1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * This program is free software; you can redistribute it and/or modify
14 * it under the terms of the GNU General Public License as published by
15 * the Free Software Foundation; either version 2 of the License, or
16 * (at your option) any later version.
18 * This program is distributed in the hope that it will be useful,
19 * but WITHOUT ANY WARRANTY; without even the implied warranty of
20 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
21 * GNU General Public License for more details.
24 #include <linux/res_counter.h>
25 #include <linux/memcontrol.h>
26 #include <linux/cgroup.h>
28 #include <linux/hugetlb.h>
29 #include <linux/pagemap.h>
30 #include <linux/smp.h>
31 #include <linux/page-flags.h>
32 #include <linux/backing-dev.h>
33 #include <linux/bit_spinlock.h>
34 #include <linux/rcupdate.h>
35 #include <linux/limits.h>
36 #include <linux/export.h>
37 #include <linux/mutex.h>
38 #include <linux/rbtree.h>
39 #include <linux/slab.h>
40 #include <linux/swap.h>
41 #include <linux/swapops.h>
42 #include <linux/spinlock.h>
43 #include <linux/eventfd.h>
44 #include <linux/sort.h>
46 #include <linux/seq_file.h>
47 #include <linux/vmalloc.h>
48 #include <linux/mm_inline.h>
49 #include <linux/page_cgroup.h>
50 #include <linux/cpu.h>
51 #include <linux/oom.h>
54 #include <net/tcp_memcontrol.h>
56 #include <asm/uaccess.h>
58 #include <trace/events/vmscan.h>
60 struct cgroup_subsys mem_cgroup_subsys __read_mostly;
61 #define MEM_CGROUP_RECLAIM_RETRIES 5
62 struct mem_cgroup *root_mem_cgroup __read_mostly;
64 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
65 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
66 int do_swap_account __read_mostly;
68 /* for remember boot option*/
69 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP_ENABLED
70 static int really_do_swap_account __initdata = 1;
72 static int really_do_swap_account __initdata = 0;
76 #define do_swap_account (0)
81 * Statistics for memory cgroup.
83 enum mem_cgroup_stat_index {
85 * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
87 MEM_CGROUP_STAT_CACHE, /* # of pages charged as cache */
88 MEM_CGROUP_STAT_RSS, /* # of pages charged as anon rss */
89 MEM_CGROUP_STAT_FILE_MAPPED, /* # of pages charged as file rss */
90 MEM_CGROUP_STAT_SWAPOUT, /* # of pages, swapped out */
91 MEM_CGROUP_STAT_DATA, /* end of data requires synchronization */
92 MEM_CGROUP_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 zone_reclaim_stat reclaim_stat;
142 struct rb_node tree_node; /* RB tree node */
143 unsigned long long usage_in_excess;/* Set to the value by which */
144 /* the soft limit is exceeded*/
146 struct mem_cgroup *memcg; /* Back pointer, we cannot */
147 /* use container_of */
150 struct mem_cgroup_per_node {
151 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
154 struct mem_cgroup_lru_info {
155 struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
159 * Cgroups above their limits are maintained in a RB-Tree, independent of
160 * their hierarchy representation
163 struct mem_cgroup_tree_per_zone {
164 struct rb_root rb_root;
168 struct mem_cgroup_tree_per_node {
169 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
172 struct mem_cgroup_tree {
173 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
176 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
178 struct mem_cgroup_threshold {
179 struct eventfd_ctx *eventfd;
184 struct mem_cgroup_threshold_ary {
185 /* An array index points to threshold just below usage. */
186 int current_threshold;
187 /* Size of entries[] */
189 /* Array of thresholds */
190 struct mem_cgroup_threshold entries[0];
193 struct mem_cgroup_thresholds {
194 /* Primary thresholds array */
195 struct mem_cgroup_threshold_ary *primary;
197 * Spare threshold array.
198 * This is needed to make mem_cgroup_unregister_event() "never fail".
199 * It must be able to store at least primary->size - 1 entries.
201 struct mem_cgroup_threshold_ary *spare;
205 struct mem_cgroup_eventfd_list {
206 struct list_head list;
207 struct eventfd_ctx *eventfd;
210 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
211 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
214 * The memory controller data structure. The memory controller controls both
215 * page cache and RSS per cgroup. We would eventually like to provide
216 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
217 * to help the administrator determine what knobs to tune.
219 * TODO: Add a water mark for the memory controller. Reclaim will begin when
220 * we hit the water mark. May be even add a low water mark, such that
221 * no reclaim occurs from a cgroup at it's low water mark, this is
222 * a feature that will be implemented much later in the future.
225 struct cgroup_subsys_state css;
227 * the counter to account for memory usage
229 struct res_counter res;
233 * the counter to account for mem+swap usage.
235 struct res_counter memsw;
238 * rcu_freeing is used only when freeing struct mem_cgroup,
239 * so put it into a union to avoid wasting more memory.
240 * It must be disjoint from the css field. It could be
241 * in a union with the res field, but res plays a much
242 * larger part in mem_cgroup life than memsw, and might
243 * be of interest, even at time of free, when debugging.
244 * So share rcu_head with the less interesting memsw.
246 struct rcu_head rcu_freeing;
248 * But when using vfree(), that cannot be done at
249 * interrupt time, so we must then queue the work.
251 struct work_struct work_freeing;
255 * Per cgroup active and inactive list, similar to the
256 * per zone LRU lists.
258 struct mem_cgroup_lru_info info;
259 int last_scanned_node;
261 nodemask_t scan_nodes;
262 atomic_t numainfo_events;
263 atomic_t numainfo_updating;
266 * Should the accounting and control be hierarchical, per subtree?
276 /* OOM-Killer disable */
277 int oom_kill_disable;
279 /* set when res.limit == memsw.limit */
280 bool memsw_is_minimum;
282 /* protect arrays of thresholds */
283 struct mutex thresholds_lock;
285 /* thresholds for memory usage. RCU-protected */
286 struct mem_cgroup_thresholds thresholds;
288 /* thresholds for mem+swap usage. RCU-protected */
289 struct mem_cgroup_thresholds memsw_thresholds;
291 /* For oom notifier event fd */
292 struct list_head oom_notify;
295 * Should we move charges of a task when a task is moved into this
296 * mem_cgroup ? And what type of charges should we move ?
298 unsigned long move_charge_at_immigrate;
300 * set > 0 if pages under this cgroup are moving to other cgroup.
302 atomic_t moving_account;
303 /* taken only while moving_account > 0 */
304 spinlock_t move_lock;
308 struct mem_cgroup_stat_cpu *stat;
310 * used when a cpu is offlined or other synchronizations
311 * See mem_cgroup_read_stat().
313 struct mem_cgroup_stat_cpu nocpu_base;
314 spinlock_t pcp_counter_lock;
317 struct tcp_memcontrol tcp_mem;
321 /* Stuffs for move charges at task migration. */
323 * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
324 * left-shifted bitmap of these types.
327 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
328 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
332 /* "mc" and its members are protected by cgroup_mutex */
333 static struct move_charge_struct {
334 spinlock_t lock; /* for from, to */
335 struct mem_cgroup *from;
336 struct mem_cgroup *to;
337 unsigned long precharge;
338 unsigned long moved_charge;
339 unsigned long moved_swap;
340 struct task_struct *moving_task; /* a task moving charges */
341 wait_queue_head_t waitq; /* a waitq for other context */
343 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
344 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
347 static bool move_anon(void)
349 return test_bit(MOVE_CHARGE_TYPE_ANON,
350 &mc.to->move_charge_at_immigrate);
353 static bool move_file(void)
355 return test_bit(MOVE_CHARGE_TYPE_FILE,
356 &mc.to->move_charge_at_immigrate);
360 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
361 * limit reclaim to prevent infinite loops, if they ever occur.
363 #define MEM_CGROUP_MAX_RECLAIM_LOOPS (100)
364 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2)
367 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
368 MEM_CGROUP_CHARGE_TYPE_MAPPED,
369 MEM_CGROUP_CHARGE_TYPE_SHMEM, /* used by page migration of shmem */
370 MEM_CGROUP_CHARGE_TYPE_FORCE, /* used by force_empty */
371 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
372 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
376 /* for encoding cft->private value on file */
379 #define _OOM_TYPE (2)
380 #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
381 #define MEMFILE_TYPE(val) (((val) >> 16) & 0xffff)
382 #define MEMFILE_ATTR(val) ((val) & 0xffff)
383 /* Used for OOM nofiier */
384 #define OOM_CONTROL (0)
387 * Reclaim flags for mem_cgroup_hierarchical_reclaim
389 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
390 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
391 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
392 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
394 static void mem_cgroup_get(struct mem_cgroup *memcg);
395 static void mem_cgroup_put(struct mem_cgroup *memcg);
397 /* Writing them here to avoid exposing memcg's inner layout */
398 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_KMEM
399 #include <net/sock.h>
402 static bool mem_cgroup_is_root(struct mem_cgroup *memcg);
403 void sock_update_memcg(struct sock *sk)
405 if (mem_cgroup_sockets_enabled) {
406 struct mem_cgroup *memcg;
408 BUG_ON(!sk->sk_prot->proto_cgroup);
410 /* Socket cloning can throw us here with sk_cgrp already
411 * filled. It won't however, necessarily happen from
412 * process context. So the test for root memcg given
413 * the current task's memcg won't help us in this case.
415 * Respecting the original socket's memcg is a better
416 * decision in this case.
419 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
420 mem_cgroup_get(sk->sk_cgrp->memcg);
425 memcg = mem_cgroup_from_task(current);
426 if (!mem_cgroup_is_root(memcg)) {
427 mem_cgroup_get(memcg);
428 sk->sk_cgrp = sk->sk_prot->proto_cgroup(memcg);
433 EXPORT_SYMBOL(sock_update_memcg);
435 void sock_release_memcg(struct sock *sk)
437 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
438 struct mem_cgroup *memcg;
439 WARN_ON(!sk->sk_cgrp->memcg);
440 memcg = sk->sk_cgrp->memcg;
441 mem_cgroup_put(memcg);
446 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
448 if (!memcg || mem_cgroup_is_root(memcg))
451 return &memcg->tcp_mem.cg_proto;
453 EXPORT_SYMBOL(tcp_proto_cgroup);
454 #endif /* CONFIG_INET */
455 #endif /* CONFIG_CGROUP_MEM_RES_CTLR_KMEM */
457 static void drain_all_stock_async(struct mem_cgroup *memcg);
459 static struct mem_cgroup_per_zone *
460 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
462 return &memcg->info.nodeinfo[nid]->zoneinfo[zid];
465 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
470 static struct mem_cgroup_per_zone *
471 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
473 int nid = page_to_nid(page);
474 int zid = page_zonenum(page);
476 return mem_cgroup_zoneinfo(memcg, nid, zid);
479 static struct mem_cgroup_tree_per_zone *
480 soft_limit_tree_node_zone(int nid, int zid)
482 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
485 static struct mem_cgroup_tree_per_zone *
486 soft_limit_tree_from_page(struct page *page)
488 int nid = page_to_nid(page);
489 int zid = page_zonenum(page);
491 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
495 __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
496 struct mem_cgroup_per_zone *mz,
497 struct mem_cgroup_tree_per_zone *mctz,
498 unsigned long long new_usage_in_excess)
500 struct rb_node **p = &mctz->rb_root.rb_node;
501 struct rb_node *parent = NULL;
502 struct mem_cgroup_per_zone *mz_node;
507 mz->usage_in_excess = new_usage_in_excess;
508 if (!mz->usage_in_excess)
512 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
514 if (mz->usage_in_excess < mz_node->usage_in_excess)
517 * We can't avoid mem cgroups that are over their soft
518 * limit by the same amount
520 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
523 rb_link_node(&mz->tree_node, parent, p);
524 rb_insert_color(&mz->tree_node, &mctz->rb_root);
529 __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
530 struct mem_cgroup_per_zone *mz,
531 struct mem_cgroup_tree_per_zone *mctz)
535 rb_erase(&mz->tree_node, &mctz->rb_root);
540 mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
541 struct mem_cgroup_per_zone *mz,
542 struct mem_cgroup_tree_per_zone *mctz)
544 spin_lock(&mctz->lock);
545 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
546 spin_unlock(&mctz->lock);
550 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
552 unsigned long long excess;
553 struct mem_cgroup_per_zone *mz;
554 struct mem_cgroup_tree_per_zone *mctz;
555 int nid = page_to_nid(page);
556 int zid = page_zonenum(page);
557 mctz = soft_limit_tree_from_page(page);
560 * Necessary to update all ancestors when hierarchy is used.
561 * because their event counter is not touched.
563 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
564 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
565 excess = res_counter_soft_limit_excess(&memcg->res);
567 * We have to update the tree if mz is on RB-tree or
568 * mem is over its softlimit.
570 if (excess || mz->on_tree) {
571 spin_lock(&mctz->lock);
572 /* if on-tree, remove it */
574 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
576 * Insert again. mz->usage_in_excess will be updated.
577 * If excess is 0, no tree ops.
579 __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
580 spin_unlock(&mctz->lock);
585 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
588 struct mem_cgroup_per_zone *mz;
589 struct mem_cgroup_tree_per_zone *mctz;
591 for_each_node(node) {
592 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
593 mz = mem_cgroup_zoneinfo(memcg, node, zone);
594 mctz = soft_limit_tree_node_zone(node, zone);
595 mem_cgroup_remove_exceeded(memcg, mz, mctz);
600 static struct mem_cgroup_per_zone *
601 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
603 struct rb_node *rightmost = NULL;
604 struct mem_cgroup_per_zone *mz;
608 rightmost = rb_last(&mctz->rb_root);
610 goto done; /* Nothing to reclaim from */
612 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
614 * Remove the node now but someone else can add it back,
615 * we will to add it back at the end of reclaim to its correct
616 * position in the tree.
618 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
619 if (!res_counter_soft_limit_excess(&mz->memcg->res) ||
620 !css_tryget(&mz->memcg->css))
626 static struct mem_cgroup_per_zone *
627 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
629 struct mem_cgroup_per_zone *mz;
631 spin_lock(&mctz->lock);
632 mz = __mem_cgroup_largest_soft_limit_node(mctz);
633 spin_unlock(&mctz->lock);
638 * Implementation Note: reading percpu statistics for memcg.
640 * Both of vmstat[] and percpu_counter has threshold and do periodic
641 * synchronization to implement "quick" read. There are trade-off between
642 * reading cost and precision of value. Then, we may have a chance to implement
643 * a periodic synchronizion of counter in memcg's counter.
645 * But this _read() function is used for user interface now. The user accounts
646 * memory usage by memory cgroup and he _always_ requires exact value because
647 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
648 * have to visit all online cpus and make sum. So, for now, unnecessary
649 * synchronization is not implemented. (just implemented for cpu hotplug)
651 * If there are kernel internal actions which can make use of some not-exact
652 * value, and reading all cpu value can be performance bottleneck in some
653 * common workload, threashold and synchonization as vmstat[] should be
656 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
657 enum mem_cgroup_stat_index idx)
663 for_each_online_cpu(cpu)
664 val += per_cpu(memcg->stat->count[idx], cpu);
665 #ifdef CONFIG_HOTPLUG_CPU
666 spin_lock(&memcg->pcp_counter_lock);
667 val += memcg->nocpu_base.count[idx];
668 spin_unlock(&memcg->pcp_counter_lock);
674 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
677 int val = (charge) ? 1 : -1;
678 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAPOUT], val);
681 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
682 enum mem_cgroup_events_index idx)
684 unsigned long val = 0;
687 for_each_online_cpu(cpu)
688 val += per_cpu(memcg->stat->events[idx], cpu);
689 #ifdef CONFIG_HOTPLUG_CPU
690 spin_lock(&memcg->pcp_counter_lock);
691 val += memcg->nocpu_base.events[idx];
692 spin_unlock(&memcg->pcp_counter_lock);
697 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
698 bool anon, int nr_pages)
703 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
704 * counted as CACHE even if it's on ANON LRU.
707 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
710 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
713 /* pagein of a big page is an event. So, ignore page size */
715 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
717 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
718 nr_pages = -nr_pages; /* for event */
721 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT], nr_pages);
727 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
728 unsigned int lru_mask)
730 struct mem_cgroup_per_zone *mz;
732 unsigned long ret = 0;
734 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
737 if (BIT(lru) & lru_mask)
738 ret += mz->lru_size[lru];
744 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
745 int nid, unsigned int lru_mask)
750 for (zid = 0; zid < MAX_NR_ZONES; zid++)
751 total += mem_cgroup_zone_nr_lru_pages(memcg,
757 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
758 unsigned int lru_mask)
763 for_each_node_state(nid, N_HIGH_MEMORY)
764 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
768 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
769 enum mem_cgroup_events_target target)
771 unsigned long val, next;
773 val = __this_cpu_read(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT]);
774 next = __this_cpu_read(memcg->stat->targets[target]);
775 /* from time_after() in jiffies.h */
776 if ((long)next - (long)val < 0) {
778 case MEM_CGROUP_TARGET_THRESH:
779 next = val + THRESHOLDS_EVENTS_TARGET;
781 case MEM_CGROUP_TARGET_SOFTLIMIT:
782 next = val + SOFTLIMIT_EVENTS_TARGET;
784 case MEM_CGROUP_TARGET_NUMAINFO:
785 next = val + NUMAINFO_EVENTS_TARGET;
790 __this_cpu_write(memcg->stat->targets[target], next);
797 * Check events in order.
800 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
803 /* threshold event is triggered in finer grain than soft limit */
804 if (unlikely(mem_cgroup_event_ratelimit(memcg,
805 MEM_CGROUP_TARGET_THRESH))) {
807 bool do_numainfo __maybe_unused;
809 do_softlimit = mem_cgroup_event_ratelimit(memcg,
810 MEM_CGROUP_TARGET_SOFTLIMIT);
812 do_numainfo = mem_cgroup_event_ratelimit(memcg,
813 MEM_CGROUP_TARGET_NUMAINFO);
817 mem_cgroup_threshold(memcg);
818 if (unlikely(do_softlimit))
819 mem_cgroup_update_tree(memcg, page);
821 if (unlikely(do_numainfo))
822 atomic_inc(&memcg->numainfo_events);
828 struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
830 return container_of(cgroup_subsys_state(cont,
831 mem_cgroup_subsys_id), struct mem_cgroup,
835 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
838 * mm_update_next_owner() may clear mm->owner to NULL
839 * if it races with swapoff, page migration, etc.
840 * So this can be called with p == NULL.
845 return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
846 struct mem_cgroup, css);
849 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
851 struct mem_cgroup *memcg = NULL;
856 * Because we have no locks, mm->owner's may be being moved to other
857 * cgroup. We use css_tryget() here even if this looks
858 * pessimistic (rather than adding locks here).
862 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
863 if (unlikely(!memcg))
865 } while (!css_tryget(&memcg->css));
871 * mem_cgroup_iter - iterate over memory cgroup hierarchy
872 * @root: hierarchy root
873 * @prev: previously returned memcg, NULL on first invocation
874 * @reclaim: cookie for shared reclaim walks, NULL for full walks
876 * Returns references to children of the hierarchy below @root, or
877 * @root itself, or %NULL after a full round-trip.
879 * Caller must pass the return value in @prev on subsequent
880 * invocations for reference counting, or use mem_cgroup_iter_break()
881 * to cancel a hierarchy walk before the round-trip is complete.
883 * Reclaimers can specify a zone and a priority level in @reclaim to
884 * divide up the memcgs in the hierarchy among all concurrent
885 * reclaimers operating on the same zone and priority.
887 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
888 struct mem_cgroup *prev,
889 struct mem_cgroup_reclaim_cookie *reclaim)
891 struct mem_cgroup *memcg = NULL;
894 if (mem_cgroup_disabled())
898 root = root_mem_cgroup;
900 if (prev && !reclaim)
901 id = css_id(&prev->css);
903 if (prev && prev != root)
906 if (!root->use_hierarchy && root != root_mem_cgroup) {
913 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
914 struct cgroup_subsys_state *css;
917 int nid = zone_to_nid(reclaim->zone);
918 int zid = zone_idx(reclaim->zone);
919 struct mem_cgroup_per_zone *mz;
921 mz = mem_cgroup_zoneinfo(root, nid, zid);
922 iter = &mz->reclaim_iter[reclaim->priority];
923 if (prev && reclaim->generation != iter->generation)
929 css = css_get_next(&mem_cgroup_subsys, id + 1, &root->css, &id);
931 if (css == &root->css || css_tryget(css))
932 memcg = container_of(css,
933 struct mem_cgroup, css);
942 else if (!prev && memcg)
943 reclaim->generation = iter->generation;
953 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
954 * @root: hierarchy root
955 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
957 void mem_cgroup_iter_break(struct mem_cgroup *root,
958 struct mem_cgroup *prev)
961 root = root_mem_cgroup;
962 if (prev && prev != root)
967 * Iteration constructs for visiting all cgroups (under a tree). If
968 * loops are exited prematurely (break), mem_cgroup_iter_break() must
969 * be used for reference counting.
971 #define for_each_mem_cgroup_tree(iter, root) \
972 for (iter = mem_cgroup_iter(root, NULL, NULL); \
974 iter = mem_cgroup_iter(root, iter, NULL))
976 #define for_each_mem_cgroup(iter) \
977 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
979 iter = mem_cgroup_iter(NULL, iter, NULL))
981 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
983 return (memcg == root_mem_cgroup);
986 void mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
988 struct mem_cgroup *memcg;
994 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
995 if (unlikely(!memcg))
1000 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1003 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1011 EXPORT_SYMBOL(mem_cgroup_count_vm_event);
1014 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1015 * @zone: zone of the wanted lruvec
1016 * @mem: memcg of the wanted lruvec
1018 * Returns the lru list vector holding pages for the given @zone and
1019 * @mem. This can be the global zone lruvec, if the memory controller
1022 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1023 struct mem_cgroup *memcg)
1025 struct mem_cgroup_per_zone *mz;
1027 if (mem_cgroup_disabled())
1028 return &zone->lruvec;
1030 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1035 * Following LRU functions are allowed to be used without PCG_LOCK.
1036 * Operations are called by routine of global LRU independently from memcg.
1037 * What we have to take care of here is validness of pc->mem_cgroup.
1039 * Changes to pc->mem_cgroup happens when
1042 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1043 * It is added to LRU before charge.
1044 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1045 * When moving account, the page is not on LRU. It's isolated.
1049 * mem_cgroup_lru_add_list - account for adding an lru page and return lruvec
1050 * @zone: zone of the page
1054 * This function accounts for @page being added to @lru, and returns
1055 * the lruvec for the given @zone and the memcg @page is charged to.
1057 * The callsite is then responsible for physically linking the page to
1058 * the returned lruvec->lists[@lru].
1060 struct lruvec *mem_cgroup_lru_add_list(struct zone *zone, struct page *page,
1063 struct mem_cgroup_per_zone *mz;
1064 struct mem_cgroup *memcg;
1065 struct page_cgroup *pc;
1067 if (mem_cgroup_disabled())
1068 return &zone->lruvec;
1070 pc = lookup_page_cgroup(page);
1071 memcg = pc->mem_cgroup;
1074 * Surreptitiously switch any uncharged page to root:
1075 * an uncharged page off lru does nothing to secure
1076 * its former mem_cgroup from sudden removal.
1078 * Our caller holds lru_lock, and PageCgroupUsed is updated
1079 * under page_cgroup lock: between them, they make all uses
1080 * of pc->mem_cgroup safe.
1082 if (!PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1083 pc->mem_cgroup = memcg = root_mem_cgroup;
1085 mz = page_cgroup_zoneinfo(memcg, page);
1086 /* compound_order() is stabilized through lru_lock */
1087 mz->lru_size[lru] += 1 << compound_order(page);
1092 * mem_cgroup_lru_del_list - account for removing an lru page
1096 * This function accounts for @page being removed from @lru.
1098 * The callsite is then responsible for physically unlinking
1101 void mem_cgroup_lru_del_list(struct page *page, enum lru_list lru)
1103 struct mem_cgroup_per_zone *mz;
1104 struct mem_cgroup *memcg;
1105 struct page_cgroup *pc;
1107 if (mem_cgroup_disabled())
1110 pc = lookup_page_cgroup(page);
1111 memcg = pc->mem_cgroup;
1113 mz = page_cgroup_zoneinfo(memcg, page);
1114 /* huge page split is done under lru_lock. so, we have no races. */
1115 VM_BUG_ON(mz->lru_size[lru] < (1 << compound_order(page)));
1116 mz->lru_size[lru] -= 1 << compound_order(page);
1119 void mem_cgroup_lru_del(struct page *page)
1121 mem_cgroup_lru_del_list(page, page_lru(page));
1125 * mem_cgroup_lru_move_lists - account for moving a page between lrus
1126 * @zone: zone of the page
1128 * @from: current lru
1131 * This function accounts for @page being moved between the lrus @from
1132 * and @to, and returns the lruvec for the given @zone and the memcg
1133 * @page is charged to.
1135 * The callsite is then responsible for physically relinking
1136 * @page->lru to the returned lruvec->lists[@to].
1138 struct lruvec *mem_cgroup_lru_move_lists(struct zone *zone,
1143 /* XXX: Optimize this, especially for @from == @to */
1144 mem_cgroup_lru_del_list(page, from);
1145 return mem_cgroup_lru_add_list(zone, page, to);
1149 * Checks whether given mem is same or in the root_mem_cgroup's
1152 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1153 struct mem_cgroup *memcg)
1155 if (root_memcg == memcg)
1157 if (!root_memcg->use_hierarchy)
1159 return css_is_ancestor(&memcg->css, &root_memcg->css);
1162 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1163 struct mem_cgroup *memcg)
1168 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1173 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *memcg)
1176 struct mem_cgroup *curr = NULL;
1177 struct task_struct *p;
1179 p = find_lock_task_mm(task);
1181 curr = try_get_mem_cgroup_from_mm(p->mm);
1185 * All threads may have already detached their mm's, but the oom
1186 * killer still needs to detect if they have already been oom
1187 * killed to prevent needlessly killing additional tasks.
1190 curr = mem_cgroup_from_task(task);
1192 css_get(&curr->css);
1198 * We should check use_hierarchy of "memcg" not "curr". Because checking
1199 * use_hierarchy of "curr" here make this function true if hierarchy is
1200 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1201 * hierarchy(even if use_hierarchy is disabled in "memcg").
1203 ret = mem_cgroup_same_or_subtree(memcg, curr);
1204 css_put(&curr->css);
1208 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg, struct zone *zone)
1210 unsigned long inactive_ratio;
1211 int nid = zone_to_nid(zone);
1212 int zid = zone_idx(zone);
1213 unsigned long inactive;
1214 unsigned long active;
1217 inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1218 BIT(LRU_INACTIVE_ANON));
1219 active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1220 BIT(LRU_ACTIVE_ANON));
1222 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1224 inactive_ratio = int_sqrt(10 * gb);
1228 return inactive * inactive_ratio < active;
1231 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg, struct zone *zone)
1233 unsigned long active;
1234 unsigned long inactive;
1235 int zid = zone_idx(zone);
1236 int nid = zone_to_nid(zone);
1238 inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1239 BIT(LRU_INACTIVE_FILE));
1240 active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1241 BIT(LRU_ACTIVE_FILE));
1243 return (active > inactive);
1246 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
1249 int nid = zone_to_nid(zone);
1250 int zid = zone_idx(zone);
1251 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
1253 return &mz->reclaim_stat;
1256 struct zone_reclaim_stat *
1257 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
1259 struct page_cgroup *pc;
1260 struct mem_cgroup_per_zone *mz;
1262 if (mem_cgroup_disabled())
1265 pc = lookup_page_cgroup(page);
1266 if (!PageCgroupUsed(pc))
1268 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1270 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1271 return &mz->reclaim_stat;
1274 #define mem_cgroup_from_res_counter(counter, member) \
1275 container_of(counter, struct mem_cgroup, member)
1278 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1279 * @mem: the memory cgroup
1281 * Returns the maximum amount of memory @mem can be charged with, in
1284 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1286 unsigned long long margin;
1288 margin = res_counter_margin(&memcg->res);
1289 if (do_swap_account)
1290 margin = min(margin, res_counter_margin(&memcg->memsw));
1291 return margin >> PAGE_SHIFT;
1294 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1296 struct cgroup *cgrp = memcg->css.cgroup;
1299 if (cgrp->parent == NULL)
1300 return vm_swappiness;
1302 return memcg->swappiness;
1306 * memcg->moving_account is used for checking possibility that some thread is
1307 * calling move_account(). When a thread on CPU-A starts moving pages under
1308 * a memcg, other threads should check memcg->moving_account under
1309 * rcu_read_lock(), like this:
1313 * memcg->moving_account+1 if (memcg->mocing_account)
1315 * synchronize_rcu() update something.
1320 /* for quick checking without looking up memcg */
1321 atomic_t memcg_moving __read_mostly;
1323 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1325 atomic_inc(&memcg_moving);
1326 atomic_inc(&memcg->moving_account);
1330 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1333 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1334 * We check NULL in callee rather than caller.
1337 atomic_dec(&memcg_moving);
1338 atomic_dec(&memcg->moving_account);
1343 * 2 routines for checking "mem" is under move_account() or not.
1345 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1346 * is used for avoiding races in accounting. If true,
1347 * pc->mem_cgroup may be overwritten.
1349 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1350 * under hierarchy of moving cgroups. This is for
1351 * waiting at hith-memory prressure caused by "move".
1354 static bool mem_cgroup_stolen(struct mem_cgroup *memcg)
1356 VM_BUG_ON(!rcu_read_lock_held());
1357 return atomic_read(&memcg->moving_account) > 0;
1360 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1362 struct mem_cgroup *from;
1363 struct mem_cgroup *to;
1366 * Unlike task_move routines, we access mc.to, mc.from not under
1367 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1369 spin_lock(&mc.lock);
1375 ret = mem_cgroup_same_or_subtree(memcg, from)
1376 || mem_cgroup_same_or_subtree(memcg, to);
1378 spin_unlock(&mc.lock);
1382 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1384 if (mc.moving_task && current != mc.moving_task) {
1385 if (mem_cgroup_under_move(memcg)) {
1387 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1388 /* moving charge context might have finished. */
1391 finish_wait(&mc.waitq, &wait);
1399 * Take this lock when
1400 * - a code tries to modify page's memcg while it's USED.
1401 * - a code tries to modify page state accounting in a memcg.
1402 * see mem_cgroup_stolen(), too.
1404 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1405 unsigned long *flags)
1407 spin_lock_irqsave(&memcg->move_lock, *flags);
1410 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1411 unsigned long *flags)
1413 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1417 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1418 * @memcg: The memory cgroup that went over limit
1419 * @p: Task that is going to be killed
1421 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1424 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1426 struct cgroup *task_cgrp;
1427 struct cgroup *mem_cgrp;
1429 * Need a buffer in BSS, can't rely on allocations. The code relies
1430 * on the assumption that OOM is serialized for memory controller.
1431 * If this assumption is broken, revisit this code.
1433 static char memcg_name[PATH_MAX];
1441 mem_cgrp = memcg->css.cgroup;
1442 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1444 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1447 * Unfortunately, we are unable to convert to a useful name
1448 * But we'll still print out the usage information
1455 printk(KERN_INFO "Task in %s killed", memcg_name);
1458 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1466 * Continues from above, so we don't need an KERN_ level
1468 printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1471 printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1472 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1473 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1474 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1475 printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1477 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1478 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1479 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1483 * This function returns the number of memcg under hierarchy tree. Returns
1484 * 1(self count) if no children.
1486 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1489 struct mem_cgroup *iter;
1491 for_each_mem_cgroup_tree(iter, memcg)
1497 * Return the memory (and swap, if configured) limit for a memcg.
1499 u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1504 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1505 limit += total_swap_pages << PAGE_SHIFT;
1507 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1509 * If memsw is finite and limits the amount of swap space available
1510 * to this memcg, return that limit.
1512 return min(limit, memsw);
1515 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1517 unsigned long flags)
1519 unsigned long total = 0;
1520 bool noswap = false;
1523 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1525 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1528 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1530 drain_all_stock_async(memcg);
1531 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1533 * Allow limit shrinkers, which are triggered directly
1534 * by userspace, to catch signals and stop reclaim
1535 * after minimal progress, regardless of the margin.
1537 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1539 if (mem_cgroup_margin(memcg))
1542 * If nothing was reclaimed after two attempts, there
1543 * may be no reclaimable pages in this hierarchy.
1552 * test_mem_cgroup_node_reclaimable
1553 * @mem: the target memcg
1554 * @nid: the node ID to be checked.
1555 * @noswap : specify true here if the user wants flle only information.
1557 * This function returns whether the specified memcg contains any
1558 * reclaimable pages on a node. Returns true if there are any reclaimable
1559 * pages in the node.
1561 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1562 int nid, bool noswap)
1564 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1566 if (noswap || !total_swap_pages)
1568 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1573 #if MAX_NUMNODES > 1
1576 * Always updating the nodemask is not very good - even if we have an empty
1577 * list or the wrong list here, we can start from some node and traverse all
1578 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1581 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1585 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1586 * pagein/pageout changes since the last update.
1588 if (!atomic_read(&memcg->numainfo_events))
1590 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1593 /* make a nodemask where this memcg uses memory from */
1594 memcg->scan_nodes = node_states[N_HIGH_MEMORY];
1596 for_each_node_mask(nid, node_states[N_HIGH_MEMORY]) {
1598 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1599 node_clear(nid, memcg->scan_nodes);
1602 atomic_set(&memcg->numainfo_events, 0);
1603 atomic_set(&memcg->numainfo_updating, 0);
1607 * Selecting a node where we start reclaim from. Because what we need is just
1608 * reducing usage counter, start from anywhere is O,K. Considering
1609 * memory reclaim from current node, there are pros. and cons.
1611 * Freeing memory from current node means freeing memory from a node which
1612 * we'll use or we've used. So, it may make LRU bad. And if several threads
1613 * hit limits, it will see a contention on a node. But freeing from remote
1614 * node means more costs for memory reclaim because of memory latency.
1616 * Now, we use round-robin. Better algorithm is welcomed.
1618 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1622 mem_cgroup_may_update_nodemask(memcg);
1623 node = memcg->last_scanned_node;
1625 node = next_node(node, memcg->scan_nodes);
1626 if (node == MAX_NUMNODES)
1627 node = first_node(memcg->scan_nodes);
1629 * We call this when we hit limit, not when pages are added to LRU.
1630 * No LRU may hold pages because all pages are UNEVICTABLE or
1631 * memcg is too small and all pages are not on LRU. In that case,
1632 * we use curret node.
1634 if (unlikely(node == MAX_NUMNODES))
1635 node = numa_node_id();
1637 memcg->last_scanned_node = node;
1642 * Check all nodes whether it contains reclaimable pages or not.
1643 * For quick scan, we make use of scan_nodes. This will allow us to skip
1644 * unused nodes. But scan_nodes is lazily updated and may not cotain
1645 * enough new information. We need to do double check.
1647 bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1652 * quick check...making use of scan_node.
1653 * We can skip unused nodes.
1655 if (!nodes_empty(memcg->scan_nodes)) {
1656 for (nid = first_node(memcg->scan_nodes);
1658 nid = next_node(nid, memcg->scan_nodes)) {
1660 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1665 * Check rest of nodes.
1667 for_each_node_state(nid, N_HIGH_MEMORY) {
1668 if (node_isset(nid, memcg->scan_nodes))
1670 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1677 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1682 bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1684 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
1688 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1691 unsigned long *total_scanned)
1693 struct mem_cgroup *victim = NULL;
1696 unsigned long excess;
1697 unsigned long nr_scanned;
1698 struct mem_cgroup_reclaim_cookie reclaim = {
1703 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
1706 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1711 * If we have not been able to reclaim
1712 * anything, it might because there are
1713 * no reclaimable pages under this hierarchy
1718 * We want to do more targeted reclaim.
1719 * excess >> 2 is not to excessive so as to
1720 * reclaim too much, nor too less that we keep
1721 * coming back to reclaim from this cgroup
1723 if (total >= (excess >> 2) ||
1724 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1729 if (!mem_cgroup_reclaimable(victim, false))
1731 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1733 *total_scanned += nr_scanned;
1734 if (!res_counter_soft_limit_excess(&root_memcg->res))
1737 mem_cgroup_iter_break(root_memcg, victim);
1742 * Check OOM-Killer is already running under our hierarchy.
1743 * If someone is running, return false.
1744 * Has to be called with memcg_oom_lock
1746 static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg)
1748 struct mem_cgroup *iter, *failed = NULL;
1750 for_each_mem_cgroup_tree(iter, memcg) {
1751 if (iter->oom_lock) {
1753 * this subtree of our hierarchy is already locked
1754 * so we cannot give a lock.
1757 mem_cgroup_iter_break(memcg, iter);
1760 iter->oom_lock = true;
1767 * OK, we failed to lock the whole subtree so we have to clean up
1768 * what we set up to the failing subtree
1770 for_each_mem_cgroup_tree(iter, memcg) {
1771 if (iter == failed) {
1772 mem_cgroup_iter_break(memcg, iter);
1775 iter->oom_lock = false;
1781 * Has to be called with memcg_oom_lock
1783 static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1785 struct mem_cgroup *iter;
1787 for_each_mem_cgroup_tree(iter, memcg)
1788 iter->oom_lock = false;
1792 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1794 struct mem_cgroup *iter;
1796 for_each_mem_cgroup_tree(iter, memcg)
1797 atomic_inc(&iter->under_oom);
1800 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1802 struct mem_cgroup *iter;
1805 * When a new child is created while the hierarchy is under oom,
1806 * mem_cgroup_oom_lock() may not be called. We have to use
1807 * atomic_add_unless() here.
1809 for_each_mem_cgroup_tree(iter, memcg)
1810 atomic_add_unless(&iter->under_oom, -1, 0);
1813 static DEFINE_SPINLOCK(memcg_oom_lock);
1814 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1816 struct oom_wait_info {
1817 struct mem_cgroup *memcg;
1821 static int memcg_oom_wake_function(wait_queue_t *wait,
1822 unsigned mode, int sync, void *arg)
1824 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1825 struct mem_cgroup *oom_wait_memcg;
1826 struct oom_wait_info *oom_wait_info;
1828 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1829 oom_wait_memcg = oom_wait_info->memcg;
1832 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
1833 * Then we can use css_is_ancestor without taking care of RCU.
1835 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
1836 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
1838 return autoremove_wake_function(wait, mode, sync, arg);
1841 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
1843 /* for filtering, pass "memcg" as argument. */
1844 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1847 static void memcg_oom_recover(struct mem_cgroup *memcg)
1849 if (memcg && atomic_read(&memcg->under_oom))
1850 memcg_wakeup_oom(memcg);
1854 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1856 bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1858 struct oom_wait_info owait;
1859 bool locked, need_to_kill;
1861 owait.memcg = memcg;
1862 owait.wait.flags = 0;
1863 owait.wait.func = memcg_oom_wake_function;
1864 owait.wait.private = current;
1865 INIT_LIST_HEAD(&owait.wait.task_list);
1866 need_to_kill = true;
1867 mem_cgroup_mark_under_oom(memcg);
1869 /* At first, try to OOM lock hierarchy under memcg.*/
1870 spin_lock(&memcg_oom_lock);
1871 locked = mem_cgroup_oom_lock(memcg);
1873 * Even if signal_pending(), we can't quit charge() loop without
1874 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1875 * under OOM is always welcomed, use TASK_KILLABLE here.
1877 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1878 if (!locked || memcg->oom_kill_disable)
1879 need_to_kill = false;
1881 mem_cgroup_oom_notify(memcg);
1882 spin_unlock(&memcg_oom_lock);
1885 finish_wait(&memcg_oom_waitq, &owait.wait);
1886 mem_cgroup_out_of_memory(memcg, mask, order);
1889 finish_wait(&memcg_oom_waitq, &owait.wait);
1891 spin_lock(&memcg_oom_lock);
1893 mem_cgroup_oom_unlock(memcg);
1894 memcg_wakeup_oom(memcg);
1895 spin_unlock(&memcg_oom_lock);
1897 mem_cgroup_unmark_under_oom(memcg);
1899 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
1901 /* Give chance to dying process */
1902 schedule_timeout_uninterruptible(1);
1907 * Currently used to update mapped file statistics, but the routine can be
1908 * generalized to update other statistics as well.
1910 * Notes: Race condition
1912 * We usually use page_cgroup_lock() for accessing page_cgroup member but
1913 * it tends to be costly. But considering some conditions, we doesn't need
1914 * to do so _always_.
1916 * Considering "charge", lock_page_cgroup() is not required because all
1917 * file-stat operations happen after a page is attached to radix-tree. There
1918 * are no race with "charge".
1920 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
1921 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
1922 * if there are race with "uncharge". Statistics itself is properly handled
1925 * Considering "move", this is an only case we see a race. To make the race
1926 * small, we check mm->moving_account and detect there are possibility of race
1927 * If there is, we take a lock.
1930 void __mem_cgroup_begin_update_page_stat(struct page *page,
1931 bool *locked, unsigned long *flags)
1933 struct mem_cgroup *memcg;
1934 struct page_cgroup *pc;
1936 pc = lookup_page_cgroup(page);
1938 memcg = pc->mem_cgroup;
1939 if (unlikely(!memcg || !PageCgroupUsed(pc)))
1942 * If this memory cgroup is not under account moving, we don't
1943 * need to take move_lock_page_cgroup(). Because we already hold
1944 * rcu_read_lock(), any calls to move_account will be delayed until
1945 * rcu_read_unlock() if mem_cgroup_stolen() == true.
1947 if (!mem_cgroup_stolen(memcg))
1950 move_lock_mem_cgroup(memcg, flags);
1951 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
1952 move_unlock_mem_cgroup(memcg, flags);
1958 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
1960 struct page_cgroup *pc = lookup_page_cgroup(page);
1963 * It's guaranteed that pc->mem_cgroup never changes while
1964 * lock is held because a routine modifies pc->mem_cgroup
1965 * should take move_lock_page_cgroup().
1967 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
1970 void mem_cgroup_update_page_stat(struct page *page,
1971 enum mem_cgroup_page_stat_item idx, int val)
1973 struct mem_cgroup *memcg;
1974 struct page_cgroup *pc = lookup_page_cgroup(page);
1975 unsigned long uninitialized_var(flags);
1977 if (mem_cgroup_disabled())
1980 memcg = pc->mem_cgroup;
1981 if (unlikely(!memcg || !PageCgroupUsed(pc)))
1985 case MEMCG_NR_FILE_MAPPED:
1986 idx = MEM_CGROUP_STAT_FILE_MAPPED;
1992 this_cpu_add(memcg->stat->count[idx], val);
1996 * size of first charge trial. "32" comes from vmscan.c's magic value.
1997 * TODO: maybe necessary to use big numbers in big irons.
1999 #define CHARGE_BATCH 32U
2000 struct memcg_stock_pcp {
2001 struct mem_cgroup *cached; /* this never be root cgroup */
2002 unsigned int nr_pages;
2003 struct work_struct work;
2004 unsigned long flags;
2005 #define FLUSHING_CACHED_CHARGE (0)
2007 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2008 static DEFINE_MUTEX(percpu_charge_mutex);
2011 * Try to consume stocked charge on this cpu. If success, one page is consumed
2012 * from local stock and true is returned. If the stock is 0 or charges from a
2013 * cgroup which is not current target, returns false. This stock will be
2016 static bool consume_stock(struct mem_cgroup *memcg)
2018 struct memcg_stock_pcp *stock;
2021 stock = &get_cpu_var(memcg_stock);
2022 if (memcg == stock->cached && stock->nr_pages)
2024 else /* need to call res_counter_charge */
2026 put_cpu_var(memcg_stock);
2031 * Returns stocks cached in percpu to res_counter and reset cached information.
2033 static void drain_stock(struct memcg_stock_pcp *stock)
2035 struct mem_cgroup *old = stock->cached;
2037 if (stock->nr_pages) {
2038 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2040 res_counter_uncharge(&old->res, bytes);
2041 if (do_swap_account)
2042 res_counter_uncharge(&old->memsw, bytes);
2043 stock->nr_pages = 0;
2045 stock->cached = NULL;
2049 * This must be called under preempt disabled or must be called by
2050 * a thread which is pinned to local cpu.
2052 static void drain_local_stock(struct work_struct *dummy)
2054 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2056 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2060 * Cache charges(val) which is from res_counter, to local per_cpu area.
2061 * This will be consumed by consume_stock() function, later.
2063 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2065 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2067 if (stock->cached != memcg) { /* reset if necessary */
2069 stock->cached = memcg;
2071 stock->nr_pages += nr_pages;
2072 put_cpu_var(memcg_stock);
2076 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2077 * of the hierarchy under it. sync flag says whether we should block
2078 * until the work is done.
2080 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2084 /* Notify other cpus that system-wide "drain" is running */
2087 for_each_online_cpu(cpu) {
2088 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2089 struct mem_cgroup *memcg;
2091 memcg = stock->cached;
2092 if (!memcg || !stock->nr_pages)
2094 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2096 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2098 drain_local_stock(&stock->work);
2100 schedule_work_on(cpu, &stock->work);
2108 for_each_online_cpu(cpu) {
2109 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2110 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2111 flush_work(&stock->work);
2118 * Tries to drain stocked charges in other cpus. This function is asynchronous
2119 * and just put a work per cpu for draining localy on each cpu. Caller can
2120 * expects some charges will be back to res_counter later but cannot wait for
2123 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2126 * If someone calls draining, avoid adding more kworker runs.
2128 if (!mutex_trylock(&percpu_charge_mutex))
2130 drain_all_stock(root_memcg, false);
2131 mutex_unlock(&percpu_charge_mutex);
2134 /* This is a synchronous drain interface. */
2135 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2137 /* called when force_empty is called */
2138 mutex_lock(&percpu_charge_mutex);
2139 drain_all_stock(root_memcg, true);
2140 mutex_unlock(&percpu_charge_mutex);
2144 * This function drains percpu counter value from DEAD cpu and
2145 * move it to local cpu. Note that this function can be preempted.
2147 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2151 spin_lock(&memcg->pcp_counter_lock);
2152 for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) {
2153 long x = per_cpu(memcg->stat->count[i], cpu);
2155 per_cpu(memcg->stat->count[i], cpu) = 0;
2156 memcg->nocpu_base.count[i] += x;
2158 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2159 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2161 per_cpu(memcg->stat->events[i], cpu) = 0;
2162 memcg->nocpu_base.events[i] += x;
2164 spin_unlock(&memcg->pcp_counter_lock);
2167 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2168 unsigned long action,
2171 int cpu = (unsigned long)hcpu;
2172 struct memcg_stock_pcp *stock;
2173 struct mem_cgroup *iter;
2175 if (action == CPU_ONLINE)
2178 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2181 for_each_mem_cgroup(iter)
2182 mem_cgroup_drain_pcp_counter(iter, cpu);
2184 stock = &per_cpu(memcg_stock, cpu);
2190 /* See __mem_cgroup_try_charge() for details */
2192 CHARGE_OK, /* success */
2193 CHARGE_RETRY, /* need to retry but retry is not bad */
2194 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2195 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2196 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2199 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2200 unsigned int nr_pages, bool oom_check)
2202 unsigned long csize = nr_pages * PAGE_SIZE;
2203 struct mem_cgroup *mem_over_limit;
2204 struct res_counter *fail_res;
2205 unsigned long flags = 0;
2208 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2211 if (!do_swap_account)
2213 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2217 res_counter_uncharge(&memcg->res, csize);
2218 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2219 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2221 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2223 * nr_pages can be either a huge page (HPAGE_PMD_NR), a batch
2224 * of regular pages (CHARGE_BATCH), or a single regular page (1).
2226 * Never reclaim on behalf of optional batching, retry with a
2227 * single page instead.
2229 if (nr_pages == CHARGE_BATCH)
2230 return CHARGE_RETRY;
2232 if (!(gfp_mask & __GFP_WAIT))
2233 return CHARGE_WOULDBLOCK;
2235 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2236 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2237 return CHARGE_RETRY;
2239 * Even though the limit is exceeded at this point, reclaim
2240 * may have been able to free some pages. Retry the charge
2241 * before killing the task.
2243 * Only for regular pages, though: huge pages are rather
2244 * unlikely to succeed so close to the limit, and we fall back
2245 * to regular pages anyway in case of failure.
2247 if (nr_pages == 1 && ret)
2248 return CHARGE_RETRY;
2251 * At task move, charge accounts can be doubly counted. So, it's
2252 * better to wait until the end of task_move if something is going on.
2254 if (mem_cgroup_wait_acct_move(mem_over_limit))
2255 return CHARGE_RETRY;
2257 /* If we don't need to call oom-killer at el, return immediately */
2259 return CHARGE_NOMEM;
2261 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask, get_order(csize)))
2262 return CHARGE_OOM_DIE;
2264 return CHARGE_RETRY;
2268 * __mem_cgroup_try_charge() does
2269 * 1. detect memcg to be charged against from passed *mm and *ptr,
2270 * 2. update res_counter
2271 * 3. call memory reclaim if necessary.
2273 * In some special case, if the task is fatal, fatal_signal_pending() or
2274 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2275 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2276 * as possible without any hazards. 2: all pages should have a valid
2277 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2278 * pointer, that is treated as a charge to root_mem_cgroup.
2280 * So __mem_cgroup_try_charge() will return
2281 * 0 ... on success, filling *ptr with a valid memcg pointer.
2282 * -ENOMEM ... charge failure because of resource limits.
2283 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2285 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2286 * the oom-killer can be invoked.
2288 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2290 unsigned int nr_pages,
2291 struct mem_cgroup **ptr,
2294 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2295 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2296 struct mem_cgroup *memcg = NULL;
2300 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2301 * in system level. So, allow to go ahead dying process in addition to
2304 if (unlikely(test_thread_flag(TIF_MEMDIE)
2305 || fatal_signal_pending(current)))
2309 * We always charge the cgroup the mm_struct belongs to.
2310 * The mm_struct's mem_cgroup changes on task migration if the
2311 * thread group leader migrates. It's possible that mm is not
2312 * set, if so charge the init_mm (happens for pagecache usage).
2315 *ptr = root_mem_cgroup;
2317 if (*ptr) { /* css should be a valid one */
2319 VM_BUG_ON(css_is_removed(&memcg->css));
2320 if (mem_cgroup_is_root(memcg))
2322 if (nr_pages == 1 && consume_stock(memcg))
2324 css_get(&memcg->css);
2326 struct task_struct *p;
2329 p = rcu_dereference(mm->owner);
2331 * Because we don't have task_lock(), "p" can exit.
2332 * In that case, "memcg" can point to root or p can be NULL with
2333 * race with swapoff. Then, we have small risk of mis-accouning.
2334 * But such kind of mis-account by race always happens because
2335 * we don't have cgroup_mutex(). It's overkill and we allo that
2337 * (*) swapoff at el will charge against mm-struct not against
2338 * task-struct. So, mm->owner can be NULL.
2340 memcg = mem_cgroup_from_task(p);
2342 memcg = root_mem_cgroup;
2343 if (mem_cgroup_is_root(memcg)) {
2347 if (nr_pages == 1 && consume_stock(memcg)) {
2349 * It seems dagerous to access memcg without css_get().
2350 * But considering how consume_stok works, it's not
2351 * necessary. If consume_stock success, some charges
2352 * from this memcg are cached on this cpu. So, we
2353 * don't need to call css_get()/css_tryget() before
2354 * calling consume_stock().
2359 /* after here, we may be blocked. we need to get refcnt */
2360 if (!css_tryget(&memcg->css)) {
2370 /* If killed, bypass charge */
2371 if (fatal_signal_pending(current)) {
2372 css_put(&memcg->css);
2377 if (oom && !nr_oom_retries) {
2379 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2382 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, oom_check);
2386 case CHARGE_RETRY: /* not in OOM situation but retry */
2388 css_put(&memcg->css);
2391 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2392 css_put(&memcg->css);
2394 case CHARGE_NOMEM: /* OOM routine works */
2396 css_put(&memcg->css);
2399 /* If oom, we never return -ENOMEM */
2402 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2403 css_put(&memcg->css);
2406 } while (ret != CHARGE_OK);
2408 if (batch > nr_pages)
2409 refill_stock(memcg, batch - nr_pages);
2410 css_put(&memcg->css);
2418 *ptr = root_mem_cgroup;
2423 * Somemtimes we have to undo a charge we got by try_charge().
2424 * This function is for that and do uncharge, put css's refcnt.
2425 * gotten by try_charge().
2427 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2428 unsigned int nr_pages)
2430 if (!mem_cgroup_is_root(memcg)) {
2431 unsigned long bytes = nr_pages * PAGE_SIZE;
2433 res_counter_uncharge(&memcg->res, bytes);
2434 if (do_swap_account)
2435 res_counter_uncharge(&memcg->memsw, bytes);
2440 * A helper function to get mem_cgroup from ID. must be called under
2441 * rcu_read_lock(). The caller must check css_is_removed() or some if
2442 * it's concern. (dropping refcnt from swap can be called against removed
2445 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2447 struct cgroup_subsys_state *css;
2449 /* ID 0 is unused ID */
2452 css = css_lookup(&mem_cgroup_subsys, id);
2455 return container_of(css, struct mem_cgroup, css);
2458 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2460 struct mem_cgroup *memcg = NULL;
2461 struct page_cgroup *pc;
2465 VM_BUG_ON(!PageLocked(page));
2467 pc = lookup_page_cgroup(page);
2468 lock_page_cgroup(pc);
2469 if (PageCgroupUsed(pc)) {
2470 memcg = pc->mem_cgroup;
2471 if (memcg && !css_tryget(&memcg->css))
2473 } else if (PageSwapCache(page)) {
2474 ent.val = page_private(page);
2475 id = lookup_swap_cgroup_id(ent);
2477 memcg = mem_cgroup_lookup(id);
2478 if (memcg && !css_tryget(&memcg->css))
2482 unlock_page_cgroup(pc);
2486 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2488 unsigned int nr_pages,
2489 enum charge_type ctype,
2492 struct page_cgroup *pc = lookup_page_cgroup(page);
2493 struct zone *uninitialized_var(zone);
2494 bool was_on_lru = false;
2497 lock_page_cgroup(pc);
2498 if (unlikely(PageCgroupUsed(pc))) {
2499 unlock_page_cgroup(pc);
2500 __mem_cgroup_cancel_charge(memcg, nr_pages);
2504 * we don't need page_cgroup_lock about tail pages, becase they are not
2505 * accessed by any other context at this point.
2509 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2510 * may already be on some other mem_cgroup's LRU. Take care of it.
2513 zone = page_zone(page);
2514 spin_lock_irq(&zone->lru_lock);
2515 if (PageLRU(page)) {
2517 del_page_from_lru_list(zone, page, page_lru(page));
2522 pc->mem_cgroup = memcg;
2524 * We access a page_cgroup asynchronously without lock_page_cgroup().
2525 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2526 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2527 * before USED bit, we need memory barrier here.
2528 * See mem_cgroup_add_lru_list(), etc.
2531 SetPageCgroupUsed(pc);
2535 VM_BUG_ON(PageLRU(page));
2537 add_page_to_lru_list(zone, page, page_lru(page));
2539 spin_unlock_irq(&zone->lru_lock);
2542 if (ctype == MEM_CGROUP_CHARGE_TYPE_MAPPED)
2547 mem_cgroup_charge_statistics(memcg, anon, nr_pages);
2548 unlock_page_cgroup(pc);
2551 * "charge_statistics" updated event counter. Then, check it.
2552 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2553 * if they exceeds softlimit.
2555 memcg_check_events(memcg, page);
2558 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2560 #define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MIGRATION))
2562 * Because tail pages are not marked as "used", set it. We're under
2563 * zone->lru_lock, 'splitting on pmd' and compound_lock.
2564 * charge/uncharge will be never happen and move_account() is done under
2565 * compound_lock(), so we don't have to take care of races.
2567 void mem_cgroup_split_huge_fixup(struct page *head)
2569 struct page_cgroup *head_pc = lookup_page_cgroup(head);
2570 struct page_cgroup *pc;
2573 if (mem_cgroup_disabled())
2575 for (i = 1; i < HPAGE_PMD_NR; i++) {
2577 pc->mem_cgroup = head_pc->mem_cgroup;
2578 smp_wmb();/* see __commit_charge() */
2579 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2582 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2585 * mem_cgroup_move_account - move account of the page
2587 * @nr_pages: number of regular pages (>1 for huge pages)
2588 * @pc: page_cgroup of the page.
2589 * @from: mem_cgroup which the page is moved from.
2590 * @to: mem_cgroup which the page is moved to. @from != @to.
2591 * @uncharge: whether we should call uncharge and css_put against @from.
2593 * The caller must confirm following.
2594 * - page is not on LRU (isolate_page() is useful.)
2595 * - compound_lock is held when nr_pages > 1
2597 * This function doesn't do "charge" nor css_get to new cgroup. It should be
2598 * done by a caller(__mem_cgroup_try_charge would be useful). If @uncharge is
2599 * true, this function does "uncharge" from old cgroup, but it doesn't if
2600 * @uncharge is false, so a caller should do "uncharge".
2602 static int mem_cgroup_move_account(struct page *page,
2603 unsigned int nr_pages,
2604 struct page_cgroup *pc,
2605 struct mem_cgroup *from,
2606 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 /* This is not "cancel", but cancel_charge does all we need. */
2643 __mem_cgroup_cancel_charge(from, nr_pages);
2645 /* caller should have done css_get */
2646 pc->mem_cgroup = to;
2647 mem_cgroup_charge_statistics(to, anon, nr_pages);
2649 * We charges against "to" which may not have any tasks. Then, "to"
2650 * can be under rmdir(). But in current implementation, caller of
2651 * this function is just force_empty() and move charge, so it's
2652 * guaranteed that "to" is never removed. So, we don't check rmdir
2655 move_unlock_mem_cgroup(from, &flags);
2658 unlock_page_cgroup(pc);
2662 memcg_check_events(to, page);
2663 memcg_check_events(from, page);
2669 * move charges to its parent.
2672 static int mem_cgroup_move_parent(struct page *page,
2673 struct page_cgroup *pc,
2674 struct mem_cgroup *child,
2677 struct cgroup *cg = child->css.cgroup;
2678 struct cgroup *pcg = cg->parent;
2679 struct mem_cgroup *parent;
2680 unsigned int nr_pages;
2681 unsigned long uninitialized_var(flags);
2689 if (!get_page_unless_zero(page))
2691 if (isolate_lru_page(page))
2694 nr_pages = hpage_nr_pages(page);
2696 parent = mem_cgroup_from_cont(pcg);
2697 ret = __mem_cgroup_try_charge(NULL, gfp_mask, nr_pages, &parent, false);
2702 flags = compound_lock_irqsave(page);
2704 ret = mem_cgroup_move_account(page, nr_pages, pc, child, parent, true);
2706 __mem_cgroup_cancel_charge(parent, nr_pages);
2709 compound_unlock_irqrestore(page, flags);
2711 putback_lru_page(page);
2719 * Charge the memory controller for page usage.
2721 * 0 if the charge was successful
2722 * < 0 if the cgroup is over its limit
2724 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2725 gfp_t gfp_mask, enum charge_type ctype)
2727 struct mem_cgroup *memcg = NULL;
2728 unsigned int nr_pages = 1;
2732 if (PageTransHuge(page)) {
2733 nr_pages <<= compound_order(page);
2734 VM_BUG_ON(!PageTransHuge(page));
2736 * Never OOM-kill a process for a huge page. The
2737 * fault handler will fall back to regular pages.
2742 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
2745 __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false);
2749 int mem_cgroup_newpage_charge(struct page *page,
2750 struct mm_struct *mm, gfp_t gfp_mask)
2752 if (mem_cgroup_disabled())
2754 VM_BUG_ON(page_mapped(page));
2755 VM_BUG_ON(page->mapping && !PageAnon(page));
2757 return mem_cgroup_charge_common(page, mm, gfp_mask,
2758 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2762 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2763 enum charge_type ctype);
2765 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2768 struct mem_cgroup *memcg = NULL;
2769 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
2772 if (mem_cgroup_disabled())
2774 if (PageCompound(page))
2779 if (!page_is_file_cache(page))
2780 type = MEM_CGROUP_CHARGE_TYPE_SHMEM;
2782 if (!PageSwapCache(page))
2783 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
2784 else { /* page is swapcache/shmem */
2785 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &memcg);
2787 __mem_cgroup_commit_charge_swapin(page, memcg, type);
2793 * While swap-in, try_charge -> commit or cancel, the page is locked.
2794 * And when try_charge() successfully returns, one refcnt to memcg without
2795 * struct page_cgroup is acquired. This refcnt will be consumed by
2796 * "commit()" or removed by "cancel()"
2798 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2800 gfp_t mask, struct mem_cgroup **memcgp)
2802 struct mem_cgroup *memcg;
2807 if (mem_cgroup_disabled())
2810 if (!do_swap_account)
2813 * A racing thread's fault, or swapoff, may have already updated
2814 * the pte, and even removed page from swap cache: in those cases
2815 * do_swap_page()'s pte_same() test will fail; but there's also a
2816 * KSM case which does need to charge the page.
2818 if (!PageSwapCache(page))
2820 memcg = try_get_mem_cgroup_from_page(page);
2824 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
2825 css_put(&memcg->css);
2832 ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
2839 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
2840 enum charge_type ctype)
2842 if (mem_cgroup_disabled())
2846 cgroup_exclude_rmdir(&memcg->css);
2848 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
2850 * Now swap is on-memory. This means this page may be
2851 * counted both as mem and swap....double count.
2852 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2853 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2854 * may call delete_from_swap_cache() before reach here.
2856 if (do_swap_account && PageSwapCache(page)) {
2857 swp_entry_t ent = {.val = page_private(page)};
2858 struct mem_cgroup *swap_memcg;
2861 id = swap_cgroup_record(ent, 0);
2863 swap_memcg = mem_cgroup_lookup(id);
2866 * This recorded memcg can be obsolete one. So, avoid
2867 * calling css_tryget
2869 if (!mem_cgroup_is_root(swap_memcg))
2870 res_counter_uncharge(&swap_memcg->memsw,
2872 mem_cgroup_swap_statistics(swap_memcg, false);
2873 mem_cgroup_put(swap_memcg);
2878 * At swapin, we may charge account against cgroup which has no tasks.
2879 * So, rmdir()->pre_destroy() can be called while we do this charge.
2880 * In that case, we need to call pre_destroy() again. check it here.
2882 cgroup_release_and_wakeup_rmdir(&memcg->css);
2885 void mem_cgroup_commit_charge_swapin(struct page *page,
2886 struct mem_cgroup *memcg)
2888 __mem_cgroup_commit_charge_swapin(page, memcg,
2889 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2892 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
2894 if (mem_cgroup_disabled())
2898 __mem_cgroup_cancel_charge(memcg, 1);
2901 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
2902 unsigned int nr_pages,
2903 const enum charge_type ctype)
2905 struct memcg_batch_info *batch = NULL;
2906 bool uncharge_memsw = true;
2908 /* If swapout, usage of swap doesn't decrease */
2909 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2910 uncharge_memsw = false;
2912 batch = ¤t->memcg_batch;
2914 * In usual, we do css_get() when we remember memcg pointer.
2915 * But in this case, we keep res->usage until end of a series of
2916 * uncharges. Then, it's ok to ignore memcg's refcnt.
2919 batch->memcg = memcg;
2921 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2922 * In those cases, all pages freed continuously can be expected to be in
2923 * the same cgroup and we have chance to coalesce uncharges.
2924 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2925 * because we want to do uncharge as soon as possible.
2928 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
2929 goto direct_uncharge;
2932 goto direct_uncharge;
2935 * In typical case, batch->memcg == mem. This means we can
2936 * merge a series of uncharges to an uncharge of res_counter.
2937 * If not, we uncharge res_counter ony by one.
2939 if (batch->memcg != memcg)
2940 goto direct_uncharge;
2941 /* remember freed charge and uncharge it later */
2944 batch->memsw_nr_pages++;
2947 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
2949 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
2950 if (unlikely(batch->memcg != memcg))
2951 memcg_oom_recover(memcg);
2955 * uncharge if !page_mapped(page)
2957 static struct mem_cgroup *
2958 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
2960 struct mem_cgroup *memcg = NULL;
2961 unsigned int nr_pages = 1;
2962 struct page_cgroup *pc;
2965 if (mem_cgroup_disabled())
2968 if (PageSwapCache(page))
2971 if (PageTransHuge(page)) {
2972 nr_pages <<= compound_order(page);
2973 VM_BUG_ON(!PageTransHuge(page));
2976 * Check if our page_cgroup is valid
2978 pc = lookup_page_cgroup(page);
2979 if (unlikely(!PageCgroupUsed(pc)))
2982 lock_page_cgroup(pc);
2984 memcg = pc->mem_cgroup;
2986 if (!PageCgroupUsed(pc))
2989 anon = PageAnon(page);
2992 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2994 * Generally PageAnon tells if it's the anon statistics to be
2995 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
2996 * used before page reached the stage of being marked PageAnon.
3000 case MEM_CGROUP_CHARGE_TYPE_DROP:
3001 /* See mem_cgroup_prepare_migration() */
3002 if (page_mapped(page) || PageCgroupMigration(pc))
3005 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
3006 if (!PageAnon(page)) { /* Shared memory */
3007 if (page->mapping && !page_is_file_cache(page))
3009 } else if (page_mapped(page)) /* Anon */
3016 mem_cgroup_charge_statistics(memcg, anon, -nr_pages);
3018 ClearPageCgroupUsed(pc);
3020 * pc->mem_cgroup is not cleared here. It will be accessed when it's
3021 * freed from LRU. This is safe because uncharged page is expected not
3022 * to be reused (freed soon). Exception is SwapCache, it's handled by
3023 * special functions.
3026 unlock_page_cgroup(pc);
3028 * even after unlock, we have memcg->res.usage here and this memcg
3029 * will never be freed.
3031 memcg_check_events(memcg, page);
3032 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
3033 mem_cgroup_swap_statistics(memcg, true);
3034 mem_cgroup_get(memcg);
3036 if (!mem_cgroup_is_root(memcg))
3037 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
3042 unlock_page_cgroup(pc);
3046 void mem_cgroup_uncharge_page(struct page *page)
3049 if (page_mapped(page))
3051 VM_BUG_ON(page->mapping && !PageAnon(page));
3052 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
3055 void mem_cgroup_uncharge_cache_page(struct page *page)
3057 VM_BUG_ON(page_mapped(page));
3058 VM_BUG_ON(page->mapping);
3059 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
3063 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
3064 * In that cases, pages are freed continuously and we can expect pages
3065 * are in the same memcg. All these calls itself limits the number of
3066 * pages freed at once, then uncharge_start/end() is called properly.
3067 * This may be called prural(2) times in a context,
3070 void mem_cgroup_uncharge_start(void)
3072 current->memcg_batch.do_batch++;
3073 /* We can do nest. */
3074 if (current->memcg_batch.do_batch == 1) {
3075 current->memcg_batch.memcg = NULL;
3076 current->memcg_batch.nr_pages = 0;
3077 current->memcg_batch.memsw_nr_pages = 0;
3081 void mem_cgroup_uncharge_end(void)
3083 struct memcg_batch_info *batch = ¤t->memcg_batch;
3085 if (!batch->do_batch)
3089 if (batch->do_batch) /* If stacked, do nothing. */
3095 * This "batch->memcg" is valid without any css_get/put etc...
3096 * bacause we hide charges behind us.
3098 if (batch->nr_pages)
3099 res_counter_uncharge(&batch->memcg->res,
3100 batch->nr_pages * PAGE_SIZE);
3101 if (batch->memsw_nr_pages)
3102 res_counter_uncharge(&batch->memcg->memsw,
3103 batch->memsw_nr_pages * PAGE_SIZE);
3104 memcg_oom_recover(batch->memcg);
3105 /* forget this pointer (for sanity check) */
3106 batch->memcg = NULL;
3111 * called after __delete_from_swap_cache() and drop "page" account.
3112 * memcg information is recorded to swap_cgroup of "ent"
3115 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
3117 struct mem_cgroup *memcg;
3118 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
3120 if (!swapout) /* this was a swap cache but the swap is unused ! */
3121 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
3123 memcg = __mem_cgroup_uncharge_common(page, ctype);
3126 * record memcg information, if swapout && memcg != NULL,
3127 * mem_cgroup_get() was called in uncharge().
3129 if (do_swap_account && swapout && memcg)
3130 swap_cgroup_record(ent, css_id(&memcg->css));
3134 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3136 * called from swap_entry_free(). remove record in swap_cgroup and
3137 * uncharge "memsw" account.
3139 void mem_cgroup_uncharge_swap(swp_entry_t ent)
3141 struct mem_cgroup *memcg;
3144 if (!do_swap_account)
3147 id = swap_cgroup_record(ent, 0);
3149 memcg = mem_cgroup_lookup(id);
3152 * We uncharge this because swap is freed.
3153 * This memcg can be obsolete one. We avoid calling css_tryget
3155 if (!mem_cgroup_is_root(memcg))
3156 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
3157 mem_cgroup_swap_statistics(memcg, false);
3158 mem_cgroup_put(memcg);
3164 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3165 * @entry: swap entry to be moved
3166 * @from: mem_cgroup which the entry is moved from
3167 * @to: mem_cgroup which the entry is moved to
3168 * @need_fixup: whether we should fixup res_counters and refcounts.
3170 * It succeeds only when the swap_cgroup's record for this entry is the same
3171 * as the mem_cgroup's id of @from.
3173 * Returns 0 on success, -EINVAL on failure.
3175 * The caller must have charged to @to, IOW, called res_counter_charge() about
3176 * both res and memsw, and called css_get().
3178 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3179 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3181 unsigned short old_id, new_id;
3183 old_id = css_id(&from->css);
3184 new_id = css_id(&to->css);
3186 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3187 mem_cgroup_swap_statistics(from, false);
3188 mem_cgroup_swap_statistics(to, true);
3190 * This function is only called from task migration context now.
3191 * It postpones res_counter and refcount handling till the end
3192 * of task migration(mem_cgroup_clear_mc()) for performance
3193 * improvement. But we cannot postpone mem_cgroup_get(to)
3194 * because if the process that has been moved to @to does
3195 * swap-in, the refcount of @to might be decreased to 0.
3199 if (!mem_cgroup_is_root(from))
3200 res_counter_uncharge(&from->memsw, PAGE_SIZE);
3201 mem_cgroup_put(from);
3203 * we charged both to->res and to->memsw, so we should
3206 if (!mem_cgroup_is_root(to))
3207 res_counter_uncharge(&to->res, PAGE_SIZE);
3214 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3215 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3222 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
3225 int mem_cgroup_prepare_migration(struct page *page,
3226 struct page *newpage, struct mem_cgroup **memcgp, gfp_t gfp_mask)
3228 struct mem_cgroup *memcg = NULL;
3229 struct page_cgroup *pc;
3230 enum charge_type ctype;
3235 VM_BUG_ON(PageTransHuge(page));
3236 if (mem_cgroup_disabled())
3239 pc = lookup_page_cgroup(page);
3240 lock_page_cgroup(pc);
3241 if (PageCgroupUsed(pc)) {
3242 memcg = pc->mem_cgroup;
3243 css_get(&memcg->css);
3245 * At migrating an anonymous page, its mapcount goes down
3246 * to 0 and uncharge() will be called. But, even if it's fully
3247 * unmapped, migration may fail and this page has to be
3248 * charged again. We set MIGRATION flag here and delay uncharge
3249 * until end_migration() is called
3251 * Corner Case Thinking
3253 * When the old page was mapped as Anon and it's unmap-and-freed
3254 * while migration was ongoing.
3255 * If unmap finds the old page, uncharge() of it will be delayed
3256 * until end_migration(). If unmap finds a new page, it's
3257 * uncharged when it make mapcount to be 1->0. If unmap code
3258 * finds swap_migration_entry, the new page will not be mapped
3259 * and end_migration() will find it(mapcount==0).
3262 * When the old page was mapped but migraion fails, the kernel
3263 * remaps it. A charge for it is kept by MIGRATION flag even
3264 * if mapcount goes down to 0. We can do remap successfully
3265 * without charging it again.
3268 * The "old" page is under lock_page() until the end of
3269 * migration, so, the old page itself will not be swapped-out.
3270 * If the new page is swapped out before end_migraton, our
3271 * hook to usual swap-out path will catch the event.
3274 SetPageCgroupMigration(pc);
3276 unlock_page_cgroup(pc);
3278 * If the page is not charged at this point,
3285 ret = __mem_cgroup_try_charge(NULL, gfp_mask, 1, memcgp, false);
3286 css_put(&memcg->css);/* drop extra refcnt */
3288 if (PageAnon(page)) {
3289 lock_page_cgroup(pc);
3290 ClearPageCgroupMigration(pc);
3291 unlock_page_cgroup(pc);
3293 * The old page may be fully unmapped while we kept it.
3295 mem_cgroup_uncharge_page(page);
3297 /* we'll need to revisit this error code (we have -EINTR) */
3301 * We charge new page before it's used/mapped. So, even if unlock_page()
3302 * is called before end_migration, we can catch all events on this new
3303 * page. In the case new page is migrated but not remapped, new page's
3304 * mapcount will be finally 0 and we call uncharge in end_migration().
3307 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
3308 else if (page_is_file_cache(page))
3309 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
3311 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3312 __mem_cgroup_commit_charge(memcg, newpage, 1, ctype, false);
3316 /* remove redundant charge if migration failed*/
3317 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
3318 struct page *oldpage, struct page *newpage, bool migration_ok)
3320 struct page *used, *unused;
3321 struct page_cgroup *pc;
3326 /* blocks rmdir() */
3327 cgroup_exclude_rmdir(&memcg->css);
3328 if (!migration_ok) {
3336 * We disallowed uncharge of pages under migration because mapcount
3337 * of the page goes down to zero, temporarly.
3338 * Clear the flag and check the page should be charged.
3340 pc = lookup_page_cgroup(oldpage);
3341 lock_page_cgroup(pc);
3342 ClearPageCgroupMigration(pc);
3343 unlock_page_cgroup(pc);
3344 anon = PageAnon(used);
3345 __mem_cgroup_uncharge_common(unused,
3346 anon ? MEM_CGROUP_CHARGE_TYPE_MAPPED
3347 : MEM_CGROUP_CHARGE_TYPE_CACHE);
3350 * If a page is a file cache, radix-tree replacement is very atomic
3351 * and we can skip this check. When it was an Anon page, its mapcount
3352 * goes down to 0. But because we added MIGRATION flage, it's not
3353 * uncharged yet. There are several case but page->mapcount check
3354 * and USED bit check in mem_cgroup_uncharge_page() will do enough
3355 * check. (see prepare_charge() also)
3358 mem_cgroup_uncharge_page(used);
3360 * At migration, we may charge account against cgroup which has no
3362 * So, rmdir()->pre_destroy() can be called while we do this charge.
3363 * In that case, we need to call pre_destroy() again. check it here.
3365 cgroup_release_and_wakeup_rmdir(&memcg->css);
3369 * At replace page cache, newpage is not under any memcg but it's on
3370 * LRU. So, this function doesn't touch res_counter but handles LRU
3371 * in correct way. Both pages are locked so we cannot race with uncharge.
3373 void mem_cgroup_replace_page_cache(struct page *oldpage,
3374 struct page *newpage)
3376 struct mem_cgroup *memcg = NULL;
3377 struct page_cgroup *pc;
3378 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
3380 if (mem_cgroup_disabled())
3383 pc = lookup_page_cgroup(oldpage);
3384 /* fix accounting on old pages */
3385 lock_page_cgroup(pc);
3386 if (PageCgroupUsed(pc)) {
3387 memcg = pc->mem_cgroup;
3388 mem_cgroup_charge_statistics(memcg, false, -1);
3389 ClearPageCgroupUsed(pc);
3391 unlock_page_cgroup(pc);
3394 * When called from shmem_replace_page(), in some cases the
3395 * oldpage has already been charged, and in some cases not.
3400 if (PageSwapBacked(oldpage))
3401 type = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3404 * Even if newpage->mapping was NULL before starting replacement,
3405 * the newpage may be on LRU(or pagevec for LRU) already. We lock
3406 * LRU while we overwrite pc->mem_cgroup.
3408 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
3411 #ifdef CONFIG_DEBUG_VM
3412 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3414 struct page_cgroup *pc;
3416 pc = lookup_page_cgroup(page);
3418 * Can be NULL while feeding pages into the page allocator for
3419 * the first time, i.e. during boot or memory hotplug;
3420 * or when mem_cgroup_disabled().
3422 if (likely(pc) && PageCgroupUsed(pc))
3427 bool mem_cgroup_bad_page_check(struct page *page)
3429 if (mem_cgroup_disabled())
3432 return lookup_page_cgroup_used(page) != NULL;
3435 void mem_cgroup_print_bad_page(struct page *page)
3437 struct page_cgroup *pc;
3439 pc = lookup_page_cgroup_used(page);
3441 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
3442 pc, pc->flags, pc->mem_cgroup);
3447 static DEFINE_MUTEX(set_limit_mutex);
3449 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3450 unsigned long long val)
3453 u64 memswlimit, memlimit;
3455 int children = mem_cgroup_count_children(memcg);
3456 u64 curusage, oldusage;
3460 * For keeping hierarchical_reclaim simple, how long we should retry
3461 * is depends on callers. We set our retry-count to be function
3462 * of # of children which we should visit in this loop.
3464 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3466 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3469 while (retry_count) {
3470 if (signal_pending(current)) {
3475 * Rather than hide all in some function, I do this in
3476 * open coded manner. You see what this really does.
3477 * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3479 mutex_lock(&set_limit_mutex);
3480 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3481 if (memswlimit < val) {
3483 mutex_unlock(&set_limit_mutex);
3487 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3491 ret = res_counter_set_limit(&memcg->res, val);
3493 if (memswlimit == val)
3494 memcg->memsw_is_minimum = true;
3496 memcg->memsw_is_minimum = false;
3498 mutex_unlock(&set_limit_mutex);
3503 mem_cgroup_reclaim(memcg, GFP_KERNEL,
3504 MEM_CGROUP_RECLAIM_SHRINK);
3505 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3506 /* Usage is reduced ? */
3507 if (curusage >= oldusage)
3510 oldusage = curusage;
3512 if (!ret && enlarge)
3513 memcg_oom_recover(memcg);
3518 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3519 unsigned long long val)
3522 u64 memlimit, memswlimit, oldusage, curusage;
3523 int children = mem_cgroup_count_children(memcg);
3527 /* see mem_cgroup_resize_res_limit */
3528 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3529 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3530 while (retry_count) {
3531 if (signal_pending(current)) {
3536 * Rather than hide all in some function, I do this in
3537 * open coded manner. You see what this really does.
3538 * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3540 mutex_lock(&set_limit_mutex);
3541 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3542 if (memlimit > val) {
3544 mutex_unlock(&set_limit_mutex);
3547 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3548 if (memswlimit < val)
3550 ret = res_counter_set_limit(&memcg->memsw, val);
3552 if (memlimit == val)
3553 memcg->memsw_is_minimum = true;
3555 memcg->memsw_is_minimum = false;
3557 mutex_unlock(&set_limit_mutex);
3562 mem_cgroup_reclaim(memcg, GFP_KERNEL,
3563 MEM_CGROUP_RECLAIM_NOSWAP |
3564 MEM_CGROUP_RECLAIM_SHRINK);
3565 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3566 /* Usage is reduced ? */
3567 if (curusage >= oldusage)
3570 oldusage = curusage;
3572 if (!ret && enlarge)
3573 memcg_oom_recover(memcg);
3577 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3579 unsigned long *total_scanned)
3581 unsigned long nr_reclaimed = 0;
3582 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3583 unsigned long reclaimed;
3585 struct mem_cgroup_tree_per_zone *mctz;
3586 unsigned long long excess;
3587 unsigned long nr_scanned;
3592 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3594 * This loop can run a while, specially if mem_cgroup's continuously
3595 * keep exceeding their soft limit and putting the system under
3602 mz = mem_cgroup_largest_soft_limit_node(mctz);
3607 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
3608 gfp_mask, &nr_scanned);
3609 nr_reclaimed += reclaimed;
3610 *total_scanned += nr_scanned;
3611 spin_lock(&mctz->lock);
3614 * If we failed to reclaim anything from this memory cgroup
3615 * it is time to move on to the next cgroup
3621 * Loop until we find yet another one.
3623 * By the time we get the soft_limit lock
3624 * again, someone might have aded the
3625 * group back on the RB tree. Iterate to
3626 * make sure we get a different mem.
3627 * mem_cgroup_largest_soft_limit_node returns
3628 * NULL if no other cgroup is present on
3632 __mem_cgroup_largest_soft_limit_node(mctz);
3634 css_put(&next_mz->memcg->css);
3635 else /* next_mz == NULL or other memcg */
3639 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
3640 excess = res_counter_soft_limit_excess(&mz->memcg->res);
3642 * One school of thought says that we should not add
3643 * back the node to the tree if reclaim returns 0.
3644 * But our reclaim could return 0, simply because due
3645 * to priority we are exposing a smaller subset of
3646 * memory to reclaim from. Consider this as a longer
3649 /* If excess == 0, no tree ops */
3650 __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess);
3651 spin_unlock(&mctz->lock);
3652 css_put(&mz->memcg->css);
3655 * Could not reclaim anything and there are no more
3656 * mem cgroups to try or we seem to be looping without
3657 * reclaiming anything.
3659 if (!nr_reclaimed &&
3661 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3663 } while (!nr_reclaimed);
3665 css_put(&next_mz->memcg->css);
3666 return nr_reclaimed;
3670 * This routine traverse page_cgroup in given list and drop them all.
3671 * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3673 static int mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
3674 int node, int zid, enum lru_list lru)
3676 struct mem_cgroup_per_zone *mz;
3677 unsigned long flags, loop;
3678 struct list_head *list;
3683 zone = &NODE_DATA(node)->node_zones[zid];
3684 mz = mem_cgroup_zoneinfo(memcg, node, zid);
3685 list = &mz->lruvec.lists[lru];
3687 loop = mz->lru_size[lru];
3688 /* give some margin against EBUSY etc...*/
3692 struct page_cgroup *pc;
3696 spin_lock_irqsave(&zone->lru_lock, flags);
3697 if (list_empty(list)) {
3698 spin_unlock_irqrestore(&zone->lru_lock, flags);
3701 page = list_entry(list->prev, struct page, lru);
3703 list_move(&page->lru, list);
3705 spin_unlock_irqrestore(&zone->lru_lock, flags);
3708 spin_unlock_irqrestore(&zone->lru_lock, flags);
3710 pc = lookup_page_cgroup(page);
3712 ret = mem_cgroup_move_parent(page, pc, memcg, GFP_KERNEL);
3713 if (ret == -ENOMEM || ret == -EINTR)
3716 if (ret == -EBUSY || ret == -EINVAL) {
3717 /* found lock contention or "pc" is obsolete. */
3724 if (!ret && !list_empty(list))
3730 * make mem_cgroup's charge to be 0 if there is no task.
3731 * This enables deleting this mem_cgroup.
3733 static int mem_cgroup_force_empty(struct mem_cgroup *memcg, bool free_all)
3736 int node, zid, shrink;
3737 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3738 struct cgroup *cgrp = memcg->css.cgroup;
3740 css_get(&memcg->css);
3743 /* should free all ? */
3749 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3752 if (signal_pending(current))
3754 /* This is for making all *used* pages to be on LRU. */
3755 lru_add_drain_all();
3756 drain_all_stock_sync(memcg);
3758 mem_cgroup_start_move(memcg);
3759 for_each_node_state(node, N_HIGH_MEMORY) {
3760 for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3763 ret = mem_cgroup_force_empty_list(memcg,
3772 mem_cgroup_end_move(memcg);
3773 memcg_oom_recover(memcg);
3774 /* it seems parent cgroup doesn't have enough mem */
3778 /* "ret" should also be checked to ensure all lists are empty. */
3779 } while (res_counter_read_u64(&memcg->res, RES_USAGE) > 0 || ret);
3781 css_put(&memcg->css);
3785 /* returns EBUSY if there is a task or if we come here twice. */
3786 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3790 /* we call try-to-free pages for make this cgroup empty */
3791 lru_add_drain_all();
3792 /* try to free all pages in this cgroup */
3794 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
3797 if (signal_pending(current)) {
3801 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
3805 /* maybe some writeback is necessary */
3806 congestion_wait(BLK_RW_ASYNC, HZ/10);
3811 /* try move_account...there may be some *locked* pages. */
3815 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3817 return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3821 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3823 return mem_cgroup_from_cont(cont)->use_hierarchy;
3826 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3830 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3831 struct cgroup *parent = cont->parent;
3832 struct mem_cgroup *parent_memcg = NULL;
3835 parent_memcg = mem_cgroup_from_cont(parent);
3839 * If parent's use_hierarchy is set, we can't make any modifications
3840 * in the child subtrees. If it is unset, then the change can
3841 * occur, provided the current cgroup has no children.
3843 * For the root cgroup, parent_mem is NULL, we allow value to be
3844 * set if there are no children.
3846 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3847 (val == 1 || val == 0)) {
3848 if (list_empty(&cont->children))
3849 memcg->use_hierarchy = val;
3860 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
3861 enum mem_cgroup_stat_index idx)
3863 struct mem_cgroup *iter;
3866 /* Per-cpu values can be negative, use a signed accumulator */
3867 for_each_mem_cgroup_tree(iter, memcg)
3868 val += mem_cgroup_read_stat(iter, idx);
3870 if (val < 0) /* race ? */
3875 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3879 if (!mem_cgroup_is_root(memcg)) {
3881 return res_counter_read_u64(&memcg->res, RES_USAGE);
3883 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
3886 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
3887 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
3890 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
3892 return val << PAGE_SHIFT;
3895 static ssize_t mem_cgroup_read(struct cgroup *cont, struct cftype *cft,
3896 struct file *file, char __user *buf,
3897 size_t nbytes, loff_t *ppos)
3899 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3902 int type, name, len;
3904 type = MEMFILE_TYPE(cft->private);
3905 name = MEMFILE_ATTR(cft->private);
3907 if (!do_swap_account && type == _MEMSWAP)
3912 if (name == RES_USAGE)
3913 val = mem_cgroup_usage(memcg, false);
3915 val = res_counter_read_u64(&memcg->res, name);
3918 if (name == RES_USAGE)
3919 val = mem_cgroup_usage(memcg, true);
3921 val = res_counter_read_u64(&memcg->memsw, name);
3927 len = scnprintf(str, sizeof(str), "%llu\n", (unsigned long long)val);
3928 return simple_read_from_buffer(buf, nbytes, ppos, str, len);
3931 * The user of this function is...
3934 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
3937 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3939 unsigned long long val;
3942 type = MEMFILE_TYPE(cft->private);
3943 name = MEMFILE_ATTR(cft->private);
3945 if (!do_swap_account && type == _MEMSWAP)
3950 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3954 /* This function does all necessary parse...reuse it */
3955 ret = res_counter_memparse_write_strategy(buffer, &val);
3959 ret = mem_cgroup_resize_limit(memcg, val);
3961 ret = mem_cgroup_resize_memsw_limit(memcg, val);
3963 case RES_SOFT_LIMIT:
3964 ret = res_counter_memparse_write_strategy(buffer, &val);
3968 * For memsw, soft limits are hard to implement in terms
3969 * of semantics, for now, we support soft limits for
3970 * control without swap
3973 ret = res_counter_set_soft_limit(&memcg->res, val);
3978 ret = -EINVAL; /* should be BUG() ? */
3984 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
3985 unsigned long long *mem_limit, unsigned long long *memsw_limit)
3987 struct cgroup *cgroup;
3988 unsigned long long min_limit, min_memsw_limit, tmp;
3990 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3991 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3992 cgroup = memcg->css.cgroup;
3993 if (!memcg->use_hierarchy)
3996 while (cgroup->parent) {
3997 cgroup = cgroup->parent;
3998 memcg = mem_cgroup_from_cont(cgroup);
3999 if (!memcg->use_hierarchy)
4001 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
4002 min_limit = min(min_limit, tmp);
4003 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4004 min_memsw_limit = min(min_memsw_limit, tmp);
4007 *mem_limit = min_limit;
4008 *memsw_limit = min_memsw_limit;
4011 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
4013 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4016 type = MEMFILE_TYPE(event);
4017 name = MEMFILE_ATTR(event);
4019 if (!do_swap_account && type == _MEMSWAP)
4025 res_counter_reset_max(&memcg->res);
4027 res_counter_reset_max(&memcg->memsw);
4031 res_counter_reset_failcnt(&memcg->res);
4033 res_counter_reset_failcnt(&memcg->memsw);
4040 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
4043 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
4047 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4048 struct cftype *cft, u64 val)
4050 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4052 if (val >= (1 << NR_MOVE_TYPE))
4055 * We check this value several times in both in can_attach() and
4056 * attach(), so we need cgroup lock to prevent this value from being
4060 memcg->move_charge_at_immigrate = val;
4066 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4067 struct cftype *cft, u64 val)
4074 /* For read statistics */
4092 struct mcs_total_stat {
4093 s64 stat[NR_MCS_STAT];
4099 } memcg_stat_strings[NR_MCS_STAT] = {
4100 {"cache", "total_cache"},
4101 {"rss", "total_rss"},
4102 {"mapped_file", "total_mapped_file"},
4103 {"pgpgin", "total_pgpgin"},
4104 {"pgpgout", "total_pgpgout"},
4105 {"swap", "total_swap"},
4106 {"pgfault", "total_pgfault"},
4107 {"pgmajfault", "total_pgmajfault"},
4108 {"inactive_anon", "total_inactive_anon"},
4109 {"active_anon", "total_active_anon"},
4110 {"inactive_file", "total_inactive_file"},
4111 {"active_file", "total_active_file"},
4112 {"unevictable", "total_unevictable"}
4117 mem_cgroup_get_local_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4122 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_CACHE);
4123 s->stat[MCS_CACHE] += val * PAGE_SIZE;
4124 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_RSS);
4125 s->stat[MCS_RSS] += val * PAGE_SIZE;
4126 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_FILE_MAPPED);
4127 s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
4128 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGIN);
4129 s->stat[MCS_PGPGIN] += val;
4130 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGOUT);
4131 s->stat[MCS_PGPGOUT] += val;
4132 if (do_swap_account) {
4133 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
4134 s->stat[MCS_SWAP] += val * PAGE_SIZE;
4136 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGFAULT);
4137 s->stat[MCS_PGFAULT] += val;
4138 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGMAJFAULT);
4139 s->stat[MCS_PGMAJFAULT] += val;
4142 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_ANON));
4143 s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
4144 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_ANON));
4145 s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
4146 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_FILE));
4147 s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
4148 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_FILE));
4149 s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
4150 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
4151 s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
4155 mem_cgroup_get_total_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4157 struct mem_cgroup *iter;
4159 for_each_mem_cgroup_tree(iter, memcg)
4160 mem_cgroup_get_local_stat(iter, s);
4164 static int mem_control_numa_stat_show(struct seq_file *m, void *arg)
4167 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
4168 unsigned long node_nr;
4169 struct cgroup *cont = m->private;
4170 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4172 total_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL);
4173 seq_printf(m, "total=%lu", total_nr);
4174 for_each_node_state(nid, N_HIGH_MEMORY) {
4175 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL);
4176 seq_printf(m, " N%d=%lu", nid, node_nr);
4180 file_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_FILE);
4181 seq_printf(m, "file=%lu", file_nr);
4182 for_each_node_state(nid, N_HIGH_MEMORY) {
4183 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4185 seq_printf(m, " N%d=%lu", nid, node_nr);
4189 anon_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_ANON);
4190 seq_printf(m, "anon=%lu", anon_nr);
4191 for_each_node_state(nid, N_HIGH_MEMORY) {
4192 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4194 seq_printf(m, " N%d=%lu", nid, node_nr);
4198 unevictable_nr = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
4199 seq_printf(m, "unevictable=%lu", unevictable_nr);
4200 for_each_node_state(nid, N_HIGH_MEMORY) {
4201 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4202 BIT(LRU_UNEVICTABLE));
4203 seq_printf(m, " N%d=%lu", nid, node_nr);
4208 #endif /* CONFIG_NUMA */
4210 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
4211 struct cgroup_map_cb *cb)
4213 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4214 struct mcs_total_stat mystat;
4217 memset(&mystat, 0, sizeof(mystat));
4218 mem_cgroup_get_local_stat(memcg, &mystat);
4221 for (i = 0; i < NR_MCS_STAT; i++) {
4222 if (i == MCS_SWAP && !do_swap_account)
4224 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
4227 /* Hierarchical information */
4229 unsigned long long limit, memsw_limit;
4230 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
4231 cb->fill(cb, "hierarchical_memory_limit", limit);
4232 if (do_swap_account)
4233 cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
4236 memset(&mystat, 0, sizeof(mystat));
4237 mem_cgroup_get_total_stat(memcg, &mystat);
4238 for (i = 0; i < NR_MCS_STAT; i++) {
4239 if (i == MCS_SWAP && !do_swap_account)
4241 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
4244 #ifdef CONFIG_DEBUG_VM
4247 struct mem_cgroup_per_zone *mz;
4248 unsigned long recent_rotated[2] = {0, 0};
4249 unsigned long recent_scanned[2] = {0, 0};
4251 for_each_online_node(nid)
4252 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4253 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
4255 recent_rotated[0] +=
4256 mz->reclaim_stat.recent_rotated[0];
4257 recent_rotated[1] +=
4258 mz->reclaim_stat.recent_rotated[1];
4259 recent_scanned[0] +=
4260 mz->reclaim_stat.recent_scanned[0];
4261 recent_scanned[1] +=
4262 mz->reclaim_stat.recent_scanned[1];
4264 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
4265 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
4266 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
4267 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
4274 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
4276 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4278 return mem_cgroup_swappiness(memcg);
4281 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
4284 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4285 struct mem_cgroup *parent;
4290 if (cgrp->parent == NULL)
4293 parent = mem_cgroup_from_cont(cgrp->parent);
4297 /* If under hierarchy, only empty-root can set this value */
4298 if ((parent->use_hierarchy) ||
4299 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4304 memcg->swappiness = val;
4311 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4313 struct mem_cgroup_threshold_ary *t;
4319 t = rcu_dereference(memcg->thresholds.primary);
4321 t = rcu_dereference(memcg->memsw_thresholds.primary);
4326 usage = mem_cgroup_usage(memcg, swap);
4329 * current_threshold points to threshold just below usage.
4330 * If it's not true, a threshold was crossed after last
4331 * call of __mem_cgroup_threshold().
4333 i = t->current_threshold;
4336 * Iterate backward over array of thresholds starting from
4337 * current_threshold and check if a threshold is crossed.
4338 * If none of thresholds below usage is crossed, we read
4339 * only one element of the array here.
4341 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4342 eventfd_signal(t->entries[i].eventfd, 1);
4344 /* i = current_threshold + 1 */
4348 * Iterate forward over array of thresholds starting from
4349 * current_threshold+1 and check if a threshold is crossed.
4350 * If none of thresholds above usage is crossed, we read
4351 * only one element of the array here.
4353 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4354 eventfd_signal(t->entries[i].eventfd, 1);
4356 /* Update current_threshold */
4357 t->current_threshold = i - 1;
4362 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4365 __mem_cgroup_threshold(memcg, false);
4366 if (do_swap_account)
4367 __mem_cgroup_threshold(memcg, true);
4369 memcg = parent_mem_cgroup(memcg);
4373 static int compare_thresholds(const void *a, const void *b)
4375 const struct mem_cgroup_threshold *_a = a;
4376 const struct mem_cgroup_threshold *_b = b;
4378 return _a->threshold - _b->threshold;
4381 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4383 struct mem_cgroup_eventfd_list *ev;
4385 list_for_each_entry(ev, &memcg->oom_notify, list)
4386 eventfd_signal(ev->eventfd, 1);
4390 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4392 struct mem_cgroup *iter;
4394 for_each_mem_cgroup_tree(iter, memcg)
4395 mem_cgroup_oom_notify_cb(iter);
4398 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
4399 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4401 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4402 struct mem_cgroup_thresholds *thresholds;
4403 struct mem_cgroup_threshold_ary *new;
4404 int type = MEMFILE_TYPE(cft->private);
4405 u64 threshold, usage;
4408 ret = res_counter_memparse_write_strategy(args, &threshold);
4412 mutex_lock(&memcg->thresholds_lock);
4415 thresholds = &memcg->thresholds;
4416 else if (type == _MEMSWAP)
4417 thresholds = &memcg->memsw_thresholds;
4421 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4423 /* Check if a threshold crossed before adding a new one */
4424 if (thresholds->primary)
4425 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4427 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4429 /* Allocate memory for new array of thresholds */
4430 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4438 /* Copy thresholds (if any) to new array */
4439 if (thresholds->primary) {
4440 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4441 sizeof(struct mem_cgroup_threshold));
4444 /* Add new threshold */
4445 new->entries[size - 1].eventfd = eventfd;
4446 new->entries[size - 1].threshold = threshold;
4448 /* Sort thresholds. Registering of new threshold isn't time-critical */
4449 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4450 compare_thresholds, NULL);
4452 /* Find current threshold */
4453 new->current_threshold = -1;
4454 for (i = 0; i < size; i++) {
4455 if (new->entries[i].threshold < usage) {
4457 * new->current_threshold will not be used until
4458 * rcu_assign_pointer(), so it's safe to increment
4461 ++new->current_threshold;
4465 /* Free old spare buffer and save old primary buffer as spare */
4466 kfree(thresholds->spare);
4467 thresholds->spare = thresholds->primary;
4469 rcu_assign_pointer(thresholds->primary, new);
4471 /* To be sure that nobody uses thresholds */
4475 mutex_unlock(&memcg->thresholds_lock);
4480 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
4481 struct cftype *cft, struct eventfd_ctx *eventfd)
4483 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4484 struct mem_cgroup_thresholds *thresholds;
4485 struct mem_cgroup_threshold_ary *new;
4486 int type = MEMFILE_TYPE(cft->private);
4490 mutex_lock(&memcg->thresholds_lock);
4492 thresholds = &memcg->thresholds;
4493 else if (type == _MEMSWAP)
4494 thresholds = &memcg->memsw_thresholds;
4498 if (!thresholds->primary)
4501 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4503 /* Check if a threshold crossed before removing */
4504 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4506 /* Calculate new number of threshold */
4508 for (i = 0; i < thresholds->primary->size; i++) {
4509 if (thresholds->primary->entries[i].eventfd != eventfd)
4513 new = thresholds->spare;
4515 /* Set thresholds array to NULL if we don't have thresholds */
4524 /* Copy thresholds and find current threshold */
4525 new->current_threshold = -1;
4526 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4527 if (thresholds->primary->entries[i].eventfd == eventfd)
4530 new->entries[j] = thresholds->primary->entries[i];
4531 if (new->entries[j].threshold < usage) {
4533 * new->current_threshold will not be used
4534 * until rcu_assign_pointer(), so it's safe to increment
4537 ++new->current_threshold;
4543 /* Swap primary and spare array */
4544 thresholds->spare = thresholds->primary;
4545 /* If all events are unregistered, free the spare array */
4547 kfree(thresholds->spare);
4548 thresholds->spare = NULL;
4551 rcu_assign_pointer(thresholds->primary, new);
4553 /* To be sure that nobody uses thresholds */
4556 mutex_unlock(&memcg->thresholds_lock);
4559 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4560 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4562 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4563 struct mem_cgroup_eventfd_list *event;
4564 int type = MEMFILE_TYPE(cft->private);
4566 BUG_ON(type != _OOM_TYPE);
4567 event = kmalloc(sizeof(*event), GFP_KERNEL);
4571 spin_lock(&memcg_oom_lock);
4573 event->eventfd = eventfd;
4574 list_add(&event->list, &memcg->oom_notify);
4576 /* already in OOM ? */
4577 if (atomic_read(&memcg->under_oom))
4578 eventfd_signal(eventfd, 1);
4579 spin_unlock(&memcg_oom_lock);
4584 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4585 struct cftype *cft, struct eventfd_ctx *eventfd)
4587 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4588 struct mem_cgroup_eventfd_list *ev, *tmp;
4589 int type = MEMFILE_TYPE(cft->private);
4591 BUG_ON(type != _OOM_TYPE);
4593 spin_lock(&memcg_oom_lock);
4595 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4596 if (ev->eventfd == eventfd) {
4597 list_del(&ev->list);
4602 spin_unlock(&memcg_oom_lock);
4605 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4606 struct cftype *cft, struct cgroup_map_cb *cb)
4608 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4610 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
4612 if (atomic_read(&memcg->under_oom))
4613 cb->fill(cb, "under_oom", 1);
4615 cb->fill(cb, "under_oom", 0);
4619 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4620 struct cftype *cft, u64 val)
4622 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4623 struct mem_cgroup *parent;
4625 /* cannot set to root cgroup and only 0 and 1 are allowed */
4626 if (!cgrp->parent || !((val == 0) || (val == 1)))
4629 parent = mem_cgroup_from_cont(cgrp->parent);
4632 /* oom-kill-disable is a flag for subhierarchy. */
4633 if ((parent->use_hierarchy) ||
4634 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4638 memcg->oom_kill_disable = val;
4640 memcg_oom_recover(memcg);
4646 static const struct file_operations mem_control_numa_stat_file_operations = {
4648 .llseek = seq_lseek,
4649 .release = single_release,
4652 static int mem_control_numa_stat_open(struct inode *unused, struct file *file)
4654 struct cgroup *cont = file->f_dentry->d_parent->d_fsdata;
4656 file->f_op = &mem_control_numa_stat_file_operations;
4657 return single_open(file, mem_control_numa_stat_show, cont);
4659 #endif /* CONFIG_NUMA */
4661 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_KMEM
4662 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
4664 return mem_cgroup_sockets_init(memcg, ss);
4667 static void kmem_cgroup_destroy(struct mem_cgroup *memcg)
4669 mem_cgroup_sockets_destroy(memcg);
4672 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
4677 static void kmem_cgroup_destroy(struct mem_cgroup *memcg)
4682 static struct cftype mem_cgroup_files[] = {
4684 .name = "usage_in_bytes",
4685 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4686 .read = mem_cgroup_read,
4687 .register_event = mem_cgroup_usage_register_event,
4688 .unregister_event = mem_cgroup_usage_unregister_event,
4691 .name = "max_usage_in_bytes",
4692 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4693 .trigger = mem_cgroup_reset,
4694 .read = mem_cgroup_read,
4697 .name = "limit_in_bytes",
4698 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4699 .write_string = mem_cgroup_write,
4700 .read = mem_cgroup_read,
4703 .name = "soft_limit_in_bytes",
4704 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4705 .write_string = mem_cgroup_write,
4706 .read = mem_cgroup_read,
4710 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4711 .trigger = mem_cgroup_reset,
4712 .read = mem_cgroup_read,
4716 .read_map = mem_control_stat_show,
4719 .name = "force_empty",
4720 .trigger = mem_cgroup_force_empty_write,
4723 .name = "use_hierarchy",
4724 .write_u64 = mem_cgroup_hierarchy_write,
4725 .read_u64 = mem_cgroup_hierarchy_read,
4728 .name = "swappiness",
4729 .read_u64 = mem_cgroup_swappiness_read,
4730 .write_u64 = mem_cgroup_swappiness_write,
4733 .name = "move_charge_at_immigrate",
4734 .read_u64 = mem_cgroup_move_charge_read,
4735 .write_u64 = mem_cgroup_move_charge_write,
4738 .name = "oom_control",
4739 .read_map = mem_cgroup_oom_control_read,
4740 .write_u64 = mem_cgroup_oom_control_write,
4741 .register_event = mem_cgroup_oom_register_event,
4742 .unregister_event = mem_cgroup_oom_unregister_event,
4743 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4747 .name = "numa_stat",
4748 .open = mem_control_numa_stat_open,
4752 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4754 .name = "memsw.usage_in_bytes",
4755 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4756 .read = mem_cgroup_read,
4757 .register_event = mem_cgroup_usage_register_event,
4758 .unregister_event = mem_cgroup_usage_unregister_event,
4761 .name = "memsw.max_usage_in_bytes",
4762 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4763 .trigger = mem_cgroup_reset,
4764 .read = mem_cgroup_read,
4767 .name = "memsw.limit_in_bytes",
4768 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4769 .write_string = mem_cgroup_write,
4770 .read = mem_cgroup_read,
4773 .name = "memsw.failcnt",
4774 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4775 .trigger = mem_cgroup_reset,
4776 .read = mem_cgroup_read,
4779 { }, /* terminate */
4782 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4784 struct mem_cgroup_per_node *pn;
4785 struct mem_cgroup_per_zone *mz;
4787 int zone, tmp = node;
4789 * This routine is called against possible nodes.
4790 * But it's BUG to call kmalloc() against offline node.
4792 * TODO: this routine can waste much memory for nodes which will
4793 * never be onlined. It's better to use memory hotplug callback
4796 if (!node_state(node, N_NORMAL_MEMORY))
4798 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4802 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4803 mz = &pn->zoneinfo[zone];
4805 INIT_LIST_HEAD(&mz->lruvec.lists[lru]);
4806 mz->usage_in_excess = 0;
4807 mz->on_tree = false;
4810 memcg->info.nodeinfo[node] = pn;
4814 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4816 kfree(memcg->info.nodeinfo[node]);
4819 static struct mem_cgroup *mem_cgroup_alloc(void)
4821 struct mem_cgroup *memcg;
4822 int size = sizeof(struct mem_cgroup);
4824 /* Can be very big if MAX_NUMNODES is very big */
4825 if (size < PAGE_SIZE)
4826 memcg = kzalloc(size, GFP_KERNEL);
4828 memcg = vzalloc(size);
4833 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4836 spin_lock_init(&memcg->pcp_counter_lock);
4840 if (size < PAGE_SIZE)
4848 * Helpers for freeing a vzalloc()ed mem_cgroup by RCU,
4849 * but in process context. The work_freeing structure is overlaid
4850 * on the rcu_freeing structure, which itself is overlaid on memsw.
4852 static void vfree_work(struct work_struct *work)
4854 struct mem_cgroup *memcg;
4856 memcg = container_of(work, struct mem_cgroup, work_freeing);
4859 static void vfree_rcu(struct rcu_head *rcu_head)
4861 struct mem_cgroup *memcg;
4863 memcg = container_of(rcu_head, struct mem_cgroup, rcu_freeing);
4864 INIT_WORK(&memcg->work_freeing, vfree_work);
4865 schedule_work(&memcg->work_freeing);
4869 * At destroying mem_cgroup, references from swap_cgroup can remain.
4870 * (scanning all at force_empty is too costly...)
4872 * Instead of clearing all references at force_empty, we remember
4873 * the number of reference from swap_cgroup and free mem_cgroup when
4874 * it goes down to 0.
4876 * Removal of cgroup itself succeeds regardless of refs from swap.
4879 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4883 mem_cgroup_remove_from_trees(memcg);
4884 free_css_id(&mem_cgroup_subsys, &memcg->css);
4887 free_mem_cgroup_per_zone_info(memcg, node);
4889 free_percpu(memcg->stat);
4890 if (sizeof(struct mem_cgroup) < PAGE_SIZE)
4891 kfree_rcu(memcg, rcu_freeing);
4893 call_rcu(&memcg->rcu_freeing, vfree_rcu);
4896 static void mem_cgroup_get(struct mem_cgroup *memcg)
4898 atomic_inc(&memcg->refcnt);
4901 static void __mem_cgroup_put(struct mem_cgroup *memcg, int count)
4903 if (atomic_sub_and_test(count, &memcg->refcnt)) {
4904 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4905 __mem_cgroup_free(memcg);
4907 mem_cgroup_put(parent);
4911 static void mem_cgroup_put(struct mem_cgroup *memcg)
4913 __mem_cgroup_put(memcg, 1);
4917 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4919 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4921 if (!memcg->res.parent)
4923 return mem_cgroup_from_res_counter(memcg->res.parent, res);
4925 EXPORT_SYMBOL(parent_mem_cgroup);
4927 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4928 static void __init enable_swap_cgroup(void)
4930 if (!mem_cgroup_disabled() && really_do_swap_account)
4931 do_swap_account = 1;
4934 static void __init enable_swap_cgroup(void)
4939 static int mem_cgroup_soft_limit_tree_init(void)
4941 struct mem_cgroup_tree_per_node *rtpn;
4942 struct mem_cgroup_tree_per_zone *rtpz;
4943 int tmp, node, zone;
4945 for_each_node(node) {
4947 if (!node_state(node, N_NORMAL_MEMORY))
4949 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
4953 soft_limit_tree.rb_tree_per_node[node] = rtpn;
4955 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4956 rtpz = &rtpn->rb_tree_per_zone[zone];
4957 rtpz->rb_root = RB_ROOT;
4958 spin_lock_init(&rtpz->lock);
4964 for_each_node(node) {
4965 if (!soft_limit_tree.rb_tree_per_node[node])
4967 kfree(soft_limit_tree.rb_tree_per_node[node]);
4968 soft_limit_tree.rb_tree_per_node[node] = NULL;
4974 static struct cgroup_subsys_state * __ref
4975 mem_cgroup_create(struct cgroup *cont)
4977 struct mem_cgroup *memcg, *parent;
4978 long error = -ENOMEM;
4981 memcg = mem_cgroup_alloc();
4983 return ERR_PTR(error);
4986 if (alloc_mem_cgroup_per_zone_info(memcg, node))
4990 if (cont->parent == NULL) {
4992 enable_swap_cgroup();
4994 if (mem_cgroup_soft_limit_tree_init())
4996 root_mem_cgroup = memcg;
4997 for_each_possible_cpu(cpu) {
4998 struct memcg_stock_pcp *stock =
4999 &per_cpu(memcg_stock, cpu);
5000 INIT_WORK(&stock->work, drain_local_stock);
5002 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5004 parent = mem_cgroup_from_cont(cont->parent);
5005 memcg->use_hierarchy = parent->use_hierarchy;
5006 memcg->oom_kill_disable = parent->oom_kill_disable;
5009 if (parent && parent->use_hierarchy) {
5010 res_counter_init(&memcg->res, &parent->res);
5011 res_counter_init(&memcg->memsw, &parent->memsw);
5013 * We increment refcnt of the parent to ensure that we can
5014 * safely access it on res_counter_charge/uncharge.
5015 * This refcnt will be decremented when freeing this
5016 * mem_cgroup(see mem_cgroup_put).
5018 mem_cgroup_get(parent);
5020 res_counter_init(&memcg->res, NULL);
5021 res_counter_init(&memcg->memsw, NULL);
5023 memcg->last_scanned_node = MAX_NUMNODES;
5024 INIT_LIST_HEAD(&memcg->oom_notify);
5027 memcg->swappiness = mem_cgroup_swappiness(parent);
5028 atomic_set(&memcg->refcnt, 1);
5029 memcg->move_charge_at_immigrate = 0;
5030 mutex_init(&memcg->thresholds_lock);
5031 spin_lock_init(&memcg->move_lock);
5033 error = memcg_init_kmem(memcg, &mem_cgroup_subsys);
5036 * We call put now because our (and parent's) refcnts
5037 * are already in place. mem_cgroup_put() will internally
5038 * call __mem_cgroup_free, so return directly
5040 mem_cgroup_put(memcg);
5041 return ERR_PTR(error);
5045 __mem_cgroup_free(memcg);
5046 return ERR_PTR(error);
5049 static int mem_cgroup_pre_destroy(struct cgroup *cont)
5051 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5053 return mem_cgroup_force_empty(memcg, false);
5056 static void mem_cgroup_destroy(struct cgroup *cont)
5058 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5060 kmem_cgroup_destroy(memcg);
5062 mem_cgroup_put(memcg);
5066 /* Handlers for move charge at task migration. */
5067 #define PRECHARGE_COUNT_AT_ONCE 256
5068 static int mem_cgroup_do_precharge(unsigned long count)
5071 int batch_count = PRECHARGE_COUNT_AT_ONCE;
5072 struct mem_cgroup *memcg = mc.to;
5074 if (mem_cgroup_is_root(memcg)) {
5075 mc.precharge += count;
5076 /* we don't need css_get for root */
5079 /* try to charge at once */
5081 struct res_counter *dummy;
5083 * "memcg" cannot be under rmdir() because we've already checked
5084 * by cgroup_lock_live_cgroup() that it is not removed and we
5085 * are still under the same cgroup_mutex. So we can postpone
5088 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
5090 if (do_swap_account && res_counter_charge(&memcg->memsw,
5091 PAGE_SIZE * count, &dummy)) {
5092 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
5095 mc.precharge += count;
5099 /* fall back to one by one charge */
5101 if (signal_pending(current)) {
5105 if (!batch_count--) {
5106 batch_count = PRECHARGE_COUNT_AT_ONCE;
5109 ret = __mem_cgroup_try_charge(NULL,
5110 GFP_KERNEL, 1, &memcg, false);
5112 /* mem_cgroup_clear_mc() will do uncharge later */
5120 * get_mctgt_type - get target type of moving charge
5121 * @vma: the vma the pte to be checked belongs
5122 * @addr: the address corresponding to the pte to be checked
5123 * @ptent: the pte to be checked
5124 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5127 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5128 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5129 * move charge. if @target is not NULL, the page is stored in target->page
5130 * with extra refcnt got(Callers should handle it).
5131 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5132 * target for charge migration. if @target is not NULL, the entry is stored
5135 * Called with pte lock held.
5142 enum mc_target_type {
5148 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5149 unsigned long addr, pte_t ptent)
5151 struct page *page = vm_normal_page(vma, addr, ptent);
5153 if (!page || !page_mapped(page))
5155 if (PageAnon(page)) {
5156 /* we don't move shared anon */
5157 if (!move_anon() || page_mapcount(page) > 2)
5159 } else if (!move_file())
5160 /* we ignore mapcount for file pages */
5162 if (!get_page_unless_zero(page))
5168 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5169 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5172 struct page *page = NULL;
5173 swp_entry_t ent = pte_to_swp_entry(ptent);
5175 if (!move_anon() || non_swap_entry(ent))
5177 usage_count = mem_cgroup_count_swap_user(ent, &page);
5178 if (usage_count > 1) { /* we don't move shared anon */
5183 if (do_swap_account)
5184 entry->val = ent.val;
5189 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5190 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5192 struct page *page = NULL;
5193 struct inode *inode;
5194 struct address_space *mapping;
5197 if (!vma->vm_file) /* anonymous vma */
5202 inode = vma->vm_file->f_path.dentry->d_inode;
5203 mapping = vma->vm_file->f_mapping;
5204 if (pte_none(ptent))
5205 pgoff = linear_page_index(vma, addr);
5206 else /* pte_file(ptent) is true */
5207 pgoff = pte_to_pgoff(ptent);
5209 /* page is moved even if it's not RSS of this task(page-faulted). */
5210 page = find_get_page(mapping, pgoff);
5213 /* shmem/tmpfs may report page out on swap: account for that too. */
5214 if (radix_tree_exceptional_entry(page)) {
5215 swp_entry_t swap = radix_to_swp_entry(page);
5216 if (do_swap_account)
5218 page = find_get_page(&swapper_space, swap.val);
5224 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5225 unsigned long addr, pte_t ptent, union mc_target *target)
5227 struct page *page = NULL;
5228 struct page_cgroup *pc;
5229 enum mc_target_type ret = MC_TARGET_NONE;
5230 swp_entry_t ent = { .val = 0 };
5232 if (pte_present(ptent))
5233 page = mc_handle_present_pte(vma, addr, ptent);
5234 else if (is_swap_pte(ptent))
5235 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5236 else if (pte_none(ptent) || pte_file(ptent))
5237 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5239 if (!page && !ent.val)
5242 pc = lookup_page_cgroup(page);
5244 * Do only loose check w/o page_cgroup lock.
5245 * mem_cgroup_move_account() checks the pc is valid or not under
5248 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5249 ret = MC_TARGET_PAGE;
5251 target->page = page;
5253 if (!ret || !target)
5256 /* There is a swap entry and a page doesn't exist or isn't charged */
5257 if (ent.val && !ret &&
5258 css_id(&mc.from->css) == lookup_swap_cgroup_id(ent)) {
5259 ret = MC_TARGET_SWAP;
5266 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5268 * We don't consider swapping or file mapped pages because THP does not
5269 * support them for now.
5270 * Caller should make sure that pmd_trans_huge(pmd) is true.
5272 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5273 unsigned long addr, pmd_t pmd, union mc_target *target)
5275 struct page *page = NULL;
5276 struct page_cgroup *pc;
5277 enum mc_target_type ret = MC_TARGET_NONE;
5279 page = pmd_page(pmd);
5280 VM_BUG_ON(!page || !PageHead(page));
5283 pc = lookup_page_cgroup(page);
5284 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5285 ret = MC_TARGET_PAGE;
5288 target->page = page;
5294 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5295 unsigned long addr, pmd_t pmd, union mc_target *target)
5297 return MC_TARGET_NONE;
5301 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5302 unsigned long addr, unsigned long end,
5303 struct mm_walk *walk)
5305 struct vm_area_struct *vma = walk->private;
5309 if (pmd_trans_huge_lock(pmd, vma) == 1) {
5310 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5311 mc.precharge += HPAGE_PMD_NR;
5312 spin_unlock(&vma->vm_mm->page_table_lock);
5316 if (pmd_trans_unstable(pmd))
5318 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5319 for (; addr != end; pte++, addr += PAGE_SIZE)
5320 if (get_mctgt_type(vma, addr, *pte, NULL))
5321 mc.precharge++; /* increment precharge temporarily */
5322 pte_unmap_unlock(pte - 1, ptl);
5328 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5330 unsigned long precharge;
5331 struct vm_area_struct *vma;
5333 down_read(&mm->mmap_sem);
5334 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5335 struct mm_walk mem_cgroup_count_precharge_walk = {
5336 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5340 if (is_vm_hugetlb_page(vma))
5342 walk_page_range(vma->vm_start, vma->vm_end,
5343 &mem_cgroup_count_precharge_walk);
5345 up_read(&mm->mmap_sem);
5347 precharge = mc.precharge;
5353 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5355 unsigned long precharge = mem_cgroup_count_precharge(mm);
5357 VM_BUG_ON(mc.moving_task);
5358 mc.moving_task = current;
5359 return mem_cgroup_do_precharge(precharge);
5362 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5363 static void __mem_cgroup_clear_mc(void)
5365 struct mem_cgroup *from = mc.from;
5366 struct mem_cgroup *to = mc.to;
5368 /* we must uncharge all the leftover precharges from mc.to */
5370 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
5374 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5375 * we must uncharge here.
5377 if (mc.moved_charge) {
5378 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
5379 mc.moved_charge = 0;
5381 /* we must fixup refcnts and charges */
5382 if (mc.moved_swap) {
5383 /* uncharge swap account from the old cgroup */
5384 if (!mem_cgroup_is_root(mc.from))
5385 res_counter_uncharge(&mc.from->memsw,
5386 PAGE_SIZE * mc.moved_swap);
5387 __mem_cgroup_put(mc.from, mc.moved_swap);
5389 if (!mem_cgroup_is_root(mc.to)) {
5391 * we charged both to->res and to->memsw, so we should
5394 res_counter_uncharge(&mc.to->res,
5395 PAGE_SIZE * mc.moved_swap);
5397 /* we've already done mem_cgroup_get(mc.to) */
5400 memcg_oom_recover(from);
5401 memcg_oom_recover(to);
5402 wake_up_all(&mc.waitq);
5405 static void mem_cgroup_clear_mc(void)
5407 struct mem_cgroup *from = mc.from;
5410 * we must clear moving_task before waking up waiters at the end of
5413 mc.moving_task = NULL;
5414 __mem_cgroup_clear_mc();
5415 spin_lock(&mc.lock);
5418 spin_unlock(&mc.lock);
5419 mem_cgroup_end_move(from);
5422 static int mem_cgroup_can_attach(struct cgroup *cgroup,
5423 struct cgroup_taskset *tset)
5425 struct task_struct *p = cgroup_taskset_first(tset);
5427 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup);
5429 if (memcg->move_charge_at_immigrate) {
5430 struct mm_struct *mm;
5431 struct mem_cgroup *from = mem_cgroup_from_task(p);
5433 VM_BUG_ON(from == memcg);
5435 mm = get_task_mm(p);
5438 /* We move charges only when we move a owner of the mm */
5439 if (mm->owner == p) {
5442 VM_BUG_ON(mc.precharge);
5443 VM_BUG_ON(mc.moved_charge);
5444 VM_BUG_ON(mc.moved_swap);
5445 mem_cgroup_start_move(from);
5446 spin_lock(&mc.lock);
5449 spin_unlock(&mc.lock);
5450 /* We set mc.moving_task later */
5452 ret = mem_cgroup_precharge_mc(mm);
5454 mem_cgroup_clear_mc();
5461 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
5462 struct cgroup_taskset *tset)
5464 mem_cgroup_clear_mc();
5467 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5468 unsigned long addr, unsigned long end,
5469 struct mm_walk *walk)
5472 struct vm_area_struct *vma = walk->private;
5475 enum mc_target_type target_type;
5476 union mc_target target;
5478 struct page_cgroup *pc;
5481 * We don't take compound_lock() here but no race with splitting thp
5483 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
5484 * under splitting, which means there's no concurrent thp split,
5485 * - if another thread runs into split_huge_page() just after we
5486 * entered this if-block, the thread must wait for page table lock
5487 * to be unlocked in __split_huge_page_splitting(), where the main
5488 * part of thp split is not executed yet.
5490 if (pmd_trans_huge_lock(pmd, vma) == 1) {
5491 if (mc.precharge < HPAGE_PMD_NR) {
5492 spin_unlock(&vma->vm_mm->page_table_lock);
5495 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5496 if (target_type == MC_TARGET_PAGE) {
5498 if (!isolate_lru_page(page)) {
5499 pc = lookup_page_cgroup(page);
5500 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
5503 mc.precharge -= HPAGE_PMD_NR;
5504 mc.moved_charge += HPAGE_PMD_NR;
5506 putback_lru_page(page);
5510 spin_unlock(&vma->vm_mm->page_table_lock);
5514 if (pmd_trans_unstable(pmd))
5517 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5518 for (; addr != end; addr += PAGE_SIZE) {
5519 pte_t ptent = *(pte++);
5525 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5526 case MC_TARGET_PAGE:
5528 if (isolate_lru_page(page))
5530 pc = lookup_page_cgroup(page);
5531 if (!mem_cgroup_move_account(page, 1, pc,
5532 mc.from, mc.to, false)) {
5534 /* we uncharge from mc.from later. */
5537 putback_lru_page(page);
5538 put: /* get_mctgt_type() gets the page */
5541 case MC_TARGET_SWAP:
5543 if (!mem_cgroup_move_swap_account(ent,
5544 mc.from, mc.to, false)) {
5546 /* we fixup refcnts and charges later. */
5554 pte_unmap_unlock(pte - 1, ptl);
5559 * We have consumed all precharges we got in can_attach().
5560 * We try charge one by one, but don't do any additional
5561 * charges to mc.to if we have failed in charge once in attach()
5564 ret = mem_cgroup_do_precharge(1);
5572 static void mem_cgroup_move_charge(struct mm_struct *mm)
5574 struct vm_area_struct *vma;
5576 lru_add_drain_all();
5578 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5580 * Someone who are holding the mmap_sem might be waiting in
5581 * waitq. So we cancel all extra charges, wake up all waiters,
5582 * and retry. Because we cancel precharges, we might not be able
5583 * to move enough charges, but moving charge is a best-effort
5584 * feature anyway, so it wouldn't be a big problem.
5586 __mem_cgroup_clear_mc();
5590 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5592 struct mm_walk mem_cgroup_move_charge_walk = {
5593 .pmd_entry = mem_cgroup_move_charge_pte_range,
5597 if (is_vm_hugetlb_page(vma))
5599 ret = walk_page_range(vma->vm_start, vma->vm_end,
5600 &mem_cgroup_move_charge_walk);
5603 * means we have consumed all precharges and failed in
5604 * doing additional charge. Just abandon here.
5608 up_read(&mm->mmap_sem);
5611 static void mem_cgroup_move_task(struct cgroup *cont,
5612 struct cgroup_taskset *tset)
5614 struct task_struct *p = cgroup_taskset_first(tset);
5615 struct mm_struct *mm = get_task_mm(p);
5619 mem_cgroup_move_charge(mm);
5623 mem_cgroup_clear_mc();
5625 #else /* !CONFIG_MMU */
5626 static int mem_cgroup_can_attach(struct cgroup *cgroup,
5627 struct cgroup_taskset *tset)
5631 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
5632 struct cgroup_taskset *tset)
5635 static void mem_cgroup_move_task(struct cgroup *cont,
5636 struct cgroup_taskset *tset)
5641 struct cgroup_subsys mem_cgroup_subsys = {
5643 .subsys_id = mem_cgroup_subsys_id,
5644 .create = mem_cgroup_create,
5645 .pre_destroy = mem_cgroup_pre_destroy,
5646 .destroy = mem_cgroup_destroy,
5647 .can_attach = mem_cgroup_can_attach,
5648 .cancel_attach = mem_cgroup_cancel_attach,
5649 .attach = mem_cgroup_move_task,
5650 .base_cftypes = mem_cgroup_files,
5653 .__DEPRECATED_clear_css_refs = true,
5656 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5657 static int __init enable_swap_account(char *s)
5659 /* consider enabled if no parameter or 1 is given */
5660 if (!strcmp(s, "1"))
5661 really_do_swap_account = 1;
5662 else if (!strcmp(s, "0"))
5663 really_do_swap_account = 0;
5666 __setup("swapaccount=", enable_swap_account);