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 lruvec *lruvec)
1213 unsigned long inactive_ratio;
1214 unsigned long inactive;
1215 unsigned long active;
1218 inactive = mem_cgroup_get_lruvec_size(lruvec, LRU_INACTIVE_ANON);
1219 active = mem_cgroup_get_lruvec_size(lruvec, LRU_ACTIVE_ANON);
1221 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1223 inactive_ratio = int_sqrt(10 * gb);
1227 return inactive * inactive_ratio < active;
1230 int mem_cgroup_inactive_file_is_low(struct lruvec *lruvec)
1232 unsigned long active;
1233 unsigned long inactive;
1235 inactive = mem_cgroup_get_lruvec_size(lruvec, LRU_INACTIVE_FILE);
1236 active = mem_cgroup_get_lruvec_size(lruvec, LRU_ACTIVE_FILE);
1238 return (active > inactive);
1241 struct zone_reclaim_stat *
1242 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
1244 struct page_cgroup *pc;
1245 struct mem_cgroup_per_zone *mz;
1247 if (mem_cgroup_disabled())
1250 pc = lookup_page_cgroup(page);
1251 if (!PageCgroupUsed(pc))
1253 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1255 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1256 return &mz->lruvec.reclaim_stat;
1259 #define mem_cgroup_from_res_counter(counter, member) \
1260 container_of(counter, struct mem_cgroup, member)
1263 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1264 * @mem: the memory cgroup
1266 * Returns the maximum amount of memory @mem can be charged with, in
1269 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1271 unsigned long long margin;
1273 margin = res_counter_margin(&memcg->res);
1274 if (do_swap_account)
1275 margin = min(margin, res_counter_margin(&memcg->memsw));
1276 return margin >> PAGE_SHIFT;
1279 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1281 struct cgroup *cgrp = memcg->css.cgroup;
1284 if (cgrp->parent == NULL)
1285 return vm_swappiness;
1287 return memcg->swappiness;
1291 * memcg->moving_account is used for checking possibility that some thread is
1292 * calling move_account(). When a thread on CPU-A starts moving pages under
1293 * a memcg, other threads should check memcg->moving_account under
1294 * rcu_read_lock(), like this:
1298 * memcg->moving_account+1 if (memcg->mocing_account)
1300 * synchronize_rcu() update something.
1305 /* for quick checking without looking up memcg */
1306 atomic_t memcg_moving __read_mostly;
1308 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1310 atomic_inc(&memcg_moving);
1311 atomic_inc(&memcg->moving_account);
1315 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1318 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1319 * We check NULL in callee rather than caller.
1322 atomic_dec(&memcg_moving);
1323 atomic_dec(&memcg->moving_account);
1328 * 2 routines for checking "mem" is under move_account() or not.
1330 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1331 * is used for avoiding races in accounting. If true,
1332 * pc->mem_cgroup may be overwritten.
1334 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1335 * under hierarchy of moving cgroups. This is for
1336 * waiting at hith-memory prressure caused by "move".
1339 static bool mem_cgroup_stolen(struct mem_cgroup *memcg)
1341 VM_BUG_ON(!rcu_read_lock_held());
1342 return atomic_read(&memcg->moving_account) > 0;
1345 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1347 struct mem_cgroup *from;
1348 struct mem_cgroup *to;
1351 * Unlike task_move routines, we access mc.to, mc.from not under
1352 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1354 spin_lock(&mc.lock);
1360 ret = mem_cgroup_same_or_subtree(memcg, from)
1361 || mem_cgroup_same_or_subtree(memcg, to);
1363 spin_unlock(&mc.lock);
1367 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1369 if (mc.moving_task && current != mc.moving_task) {
1370 if (mem_cgroup_under_move(memcg)) {
1372 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1373 /* moving charge context might have finished. */
1376 finish_wait(&mc.waitq, &wait);
1384 * Take this lock when
1385 * - a code tries to modify page's memcg while it's USED.
1386 * - a code tries to modify page state accounting in a memcg.
1387 * see mem_cgroup_stolen(), too.
1389 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1390 unsigned long *flags)
1392 spin_lock_irqsave(&memcg->move_lock, *flags);
1395 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1396 unsigned long *flags)
1398 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1402 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1403 * @memcg: The memory cgroup that went over limit
1404 * @p: Task that is going to be killed
1406 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1409 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1411 struct cgroup *task_cgrp;
1412 struct cgroup *mem_cgrp;
1414 * Need a buffer in BSS, can't rely on allocations. The code relies
1415 * on the assumption that OOM is serialized for memory controller.
1416 * If this assumption is broken, revisit this code.
1418 static char memcg_name[PATH_MAX];
1426 mem_cgrp = memcg->css.cgroup;
1427 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1429 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1432 * Unfortunately, we are unable to convert to a useful name
1433 * But we'll still print out the usage information
1440 printk(KERN_INFO "Task in %s killed", memcg_name);
1443 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1451 * Continues from above, so we don't need an KERN_ level
1453 printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1456 printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1457 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1458 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1459 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1460 printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1462 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1463 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1464 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1468 * This function returns the number of memcg under hierarchy tree. Returns
1469 * 1(self count) if no children.
1471 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1474 struct mem_cgroup *iter;
1476 for_each_mem_cgroup_tree(iter, memcg)
1482 * Return the memory (and swap, if configured) limit for a memcg.
1484 u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1489 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1490 limit += total_swap_pages << PAGE_SHIFT;
1492 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1494 * If memsw is finite and limits the amount of swap space available
1495 * to this memcg, return that limit.
1497 return min(limit, memsw);
1500 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1502 unsigned long flags)
1504 unsigned long total = 0;
1505 bool noswap = false;
1508 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1510 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1513 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1515 drain_all_stock_async(memcg);
1516 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1518 * Allow limit shrinkers, which are triggered directly
1519 * by userspace, to catch signals and stop reclaim
1520 * after minimal progress, regardless of the margin.
1522 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1524 if (mem_cgroup_margin(memcg))
1527 * If nothing was reclaimed after two attempts, there
1528 * may be no reclaimable pages in this hierarchy.
1537 * test_mem_cgroup_node_reclaimable
1538 * @mem: the target memcg
1539 * @nid: the node ID to be checked.
1540 * @noswap : specify true here if the user wants flle only information.
1542 * This function returns whether the specified memcg contains any
1543 * reclaimable pages on a node. Returns true if there are any reclaimable
1544 * pages in the node.
1546 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1547 int nid, bool noswap)
1549 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1551 if (noswap || !total_swap_pages)
1553 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1558 #if MAX_NUMNODES > 1
1561 * Always updating the nodemask is not very good - even if we have an empty
1562 * list or the wrong list here, we can start from some node and traverse all
1563 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1566 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1570 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1571 * pagein/pageout changes since the last update.
1573 if (!atomic_read(&memcg->numainfo_events))
1575 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1578 /* make a nodemask where this memcg uses memory from */
1579 memcg->scan_nodes = node_states[N_HIGH_MEMORY];
1581 for_each_node_mask(nid, node_states[N_HIGH_MEMORY]) {
1583 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1584 node_clear(nid, memcg->scan_nodes);
1587 atomic_set(&memcg->numainfo_events, 0);
1588 atomic_set(&memcg->numainfo_updating, 0);
1592 * Selecting a node where we start reclaim from. Because what we need is just
1593 * reducing usage counter, start from anywhere is O,K. Considering
1594 * memory reclaim from current node, there are pros. and cons.
1596 * Freeing memory from current node means freeing memory from a node which
1597 * we'll use or we've used. So, it may make LRU bad. And if several threads
1598 * hit limits, it will see a contention on a node. But freeing from remote
1599 * node means more costs for memory reclaim because of memory latency.
1601 * Now, we use round-robin. Better algorithm is welcomed.
1603 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1607 mem_cgroup_may_update_nodemask(memcg);
1608 node = memcg->last_scanned_node;
1610 node = next_node(node, memcg->scan_nodes);
1611 if (node == MAX_NUMNODES)
1612 node = first_node(memcg->scan_nodes);
1614 * We call this when we hit limit, not when pages are added to LRU.
1615 * No LRU may hold pages because all pages are UNEVICTABLE or
1616 * memcg is too small and all pages are not on LRU. In that case,
1617 * we use curret node.
1619 if (unlikely(node == MAX_NUMNODES))
1620 node = numa_node_id();
1622 memcg->last_scanned_node = node;
1627 * Check all nodes whether it contains reclaimable pages or not.
1628 * For quick scan, we make use of scan_nodes. This will allow us to skip
1629 * unused nodes. But scan_nodes is lazily updated and may not cotain
1630 * enough new information. We need to do double check.
1632 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1637 * quick check...making use of scan_node.
1638 * We can skip unused nodes.
1640 if (!nodes_empty(memcg->scan_nodes)) {
1641 for (nid = first_node(memcg->scan_nodes);
1643 nid = next_node(nid, memcg->scan_nodes)) {
1645 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1650 * Check rest of nodes.
1652 for_each_node_state(nid, N_HIGH_MEMORY) {
1653 if (node_isset(nid, memcg->scan_nodes))
1655 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1662 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1667 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1669 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
1673 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1676 unsigned long *total_scanned)
1678 struct mem_cgroup *victim = NULL;
1681 unsigned long excess;
1682 unsigned long nr_scanned;
1683 struct mem_cgroup_reclaim_cookie reclaim = {
1688 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
1691 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1696 * If we have not been able to reclaim
1697 * anything, it might because there are
1698 * no reclaimable pages under this hierarchy
1703 * We want to do more targeted reclaim.
1704 * excess >> 2 is not to excessive so as to
1705 * reclaim too much, nor too less that we keep
1706 * coming back to reclaim from this cgroup
1708 if (total >= (excess >> 2) ||
1709 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1714 if (!mem_cgroup_reclaimable(victim, false))
1716 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1718 *total_scanned += nr_scanned;
1719 if (!res_counter_soft_limit_excess(&root_memcg->res))
1722 mem_cgroup_iter_break(root_memcg, victim);
1727 * Check OOM-Killer is already running under our hierarchy.
1728 * If someone is running, return false.
1729 * Has to be called with memcg_oom_lock
1731 static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg)
1733 struct mem_cgroup *iter, *failed = NULL;
1735 for_each_mem_cgroup_tree(iter, memcg) {
1736 if (iter->oom_lock) {
1738 * this subtree of our hierarchy is already locked
1739 * so we cannot give a lock.
1742 mem_cgroup_iter_break(memcg, iter);
1745 iter->oom_lock = true;
1752 * OK, we failed to lock the whole subtree so we have to clean up
1753 * what we set up to the failing subtree
1755 for_each_mem_cgroup_tree(iter, memcg) {
1756 if (iter == failed) {
1757 mem_cgroup_iter_break(memcg, iter);
1760 iter->oom_lock = false;
1766 * Has to be called with memcg_oom_lock
1768 static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1770 struct mem_cgroup *iter;
1772 for_each_mem_cgroup_tree(iter, memcg)
1773 iter->oom_lock = false;
1777 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1779 struct mem_cgroup *iter;
1781 for_each_mem_cgroup_tree(iter, memcg)
1782 atomic_inc(&iter->under_oom);
1785 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1787 struct mem_cgroup *iter;
1790 * When a new child is created while the hierarchy is under oom,
1791 * mem_cgroup_oom_lock() may not be called. We have to use
1792 * atomic_add_unless() here.
1794 for_each_mem_cgroup_tree(iter, memcg)
1795 atomic_add_unless(&iter->under_oom, -1, 0);
1798 static DEFINE_SPINLOCK(memcg_oom_lock);
1799 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1801 struct oom_wait_info {
1802 struct mem_cgroup *memcg;
1806 static int memcg_oom_wake_function(wait_queue_t *wait,
1807 unsigned mode, int sync, void *arg)
1809 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1810 struct mem_cgroup *oom_wait_memcg;
1811 struct oom_wait_info *oom_wait_info;
1813 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1814 oom_wait_memcg = oom_wait_info->memcg;
1817 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
1818 * Then we can use css_is_ancestor without taking care of RCU.
1820 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
1821 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
1823 return autoremove_wake_function(wait, mode, sync, arg);
1826 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
1828 /* for filtering, pass "memcg" as argument. */
1829 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1832 static void memcg_oom_recover(struct mem_cgroup *memcg)
1834 if (memcg && atomic_read(&memcg->under_oom))
1835 memcg_wakeup_oom(memcg);
1839 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1841 static bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask,
1844 struct oom_wait_info owait;
1845 bool locked, need_to_kill;
1847 owait.memcg = memcg;
1848 owait.wait.flags = 0;
1849 owait.wait.func = memcg_oom_wake_function;
1850 owait.wait.private = current;
1851 INIT_LIST_HEAD(&owait.wait.task_list);
1852 need_to_kill = true;
1853 mem_cgroup_mark_under_oom(memcg);
1855 /* At first, try to OOM lock hierarchy under memcg.*/
1856 spin_lock(&memcg_oom_lock);
1857 locked = mem_cgroup_oom_lock(memcg);
1859 * Even if signal_pending(), we can't quit charge() loop without
1860 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1861 * under OOM is always welcomed, use TASK_KILLABLE here.
1863 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1864 if (!locked || memcg->oom_kill_disable)
1865 need_to_kill = false;
1867 mem_cgroup_oom_notify(memcg);
1868 spin_unlock(&memcg_oom_lock);
1871 finish_wait(&memcg_oom_waitq, &owait.wait);
1872 mem_cgroup_out_of_memory(memcg, mask, order);
1875 finish_wait(&memcg_oom_waitq, &owait.wait);
1877 spin_lock(&memcg_oom_lock);
1879 mem_cgroup_oom_unlock(memcg);
1880 memcg_wakeup_oom(memcg);
1881 spin_unlock(&memcg_oom_lock);
1883 mem_cgroup_unmark_under_oom(memcg);
1885 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
1887 /* Give chance to dying process */
1888 schedule_timeout_uninterruptible(1);
1893 * Currently used to update mapped file statistics, but the routine can be
1894 * generalized to update other statistics as well.
1896 * Notes: Race condition
1898 * We usually use page_cgroup_lock() for accessing page_cgroup member but
1899 * it tends to be costly. But considering some conditions, we doesn't need
1900 * to do so _always_.
1902 * Considering "charge", lock_page_cgroup() is not required because all
1903 * file-stat operations happen after a page is attached to radix-tree. There
1904 * are no race with "charge".
1906 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
1907 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
1908 * if there are race with "uncharge". Statistics itself is properly handled
1911 * Considering "move", this is an only case we see a race. To make the race
1912 * small, we check mm->moving_account and detect there are possibility of race
1913 * If there is, we take a lock.
1916 void __mem_cgroup_begin_update_page_stat(struct page *page,
1917 bool *locked, unsigned long *flags)
1919 struct mem_cgroup *memcg;
1920 struct page_cgroup *pc;
1922 pc = lookup_page_cgroup(page);
1924 memcg = pc->mem_cgroup;
1925 if (unlikely(!memcg || !PageCgroupUsed(pc)))
1928 * If this memory cgroup is not under account moving, we don't
1929 * need to take move_lock_page_cgroup(). Because we already hold
1930 * rcu_read_lock(), any calls to move_account will be delayed until
1931 * rcu_read_unlock() if mem_cgroup_stolen() == true.
1933 if (!mem_cgroup_stolen(memcg))
1936 move_lock_mem_cgroup(memcg, flags);
1937 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
1938 move_unlock_mem_cgroup(memcg, flags);
1944 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
1946 struct page_cgroup *pc = lookup_page_cgroup(page);
1949 * It's guaranteed that pc->mem_cgroup never changes while
1950 * lock is held because a routine modifies pc->mem_cgroup
1951 * should take move_lock_page_cgroup().
1953 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
1956 void mem_cgroup_update_page_stat(struct page *page,
1957 enum mem_cgroup_page_stat_item idx, int val)
1959 struct mem_cgroup *memcg;
1960 struct page_cgroup *pc = lookup_page_cgroup(page);
1961 unsigned long uninitialized_var(flags);
1963 if (mem_cgroup_disabled())
1966 memcg = pc->mem_cgroup;
1967 if (unlikely(!memcg || !PageCgroupUsed(pc)))
1971 case MEMCG_NR_FILE_MAPPED:
1972 idx = MEM_CGROUP_STAT_FILE_MAPPED;
1978 this_cpu_add(memcg->stat->count[idx], val);
1982 * size of first charge trial. "32" comes from vmscan.c's magic value.
1983 * TODO: maybe necessary to use big numbers in big irons.
1985 #define CHARGE_BATCH 32U
1986 struct memcg_stock_pcp {
1987 struct mem_cgroup *cached; /* this never be root cgroup */
1988 unsigned int nr_pages;
1989 struct work_struct work;
1990 unsigned long flags;
1991 #define FLUSHING_CACHED_CHARGE 0
1993 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1994 static DEFINE_MUTEX(percpu_charge_mutex);
1997 * Try to consume stocked charge on this cpu. If success, one page is consumed
1998 * from local stock and true is returned. If the stock is 0 or charges from a
1999 * cgroup which is not current target, returns false. This stock will be
2002 static bool consume_stock(struct mem_cgroup *memcg)
2004 struct memcg_stock_pcp *stock;
2007 stock = &get_cpu_var(memcg_stock);
2008 if (memcg == stock->cached && stock->nr_pages)
2010 else /* need to call res_counter_charge */
2012 put_cpu_var(memcg_stock);
2017 * Returns stocks cached in percpu to res_counter and reset cached information.
2019 static void drain_stock(struct memcg_stock_pcp *stock)
2021 struct mem_cgroup *old = stock->cached;
2023 if (stock->nr_pages) {
2024 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2026 res_counter_uncharge(&old->res, bytes);
2027 if (do_swap_account)
2028 res_counter_uncharge(&old->memsw, bytes);
2029 stock->nr_pages = 0;
2031 stock->cached = NULL;
2035 * This must be called under preempt disabled or must be called by
2036 * a thread which is pinned to local cpu.
2038 static void drain_local_stock(struct work_struct *dummy)
2040 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2042 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2046 * Cache charges(val) which is from res_counter, to local per_cpu area.
2047 * This will be consumed by consume_stock() function, later.
2049 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2051 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2053 if (stock->cached != memcg) { /* reset if necessary */
2055 stock->cached = memcg;
2057 stock->nr_pages += nr_pages;
2058 put_cpu_var(memcg_stock);
2062 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2063 * of the hierarchy under it. sync flag says whether we should block
2064 * until the work is done.
2066 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2070 /* Notify other cpus that system-wide "drain" is running */
2073 for_each_online_cpu(cpu) {
2074 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2075 struct mem_cgroup *memcg;
2077 memcg = stock->cached;
2078 if (!memcg || !stock->nr_pages)
2080 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2082 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2084 drain_local_stock(&stock->work);
2086 schedule_work_on(cpu, &stock->work);
2094 for_each_online_cpu(cpu) {
2095 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2096 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2097 flush_work(&stock->work);
2104 * Tries to drain stocked charges in other cpus. This function is asynchronous
2105 * and just put a work per cpu for draining localy on each cpu. Caller can
2106 * expects some charges will be back to res_counter later but cannot wait for
2109 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2112 * If someone calls draining, avoid adding more kworker runs.
2114 if (!mutex_trylock(&percpu_charge_mutex))
2116 drain_all_stock(root_memcg, false);
2117 mutex_unlock(&percpu_charge_mutex);
2120 /* This is a synchronous drain interface. */
2121 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2123 /* called when force_empty is called */
2124 mutex_lock(&percpu_charge_mutex);
2125 drain_all_stock(root_memcg, true);
2126 mutex_unlock(&percpu_charge_mutex);
2130 * This function drains percpu counter value from DEAD cpu and
2131 * move it to local cpu. Note that this function can be preempted.
2133 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2137 spin_lock(&memcg->pcp_counter_lock);
2138 for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) {
2139 long x = per_cpu(memcg->stat->count[i], cpu);
2141 per_cpu(memcg->stat->count[i], cpu) = 0;
2142 memcg->nocpu_base.count[i] += x;
2144 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2145 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2147 per_cpu(memcg->stat->events[i], cpu) = 0;
2148 memcg->nocpu_base.events[i] += x;
2150 spin_unlock(&memcg->pcp_counter_lock);
2153 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2154 unsigned long action,
2157 int cpu = (unsigned long)hcpu;
2158 struct memcg_stock_pcp *stock;
2159 struct mem_cgroup *iter;
2161 if (action == CPU_ONLINE)
2164 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2167 for_each_mem_cgroup(iter)
2168 mem_cgroup_drain_pcp_counter(iter, cpu);
2170 stock = &per_cpu(memcg_stock, cpu);
2176 /* See __mem_cgroup_try_charge() for details */
2178 CHARGE_OK, /* success */
2179 CHARGE_RETRY, /* need to retry but retry is not bad */
2180 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2181 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2182 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2185 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2186 unsigned int nr_pages, bool oom_check)
2188 unsigned long csize = nr_pages * PAGE_SIZE;
2189 struct mem_cgroup *mem_over_limit;
2190 struct res_counter *fail_res;
2191 unsigned long flags = 0;
2194 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2197 if (!do_swap_account)
2199 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2203 res_counter_uncharge(&memcg->res, csize);
2204 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2205 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2207 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2209 * nr_pages can be either a huge page (HPAGE_PMD_NR), a batch
2210 * of regular pages (CHARGE_BATCH), or a single regular page (1).
2212 * Never reclaim on behalf of optional batching, retry with a
2213 * single page instead.
2215 if (nr_pages == CHARGE_BATCH)
2216 return CHARGE_RETRY;
2218 if (!(gfp_mask & __GFP_WAIT))
2219 return CHARGE_WOULDBLOCK;
2221 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2222 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2223 return CHARGE_RETRY;
2225 * Even though the limit is exceeded at this point, reclaim
2226 * may have been able to free some pages. Retry the charge
2227 * before killing the task.
2229 * Only for regular pages, though: huge pages are rather
2230 * unlikely to succeed so close to the limit, and we fall back
2231 * to regular pages anyway in case of failure.
2233 if (nr_pages == 1 && ret)
2234 return CHARGE_RETRY;
2237 * At task move, charge accounts can be doubly counted. So, it's
2238 * better to wait until the end of task_move if something is going on.
2240 if (mem_cgroup_wait_acct_move(mem_over_limit))
2241 return CHARGE_RETRY;
2243 /* If we don't need to call oom-killer at el, return immediately */
2245 return CHARGE_NOMEM;
2247 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask, get_order(csize)))
2248 return CHARGE_OOM_DIE;
2250 return CHARGE_RETRY;
2254 * __mem_cgroup_try_charge() does
2255 * 1. detect memcg to be charged against from passed *mm and *ptr,
2256 * 2. update res_counter
2257 * 3. call memory reclaim if necessary.
2259 * In some special case, if the task is fatal, fatal_signal_pending() or
2260 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2261 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2262 * as possible without any hazards. 2: all pages should have a valid
2263 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2264 * pointer, that is treated as a charge to root_mem_cgroup.
2266 * So __mem_cgroup_try_charge() will return
2267 * 0 ... on success, filling *ptr with a valid memcg pointer.
2268 * -ENOMEM ... charge failure because of resource limits.
2269 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2271 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2272 * the oom-killer can be invoked.
2274 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2276 unsigned int nr_pages,
2277 struct mem_cgroup **ptr,
2280 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2281 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2282 struct mem_cgroup *memcg = NULL;
2286 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2287 * in system level. So, allow to go ahead dying process in addition to
2290 if (unlikely(test_thread_flag(TIF_MEMDIE)
2291 || fatal_signal_pending(current)))
2295 * We always charge the cgroup the mm_struct belongs to.
2296 * The mm_struct's mem_cgroup changes on task migration if the
2297 * thread group leader migrates. It's possible that mm is not
2298 * set, if so charge the init_mm (happens for pagecache usage).
2301 *ptr = root_mem_cgroup;
2303 if (*ptr) { /* css should be a valid one */
2305 VM_BUG_ON(css_is_removed(&memcg->css));
2306 if (mem_cgroup_is_root(memcg))
2308 if (nr_pages == 1 && consume_stock(memcg))
2310 css_get(&memcg->css);
2312 struct task_struct *p;
2315 p = rcu_dereference(mm->owner);
2317 * Because we don't have task_lock(), "p" can exit.
2318 * In that case, "memcg" can point to root or p can be NULL with
2319 * race with swapoff. Then, we have small risk of mis-accouning.
2320 * But such kind of mis-account by race always happens because
2321 * we don't have cgroup_mutex(). It's overkill and we allo that
2323 * (*) swapoff at el will charge against mm-struct not against
2324 * task-struct. So, mm->owner can be NULL.
2326 memcg = mem_cgroup_from_task(p);
2328 memcg = root_mem_cgroup;
2329 if (mem_cgroup_is_root(memcg)) {
2333 if (nr_pages == 1 && consume_stock(memcg)) {
2335 * It seems dagerous to access memcg without css_get().
2336 * But considering how consume_stok works, it's not
2337 * necessary. If consume_stock success, some charges
2338 * from this memcg are cached on this cpu. So, we
2339 * don't need to call css_get()/css_tryget() before
2340 * calling consume_stock().
2345 /* after here, we may be blocked. we need to get refcnt */
2346 if (!css_tryget(&memcg->css)) {
2356 /* If killed, bypass charge */
2357 if (fatal_signal_pending(current)) {
2358 css_put(&memcg->css);
2363 if (oom && !nr_oom_retries) {
2365 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2368 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, oom_check);
2372 case CHARGE_RETRY: /* not in OOM situation but retry */
2374 css_put(&memcg->css);
2377 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2378 css_put(&memcg->css);
2380 case CHARGE_NOMEM: /* OOM routine works */
2382 css_put(&memcg->css);
2385 /* If oom, we never return -ENOMEM */
2388 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2389 css_put(&memcg->css);
2392 } while (ret != CHARGE_OK);
2394 if (batch > nr_pages)
2395 refill_stock(memcg, batch - nr_pages);
2396 css_put(&memcg->css);
2404 *ptr = root_mem_cgroup;
2409 * Somemtimes we have to undo a charge we got by try_charge().
2410 * This function is for that and do uncharge, put css's refcnt.
2411 * gotten by try_charge().
2413 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2414 unsigned int nr_pages)
2416 if (!mem_cgroup_is_root(memcg)) {
2417 unsigned long bytes = nr_pages * PAGE_SIZE;
2419 res_counter_uncharge(&memcg->res, bytes);
2420 if (do_swap_account)
2421 res_counter_uncharge(&memcg->memsw, bytes);
2426 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2427 * This is useful when moving usage to parent cgroup.
2429 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2430 unsigned int nr_pages)
2432 unsigned long bytes = nr_pages * PAGE_SIZE;
2434 if (mem_cgroup_is_root(memcg))
2437 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2438 if (do_swap_account)
2439 res_counter_uncharge_until(&memcg->memsw,
2440 memcg->memsw.parent, bytes);
2444 * A helper function to get mem_cgroup from ID. must be called under
2445 * rcu_read_lock(). The caller must check css_is_removed() or some if
2446 * it's concern. (dropping refcnt from swap can be called against removed
2449 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2451 struct cgroup_subsys_state *css;
2453 /* ID 0 is unused ID */
2456 css = css_lookup(&mem_cgroup_subsys, id);
2459 return container_of(css, struct mem_cgroup, css);
2462 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2464 struct mem_cgroup *memcg = NULL;
2465 struct page_cgroup *pc;
2469 VM_BUG_ON(!PageLocked(page));
2471 pc = lookup_page_cgroup(page);
2472 lock_page_cgroup(pc);
2473 if (PageCgroupUsed(pc)) {
2474 memcg = pc->mem_cgroup;
2475 if (memcg && !css_tryget(&memcg->css))
2477 } else if (PageSwapCache(page)) {
2478 ent.val = page_private(page);
2479 id = lookup_swap_cgroup_id(ent);
2481 memcg = mem_cgroup_lookup(id);
2482 if (memcg && !css_tryget(&memcg->css))
2486 unlock_page_cgroup(pc);
2490 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2492 unsigned int nr_pages,
2493 enum charge_type ctype,
2496 struct page_cgroup *pc = lookup_page_cgroup(page);
2497 struct zone *uninitialized_var(zone);
2498 bool was_on_lru = false;
2501 lock_page_cgroup(pc);
2502 if (unlikely(PageCgroupUsed(pc))) {
2503 unlock_page_cgroup(pc);
2504 __mem_cgroup_cancel_charge(memcg, nr_pages);
2508 * we don't need page_cgroup_lock about tail pages, becase they are not
2509 * accessed by any other context at this point.
2513 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2514 * may already be on some other mem_cgroup's LRU. Take care of it.
2517 zone = page_zone(page);
2518 spin_lock_irq(&zone->lru_lock);
2519 if (PageLRU(page)) {
2521 del_page_from_lru_list(zone, page, page_lru(page));
2526 pc->mem_cgroup = memcg;
2528 * We access a page_cgroup asynchronously without lock_page_cgroup().
2529 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2530 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2531 * before USED bit, we need memory barrier here.
2532 * See mem_cgroup_add_lru_list(), etc.
2535 SetPageCgroupUsed(pc);
2539 VM_BUG_ON(PageLRU(page));
2541 add_page_to_lru_list(zone, page, page_lru(page));
2543 spin_unlock_irq(&zone->lru_lock);
2546 if (ctype == MEM_CGROUP_CHARGE_TYPE_MAPPED)
2551 mem_cgroup_charge_statistics(memcg, anon, nr_pages);
2552 unlock_page_cgroup(pc);
2555 * "charge_statistics" updated event counter. Then, check it.
2556 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2557 * if they exceeds softlimit.
2559 memcg_check_events(memcg, page);
2562 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2564 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
2566 * Because tail pages are not marked as "used", set it. We're under
2567 * zone->lru_lock, 'splitting on pmd' and compound_lock.
2568 * charge/uncharge will be never happen and move_account() is done under
2569 * compound_lock(), so we don't have to take care of races.
2571 void mem_cgroup_split_huge_fixup(struct page *head)
2573 struct page_cgroup *head_pc = lookup_page_cgroup(head);
2574 struct page_cgroup *pc;
2577 if (mem_cgroup_disabled())
2579 for (i = 1; i < HPAGE_PMD_NR; i++) {
2581 pc->mem_cgroup = head_pc->mem_cgroup;
2582 smp_wmb();/* see __commit_charge() */
2583 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2586 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2589 * mem_cgroup_move_account - move account of the page
2591 * @nr_pages: number of regular pages (>1 for huge pages)
2592 * @pc: page_cgroup of the page.
2593 * @from: mem_cgroup which the page is moved from.
2594 * @to: mem_cgroup which the page is moved to. @from != @to.
2596 * The caller must confirm following.
2597 * - page is not on LRU (isolate_page() is useful.)
2598 * - compound_lock is held when nr_pages > 1
2600 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
2603 static int mem_cgroup_move_account(struct page *page,
2604 unsigned int nr_pages,
2605 struct page_cgroup *pc,
2606 struct mem_cgroup *from,
2607 struct mem_cgroup *to)
2609 unsigned long flags;
2611 bool anon = PageAnon(page);
2613 VM_BUG_ON(from == to);
2614 VM_BUG_ON(PageLRU(page));
2616 * The page is isolated from LRU. So, collapse function
2617 * will not handle this page. But page splitting can happen.
2618 * Do this check under compound_page_lock(). The caller should
2622 if (nr_pages > 1 && !PageTransHuge(page))
2625 lock_page_cgroup(pc);
2628 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
2631 move_lock_mem_cgroup(from, &flags);
2633 if (!anon && page_mapped(page)) {
2634 /* Update mapped_file data for mem_cgroup */
2636 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2637 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2640 mem_cgroup_charge_statistics(from, anon, -nr_pages);
2642 /* caller should have done css_get */
2643 pc->mem_cgroup = to;
2644 mem_cgroup_charge_statistics(to, anon, nr_pages);
2646 * We charges against "to" which may not have any tasks. Then, "to"
2647 * can be under rmdir(). But in current implementation, caller of
2648 * this function is just force_empty() and move charge, so it's
2649 * guaranteed that "to" is never removed. So, we don't check rmdir
2652 move_unlock_mem_cgroup(from, &flags);
2655 unlock_page_cgroup(pc);
2659 memcg_check_events(to, page);
2660 memcg_check_events(from, page);
2666 * move charges to its parent.
2669 static int mem_cgroup_move_parent(struct page *page,
2670 struct page_cgroup *pc,
2671 struct mem_cgroup *child,
2674 struct mem_cgroup *parent;
2675 unsigned int nr_pages;
2676 unsigned long uninitialized_var(flags);
2680 if (mem_cgroup_is_root(child))
2684 if (!get_page_unless_zero(page))
2686 if (isolate_lru_page(page))
2689 nr_pages = hpage_nr_pages(page);
2691 parent = parent_mem_cgroup(child);
2693 * If no parent, move charges to root cgroup.
2696 parent = root_mem_cgroup;
2699 flags = compound_lock_irqsave(page);
2701 ret = mem_cgroup_move_account(page, nr_pages,
2704 __mem_cgroup_cancel_local_charge(child, nr_pages);
2707 compound_unlock_irqrestore(page, flags);
2708 putback_lru_page(page);
2716 * Charge the memory controller for page usage.
2718 * 0 if the charge was successful
2719 * < 0 if the cgroup is over its limit
2721 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2722 gfp_t gfp_mask, enum charge_type ctype)
2724 struct mem_cgroup *memcg = NULL;
2725 unsigned int nr_pages = 1;
2729 if (PageTransHuge(page)) {
2730 nr_pages <<= compound_order(page);
2731 VM_BUG_ON(!PageTransHuge(page));
2733 * Never OOM-kill a process for a huge page. The
2734 * fault handler will fall back to regular pages.
2739 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
2742 __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false);
2746 int mem_cgroup_newpage_charge(struct page *page,
2747 struct mm_struct *mm, gfp_t gfp_mask)
2749 if (mem_cgroup_disabled())
2751 VM_BUG_ON(page_mapped(page));
2752 VM_BUG_ON(page->mapping && !PageAnon(page));
2754 return mem_cgroup_charge_common(page, mm, gfp_mask,
2755 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2759 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2760 enum charge_type ctype);
2762 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2765 struct mem_cgroup *memcg = NULL;
2766 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
2769 if (mem_cgroup_disabled())
2771 if (PageCompound(page))
2776 if (!page_is_file_cache(page))
2777 type = MEM_CGROUP_CHARGE_TYPE_SHMEM;
2779 if (!PageSwapCache(page))
2780 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
2781 else { /* page is swapcache/shmem */
2782 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &memcg);
2784 __mem_cgroup_commit_charge_swapin(page, memcg, type);
2790 * While swap-in, try_charge -> commit or cancel, the page is locked.
2791 * And when try_charge() successfully returns, one refcnt to memcg without
2792 * struct page_cgroup is acquired. This refcnt will be consumed by
2793 * "commit()" or removed by "cancel()"
2795 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2797 gfp_t mask, struct mem_cgroup **memcgp)
2799 struct mem_cgroup *memcg;
2804 if (mem_cgroup_disabled())
2807 if (!do_swap_account)
2810 * A racing thread's fault, or swapoff, may have already updated
2811 * the pte, and even removed page from swap cache: in those cases
2812 * do_swap_page()'s pte_same() test will fail; but there's also a
2813 * KSM case which does need to charge the page.
2815 if (!PageSwapCache(page))
2817 memcg = try_get_mem_cgroup_from_page(page);
2821 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
2822 css_put(&memcg->css);
2829 ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
2836 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
2837 enum charge_type ctype)
2839 if (mem_cgroup_disabled())
2843 cgroup_exclude_rmdir(&memcg->css);
2845 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
2847 * Now swap is on-memory. This means this page may be
2848 * counted both as mem and swap....double count.
2849 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2850 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2851 * may call delete_from_swap_cache() before reach here.
2853 if (do_swap_account && PageSwapCache(page)) {
2854 swp_entry_t ent = {.val = page_private(page)};
2855 mem_cgroup_uncharge_swap(ent);
2858 * At swapin, we may charge account against cgroup which has no tasks.
2859 * So, rmdir()->pre_destroy() can be called while we do this charge.
2860 * In that case, we need to call pre_destroy() again. check it here.
2862 cgroup_release_and_wakeup_rmdir(&memcg->css);
2865 void mem_cgroup_commit_charge_swapin(struct page *page,
2866 struct mem_cgroup *memcg)
2868 __mem_cgroup_commit_charge_swapin(page, memcg,
2869 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2872 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
2874 if (mem_cgroup_disabled())
2878 __mem_cgroup_cancel_charge(memcg, 1);
2881 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
2882 unsigned int nr_pages,
2883 const enum charge_type ctype)
2885 struct memcg_batch_info *batch = NULL;
2886 bool uncharge_memsw = true;
2888 /* If swapout, usage of swap doesn't decrease */
2889 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2890 uncharge_memsw = false;
2892 batch = ¤t->memcg_batch;
2894 * In usual, we do css_get() when we remember memcg pointer.
2895 * But in this case, we keep res->usage until end of a series of
2896 * uncharges. Then, it's ok to ignore memcg's refcnt.
2899 batch->memcg = memcg;
2901 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2902 * In those cases, all pages freed continuously can be expected to be in
2903 * the same cgroup and we have chance to coalesce uncharges.
2904 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2905 * because we want to do uncharge as soon as possible.
2908 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
2909 goto direct_uncharge;
2912 goto direct_uncharge;
2915 * In typical case, batch->memcg == mem. This means we can
2916 * merge a series of uncharges to an uncharge of res_counter.
2917 * If not, we uncharge res_counter ony by one.
2919 if (batch->memcg != memcg)
2920 goto direct_uncharge;
2921 /* remember freed charge and uncharge it later */
2924 batch->memsw_nr_pages++;
2927 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
2929 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
2930 if (unlikely(batch->memcg != memcg))
2931 memcg_oom_recover(memcg);
2935 * uncharge if !page_mapped(page)
2937 static struct mem_cgroup *
2938 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
2940 struct mem_cgroup *memcg = NULL;
2941 unsigned int nr_pages = 1;
2942 struct page_cgroup *pc;
2945 if (mem_cgroup_disabled())
2948 if (PageSwapCache(page))
2951 if (PageTransHuge(page)) {
2952 nr_pages <<= compound_order(page);
2953 VM_BUG_ON(!PageTransHuge(page));
2956 * Check if our page_cgroup is valid
2958 pc = lookup_page_cgroup(page);
2959 if (unlikely(!PageCgroupUsed(pc)))
2962 lock_page_cgroup(pc);
2964 memcg = pc->mem_cgroup;
2966 if (!PageCgroupUsed(pc))
2969 anon = PageAnon(page);
2972 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2974 * Generally PageAnon tells if it's the anon statistics to be
2975 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
2976 * used before page reached the stage of being marked PageAnon.
2980 case MEM_CGROUP_CHARGE_TYPE_DROP:
2981 /* See mem_cgroup_prepare_migration() */
2982 if (page_mapped(page) || PageCgroupMigration(pc))
2985 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
2986 if (!PageAnon(page)) { /* Shared memory */
2987 if (page->mapping && !page_is_file_cache(page))
2989 } else if (page_mapped(page)) /* Anon */
2996 mem_cgroup_charge_statistics(memcg, anon, -nr_pages);
2998 ClearPageCgroupUsed(pc);
3000 * pc->mem_cgroup is not cleared here. It will be accessed when it's
3001 * freed from LRU. This is safe because uncharged page is expected not
3002 * to be reused (freed soon). Exception is SwapCache, it's handled by
3003 * special functions.
3006 unlock_page_cgroup(pc);
3008 * even after unlock, we have memcg->res.usage here and this memcg
3009 * will never be freed.
3011 memcg_check_events(memcg, page);
3012 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
3013 mem_cgroup_swap_statistics(memcg, true);
3014 mem_cgroup_get(memcg);
3016 if (!mem_cgroup_is_root(memcg))
3017 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
3022 unlock_page_cgroup(pc);
3026 void mem_cgroup_uncharge_page(struct page *page)
3029 if (page_mapped(page))
3031 VM_BUG_ON(page->mapping && !PageAnon(page));
3032 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
3035 void mem_cgroup_uncharge_cache_page(struct page *page)
3037 VM_BUG_ON(page_mapped(page));
3038 VM_BUG_ON(page->mapping);
3039 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
3043 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
3044 * In that cases, pages are freed continuously and we can expect pages
3045 * are in the same memcg. All these calls itself limits the number of
3046 * pages freed at once, then uncharge_start/end() is called properly.
3047 * This may be called prural(2) times in a context,
3050 void mem_cgroup_uncharge_start(void)
3052 current->memcg_batch.do_batch++;
3053 /* We can do nest. */
3054 if (current->memcg_batch.do_batch == 1) {
3055 current->memcg_batch.memcg = NULL;
3056 current->memcg_batch.nr_pages = 0;
3057 current->memcg_batch.memsw_nr_pages = 0;
3061 void mem_cgroup_uncharge_end(void)
3063 struct memcg_batch_info *batch = ¤t->memcg_batch;
3065 if (!batch->do_batch)
3069 if (batch->do_batch) /* If stacked, do nothing. */
3075 * This "batch->memcg" is valid without any css_get/put etc...
3076 * bacause we hide charges behind us.
3078 if (batch->nr_pages)
3079 res_counter_uncharge(&batch->memcg->res,
3080 batch->nr_pages * PAGE_SIZE);
3081 if (batch->memsw_nr_pages)
3082 res_counter_uncharge(&batch->memcg->memsw,
3083 batch->memsw_nr_pages * PAGE_SIZE);
3084 memcg_oom_recover(batch->memcg);
3085 /* forget this pointer (for sanity check) */
3086 batch->memcg = NULL;
3091 * called after __delete_from_swap_cache() and drop "page" account.
3092 * memcg information is recorded to swap_cgroup of "ent"
3095 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
3097 struct mem_cgroup *memcg;
3098 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
3100 if (!swapout) /* this was a swap cache but the swap is unused ! */
3101 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
3103 memcg = __mem_cgroup_uncharge_common(page, ctype);
3106 * record memcg information, if swapout && memcg != NULL,
3107 * mem_cgroup_get() was called in uncharge().
3109 if (do_swap_account && swapout && memcg)
3110 swap_cgroup_record(ent, css_id(&memcg->css));
3114 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3116 * called from swap_entry_free(). remove record in swap_cgroup and
3117 * uncharge "memsw" account.
3119 void mem_cgroup_uncharge_swap(swp_entry_t ent)
3121 struct mem_cgroup *memcg;
3124 if (!do_swap_account)
3127 id = swap_cgroup_record(ent, 0);
3129 memcg = mem_cgroup_lookup(id);
3132 * We uncharge this because swap is freed.
3133 * This memcg can be obsolete one. We avoid calling css_tryget
3135 if (!mem_cgroup_is_root(memcg))
3136 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
3137 mem_cgroup_swap_statistics(memcg, false);
3138 mem_cgroup_put(memcg);
3144 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3145 * @entry: swap entry to be moved
3146 * @from: mem_cgroup which the entry is moved from
3147 * @to: mem_cgroup which the entry is moved to
3149 * It succeeds only when the swap_cgroup's record for this entry is the same
3150 * as the mem_cgroup's id of @from.
3152 * Returns 0 on success, -EINVAL on failure.
3154 * The caller must have charged to @to, IOW, called res_counter_charge() about
3155 * both res and memsw, and called css_get().
3157 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3158 struct mem_cgroup *from, struct mem_cgroup *to)
3160 unsigned short old_id, new_id;
3162 old_id = css_id(&from->css);
3163 new_id = css_id(&to->css);
3165 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3166 mem_cgroup_swap_statistics(from, false);
3167 mem_cgroup_swap_statistics(to, true);
3169 * This function is only called from task migration context now.
3170 * It postpones res_counter and refcount handling till the end
3171 * of task migration(mem_cgroup_clear_mc()) for performance
3172 * improvement. But we cannot postpone mem_cgroup_get(to)
3173 * because if the process that has been moved to @to does
3174 * swap-in, the refcount of @to might be decreased to 0.
3182 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3183 struct mem_cgroup *from, struct mem_cgroup *to)
3190 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
3193 int mem_cgroup_prepare_migration(struct page *page,
3194 struct page *newpage, struct mem_cgroup **memcgp, gfp_t gfp_mask)
3196 struct mem_cgroup *memcg = NULL;
3197 struct page_cgroup *pc;
3198 enum charge_type ctype;
3203 VM_BUG_ON(PageTransHuge(page));
3204 if (mem_cgroup_disabled())
3207 pc = lookup_page_cgroup(page);
3208 lock_page_cgroup(pc);
3209 if (PageCgroupUsed(pc)) {
3210 memcg = pc->mem_cgroup;
3211 css_get(&memcg->css);
3213 * At migrating an anonymous page, its mapcount goes down
3214 * to 0 and uncharge() will be called. But, even if it's fully
3215 * unmapped, migration may fail and this page has to be
3216 * charged again. We set MIGRATION flag here and delay uncharge
3217 * until end_migration() is called
3219 * Corner Case Thinking
3221 * When the old page was mapped as Anon and it's unmap-and-freed
3222 * while migration was ongoing.
3223 * If unmap finds the old page, uncharge() of it will be delayed
3224 * until end_migration(). If unmap finds a new page, it's
3225 * uncharged when it make mapcount to be 1->0. If unmap code
3226 * finds swap_migration_entry, the new page will not be mapped
3227 * and end_migration() will find it(mapcount==0).
3230 * When the old page was mapped but migraion fails, the kernel
3231 * remaps it. A charge for it is kept by MIGRATION flag even
3232 * if mapcount goes down to 0. We can do remap successfully
3233 * without charging it again.
3236 * The "old" page is under lock_page() until the end of
3237 * migration, so, the old page itself will not be swapped-out.
3238 * If the new page is swapped out before end_migraton, our
3239 * hook to usual swap-out path will catch the event.
3242 SetPageCgroupMigration(pc);
3244 unlock_page_cgroup(pc);
3246 * If the page is not charged at this point,
3253 ret = __mem_cgroup_try_charge(NULL, gfp_mask, 1, memcgp, false);
3254 css_put(&memcg->css);/* drop extra refcnt */
3256 if (PageAnon(page)) {
3257 lock_page_cgroup(pc);
3258 ClearPageCgroupMigration(pc);
3259 unlock_page_cgroup(pc);
3261 * The old page may be fully unmapped while we kept it.
3263 mem_cgroup_uncharge_page(page);
3265 /* we'll need to revisit this error code (we have -EINTR) */
3269 * We charge new page before it's used/mapped. So, even if unlock_page()
3270 * is called before end_migration, we can catch all events on this new
3271 * page. In the case new page is migrated but not remapped, new page's
3272 * mapcount will be finally 0 and we call uncharge in end_migration().
3275 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
3276 else if (page_is_file_cache(page))
3277 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
3279 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3280 __mem_cgroup_commit_charge(memcg, newpage, 1, ctype, false);
3284 /* remove redundant charge if migration failed*/
3285 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
3286 struct page *oldpage, struct page *newpage, bool migration_ok)
3288 struct page *used, *unused;
3289 struct page_cgroup *pc;
3294 /* blocks rmdir() */
3295 cgroup_exclude_rmdir(&memcg->css);
3296 if (!migration_ok) {
3304 * We disallowed uncharge of pages under migration because mapcount
3305 * of the page goes down to zero, temporarly.
3306 * Clear the flag and check the page should be charged.
3308 pc = lookup_page_cgroup(oldpage);
3309 lock_page_cgroup(pc);
3310 ClearPageCgroupMigration(pc);
3311 unlock_page_cgroup(pc);
3312 anon = PageAnon(used);
3313 __mem_cgroup_uncharge_common(unused,
3314 anon ? MEM_CGROUP_CHARGE_TYPE_MAPPED
3315 : MEM_CGROUP_CHARGE_TYPE_CACHE);
3318 * If a page is a file cache, radix-tree replacement is very atomic
3319 * and we can skip this check. When it was an Anon page, its mapcount
3320 * goes down to 0. But because we added MIGRATION flage, it's not
3321 * uncharged yet. There are several case but page->mapcount check
3322 * and USED bit check in mem_cgroup_uncharge_page() will do enough
3323 * check. (see prepare_charge() also)
3326 mem_cgroup_uncharge_page(used);
3328 * At migration, we may charge account against cgroup which has no
3330 * So, rmdir()->pre_destroy() can be called while we do this charge.
3331 * In that case, we need to call pre_destroy() again. check it here.
3333 cgroup_release_and_wakeup_rmdir(&memcg->css);
3337 * At replace page cache, newpage is not under any memcg but it's on
3338 * LRU. So, this function doesn't touch res_counter but handles LRU
3339 * in correct way. Both pages are locked so we cannot race with uncharge.
3341 void mem_cgroup_replace_page_cache(struct page *oldpage,
3342 struct page *newpage)
3344 struct mem_cgroup *memcg = NULL;
3345 struct page_cgroup *pc;
3346 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
3348 if (mem_cgroup_disabled())
3351 pc = lookup_page_cgroup(oldpage);
3352 /* fix accounting on old pages */
3353 lock_page_cgroup(pc);
3354 if (PageCgroupUsed(pc)) {
3355 memcg = pc->mem_cgroup;
3356 mem_cgroup_charge_statistics(memcg, false, -1);
3357 ClearPageCgroupUsed(pc);
3359 unlock_page_cgroup(pc);
3362 * When called from shmem_replace_page(), in some cases the
3363 * oldpage has already been charged, and in some cases not.
3368 if (PageSwapBacked(oldpage))
3369 type = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3372 * Even if newpage->mapping was NULL before starting replacement,
3373 * the newpage may be on LRU(or pagevec for LRU) already. We lock
3374 * LRU while we overwrite pc->mem_cgroup.
3376 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
3379 #ifdef CONFIG_DEBUG_VM
3380 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3382 struct page_cgroup *pc;
3384 pc = lookup_page_cgroup(page);
3386 * Can be NULL while feeding pages into the page allocator for
3387 * the first time, i.e. during boot or memory hotplug;
3388 * or when mem_cgroup_disabled().
3390 if (likely(pc) && PageCgroupUsed(pc))
3395 bool mem_cgroup_bad_page_check(struct page *page)
3397 if (mem_cgroup_disabled())
3400 return lookup_page_cgroup_used(page) != NULL;
3403 void mem_cgroup_print_bad_page(struct page *page)
3405 struct page_cgroup *pc;
3407 pc = lookup_page_cgroup_used(page);
3409 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
3410 pc, pc->flags, pc->mem_cgroup);
3415 static DEFINE_MUTEX(set_limit_mutex);
3417 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3418 unsigned long long val)
3421 u64 memswlimit, memlimit;
3423 int children = mem_cgroup_count_children(memcg);
3424 u64 curusage, oldusage;
3428 * For keeping hierarchical_reclaim simple, how long we should retry
3429 * is depends on callers. We set our retry-count to be function
3430 * of # of children which we should visit in this loop.
3432 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3434 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3437 while (retry_count) {
3438 if (signal_pending(current)) {
3443 * Rather than hide all in some function, I do this in
3444 * open coded manner. You see what this really does.
3445 * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3447 mutex_lock(&set_limit_mutex);
3448 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3449 if (memswlimit < val) {
3451 mutex_unlock(&set_limit_mutex);
3455 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3459 ret = res_counter_set_limit(&memcg->res, val);
3461 if (memswlimit == val)
3462 memcg->memsw_is_minimum = true;
3464 memcg->memsw_is_minimum = false;
3466 mutex_unlock(&set_limit_mutex);
3471 mem_cgroup_reclaim(memcg, GFP_KERNEL,
3472 MEM_CGROUP_RECLAIM_SHRINK);
3473 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3474 /* Usage is reduced ? */
3475 if (curusage >= oldusage)
3478 oldusage = curusage;
3480 if (!ret && enlarge)
3481 memcg_oom_recover(memcg);
3486 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3487 unsigned long long val)
3490 u64 memlimit, memswlimit, oldusage, curusage;
3491 int children = mem_cgroup_count_children(memcg);
3495 /* see mem_cgroup_resize_res_limit */
3496 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3497 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3498 while (retry_count) {
3499 if (signal_pending(current)) {
3504 * Rather than hide all in some function, I do this in
3505 * open coded manner. You see what this really does.
3506 * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3508 mutex_lock(&set_limit_mutex);
3509 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3510 if (memlimit > val) {
3512 mutex_unlock(&set_limit_mutex);
3515 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3516 if (memswlimit < val)
3518 ret = res_counter_set_limit(&memcg->memsw, val);
3520 if (memlimit == val)
3521 memcg->memsw_is_minimum = true;
3523 memcg->memsw_is_minimum = false;
3525 mutex_unlock(&set_limit_mutex);
3530 mem_cgroup_reclaim(memcg, GFP_KERNEL,
3531 MEM_CGROUP_RECLAIM_NOSWAP |
3532 MEM_CGROUP_RECLAIM_SHRINK);
3533 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3534 /* Usage is reduced ? */
3535 if (curusage >= oldusage)
3538 oldusage = curusage;
3540 if (!ret && enlarge)
3541 memcg_oom_recover(memcg);
3545 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3547 unsigned long *total_scanned)
3549 unsigned long nr_reclaimed = 0;
3550 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3551 unsigned long reclaimed;
3553 struct mem_cgroup_tree_per_zone *mctz;
3554 unsigned long long excess;
3555 unsigned long nr_scanned;
3560 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3562 * This loop can run a while, specially if mem_cgroup's continuously
3563 * keep exceeding their soft limit and putting the system under
3570 mz = mem_cgroup_largest_soft_limit_node(mctz);
3575 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
3576 gfp_mask, &nr_scanned);
3577 nr_reclaimed += reclaimed;
3578 *total_scanned += nr_scanned;
3579 spin_lock(&mctz->lock);
3582 * If we failed to reclaim anything from this memory cgroup
3583 * it is time to move on to the next cgroup
3589 * Loop until we find yet another one.
3591 * By the time we get the soft_limit lock
3592 * again, someone might have aded the
3593 * group back on the RB tree. Iterate to
3594 * make sure we get a different mem.
3595 * mem_cgroup_largest_soft_limit_node returns
3596 * NULL if no other cgroup is present on
3600 __mem_cgroup_largest_soft_limit_node(mctz);
3602 css_put(&next_mz->memcg->css);
3603 else /* next_mz == NULL or other memcg */
3607 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
3608 excess = res_counter_soft_limit_excess(&mz->memcg->res);
3610 * One school of thought says that we should not add
3611 * back the node to the tree if reclaim returns 0.
3612 * But our reclaim could return 0, simply because due
3613 * to priority we are exposing a smaller subset of
3614 * memory to reclaim from. Consider this as a longer
3617 /* If excess == 0, no tree ops */
3618 __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess);
3619 spin_unlock(&mctz->lock);
3620 css_put(&mz->memcg->css);
3623 * Could not reclaim anything and there are no more
3624 * mem cgroups to try or we seem to be looping without
3625 * reclaiming anything.
3627 if (!nr_reclaimed &&
3629 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3631 } while (!nr_reclaimed);
3633 css_put(&next_mz->memcg->css);
3634 return nr_reclaimed;
3638 * This routine traverse page_cgroup in given list and drop them all.
3639 * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3641 static int mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
3642 int node, int zid, enum lru_list lru)
3644 struct mem_cgroup_per_zone *mz;
3645 unsigned long flags, loop;
3646 struct list_head *list;
3651 zone = &NODE_DATA(node)->node_zones[zid];
3652 mz = mem_cgroup_zoneinfo(memcg, node, zid);
3653 list = &mz->lruvec.lists[lru];
3655 loop = mz->lru_size[lru];
3656 /* give some margin against EBUSY etc...*/
3660 struct page_cgroup *pc;
3664 spin_lock_irqsave(&zone->lru_lock, flags);
3665 if (list_empty(list)) {
3666 spin_unlock_irqrestore(&zone->lru_lock, flags);
3669 page = list_entry(list->prev, struct page, lru);
3671 list_move(&page->lru, list);
3673 spin_unlock_irqrestore(&zone->lru_lock, flags);
3676 spin_unlock_irqrestore(&zone->lru_lock, flags);
3678 pc = lookup_page_cgroup(page);
3680 ret = mem_cgroup_move_parent(page, pc, memcg, GFP_KERNEL);
3681 if (ret == -ENOMEM || ret == -EINTR)
3684 if (ret == -EBUSY || ret == -EINVAL) {
3685 /* found lock contention or "pc" is obsolete. */
3692 if (!ret && !list_empty(list))
3698 * make mem_cgroup's charge to be 0 if there is no task.
3699 * This enables deleting this mem_cgroup.
3701 static int mem_cgroup_force_empty(struct mem_cgroup *memcg, bool free_all)
3704 int node, zid, shrink;
3705 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3706 struct cgroup *cgrp = memcg->css.cgroup;
3708 css_get(&memcg->css);
3711 /* should free all ? */
3717 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3720 if (signal_pending(current))
3722 /* This is for making all *used* pages to be on LRU. */
3723 lru_add_drain_all();
3724 drain_all_stock_sync(memcg);
3726 mem_cgroup_start_move(memcg);
3727 for_each_node_state(node, N_HIGH_MEMORY) {
3728 for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3731 ret = mem_cgroup_force_empty_list(memcg,
3740 mem_cgroup_end_move(memcg);
3741 memcg_oom_recover(memcg);
3742 /* it seems parent cgroup doesn't have enough mem */
3746 /* "ret" should also be checked to ensure all lists are empty. */
3747 } while (res_counter_read_u64(&memcg->res, RES_USAGE) > 0 || ret);
3749 css_put(&memcg->css);
3753 /* returns EBUSY if there is a task or if we come here twice. */
3754 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3758 /* we call try-to-free pages for make this cgroup empty */
3759 lru_add_drain_all();
3760 /* try to free all pages in this cgroup */
3762 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
3765 if (signal_pending(current)) {
3769 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
3773 /* maybe some writeback is necessary */
3774 congestion_wait(BLK_RW_ASYNC, HZ/10);
3779 /* try move_account...there may be some *locked* pages. */
3783 static int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3785 return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3789 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3791 return mem_cgroup_from_cont(cont)->use_hierarchy;
3794 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3798 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3799 struct cgroup *parent = cont->parent;
3800 struct mem_cgroup *parent_memcg = NULL;
3803 parent_memcg = mem_cgroup_from_cont(parent);
3807 * If parent's use_hierarchy is set, we can't make any modifications
3808 * in the child subtrees. If it is unset, then the change can
3809 * occur, provided the current cgroup has no children.
3811 * For the root cgroup, parent_mem is NULL, we allow value to be
3812 * set if there are no children.
3814 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3815 (val == 1 || val == 0)) {
3816 if (list_empty(&cont->children))
3817 memcg->use_hierarchy = val;
3828 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
3829 enum mem_cgroup_stat_index idx)
3831 struct mem_cgroup *iter;
3834 /* Per-cpu values can be negative, use a signed accumulator */
3835 for_each_mem_cgroup_tree(iter, memcg)
3836 val += mem_cgroup_read_stat(iter, idx);
3838 if (val < 0) /* race ? */
3843 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3847 if (!mem_cgroup_is_root(memcg)) {
3849 return res_counter_read_u64(&memcg->res, RES_USAGE);
3851 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
3854 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
3855 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
3858 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
3860 return val << PAGE_SHIFT;
3863 static ssize_t mem_cgroup_read(struct cgroup *cont, struct cftype *cft,
3864 struct file *file, char __user *buf,
3865 size_t nbytes, loff_t *ppos)
3867 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3870 int type, name, len;
3872 type = MEMFILE_TYPE(cft->private);
3873 name = MEMFILE_ATTR(cft->private);
3875 if (!do_swap_account && type == _MEMSWAP)
3880 if (name == RES_USAGE)
3881 val = mem_cgroup_usage(memcg, false);
3883 val = res_counter_read_u64(&memcg->res, name);
3886 if (name == RES_USAGE)
3887 val = mem_cgroup_usage(memcg, true);
3889 val = res_counter_read_u64(&memcg->memsw, name);
3895 len = scnprintf(str, sizeof(str), "%llu\n", (unsigned long long)val);
3896 return simple_read_from_buffer(buf, nbytes, ppos, str, len);
3899 * The user of this function is...
3902 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
3905 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3907 unsigned long long val;
3910 type = MEMFILE_TYPE(cft->private);
3911 name = MEMFILE_ATTR(cft->private);
3913 if (!do_swap_account && type == _MEMSWAP)
3918 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3922 /* This function does all necessary parse...reuse it */
3923 ret = res_counter_memparse_write_strategy(buffer, &val);
3927 ret = mem_cgroup_resize_limit(memcg, val);
3929 ret = mem_cgroup_resize_memsw_limit(memcg, val);
3931 case RES_SOFT_LIMIT:
3932 ret = res_counter_memparse_write_strategy(buffer, &val);
3936 * For memsw, soft limits are hard to implement in terms
3937 * of semantics, for now, we support soft limits for
3938 * control without swap
3941 ret = res_counter_set_soft_limit(&memcg->res, val);
3946 ret = -EINVAL; /* should be BUG() ? */
3952 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
3953 unsigned long long *mem_limit, unsigned long long *memsw_limit)
3955 struct cgroup *cgroup;
3956 unsigned long long min_limit, min_memsw_limit, tmp;
3958 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3959 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3960 cgroup = memcg->css.cgroup;
3961 if (!memcg->use_hierarchy)
3964 while (cgroup->parent) {
3965 cgroup = cgroup->parent;
3966 memcg = mem_cgroup_from_cont(cgroup);
3967 if (!memcg->use_hierarchy)
3969 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
3970 min_limit = min(min_limit, tmp);
3971 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3972 min_memsw_limit = min(min_memsw_limit, tmp);
3975 *mem_limit = min_limit;
3976 *memsw_limit = min_memsw_limit;
3979 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
3981 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3984 type = MEMFILE_TYPE(event);
3985 name = MEMFILE_ATTR(event);
3987 if (!do_swap_account && type == _MEMSWAP)
3993 res_counter_reset_max(&memcg->res);
3995 res_counter_reset_max(&memcg->memsw);
3999 res_counter_reset_failcnt(&memcg->res);
4001 res_counter_reset_failcnt(&memcg->memsw);
4008 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
4011 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
4015 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4016 struct cftype *cft, u64 val)
4018 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4020 if (val >= (1 << NR_MOVE_TYPE))
4023 * We check this value several times in both in can_attach() and
4024 * attach(), so we need cgroup lock to prevent this value from being
4028 memcg->move_charge_at_immigrate = val;
4034 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4035 struct cftype *cft, u64 val)
4042 /* For read statistics */
4060 struct mcs_total_stat {
4061 s64 stat[NR_MCS_STAT];
4067 } memcg_stat_strings[NR_MCS_STAT] = {
4068 {"cache", "total_cache"},
4069 {"rss", "total_rss"},
4070 {"mapped_file", "total_mapped_file"},
4071 {"pgpgin", "total_pgpgin"},
4072 {"pgpgout", "total_pgpgout"},
4073 {"swap", "total_swap"},
4074 {"pgfault", "total_pgfault"},
4075 {"pgmajfault", "total_pgmajfault"},
4076 {"inactive_anon", "total_inactive_anon"},
4077 {"active_anon", "total_active_anon"},
4078 {"inactive_file", "total_inactive_file"},
4079 {"active_file", "total_active_file"},
4080 {"unevictable", "total_unevictable"}
4085 mem_cgroup_get_local_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4090 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_CACHE);
4091 s->stat[MCS_CACHE] += val * PAGE_SIZE;
4092 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_RSS);
4093 s->stat[MCS_RSS] += val * PAGE_SIZE;
4094 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_FILE_MAPPED);
4095 s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
4096 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGIN);
4097 s->stat[MCS_PGPGIN] += val;
4098 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGOUT);
4099 s->stat[MCS_PGPGOUT] += val;
4100 if (do_swap_account) {
4101 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
4102 s->stat[MCS_SWAP] += val * PAGE_SIZE;
4104 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGFAULT);
4105 s->stat[MCS_PGFAULT] += val;
4106 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGMAJFAULT);
4107 s->stat[MCS_PGMAJFAULT] += val;
4110 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_ANON));
4111 s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
4112 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_ANON));
4113 s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
4114 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_FILE));
4115 s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
4116 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_FILE));
4117 s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
4118 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
4119 s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
4123 mem_cgroup_get_total_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4125 struct mem_cgroup *iter;
4127 for_each_mem_cgroup_tree(iter, memcg)
4128 mem_cgroup_get_local_stat(iter, s);
4132 static int mem_control_numa_stat_show(struct seq_file *m, void *arg)
4135 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
4136 unsigned long node_nr;
4137 struct cgroup *cont = m->private;
4138 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4140 total_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL);
4141 seq_printf(m, "total=%lu", total_nr);
4142 for_each_node_state(nid, N_HIGH_MEMORY) {
4143 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL);
4144 seq_printf(m, " N%d=%lu", nid, node_nr);
4148 file_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_FILE);
4149 seq_printf(m, "file=%lu", file_nr);
4150 for_each_node_state(nid, N_HIGH_MEMORY) {
4151 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4153 seq_printf(m, " N%d=%lu", nid, node_nr);
4157 anon_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_ANON);
4158 seq_printf(m, "anon=%lu", anon_nr);
4159 for_each_node_state(nid, N_HIGH_MEMORY) {
4160 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4162 seq_printf(m, " N%d=%lu", nid, node_nr);
4166 unevictable_nr = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
4167 seq_printf(m, "unevictable=%lu", unevictable_nr);
4168 for_each_node_state(nid, N_HIGH_MEMORY) {
4169 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4170 BIT(LRU_UNEVICTABLE));
4171 seq_printf(m, " N%d=%lu", nid, node_nr);
4176 #endif /* CONFIG_NUMA */
4178 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
4179 struct cgroup_map_cb *cb)
4181 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4182 struct mcs_total_stat mystat;
4185 memset(&mystat, 0, sizeof(mystat));
4186 mem_cgroup_get_local_stat(memcg, &mystat);
4189 for (i = 0; i < NR_MCS_STAT; i++) {
4190 if (i == MCS_SWAP && !do_swap_account)
4192 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
4195 /* Hierarchical information */
4197 unsigned long long limit, memsw_limit;
4198 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
4199 cb->fill(cb, "hierarchical_memory_limit", limit);
4200 if (do_swap_account)
4201 cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
4204 memset(&mystat, 0, sizeof(mystat));
4205 mem_cgroup_get_total_stat(memcg, &mystat);
4206 for (i = 0; i < NR_MCS_STAT; i++) {
4207 if (i == MCS_SWAP && !do_swap_account)
4209 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
4212 #ifdef CONFIG_DEBUG_VM
4215 struct mem_cgroup_per_zone *mz;
4216 struct zone_reclaim_stat *rstat;
4217 unsigned long recent_rotated[2] = {0, 0};
4218 unsigned long recent_scanned[2] = {0, 0};
4220 for_each_online_node(nid)
4221 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4222 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
4223 rstat = &mz->lruvec.reclaim_stat;
4225 recent_rotated[0] += rstat->recent_rotated[0];
4226 recent_rotated[1] += rstat->recent_rotated[1];
4227 recent_scanned[0] += rstat->recent_scanned[0];
4228 recent_scanned[1] += rstat->recent_scanned[1];
4230 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
4231 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
4232 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
4233 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
4240 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
4242 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4244 return mem_cgroup_swappiness(memcg);
4247 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
4250 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4251 struct mem_cgroup *parent;
4256 if (cgrp->parent == NULL)
4259 parent = mem_cgroup_from_cont(cgrp->parent);
4263 /* If under hierarchy, only empty-root can set this value */
4264 if ((parent->use_hierarchy) ||
4265 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4270 memcg->swappiness = val;
4277 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4279 struct mem_cgroup_threshold_ary *t;
4285 t = rcu_dereference(memcg->thresholds.primary);
4287 t = rcu_dereference(memcg->memsw_thresholds.primary);
4292 usage = mem_cgroup_usage(memcg, swap);
4295 * current_threshold points to threshold just below or equal to usage.
4296 * If it's not true, a threshold was crossed after last
4297 * call of __mem_cgroup_threshold().
4299 i = t->current_threshold;
4302 * Iterate backward over array of thresholds starting from
4303 * current_threshold and check if a threshold is crossed.
4304 * If none of thresholds below usage is crossed, we read
4305 * only one element of the array here.
4307 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4308 eventfd_signal(t->entries[i].eventfd, 1);
4310 /* i = current_threshold + 1 */
4314 * Iterate forward over array of thresholds starting from
4315 * current_threshold+1 and check if a threshold is crossed.
4316 * If none of thresholds above usage is crossed, we read
4317 * only one element of the array here.
4319 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4320 eventfd_signal(t->entries[i].eventfd, 1);
4322 /* Update current_threshold */
4323 t->current_threshold = i - 1;
4328 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4331 __mem_cgroup_threshold(memcg, false);
4332 if (do_swap_account)
4333 __mem_cgroup_threshold(memcg, true);
4335 memcg = parent_mem_cgroup(memcg);
4339 static int compare_thresholds(const void *a, const void *b)
4341 const struct mem_cgroup_threshold *_a = a;
4342 const struct mem_cgroup_threshold *_b = b;
4344 return _a->threshold - _b->threshold;
4347 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4349 struct mem_cgroup_eventfd_list *ev;
4351 list_for_each_entry(ev, &memcg->oom_notify, list)
4352 eventfd_signal(ev->eventfd, 1);
4356 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4358 struct mem_cgroup *iter;
4360 for_each_mem_cgroup_tree(iter, memcg)
4361 mem_cgroup_oom_notify_cb(iter);
4364 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
4365 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4367 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4368 struct mem_cgroup_thresholds *thresholds;
4369 struct mem_cgroup_threshold_ary *new;
4370 int type = MEMFILE_TYPE(cft->private);
4371 u64 threshold, usage;
4374 ret = res_counter_memparse_write_strategy(args, &threshold);
4378 mutex_lock(&memcg->thresholds_lock);
4381 thresholds = &memcg->thresholds;
4382 else if (type == _MEMSWAP)
4383 thresholds = &memcg->memsw_thresholds;
4387 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4389 /* Check if a threshold crossed before adding a new one */
4390 if (thresholds->primary)
4391 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4393 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4395 /* Allocate memory for new array of thresholds */
4396 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4404 /* Copy thresholds (if any) to new array */
4405 if (thresholds->primary) {
4406 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4407 sizeof(struct mem_cgroup_threshold));
4410 /* Add new threshold */
4411 new->entries[size - 1].eventfd = eventfd;
4412 new->entries[size - 1].threshold = threshold;
4414 /* Sort thresholds. Registering of new threshold isn't time-critical */
4415 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4416 compare_thresholds, NULL);
4418 /* Find current threshold */
4419 new->current_threshold = -1;
4420 for (i = 0; i < size; i++) {
4421 if (new->entries[i].threshold <= usage) {
4423 * new->current_threshold will not be used until
4424 * rcu_assign_pointer(), so it's safe to increment
4427 ++new->current_threshold;
4432 /* Free old spare buffer and save old primary buffer as spare */
4433 kfree(thresholds->spare);
4434 thresholds->spare = thresholds->primary;
4436 rcu_assign_pointer(thresholds->primary, new);
4438 /* To be sure that nobody uses thresholds */
4442 mutex_unlock(&memcg->thresholds_lock);
4447 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
4448 struct cftype *cft, struct eventfd_ctx *eventfd)
4450 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4451 struct mem_cgroup_thresholds *thresholds;
4452 struct mem_cgroup_threshold_ary *new;
4453 int type = MEMFILE_TYPE(cft->private);
4457 mutex_lock(&memcg->thresholds_lock);
4459 thresholds = &memcg->thresholds;
4460 else if (type == _MEMSWAP)
4461 thresholds = &memcg->memsw_thresholds;
4465 if (!thresholds->primary)
4468 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4470 /* Check if a threshold crossed before removing */
4471 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4473 /* Calculate new number of threshold */
4475 for (i = 0; i < thresholds->primary->size; i++) {
4476 if (thresholds->primary->entries[i].eventfd != eventfd)
4480 new = thresholds->spare;
4482 /* Set thresholds array to NULL if we don't have thresholds */
4491 /* Copy thresholds and find current threshold */
4492 new->current_threshold = -1;
4493 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4494 if (thresholds->primary->entries[i].eventfd == eventfd)
4497 new->entries[j] = thresholds->primary->entries[i];
4498 if (new->entries[j].threshold <= usage) {
4500 * new->current_threshold will not be used
4501 * until rcu_assign_pointer(), so it's safe to increment
4504 ++new->current_threshold;
4510 /* Swap primary and spare array */
4511 thresholds->spare = thresholds->primary;
4512 /* If all events are unregistered, free the spare array */
4514 kfree(thresholds->spare);
4515 thresholds->spare = NULL;
4518 rcu_assign_pointer(thresholds->primary, new);
4520 /* To be sure that nobody uses thresholds */
4523 mutex_unlock(&memcg->thresholds_lock);
4526 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4527 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4529 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4530 struct mem_cgroup_eventfd_list *event;
4531 int type = MEMFILE_TYPE(cft->private);
4533 BUG_ON(type != _OOM_TYPE);
4534 event = kmalloc(sizeof(*event), GFP_KERNEL);
4538 spin_lock(&memcg_oom_lock);
4540 event->eventfd = eventfd;
4541 list_add(&event->list, &memcg->oom_notify);
4543 /* already in OOM ? */
4544 if (atomic_read(&memcg->under_oom))
4545 eventfd_signal(eventfd, 1);
4546 spin_unlock(&memcg_oom_lock);
4551 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4552 struct cftype *cft, struct eventfd_ctx *eventfd)
4554 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4555 struct mem_cgroup_eventfd_list *ev, *tmp;
4556 int type = MEMFILE_TYPE(cft->private);
4558 BUG_ON(type != _OOM_TYPE);
4560 spin_lock(&memcg_oom_lock);
4562 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4563 if (ev->eventfd == eventfd) {
4564 list_del(&ev->list);
4569 spin_unlock(&memcg_oom_lock);
4572 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4573 struct cftype *cft, struct cgroup_map_cb *cb)
4575 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4577 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
4579 if (atomic_read(&memcg->under_oom))
4580 cb->fill(cb, "under_oom", 1);
4582 cb->fill(cb, "under_oom", 0);
4586 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4587 struct cftype *cft, u64 val)
4589 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4590 struct mem_cgroup *parent;
4592 /* cannot set to root cgroup and only 0 and 1 are allowed */
4593 if (!cgrp->parent || !((val == 0) || (val == 1)))
4596 parent = mem_cgroup_from_cont(cgrp->parent);
4599 /* oom-kill-disable is a flag for subhierarchy. */
4600 if ((parent->use_hierarchy) ||
4601 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4605 memcg->oom_kill_disable = val;
4607 memcg_oom_recover(memcg);
4613 static const struct file_operations mem_control_numa_stat_file_operations = {
4615 .llseek = seq_lseek,
4616 .release = single_release,
4619 static int mem_control_numa_stat_open(struct inode *unused, struct file *file)
4621 struct cgroup *cont = file->f_dentry->d_parent->d_fsdata;
4623 file->f_op = &mem_control_numa_stat_file_operations;
4624 return single_open(file, mem_control_numa_stat_show, cont);
4626 #endif /* CONFIG_NUMA */
4628 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_KMEM
4629 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
4631 return mem_cgroup_sockets_init(memcg, ss);
4634 static void kmem_cgroup_destroy(struct mem_cgroup *memcg)
4636 mem_cgroup_sockets_destroy(memcg);
4639 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
4644 static void kmem_cgroup_destroy(struct mem_cgroup *memcg)
4649 static struct cftype mem_cgroup_files[] = {
4651 .name = "usage_in_bytes",
4652 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4653 .read = mem_cgroup_read,
4654 .register_event = mem_cgroup_usage_register_event,
4655 .unregister_event = mem_cgroup_usage_unregister_event,
4658 .name = "max_usage_in_bytes",
4659 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4660 .trigger = mem_cgroup_reset,
4661 .read = mem_cgroup_read,
4664 .name = "limit_in_bytes",
4665 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4666 .write_string = mem_cgroup_write,
4667 .read = mem_cgroup_read,
4670 .name = "soft_limit_in_bytes",
4671 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4672 .write_string = mem_cgroup_write,
4673 .read = mem_cgroup_read,
4677 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4678 .trigger = mem_cgroup_reset,
4679 .read = mem_cgroup_read,
4683 .read_map = mem_control_stat_show,
4686 .name = "force_empty",
4687 .trigger = mem_cgroup_force_empty_write,
4690 .name = "use_hierarchy",
4691 .write_u64 = mem_cgroup_hierarchy_write,
4692 .read_u64 = mem_cgroup_hierarchy_read,
4695 .name = "swappiness",
4696 .read_u64 = mem_cgroup_swappiness_read,
4697 .write_u64 = mem_cgroup_swappiness_write,
4700 .name = "move_charge_at_immigrate",
4701 .read_u64 = mem_cgroup_move_charge_read,
4702 .write_u64 = mem_cgroup_move_charge_write,
4705 .name = "oom_control",
4706 .read_map = mem_cgroup_oom_control_read,
4707 .write_u64 = mem_cgroup_oom_control_write,
4708 .register_event = mem_cgroup_oom_register_event,
4709 .unregister_event = mem_cgroup_oom_unregister_event,
4710 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4714 .name = "numa_stat",
4715 .open = mem_control_numa_stat_open,
4719 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4721 .name = "memsw.usage_in_bytes",
4722 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4723 .read = mem_cgroup_read,
4724 .register_event = mem_cgroup_usage_register_event,
4725 .unregister_event = mem_cgroup_usage_unregister_event,
4728 .name = "memsw.max_usage_in_bytes",
4729 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4730 .trigger = mem_cgroup_reset,
4731 .read = mem_cgroup_read,
4734 .name = "memsw.limit_in_bytes",
4735 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4736 .write_string = mem_cgroup_write,
4737 .read = mem_cgroup_read,
4740 .name = "memsw.failcnt",
4741 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4742 .trigger = mem_cgroup_reset,
4743 .read = mem_cgroup_read,
4746 { }, /* terminate */
4749 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4751 struct mem_cgroup_per_node *pn;
4752 struct mem_cgroup_per_zone *mz;
4753 int zone, tmp = node;
4755 * This routine is called against possible nodes.
4756 * But it's BUG to call kmalloc() against offline node.
4758 * TODO: this routine can waste much memory for nodes which will
4759 * never be onlined. It's better to use memory hotplug callback
4762 if (!node_state(node, N_NORMAL_MEMORY))
4764 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4768 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4769 mz = &pn->zoneinfo[zone];
4770 lruvec_init(&mz->lruvec, &NODE_DATA(node)->node_zones[zone]);
4771 mz->usage_in_excess = 0;
4772 mz->on_tree = false;
4775 memcg->info.nodeinfo[node] = pn;
4779 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4781 kfree(memcg->info.nodeinfo[node]);
4784 static struct mem_cgroup *mem_cgroup_alloc(void)
4786 struct mem_cgroup *memcg;
4787 int size = sizeof(struct mem_cgroup);
4789 /* Can be very big if MAX_NUMNODES is very big */
4790 if (size < PAGE_SIZE)
4791 memcg = kzalloc(size, GFP_KERNEL);
4793 memcg = vzalloc(size);
4798 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4801 spin_lock_init(&memcg->pcp_counter_lock);
4805 if (size < PAGE_SIZE)
4813 * Helpers for freeing a vzalloc()ed mem_cgroup by RCU,
4814 * but in process context. The work_freeing structure is overlaid
4815 * on the rcu_freeing structure, which itself is overlaid on memsw.
4817 static void vfree_work(struct work_struct *work)
4819 struct mem_cgroup *memcg;
4821 memcg = container_of(work, struct mem_cgroup, work_freeing);
4824 static void vfree_rcu(struct rcu_head *rcu_head)
4826 struct mem_cgroup *memcg;
4828 memcg = container_of(rcu_head, struct mem_cgroup, rcu_freeing);
4829 INIT_WORK(&memcg->work_freeing, vfree_work);
4830 schedule_work(&memcg->work_freeing);
4834 * At destroying mem_cgroup, references from swap_cgroup can remain.
4835 * (scanning all at force_empty is too costly...)
4837 * Instead of clearing all references at force_empty, we remember
4838 * the number of reference from swap_cgroup and free mem_cgroup when
4839 * it goes down to 0.
4841 * Removal of cgroup itself succeeds regardless of refs from swap.
4844 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4848 mem_cgroup_remove_from_trees(memcg);
4849 free_css_id(&mem_cgroup_subsys, &memcg->css);
4852 free_mem_cgroup_per_zone_info(memcg, node);
4854 free_percpu(memcg->stat);
4855 if (sizeof(struct mem_cgroup) < PAGE_SIZE)
4856 kfree_rcu(memcg, rcu_freeing);
4858 call_rcu(&memcg->rcu_freeing, vfree_rcu);
4861 static void mem_cgroup_get(struct mem_cgroup *memcg)
4863 atomic_inc(&memcg->refcnt);
4866 static void __mem_cgroup_put(struct mem_cgroup *memcg, int count)
4868 if (atomic_sub_and_test(count, &memcg->refcnt)) {
4869 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4870 __mem_cgroup_free(memcg);
4872 mem_cgroup_put(parent);
4876 static void mem_cgroup_put(struct mem_cgroup *memcg)
4878 __mem_cgroup_put(memcg, 1);
4882 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4884 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4886 if (!memcg->res.parent)
4888 return mem_cgroup_from_res_counter(memcg->res.parent, res);
4890 EXPORT_SYMBOL(parent_mem_cgroup);
4892 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4893 static void __init enable_swap_cgroup(void)
4895 if (!mem_cgroup_disabled() && really_do_swap_account)
4896 do_swap_account = 1;
4899 static void __init enable_swap_cgroup(void)
4904 static int mem_cgroup_soft_limit_tree_init(void)
4906 struct mem_cgroup_tree_per_node *rtpn;
4907 struct mem_cgroup_tree_per_zone *rtpz;
4908 int tmp, node, zone;
4910 for_each_node(node) {
4912 if (!node_state(node, N_NORMAL_MEMORY))
4914 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
4918 soft_limit_tree.rb_tree_per_node[node] = rtpn;
4920 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4921 rtpz = &rtpn->rb_tree_per_zone[zone];
4922 rtpz->rb_root = RB_ROOT;
4923 spin_lock_init(&rtpz->lock);
4929 for_each_node(node) {
4930 if (!soft_limit_tree.rb_tree_per_node[node])
4932 kfree(soft_limit_tree.rb_tree_per_node[node]);
4933 soft_limit_tree.rb_tree_per_node[node] = NULL;
4939 static struct cgroup_subsys_state * __ref
4940 mem_cgroup_create(struct cgroup *cont)
4942 struct mem_cgroup *memcg, *parent;
4943 long error = -ENOMEM;
4946 memcg = mem_cgroup_alloc();
4948 return ERR_PTR(error);
4951 if (alloc_mem_cgroup_per_zone_info(memcg, node))
4955 if (cont->parent == NULL) {
4957 enable_swap_cgroup();
4959 if (mem_cgroup_soft_limit_tree_init())
4961 root_mem_cgroup = memcg;
4962 for_each_possible_cpu(cpu) {
4963 struct memcg_stock_pcp *stock =
4964 &per_cpu(memcg_stock, cpu);
4965 INIT_WORK(&stock->work, drain_local_stock);
4967 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
4969 parent = mem_cgroup_from_cont(cont->parent);
4970 memcg->use_hierarchy = parent->use_hierarchy;
4971 memcg->oom_kill_disable = parent->oom_kill_disable;
4974 if (parent && parent->use_hierarchy) {
4975 res_counter_init(&memcg->res, &parent->res);
4976 res_counter_init(&memcg->memsw, &parent->memsw);
4978 * We increment refcnt of the parent to ensure that we can
4979 * safely access it on res_counter_charge/uncharge.
4980 * This refcnt will be decremented when freeing this
4981 * mem_cgroup(see mem_cgroup_put).
4983 mem_cgroup_get(parent);
4985 res_counter_init(&memcg->res, NULL);
4986 res_counter_init(&memcg->memsw, NULL);
4988 memcg->last_scanned_node = MAX_NUMNODES;
4989 INIT_LIST_HEAD(&memcg->oom_notify);
4992 memcg->swappiness = mem_cgroup_swappiness(parent);
4993 atomic_set(&memcg->refcnt, 1);
4994 memcg->move_charge_at_immigrate = 0;
4995 mutex_init(&memcg->thresholds_lock);
4996 spin_lock_init(&memcg->move_lock);
4998 error = memcg_init_kmem(memcg, &mem_cgroup_subsys);
5001 * We call put now because our (and parent's) refcnts
5002 * are already in place. mem_cgroup_put() will internally
5003 * call __mem_cgroup_free, so return directly
5005 mem_cgroup_put(memcg);
5006 return ERR_PTR(error);
5010 __mem_cgroup_free(memcg);
5011 return ERR_PTR(error);
5014 static int mem_cgroup_pre_destroy(struct cgroup *cont)
5016 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5018 return mem_cgroup_force_empty(memcg, false);
5021 static void mem_cgroup_destroy(struct cgroup *cont)
5023 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5025 kmem_cgroup_destroy(memcg);
5027 mem_cgroup_put(memcg);
5031 /* Handlers for move charge at task migration. */
5032 #define PRECHARGE_COUNT_AT_ONCE 256
5033 static int mem_cgroup_do_precharge(unsigned long count)
5036 int batch_count = PRECHARGE_COUNT_AT_ONCE;
5037 struct mem_cgroup *memcg = mc.to;
5039 if (mem_cgroup_is_root(memcg)) {
5040 mc.precharge += count;
5041 /* we don't need css_get for root */
5044 /* try to charge at once */
5046 struct res_counter *dummy;
5048 * "memcg" cannot be under rmdir() because we've already checked
5049 * by cgroup_lock_live_cgroup() that it is not removed and we
5050 * are still under the same cgroup_mutex. So we can postpone
5053 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
5055 if (do_swap_account && res_counter_charge(&memcg->memsw,
5056 PAGE_SIZE * count, &dummy)) {
5057 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
5060 mc.precharge += count;
5064 /* fall back to one by one charge */
5066 if (signal_pending(current)) {
5070 if (!batch_count--) {
5071 batch_count = PRECHARGE_COUNT_AT_ONCE;
5074 ret = __mem_cgroup_try_charge(NULL,
5075 GFP_KERNEL, 1, &memcg, false);
5077 /* mem_cgroup_clear_mc() will do uncharge later */
5085 * get_mctgt_type - get target type of moving charge
5086 * @vma: the vma the pte to be checked belongs
5087 * @addr: the address corresponding to the pte to be checked
5088 * @ptent: the pte to be checked
5089 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5092 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5093 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5094 * move charge. if @target is not NULL, the page is stored in target->page
5095 * with extra refcnt got(Callers should handle it).
5096 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5097 * target for charge migration. if @target is not NULL, the entry is stored
5100 * Called with pte lock held.
5107 enum mc_target_type {
5113 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5114 unsigned long addr, pte_t ptent)
5116 struct page *page = vm_normal_page(vma, addr, ptent);
5118 if (!page || !page_mapped(page))
5120 if (PageAnon(page)) {
5121 /* we don't move shared anon */
5124 } else if (!move_file())
5125 /* we ignore mapcount for file pages */
5127 if (!get_page_unless_zero(page))
5134 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5135 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5137 struct page *page = NULL;
5138 swp_entry_t ent = pte_to_swp_entry(ptent);
5140 if (!move_anon() || non_swap_entry(ent))
5143 * Because lookup_swap_cache() updates some statistics counter,
5144 * we call find_get_page() with swapper_space directly.
5146 page = find_get_page(&swapper_space, ent.val);
5147 if (do_swap_account)
5148 entry->val = ent.val;
5153 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5154 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5160 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5161 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5163 struct page *page = NULL;
5164 struct address_space *mapping;
5167 if (!vma->vm_file) /* anonymous vma */
5172 mapping = vma->vm_file->f_mapping;
5173 if (pte_none(ptent))
5174 pgoff = linear_page_index(vma, addr);
5175 else /* pte_file(ptent) is true */
5176 pgoff = pte_to_pgoff(ptent);
5178 /* page is moved even if it's not RSS of this task(page-faulted). */
5179 page = find_get_page(mapping, pgoff);
5182 /* shmem/tmpfs may report page out on swap: account for that too. */
5183 if (radix_tree_exceptional_entry(page)) {
5184 swp_entry_t swap = radix_to_swp_entry(page);
5185 if (do_swap_account)
5187 page = find_get_page(&swapper_space, swap.val);
5193 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5194 unsigned long addr, pte_t ptent, union mc_target *target)
5196 struct page *page = NULL;
5197 struct page_cgroup *pc;
5198 enum mc_target_type ret = MC_TARGET_NONE;
5199 swp_entry_t ent = { .val = 0 };
5201 if (pte_present(ptent))
5202 page = mc_handle_present_pte(vma, addr, ptent);
5203 else if (is_swap_pte(ptent))
5204 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5205 else if (pte_none(ptent) || pte_file(ptent))
5206 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5208 if (!page && !ent.val)
5211 pc = lookup_page_cgroup(page);
5213 * Do only loose check w/o page_cgroup lock.
5214 * mem_cgroup_move_account() checks the pc is valid or not under
5217 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5218 ret = MC_TARGET_PAGE;
5220 target->page = page;
5222 if (!ret || !target)
5225 /* There is a swap entry and a page doesn't exist or isn't charged */
5226 if (ent.val && !ret &&
5227 css_id(&mc.from->css) == lookup_swap_cgroup_id(ent)) {
5228 ret = MC_TARGET_SWAP;
5235 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5237 * We don't consider swapping or file mapped pages because THP does not
5238 * support them for now.
5239 * Caller should make sure that pmd_trans_huge(pmd) is true.
5241 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5242 unsigned long addr, pmd_t pmd, union mc_target *target)
5244 struct page *page = NULL;
5245 struct page_cgroup *pc;
5246 enum mc_target_type ret = MC_TARGET_NONE;
5248 page = pmd_page(pmd);
5249 VM_BUG_ON(!page || !PageHead(page));
5252 pc = lookup_page_cgroup(page);
5253 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5254 ret = MC_TARGET_PAGE;
5257 target->page = page;
5263 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5264 unsigned long addr, pmd_t pmd, union mc_target *target)
5266 return MC_TARGET_NONE;
5270 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5271 unsigned long addr, unsigned long end,
5272 struct mm_walk *walk)
5274 struct vm_area_struct *vma = walk->private;
5278 if (pmd_trans_huge_lock(pmd, vma) == 1) {
5279 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5280 mc.precharge += HPAGE_PMD_NR;
5281 spin_unlock(&vma->vm_mm->page_table_lock);
5285 if (pmd_trans_unstable(pmd))
5287 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5288 for (; addr != end; pte++, addr += PAGE_SIZE)
5289 if (get_mctgt_type(vma, addr, *pte, NULL))
5290 mc.precharge++; /* increment precharge temporarily */
5291 pte_unmap_unlock(pte - 1, ptl);
5297 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5299 unsigned long precharge;
5300 struct vm_area_struct *vma;
5302 down_read(&mm->mmap_sem);
5303 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5304 struct mm_walk mem_cgroup_count_precharge_walk = {
5305 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5309 if (is_vm_hugetlb_page(vma))
5311 walk_page_range(vma->vm_start, vma->vm_end,
5312 &mem_cgroup_count_precharge_walk);
5314 up_read(&mm->mmap_sem);
5316 precharge = mc.precharge;
5322 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5324 unsigned long precharge = mem_cgroup_count_precharge(mm);
5326 VM_BUG_ON(mc.moving_task);
5327 mc.moving_task = current;
5328 return mem_cgroup_do_precharge(precharge);
5331 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5332 static void __mem_cgroup_clear_mc(void)
5334 struct mem_cgroup *from = mc.from;
5335 struct mem_cgroup *to = mc.to;
5337 /* we must uncharge all the leftover precharges from mc.to */
5339 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
5343 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5344 * we must uncharge here.
5346 if (mc.moved_charge) {
5347 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
5348 mc.moved_charge = 0;
5350 /* we must fixup refcnts and charges */
5351 if (mc.moved_swap) {
5352 /* uncharge swap account from the old cgroup */
5353 if (!mem_cgroup_is_root(mc.from))
5354 res_counter_uncharge(&mc.from->memsw,
5355 PAGE_SIZE * mc.moved_swap);
5356 __mem_cgroup_put(mc.from, mc.moved_swap);
5358 if (!mem_cgroup_is_root(mc.to)) {
5360 * we charged both to->res and to->memsw, so we should
5363 res_counter_uncharge(&mc.to->res,
5364 PAGE_SIZE * mc.moved_swap);
5366 /* we've already done mem_cgroup_get(mc.to) */
5369 memcg_oom_recover(from);
5370 memcg_oom_recover(to);
5371 wake_up_all(&mc.waitq);
5374 static void mem_cgroup_clear_mc(void)
5376 struct mem_cgroup *from = mc.from;
5379 * we must clear moving_task before waking up waiters at the end of
5382 mc.moving_task = NULL;
5383 __mem_cgroup_clear_mc();
5384 spin_lock(&mc.lock);
5387 spin_unlock(&mc.lock);
5388 mem_cgroup_end_move(from);
5391 static int mem_cgroup_can_attach(struct cgroup *cgroup,
5392 struct cgroup_taskset *tset)
5394 struct task_struct *p = cgroup_taskset_first(tset);
5396 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup);
5398 if (memcg->move_charge_at_immigrate) {
5399 struct mm_struct *mm;
5400 struct mem_cgroup *from = mem_cgroup_from_task(p);
5402 VM_BUG_ON(from == memcg);
5404 mm = get_task_mm(p);
5407 /* We move charges only when we move a owner of the mm */
5408 if (mm->owner == p) {
5411 VM_BUG_ON(mc.precharge);
5412 VM_BUG_ON(mc.moved_charge);
5413 VM_BUG_ON(mc.moved_swap);
5414 mem_cgroup_start_move(from);
5415 spin_lock(&mc.lock);
5418 spin_unlock(&mc.lock);
5419 /* We set mc.moving_task later */
5421 ret = mem_cgroup_precharge_mc(mm);
5423 mem_cgroup_clear_mc();
5430 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
5431 struct cgroup_taskset *tset)
5433 mem_cgroup_clear_mc();
5436 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5437 unsigned long addr, unsigned long end,
5438 struct mm_walk *walk)
5441 struct vm_area_struct *vma = walk->private;
5444 enum mc_target_type target_type;
5445 union mc_target target;
5447 struct page_cgroup *pc;
5450 * We don't take compound_lock() here but no race with splitting thp
5452 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
5453 * under splitting, which means there's no concurrent thp split,
5454 * - if another thread runs into split_huge_page() just after we
5455 * entered this if-block, the thread must wait for page table lock
5456 * to be unlocked in __split_huge_page_splitting(), where the main
5457 * part of thp split is not executed yet.
5459 if (pmd_trans_huge_lock(pmd, vma) == 1) {
5460 if (mc.precharge < HPAGE_PMD_NR) {
5461 spin_unlock(&vma->vm_mm->page_table_lock);
5464 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5465 if (target_type == MC_TARGET_PAGE) {
5467 if (!isolate_lru_page(page)) {
5468 pc = lookup_page_cgroup(page);
5469 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
5470 pc, mc.from, mc.to)) {
5471 mc.precharge -= HPAGE_PMD_NR;
5472 mc.moved_charge += HPAGE_PMD_NR;
5474 putback_lru_page(page);
5478 spin_unlock(&vma->vm_mm->page_table_lock);
5482 if (pmd_trans_unstable(pmd))
5485 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5486 for (; addr != end; addr += PAGE_SIZE) {
5487 pte_t ptent = *(pte++);
5493 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5494 case MC_TARGET_PAGE:
5496 if (isolate_lru_page(page))
5498 pc = lookup_page_cgroup(page);
5499 if (!mem_cgroup_move_account(page, 1, pc,
5502 /* we uncharge from mc.from later. */
5505 putback_lru_page(page);
5506 put: /* get_mctgt_type() gets the page */
5509 case MC_TARGET_SWAP:
5511 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5513 /* we fixup refcnts and charges later. */
5521 pte_unmap_unlock(pte - 1, ptl);
5526 * We have consumed all precharges we got in can_attach().
5527 * We try charge one by one, but don't do any additional
5528 * charges to mc.to if we have failed in charge once in attach()
5531 ret = mem_cgroup_do_precharge(1);
5539 static void mem_cgroup_move_charge(struct mm_struct *mm)
5541 struct vm_area_struct *vma;
5543 lru_add_drain_all();
5545 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5547 * Someone who are holding the mmap_sem might be waiting in
5548 * waitq. So we cancel all extra charges, wake up all waiters,
5549 * and retry. Because we cancel precharges, we might not be able
5550 * to move enough charges, but moving charge is a best-effort
5551 * feature anyway, so it wouldn't be a big problem.
5553 __mem_cgroup_clear_mc();
5557 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5559 struct mm_walk mem_cgroup_move_charge_walk = {
5560 .pmd_entry = mem_cgroup_move_charge_pte_range,
5564 if (is_vm_hugetlb_page(vma))
5566 ret = walk_page_range(vma->vm_start, vma->vm_end,
5567 &mem_cgroup_move_charge_walk);
5570 * means we have consumed all precharges and failed in
5571 * doing additional charge. Just abandon here.
5575 up_read(&mm->mmap_sem);
5578 static void mem_cgroup_move_task(struct cgroup *cont,
5579 struct cgroup_taskset *tset)
5581 struct task_struct *p = cgroup_taskset_first(tset);
5582 struct mm_struct *mm = get_task_mm(p);
5586 mem_cgroup_move_charge(mm);
5590 mem_cgroup_clear_mc();
5592 #else /* !CONFIG_MMU */
5593 static int mem_cgroup_can_attach(struct cgroup *cgroup,
5594 struct cgroup_taskset *tset)
5598 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
5599 struct cgroup_taskset *tset)
5602 static void mem_cgroup_move_task(struct cgroup *cont,
5603 struct cgroup_taskset *tset)
5608 struct cgroup_subsys mem_cgroup_subsys = {
5610 .subsys_id = mem_cgroup_subsys_id,
5611 .create = mem_cgroup_create,
5612 .pre_destroy = mem_cgroup_pre_destroy,
5613 .destroy = mem_cgroup_destroy,
5614 .can_attach = mem_cgroup_can_attach,
5615 .cancel_attach = mem_cgroup_cancel_attach,
5616 .attach = mem_cgroup_move_task,
5617 .base_cftypes = mem_cgroup_files,
5620 .__DEPRECATED_clear_css_refs = true,
5623 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5624 static int __init enable_swap_account(char *s)
5626 /* consider enabled if no parameter or 1 is given */
5627 if (!strcmp(s, "1"))
5628 really_do_swap_account = 1;
5629 else if (!strcmp(s, "0"))
5630 really_do_swap_account = 0;
5633 __setup("swapaccount=", enable_swap_account);