroup_subsys *ss)
4616 static void kmem_cgroup_destroy(struct cgroup *cont)
4621 static struct cftype mem_cgroup_files[] = {
4623 .name = "usage_in_bytes",
4624 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4625 .read_u64 = mem_cgroup_read,
4626 .register_event = mem_cgroup_usage_register_event,
4627 .unregister_event = mem_cgroup_usage_unregister_event,
4630 .name = "max_usage_in_bytes",
4631 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4632 .trigger = mem_cgroup_reset,
4633 .read_u64 = mem_cgroup_read,
4636 .name = "limit_in_bytes",
4637 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4638 .write_string = mem_cgroup_write,
4639 .read_u64 = mem_cgroup_read,
4642 .name = "soft_limit_in_bytes",
4643 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4644 .write_string = mem_cgroup_write,
4645 .read_u64 = mem_cgroup_read,
4649 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4650 .trigger = mem_cgroup_reset,
4651 .read_u64 = mem_cgroup_read,
4655 .read_map = mem_control_stat_show,
4658 .name = "force_empty",
4659 .trigger = mem_cgroup_force_empty_write,
4662 .name = "use_hierarchy",
4663 .write_u64 = mem_cgroup_hierarchy_write,
4664 .read_u64 = mem_cgroup_hierarchy_read,
4667 .name = "swappiness",
4668 .read_u64 = mem_cgroup_swappiness_read,
4669 .write_u64 = mem_cgroup_swappiness_write,
4672 .name = "move_charge_at_immigrate",
4673 .read_u64 = mem_cgroup_move_charge_read,
4674 .write_u64 = mem_cgroup_move_charge_write,
4677 .name = "oom_control",
4678 .read_map = mem_cgroup_oom_control_read,
4679 .write_u64 = mem_cgroup_oom_control_write,
4680 .register_event = mem_cgroup_oom_register_event,
4681 .unregister_event = mem_cgroup_oom_unregister_event,
4682 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4686 .name = "numa_stat",
4687 .open = mem_control_numa_stat_open,
4693 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4694 static struct cftype memsw_cgroup_files[] = {
4696 .name = "memsw.usage_in_bytes",
4697 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4698 .read_u64 = mem_cgroup_read,
4699 .register_event = mem_cgroup_usage_register_event,
4700 .unregister_event = mem_cgroup_usage_unregister_event,
4703 .name = "memsw.max_usage_in_bytes",
4704 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4705 .trigger = mem_cgroup_reset,
4706 .read_u64 = mem_cgroup_read,
4709 .name = "memsw.limit_in_bytes",
4710 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4711 .write_string = mem_cgroup_write,
4712 .read_u64 = mem_cgroup_read,
4715 .name = "memsw.failcnt",
4716 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4717 .trigger = mem_cgroup_reset,
4718 .read_u64 = mem_cgroup_read,
4722 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4724 if (!do_swap_account)
4726 return cgroup_add_files(cont, ss, memsw_cgroup_files,
4727 ARRAY_SIZE(memsw_cgroup_files));
4730 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4736 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4738 struct mem_cgroup_per_node *pn;
4739 struct mem_cgroup_per_zone *mz;
4741 int zone, tmp = node;
4743 * This routine is called against possible nodes.
4744 * But it's BUG to call kmalloc() against offline node.
4746 * TODO: this routine can waste much memory for nodes which will
4747 * never be onlined. It's better to use memory hotplug callback
4750 if (!node_state(node, N_NORMAL_MEMORY))
4752 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4756 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4757 mz = &pn->zoneinfo[zone];
4759 INIT_LIST_HEAD(&mz->lruvec.lists[lru]);
4760 mz->usage_in_excess = 0;
4761 mz->on_tree = false;
4764 memcg->info.nodeinfo[node] = pn;
4768 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4770 kfree(memcg->info.nodeinfo[node]);
4773 static struct mem_cgroup *mem_cgroup_alloc(void)
4775 struct mem_cgroup *memcg;
4776 int size = sizeof(struct mem_cgroup);
4778 /* Can be very big if MAX_NUMNODES is very big */
4779 if (size < PAGE_SIZE)
4780 memcg = kzalloc(size, GFP_KERNEL);
4782 memcg = vzalloc(size);
4787 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4790 spin_lock_init(&memcg->pcp_counter_lock);
4794 if (size < PAGE_SIZE)
4802 * Helpers for freeing a vzalloc()ed mem_cgroup by RCU,
4803 * but in process context. The work_freeing structure is overlaid
4804 * on the rcu_freeing structure, which itself is overlaid on memsw.
4806 static void vfree_work(struct work_struct *work)
4808 struct mem_cgroup *memcg;
4810 memcg = container_of(work, struct mem_cgroup, work_freeing);
4813 static void vfree_rcu(struct rcu_head *rcu_head)
4815 struct mem_cgroup *memcg;
4817 memcg = container_of(rcu_head, struct mem_cgroup, rcu_freeing);
4818 INIT_WORK(&memcg->work_freeing, vfree_work);
4819 schedule_work(&memcg->work_freeing);
4823 * At destroying mem_cgroup, references from swap_cgroup can remain.
4824 * (scanning all at force_empty is too costly...)
4826 * Instead of clearing all references at force_empty, we remember
4827 * the number of reference from swap_cgroup and free mem_cgroup when
4828 * it goes down to 0.
4830 * Removal of cgroup itself succeeds regardless of refs from swap.
4833 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4837 mem_cgroup_remove_from_trees(memcg);
4838 free_css_id(&mem_cgroup_subsys, &memcg->css);
4841 free_mem_cgroup_per_zone_info(memcg, node);
4843 free_percpu(memcg->stat);
4844 if (sizeof(struct mem_cgroup) < PAGE_SIZE)
4845 kfree_rcu(memcg, rcu_freeing);
4847 call_rcu(&memcg->rcu_freeing, vfree_rcu);
4850 static void mem_cgroup_get(struct mem_cgroup *memcg)
4852 atomic_inc(&memcg->refcnt);
4855 static void __mem_cgroup_put(struct mem_cgroup *memcg, int count)
4857 if (atomic_sub_and_test(count, &memcg->refcnt)) {
4858 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4859 __mem_cgroup_free(memcg);
4861 mem_cgroup_put(parent);
4865 static void mem_cgroup_put(struct mem_cgroup *memcg)
4867 __mem_cgroup_put(memcg, 1);
4871 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4873 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4875 if (!memcg->res.parent)
4877 return mem_cgroup_from_res_counter(memcg->res.parent, res);
4879 EXPORT_SYMBOL(parent_mem_cgroup);
4881 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4882 static void __init enable_swap_cgroup(void)
4884 if (!mem_cgroup_disabled() && really_do_swap_account)
4885 do_swap_account = 1;
4888 static void __init enable_swap_cgroup(void)
4893 static int mem_cgroup_soft_limit_tree_init(void)
4895 struct mem_cgroup_tree_per_node *rtpn;
4896 struct mem_cgroup_tree_per_zone *rtpz;
4897 int tmp, node, zone;
4899 for_each_node(node) {
4901 if (!node_state(node, N_NORMAL_MEMORY))
4903 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
4907 soft_limit_tree.rb_tree_per_node[node] = rtpn;
4909 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4910 rtpz = &rtpn->rb_tree_per_zone[zone];
4911 rtpz->rb_root = RB_ROOT;
4912 spin_lock_init(&rtpz->lock);
4918 for_each_node(node) {
4919 if (!soft_limit_tree.rb_tree_per_node[node])
4921 kfree(soft_limit_tree.rb_tree_per_node[node]);
4922 soft_limit_tree.rb_tree_per_node[node] = NULL;
4928 static struct cgroup_subsys_state * __ref
4929 mem_cgroup_create(struct cgroup *cont)
4931 struct mem_cgroup *memcg, *parent;
4932 long error = -ENOMEM;
4935 memcg = mem_cgroup_alloc();
4937 return ERR_PTR(error);
4940 if (alloc_mem_cgroup_per_zone_info(memcg, node))
4944 if (cont->parent == NULL) {
4946 enable_swap_cgroup();
4948 if (mem_cgroup_soft_limit_tree_init())
4950 root_mem_cgroup = memcg;
4951 for_each_possible_cpu(cpu) {
4952 struct memcg_stock_pcp *stock =
4953 &per_cpu(memcg_stock, cpu);
4954 INIT_WORK(&stock->work, drain_local_stock);
4956 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
4958 parent = mem_cgroup_from_cont(cont->parent);
4959 memcg->use_hierarchy = parent->use_hierarchy;
4960 memcg->oom_kill_disable = parent->oom_kill_disable;
4963 if (parent && parent->use_hierarchy) {
4964 res_counter_init(&memcg->res, &parent->res);
4965 res_counter_init(&memcg->memsw, &parent->memsw);
4967 * We increment refcnt of the parent to ensure that we can
4968 * safely access it on res_counter_charge/uncharge.
4969 * This refcnt will be decremented when freeing this
4970 * mem_cgroup(see mem_cgroup_put).
4972 mem_cgroup_get(parent);
4974 res_counter_init(&memcg->res, NULL);
4975 res_counter_init(&memcg->memsw, NULL);
4977 memcg->last_scanned_node = MAX_NUMNODES;
4978 INIT_LIST_HEAD(&memcg->oom_notify);
4981 memcg->swappiness = mem_cgroup_swappiness(parent);
4982 atomic_set(&memcg->refcnt, 1);
4983 memcg->move_charge_at_immigrate = 0;
4984 mutex_init(&memcg->thresholds_lock);
4987 __mem_cgroup_free(memcg);
4988 return ERR_PTR(error);
4991 static int mem_cgroup_pre_destroy(struct cgroup *cont)
4993 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4995 return mem_cgroup_force_empty(memcg, false);
4998 static void mem_cgroup_destroy(struct cgroup *cont)
5000 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5002 kmem_cgroup_destroy(cont);
5004 mem_cgroup_put(memcg);
5007 static int mem_cgroup_populate(struct cgroup_subsys *ss,
5008 struct cgroup *cont)
5012 ret = cgroup_add_files(cont, ss, mem_cgroup_files,
5013 ARRAY_SIZE(mem_cgroup_files));
5016 ret = register_memsw_files(cont, ss);
5019 ret = register_kmem_files(cont, ss);
5025 /* Handlers for move charge at task migration. */
5026 #define PRECHARGE_COUNT_AT_ONCE 256
5027 static int mem_cgroup_do_precharge(unsigned long count)
5030 int batch_count = PRECHARGE_COUNT_AT_ONCE;
5031 struct mem_cgroup *memcg = mc.to;
5033 if (mem_cgroup_is_root(memcg)) {
5034 mc.precharge += count;
5035 /* we don't need css_get for root */
5038 /* try to charge at once */
5040 struct res_counter *dummy;
5042 * "memcg" cannot be under rmdir() because we've already checked
5043 * by cgroup_lock_live_cgroup() that it is not removed and we
5044 * are still under the same cgroup_mutex. So we can postpone
5047 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
5049 if (do_swap_account && res_counter_charge(&memcg->memsw,
5050 PAGE_SIZE * count, &dummy)) {
5051 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
5054 mc.precharge += count;
5058 /* fall back to one by one charge */
5060 if (signal_pending(current)) {
5064 if (!batch_count--) {
5065 batch_count = PRECHARGE_COUNT_AT_ONCE;
5068 ret = __mem_cgroup_try_charge(NULL,
5069 GFP_KERNEL, 1, &memcg, false);
5071 /* mem_cgroup_clear_mc() will do uncharge later */
5079 * is_target_pte_for_mc - check a pte whether it is valid for move charge
5080 * @vma: the vma the pte to be checked belongs
5081 * @addr: the address corresponding to the pte to be checked
5082 * @ptent: the pte to be checked
5083 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5086 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5087 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5088 * move charge. if @target is not NULL, the page is stored in target->page
5089 * with extra refcnt got(Callers should handle it).
5090 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5091 * target for charge migration. if @target is not NULL, the entry is stored
5094 * Called with pte lock held.
5101 enum mc_target_type {
5102 MC_TARGET_NONE, /* not used */
5107 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5108 unsigned long addr, pte_t ptent)
5110 struct page *page = vm_normal_page(vma, addr, ptent);
5112 if (!page || !page_mapped(page))
5114 if (PageAnon(page)) {
5115 /* we don't move shared anon */
5116 if (!move_anon() || page_mapcount(page) > 2)
5118 } else if (!move_file())
5119 /* we ignore mapcount for file pages */
5121 if (!get_page_unless_zero(page))
5127 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5128 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5131 struct page *page = NULL;
5132 swp_entry_t ent = pte_to_swp_entry(ptent);
5134 if (!move_anon() || non_swap_entry(ent))
5136 usage_count = mem_cgroup_count_swap_user(ent, &page);
5137 if (usage_count > 1) { /* we don't move shared anon */
5142 if (do_swap_account)
5143 entry->val = ent.val;
5148 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5149 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5151 struct page *page = NULL;
5152 struct inode *inode;
5153 struct address_space *mapping;
5156 if (!vma->vm_file) /* anonymous vma */
5161 inode = vma->vm_file->f_path.dentry->d_inode;
5162 mapping = vma->vm_file->f_mapping;
5163 if (pte_none(ptent))
5164 pgoff = linear_page_index(vma, addr);
5165 else /* pte_file(ptent) is true */
5166 pgoff = pte_to_pgoff(ptent);
5168 /* page is moved even if it's not RSS of this task(page-faulted). */
5169 page = find_get_page(mapping, pgoff);
5172 /* shmem/tmpfs may report page out on swap: account for that too. */
5173 if (radix_tree_exceptional_entry(page)) {
5174 swp_entry_t swap = radix_to_swp_entry(page);
5175 if (do_swap_account)
5177 page = find_get_page(&swapper_space, swap.val);
5183 static int is_target_pte_for_mc(struct vm_area_struct *vma,
5184 unsigned long addr, pte_t ptent, union mc_target *target)
5186 struct page *page = NULL;
5187 struct page_cgroup *pc;
5189 swp_entry_t ent = { .val = 0 };
5191 if (pte_present(ptent))
5192 page = mc_handle_present_pte(vma, addr, ptent);
5193 else if (is_swap_pte(ptent))
5194 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5195 else if (pte_none(ptent) || pte_file(ptent))
5196 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5198 if (!page && !ent.val)
5201 pc = lookup_page_cgroup(page);
5203 * Do only loose check w/o page_cgroup lock.
5204 * mem_cgroup_move_account() checks the pc is valid or not under
5207 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5208 ret = MC_TARGET_PAGE;
5210 target->page = page;
5212 if (!ret || !target)
5215 /* There is a swap entry and a page doesn't exist or isn't charged */
5216 if (ent.val && !ret &&
5217 css_id(&mc.from->css) == lookup_swap_cgroup_id(ent)) {
5218 ret = MC_TARGET_SWAP;
5225 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5226 unsigned long addr, unsigned long end,
5227 struct mm_walk *walk)
5229 struct vm_area_struct *vma = walk->private;
5233 split_huge_page_pmd(walk->mm, pmd);
5234 if (pmd_trans_unstable(pmd))
5237 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5238 for (; addr != end; pte++, addr += PAGE_SIZE)
5239 if (is_target_pte_for_mc(vma, addr, *pte, NULL))
5240 mc.precharge++; /* increment precharge temporarily */
5241 pte_unmap_unlock(pte - 1, ptl);
5247 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5249 unsigned long precharge;
5250 struct vm_area_struct *vma;
5252 down_read(&mm->mmap_sem);
5253 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5254 struct mm_walk mem_cgroup_count_precharge_walk = {
5255 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5259 if (is_vm_hugetlb_page(vma))
5261 walk_page_range(vma->vm_start, vma->vm_end,
5262 &mem_cgroup_count_precharge_walk);
5264 up_read(&mm->mmap_sem);
5266 precharge = mc.precharge;
5272 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5274 unsigned long precharge = mem_cgroup_count_precharge(mm);
5276 VM_BUG_ON(mc.moving_task);
5277 mc.moving_task = current;
5278 return mem_cgroup_do_precharge(precharge);
5281 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5282 static void __mem_cgroup_clear_mc(void)
5284 struct mem_cgroup *from = mc.from;
5285 struct mem_cgroup *to = mc.to;
5287 /* we must uncharge all the leftover precharges from mc.to */
5289 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
5293 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5294 * we must uncharge here.
5296 if (mc.moved_charge) {
5297 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
5298 mc.moved_charge = 0;
5300 /* we must fixup refcnts and charges */
5301 if (mc.moved_swap) {
5302 /* uncharge swap account from the old cgroup */
5303 if (!mem_cgroup_is_root(mc.from))
5304 res_counter_uncharge(&mc.from->memsw,
5305 PAGE_SIZE * mc.moved_swap);
5306 __mem_cgroup_put(mc.from, mc.moved_swap);
5308 if (!mem_cgroup_is_root(mc.to)) {
5310 * we charged both to->res and to->memsw, so we should
5313 res_counter_uncharge(&mc.to->res,
5314 PAGE_SIZE * mc.moved_swap);
5316 /* we've already done mem_cgroup_get(mc.to) */
5319 memcg_oom_recover(from);
5320 memcg_oom_recover(to);
5321 wake_up_all(&mc.waitq);
5324 static void mem_cgroup_clear_mc(void)
5326 struct mem_cgroup *from = mc.from;
5329 * we must clear moving_task before waking up waiters at the end of
5332 mc.moving_task = NULL;
5333 __mem_cgroup_clear_mc();
5334 spin_lock(&mc.lock);
5337 spin_unlock(&mc.lock);
5338 mem_cgroup_end_move(from);
5341 static int mem_cgroup_can_attach(struct cgroup *cgroup,
5342 struct cgroup_taskset *tset)
5344 struct task_struct *p = cgroup_taskset_first(tset);
5346 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup);
5348 if (memcg->move_charge_at_immigrate) {
5349 struct mm_struct *mm;
5350 struct mem_cgroup *from = mem_cgroup_from_task(p);
5352 VM_BUG_ON(from == memcg);
5354 mm = get_task_mm(p);
5357 /* We move charges only when we move a owner of the mm */
5358 if (mm->owner == p) {
5361 VM_BUG_ON(mc.precharge);
5362 VM_BUG_ON(mc.moved_charge);
5363 VM_BUG_ON(mc.moved_swap);
5364 mem_cgroup_start_move(from);
5365 spin_lock(&mc.lock);
5368 spin_unlock(&mc.lock);
5369 /* We set mc.moving_task later */
5371 ret = mem_cgroup_precharge_mc(mm);
5373 mem_cgroup_clear_mc();
5380 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
5381 struct cgroup_taskset *tset)
5383 mem_cgroup_clear_mc();
5386 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5387 unsigned long addr, unsigned long end,
5388 struct mm_walk *walk)
5391 struct vm_area_struct *vma = walk->private;
5395 split_huge_page_pmd(walk->mm, pmd);
5396 if (pmd_trans_unstable(pmd))
5399 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5400 for (; addr != end; addr += PAGE_SIZE) {
5401 pte_t ptent = *(pte++);
5402 union mc_target target;
5405 struct page_cgroup *pc;
5411 type = is_target_pte_for_mc(vma, addr, ptent, &target);
5413 case MC_TARGET_PAGE:
5415 if (isolate_lru_page(page))
5417 pc = lookup_page_cgroup(page);
5418 if (!mem_cgroup_move_account(page, 1, pc,
5419 mc.from, mc.to, false)) {
5421 /* we uncharge from mc.from later. */
5424 putback_lru_page(page);
5425 put: /* is_target_pte_for_mc() gets the page */
5428 case MC_TARGET_SWAP:
5430 if (!mem_cgroup_move_swap_account(ent,
5431 mc.from, mc.to, false)) {
5433 /* we fixup refcnts and charges later. */
5441 pte_unmap_unlock(pte - 1, ptl);
5446 * We have consumed all precharges we got in can_attach().
5447 * We try charge one by one, but don't do any additional
5448 * charges to mc.to if we have failed in charge once in attach()
5451 ret = mem_cgroup_do_precharge(1);
5459 static void mem_cgroup_move_charge(struct mm_struct *mm)
5461 struct vm_area_struct *vma;
5463 lru_add_drain_all();
5465 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5467 * Someone who are holding the mmap_sem might be waiting in
5468 * waitq. So we cancel all extra charges, wake up all waiters,
5469 * and retry. Because we cancel precharges, we might not be able
5470 * to move enough charges, but moving charge is a best-effort
5471 * feature anyway, so it wouldn't be a big problem.
5473 __mem_cgroup_clear_mc();
5477 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5479 struct mm_walk mem_cgroup_move_charge_walk = {
5480 .pmd_entry = mem_cgroup_move_charge_pte_range,
5484 if (is_vm_hugetlb_page(vma))
5486 ret = walk_page_range(vma->vm_start, vma->vm_end,
5487 &mem_cgroup_move_charge_walk);
5490 * means we have consumed all precharges and failed in
5491 * doing additional charge. Just abandon here.
5495 up_read(&mm->mmap_sem);
5498 static void mem_cgroup_move_task(struct cgroup *cont,
5499 struct cgroup_taskset *tset)
5501 struct task_struct *p = cgroup_taskset_first(tset);
5502 struct mm_struct *mm = get_task_mm(p);
5506 mem_cgroup_move_charge(mm);
5511 mem_cgroup_clear_mc();
5513 #else /* !CONFIG_MMU */
5514 static int mem_cgroup_can_attach(struct cgroup *cgroup,
5515 struct cgroup_taskset *tset)
5519 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
5520 struct cgroup_taskset *tset)
5523 static void mem_cgroup_move_task(struct cgroup *cont,
5524 struct cgroup_taskset *tset)
5529 struct cgroup_subsys mem_cgroup_subsys = {
5531 .subsys_id = mem_cgroup_subsys_id,
5532 .create = mem_cgroup_create,
5533 .pre_destroy = mem_cgroup_pre_destroy,
5534 .destroy = mem_cgroup_destroy,
5535 .populate = mem_cgroup_populate,
5536 .can_attach = mem_cgroup_can_attach,
5537 .cancel_attach = mem_cgroup_cancel_attach,
5538 .attach = mem_cgroup_move_task,
5543 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5544 static int __init enable_swap_account(char *s)
5546 /* consider enabled if no parameter or 1 is given */
5547 if (!strcmp(s, "1"))
5548 really_do_swap_account = 1;
5549 else if (!strcmp(s, "0"))
5550 really_do_swap_account = 0;
5553 __setup("swapaccount=", enable_swap_account);