1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * This program is free software; you can redistribute it and/or modify
14 * it under the terms of the GNU General Public License as published by
15 * the Free Software Foundation; either version 2 of the License, or
16 * (at your option) any later version.
18 * This program is distributed in the hope that it will be useful,
19 * but WITHOUT ANY WARRANTY; without even the implied warranty of
20 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
21 * GNU General Public License for more details.
24 #include <linux/res_counter.h>
25 #include <linux/memcontrol.h>
26 #include <linux/cgroup.h>
28 #include <linux/hugetlb.h>
29 #include <linux/pagemap.h>
30 #include <linux/smp.h>
31 #include <linux/page-flags.h>
32 #include <linux/backing-dev.h>
33 #include <linux/bit_spinlock.h>
34 #include <linux/rcupdate.h>
35 #include <linux/limits.h>
36 #include <linux/export.h>
37 #include <linux/mutex.h>
38 #include <linux/rbtree.h>
39 #include <linux/slab.h>
40 #include <linux/swap.h>
41 #include <linux/swapops.h>
42 #include <linux/spinlock.h>
43 #include <linux/eventfd.h>
44 #include <linux/sort.h>
46 #include <linux/seq_file.h>
47 #include <linux/vmalloc.h>
48 #include <linux/mm_inline.h>
49 #include <linux/page_cgroup.h>
50 #include <linux/cpu.h>
51 #include <linux/oom.h>
54 #include <net/tcp_memcontrol.h>
56 #include <asm/uaccess.h>
58 #include <trace/events/vmscan.h>
60 struct cgroup_subsys mem_cgroup_subsys __read_mostly;
61 #define MEM_CGROUP_RECLAIM_RETRIES 5
62 static struct mem_cgroup *root_mem_cgroup __read_mostly;
64 #ifdef CONFIG_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_STAT_NSTATS,
95 enum mem_cgroup_events_index {
96 MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */
97 MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */
98 MEM_CGROUP_EVENTS_COUNT, /* # of pages paged in/out */
99 MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */
100 MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */
101 MEM_CGROUP_EVENTS_NSTATS,
104 * Per memcg event counter is incremented at every pagein/pageout. With THP,
105 * it will be incremated by the number of pages. This counter is used for
106 * for trigger some periodic events. This is straightforward and better
107 * than using jiffies etc. to handle periodic memcg event.
109 enum mem_cgroup_events_target {
110 MEM_CGROUP_TARGET_THRESH,
111 MEM_CGROUP_TARGET_SOFTLIMIT,
112 MEM_CGROUP_TARGET_NUMAINFO,
115 #define THRESHOLDS_EVENTS_TARGET 128
116 #define SOFTLIMIT_EVENTS_TARGET 1024
117 #define NUMAINFO_EVENTS_TARGET 1024
119 struct mem_cgroup_stat_cpu {
120 long count[MEM_CGROUP_STAT_NSTATS];
121 unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
122 unsigned long targets[MEM_CGROUP_NTARGETS];
125 struct mem_cgroup_reclaim_iter {
126 /* css_id of the last scanned hierarchy member */
128 /* scan generation, increased every round-trip */
129 unsigned int generation;
133 * per-zone information in memory controller.
135 struct mem_cgroup_per_zone {
136 struct lruvec lruvec;
137 unsigned long lru_size[NR_LRU_LISTS];
139 struct mem_cgroup_reclaim_iter reclaim_iter[DEF_PRIORITY + 1];
141 struct rb_node tree_node; /* RB tree node */
142 unsigned long long usage_in_excess;/* Set to the value by which */
143 /* the soft limit is exceeded*/
145 struct mem_cgroup *memcg; /* Back pointer, we cannot */
146 /* use container_of */
149 struct mem_cgroup_per_node {
150 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
153 struct mem_cgroup_lru_info {
154 struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
158 * Cgroups above their limits are maintained in a RB-Tree, independent of
159 * their hierarchy representation
162 struct mem_cgroup_tree_per_zone {
163 struct rb_root rb_root;
167 struct mem_cgroup_tree_per_node {
168 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
171 struct mem_cgroup_tree {
172 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
175 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
177 struct mem_cgroup_threshold {
178 struct eventfd_ctx *eventfd;
183 struct mem_cgroup_threshold_ary {
184 /* An array index points to threshold just below or equal to usage. */
185 int current_threshold;
186 /* Size of entries[] */
188 /* Array of thresholds */
189 struct mem_cgroup_threshold entries[0];
192 struct mem_cgroup_thresholds {
193 /* Primary thresholds array */
194 struct mem_cgroup_threshold_ary *primary;
196 * Spare threshold array.
197 * This is needed to make mem_cgroup_unregister_event() "never fail".
198 * It must be able to store at least primary->size - 1 entries.
200 struct mem_cgroup_threshold_ary *spare;
204 struct mem_cgroup_eventfd_list {
205 struct list_head list;
206 struct eventfd_ctx *eventfd;
209 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
210 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
213 * The memory controller data structure. The memory controller controls both
214 * page cache and RSS per cgroup. We would eventually like to provide
215 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
216 * to help the administrator determine what knobs to tune.
218 * TODO: Add a water mark for the memory controller. Reclaim will begin when
219 * we hit the water mark. May be even add a low water mark, such that
220 * no reclaim occurs from a cgroup at it's low water mark, this is
221 * a feature that will be implemented much later in the future.
224 struct cgroup_subsys_state css;
226 * the counter to account for memory usage
228 struct res_counter res;
232 * the counter to account for mem+swap usage.
234 struct res_counter memsw;
237 * rcu_freeing is used only when freeing struct mem_cgroup,
238 * so put it into a union to avoid wasting more memory.
239 * It must be disjoint from the css field. It could be
240 * in a union with the res field, but res plays a much
241 * larger part in mem_cgroup life than memsw, and might
242 * be of interest, even at time of free, when debugging.
243 * So share rcu_head with the less interesting memsw.
245 struct rcu_head rcu_freeing;
247 * But when using vfree(), that cannot be done at
248 * interrupt time, so we must then queue the work.
250 struct work_struct work_freeing;
254 * Per cgroup active and inactive list, similar to the
255 * per zone LRU lists.
257 struct mem_cgroup_lru_info info;
258 int last_scanned_node;
260 nodemask_t scan_nodes;
261 atomic_t numainfo_events;
262 atomic_t numainfo_updating;
265 * Should the accounting and control be hierarchical, per subtree?
275 /* OOM-Killer disable */
276 int oom_kill_disable;
278 /* set when res.limit == memsw.limit */
279 bool memsw_is_minimum;
281 /* protect arrays of thresholds */
282 struct mutex thresholds_lock;
284 /* thresholds for memory usage. RCU-protected */
285 struct mem_cgroup_thresholds thresholds;
287 /* thresholds for mem+swap usage. RCU-protected */
288 struct mem_cgroup_thresholds memsw_thresholds;
290 /* For oom notifier event fd */
291 struct list_head oom_notify;
294 * Should we move charges of a task when a task is moved into this
295 * mem_cgroup ? And what type of charges should we move ?
297 unsigned long move_charge_at_immigrate;
299 * set > 0 if pages under this cgroup are moving to other cgroup.
301 atomic_t moving_account;
302 /* taken only while moving_account > 0 */
303 spinlock_t move_lock;
307 struct mem_cgroup_stat_cpu __percpu *stat;
309 * used when a cpu is offlined or other synchronizations
310 * See mem_cgroup_read_stat().
312 struct mem_cgroup_stat_cpu nocpu_base;
313 spinlock_t pcp_counter_lock;
316 struct tcp_memcontrol tcp_mem;
320 /* Stuffs for move charges at task migration. */
322 * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
323 * left-shifted bitmap of these types.
326 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
327 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
331 /* "mc" and its members are protected by cgroup_mutex */
332 static struct move_charge_struct {
333 spinlock_t lock; /* for from, to */
334 struct mem_cgroup *from;
335 struct mem_cgroup *to;
336 unsigned long precharge;
337 unsigned long moved_charge;
338 unsigned long moved_swap;
339 struct task_struct *moving_task; /* a task moving charges */
340 wait_queue_head_t waitq; /* a waitq for other context */
342 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
343 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
346 static bool move_anon(void)
348 return test_bit(MOVE_CHARGE_TYPE_ANON,
349 &mc.to->move_charge_at_immigrate);
352 static bool move_file(void)
354 return test_bit(MOVE_CHARGE_TYPE_FILE,
355 &mc.to->move_charge_at_immigrate);
359 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
360 * limit reclaim to prevent infinite loops, if they ever occur.
362 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
363 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
366 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
367 MEM_CGROUP_CHARGE_TYPE_MAPPED,
368 MEM_CGROUP_CHARGE_TYPE_SHMEM, /* used by page migration of shmem */
369 MEM_CGROUP_CHARGE_TYPE_FORCE, /* used by force_empty */
370 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
371 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
375 /* for encoding cft->private value on file */
378 #define _OOM_TYPE (2)
379 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
380 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
381 #define MEMFILE_ATTR(val) ((val) & 0xffff)
382 /* Used for OOM nofiier */
383 #define OOM_CONTROL (0)
386 * Reclaim flags for mem_cgroup_hierarchical_reclaim
388 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
389 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
390 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
391 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
393 static void mem_cgroup_get(struct mem_cgroup *memcg);
394 static void mem_cgroup_put(struct mem_cgroup *memcg);
396 /* Writing them here to avoid exposing memcg's inner layout */
397 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_KMEM
398 #include <net/sock.h>
401 static bool mem_cgroup_is_root(struct mem_cgroup *memcg);
402 void sock_update_memcg(struct sock *sk)
404 if (mem_cgroup_sockets_enabled) {
405 struct mem_cgroup *memcg;
407 BUG_ON(!sk->sk_prot->proto_cgroup);
409 /* Socket cloning can throw us here with sk_cgrp already
410 * filled. It won't however, necessarily happen from
411 * process context. So the test for root memcg given
412 * the current task's memcg won't help us in this case.
414 * Respecting the original socket's memcg is a better
415 * decision in this case.
418 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
419 mem_cgroup_get(sk->sk_cgrp->memcg);
424 memcg = mem_cgroup_from_task(current);
425 if (!mem_cgroup_is_root(memcg)) {
426 mem_cgroup_get(memcg);
427 sk->sk_cgrp = sk->sk_prot->proto_cgroup(memcg);
432 EXPORT_SYMBOL(sock_update_memcg);
434 void sock_release_memcg(struct sock *sk)
436 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
437 struct mem_cgroup *memcg;
438 WARN_ON(!sk->sk_cgrp->memcg);
439 memcg = sk->sk_cgrp->memcg;
440 mem_cgroup_put(memcg);
445 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
447 if (!memcg || mem_cgroup_is_root(memcg))
450 return &memcg->tcp_mem.cg_proto;
452 EXPORT_SYMBOL(tcp_proto_cgroup);
453 #endif /* CONFIG_INET */
454 #endif /* CONFIG_CGROUP_MEM_RES_CTLR_KMEM */
456 static void drain_all_stock_async(struct mem_cgroup *memcg);
458 static struct mem_cgroup_per_zone *
459 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
461 return &memcg->info.nodeinfo[nid]->zoneinfo[zid];
464 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
469 static struct mem_cgroup_per_zone *
470 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
472 int nid = page_to_nid(page);
473 int zid = page_zonenum(page);
475 return mem_cgroup_zoneinfo(memcg, nid, zid);
478 static struct mem_cgroup_tree_per_zone *
479 soft_limit_tree_node_zone(int nid, int zid)
481 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
484 static struct mem_cgroup_tree_per_zone *
485 soft_limit_tree_from_page(struct page *page)
487 int nid = page_to_nid(page);
488 int zid = page_zonenum(page);
490 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
494 __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
495 struct mem_cgroup_per_zone *mz,
496 struct mem_cgroup_tree_per_zone *mctz,
497 unsigned long long new_usage_in_excess)
499 struct rb_node **p = &mctz->rb_root.rb_node;
500 struct rb_node *parent = NULL;
501 struct mem_cgroup_per_zone *mz_node;
506 mz->usage_in_excess = new_usage_in_excess;
507 if (!mz->usage_in_excess)
511 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
513 if (mz->usage_in_excess < mz_node->usage_in_excess)
516 * We can't avoid mem cgroups that are over their soft
517 * limit by the same amount
519 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
522 rb_link_node(&mz->tree_node, parent, p);
523 rb_insert_color(&mz->tree_node, &mctz->rb_root);
528 __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
529 struct mem_cgroup_per_zone *mz,
530 struct mem_cgroup_tree_per_zone *mctz)
534 rb_erase(&mz->tree_node, &mctz->rb_root);
539 mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
540 struct mem_cgroup_per_zone *mz,
541 struct mem_cgroup_tree_per_zone *mctz)
543 spin_lock(&mctz->lock);
544 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
545 spin_unlock(&mctz->lock);
549 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
551 unsigned long long excess;
552 struct mem_cgroup_per_zone *mz;
553 struct mem_cgroup_tree_per_zone *mctz;
554 int nid = page_to_nid(page);
555 int zid = page_zonenum(page);
556 mctz = soft_limit_tree_from_page(page);
559 * Necessary to update all ancestors when hierarchy is used.
560 * because their event counter is not touched.
562 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
563 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
564 excess = res_counter_soft_limit_excess(&memcg->res);
566 * We have to update the tree if mz is on RB-tree or
567 * mem is over its softlimit.
569 if (excess || mz->on_tree) {
570 spin_lock(&mctz->lock);
571 /* if on-tree, remove it */
573 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
575 * Insert again. mz->usage_in_excess will be updated.
576 * If excess is 0, no tree ops.
578 __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
579 spin_unlock(&mctz->lock);
584 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
587 struct mem_cgroup_per_zone *mz;
588 struct mem_cgroup_tree_per_zone *mctz;
590 for_each_node(node) {
591 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
592 mz = mem_cgroup_zoneinfo(memcg, node, zone);
593 mctz = soft_limit_tree_node_zone(node, zone);
594 mem_cgroup_remove_exceeded(memcg, mz, mctz);
599 static struct mem_cgroup_per_zone *
600 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
602 struct rb_node *rightmost = NULL;
603 struct mem_cgroup_per_zone *mz;
607 rightmost = rb_last(&mctz->rb_root);
609 goto done; /* Nothing to reclaim from */
611 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
613 * Remove the node now but someone else can add it back,
614 * we will to add it back at the end of reclaim to its correct
615 * position in the tree.
617 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
618 if (!res_counter_soft_limit_excess(&mz->memcg->res) ||
619 !css_tryget(&mz->memcg->css))
625 static struct mem_cgroup_per_zone *
626 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
628 struct mem_cgroup_per_zone *mz;
630 spin_lock(&mctz->lock);
631 mz = __mem_cgroup_largest_soft_limit_node(mctz);
632 spin_unlock(&mctz->lock);
637 * Implementation Note: reading percpu statistics for memcg.
639 * Both of vmstat[] and percpu_counter has threshold and do periodic
640 * synchronization to implement "quick" read. There are trade-off between
641 * reading cost and precision of value. Then, we may have a chance to implement
642 * a periodic synchronizion of counter in memcg's counter.
644 * But this _read() function is used for user interface now. The user accounts
645 * memory usage by memory cgroup and he _always_ requires exact value because
646 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
647 * have to visit all online cpus and make sum. So, for now, unnecessary
648 * synchronization is not implemented. (just implemented for cpu hotplug)
650 * If there are kernel internal actions which can make use of some not-exact
651 * value, and reading all cpu value can be performance bottleneck in some
652 * common workload, threashold and synchonization as vmstat[] should be
655 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
656 enum mem_cgroup_stat_index idx)
662 for_each_online_cpu(cpu)
663 val += per_cpu(memcg->stat->count[idx], cpu);
664 #ifdef CONFIG_HOTPLUG_CPU
665 spin_lock(&memcg->pcp_counter_lock);
666 val += memcg->nocpu_base.count[idx];
667 spin_unlock(&memcg->pcp_counter_lock);
673 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
676 int val = (charge) ? 1 : -1;
677 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAPOUT], val);
680 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
681 enum mem_cgroup_events_index idx)
683 unsigned long val = 0;
686 for_each_online_cpu(cpu)
687 val += per_cpu(memcg->stat->events[idx], cpu);
688 #ifdef CONFIG_HOTPLUG_CPU
689 spin_lock(&memcg->pcp_counter_lock);
690 val += memcg->nocpu_base.events[idx];
691 spin_unlock(&memcg->pcp_counter_lock);
696 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
697 bool anon, int nr_pages)
702 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
703 * counted as CACHE even if it's on ANON LRU.
706 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
709 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
712 /* pagein of a big page is an event. So, ignore page size */
714 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
716 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
717 nr_pages = -nr_pages; /* for event */
720 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT], nr_pages);
726 mem_cgroup_get_lruvec_size(struct lruvec *lruvec, enum lru_list lru)
728 struct mem_cgroup_per_zone *mz;
730 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
731 return mz->lru_size[lru];
735 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
736 unsigned int lru_mask)
738 struct mem_cgroup_per_zone *mz;
740 unsigned long ret = 0;
742 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
745 if (BIT(lru) & lru_mask)
746 ret += mz->lru_size[lru];
752 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
753 int nid, unsigned int lru_mask)
758 for (zid = 0; zid < MAX_NR_ZONES; zid++)
759 total += mem_cgroup_zone_nr_lru_pages(memcg,
765 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
766 unsigned int lru_mask)
771 for_each_node_state(nid, N_HIGH_MEMORY)
772 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
776 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
777 enum mem_cgroup_events_target target)
779 unsigned long val, next;
781 val = __this_cpu_read(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT]);
782 next = __this_cpu_read(memcg->stat->targets[target]);
783 /* from time_after() in jiffies.h */
784 if ((long)next - (long)val < 0) {
786 case MEM_CGROUP_TARGET_THRESH:
787 next = val + THRESHOLDS_EVENTS_TARGET;
789 case MEM_CGROUP_TARGET_SOFTLIMIT:
790 next = val + SOFTLIMIT_EVENTS_TARGET;
792 case MEM_CGROUP_TARGET_NUMAINFO:
793 next = val + NUMAINFO_EVENTS_TARGET;
798 __this_cpu_write(memcg->stat->targets[target], next);
805 * Check events in order.
808 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
811 /* threshold event is triggered in finer grain than soft limit */
812 if (unlikely(mem_cgroup_event_ratelimit(memcg,
813 MEM_CGROUP_TARGET_THRESH))) {
815 bool do_numainfo __maybe_unused;
817 do_softlimit = mem_cgroup_event_ratelimit(memcg,
818 MEM_CGROUP_TARGET_SOFTLIMIT);
820 do_numainfo = mem_cgroup_event_ratelimit(memcg,
821 MEM_CGROUP_TARGET_NUMAINFO);
825 mem_cgroup_threshold(memcg);
826 if (unlikely(do_softlimit))
827 mem_cgroup_update_tree(memcg, page);
829 if (unlikely(do_numainfo))
830 atomic_inc(&memcg->numainfo_events);
836 struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
838 return container_of(cgroup_subsys_state(cont,
839 mem_cgroup_subsys_id), struct mem_cgroup,
843 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
846 * mm_update_next_owner() may clear mm->owner to NULL
847 * if it races with swapoff, page migration, etc.
848 * So this can be called with p == NULL.
853 return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
854 struct mem_cgroup, css);
857 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
859 struct mem_cgroup *memcg = NULL;
864 * Because we have no locks, mm->owner's may be being moved to other
865 * cgroup. We use css_tryget() here even if this looks
866 * pessimistic (rather than adding locks here).
870 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
871 if (unlikely(!memcg))
873 } while (!css_tryget(&memcg->css));
879 * mem_cgroup_iter - iterate over memory cgroup hierarchy
880 * @root: hierarchy root
881 * @prev: previously returned memcg, NULL on first invocation
882 * @reclaim: cookie for shared reclaim walks, NULL for full walks
884 * Returns references to children of the hierarchy below @root, or
885 * @root itself, or %NULL after a full round-trip.
887 * Caller must pass the return value in @prev on subsequent
888 * invocations for reference counting, or use mem_cgroup_iter_break()
889 * to cancel a hierarchy walk before the round-trip is complete.
891 * Reclaimers can specify a zone and a priority level in @reclaim to
892 * divide up the memcgs in the hierarchy among all concurrent
893 * reclaimers operating on the same zone and priority.
895 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
896 struct mem_cgroup *prev,
897 struct mem_cgroup_reclaim_cookie *reclaim)
899 struct mem_cgroup *memcg = NULL;
902 if (mem_cgroup_disabled())
906 root = root_mem_cgroup;
908 if (prev && !reclaim)
909 id = css_id(&prev->css);
911 if (prev && prev != root)
914 if (!root->use_hierarchy && root != root_mem_cgroup) {
921 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
922 struct cgroup_subsys_state *css;
925 int nid = zone_to_nid(reclaim->zone);
926 int zid = zone_idx(reclaim->zone);
927 struct mem_cgroup_per_zone *mz;
929 mz = mem_cgroup_zoneinfo(root, nid, zid);
930 iter = &mz->reclaim_iter[reclaim->priority];
931 if (prev && reclaim->generation != iter->generation)
937 css = css_get_next(&mem_cgroup_subsys, id + 1, &root->css, &id);
939 if (css == &root->css || css_tryget(css))
940 memcg = container_of(css,
941 struct mem_cgroup, css);
950 else if (!prev && memcg)
951 reclaim->generation = iter->generation;
961 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
962 * @root: hierarchy root
963 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
965 void mem_cgroup_iter_break(struct mem_cgroup *root,
966 struct mem_cgroup *prev)
969 root = root_mem_cgroup;
970 if (prev && prev != root)
975 * Iteration constructs for visiting all cgroups (under a tree). If
976 * loops are exited prematurely (break), mem_cgroup_iter_break() must
977 * be used for reference counting.
979 #define for_each_mem_cgroup_tree(iter, root) \
980 for (iter = mem_cgroup_iter(root, NULL, NULL); \
982 iter = mem_cgroup_iter(root, iter, NULL))
984 #define for_each_mem_cgroup(iter) \
985 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
987 iter = mem_cgroup_iter(NULL, iter, NULL))
989 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
991 return (memcg == root_mem_cgroup);
994 void mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
996 struct mem_cgroup *memcg;
1002 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1003 if (unlikely(!memcg))
1008 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1011 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1019 EXPORT_SYMBOL(mem_cgroup_count_vm_event);
1022 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1023 * @zone: zone of the wanted lruvec
1024 * @mem: memcg of the wanted lruvec
1026 * Returns the lru list vector holding pages for the given @zone and
1027 * @mem. This can be the global zone lruvec, if the memory controller
1030 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1031 struct mem_cgroup *memcg)
1033 struct mem_cgroup_per_zone *mz;
1035 if (mem_cgroup_disabled())
1036 return &zone->lruvec;
1038 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1043 * Following LRU functions are allowed to be used without PCG_LOCK.
1044 * Operations are called by routine of global LRU independently from memcg.
1045 * What we have to take care of here is validness of pc->mem_cgroup.
1047 * Changes to pc->mem_cgroup happens when
1050 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1051 * It is added to LRU before charge.
1052 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1053 * When moving account, the page is not on LRU. It's isolated.
1057 * mem_cgroup_lru_add_list - account for adding an lru page and return lruvec
1058 * @zone: zone of the page
1062 * This function accounts for @page being added to @lru, and returns
1063 * the lruvec for the given @zone and the memcg @page is charged to.
1065 * The callsite is then responsible for physically linking the page to
1066 * the returned lruvec->lists[@lru].
1068 struct lruvec *mem_cgroup_lru_add_list(struct zone *zone, struct page *page,
1071 struct mem_cgroup_per_zone *mz;
1072 struct mem_cgroup *memcg;
1073 struct page_cgroup *pc;
1075 if (mem_cgroup_disabled())
1076 return &zone->lruvec;
1078 pc = lookup_page_cgroup(page);
1079 memcg = pc->mem_cgroup;
1082 * Surreptitiously switch any uncharged page to root:
1083 * an uncharged page off lru does nothing to secure
1084 * its former mem_cgroup from sudden removal.
1086 * Our caller holds lru_lock, and PageCgroupUsed is updated
1087 * under page_cgroup lock: between them, they make all uses
1088 * of pc->mem_cgroup safe.
1090 if (!PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1091 pc->mem_cgroup = memcg = root_mem_cgroup;
1093 mz = page_cgroup_zoneinfo(memcg, page);
1094 /* compound_order() is stabilized through lru_lock */
1095 mz->lru_size[lru] += 1 << compound_order(page);
1100 * mem_cgroup_lru_del_list - account for removing an lru page
1104 * This function accounts for @page being removed from @lru.
1106 * The callsite is then responsible for physically unlinking
1109 void mem_cgroup_lru_del_list(struct page *page, enum lru_list lru)
1111 struct mem_cgroup_per_zone *mz;
1112 struct mem_cgroup *memcg;
1113 struct page_cgroup *pc;
1115 if (mem_cgroup_disabled())
1118 pc = lookup_page_cgroup(page);
1119 memcg = pc->mem_cgroup;
1121 mz = page_cgroup_zoneinfo(memcg, page);
1122 /* huge page split is done under lru_lock. so, we have no races. */
1123 VM_BUG_ON(mz->lru_size[lru] < (1 << compound_order(page)));
1124 mz->lru_size[lru] -= 1 << compound_order(page);
1128 * mem_cgroup_lru_move_lists - account for moving a page between lrus
1129 * @zone: zone of the page
1131 * @from: current lru
1134 * This function accounts for @page being moved between the lrus @from
1135 * and @to, and returns the lruvec for the given @zone and the memcg
1136 * @page is charged to.
1138 * The callsite is then responsible for physically relinking
1139 * @page->lru to the returned lruvec->lists[@to].
1141 struct lruvec *mem_cgroup_lru_move_lists(struct zone *zone,
1146 /* XXX: Optimize this, especially for @from == @to */
1147 mem_cgroup_lru_del_list(page, from);
1148 return mem_cgroup_lru_add_list(zone, page, to);
1152 * Checks whether given mem is same or in the root_mem_cgroup's
1155 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1156 struct mem_cgroup *memcg)
1158 if (root_memcg == memcg)
1160 if (!root_memcg->use_hierarchy)
1162 return css_is_ancestor(&memcg->css, &root_memcg->css);
1165 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1166 struct mem_cgroup *memcg)
1171 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1176 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *memcg)
1179 struct mem_cgroup *curr = NULL;
1180 struct task_struct *p;
1182 p = find_lock_task_mm(task);
1184 curr = try_get_mem_cgroup_from_mm(p->mm);
1188 * All threads may have already detached their mm's, but the oom
1189 * killer still needs to detect if they have already been oom
1190 * killed to prevent needlessly killing additional tasks.
1193 curr = mem_cgroup_from_task(task);
1195 css_get(&curr->css);
1201 * We should check use_hierarchy of "memcg" not "curr". Because checking
1202 * use_hierarchy of "curr" here make this function true if hierarchy is
1203 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1204 * hierarchy(even if use_hierarchy is disabled in "memcg").
1206 ret = mem_cgroup_same_or_subtree(memcg, curr);
1207 css_put(&curr->css);
1211 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg, struct zone *zone)
1213 unsigned long inactive_ratio;
1214 int nid = zone_to_nid(zone);
1215 int zid = zone_idx(zone);
1216 unsigned long inactive;
1217 unsigned long active;
1220 inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1221 BIT(LRU_INACTIVE_ANON));
1222 active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1223 BIT(LRU_ACTIVE_ANON));
1225 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1227 inactive_ratio = int_sqrt(10 * gb);
1231 return inactive * inactive_ratio < active;
1234 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg, struct zone *zone)
1236 unsigned long active;
1237 unsigned long inactive;
1238 int zid = zone_idx(zone);
1239 int nid = zone_to_nid(zone);
1241 inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1242 BIT(LRU_INACTIVE_FILE));
1243 active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1244 BIT(LRU_ACTIVE_FILE));
1246 return (active > inactive);
1249 struct zone_reclaim_stat *
1250 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
1252 struct page_cgroup *pc;
1253 struct mem_cgroup_per_zone *mz;
1255 if (mem_cgroup_disabled())
1258 pc = lookup_page_cgroup(page);
1259 if (!PageCgroupUsed(pc))
1261 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1263 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1264 return &mz->lruvec.reclaim_stat;
1267 #define mem_cgroup_from_res_counter(counter, member) \
1268 container_of(counter, struct mem_cgroup, member)
1271 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1272 * @mem: the memory cgroup
1274 * Returns the maximum amount of memory @mem can be charged with, in
1277 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1279 unsigned long long margin;
1281 margin = res_counter_margin(&memcg->res);
1282 if (do_swap_account)
1283 margin = min(margin, res_counter_margin(&memcg->memsw));
1284 return margin >> PAGE_SHIFT;
1287 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1289 struct cgroup *cgrp = memcg->css.cgroup;
1292 if (cgrp->parent == NULL)
1293 return vm_swappiness;
1295 return memcg->swappiness;
1299 * memcg->moving_account is used for checking possibility that some thread is
1300 * calling move_account(). When a thread on CPU-A starts moving pages under
1301 * a memcg, other threads should check memcg->moving_account under
1302 * rcu_read_lock(), like this:
1306 * memcg->moving_account+1 if (memcg->mocing_account)
1308 * synchronize_rcu() update something.
1313 /* for quick checking without looking up memcg */
1314 atomic_t memcg_moving __read_mostly;
1316 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1318 atomic_inc(&memcg_moving);
1319 atomic_inc(&memcg->moving_account);
1323 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1326 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1327 * We check NULL in callee rather than caller.
1330 atomic_dec(&memcg_moving);
1331 atomic_dec(&memcg->moving_account);
1336 * 2 routines for checking "mem" is under move_account() or not.
1338 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1339 * is used for avoiding races in accounting. If true,
1340 * pc->mem_cgroup may be overwritten.
1342 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1343 * under hierarchy of moving cgroups. This is for
1344 * waiting at hith-memory prressure caused by "move".
1347 static bool mem_cgroup_stolen(struct mem_cgroup *memcg)
1349 VM_BUG_ON(!rcu_read_lock_held());
1350 return atomic_read(&memcg->moving_account) > 0;
1353 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1355 struct mem_cgroup *from;
1356 struct mem_cgroup *to;
1359 * Unlike task_move routines, we access mc.to, mc.from not under
1360 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1362 spin_lock(&mc.lock);
1368 ret = mem_cgroup_same_or_subtree(memcg, from)
1369 || mem_cgroup_same_or_subtree(memcg, to);
1371 spin_unlock(&mc.lock);
1375 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1377 if (mc.moving_task && current != mc.moving_task) {
1378 if (mem_cgroup_under_move(memcg)) {
1380 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1381 /* moving charge context might have finished. */
1384 finish_wait(&mc.waitq, &wait);
1392 * Take this lock when
1393 * - a code tries to modify page's memcg while it's USED.
1394 * - a code tries to modify page state accounting in a memcg.
1395 * see mem_cgroup_stolen(), too.
1397 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1398 unsigned long *flags)
1400 spin_lock_irqsave(&memcg->move_lock, *flags);
1403 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1404 unsigned long *flags)
1406 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1410 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1411 * @memcg: The memory cgroup that went over limit
1412 * @p: Task that is going to be killed
1414 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1417 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1419 struct cgroup *task_cgrp;
1420 struct cgroup *mem_cgrp;
1422 * Need a buffer in BSS, can't rely on allocations. The code relies
1423 * on the assumption that OOM is serialized for memory controller.
1424 * If this assumption is broken, revisit this code.
1426 static char memcg_name[PATH_MAX];
1434 mem_cgrp = memcg->css.cgroup;
1435 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1437 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1440 * Unfortunately, we are unable to convert to a useful name
1441 * But we'll still print out the usage information
1448 printk(KERN_INFO "Task in %s killed", memcg_name);
1451 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1459 * Continues from above, so we don't need an KERN_ level
1461 printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1464 printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1465 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1466 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1467 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1468 printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1470 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1471 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1472 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1476 * This function returns the number of memcg under hierarchy tree. Returns
1477 * 1(self count) if no children.
1479 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1482 struct mem_cgroup *iter;
1484 for_each_mem_cgroup_tree(iter, memcg)
1490 * Return the memory (and swap, if configured) limit for a memcg.
1492 u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1497 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1498 limit += total_swap_pages << PAGE_SHIFT;
1500 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1502 * If memsw is finite and limits the amount of swap space available
1503 * to this memcg, return that limit.
1505 return min(limit, memsw);
1508 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1510 unsigned long flags)
1512 unsigned long total = 0;
1513 bool noswap = false;
1516 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1518 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1521 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1523 drain_all_stock_async(memcg);
1524 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1526 * Allow limit shrinkers, which are triggered directly
1527 * by userspace, to catch signals and stop reclaim
1528 * after minimal progress, regardless of the margin.
1530 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1532 if (mem_cgroup_margin(memcg))
1535 * If nothing was reclaimed after two attempts, there
1536 * may be no reclaimable pages in this hierarchy.
1545 * test_mem_cgroup_node_reclaimable
1546 * @mem: the target memcg
1547 * @nid: the node ID to be checked.
1548 * @noswap : specify true here if the user wants flle only information.
1550 * This function returns whether the specified memcg contains any
1551 * reclaimable pages on a node. Returns true if there are any reclaimable
1552 * pages in the node.
1554 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1555 int nid, bool noswap)
1557 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1559 if (noswap || !total_swap_pages)
1561 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1566 #if MAX_NUMNODES > 1
1569 * Always updating the nodemask is not very good - even if we have an empty
1570 * list or the wrong list here, we can start from some node and traverse all
1571 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1574 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1578 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1579 * pagein/pageout changes since the last update.
1581 if (!atomic_read(&memcg->numainfo_events))
1583 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1586 /* make a nodemask where this memcg uses memory from */
1587 memcg->scan_nodes = node_states[N_HIGH_MEMORY];
1589 for_each_node_mask(nid, node_states[N_HIGH_MEMORY]) {
1591 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1592 node_clear(nid, memcg->scan_nodes);
1595 atomic_set(&memcg->numainfo_events, 0);
1596 atomic_set(&memcg->numainfo_updating, 0);
1600 * Selecting a node where we start reclaim from. Because what we need is just
1601 * reducing usage counter, start from anywhere is O,K. Considering
1602 * memory reclaim from current node, there are pros. and cons.
1604 * Freeing memory from current node means freeing memory from a node which
1605 * we'll use or we've used. So, it may make LRU bad. And if several threads
1606 * hit limits, it will see a contention on a node. But freeing from remote
1607 * node means more costs for memory reclaim because of memory latency.
1609 * Now, we use round-robin. Better algorithm is welcomed.
1611 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1615 mem_cgroup_may_update_nodemask(memcg);
1616 node = memcg->last_scanned_node;
1618 node = next_node(node, memcg->scan_nodes);
1619 if (node == MAX_NUMNODES)
1620 node = first_node(memcg->scan_nodes);
1622 * We call this when we hit limit, not when pages are added to LRU.
1623 * No LRU may hold pages because all pages are UNEVICTABLE or
1624 * memcg is too small and all pages are not on LRU. In that case,
1625 * we use curret node.
1627 if (unlikely(node == MAX_NUMNODES))
1628 node = numa_node_id();
1630 memcg->last_scanned_node = node;
1635 * Check all nodes whether it contains reclaimable pages or not.
1636 * For quick scan, we make use of scan_nodes. This will allow us to skip
1637 * unused nodes. But scan_nodes is lazily updated and may not cotain
1638 * enough new information. We need to do double check.
1640 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1645 * quick check...making use of scan_node.
1646 * We can skip unused nodes.
1648 if (!nodes_empty(memcg->scan_nodes)) {
1649 for (nid = first_node(memcg->scan_nodes);
1651 nid = next_node(nid, memcg->scan_nodes)) {
1653 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1658 * Check rest of nodes.
1660 for_each_node_state(nid, N_HIGH_MEMORY) {
1661 if (node_isset(nid, memcg->scan_nodes))
1663 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1670 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1675 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1677 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
1681 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1684 unsigned long *total_scanned)
1686 struct mem_cgroup *victim = NULL;
1689 unsigned long excess;
1690 unsigned long nr_scanned;
1691 struct mem_cgroup_reclaim_cookie reclaim = {
1696 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
1699 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1704 * If we have not been able to reclaim
1705 * anything, it might because there are
1706 * no reclaimable pages under this hierarchy
1711 * We want to do more targeted reclaim.
1712 * excess >> 2 is not to excessive so as to
1713 * reclaim too much, nor too less that we keep
1714 * coming back to reclaim from this cgroup
1716 if (total >= (excess >> 2) ||
1717 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1722 if (!mem_cgroup_reclaimable(victim, false))
1724 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1726 *total_scanned += nr_scanned;
1727 if (!res_counter_soft_limit_excess(&root_memcg->res))
1730 mem_cgroup_iter_break(root_memcg, victim);
1735 * Check OOM-Killer is already running under our hierarchy.
1736 * If someone is running, return false.
1737 * Has to be called with memcg_oom_lock
1739 static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg)
1741 struct mem_cgroup *iter, *failed = NULL;
1743 for_each_mem_cgroup_tree(iter, memcg) {
1744 if (iter->oom_lock) {
1746 * this subtree of our hierarchy is already locked
1747 * so we cannot give a lock.
1750 mem_cgroup_iter_break(memcg, iter);
1753 iter->oom_lock = true;
1760 * OK, we failed to lock the whole subtree so we have to clean up
1761 * what we set up to the failing subtree
1763 for_each_mem_cgroup_tree(iter, memcg) {
1764 if (iter == failed) {
1765 mem_cgroup_iter_break(memcg, iter);
1768 iter->oom_lock = false;
1774 * Has to be called with memcg_oom_lock
1776 static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1778 struct mem_cgroup *iter;
1780 for_each_mem_cgroup_tree(iter, memcg)
1781 iter->oom_lock = false;
1785 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1787 struct mem_cgroup *iter;
1789 for_each_mem_cgroup_tree(iter, memcg)
1790 atomic_inc(&iter->under_oom);
1793 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1795 struct mem_cgroup *iter;
1798 * When a new child is created while the hierarchy is under oom,
1799 * mem_cgroup_oom_lock() may not be called. We have to use
1800 * atomic_add_unless() here.
1802 for_each_mem_cgroup_tree(iter, memcg)
1803 atomic_add_unless(&iter->under_oom, -1, 0);
1806 static DEFINE_SPINLOCK(memcg_oom_lock);
1807 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1809 struct oom_wait_info {
1810 struct mem_cgroup *memcg;
1814 static int memcg_oom_wake_function(wait_queue_t *wait,
1815 unsigned mode, int sync, void *arg)
1817 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1818 struct mem_cgroup *oom_wait_memcg;
1819 struct oom_wait_info *oom_wait_info;
1821 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1822 oom_wait_memcg = oom_wait_info->memcg;
1825 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
1826 * Then we can use css_is_ancestor without taking care of RCU.
1828 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
1829 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
1831 return autoremove_wake_function(wait, mode, sync, arg);
1834 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
1836 /* for filtering, pass "memcg" as argument. */
1837 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1840 static void memcg_oom_recover(struct mem_cgroup *memcg)
1842 if (memcg && atomic_read(&memcg->under_oom))
1843 memcg_wakeup_oom(memcg);
1847 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1849 static bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask,
1852 struct oom_wait_info owait;
1853 bool locked, need_to_kill;
1855 owait.memcg = memcg;
1856 owait.wait.flags = 0;
1857 owait.wait.func = memcg_oom_wake_function;
1858 owait.wait.private = current;
1859 INIT_LIST_HEAD(&owait.wait.task_list);
1860 need_to_kill = true;
1861 mem_cgroup_mark_under_oom(memcg);
1863 /* At first, try to OOM lock hierarchy under memcg.*/
1864 spin_lock(&memcg_oom_lock);
1865 locked = mem_cgroup_oom_lock(memcg);
1867 * Even if signal_pending(), we can't quit charge() loop without
1868 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1869 * under OOM is always welcomed, use TASK_KILLABLE here.
1871 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1872 if (!locked || memcg->oom_kill_disable)
1873 need_to_kill = false;
1875 mem_cgroup_oom_notify(memcg);
1876 spin_unlock(&memcg_oom_lock);
1879 finish_wait(&memcg_oom_waitq, &owait.wait);
1880 mem_cgroup_out_of_memory(memcg, mask, order);
1883 finish_wait(&memcg_oom_waitq, &owait.wait);
1885 spin_lock(&memcg_oom_lock);
1887 mem_cgroup_oom_unlock(memcg);
1888 memcg_wakeup_oom(memcg);
1889 spin_unlock(&memcg_oom_lock);
1891 mem_cgroup_unmark_under_oom(memcg);
1893 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
1895 /* Give chance to dying process */
1896 schedule_timeout_uninterruptible(1);
1901 * Currently used to update mapped file statistics, but the routine can be
1902 * generalized to update other statistics as well.
1904 * Notes: Race condition
1906 * We usually use page_cgroup_lock() for accessing page_cgroup member but
1907 * it tends to be costly. But considering some conditions, we doesn't need
1908 * to do so _always_.
1910 * Considering "charge", lock_page_cgroup() is not required because all
1911 * file-stat operations happen after a page is attached to radix-tree. There
1912 * are no race with "charge".
1914 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
1915 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
1916 * if there are race with "uncharge". Statistics itself is properly handled
1919 * Considering "move", this is an only case we see a race. To make the race
1920 * small, we check mm->moving_account and detect there are possibility of race
1921 * If there is, we take a lock.
1924 void __mem_cgroup_begin_update_page_stat(struct page *page,
1925 bool *locked, unsigned long *flags)
1927 struct mem_cgroup *memcg;
1928 struct page_cgroup *pc;
1930 pc = lookup_page_cgroup(page);
1932 memcg = pc->mem_cgroup;
1933 if (unlikely(!memcg || !PageCgroupUsed(pc)))
1936 * If this memory cgroup is not under account moving, we don't
1937 * need to take move_lock_page_cgroup(). Because we already hold
1938 * rcu_read_lock(), any calls to move_account will be delayed until
1939 * rcu_read_unlock() if mem_cgroup_stolen() == true.
1941 if (!mem_cgroup_stolen(memcg))
1944 move_lock_mem_cgroup(memcg, flags);
1945 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
1946 move_unlock_mem_cgroup(memcg, flags);
1952 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
1954 struct page_cgroup *pc = lookup_page_cgroup(page);
1957 * It's guaranteed that pc->mem_cgroup never changes while
1958 * lock is held because a routine modifies pc->mem_cgroup
1959 * should take move_lock_page_cgroup().
1961 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
1964 void mem_cgroup_update_page_stat(struct page *page,
1965 enum mem_cgroup_page_stat_item idx, int val)
1967 struct mem_cgroup *memcg;
1968 struct page_cgroup *pc = lookup_page_cgroup(page);
1969 unsigned long uninitialized_var(flags);
1971 if (mem_cgroup_disabled())
1974 memcg = pc->mem_cgroup;
1975 if (unlikely(!memcg || !PageCgroupUsed(pc)))
1979 case MEMCG_NR_FILE_MAPPED:
1980 idx = MEM_CGROUP_STAT_FILE_MAPPED;
1986 this_cpu_add(memcg->stat->count[idx], val);
1990 * size of first charge trial. "32" comes from vmscan.c's magic value.
1991 * TODO: maybe necessary to use big numbers in big irons.
1993 #define CHARGE_BATCH 32U
1994 struct memcg_stock_pcp {
1995 struct mem_cgroup *cached; /* this never be root cgroup */
1996 unsigned int nr_pages;
1997 struct work_struct work;
1998 unsigned long flags;
1999 #define FLUSHING_CACHED_CHARGE 0
2001 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2002 static DEFINE_MUTEX(percpu_charge_mutex);
2005 * Try to consume stocked charge on this cpu. If success, one page is consumed
2006 * from local stock and true is returned. If the stock is 0 or charges from a
2007 * cgroup which is not current target, returns false. This stock will be
2010 static bool consume_stock(struct mem_cgroup *memcg)
2012 struct memcg_stock_pcp *stock;
2015 stock = &get_cpu_var(memcg_stock);
2016 if (memcg == stock->cached && stock->nr_pages)
2018 else /* need to call res_counter_charge */
2020 put_cpu_var(memcg_stock);
2025 * Returns stocks cached in percpu to res_counter and reset cached information.
2027 static void drain_stock(struct memcg_stock_pcp *stock)
2029 struct mem_cgroup *old = stock->cached;
2031 if (stock->nr_pages) {
2032 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2034 res_counter_uncharge(&old->res, bytes);
2035 if (do_swap_account)
2036 res_counter_uncharge(&old->memsw, bytes);
2037 stock->nr_pages = 0;
2039 stock->cached = NULL;
2043 * This must be called under preempt disabled or must be called by
2044 * a thread which is pinned to local cpu.
2046 static void drain_local_stock(struct work_struct *dummy)
2048 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2050 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2054 * Cache charges(val) which is from res_counter, to local per_cpu area.
2055 * This will be consumed by consume_stock() function, later.
2057 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2059 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2061 if (stock->cached != memcg) { /* reset if necessary */
2063 stock->cached = memcg;
2065 stock->nr_pages += nr_pages;
2066 put_cpu_var(memcg_stock);
2070 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2071 * of the hierarchy under it. sync flag says whether we should block
2072 * until the work is done.
2074 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2078 /* Notify other cpus that system-wide "drain" is running */
2081 for_each_online_cpu(cpu) {
2082 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2083 struct mem_cgroup *memcg;
2085 memcg = stock->cached;
2086 if (!memcg || !stock->nr_pages)
2088 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2090 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2092 drain_local_stock(&stock->work);
2094 schedule_work_on(cpu, &stock->work);
2102 for_each_online_cpu(cpu) {
2103 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2104 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2105 flush_work(&stock->work);
2112 * Tries to drain stocked charges in other cpus. This function is asynchronous
2113 * and just put a work per cpu for draining localy on each cpu. Caller can
2114 * expects some charges will be back to res_counter later but cannot wait for
2117 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2120 * If someone calls draining, avoid adding more kworker runs.
2122 if (!mutex_trylock(&percpu_charge_mutex))
2124 drain_all_stock(root_memcg, false);
2125 mutex_unlock(&percpu_charge_mutex);
2128 /* This is a synchronous drain interface. */
2129 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2131 /* called when force_empty is called */
2132 mutex_lock(&percpu_charge_mutex);
2133 drain_all_stock(root_memcg, true);
2134 mutex_unlock(&percpu_charge_mutex);
2138 * This function drains percpu counter value from DEAD cpu and
2139 * move it to local cpu. Note that this function can be preempted.
2141 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2145 spin_lock(&memcg->pcp_counter_lock);
2146 for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) {
2147 long x = per_cpu(memcg->stat->count[i], cpu);
2149 per_cpu(memcg->stat->count[i], cpu) = 0;
2150 memcg->nocpu_base.count[i] += x;
2152 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2153 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2155 per_cpu(memcg->stat->events[i], cpu) = 0;
2156 memcg->nocpu_base.events[i] += x;
2158 spin_unlock(&memcg->pcp_counter_lock);
2161 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2162 unsigned long action,
2165 int cpu = (unsigned long)hcpu;
2166 struct memcg_stock_pcp *stock;
2167 struct mem_cgroup *iter;
2169 if (action == CPU_ONLINE)
2172 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2175 for_each_mem_cgroup(iter)
2176 mem_cgroup_drain_pcp_counter(iter, cpu);
2178 stock = &per_cpu(memcg_stock, cpu);
2184 /* See __mem_cgroup_try_charge() for details */
2186 CHARGE_OK, /* success */
2187 CHARGE_RETRY, /* need to retry but retry is not bad */
2188 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2189 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2190 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2193 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2194 unsigned int nr_pages, bool oom_check)
2196 unsigned long csize = nr_pages * PAGE_SIZE;
2197 struct mem_cgroup *mem_over_limit;
2198 struct res_counter *fail_res;
2199 unsigned long flags = 0;
2202 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2205 if (!do_swap_account)
2207 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2211 res_counter_uncharge(&memcg->res, csize);
2212 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2213 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2215 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2217 * nr_pages can be either a huge page (HPAGE_PMD_NR), a batch
2218 * of regular pages (CHARGE_BATCH), or a single regular page (1).
2220 * Never reclaim on behalf of optional batching, retry with a
2221 * single page instead.
2223 if (nr_pages == CHARGE_BATCH)
2224 return CHARGE_RETRY;
2226 if (!(gfp_mask & __GFP_WAIT))
2227 return CHARGE_WOULDBLOCK;
2229 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2230 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2231 return CHARGE_RETRY;
2233 * Even though the limit is exceeded at this point, reclaim
2234 * may have been able to free some pages. Retry the charge
2235 * before killing the task.
2237 * Only for regular pages, though: huge pages are rather
2238 * unlikely to succeed so close to the limit, and we fall back
2239 * to regular pages anyway in case of failure.
2241 if (nr_pages == 1 && ret)
2242 return CHARGE_RETRY;
2245 * At task move, charge accounts can be doubly counted. So, it's
2246 * better to wait until the end of task_move if something is going on.
2248 if (mem_cgroup_wait_acct_move(mem_over_limit))
2249 return CHARGE_RETRY;
2251 /* If we don't need to call oom-killer at el, return immediately */
2253 return CHARGE_NOMEM;
2255 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask, get_order(csize)))
2256 return CHARGE_OOM_DIE;
2258 return CHARGE_RETRY;
2262 * __mem_cgroup_try_charge() does
2263 * 1. detect memcg to be charged against from passed *mm and *ptr,
2264 * 2. update res_counter
2265 * 3. call memory reclaim if necessary.
2267 * In some special case, if the task is fatal, fatal_signal_pending() or
2268 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2269 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2270 * as possible without any hazards. 2: all pages should have a valid
2271 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2272 * pointer, that is treated as a charge to root_mem_cgroup.
2274 * So __mem_cgroup_try_charge() will return
2275 * 0 ... on success, filling *ptr with a valid memcg pointer.
2276 * -ENOMEM ... charge failure because of resource limits.
2277 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2279 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2280 * the oom-killer can be invoked.
2282 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2284 unsigned int nr_pages,
2285 struct mem_cgroup **ptr,
2288 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2289 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2290 struct mem_cgroup *memcg = NULL;
2294 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2295 * in system level. So, allow to go ahead dying process in addition to
2298 if (unlikely(test_thread_flag(TIF_MEMDIE)
2299 || fatal_signal_pending(current)))
2303 * We always charge the cgroup the mm_struct belongs to.
2304 * The mm_struct's mem_cgroup changes on task migration if the
2305 * thread group leader migrates. It's possible that mm is not
2306 * set, if so charge the init_mm (happens for pagecache usage).
2309 *ptr = root_mem_cgroup;
2311 if (*ptr) { /* css should be a valid one */
2313 VM_BUG_ON(css_is_removed(&memcg->css));
2314 if (mem_cgroup_is_root(memcg))
2316 if (nr_pages == 1 && consume_stock(memcg))
2318 css_get(&memcg->css);
2320 struct task_struct *p;
2323 p = rcu_dereference(mm->owner);
2325 * Because we don't have task_lock(), "p" can exit.
2326 * In that case, "memcg" can point to root or p can be NULL with
2327 * race with swapoff. Then, we have small risk of mis-accouning.
2328 * But such kind of mis-account by race always happens because
2329 * we don't have cgroup_mutex(). It's overkill and we allo that
2331 * (*) swapoff at el will charge against mm-struct not against
2332 * task-struct. So, mm->owner can be NULL.
2334 memcg = mem_cgroup_from_task(p);
2336 memcg = root_mem_cgroup;
2337 if (mem_cgroup_is_root(memcg)) {
2341 if (nr_pages == 1 && consume_stock(memcg)) {
2343 * It seems dagerous to access memcg without css_get().
2344 * But considering how consume_stok works, it's not
2345 * necessary. If consume_stock success, some charges
2346 * from this memcg are cached on this cpu. So, we
2347 * don't need to call css_get()/css_tryget() before
2348 * calling consume_stock().
2353 /* after here, we may be blocked. we need to get refcnt */
2354 if (!css_tryget(&memcg->css)) {
2364 /* If killed, bypass charge */
2365 if (fatal_signal_pending(current)) {
2366 css_put(&memcg->css);
2371 if (oom && !nr_oom_retries) {
2373 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2376 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, oom_check);
2380 case CHARGE_RETRY: /* not in OOM situation but retry */
2382 css_put(&memcg->css);
2385 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2386 css_put(&memcg->css);
2388 case CHARGE_NOMEM: /* OOM routine works */
2390 css_put(&memcg->css);
2393 /* If oom, we never return -ENOMEM */
2396 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2397 css_put(&memcg->css);
2400 } while (ret != CHARGE_OK);
2402 if (batch > nr_pages)
2403 refill_stock(memcg, batch - nr_pages);
2404 css_put(&memcg->css);
2412 *ptr = root_mem_cgroup;
2417 * Somemtimes we have to undo a charge we got by try_charge().
2418 * This function is for that and do uncharge, put css's refcnt.
2419 * gotten by try_charge().
2421 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2422 unsigned int nr_pages)
2424 if (!mem_cgroup_is_root(memcg)) {
2425 unsigned long bytes = nr_pages * PAGE_SIZE;
2427 res_counter_uncharge(&memcg->res, bytes);
2428 if (do_swap_account)
2429 res_counter_uncharge(&memcg->memsw, bytes);
2434 * A helper function to get mem_cgroup from ID. must be called under
2435 * rcu_read_lock(). The caller must check css_is_removed() or some if
2436 * it's concern. (dropping refcnt from swap can be called against removed
2439 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2441 struct cgroup_subsys_state *css;
2443 /* ID 0 is unused ID */
2446 css = css_lookup(&mem_cgroup_subsys, id);
2449 return container_of(css, struct mem_cgroup, css);
2452 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2454 struct mem_cgroup *memcg = NULL;
2455 struct page_cgroup *pc;
2459 VM_BUG_ON(!PageLocked(page));
2461 pc = lookup_page_cgroup(page);
2462 lock_page_cgroup(pc);
2463 if (PageCgroupUsed(pc)) {
2464 memcg = pc->mem_cgroup;
2465 if (memcg && !css_tryget(&memcg->css))
2467 } else if (PageSwapCache(page)) {
2468 ent.val = page_private(page);
2469 id = lookup_swap_cgroup_id(ent);
2471 memcg = mem_cgroup_lookup(id);
2472 if (memcg && !css_tryget(&memcg->css))
2476 unlock_page_cgroup(pc);
2480 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2482 unsigned int nr_pages,
2483 enum charge_type ctype,
2486 struct page_cgroup *pc = lookup_page_cgroup(page);
2487 struct zone *uninitialized_var(zone);
2488 bool was_on_lru = false;
2491 lock_page_cgroup(pc);
2492 if (unlikely(PageCgroupUsed(pc))) {
2493 unlock_page_cgroup(pc);
2494 __mem_cgroup_cancel_charge(memcg, nr_pages);
2498 * we don't need page_cgroup_lock about tail pages, becase they are not
2499 * accessed by any other context at this point.
2503 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2504 * may already be on some other mem_cgroup's LRU. Take care of it.
2507 zone = page_zone(page);
2508 spin_lock_irq(&zone->lru_lock);
2509 if (PageLRU(page)) {
2511 del_page_from_lru_list(zone, page, page_lru(page));
2516 pc->mem_cgroup = memcg;
2518 * We access a page_cgroup asynchronously without lock_page_cgroup().
2519 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2520 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2521 * before USED bit, we need memory barrier here.
2522 * See mem_cgroup_add_lru_list(), etc.
2525 SetPageCgroupUsed(pc);
2529 VM_BUG_ON(PageLRU(page));
2531 add_page_to_lru_list(zone, page, page_lru(page));
2533 spin_unlock_irq(&zone->lru_lock);
2536 if (ctype == MEM_CGROUP_CHARGE_TYPE_MAPPED)
2541 mem_cgroup_charge_statistics(memcg, anon, nr_pages);
2542 unlock_page_cgroup(pc);
2545 * "charge_statistics" updated event counter. Then, check it.
2546 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2547 * if they exceeds softlimit.
2549 memcg_check_events(memcg, page);
2552 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2554 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
2556 * Because tail pages are not marked as "used", set it. We're under
2557 * zone->lru_lock, 'splitting on pmd' and compound_lock.
2558 * charge/uncharge will be never happen and move_account() is done under
2559 * compound_lock(), so we don't have to take care of races.
2561 void mem_cgroup_split_huge_fixup(struct page *head)
2563 struct page_cgroup *head_pc = lookup_page_cgroup(head);
2564 struct page_cgroup *pc;
2567 if (mem_cgroup_disabled())
2569 for (i = 1; i < HPAGE_PMD_NR; i++) {
2571 pc->mem_cgroup = head_pc->mem_cgroup;
2572 smp_wmb();/* see __commit_charge() */
2573 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2576 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2579 * mem_cgroup_move_account - move account of the page
2581 * @nr_pages: number of regular pages (>1 for huge pages)
2582 * @pc: page_cgroup of the page.
2583 * @from: mem_cgroup which the page is moved from.
2584 * @to: mem_cgroup which the page is moved to. @from != @to.
2585 * @uncharge: whether we should call uncharge and css_put against @from.
2587 * The caller must confirm following.
2588 * - page is not on LRU (isolate_page() is useful.)
2589 * - compound_lock is held when nr_pages > 1
2591 * This function doesn't do "charge" nor css_get to new cgroup. It should be
2592 * done by a caller(__mem_cgroup_try_charge would be useful). If @uncharge is
2593 * true, this function does "uncharge" from old cgroup, but it doesn't if
2594 * @uncharge is false, so a caller should do "uncharge".
2596 static int mem_cgroup_move_account(struct page *page,
2597 unsigned int nr_pages,
2598 struct page_cgroup *pc,
2599 struct mem_cgroup *from,
2600 struct mem_cgroup *to,
2603 unsigned long flags;
2605 bool anon = PageAnon(page);
2607 VM_BUG_ON(from == to);
2608 VM_BUG_ON(PageLRU(page));
2610 * The page is isolated from LRU. So, collapse function
2611 * will not handle this page. But page splitting can happen.
2612 * Do this check under compound_page_lock(). The caller should
2616 if (nr_pages > 1 && !PageTransHuge(page))
2619 lock_page_cgroup(pc);
2622 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
2625 move_lock_mem_cgroup(from, &flags);
2627 if (!anon && page_mapped(page)) {
2628 /* Update mapped_file data for mem_cgroup */
2630 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2631 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2634 mem_cgroup_charge_statistics(from, anon, -nr_pages);
2636 /* This is not "cancel", but cancel_charge does all we need. */
2637 __mem_cgroup_cancel_charge(from, nr_pages);
2639 /* caller should have done css_get */
2640 pc->mem_cgroup = to;
2641 mem_cgroup_charge_statistics(to, anon, nr_pages);
2643 * We charges against "to" which may not have any tasks. Then, "to"
2644 * can be under rmdir(). But in current implementation, caller of
2645 * this function is just force_empty() and move charge, so it's
2646 * guaranteed that "to" is never removed. So, we don't check rmdir
2649 move_unlock_mem_cgroup(from, &flags);
2652 unlock_page_cgroup(pc);
2656 memcg_check_events(to, page);
2657 memcg_check_events(from, page);
2663 * move charges to its parent.
2666 static int mem_cgroup_move_parent(struct page *page,
2667 struct page_cgroup *pc,
2668 struct mem_cgroup *child,
2671 struct cgroup *cg = child->css.cgroup;
2672 struct cgroup *pcg = cg->parent;
2673 struct mem_cgroup *parent;
2674 unsigned int nr_pages;
2675 unsigned long uninitialized_var(flags);
2683 if (!get_page_unless_zero(page))
2685 if (isolate_lru_page(page))
2688 nr_pages = hpage_nr_pages(page);
2690 parent = mem_cgroup_from_cont(pcg);
2691 ret = __mem_cgroup_try_charge(NULL, gfp_mask, nr_pages, &parent, false);
2696 flags = compound_lock_irqsave(page);
2698 ret = mem_cgroup_move_account(page, nr_pages, pc, child, parent, true);
2700 __mem_cgroup_cancel_charge(parent, nr_pages);
2703 compound_unlock_irqrestore(page, flags);
2705 putback_lru_page(page);
2713 * Charge the memory controller for page usage.
2715 * 0 if the charge was successful
2716 * < 0 if the cgroup is over its limit
2718 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2719 gfp_t gfp_mask, enum charge_type ctype)
2721 struct mem_cgroup *memcg = NULL;
2722 unsigned int nr_pages = 1;
2726 if (PageTransHuge(page)) {
2727 nr_pages <<= compound_order(page);
2728 VM_BUG_ON(!PageTransHuge(page));
2730 * Never OOM-kill a process for a huge page. The
2731 * fault handler will fall back to regular pages.
2736 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
2739 __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false);
2743 int mem_cgroup_newpage_charge(struct page *page,
2744 struct mm_struct *mm, gfp_t gfp_mask)
2746 if (mem_cgroup_disabled())
2748 VM_BUG_ON(page_mapped(page));
2749 VM_BUG_ON(page->mapping && !PageAnon(page));
2751 return mem_cgroup_charge_common(page, mm, gfp_mask,
2752 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2756 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2757 enum charge_type ctype);
2759 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2762 struct mem_cgroup *memcg = NULL;
2763 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
2766 if (mem_cgroup_disabled())
2768 if (PageCompound(page))
2773 if (!page_is_file_cache(page))
2774 type = MEM_CGROUP_CHARGE_TYPE_SHMEM;
2776 if (!PageSwapCache(page))
2777 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
2778 else { /* page is swapcache/shmem */
2779 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &memcg);
2781 __mem_cgroup_commit_charge_swapin(page, memcg, type);
2787 * While swap-in, try_charge -> commit or cancel, the page is locked.
2788 * And when try_charge() successfully returns, one refcnt to memcg without
2789 * struct page_cgroup is acquired. This refcnt will be consumed by
2790 * "commit()" or removed by "cancel()"
2792 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2794 gfp_t mask, struct mem_cgroup **memcgp)
2796 struct mem_cgroup *memcg;
2801 if (mem_cgroup_disabled())
2804 if (!do_swap_account)
2807 * A racing thread's fault, or swapoff, may have already updated
2808 * the pte, and even removed page from swap cache: in those cases
2809 * do_swap_page()'s pte_same() test will fail; but there's also a
2810 * KSM case which does need to charge the page.
2812 if (!PageSwapCache(page))
2814 memcg = try_get_mem_cgroup_from_page(page);
2818 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
2819 css_put(&memcg->css);
2826 ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
2833 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
2834 enum charge_type ctype)
2836 if (mem_cgroup_disabled())
2840 cgroup_exclude_rmdir(&memcg->css);
2842 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
2844 * Now swap is on-memory. This means this page may be
2845 * counted both as mem and swap....double count.
2846 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2847 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2848 * may call delete_from_swap_cache() before reach here.
2850 if (do_swap_account && PageSwapCache(page)) {
2851 swp_entry_t ent = {.val = page_private(page)};
2852 mem_cgroup_uncharge_swap(ent);
2855 * At swapin, we may charge account against cgroup which has no tasks.
2856 * So, rmdir()->pre_destroy() can be called while we do this charge.
2857 * In that case, we need to call pre_destroy() again. check it here.
2859 cgroup_release_and_wakeup_rmdir(&memcg->css);
2862 void mem_cgroup_commit_charge_swapin(struct page *page,
2863 struct mem_cgroup *memcg)
2865 __mem_cgroup_commit_charge_swapin(page, memcg,
2866 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2869 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
2871 if (mem_cgroup_disabled())
2875 __mem_cgroup_cancel_charge(memcg, 1);
2878 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
2879 unsigned int nr_pages,
2880 const enum charge_type ctype)
2882 struct memcg_batch_info *batch = NULL;
2883 bool uncharge_memsw = true;
2885 /* If swapout, usage of swap doesn't decrease */
2886 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2887 uncharge_memsw = false;
2889 batch = ¤t->memcg_batch;
2891 * In usual, we do css_get() when we remember memcg pointer.
2892 * But in this case, we keep res->usage until end of a series of
2893 * uncharges. Then, it's ok to ignore memcg's refcnt.
2896 batch->memcg = memcg;
2898 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2899 * In those cases, all pages freed continuously can be expected to be in
2900 * the same cgroup and we have chance to coalesce uncharges.
2901 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2902 * because we want to do uncharge as soon as possible.
2905 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
2906 goto direct_uncharge;
2909 goto direct_uncharge;
2912 * In typical case, batch->memcg == mem. This means we can
2913 * merge a series of uncharges to an uncharge of res_counter.
2914 * If not, we uncharge res_counter ony by one.
2916 if (batch->memcg != memcg)
2917 goto direct_uncharge;
2918 /* remember freed charge and uncharge it later */
2921 batch->memsw_nr_pages++;
2924 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
2926 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
2927 if (unlikely(batch->memcg != memcg))
2928 memcg_oom_recover(memcg);
2932 * uncharge if !page_mapped(page)
2934 static struct mem_cgroup *
2935 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
2937 struct mem_cgroup *memcg = NULL;
2938 unsigned int nr_pages = 1;
2939 struct page_cgroup *pc;
2942 if (mem_cgroup_disabled())
2945 if (PageSwapCache(page))
2948 if (PageTransHuge(page)) {
2949 nr_pages <<= compound_order(page);
2950 VM_BUG_ON(!PageTransHuge(page));
2953 * Check if our page_cgroup is valid
2955 pc = lookup_page_cgroup(page);
2956 if (unlikely(!PageCgroupUsed(pc)))
2959 lock_page_cgroup(pc);
2961 memcg = pc->mem_cgroup;
2963 if (!PageCgroupUsed(pc))
2966 anon = PageAnon(page);
2969 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2971 * Generally PageAnon tells if it's the anon statistics to be
2972 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
2973 * used before page reached the stage of being marked PageAnon.
2977 case MEM_CGROUP_CHARGE_TYPE_DROP:
2978 /* See mem_cgroup_prepare_migration() */
2979 if (page_mapped(page) || PageCgroupMigration(pc))
2982 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
2983 if (!PageAnon(page)) { /* Shared memory */
2984 if (page->mapping && !page_is_file_cache(page))
2986 } else if (page_mapped(page)) /* Anon */
2993 mem_cgroup_charge_statistics(memcg, anon, -nr_pages);
2995 ClearPageCgroupUsed(pc);
2997 * pc->mem_cgroup is not cleared here. It will be accessed when it's
2998 * freed from LRU. This is safe because uncharged page is expected not
2999 * to be reused (freed soon). Exception is SwapCache, it's handled by
3000 * special functions.
3003 unlock_page_cgroup(pc);
3005 * even after unlock, we have memcg->res.usage here and this memcg
3006 * will never be freed.
3008 memcg_check_events(memcg, page);
3009 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
3010 mem_cgroup_swap_statistics(memcg, true);
3011 mem_cgroup_get(memcg);
3013 if (!mem_cgroup_is_root(memcg))
3014 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
3019 unlock_page_cgroup(pc);
3023 void mem_cgroup_uncharge_page(struct page *page)
3026 if (page_mapped(page))
3028 VM_BUG_ON(page->mapping && !PageAnon(page));
3029 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
3032 void mem_cgroup_uncharge_cache_page(struct page *page)
3034 VM_BUG_ON(page_mapped(page));
3035 VM_BUG_ON(page->mapping);
3036 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
3040 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
3041 * In that cases, pages are freed continuously and we can expect pages
3042 * are in the same memcg. All these calls itself limits the number of
3043 * pages freed at once, then uncharge_start/end() is called properly.
3044 * This may be called prural(2) times in a context,
3047 void mem_cgroup_uncharge_start(void)
3049 current->memcg_batch.do_batch++;
3050 /* We can do nest. */
3051 if (current->memcg_batch.do_batch == 1) {
3052 current->memcg_batch.memcg = NULL;
3053 current->memcg_batch.nr_pages = 0;
3054 current->memcg_batch.memsw_nr_pages = 0;
3058 void mem_cgroup_uncharge_end(void)
3060 struct memcg_batch_info *batch = ¤t->memcg_batch;
3062 if (!batch->do_batch)
3066 if (batch->do_batch) /* If stacked, do nothing. */
3072 * This "batch->memcg" is valid without any css_get/put etc...
3073 * bacause we hide charges behind us.
3075 if (batch->nr_pages)
3076 res_counter_uncharge(&batch->memcg->res,
3077 batch->nr_pages * PAGE_SIZE);
3078 if (batch->memsw_nr_pages)
3079 res_counter_uncharge(&batch->memcg->memsw,
3080 batch->memsw_nr_pages * PAGE_SIZE);
3081 memcg_oom_recover(batch->memcg);
3082 /* forget this pointer (for sanity check) */
3083 batch->memcg = NULL;
3088 * called after __delete_from_swap_cache() and drop "page" account.
3089 * memcg information is recorded to swap_cgroup of "ent"
3092 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
3094 struct mem_cgroup *memcg;
3095 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
3097 if (!swapout) /* this was a swap cache but the swap is unused ! */
3098 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
3100 memcg = __mem_cgroup_uncharge_common(page, ctype);
3103 * record memcg information, if swapout && memcg != NULL,
3104 * mem_cgroup_get() was called in uncharge().
3106 if (do_swap_account && swapout && memcg)
3107 swap_cgroup_record(ent, css_id(&memcg->css));
3111 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3113 * called from swap_entry_free(). remove record in swap_cgroup and
3114 * uncharge "memsw" account.
3116 void mem_cgroup_uncharge_swap(swp_entry_t ent)
3118 struct mem_cgroup *memcg;
3121 if (!do_swap_account)
3124 id = swap_cgroup_record(ent, 0);
3126 memcg = mem_cgroup_lookup(id);
3129 * We uncharge this because swap is freed.
3130 * This memcg can be obsolete one. We avoid calling css_tryget
3132 if (!mem_cgroup_is_root(memcg))
3133 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
3134 mem_cgroup_swap_statistics(memcg, false);
3135 mem_cgroup_put(memcg);
3141 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3142 * @entry: swap entry to be moved
3143 * @from: mem_cgroup which the entry is moved from
3144 * @to: mem_cgroup which the entry is moved to
3146 * It succeeds only when the swap_cgroup's record for this entry is the same
3147 * as the mem_cgroup's id of @from.
3149 * Returns 0 on success, -EINVAL on failure.
3151 * The caller must have charged to @to, IOW, called res_counter_charge() about
3152 * both res and memsw, and called css_get().
3154 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3155 struct mem_cgroup *from, struct mem_cgroup *to)
3157 unsigned short old_id, new_id;
3159 old_id = css_id(&from->css);
3160 new_id = css_id(&to->css);
3162 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3163 mem_cgroup_swap_statistics(from, false);
3164 mem_cgroup_swap_statistics(to, true);
3166 * This function is only called from task migration context now.
3167 * It postpones res_counter and refcount handling till the end
3168 * of task migration(mem_cgroup_clear_mc()) for performance
3169 * improvement. But we cannot postpone mem_cgroup_get(to)
3170 * because if the process that has been moved to @to does
3171 * swap-in, the refcount of @to might be decreased to 0.
3179 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3180 struct mem_cgroup *from, struct mem_cgroup *to)
3187 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
3190 int mem_cgroup_prepare_migration(struct page *page,
3191 struct page *newpage, struct mem_cgroup **memcgp, gfp_t gfp_mask)
3193 struct mem_cgroup *memcg = NULL;
3194 struct page_cgroup *pc;
3195 enum charge_type ctype;
3200 VM_BUG_ON(PageTransHuge(page));
3201 if (mem_cgroup_disabled())
3204 pc = lookup_page_cgroup(page);
3205 lock_page_cgroup(pc);
3206 if (PageCgroupUsed(pc)) {
3207 memcg = pc->mem_cgroup;
3208 css_get(&memcg->css);
3210 * At migrating an anonymous page, its mapcount goes down
3211 * to 0 and uncharge() will be called. But, even if it's fully
3212 * unmapped, migration may fail and this page has to be
3213 * charged again. We set MIGRATION flag here and delay uncharge
3214 * until end_migration() is called
3216 * Corner Case Thinking
3218 * When the old page was mapped as Anon and it's unmap-and-freed
3219 * while migration was ongoing.
3220 * If unmap finds the old page, uncharge() of it will be delayed
3221 * until end_migration(). If unmap finds a new page, it's
3222 * uncharged when it make mapcount to be 1->0. If unmap code
3223 * finds swap_migration_entry, the new page will not be mapped
3224 * and end_migration() will find it(mapcount==0).
3227 * When the old page was mapped but migraion fails, the kernel
3228 * remaps it. A charge for it is kept by MIGRATION flag even
3229 * if mapcount goes down to 0. We can do remap successfully
3230 * without charging it again.
3233 * The "old" page is under lock_page() until the end of
3234 * migration, so, the old page itself will not be swapped-out.
3235 * If the new page is swapped out before end_migraton, our
3236 * hook to usual swap-out path will catch the event.
3239 SetPageCgroupMigration(pc);
3241 unlock_page_cgroup(pc);
3243 * If the page is not charged at this point,
3250 ret = __mem_cgroup_try_charge(NULL, gfp_mask, 1, memcgp, false);
3251 css_put(&memcg->css);/* drop extra refcnt */
3253 if (PageAnon(page)) {
3254 lock_page_cgroup(pc);
3255 ClearPageCgroupMigration(pc);
3256 unlock_page_cgroup(pc);
3258 * The old page may be fully unmapped while we kept it.
3260 mem_cgroup_uncharge_page(page);
3262 /* we'll need to revisit this error code (we have -EINTR) */
3266 * We charge new page before it's used/mapped. So, even if unlock_page()
3267 * is called before end_migration, we can catch all events on this new
3268 * page. In the case new page is migrated but not remapped, new page's
3269 * mapcount will be finally 0 and we call uncharge in end_migration().
3272 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
3273 else if (page_is_file_cache(page))
3274 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
3276 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3277 __mem_cgroup_commit_charge(memcg, newpage, 1, ctype, false);
3281 /* remove redundant charge if migration failed*/
3282 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
3283 struct page *oldpage, struct page *newpage, bool migration_ok)
3285 struct page *used, *unused;
3286 struct page_cgroup *pc;
3291 /* blocks rmdir() */
3292 cgroup_exclude_rmdir(&memcg->css);
3293 if (!migration_ok) {
3301 * We disallowed uncharge of pages under migration because mapcount
3302 * of the page goes down to zero, temporarly.
3303 * Clear the flag and check the page should be charged.
3305 pc = lookup_page_cgroup(oldpage);
3306 lock_page_cgroup(pc);
3307 ClearPageCgroupMigration(pc);
3308 unlock_page_cgroup(pc);
3309 anon = PageAnon(used);
3310 __mem_cgroup_uncharge_common(unused,
3311 anon ? MEM_CGROUP_CHARGE_TYPE_MAPPED
3312 : MEM_CGROUP_CHARGE_TYPE_CACHE);
3315 * If a page is a file cache, radix-tree replacement is very atomic
3316 * and we can skip this check. When it was an Anon page, its mapcount
3317 * goes down to 0. But because we added MIGRATION flage, it's not
3318 * uncharged yet. There are several case but page->mapcount check
3319 * and USED bit check in mem_cgroup_uncharge_page() will do enough
3320 * check. (see prepare_charge() also)
3323 mem_cgroup_uncharge_page(used);
3325 * At migration, we may charge account against cgroup which has no
3327 * So, rmdir()->pre_destroy() can be called while we do this charge.
3328 * In that case, we need to call pre_destroy() again. check it here.
3330 cgroup_release_and_wakeup_rmdir(&memcg->css);
3334 * At replace page cache, newpage is not under any memcg but it's on
3335 * LRU. So, this function doesn't touch res_counter but handles LRU
3336 * in correct way. Both pages are locked so we cannot race with uncharge.
3338 void mem_cgroup_replace_page_cache(struct page *oldpage,
3339 struct page *newpage)
3341 struct mem_cgroup *memcg = NULL;
3342 struct page_cgroup *pc;
3343 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
3345 if (mem_cgroup_disabled())
3348 pc = lookup_page_cgroup(oldpage);
3349 /* fix accounting on old pages */
3350 lock_page_cgroup(pc);
3351 if (PageCgroupUsed(pc)) {
3352 memcg = pc->mem_cgroup;
3353 mem_cgroup_charge_statistics(memcg, false, -1);
3354 ClearPageCgroupUsed(pc);
3356 unlock_page_cgroup(pc);
3359 * When called from shmem_replace_page(), in some cases the
3360 * oldpage has already been charged, and in some cases not.
3365 if (PageSwapBacked(oldpage))
3366 type = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3369 * Even if newpage->mapping was NULL before starting replacement,
3370 * the newpage may be on LRU(or pagevec for LRU) already. We lock
3371 * LRU while we overwrite pc->mem_cgroup.
3373 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
3376 #ifdef CONFIG_DEBUG_VM
3377 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3379 struct page_cgroup *pc;
3381 pc = lookup_page_cgroup(page);
3383 * Can be NULL while feeding pages into the page allocator for
3384 * the first time, i.e. during boot or memory hotplug;
3385 * or when mem_cgroup_disabled().
3387 if (likely(pc) && PageCgroupUsed(pc))
3392 bool mem_cgroup_bad_page_check(struct page *page)
3394 if (mem_cgroup_disabled())
3397 return lookup_page_cgroup_used(page) != NULL;
3400 void mem_cgroup_print_bad_page(struct page *page)
3402 struct page_cgroup *pc;
3404 pc = lookup_page_cgroup_used(page);
3406 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
3407 pc, pc->flags, pc->mem_cgroup);
3412 static DEFINE_MUTEX(set_limit_mutex);
3414 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3415 unsigned long long val)
3418 u64 memswlimit, memlimit;
3420 int children = mem_cgroup_count_children(memcg);
3421 u64 curusage, oldusage;
3425 * For keeping hierarchical_reclaim simple, how long we should retry
3426 * is depends on callers. We set our retry-count to be function
3427 * of # of children which we should visit in this loop.
3429 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3431 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3434 while (retry_count) {
3435 if (signal_pending(current)) {
3440 * Rather than hide all in some function, I do this in
3441 * open coded manner. You see what this really does.
3442 * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3444 mutex_lock(&set_limit_mutex);
3445 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3446 if (memswlimit < val) {
3448 mutex_unlock(&set_limit_mutex);
3452 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3456 ret = res_counter_set_limit(&memcg->res, val);
3458 if (memswlimit == val)
3459 memcg->memsw_is_minimum = true;
3461 memcg->memsw_is_minimum = false;
3463 mutex_unlock(&set_limit_mutex);
3468 mem_cgroup_reclaim(memcg, GFP_KERNEL,
3469 MEM_CGROUP_RECLAIM_SHRINK);
3470 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3471 /* Usage is reduced ? */
3472 if (curusage >= oldusage)
3475 oldusage = curusage;
3477 if (!ret && enlarge)
3478 memcg_oom_recover(memcg);
3483 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3484 unsigned long long val)
3487 u64 memlimit, memswlimit, oldusage, curusage;
3488 int children = mem_cgroup_count_children(memcg);
3492 /* see mem_cgroup_resize_res_limit */
3493 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3494 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3495 while (retry_count) {
3496 if (signal_pending(current)) {
3501 * Rather than hide all in some function, I do this in
3502 * open coded manner. You see what this really does.
3503 * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3505 mutex_lock(&set_limit_mutex);
3506 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3507 if (memlimit > val) {
3509 mutex_unlock(&set_limit_mutex);
3512 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3513 if (memswlimit < val)
3515 ret = res_counter_set_limit(&memcg->memsw, val);
3517 if (memlimit == val)
3518 memcg->memsw_is_minimum = true;
3520 memcg->memsw_is_minimum = false;
3522 mutex_unlock(&set_limit_mutex);
3527 mem_cgroup_reclaim(memcg, GFP_KERNEL,
3528 MEM_CGROUP_RECLAIM_NOSWAP |
3529 MEM_CGROUP_RECLAIM_SHRINK);
3530 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3531 /* Usage is reduced ? */
3532 if (curusage >= oldusage)
3535 oldusage = curusage;
3537 if (!ret && enlarge)
3538 memcg_oom_recover(memcg);
3542 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3544 unsigned long *total_scanned)
3546 unsigned long nr_reclaimed = 0;
3547 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3548 unsigned long reclaimed;
3550 struct mem_cgroup_tree_per_zone *mctz;
3551 unsigned long long excess;
3552 unsigned long nr_scanned;
3557 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3559 * This loop can run a while, specially if mem_cgroup's continuously
3560 * keep exceeding their soft limit and putting the system under
3567 mz = mem_cgroup_largest_soft_limit_node(mctz);
3572 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
3573 gfp_mask, &nr_scanned);
3574 nr_reclaimed += reclaimed;
3575 *total_scanned += nr_scanned;
3576 spin_lock(&mctz->lock);
3579 * If we failed to reclaim anything from this memory cgroup
3580 * it is time to move on to the next cgroup
3586 * Loop until we find yet another one.
3588 * By the time we get the soft_limit lock
3589 * again, someone might have aded the
3590 * group back on the RB tree. Iterate to
3591 * make sure we get a different mem.
3592 * mem_cgroup_largest_soft_limit_node returns
3593 * NULL if no other cgroup is present on
3597 __mem_cgroup_largest_soft_limit_node(mctz);
3599 css_put(&next_mz->memcg->css);
3600 else /* next_mz == NULL or other memcg */
3604 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
3605 excess = res_counter_soft_limit_excess(&mz->memcg->res);
3607 * One school of thought says that we should not add
3608 * back the node to the tree if reclaim returns 0.
3609 * But our reclaim could return 0, simply because due
3610 * to priority we are exposing a smaller subset of
3611 * memory to reclaim from. Consider this as a longer
3614 /* If excess == 0, no tree ops */
3615 __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess);
3616 spin_unlock(&mctz->lock);
3617 css_put(&mz->memcg->css);
3620 * Could not reclaim anything and there are no more
3621 * mem cgroups to try or we seem to be looping without
3622 * reclaiming anything.
3624 if (!nr_reclaimed &&
3626 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3628 } while (!nr_reclaimed);
3630 css_put(&next_mz->memcg->css);
3631 return nr_reclaimed;
3635 * This routine traverse page_cgroup in given list and drop them all.
3636 * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3638 static int mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
3639 int node, int zid, enum lru_list lru)
3641 struct mem_cgroup_per_zone *mz;
3642 unsigned long flags, loop;
3643 struct list_head *list;
3648 zone = &NODE_DATA(node)->node_zones[zid];
3649 mz = mem_cgroup_zoneinfo(memcg, node, zid);
3650 list = &mz->lruvec.lists[lru];
3652 loop = mz->lru_size[lru];
3653 /* give some margin against EBUSY etc...*/
3657 struct page_cgroup *pc;
3661 spin_lock_irqsave(&zone->lru_lock, flags);
3662 if (list_empty(list)) {
3663 spin_unlock_irqrestore(&zone->lru_lock, flags);
3666 page = list_entry(list->prev, struct page, lru);
3668 list_move(&page->lru, list);
3670 spin_unlock_irqrestore(&zone->lru_lock, flags);
3673 spin_unlock_irqrestore(&zone->lru_lock, flags);
3675 pc = lookup_page_cgroup(page);
3677 ret = mem_cgroup_move_parent(page, pc, memcg, GFP_KERNEL);
3678 if (ret == -ENOMEM || ret == -EINTR)
3681 if (ret == -EBUSY || ret == -EINVAL) {
3682 /* found lock contention or "pc" is obsolete. */
3689 if (!ret && !list_empty(list))
3695 * make mem_cgroup's charge to be 0 if there is no task.
3696 * This enables deleting this mem_cgroup.
3698 static int mem_cgroup_force_empty(struct mem_cgroup *memcg, bool free_all)
3701 int node, zid, shrink;
3702 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3703 struct cgroup *cgrp = memcg->css.cgroup;
3705 css_get(&memcg->css);
3708 /* should free all ? */
3714 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3717 if (signal_pending(current))
3719 /* This is for making all *used* pages to be on LRU. */
3720 lru_add_drain_all();
3721 drain_all_stock_sync(memcg);
3723 mem_cgroup_start_move(memcg);
3724 for_each_node_state(node, N_HIGH_MEMORY) {
3725 for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3728 ret = mem_cgroup_force_empty_list(memcg,
3737 mem_cgroup_end_move(memcg);
3738 memcg_oom_recover(memcg);
3739 /* it seems parent cgroup doesn't have enough mem */
3743 /* "ret" should also be checked to ensure all lists are empty. */
3744 } while (res_counter_read_u64(&memcg->res, RES_USAGE) > 0 || ret);
3746 css_put(&memcg->css);
3750 /* returns EBUSY if there is a task or if we come here twice. */
3751 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3755 /* we call try-to-free pages for make this cgroup empty */
3756 lru_add_drain_all();
3757 /* try to free all pages in this cgroup */
3759 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
3762 if (signal_pending(current)) {
3766 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
3770 /* maybe some writeback is necessary */
3771 congestion_wait(BLK_RW_ASYNC, HZ/10);
3776 /* try move_account...there may be some *locked* pages. */
3780 static int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3782 return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3786 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3788 return mem_cgroup_from_cont(cont)->use_hierarchy;
3791 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3795 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3796 struct cgroup *parent = cont->parent;
3797 struct mem_cgroup *parent_memcg = NULL;
3800 parent_memcg = mem_cgroup_from_cont(parent);
3804 * If parent's use_hierarchy is set, we can't make any modifications
3805 * in the child subtrees. If it is unset, then the change can
3806 * occur, provided the current cgroup has no children.
3808 * For the root cgroup, parent_mem is NULL, we allow value to be
3809 * set if there are no children.
3811 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3812 (val == 1 || val == 0)) {
3813 if (list_empty(&cont->children))
3814 memcg->use_hierarchy = val;
3825 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
3826 enum mem_cgroup_stat_index idx)
3828 struct mem_cgroup *iter;
3831 /* Per-cpu values can be negative, use a signed accumulator */
3832 for_each_mem_cgroup_tree(iter, memcg)
3833 val += mem_cgroup_read_stat(iter, idx);
3835 if (val < 0) /* race ? */
3840 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3844 if (!mem_cgroup_is_root(memcg)) {
3846 return res_counter_read_u64(&memcg->res, RES_USAGE);
3848 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
3851 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
3852 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
3855 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
3857 return val << PAGE_SHIFT;
3860 static ssize_t mem_cgroup_read(struct cgroup *cont, struct cftype *cft,
3861 struct file *file, char __user *buf,
3862 size_t nbytes, loff_t *ppos)
3864 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3867 int type, name, len;
3869 type = MEMFILE_TYPE(cft->private);
3870 name = MEMFILE_ATTR(cft->private);
3872 if (!do_swap_account && type == _MEMSWAP)
3877 if (name == RES_USAGE)
3878 val = mem_cgroup_usage(memcg, false);
3880 val = res_counter_read_u64(&memcg->res, name);
3883 if (name == RES_USAGE)
3884 val = mem_cgroup_usage(memcg, true);
3886 val = res_counter_read_u64(&memcg->memsw, name);
3892 len = scnprintf(str, sizeof(str), "%llu\n", (unsigned long long)val);
3893 return simple_read_from_buffer(buf, nbytes, ppos, str, len);
3896 * The user of this function is...
3899 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
3902 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3904 unsigned long long val;
3907 type = MEMFILE_TYPE(cft->private);
3908 name = MEMFILE_ATTR(cft->private);
3910 if (!do_swap_account && type == _MEMSWAP)
3915 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3919 /* This function does all necessary parse...reuse it */
3920 ret = res_counter_memparse_write_strategy(buffer, &val);
3924 ret = mem_cgroup_resize_limit(memcg, val);
3926 ret = mem_cgroup_resize_memsw_limit(memcg, val);
3928 case RES_SOFT_LIMIT:
3929 ret = res_counter_memparse_write_strategy(buffer, &val);
3933 * For memsw, soft limits are hard to implement in terms
3934 * of semantics, for now, we support soft limits for
3935 * control without swap
3938 ret = res_counter_set_soft_limit(&memcg->res, val);
3943 ret = -EINVAL; /* should be BUG() ? */
3949 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
3950 unsigned long long *mem_limit, unsigned long long *memsw_limit)
3952 struct cgroup *cgroup;
3953 unsigned long long min_limit, min_memsw_limit, tmp;
3955 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3956 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3957 cgroup = memcg->css.cgroup;
3958 if (!memcg->use_hierarchy)
3961 while (cgroup->parent) {
3962 cgroup = cgroup->parent;
3963 memcg = mem_cgroup_from_cont(cgroup);
3964 if (!memcg->use_hierarchy)
3966 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
3967 min_limit = min(min_limit, tmp);
3968 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3969 min_memsw_limit = min(min_memsw_limit, tmp);
3972 *mem_limit = min_limit;
3973 *memsw_limit = min_memsw_limit;
3976 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
3978 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3981 type = MEMFILE_TYPE(event);
3982 name = MEMFILE_ATTR(event);
3984 if (!do_swap_account && type == _MEMSWAP)
3990 res_counter_reset_max(&memcg->res);
3992 res_counter_reset_max(&memcg->memsw);
3996 res_counter_reset_failcnt(&memcg->res);
3998 res_counter_reset_failcnt(&memcg->memsw);
4005 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
4008 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
4012 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4013 struct cftype *cft, u64 val)
4015 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4017 if (val >= (1 << NR_MOVE_TYPE))
4020 * We check this value several times in both in can_attach() and
4021 * attach(), so we need cgroup lock to prevent this value from being
4025 memcg->move_charge_at_immigrate = val;
4031 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4032 struct cftype *cft, u64 val)
4039 /* For read statistics */
4057 struct mcs_total_stat {
4058 s64 stat[NR_MCS_STAT];
4064 } memcg_stat_strings[NR_MCS_STAT] = {
4065 {"cache", "total_cache"},
4066 {"rss", "total_rss"},
4067 {"mapped_file", "total_mapped_file"},
4068 {"pgpgin", "total_pgpgin"},
4069 {"pgpgout", "total_pgpgout"},
4070 {"swap", "total_swap"},
4071 {"pgfault", "total_pgfault"},
4072 {"pgmajfault", "total_pgmajfault"},
4073 {"inactive_anon", "total_inactive_anon"},
4074 {"active_anon", "total_active_anon"},
4075 {"inactive_file", "total_inactive_file"},
4076 {"active_file", "total_active_file"},
4077 {"unevictable", "total_unevictable"}
4082 mem_cgroup_get_local_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4087 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_CACHE);
4088 s->stat[MCS_CACHE] += val * PAGE_SIZE;
4089 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_RSS);
4090 s->stat[MCS_RSS] += val * PAGE_SIZE;
4091 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_FILE_MAPPED);
4092 s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
4093 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGIN);
4094 s->stat[MCS_PGPGIN] += val;
4095 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGOUT);
4096 s->stat[MCS_PGPGOUT] += val;
4097 if (do_swap_account) {
4098 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
4099 s->stat[MCS_SWAP] += val * PAGE_SIZE;
4101 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGFAULT);
4102 s->stat[MCS_PGFAULT] += val;
4103 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGMAJFAULT);
4104 s->stat[MCS_PGMAJFAULT] += val;
4107 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_ANON));
4108 s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
4109 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_ANON));
4110 s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
4111 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_FILE));
4112 s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
4113 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_FILE));
4114 s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
4115 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
4116 s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
4120 mem_cgroup_get_total_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4122 struct mem_cgroup *iter;
4124 for_each_mem_cgroup_tree(iter, memcg)
4125 mem_cgroup_get_local_stat(iter, s);
4129 static int mem_control_numa_stat_show(struct seq_file *m, void *arg)
4132 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
4133 unsigned long node_nr;
4134 struct cgroup *cont = m->private;
4135 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4137 total_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL);
4138 seq_printf(m, "total=%lu", total_nr);
4139 for_each_node_state(nid, N_HIGH_MEMORY) {
4140 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL);
4141 seq_printf(m, " N%d=%lu", nid, node_nr);
4145 file_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_FILE);
4146 seq_printf(m, "file=%lu", file_nr);
4147 for_each_node_state(nid, N_HIGH_MEMORY) {
4148 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4150 seq_printf(m, " N%d=%lu", nid, node_nr);
4154 anon_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_ANON);
4155 seq_printf(m, "anon=%lu", anon_nr);
4156 for_each_node_state(nid, N_HIGH_MEMORY) {
4157 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4159 seq_printf(m, " N%d=%lu", nid, node_nr);
4163 unevictable_nr = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
4164 seq_printf(m, "unevictable=%lu", unevictable_nr);
4165 for_each_node_state(nid, N_HIGH_MEMORY) {
4166 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4167 BIT(LRU_UNEVICTABLE));
4168 seq_printf(m, " N%d=%lu", nid, node_nr);
4173 #endif /* CONFIG_NUMA */
4175 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
4176 struct cgroup_map_cb *cb)
4178 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4179 struct mcs_total_stat mystat;
4182 memset(&mystat, 0, sizeof(mystat));
4183 mem_cgroup_get_local_stat(memcg, &mystat);
4186 for (i = 0; i < NR_MCS_STAT; i++) {
4187 if (i == MCS_SWAP && !do_swap_account)
4189 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
4192 /* Hierarchical information */
4194 unsigned long long limit, memsw_limit;
4195 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
4196 cb->fill(cb, "hierarchical_memory_limit", limit);
4197 if (do_swap_account)
4198 cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
4201 memset(&mystat, 0, sizeof(mystat));
4202 mem_cgroup_get_total_stat(memcg, &mystat);
4203 for (i = 0; i < NR_MCS_STAT; i++) {
4204 if (i == MCS_SWAP && !do_swap_account)
4206 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
4209 #ifdef CONFIG_DEBUG_VM
4212 struct mem_cgroup_per_zone *mz;
4213 struct zone_reclaim_stat *rstat;
4214 unsigned long recent_rotated[2] = {0, 0};
4215 unsigned long recent_scanned[2] = {0, 0};
4217 for_each_online_node(nid)
4218 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4219 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
4220 rstat = &mz->lruvec.reclaim_stat;
4222 recent_rotated[0] += rstat->recent_rotated[0];
4223 recent_rotated[1] += rstat->recent_rotated[1];
4224 recent_scanned[0] += rstat->recent_scanned[0];
4225 recent_scanned[1] += rstat->recent_scanned[1];
4227 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
4228 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
4229 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
4230 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
4237 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
4239 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4241 return mem_cgroup_swappiness(memcg);
4244 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
4247 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4248 struct mem_cgroup *parent;
4253 if (cgrp->parent == NULL)
4256 parent = mem_cgroup_from_cont(cgrp->parent);
4260 /* If under hierarchy, only empty-root can set this value */
4261 if ((parent->use_hierarchy) ||
4262 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4267 memcg->swappiness = val;
4274 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4276 struct mem_cgroup_threshold_ary *t;
4282 t = rcu_dereference(memcg->thresholds.primary);
4284 t = rcu_dereference(memcg->memsw_thresholds.primary);
4289 usage = mem_cgroup_usage(memcg, swap);
4292 * current_threshold points to threshold just below or equal to usage.
4293 * If it's not true, a threshold was crossed after last
4294 * call of __mem_cgroup_threshold().
4296 i = t->current_threshold;
4299 * Iterate backward over array of thresholds starting from
4300 * current_threshold and check if a threshold is crossed.
4301 * If none of thresholds below usage is crossed, we read
4302 * only one element of the array here.
4304 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4305 eventfd_signal(t->entries[i].eventfd, 1);
4307 /* i = current_threshold + 1 */
4311 * Iterate forward over array of thresholds starting from
4312 * current_threshold+1 and check if a threshold is crossed.
4313 * If none of thresholds above usage is crossed, we read
4314 * only one element of the array here.
4316 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4317 eventfd_signal(t->entries[i].eventfd, 1);
4319 /* Update current_threshold */
4320 t->current_threshold = i - 1;
4325 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4328 __mem_cgroup_threshold(memcg, false);
4329 if (do_swap_account)
4330 __mem_cgroup_threshold(memcg, true);
4332 memcg = parent_mem_cgroup(memcg);
4336 static int compare_thresholds(const void *a, const void *b)
4338 const struct mem_cgroup_threshold *_a = a;
4339 const struct mem_cgroup_threshold *_b = b;
4341 return _a->threshold - _b->threshold;
4344 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4346 struct mem_cgroup_eventfd_list *ev;
4348 list_for_each_entry(ev, &memcg->oom_notify, list)
4349 eventfd_signal(ev->eventfd, 1);
4353 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4355 struct mem_cgroup *iter;
4357 for_each_mem_cgroup_tree(iter, memcg)
4358 mem_cgroup_oom_notify_cb(iter);
4361 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
4362 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4364 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4365 struct mem_cgroup_thresholds *thresholds;
4366 struct mem_cgroup_threshold_ary *new;
4367 int type = MEMFILE_TYPE(cft->private);
4368 u64 threshold, usage;
4371 ret = res_counter_memparse_write_strategy(args, &threshold);
4375 mutex_lock(&memcg->thresholds_lock);
4378 thresholds = &memcg->thresholds;
4379 else if (type == _MEMSWAP)
4380 thresholds = &memcg->memsw_thresholds;
4384 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4386 /* Check if a threshold crossed before adding a new one */
4387 if (thresholds->primary)
4388 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4390 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4392 /* Allocate memory for new array of thresholds */
4393 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4401 /* Copy thresholds (if any) to new array */
4402 if (thresholds->primary) {
4403 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4404 sizeof(struct mem_cgroup_threshold));
4407 /* Add new threshold */
4408 new->entries[size - 1].eventfd = eventfd;
4409 new->entries[size - 1].threshold = threshold;
4411 /* Sort thresholds. Registering of new threshold isn't time-critical */
4412 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4413 compare_thresholds, NULL);
4415 /* Find current threshold */
4416 new->current_threshold = -1;
4417 for (i = 0; i < size; i++) {
4418 if (new->entries[i].threshold <= usage) {
4420 * new->current_threshold will not be used until
4421 * rcu_assign_pointer(), so it's safe to increment
4424 ++new->current_threshold;
4429 /* Free old spare buffer and save old primary buffer as spare */
4430 kfree(thresholds->spare);
4431 thresholds->spare = thresholds->primary;
4433 rcu_assign_pointer(thresholds->primary, new);
4435 /* To be sure that nobody uses thresholds */
4439 mutex_unlock(&memcg->thresholds_lock);
4444 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
4445 struct cftype *cft, struct eventfd_ctx *eventfd)
4447 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4448 struct mem_cgroup_thresholds *thresholds;
4449 struct mem_cgroup_threshold_ary *new;
4450 int type = MEMFILE_TYPE(cft->private);
4454 mutex_lock(&memcg->thresholds_lock);
4456 thresholds = &memcg->thresholds;
4457 else if (type == _MEMSWAP)
4458 thresholds = &memcg->memsw_thresholds;
4462 if (!thresholds->primary)
4465 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4467 /* Check if a threshold crossed before removing */
4468 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4470 /* Calculate new number of threshold */
4472 for (i = 0; i < thresholds->primary->size; i++) {
4473 if (thresholds->primary->entries[i].eventfd != eventfd)
4477 new = thresholds->spare;
4479 /* Set thresholds array to NULL if we don't have thresholds */
4488 /* Copy thresholds and find current threshold */
4489 new->current_threshold = -1;
4490 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4491 if (thresholds->primary->entries[i].eventfd == eventfd)
4494 new->entries[j] = thresholds->primary->entries[i];
4495 if (new->entries[j].threshold <= usage) {
4497 * new->current_threshold will not be used
4498 * until rcu_assign_pointer(), so it's safe to increment
4501 ++new->current_threshold;
4507 /* Swap primary and spare array */
4508 thresholds->spare = thresholds->primary;
4509 /* If all events are unregistered, free the spare array */
4511 kfree(thresholds->spare);
4512 thresholds->spare = NULL;
4515 rcu_assign_pointer(thresholds->primary, new);
4517 /* To be sure that nobody uses thresholds */
4520 mutex_unlock(&memcg->thresholds_lock);
4523 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4524 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4526 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4527 struct mem_cgroup_eventfd_list *event;
4528 int type = MEMFILE_TYPE(cft->private);
4530 BUG_ON(type != _OOM_TYPE);
4531 event = kmalloc(sizeof(*event), GFP_KERNEL);
4535 spin_lock(&memcg_oom_lock);
4537 event->eventfd = eventfd;
4538 list_add(&event->list, &memcg->oom_notify);
4540 /* already in OOM ? */
4541 if (atomic_read(&memcg->under_oom))
4542 eventfd_signal(eventfd, 1);
4543 spin_unlock(&memcg_oom_lock);
4548 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4549 struct cftype *cft, struct eventfd_ctx *eventfd)
4551 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4552 struct mem_cgroup_eventfd_list *ev, *tmp;
4553 int type = MEMFILE_TYPE(cft->private);
4555 BUG_ON(type != _OOM_TYPE);
4557 spin_lock(&memcg_oom_lock);
4559 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4560 if (ev->eventfd == eventfd) {
4561 list_del(&ev->list);
4566 spin_unlock(&memcg_oom_lock);
4569 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4570 struct cftype *cft, struct cgroup_map_cb *cb)
4572 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4574 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
4576 if (atomic_read(&memcg->under_oom))
4577 cb->fill(cb, "under_oom", 1);
4579 cb->fill(cb, "under_oom", 0);
4583 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4584 struct cftype *cft, u64 val)
4586 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4587 struct mem_cgroup *parent;
4589 /* cannot set to root cgroup and only 0 and 1 are allowed */
4590 if (!cgrp->parent || !((val == 0) || (val == 1)))
4593 parent = mem_cgroup_from_cont(cgrp->parent);
4596 /* oom-kill-disable is a flag for subhierarchy. */
4597 if ((parent->use_hierarchy) ||
4598 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4602 memcg->oom_kill_disable = val;
4604 memcg_oom_recover(memcg);
4610 static const struct file_operations mem_control_numa_stat_file_operations = {
4612 .llseek = seq_lseek,
4613 .release = single_release,
4616 static int mem_control_numa_stat_open(struct inode *unused, struct file *file)
4618 struct cgroup *cont = file->f_dentry->d_parent->d_fsdata;
4620 file->f_op = &mem_control_numa_stat_file_operations;
4621 return single_open(file, mem_control_numa_stat_show, cont);
4623 #endif /* CONFIG_NUMA */
4625 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_KMEM
4626 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
4628 return mem_cgroup_sockets_init(memcg, ss);
4631 static void kmem_cgroup_destroy(struct mem_cgroup *memcg)
4633 mem_cgroup_sockets_destroy(memcg);
4636 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
4641 static void kmem_cgroup_destroy(struct mem_cgroup *memcg)
4646 static struct cftype mem_cgroup_files[] = {
4648 .name = "usage_in_bytes",
4649 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4650 .read = mem_cgroup_read,
4651 .register_event = mem_cgroup_usage_register_event,
4652 .unregister_event = mem_cgroup_usage_unregister_event,
4655 .name = "max_usage_in_bytes",
4656 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4657 .trigger = mem_cgroup_reset,
4658 .read = mem_cgroup_read,
4661 .name = "limit_in_bytes",
4662 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4663 .write_string = mem_cgroup_write,
4664 .read = mem_cgroup_read,
4667 .name = "soft_limit_in_bytes",
4668 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4669 .write_string = mem_cgroup_write,
4670 .read = mem_cgroup_read,
4674 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4675 .trigger = mem_cgroup_reset,
4676 .read = mem_cgroup_read,
4680 .read_map = mem_control_stat_show,
4683 .name = "force_empty",
4684 .trigger = mem_cgroup_force_empty_write,
4687 .name = "use_hierarchy",
4688 .write_u64 = mem_cgroup_hierarchy_write,
4689 .read_u64 = mem_cgroup_hierarchy_read,
4692 .name = "swappiness",
4693 .read_u64 = mem_cgroup_swappiness_read,
4694 .write_u64 = mem_cgroup_swappiness_write,
4697 .name = "move_charge_at_immigrate",
4698 .read_u64 = mem_cgroup_move_charge_read,
4699 .write_u64 = mem_cgroup_move_charge_write,
4702 .name = "oom_control",
4703 .read_map = mem_cgroup_oom_control_read,
4704 .write_u64 = mem_cgroup_oom_control_write,
4705 .register_event = mem_cgroup_oom_register_event,
4706 .unregister_event = mem_cgroup_oom_unregister_event,
4707 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4711 .name = "numa_stat",
4712 .open = mem_control_numa_stat_open,
4716 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4718 .name = "memsw.usage_in_bytes",
4719 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4720 .read = mem_cgroup_read,
4721 .register_event = mem_cgroup_usage_register_event,
4722 .unregister_event = mem_cgroup_usage_unregister_event,
4725 .name = "memsw.max_usage_in_bytes",
4726 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4727 .trigger = mem_cgroup_reset,
4728 .read = mem_cgroup_read,
4731 .name = "memsw.limit_in_bytes",
4732 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4733 .write_string = mem_cgroup_write,
4734 .read = mem_cgroup_read,
4737 .name = "memsw.failcnt",
4738 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4739 .trigger = mem_cgroup_reset,
4740 .read = mem_cgroup_read,
4743 { }, /* terminate */
4746 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4748 struct mem_cgroup_per_node *pn;
4749 struct mem_cgroup_per_zone *mz;
4750 int zone, tmp = node;
4752 * This routine is called against possible nodes.
4753 * But it's BUG to call kmalloc() against offline node.
4755 * TODO: this routine can waste much memory for nodes which will
4756 * never be onlined. It's better to use memory hotplug callback
4759 if (!node_state(node, N_NORMAL_MEMORY))
4761 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4765 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4766 mz = &pn->zoneinfo[zone];
4767 lruvec_init(&mz->lruvec, &NODE_DATA(node)->node_zones[zone]);
4768 mz->usage_in_excess = 0;
4769 mz->on_tree = false;
4772 memcg->info.nodeinfo[node] = pn;
4776 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4778 kfree(memcg->info.nodeinfo[node]);
4781 static struct mem_cgroup *mem_cgroup_alloc(void)
4783 struct mem_cgroup *memcg;
4784 int size = sizeof(struct mem_cgroup);
4786 /* Can be very big if MAX_NUMNODES is very big */
4787 if (size < PAGE_SIZE)
4788 memcg = kzalloc(size, GFP_KERNEL);
4790 memcg = vzalloc(size);
4795 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4798 spin_lock_init(&memcg->pcp_counter_lock);
4802 if (size < PAGE_SIZE)
4810 * Helpers for freeing a vzalloc()ed mem_cgroup by RCU,
4811 * but in process context. The work_freeing structure is overlaid
4812 * on the rcu_freeing structure, which itself is overlaid on memsw.
4814 static void vfree_work(struct work_struct *work)
4816 struct mem_cgroup *memcg;
4818 memcg = container_of(work, struct mem_cgroup, work_freeing);
4821 static void vfree_rcu(struct rcu_head *rcu_head)
4823 struct mem_cgroup *memcg;
4825 memcg = container_of(rcu_head, struct mem_cgroup, rcu_freeing);
4826 INIT_WORK(&memcg->work_freeing, vfree_work);
4827 schedule_work(&memcg->work_freeing);
4831 * At destroying mem_cgroup, references from swap_cgroup can remain.
4832 * (scanning all at force_empty is too costly...)
4834 * Instead of clearing all references at force_empty, we remember
4835 * the number of reference from swap_cgroup and free mem_cgroup when
4836 * it goes down to 0.
4838 * Removal of cgroup itself succeeds regardless of refs from swap.
4841 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4845 mem_cgroup_remove_from_trees(memcg);
4846 free_css_id(&mem_cgroup_subsys, &memcg->css);
4849 free_mem_cgroup_per_zone_info(memcg, node);
4851 free_percpu(memcg->stat);
4852 if (sizeof(struct mem_cgroup) < PAGE_SIZE)
4853 kfree_rcu(memcg, rcu_freeing);
4855 call_rcu(&memcg->rcu_freeing, vfree_rcu);
4858 static void mem_cgroup_get(struct mem_cgroup *memcg)
4860 atomic_inc(&memcg->refcnt);
4863 static void __mem_cgroup_put(struct mem_cgroup *memcg, int count)
4865 if (atomic_sub_and_test(count, &memcg->refcnt)) {
4866 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4867 __mem_cgroup_free(memcg);
4869 mem_cgroup_put(parent);
4873 static void mem_cgroup_put(struct mem_cgroup *memcg)
4875 __mem_cgroup_put(memcg, 1);
4879 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4881 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4883 if (!memcg->res.parent)
4885 return mem_cgroup_from_res_counter(memcg->res.parent, res);
4887 EXPORT_SYMBOL(parent_mem_cgroup);
4889 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4890 static void __init enable_swap_cgroup(void)
4892 if (!mem_cgroup_disabled() && really_do_swap_account)
4893 do_swap_account = 1;
4896 static void __init enable_swap_cgroup(void)
4901 static int mem_cgroup_soft_limit_tree_init(void)
4903 struct mem_cgroup_tree_per_node *rtpn;
4904 struct mem_cgroup_tree_per_zone *rtpz;
4905 int tmp, node, zone;
4907 for_each_node(node) {
4909 if (!node_state(node, N_NORMAL_MEMORY))
4911 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
4915 soft_limit_tree.rb_tree_per_node[node] = rtpn;
4917 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4918 rtpz = &rtpn->rb_tree_per_zone[zone];
4919 rtpz->rb_root = RB_ROOT;
4920 spin_lock_init(&rtpz->lock);
4926 for_each_node(node) {
4927 if (!soft_limit_tree.rb_tree_per_node[node])
4929 kfree(soft_limit_tree.rb_tree_per_node[node]);
4930 soft_limit_tree.rb_tree_per_node[node] = NULL;
4936 static struct cgroup_subsys_state * __ref
4937 mem_cgroup_create(struct cgroup *cont)
4939 struct mem_cgroup *memcg, *parent;
4940 long error = -ENOMEM;
4943 memcg = mem_cgroup_alloc();
4945 return ERR_PTR(error);
4948 if (alloc_mem_cgroup_per_zone_info(memcg, node))
4952 if (cont->parent == NULL) {
4954 enable_swap_cgroup();
4956 if (mem_cgroup_soft_limit_tree_init())
4958 root_mem_cgroup = memcg;
4959 for_each_possible_cpu(cpu) {
4960 struct memcg_stock_pcp *stock =
4961 &per_cpu(memcg_stock, cpu);
4962 INIT_WORK(&stock->work, drain_local_stock);
4964 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
4966 parent = mem_cgroup_from_cont(cont->parent);
4967 memcg->use_hierarchy = parent->use_hierarchy;
4968 memcg->oom_kill_disable = parent->oom_kill_disable;
4971 if (parent && parent->use_hierarchy) {
4972 res_counter_init(&memcg->res, &parent->res);
4973 res_counter_init(&memcg->memsw, &parent->memsw);
4975 * We increment refcnt of the parent to ensure that we can
4976 * safely access it on res_counter_charge/uncharge.
4977 * This refcnt will be decremented when freeing this
4978 * mem_cgroup(see mem_cgroup_put).
4980 mem_cgroup_get(parent);
4982 res_counter_init(&memcg->res, NULL);
4983 res_counter_init(&memcg->memsw, NULL);
4985 memcg->last_scanned_node = MAX_NUMNODES;
4986 INIT_LIST_HEAD(&memcg->oom_notify);
4989 memcg->swappiness = mem_cgroup_swappiness(parent);
4990 atomic_set(&memcg->refcnt, 1);
4991 memcg->move_charge_at_immigrate = 0;
4992 mutex_init(&memcg->thresholds_lock);
4993 spin_lock_init(&memcg->move_lock);
4995 error = memcg_init_kmem(memcg, &mem_cgroup_subsys);
4998 * We call put now because our (and parent's) refcnts
4999 * are already in place. mem_cgroup_put() will internally
5000 * call __mem_cgroup_free, so return directly
5002 mem_cgroup_put(memcg);
5003 return ERR_PTR(error);
5007 __mem_cgroup_free(memcg);
5008 return ERR_PTR(error);
5011 static int mem_cgroup_pre_destroy(struct cgroup *cont)
5013 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5015 return mem_cgroup_force_empty(memcg, false);
5018 static void mem_cgroup_destroy(struct cgroup *cont)
5020 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5022 kmem_cgroup_destroy(memcg);
5024 mem_cgroup_put(memcg);
5028 /* Handlers for move charge at task migration. */
5029 #define PRECHARGE_COUNT_AT_ONCE 256
5030 static int mem_cgroup_do_precharge(unsigned long count)
5033 int batch_count = PRECHARGE_COUNT_AT_ONCE;
5034 struct mem_cgroup *memcg = mc.to;
5036 if (mem_cgroup_is_root(memcg)) {
5037 mc.precharge += count;
5038 /* we don't need css_get for root */
5041 /* try to charge at once */
5043 struct res_counter *dummy;
5045 * "memcg" cannot be under rmdir() because we've already checked
5046 * by cgroup_lock_live_cgroup() that it is not removed and we
5047 * are still under the same cgroup_mutex. So we can postpone
5050 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
5052 if (do_swap_account && res_counter_charge(&memcg->memsw,
5053 PAGE_SIZE * count, &dummy)) {
5054 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
5057 mc.precharge += count;
5061 /* fall back to one by one charge */
5063 if (signal_pending(current)) {
5067 if (!batch_count--) {
5068 batch_count = PRECHARGE_COUNT_AT_ONCE;
5071 ret = __mem_cgroup_try_charge(NULL,
5072 GFP_KERNEL, 1, &memcg, false);
5074 /* mem_cgroup_clear_mc() will do uncharge later */
5082 * get_mctgt_type - get target type of moving charge
5083 * @vma: the vma the pte to be checked belongs
5084 * @addr: the address corresponding to the pte to be checked
5085 * @ptent: the pte to be checked
5086 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5089 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5090 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5091 * move charge. if @target is not NULL, the page is stored in target->page
5092 * with extra refcnt got(Callers should handle it).
5093 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5094 * target for charge migration. if @target is not NULL, the entry is stored
5097 * Called with pte lock held.
5104 enum mc_target_type {
5110 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5111 unsigned long addr, pte_t ptent)
5113 struct page *page = vm_normal_page(vma, addr, ptent);
5115 if (!page || !page_mapped(page))
5117 if (PageAnon(page)) {
5118 /* we don't move shared anon */
5121 } else if (!move_file())
5122 /* we ignore mapcount for file pages */
5124 if (!get_page_unless_zero(page))
5131 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5132 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5134 struct page *page = NULL;
5135 swp_entry_t ent = pte_to_swp_entry(ptent);
5137 if (!move_anon() || non_swap_entry(ent))
5140 * Because lookup_swap_cache() updates some statistics counter,
5141 * we call find_get_page() with swapper_space directly.
5143 page = find_get_page(&swapper_space, ent.val);
5144 if (do_swap_account)
5145 entry->val = ent.val;
5150 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5151 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5157 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5158 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5160 struct page *page = NULL;
5161 struct address_space *mapping;
5164 if (!vma->vm_file) /* anonymous vma */
5169 mapping = vma->vm_file->f_mapping;
5170 if (pte_none(ptent))
5171 pgoff = linear_page_index(vma, addr);
5172 else /* pte_file(ptent) is true */
5173 pgoff = pte_to_pgoff(ptent);
5175 /* page is moved even if it's not RSS of this task(page-faulted). */
5176 page = find_get_page(mapping, pgoff);
5179 /* shmem/tmpfs may report page out on swap: account for that too. */
5180 if (radix_tree_exceptional_entry(page)) {
5181 swp_entry_t swap = radix_to_swp_entry(page);
5182 if (do_swap_account)
5184 page = find_get_page(&swapper_space, swap.val);
5190 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5191 unsigned long addr, pte_t ptent, union mc_target *target)
5193 struct page *page = NULL;
5194 struct page_cgroup *pc;
5195 enum mc_target_type ret = MC_TARGET_NONE;
5196 swp_entry_t ent = { .val = 0 };
5198 if (pte_present(ptent))
5199 page = mc_handle_present_pte(vma, addr, ptent);
5200 else if (is_swap_pte(ptent))
5201 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5202 else if (pte_none(ptent) || pte_file(ptent))
5203 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5205 if (!page && !ent.val)
5208 pc = lookup_page_cgroup(page);
5210 * Do only loose check w/o page_cgroup lock.
5211 * mem_cgroup_move_account() checks the pc is valid or not under
5214 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5215 ret = MC_TARGET_PAGE;
5217 target->page = page;
5219 if (!ret || !target)
5222 /* There is a swap entry and a page doesn't exist or isn't charged */
5223 if (ent.val && !ret &&
5224 css_id(&mc.from->css) == lookup_swap_cgroup_id(ent)) {
5225 ret = MC_TARGET_SWAP;
5232 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5234 * We don't consider swapping or file mapped pages because THP does not
5235 * support them for now.
5236 * Caller should make sure that pmd_trans_huge(pmd) is true.
5238 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5239 unsigned long addr, pmd_t pmd, union mc_target *target)
5241 struct page *page = NULL;
5242 struct page_cgroup *pc;
5243 enum mc_target_type ret = MC_TARGET_NONE;
5245 page = pmd_page(pmd);
5246 VM_BUG_ON(!page || !PageHead(page));
5249 pc = lookup_page_cgroup(page);
5250 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5251 ret = MC_TARGET_PAGE;
5254 target->page = page;
5260 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5261 unsigned long addr, pmd_t pmd, union mc_target *target)
5263 return MC_TARGET_NONE;
5267 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5268 unsigned long addr, unsigned long end,
5269 struct mm_walk *walk)
5271 struct vm_area_struct *vma = walk->private;
5275 if (pmd_trans_huge_lock(pmd, vma) == 1) {
5276 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5277 mc.precharge += HPAGE_PMD_NR;
5278 spin_unlock(&vma->vm_mm->page_table_lock);
5282 if (pmd_trans_unstable(pmd))
5284 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5285 for (; addr != end; pte++, addr += PAGE_SIZE)
5286 if (get_mctgt_type(vma, addr, *pte, NULL))
5287 mc.precharge++; /* increment precharge temporarily */
5288 pte_unmap_unlock(pte - 1, ptl);
5294 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5296 unsigned long precharge;
5297 struct vm_area_struct *vma;
5299 down_read(&mm->mmap_sem);
5300 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5301 struct mm_walk mem_cgroup_count_precharge_walk = {
5302 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5306 if (is_vm_hugetlb_page(vma))
5308 walk_page_range(vma->vm_start, vma->vm_end,
5309 &mem_cgroup_count_precharge_walk);
5311 up_read(&mm->mmap_sem);
5313 precharge = mc.precharge;
5319 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5321 unsigned long precharge = mem_cgroup_count_precharge(mm);
5323 VM_BUG_ON(mc.moving_task);
5324 mc.moving_task = current;
5325 return mem_cgroup_do_precharge(precharge);
5328 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5329 static void __mem_cgroup_clear_mc(void)
5331 struct mem_cgroup *from = mc.from;
5332 struct mem_cgroup *to = mc.to;
5334 /* we must uncharge all the leftover precharges from mc.to */
5336 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
5340 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5341 * we must uncharge here.
5343 if (mc.moved_charge) {
5344 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
5345 mc.moved_charge = 0;
5347 /* we must fixup refcnts and charges */
5348 if (mc.moved_swap) {
5349 /* uncharge swap account from the old cgroup */
5350 if (!mem_cgroup_is_root(mc.from))
5351 res_counter_uncharge(&mc.from->memsw,
5352 PAGE_SIZE * mc.moved_swap);
5353 __mem_cgroup_put(mc.from, mc.moved_swap);
5355 if (!mem_cgroup_is_root(mc.to)) {
5357 * we charged both to->res and to->memsw, so we should
5360 res_counter_uncharge(&mc.to->res,
5361 PAGE_SIZE * mc.moved_swap);
5363 /* we've already done mem_cgroup_get(mc.to) */
5366 memcg_oom_recover(from);
5367 memcg_oom_recover(to);
5368 wake_up_all(&mc.waitq);
5371 static void mem_cgroup_clear_mc(void)
5373 struct mem_cgroup *from = mc.from;
5376 * we must clear moving_task before waking up waiters at the end of
5379 mc.moving_task = NULL;
5380 __mem_cgroup_clear_mc();
5381 spin_lock(&mc.lock);
5384 spin_unlock(&mc.lock);
5385 mem_cgroup_end_move(from);
5388 static int mem_cgroup_can_attach(struct cgroup *cgroup,
5389 struct cgroup_taskset *tset)
5391 struct task_struct *p = cgroup_taskset_first(tset);
5393 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup);
5395 if (memcg->move_charge_at_immigrate) {
5396 struct mm_struct *mm;
5397 struct mem_cgroup *from = mem_cgroup_from_task(p);
5399 VM_BUG_ON(from == memcg);
5401 mm = get_task_mm(p);
5404 /* We move charges only when we move a owner of the mm */
5405 if (mm->owner == p) {
5408 VM_BUG_ON(mc.precharge);
5409 VM_BUG_ON(mc.moved_charge);
5410 VM_BUG_ON(mc.moved_swap);
5411 mem_cgroup_start_move(from);
5412 spin_lock(&mc.lock);
5415 spin_unlock(&mc.lock);
5416 /* We set mc.moving_task later */
5418 ret = mem_cgroup_precharge_mc(mm);
5420 mem_cgroup_clear_mc();
5427 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
5428 struct cgroup_taskset *tset)
5430 mem_cgroup_clear_mc();
5433 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5434 unsigned long addr, unsigned long end,
5435 struct mm_walk *walk)
5438 struct vm_area_struct *vma = walk->private;
5441 enum mc_target_type target_type;
5442 union mc_target target;
5444 struct page_cgroup *pc;
5447 * We don't take compound_lock() here but no race with splitting thp
5449 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
5450 * under splitting, which means there's no concurrent thp split,
5451 * - if another thread runs into split_huge_page() just after we
5452 * entered this if-block, the thread must wait for page table lock
5453 * to be unlocked in __split_huge_page_splitting(), where the main
5454 * part of thp split is not executed yet.
5456 if (pmd_trans_huge_lock(pmd, vma) == 1) {
5457 if (mc.precharge < HPAGE_PMD_NR) {
5458 spin_unlock(&vma->vm_mm->page_table_lock);
5461 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5462 if (target_type == MC_TARGET_PAGE) {
5464 if (!isolate_lru_page(page)) {
5465 pc = lookup_page_cgroup(page);
5466 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
5469 mc.precharge -= HPAGE_PMD_NR;
5470 mc.moved_charge += HPAGE_PMD_NR;
5472 putback_lru_page(page);
5476 spin_unlock(&vma->vm_mm->page_table_lock);
5480 if (pmd_trans_unstable(pmd))
5483 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5484 for (; addr != end; addr += PAGE_SIZE) {
5485 pte_t ptent = *(pte++);
5491 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5492 case MC_TARGET_PAGE:
5494 if (isolate_lru_page(page))
5496 pc = lookup_page_cgroup(page);
5497 if (!mem_cgroup_move_account(page, 1, pc,
5498 mc.from, mc.to, false)) {
5500 /* we uncharge from mc.from later. */
5503 putback_lru_page(page);
5504 put: /* get_mctgt_type() gets the page */
5507 case MC_TARGET_SWAP:
5509 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5511 /* we fixup refcnts and charges later. */
5519 pte_unmap_unlock(pte - 1, ptl);
5524 * We have consumed all precharges we got in can_attach().
5525 * We try charge one by one, but don't do any additional
5526 * charges to mc.to if we have failed in charge once in attach()
5529 ret = mem_cgroup_do_precharge(1);
5537 static void mem_cgroup_move_charge(struct mm_struct *mm)
5539 struct vm_area_struct *vma;
5541 lru_add_drain_all();
5543 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5545 * Someone who are holding the mmap_sem might be waiting in
5546 * waitq. So we cancel all extra charges, wake up all waiters,
5547 * and retry. Because we cancel precharges, we might not be able
5548 * to move enough charges, but moving charge is a best-effort
5549 * feature anyway, so it wouldn't be a big problem.
5551 __mem_cgroup_clear_mc();
5555 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5557 struct mm_walk mem_cgroup_move_charge_walk = {
5558 .pmd_entry = mem_cgroup_move_charge_pte_range,
5562 if (is_vm_hugetlb_page(vma))
5564 ret = walk_page_range(vma->vm_start, vma->vm_end,
5565 &mem_cgroup_move_charge_walk);
5568 * means we have consumed all precharges and failed in
5569 * doing additional charge. Just abandon here.
5573 up_read(&mm->mmap_sem);
5576 static void mem_cgroup_move_task(struct cgroup *cont,
5577 struct cgroup_taskset *tset)
5579 struct task_struct *p = cgroup_taskset_first(tset);
5580 struct mm_struct *mm = get_task_mm(p);
5584 mem_cgroup_move_charge(mm);
5588 mem_cgroup_clear_mc();
5590 #else /* !CONFIG_MMU */
5591 static int mem_cgroup_can_attach(struct cgroup *cgroup,
5592 struct cgroup_taskset *tset)
5596 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
5597 struct cgroup_taskset *tset)
5600 static void mem_cgroup_move_task(struct cgroup *cont,
5601 struct cgroup_taskset *tset)
5606 struct cgroup_subsys mem_cgroup_subsys = {
5608 .subsys_id = mem_cgroup_subsys_id,
5609 .create = mem_cgroup_create,
5610 .pre_destroy = mem_cgroup_pre_destroy,
5611 .destroy = mem_cgroup_destroy,
5612 .can_attach = mem_cgroup_can_attach,
5613 .cancel_attach = mem_cgroup_cancel_attach,
5614 .attach = mem_cgroup_move_task,
5615 .base_cftypes = mem_cgroup_files,
5618 .__DEPRECATED_clear_css_refs = true,
5621 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5622 static int __init enable_swap_account(char *s)
5624 /* consider enabled if no parameter or 1 is given */
5625 if (!strcmp(s, "1"))
5626 really_do_swap_account = 1;
5627 else if (!strcmp(s, "0"))
5628 really_do_swap_account = 0;
5631 __setup("swapaccount=", enable_swap_account);