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 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1153 struct mem_cgroup *memcg)
1155 if (root_memcg != memcg) {
1156 return (root_memcg->use_hierarchy &&
1157 css_is_ancestor(&memcg->css, &root_memcg->css));
1163 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *memcg)
1166 struct mem_cgroup *curr = NULL;
1167 struct task_struct *p;
1169 p = find_lock_task_mm(task);
1171 curr = try_get_mem_cgroup_from_mm(p->mm);
1175 * All threads may have already detached their mm's, but the oom
1176 * killer still needs to detect if they have already been oom
1177 * killed to prevent needlessly killing additional tasks.
1180 curr = mem_cgroup_from_task(task);
1182 css_get(&curr->css);
1188 * We should check use_hierarchy of "memcg" not "curr". Because checking
1189 * use_hierarchy of "curr" here make this function true if hierarchy is
1190 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1191 * hierarchy(even if use_hierarchy is disabled in "memcg").
1193 ret = mem_cgroup_same_or_subtree(memcg, curr);
1194 css_put(&curr->css);
1198 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg, struct zone *zone)
1200 unsigned long inactive_ratio;
1201 int nid = zone_to_nid(zone);
1202 int zid = zone_idx(zone);
1203 unsigned long inactive;
1204 unsigned long active;
1207 inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1208 BIT(LRU_INACTIVE_ANON));
1209 active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1210 BIT(LRU_ACTIVE_ANON));
1212 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1214 inactive_ratio = int_sqrt(10 * gb);
1218 return inactive * inactive_ratio < active;
1221 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg, struct zone *zone)
1223 unsigned long active;
1224 unsigned long inactive;
1225 int zid = zone_idx(zone);
1226 int nid = zone_to_nid(zone);
1228 inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1229 BIT(LRU_INACTIVE_FILE));
1230 active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1231 BIT(LRU_ACTIVE_FILE));
1233 return (active > inactive);
1236 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
1239 int nid = zone_to_nid(zone);
1240 int zid = zone_idx(zone);
1241 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
1243 return &mz->reclaim_stat;
1246 struct zone_reclaim_stat *
1247 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
1249 struct page_cgroup *pc;
1250 struct mem_cgroup_per_zone *mz;
1252 if (mem_cgroup_disabled())
1255 pc = lookup_page_cgroup(page);
1256 if (!PageCgroupUsed(pc))
1258 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1260 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1261 return &mz->reclaim_stat;
1264 #define mem_cgroup_from_res_counter(counter, member) \
1265 container_of(counter, struct mem_cgroup, member)
1268 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1269 * @mem: the memory cgroup
1271 * Returns the maximum amount of memory @mem can be charged with, in
1274 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1276 unsigned long long margin;
1278 margin = res_counter_margin(&memcg->res);
1279 if (do_swap_account)
1280 margin = min(margin, res_counter_margin(&memcg->memsw));
1281 return margin >> PAGE_SHIFT;
1284 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1286 struct cgroup *cgrp = memcg->css.cgroup;
1289 if (cgrp->parent == NULL)
1290 return vm_swappiness;
1292 return memcg->swappiness;
1296 * memcg->moving_account is used for checking possibility that some thread is
1297 * calling move_account(). When a thread on CPU-A starts moving pages under
1298 * a memcg, other threads should check memcg->moving_account under
1299 * rcu_read_lock(), like this:
1303 * memcg->moving_account+1 if (memcg->mocing_account)
1305 * synchronize_rcu() update something.
1309 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1311 atomic_inc(&memcg->moving_account);
1315 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1318 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1319 * We check NULL in callee rather than caller.
1322 atomic_dec(&memcg->moving_account);
1326 * 2 routines for checking "mem" is under move_account() or not.
1328 * mem_cgroup_stealed() - checking a cgroup is mc.from or not. This is used
1329 * for avoiding race in accounting. If true,
1330 * pc->mem_cgroup may be overwritten.
1332 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1333 * under hierarchy of moving cgroups. This is for
1334 * waiting at hith-memory prressure caused by "move".
1337 static bool mem_cgroup_stealed(struct mem_cgroup *memcg)
1339 VM_BUG_ON(!rcu_read_lock_held());
1340 return atomic_read(&memcg->moving_account) > 0;
1343 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1345 struct mem_cgroup *from;
1346 struct mem_cgroup *to;
1349 * Unlike task_move routines, we access mc.to, mc.from not under
1350 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1352 spin_lock(&mc.lock);
1358 ret = mem_cgroup_same_or_subtree(memcg, from)
1359 || mem_cgroup_same_or_subtree(memcg, to);
1361 spin_unlock(&mc.lock);
1365 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1367 if (mc.moving_task && current != mc.moving_task) {
1368 if (mem_cgroup_under_move(memcg)) {
1370 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1371 /* moving charge context might have finished. */
1374 finish_wait(&mc.waitq, &wait);
1382 * Take this lock when
1383 * - a code tries to modify page's memcg while it's USED.
1384 * - a code tries to modify page state accounting in a memcg.
1385 * see mem_cgroup_stealed(), too.
1387 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1388 unsigned long *flags)
1390 spin_lock_irqsave(&memcg->move_lock, *flags);
1393 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1394 unsigned long *flags)
1396 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1400 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1401 * @memcg: The memory cgroup that went over limit
1402 * @p: Task that is going to be killed
1404 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1407 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1409 struct cgroup *task_cgrp;
1410 struct cgroup *mem_cgrp;
1412 * Need a buffer in BSS, can't rely on allocations. The code relies
1413 * on the assumption that OOM is serialized for memory controller.
1414 * If this assumption is broken, revisit this code.
1416 static char memcg_name[PATH_MAX];
1424 mem_cgrp = memcg->css.cgroup;
1425 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1427 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1430 * Unfortunately, we are unable to convert to a useful name
1431 * But we'll still print out the usage information
1438 printk(KERN_INFO "Task in %s killed", memcg_name);
1441 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1449 * Continues from above, so we don't need an KERN_ level
1451 printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1454 printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1455 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1456 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1457 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1458 printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1460 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1461 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1462 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1466 * This function returns the number of memcg under hierarchy tree. Returns
1467 * 1(self count) if no children.
1469 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1472 struct mem_cgroup *iter;
1474 for_each_mem_cgroup_tree(iter, memcg)
1480 * Return the memory (and swap, if configured) limit for a memcg.
1482 u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1487 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1488 limit += total_swap_pages << PAGE_SHIFT;
1490 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1492 * If memsw is finite and limits the amount of swap space available
1493 * to this memcg, return that limit.
1495 return min(limit, memsw);
1498 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1500 unsigned long flags)
1502 unsigned long total = 0;
1503 bool noswap = false;
1506 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1508 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1511 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1513 drain_all_stock_async(memcg);
1514 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1516 * Allow limit shrinkers, which are triggered directly
1517 * by userspace, to catch signals and stop reclaim
1518 * after minimal progress, regardless of the margin.
1520 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1522 if (mem_cgroup_margin(memcg))
1525 * If nothing was reclaimed after two attempts, there
1526 * may be no reclaimable pages in this hierarchy.
1535 * test_mem_cgroup_node_reclaimable
1536 * @mem: the target memcg
1537 * @nid: the node ID to be checked.
1538 * @noswap : specify true here if the user wants flle only information.
1540 * This function returns whether the specified memcg contains any
1541 * reclaimable pages on a node. Returns true if there are any reclaimable
1542 * pages in the node.
1544 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1545 int nid, bool noswap)
1547 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1549 if (noswap || !total_swap_pages)
1551 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1556 #if MAX_NUMNODES > 1
1559 * Always updating the nodemask is not very good - even if we have an empty
1560 * list or the wrong list here, we can start from some node and traverse all
1561 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1564 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1568 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1569 * pagein/pageout changes since the last update.
1571 if (!atomic_read(&memcg->numainfo_events))
1573 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1576 /* make a nodemask where this memcg uses memory from */
1577 memcg->scan_nodes = node_states[N_HIGH_MEMORY];
1579 for_each_node_mask(nid, node_states[N_HIGH_MEMORY]) {
1581 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1582 node_clear(nid, memcg->scan_nodes);
1585 atomic_set(&memcg->numainfo_events, 0);
1586 atomic_set(&memcg->numainfo_updating, 0);
1590 * Selecting a node where we start reclaim from. Because what we need is just
1591 * reducing usage counter, start from anywhere is O,K. Considering
1592 * memory reclaim from current node, there are pros. and cons.
1594 * Freeing memory from current node means freeing memory from a node which
1595 * we'll use or we've used. So, it may make LRU bad. And if several threads
1596 * hit limits, it will see a contention on a node. But freeing from remote
1597 * node means more costs for memory reclaim because of memory latency.
1599 * Now, we use round-robin. Better algorithm is welcomed.
1601 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1605 mem_cgroup_may_update_nodemask(memcg);
1606 node = memcg->last_scanned_node;
1608 node = next_node(node, memcg->scan_nodes);
1609 if (node == MAX_NUMNODES)
1610 node = first_node(memcg->scan_nodes);
1612 * We call this when we hit limit, not when pages are added to LRU.
1613 * No LRU may hold pages because all pages are UNEVICTABLE or
1614 * memcg is too small and all pages are not on LRU. In that case,
1615 * we use curret node.
1617 if (unlikely(node == MAX_NUMNODES))
1618 node = numa_node_id();
1620 memcg->last_scanned_node = node;
1625 * Check all nodes whether it contains reclaimable pages or not.
1626 * For quick scan, we make use of scan_nodes. This will allow us to skip
1627 * unused nodes. But scan_nodes is lazily updated and may not cotain
1628 * enough new information. We need to do double check.
1630 bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1635 * quick check...making use of scan_node.
1636 * We can skip unused nodes.
1638 if (!nodes_empty(memcg->scan_nodes)) {
1639 for (nid = first_node(memcg->scan_nodes);
1641 nid = next_node(nid, memcg->scan_nodes)) {
1643 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1648 * Check rest of nodes.
1650 for_each_node_state(nid, N_HIGH_MEMORY) {
1651 if (node_isset(nid, memcg->scan_nodes))
1653 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1660 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1665 bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1667 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
1671 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1674 unsigned long *total_scanned)
1676 struct mem_cgroup *victim = NULL;
1679 unsigned long excess;
1680 unsigned long nr_scanned;
1681 struct mem_cgroup_reclaim_cookie reclaim = {
1686 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
1689 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1694 * If we have not been able to reclaim
1695 * anything, it might because there are
1696 * no reclaimable pages under this hierarchy
1701 * We want to do more targeted reclaim.
1702 * excess >> 2 is not to excessive so as to
1703 * reclaim too much, nor too less that we keep
1704 * coming back to reclaim from this cgroup
1706 if (total >= (excess >> 2) ||
1707 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1712 if (!mem_cgroup_reclaimable(victim, false))
1714 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1716 *total_scanned += nr_scanned;
1717 if (!res_counter_soft_limit_excess(&root_memcg->res))
1720 mem_cgroup_iter_break(root_memcg, victim);
1725 * Check OOM-Killer is already running under our hierarchy.
1726 * If someone is running, return false.
1727 * Has to be called with memcg_oom_lock
1729 static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg)
1731 struct mem_cgroup *iter, *failed = NULL;
1733 for_each_mem_cgroup_tree(iter, memcg) {
1734 if (iter->oom_lock) {
1736 * this subtree of our hierarchy is already locked
1737 * so we cannot give a lock.
1740 mem_cgroup_iter_break(memcg, iter);
1743 iter->oom_lock = true;
1750 * OK, we failed to lock the whole subtree so we have to clean up
1751 * what we set up to the failing subtree
1753 for_each_mem_cgroup_tree(iter, memcg) {
1754 if (iter == failed) {
1755 mem_cgroup_iter_break(memcg, iter);
1758 iter->oom_lock = false;
1764 * Has to be called with memcg_oom_lock
1766 static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1768 struct mem_cgroup *iter;
1770 for_each_mem_cgroup_tree(iter, memcg)
1771 iter->oom_lock = false;
1775 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1777 struct mem_cgroup *iter;
1779 for_each_mem_cgroup_tree(iter, memcg)
1780 atomic_inc(&iter->under_oom);
1783 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1785 struct mem_cgroup *iter;
1788 * When a new child is created while the hierarchy is under oom,
1789 * mem_cgroup_oom_lock() may not be called. We have to use
1790 * atomic_add_unless() here.
1792 for_each_mem_cgroup_tree(iter, memcg)
1793 atomic_add_unless(&iter->under_oom, -1, 0);
1796 static DEFINE_SPINLOCK(memcg_oom_lock);
1797 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1799 struct oom_wait_info {
1800 struct mem_cgroup *memcg;
1804 static int memcg_oom_wake_function(wait_queue_t *wait,
1805 unsigned mode, int sync, void *arg)
1807 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1808 struct mem_cgroup *oom_wait_memcg;
1809 struct oom_wait_info *oom_wait_info;
1811 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1812 oom_wait_memcg = oom_wait_info->memcg;
1815 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
1816 * Then we can use css_is_ancestor without taking care of RCU.
1818 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
1819 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
1821 return autoremove_wake_function(wait, mode, sync, arg);
1824 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
1826 /* for filtering, pass "memcg" as argument. */
1827 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1830 static void memcg_oom_recover(struct mem_cgroup *memcg)
1832 if (memcg && atomic_read(&memcg->under_oom))
1833 memcg_wakeup_oom(memcg);
1837 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1839 bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1841 struct oom_wait_info owait;
1842 bool locked, need_to_kill;
1844 owait.memcg = memcg;
1845 owait.wait.flags = 0;
1846 owait.wait.func = memcg_oom_wake_function;
1847 owait.wait.private = current;
1848 INIT_LIST_HEAD(&owait.wait.task_list);
1849 need_to_kill = true;
1850 mem_cgroup_mark_under_oom(memcg);
1852 /* At first, try to OOM lock hierarchy under memcg.*/
1853 spin_lock(&memcg_oom_lock);
1854 locked = mem_cgroup_oom_lock(memcg);
1856 * Even if signal_pending(), we can't quit charge() loop without
1857 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1858 * under OOM is always welcomed, use TASK_KILLABLE here.
1860 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1861 if (!locked || memcg->oom_kill_disable)
1862 need_to_kill = false;
1864 mem_cgroup_oom_notify(memcg);
1865 spin_unlock(&memcg_oom_lock);
1868 finish_wait(&memcg_oom_waitq, &owait.wait);
1869 mem_cgroup_out_of_memory(memcg, mask, order);
1872 finish_wait(&memcg_oom_waitq, &owait.wait);
1874 spin_lock(&memcg_oom_lock);
1876 mem_cgroup_oom_unlock(memcg);
1877 memcg_wakeup_oom(memcg);
1878 spin_unlock(&memcg_oom_lock);
1880 mem_cgroup_unmark_under_oom(memcg);
1882 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
1884 /* Give chance to dying process */
1885 schedule_timeout_uninterruptible(1);
1890 * Currently used to update mapped file statistics, but the routine can be
1891 * generalized to update other statistics as well.
1893 * Notes: Race condition
1895 * We usually use page_cgroup_lock() for accessing page_cgroup member but
1896 * it tends to be costly. But considering some conditions, we doesn't need
1897 * to do so _always_.
1899 * Considering "charge", lock_page_cgroup() is not required because all
1900 * file-stat operations happen after a page is attached to radix-tree. There
1901 * are no race with "charge".
1903 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
1904 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
1905 * if there are race with "uncharge". Statistics itself is properly handled
1908 * Considering "move", this is an only case we see a race. To make the race
1909 * small, we check mm->moving_account and detect there are possibility of race
1910 * If there is, we take a lock.
1913 void __mem_cgroup_begin_update_page_stat(struct page *page,
1914 bool *locked, unsigned long *flags)
1916 struct mem_cgroup *memcg;
1917 struct page_cgroup *pc;
1919 pc = lookup_page_cgroup(page);
1921 memcg = pc->mem_cgroup;
1922 if (unlikely(!memcg || !PageCgroupUsed(pc)))
1925 * If this memory cgroup is not under account moving, we don't
1926 * need to take move_lock_page_cgroup(). Because we already hold
1927 * rcu_read_lock(), any calls to move_account will be delayed until
1928 * rcu_read_unlock() if mem_cgroup_stealed() == true.
1930 if (!mem_cgroup_stealed(memcg))
1933 move_lock_mem_cgroup(memcg, flags);
1934 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
1935 move_unlock_mem_cgroup(memcg, flags);
1941 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
1943 struct page_cgroup *pc = lookup_page_cgroup(page);
1946 * It's guaranteed that pc->mem_cgroup never changes while
1947 * lock is held because a routine modifies pc->mem_cgroup
1948 * should take move_lock_page_cgroup().
1950 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
1953 void mem_cgroup_update_page_stat(struct page *page,
1954 enum mem_cgroup_page_stat_item idx, int val)
1956 struct mem_cgroup *memcg;
1957 struct page_cgroup *pc = lookup_page_cgroup(page);
1958 unsigned long uninitialized_var(flags);
1960 if (mem_cgroup_disabled())
1963 memcg = pc->mem_cgroup;
1964 if (unlikely(!memcg || !PageCgroupUsed(pc)))
1968 case MEMCG_NR_FILE_MAPPED:
1969 idx = MEM_CGROUP_STAT_FILE_MAPPED;
1975 this_cpu_add(memcg->stat->count[idx], val);
1979 * size of first charge trial. "32" comes from vmscan.c's magic value.
1980 * TODO: maybe necessary to use big numbers in big irons.
1982 #define CHARGE_BATCH 32U
1983 struct memcg_stock_pcp {
1984 struct mem_cgroup *cached; /* this never be root cgroup */
1985 unsigned int nr_pages;
1986 struct work_struct work;
1987 unsigned long flags;
1988 #define FLUSHING_CACHED_CHARGE (0)
1990 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1991 static DEFINE_MUTEX(percpu_charge_mutex);
1994 * Try to consume stocked charge on this cpu. If success, one page is consumed
1995 * from local stock and true is returned. If the stock is 0 or charges from a
1996 * cgroup which is not current target, returns false. This stock will be
1999 static bool consume_stock(struct mem_cgroup *memcg)
2001 struct memcg_stock_pcp *stock;
2004 stock = &get_cpu_var(memcg_stock);
2005 if (memcg == stock->cached && stock->nr_pages)
2007 else /* need to call res_counter_charge */
2009 put_cpu_var(memcg_stock);
2014 * Returns stocks cached in percpu to res_counter and reset cached information.
2016 static void drain_stock(struct memcg_stock_pcp *stock)
2018 struct mem_cgroup *old = stock->cached;
2020 if (stock->nr_pages) {
2021 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2023 res_counter_uncharge(&old->res, bytes);
2024 if (do_swap_account)
2025 res_counter_uncharge(&old->memsw, bytes);
2026 stock->nr_pages = 0;
2028 stock->cached = NULL;
2032 * This must be called under preempt disabled or must be called by
2033 * a thread which is pinned to local cpu.
2035 static void drain_local_stock(struct work_struct *dummy)
2037 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2039 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2043 * Cache charges(val) which is from res_counter, to local per_cpu area.
2044 * This will be consumed by consume_stock() function, later.
2046 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2048 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2050 if (stock->cached != memcg) { /* reset if necessary */
2052 stock->cached = memcg;
2054 stock->nr_pages += nr_pages;
2055 put_cpu_var(memcg_stock);
2059 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2060 * of the hierarchy under it. sync flag says whether we should block
2061 * until the work is done.
2063 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2067 /* Notify other cpus that system-wide "drain" is running */
2070 for_each_online_cpu(cpu) {
2071 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2072 struct mem_cgroup *memcg;
2074 memcg = stock->cached;
2075 if (!memcg || !stock->nr_pages)
2077 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2079 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2081 drain_local_stock(&stock->work);
2083 schedule_work_on(cpu, &stock->work);
2091 for_each_online_cpu(cpu) {
2092 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2093 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2094 flush_work(&stock->work);
2101 * Tries to drain stocked charges in other cpus. This function is asynchronous
2102 * and just put a work per cpu for draining localy on each cpu. Caller can
2103 * expects some charges will be back to res_counter later but cannot wait for
2106 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2109 * If someone calls draining, avoid adding more kworker runs.
2111 if (!mutex_trylock(&percpu_charge_mutex))
2113 drain_all_stock(root_memcg, false);
2114 mutex_unlock(&percpu_charge_mutex);
2117 /* This is a synchronous drain interface. */
2118 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2120 /* called when force_empty is called */
2121 mutex_lock(&percpu_charge_mutex);
2122 drain_all_stock(root_memcg, true);
2123 mutex_unlock(&percpu_charge_mutex);
2127 * This function drains percpu counter value from DEAD cpu and
2128 * move it to local cpu. Note that this function can be preempted.
2130 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2134 spin_lock(&memcg->pcp_counter_lock);
2135 for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) {
2136 long x = per_cpu(memcg->stat->count[i], cpu);
2138 per_cpu(memcg->stat->count[i], cpu) = 0;
2139 memcg->nocpu_base.count[i] += x;
2141 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2142 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2144 per_cpu(memcg->stat->events[i], cpu) = 0;
2145 memcg->nocpu_base.events[i] += x;
2147 spin_unlock(&memcg->pcp_counter_lock);
2150 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2151 unsigned long action,
2154 int cpu = (unsigned long)hcpu;
2155 struct memcg_stock_pcp *stock;
2156 struct mem_cgroup *iter;
2158 if (action == CPU_ONLINE)
2161 if ((action != CPU_DEAD) || action != CPU_DEAD_FROZEN)
2164 for_each_mem_cgroup(iter)
2165 mem_cgroup_drain_pcp_counter(iter, cpu);
2167 stock = &per_cpu(memcg_stock, cpu);
2173 /* See __mem_cgroup_try_charge() for details */
2175 CHARGE_OK, /* success */
2176 CHARGE_RETRY, /* need to retry but retry is not bad */
2177 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2178 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2179 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2182 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2183 unsigned int nr_pages, bool oom_check)
2185 unsigned long csize = nr_pages * PAGE_SIZE;
2186 struct mem_cgroup *mem_over_limit;
2187 struct res_counter *fail_res;
2188 unsigned long flags = 0;
2191 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2194 if (!do_swap_account)
2196 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2200 res_counter_uncharge(&memcg->res, csize);
2201 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2202 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2204 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2206 * nr_pages can be either a huge page (HPAGE_PMD_NR), a batch
2207 * of regular pages (CHARGE_BATCH), or a single regular page (1).
2209 * Never reclaim on behalf of optional batching, retry with a
2210 * single page instead.
2212 if (nr_pages == CHARGE_BATCH)
2213 return CHARGE_RETRY;
2215 if (!(gfp_mask & __GFP_WAIT))
2216 return CHARGE_WOULDBLOCK;
2218 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2219 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2220 return CHARGE_RETRY;
2222 * Even though the limit is exceeded at this point, reclaim
2223 * may have been able to free some pages. Retry the charge
2224 * before killing the task.
2226 * Only for regular pages, though: huge pages are rather
2227 * unlikely to succeed so close to the limit, and we fall back
2228 * to regular pages anyway in case of failure.
2230 if (nr_pages == 1 && ret)
2231 return CHARGE_RETRY;
2234 * At task move, charge accounts can be doubly counted. So, it's
2235 * better to wait until the end of task_move if something is going on.
2237 if (mem_cgroup_wait_acct_move(mem_over_limit))
2238 return CHARGE_RETRY;
2240 /* If we don't need to call oom-killer at el, return immediately */
2242 return CHARGE_NOMEM;
2244 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask, get_order(csize)))
2245 return CHARGE_OOM_DIE;
2247 return CHARGE_RETRY;
2251 * __mem_cgroup_try_charge() does
2252 * 1. detect memcg to be charged against from passed *mm and *ptr,
2253 * 2. update res_counter
2254 * 3. call memory reclaim if necessary.
2256 * In some special case, if the task is fatal, fatal_signal_pending() or
2257 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2258 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2259 * as possible without any hazards. 2: all pages should have a valid
2260 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2261 * pointer, that is treated as a charge to root_mem_cgroup.
2263 * So __mem_cgroup_try_charge() will return
2264 * 0 ... on success, filling *ptr with a valid memcg pointer.
2265 * -ENOMEM ... charge failure because of resource limits.
2266 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2268 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2269 * the oom-killer can be invoked.
2271 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2273 unsigned int nr_pages,
2274 struct mem_cgroup **ptr,
2277 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2278 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2279 struct mem_cgroup *memcg = NULL;
2283 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2284 * in system level. So, allow to go ahead dying process in addition to
2287 if (unlikely(test_thread_flag(TIF_MEMDIE)
2288 || fatal_signal_pending(current)))
2292 * We always charge the cgroup the mm_struct belongs to.
2293 * The mm_struct's mem_cgroup changes on task migration if the
2294 * thread group leader migrates. It's possible that mm is not
2295 * set, if so charge the init_mm (happens for pagecache usage).
2298 *ptr = root_mem_cgroup;
2300 if (*ptr) { /* css should be a valid one */
2302 VM_BUG_ON(css_is_removed(&memcg->css));
2303 if (mem_cgroup_is_root(memcg))
2305 if (nr_pages == 1 && consume_stock(memcg))
2307 css_get(&memcg->css);
2309 struct task_struct *p;
2312 p = rcu_dereference(mm->owner);
2314 * Because we don't have task_lock(), "p" can exit.
2315 * In that case, "memcg" can point to root or p can be NULL with
2316 * race with swapoff. Then, we have small risk of mis-accouning.
2317 * But such kind of mis-account by race always happens because
2318 * we don't have cgroup_mutex(). It's overkill and we allo that
2320 * (*) swapoff at el will charge against mm-struct not against
2321 * task-struct. So, mm->owner can be NULL.
2323 memcg = mem_cgroup_from_task(p);
2325 memcg = root_mem_cgroup;
2326 if (mem_cgroup_is_root(memcg)) {
2330 if (nr_pages == 1 && consume_stock(memcg)) {
2332 * It seems dagerous to access memcg without css_get().
2333 * But considering how consume_stok works, it's not
2334 * necessary. If consume_stock success, some charges
2335 * from this memcg are cached on this cpu. So, we
2336 * don't need to call css_get()/css_tryget() before
2337 * calling consume_stock().
2342 /* after here, we may be blocked. we need to get refcnt */
2343 if (!css_tryget(&memcg->css)) {
2353 /* If killed, bypass charge */
2354 if (fatal_signal_pending(current)) {
2355 css_put(&memcg->css);
2360 if (oom && !nr_oom_retries) {
2362 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2365 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, oom_check);
2369 case CHARGE_RETRY: /* not in OOM situation but retry */
2371 css_put(&memcg->css);
2374 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2375 css_put(&memcg->css);
2377 case CHARGE_NOMEM: /* OOM routine works */
2379 css_put(&memcg->css);
2382 /* If oom, we never return -ENOMEM */
2385 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2386 css_put(&memcg->css);
2389 } while (ret != CHARGE_OK);
2391 if (batch > nr_pages)
2392 refill_stock(memcg, batch - nr_pages);
2393 css_put(&memcg->css);
2401 *ptr = root_mem_cgroup;
2406 * Somemtimes we have to undo a charge we got by try_charge().
2407 * This function is for that and do uncharge, put css's refcnt.
2408 * gotten by try_charge().
2410 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2411 unsigned int nr_pages)
2413 if (!mem_cgroup_is_root(memcg)) {
2414 unsigned long bytes = nr_pages * PAGE_SIZE;
2416 res_counter_uncharge(&memcg->res, bytes);
2417 if (do_swap_account)
2418 res_counter_uncharge(&memcg->memsw, bytes);
2423 * A helper function to get mem_cgroup from ID. must be called under
2424 * rcu_read_lock(). The caller must check css_is_removed() or some if
2425 * it's concern. (dropping refcnt from swap can be called against removed
2428 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2430 struct cgroup_subsys_state *css;
2432 /* ID 0 is unused ID */
2435 css = css_lookup(&mem_cgroup_subsys, id);
2438 return container_of(css, struct mem_cgroup, css);
2441 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2443 struct mem_cgroup *memcg = NULL;
2444 struct page_cgroup *pc;
2448 VM_BUG_ON(!PageLocked(page));
2450 pc = lookup_page_cgroup(page);
2451 lock_page_cgroup(pc);
2452 if (PageCgroupUsed(pc)) {
2453 memcg = pc->mem_cgroup;
2454 if (memcg && !css_tryget(&memcg->css))
2456 } else if (PageSwapCache(page)) {
2457 ent.val = page_private(page);
2458 id = lookup_swap_cgroup_id(ent);
2460 memcg = mem_cgroup_lookup(id);
2461 if (memcg && !css_tryget(&memcg->css))
2465 unlock_page_cgroup(pc);
2469 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2471 unsigned int nr_pages,
2472 struct page_cgroup *pc,
2473 enum charge_type ctype,
2476 struct zone *uninitialized_var(zone);
2477 bool was_on_lru = false;
2480 lock_page_cgroup(pc);
2481 if (unlikely(PageCgroupUsed(pc))) {
2482 unlock_page_cgroup(pc);
2483 __mem_cgroup_cancel_charge(memcg, nr_pages);
2487 * we don't need page_cgroup_lock about tail pages, becase they are not
2488 * accessed by any other context at this point.
2492 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2493 * may already be on some other mem_cgroup's LRU. Take care of it.
2496 zone = page_zone(page);
2497 spin_lock_irq(&zone->lru_lock);
2498 if (PageLRU(page)) {
2500 del_page_from_lru_list(zone, page, page_lru(page));
2505 pc->mem_cgroup = memcg;
2507 * We access a page_cgroup asynchronously without lock_page_cgroup().
2508 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2509 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2510 * before USED bit, we need memory barrier here.
2511 * See mem_cgroup_add_lru_list(), etc.
2514 SetPageCgroupUsed(pc);
2518 VM_BUG_ON(PageLRU(page));
2520 add_page_to_lru_list(zone, page, page_lru(page));
2522 spin_unlock_irq(&zone->lru_lock);
2525 if (ctype == MEM_CGROUP_CHARGE_TYPE_MAPPED)
2530 mem_cgroup_charge_statistics(memcg, anon, nr_pages);
2531 unlock_page_cgroup(pc);
2534 * "charge_statistics" updated event counter. Then, check it.
2535 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2536 * if they exceeds softlimit.
2538 memcg_check_events(memcg, page);
2541 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2543 #define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MIGRATION))
2545 * Because tail pages are not marked as "used", set it. We're under
2546 * zone->lru_lock, 'splitting on pmd' and compound_lock.
2547 * charge/uncharge will be never happen and move_account() is done under
2548 * compound_lock(), so we don't have to take care of races.
2550 void mem_cgroup_split_huge_fixup(struct page *head)
2552 struct page_cgroup *head_pc = lookup_page_cgroup(head);
2553 struct page_cgroup *pc;
2556 if (mem_cgroup_disabled())
2558 for (i = 1; i < HPAGE_PMD_NR; i++) {
2560 pc->mem_cgroup = head_pc->mem_cgroup;
2561 smp_wmb();/* see __commit_charge() */
2562 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2565 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2568 * mem_cgroup_move_account - move account of the page
2570 * @nr_pages: number of regular pages (>1 for huge pages)
2571 * @pc: page_cgroup of the page.
2572 * @from: mem_cgroup which the page is moved from.
2573 * @to: mem_cgroup which the page is moved to. @from != @to.
2574 * @uncharge: whether we should call uncharge and css_put against @from.
2576 * The caller must confirm following.
2577 * - page is not on LRU (isolate_page() is useful.)
2578 * - compound_lock is held when nr_pages > 1
2580 * This function doesn't do "charge" nor css_get to new cgroup. It should be
2581 * done by a caller(__mem_cgroup_try_charge would be useful). If @uncharge is
2582 * true, this function does "uncharge" from old cgroup, but it doesn't if
2583 * @uncharge is false, so a caller should do "uncharge".
2585 static int mem_cgroup_move_account(struct page *page,
2586 unsigned int nr_pages,
2587 struct page_cgroup *pc,
2588 struct mem_cgroup *from,
2589 struct mem_cgroup *to,
2592 unsigned long flags;
2594 bool anon = PageAnon(page);
2596 VM_BUG_ON(from == to);
2597 VM_BUG_ON(PageLRU(page));
2599 * The page is isolated from LRU. So, collapse function
2600 * will not handle this page. But page splitting can happen.
2601 * Do this check under compound_page_lock(). The caller should
2605 if (nr_pages > 1 && !PageTransHuge(page))
2608 lock_page_cgroup(pc);
2611 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
2614 move_lock_mem_cgroup(from, &flags);
2616 if (!anon && page_mapped(page)) {
2617 /* Update mapped_file data for mem_cgroup */
2619 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2620 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2623 mem_cgroup_charge_statistics(from, anon, -nr_pages);
2625 /* This is not "cancel", but cancel_charge does all we need. */
2626 __mem_cgroup_cancel_charge(from, nr_pages);
2628 /* caller should have done css_get */
2629 pc->mem_cgroup = to;
2630 mem_cgroup_charge_statistics(to, anon, nr_pages);
2632 * We charges against "to" which may not have any tasks. Then, "to"
2633 * can be under rmdir(). But in current implementation, caller of
2634 * this function is just force_empty() and move charge, so it's
2635 * guaranteed that "to" is never removed. So, we don't check rmdir
2638 move_unlock_mem_cgroup(from, &flags);
2641 unlock_page_cgroup(pc);
2645 memcg_check_events(to, page);
2646 memcg_check_events(from, page);
2652 * move charges to its parent.
2655 static int mem_cgroup_move_parent(struct page *page,
2656 struct page_cgroup *pc,
2657 struct mem_cgroup *child,
2660 struct cgroup *cg = child->css.cgroup;
2661 struct cgroup *pcg = cg->parent;
2662 struct mem_cgroup *parent;
2663 unsigned int nr_pages;
2664 unsigned long uninitialized_var(flags);
2672 if (!get_page_unless_zero(page))
2674 if (isolate_lru_page(page))
2677 nr_pages = hpage_nr_pages(page);
2679 parent = mem_cgroup_from_cont(pcg);
2680 ret = __mem_cgroup_try_charge(NULL, gfp_mask, nr_pages, &parent, false);
2685 flags = compound_lock_irqsave(page);
2687 ret = mem_cgroup_move_account(page, nr_pages, pc, child, parent, true);
2689 __mem_cgroup_cancel_charge(parent, nr_pages);
2692 compound_unlock_irqrestore(page, flags);
2694 putback_lru_page(page);
2702 * Charge the memory controller for page usage.
2704 * 0 if the charge was successful
2705 * < 0 if the cgroup is over its limit
2707 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2708 gfp_t gfp_mask, enum charge_type ctype)
2710 struct mem_cgroup *memcg = NULL;
2711 unsigned int nr_pages = 1;
2712 struct page_cgroup *pc;
2716 if (PageTransHuge(page)) {
2717 nr_pages <<= compound_order(page);
2718 VM_BUG_ON(!PageTransHuge(page));
2720 * Never OOM-kill a process for a huge page. The
2721 * fault handler will fall back to regular pages.
2726 pc = lookup_page_cgroup(page);
2727 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
2730 __mem_cgroup_commit_charge(memcg, page, nr_pages, pc, ctype, false);
2734 int mem_cgroup_newpage_charge(struct page *page,
2735 struct mm_struct *mm, gfp_t gfp_mask)
2737 if (mem_cgroup_disabled())
2739 VM_BUG_ON(page_mapped(page));
2740 VM_BUG_ON(page->mapping && !PageAnon(page));
2742 return mem_cgroup_charge_common(page, mm, gfp_mask,
2743 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2747 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2748 enum charge_type ctype);
2750 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2753 struct mem_cgroup *memcg = NULL;
2754 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
2757 if (mem_cgroup_disabled())
2759 if (PageCompound(page))
2764 if (!page_is_file_cache(page))
2765 type = MEM_CGROUP_CHARGE_TYPE_SHMEM;
2767 if (!PageSwapCache(page))
2768 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
2769 else { /* page is swapcache/shmem */
2770 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &memcg);
2772 __mem_cgroup_commit_charge_swapin(page, memcg, type);
2778 * While swap-in, try_charge -> commit or cancel, the page is locked.
2779 * And when try_charge() successfully returns, one refcnt to memcg without
2780 * struct page_cgroup is acquired. This refcnt will be consumed by
2781 * "commit()" or removed by "cancel()"
2783 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2785 gfp_t mask, struct mem_cgroup **memcgp)
2787 struct mem_cgroup *memcg;
2792 if (mem_cgroup_disabled())
2795 if (!do_swap_account)
2798 * A racing thread's fault, or swapoff, may have already updated
2799 * the pte, and even removed page from swap cache: in those cases
2800 * do_swap_page()'s pte_same() test will fail; but there's also a
2801 * KSM case which does need to charge the page.
2803 if (!PageSwapCache(page))
2805 memcg = try_get_mem_cgroup_from_page(page);
2809 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
2810 css_put(&memcg->css);
2817 ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
2824 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
2825 enum charge_type ctype)
2827 struct page_cgroup *pc;
2829 if (mem_cgroup_disabled())
2833 cgroup_exclude_rmdir(&memcg->css);
2835 pc = lookup_page_cgroup(page);
2836 __mem_cgroup_commit_charge(memcg, page, 1, pc, ctype, true);
2838 * Now swap is on-memory. This means this page may be
2839 * counted both as mem and swap....double count.
2840 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2841 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2842 * may call delete_from_swap_cache() before reach here.
2844 if (do_swap_account && PageSwapCache(page)) {
2845 swp_entry_t ent = {.val = page_private(page)};
2846 struct mem_cgroup *swap_memcg;
2849 id = swap_cgroup_record(ent, 0);
2851 swap_memcg = mem_cgroup_lookup(id);
2854 * This recorded memcg can be obsolete one. So, avoid
2855 * calling css_tryget
2857 if (!mem_cgroup_is_root(swap_memcg))
2858 res_counter_uncharge(&swap_memcg->memsw,
2860 mem_cgroup_swap_statistics(swap_memcg, false);
2861 mem_cgroup_put(swap_memcg);
2866 * At swapin, we may charge account against cgroup which has no tasks.
2867 * So, rmdir()->pre_destroy() can be called while we do this charge.
2868 * In that case, we need to call pre_destroy() again. check it here.
2870 cgroup_release_and_wakeup_rmdir(&memcg->css);
2873 void mem_cgroup_commit_charge_swapin(struct page *page,
2874 struct mem_cgroup *memcg)
2876 __mem_cgroup_commit_charge_swapin(page, memcg,
2877 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2880 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
2882 if (mem_cgroup_disabled())
2886 __mem_cgroup_cancel_charge(memcg, 1);
2889 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
2890 unsigned int nr_pages,
2891 const enum charge_type ctype)
2893 struct memcg_batch_info *batch = NULL;
2894 bool uncharge_memsw = true;
2896 /* If swapout, usage of swap doesn't decrease */
2897 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2898 uncharge_memsw = false;
2900 batch = ¤t->memcg_batch;
2902 * In usual, we do css_get() when we remember memcg pointer.
2903 * But in this case, we keep res->usage until end of a series of
2904 * uncharges. Then, it's ok to ignore memcg's refcnt.
2907 batch->memcg = memcg;
2909 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2910 * In those cases, all pages freed continuously can be expected to be in
2911 * the same cgroup and we have chance to coalesce uncharges.
2912 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2913 * because we want to do uncharge as soon as possible.
2916 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
2917 goto direct_uncharge;
2920 goto direct_uncharge;
2923 * In typical case, batch->memcg == mem. This means we can
2924 * merge a series of uncharges to an uncharge of res_counter.
2925 * If not, we uncharge res_counter ony by one.
2927 if (batch->memcg != memcg)
2928 goto direct_uncharge;
2929 /* remember freed charge and uncharge it later */
2932 batch->memsw_nr_pages++;
2935 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
2937 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
2938 if (unlikely(batch->memcg != memcg))
2939 memcg_oom_recover(memcg);
2943 * uncharge if !page_mapped(page)
2945 static struct mem_cgroup *
2946 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
2948 struct mem_cgroup *memcg = NULL;
2949 unsigned int nr_pages = 1;
2950 struct page_cgroup *pc;
2953 if (mem_cgroup_disabled())
2956 if (PageSwapCache(page))
2959 if (PageTransHuge(page)) {
2960 nr_pages <<= compound_order(page);
2961 VM_BUG_ON(!PageTransHuge(page));
2964 * Check if our page_cgroup is valid
2966 pc = lookup_page_cgroup(page);
2967 if (unlikely(!PageCgroupUsed(pc)))
2970 lock_page_cgroup(pc);
2972 memcg = pc->mem_cgroup;
2974 if (!PageCgroupUsed(pc))
2977 anon = PageAnon(page);
2980 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2982 * Generally PageAnon tells if it's the anon statistics to be
2983 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
2984 * used before page reached the stage of being marked PageAnon.
2988 case MEM_CGROUP_CHARGE_TYPE_DROP:
2989 /* See mem_cgroup_prepare_migration() */
2990 if (page_mapped(page) || PageCgroupMigration(pc))
2993 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
2994 if (!PageAnon(page)) { /* Shared memory */
2995 if (page->mapping && !page_is_file_cache(page))
2997 } else if (page_mapped(page)) /* Anon */
3004 mem_cgroup_charge_statistics(memcg, anon, -nr_pages);
3006 ClearPageCgroupUsed(pc);
3008 * pc->mem_cgroup is not cleared here. It will be accessed when it's
3009 * freed from LRU. This is safe because uncharged page is expected not
3010 * to be reused (freed soon). Exception is SwapCache, it's handled by
3011 * special functions.
3014 unlock_page_cgroup(pc);
3016 * even after unlock, we have memcg->res.usage here and this memcg
3017 * will never be freed.
3019 memcg_check_events(memcg, page);
3020 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
3021 mem_cgroup_swap_statistics(memcg, true);
3022 mem_cgroup_get(memcg);
3024 if (!mem_cgroup_is_root(memcg))
3025 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
3030 unlock_page_cgroup(pc);
3034 void mem_cgroup_uncharge_page(struct page *page)
3037 if (page_mapped(page))
3039 VM_BUG_ON(page->mapping && !PageAnon(page));
3040 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
3043 void mem_cgroup_uncharge_cache_page(struct page *page)
3045 VM_BUG_ON(page_mapped(page));
3046 VM_BUG_ON(page->mapping);
3047 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
3051 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
3052 * In that cases, pages are freed continuously and we can expect pages
3053 * are in the same memcg. All these calls itself limits the number of
3054 * pages freed at once, then uncharge_start/end() is called properly.
3055 * This may be called prural(2) times in a context,
3058 void mem_cgroup_uncharge_start(void)
3060 current->memcg_batch.do_batch++;
3061 /* We can do nest. */
3062 if (current->memcg_batch.do_batch == 1) {
3063 current->memcg_batch.memcg = NULL;
3064 current->memcg_batch.nr_pages = 0;
3065 current->memcg_batch.memsw_nr_pages = 0;
3069 void mem_cgroup_uncharge_end(void)
3071 struct memcg_batch_info *batch = ¤t->memcg_batch;
3073 if (!batch->do_batch)
3077 if (batch->do_batch) /* If stacked, do nothing. */
3083 * This "batch->memcg" is valid without any css_get/put etc...
3084 * bacause we hide charges behind us.
3086 if (batch->nr_pages)
3087 res_counter_uncharge(&batch->memcg->res,
3088 batch->nr_pages * PAGE_SIZE);
3089 if (batch->memsw_nr_pages)
3090 res_counter_uncharge(&batch->memcg->memsw,
3091 batch->memsw_nr_pages * PAGE_SIZE);
3092 memcg_oom_recover(batch->memcg);
3093 /* forget this pointer (for sanity check) */
3094 batch->memcg = NULL;
3099 * called after __delete_from_swap_cache() and drop "page" account.
3100 * memcg information is recorded to swap_cgroup of "ent"
3103 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
3105 struct mem_cgroup *memcg;
3106 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
3108 if (!swapout) /* this was a swap cache but the swap is unused ! */
3109 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
3111 memcg = __mem_cgroup_uncharge_common(page, ctype);
3114 * record memcg information, if swapout && memcg != NULL,
3115 * mem_cgroup_get() was called in uncharge().
3117 if (do_swap_account && swapout && memcg)
3118 swap_cgroup_record(ent, css_id(&memcg->css));
3122 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3124 * called from swap_entry_free(). remove record in swap_cgroup and
3125 * uncharge "memsw" account.
3127 void mem_cgroup_uncharge_swap(swp_entry_t ent)
3129 struct mem_cgroup *memcg;
3132 if (!do_swap_account)
3135 id = swap_cgroup_record(ent, 0);
3137 memcg = mem_cgroup_lookup(id);
3140 * We uncharge this because swap is freed.
3141 * This memcg can be obsolete one. We avoid calling css_tryget
3143 if (!mem_cgroup_is_root(memcg))
3144 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
3145 mem_cgroup_swap_statistics(memcg, false);
3146 mem_cgroup_put(memcg);
3152 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3153 * @entry: swap entry to be moved
3154 * @from: mem_cgroup which the entry is moved from
3155 * @to: mem_cgroup which the entry is moved to
3156 * @need_fixup: whether we should fixup res_counters and refcounts.
3158 * It succeeds only when the swap_cgroup's record for this entry is the same
3159 * as the mem_cgroup's id of @from.
3161 * Returns 0 on success, -EINVAL on failure.
3163 * The caller must have charged to @to, IOW, called res_counter_charge() about
3164 * both res and memsw, and called css_get().
3166 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3167 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3169 unsigned short old_id, new_id;
3171 old_id = css_id(&from->css);
3172 new_id = css_id(&to->css);
3174 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3175 mem_cgroup_swap_statistics(from, false);
3176 mem_cgroup_swap_statistics(to, true);
3178 * This function is only called from task migration context now.
3179 * It postpones res_counter and refcount handling till the end
3180 * of task migration(mem_cgroup_clear_mc()) for performance
3181 * improvement. But we cannot postpone mem_cgroup_get(to)
3182 * because if the process that has been moved to @to does
3183 * swap-in, the refcount of @to might be decreased to 0.
3187 if (!mem_cgroup_is_root(from))
3188 res_counter_uncharge(&from->memsw, PAGE_SIZE);
3189 mem_cgroup_put(from);
3191 * we charged both to->res and to->memsw, so we should
3194 if (!mem_cgroup_is_root(to))
3195 res_counter_uncharge(&to->res, PAGE_SIZE);
3202 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3203 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3210 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
3213 int mem_cgroup_prepare_migration(struct page *page,
3214 struct page *newpage, struct mem_cgroup **memcgp, gfp_t gfp_mask)
3216 struct mem_cgroup *memcg = NULL;
3217 struct page_cgroup *pc;
3218 enum charge_type ctype;
3223 VM_BUG_ON(PageTransHuge(page));
3224 if (mem_cgroup_disabled())
3227 pc = lookup_page_cgroup(page);
3228 lock_page_cgroup(pc);
3229 if (PageCgroupUsed(pc)) {
3230 memcg = pc->mem_cgroup;
3231 css_get(&memcg->css);
3233 * At migrating an anonymous page, its mapcount goes down
3234 * to 0 and uncharge() will be called. But, even if it's fully
3235 * unmapped, migration may fail and this page has to be
3236 * charged again. We set MIGRATION flag here and delay uncharge
3237 * until end_migration() is called
3239 * Corner Case Thinking
3241 * When the old page was mapped as Anon and it's unmap-and-freed
3242 * while migration was ongoing.
3243 * If unmap finds the old page, uncharge() of it will be delayed
3244 * until end_migration(). If unmap finds a new page, it's
3245 * uncharged when it make mapcount to be 1->0. If unmap code
3246 * finds swap_migration_entry, the new page will not be mapped
3247 * and end_migration() will find it(mapcount==0).
3250 * When the old page was mapped but migraion fails, the kernel
3251 * remaps it. A charge for it is kept by MIGRATION flag even
3252 * if mapcount goes down to 0. We can do remap successfully
3253 * without charging it again.
3256 * The "old" page is under lock_page() until the end of
3257 * migration, so, the old page itself will not be swapped-out.
3258 * If the new page is swapped out before end_migraton, our
3259 * hook to usual swap-out path will catch the event.
3262 SetPageCgroupMigration(pc);
3264 unlock_page_cgroup(pc);
3266 * If the page is not charged at this point,
3273 ret = __mem_cgroup_try_charge(NULL, gfp_mask, 1, memcgp, false);
3274 css_put(&memcg->css);/* drop extra refcnt */
3276 if (PageAnon(page)) {
3277 lock_page_cgroup(pc);
3278 ClearPageCgroupMigration(pc);
3279 unlock_page_cgroup(pc);
3281 * The old page may be fully unmapped while we kept it.
3283 mem_cgroup_uncharge_page(page);
3285 /* we'll need to revisit this error code (we have -EINTR) */
3289 * We charge new page before it's used/mapped. So, even if unlock_page()
3290 * is called before end_migration, we can catch all events on this new
3291 * page. In the case new page is migrated but not remapped, new page's
3292 * mapcount will be finally 0 and we call uncharge in end_migration().
3294 pc = lookup_page_cgroup(newpage);
3296 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
3297 else if (page_is_file_cache(page))
3298 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
3300 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3301 __mem_cgroup_commit_charge(memcg, newpage, 1, pc, ctype, false);
3305 /* remove redundant charge if migration failed*/
3306 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
3307 struct page *oldpage, struct page *newpage, bool migration_ok)
3309 struct page *used, *unused;
3310 struct page_cgroup *pc;
3315 /* blocks rmdir() */
3316 cgroup_exclude_rmdir(&memcg->css);
3317 if (!migration_ok) {
3325 * We disallowed uncharge of pages under migration because mapcount
3326 * of the page goes down to zero, temporarly.
3327 * Clear the flag and check the page should be charged.
3329 pc = lookup_page_cgroup(oldpage);
3330 lock_page_cgroup(pc);
3331 ClearPageCgroupMigration(pc);
3332 unlock_page_cgroup(pc);
3333 anon = PageAnon(used);
3334 __mem_cgroup_uncharge_common(unused,
3335 anon ? MEM_CGROUP_CHARGE_TYPE_MAPPED
3336 : MEM_CGROUP_CHARGE_TYPE_CACHE);
3339 * If a page is a file cache, radix-tree replacement is very atomic
3340 * and we can skip this check. When it was an Anon page, its mapcount
3341 * goes down to 0. But because we added MIGRATION flage, it's not
3342 * uncharged yet. There are several case but page->mapcount check
3343 * and USED bit check in mem_cgroup_uncharge_page() will do enough
3344 * check. (see prepare_charge() also)
3347 mem_cgroup_uncharge_page(used);
3349 * At migration, we may charge account against cgroup which has no
3351 * So, rmdir()->pre_destroy() can be called while we do this charge.
3352 * In that case, we need to call pre_destroy() again. check it here.
3354 cgroup_release_and_wakeup_rmdir(&memcg->css);
3358 * At replace page cache, newpage is not under any memcg but it's on
3359 * LRU. So, this function doesn't touch res_counter but handles LRU
3360 * in correct way. Both pages are locked so we cannot race with uncharge.
3362 void mem_cgroup_replace_page_cache(struct page *oldpage,
3363 struct page *newpage)
3365 struct mem_cgroup *memcg;
3366 struct page_cgroup *pc;
3367 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
3369 if (mem_cgroup_disabled())
3372 pc = lookup_page_cgroup(oldpage);
3373 /* fix accounting on old pages */
3374 lock_page_cgroup(pc);
3375 memcg = pc->mem_cgroup;
3376 mem_cgroup_charge_statistics(memcg, false, -1);
3377 ClearPageCgroupUsed(pc);
3378 unlock_page_cgroup(pc);
3380 if (PageSwapBacked(oldpage))
3381 type = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3384 * Even if newpage->mapping was NULL before starting replacement,
3385 * the newpage may be on LRU(or pagevec for LRU) already. We lock
3386 * LRU while we overwrite pc->mem_cgroup.
3388 __mem_cgroup_commit_charge(memcg, newpage, 1, pc, type, true);
3391 #ifdef CONFIG_DEBUG_VM
3392 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3394 struct page_cgroup *pc;
3396 pc = lookup_page_cgroup(page);
3398 * Can be NULL while feeding pages into the page allocator for
3399 * the first time, i.e. during boot or memory hotplug;
3400 * or when mem_cgroup_disabled().
3402 if (likely(pc) && PageCgroupUsed(pc))
3407 bool mem_cgroup_bad_page_check(struct page *page)
3409 if (mem_cgroup_disabled())
3412 return lookup_page_cgroup_used(page) != NULL;
3415 void mem_cgroup_print_bad_page(struct page *page)
3417 struct page_cgroup *pc;
3419 pc = lookup_page_cgroup_used(page);
3421 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
3422 pc, pc->flags, pc->mem_cgroup);
3427 static DEFINE_MUTEX(set_limit_mutex);
3429 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3430 unsigned long long val)
3433 u64 memswlimit, memlimit;
3435 int children = mem_cgroup_count_children(memcg);
3436 u64 curusage, oldusage;
3440 * For keeping hierarchical_reclaim simple, how long we should retry
3441 * is depends on callers. We set our retry-count to be function
3442 * of # of children which we should visit in this loop.
3444 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3446 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3449 while (retry_count) {
3450 if (signal_pending(current)) {
3455 * Rather than hide all in some function, I do this in
3456 * open coded manner. You see what this really does.
3457 * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3459 mutex_lock(&set_limit_mutex);
3460 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3461 if (memswlimit < val) {
3463 mutex_unlock(&set_limit_mutex);
3467 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3471 ret = res_counter_set_limit(&memcg->res, val);
3473 if (memswlimit == val)
3474 memcg->memsw_is_minimum = true;
3476 memcg->memsw_is_minimum = false;
3478 mutex_unlock(&set_limit_mutex);
3483 mem_cgroup_reclaim(memcg, GFP_KERNEL,
3484 MEM_CGROUP_RECLAIM_SHRINK);
3485 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3486 /* Usage is reduced ? */
3487 if (curusage >= oldusage)
3490 oldusage = curusage;
3492 if (!ret && enlarge)
3493 memcg_oom_recover(memcg);
3498 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3499 unsigned long long val)
3502 u64 memlimit, memswlimit, oldusage, curusage;
3503 int children = mem_cgroup_count_children(memcg);
3507 /* see mem_cgroup_resize_res_limit */
3508 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3509 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3510 while (retry_count) {
3511 if (signal_pending(current)) {
3516 * Rather than hide all in some function, I do this in
3517 * open coded manner. You see what this really does.
3518 * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3520 mutex_lock(&set_limit_mutex);
3521 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3522 if (memlimit > val) {
3524 mutex_unlock(&set_limit_mutex);
3527 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3528 if (memswlimit < val)
3530 ret = res_counter_set_limit(&memcg->memsw, val);
3532 if (memlimit == val)
3533 memcg->memsw_is_minimum = true;
3535 memcg->memsw_is_minimum = false;
3537 mutex_unlock(&set_limit_mutex);
3542 mem_cgroup_reclaim(memcg, GFP_KERNEL,
3543 MEM_CGROUP_RECLAIM_NOSWAP |
3544 MEM_CGROUP_RECLAIM_SHRINK);
3545 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3546 /* Usage is reduced ? */
3547 if (curusage >= oldusage)
3550 oldusage = curusage;
3552 if (!ret && enlarge)
3553 memcg_oom_recover(memcg);
3557 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3559 unsigned long *total_scanned)
3561 unsigned long nr_reclaimed = 0;
3562 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3563 unsigned long reclaimed;
3565 struct mem_cgroup_tree_per_zone *mctz;
3566 unsigned long long excess;
3567 unsigned long nr_scanned;
3572 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3574 * This loop can run a while, specially if mem_cgroup's continuously
3575 * keep exceeding their soft limit and putting the system under
3582 mz = mem_cgroup_largest_soft_limit_node(mctz);
3587 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
3588 gfp_mask, &nr_scanned);
3589 nr_reclaimed += reclaimed;
3590 *total_scanned += nr_scanned;
3591 spin_lock(&mctz->lock);
3594 * If we failed to reclaim anything from this memory cgroup
3595 * it is time to move on to the next cgroup
3601 * Loop until we find yet another one.
3603 * By the time we get the soft_limit lock
3604 * again, someone might have aded the
3605 * group back on the RB tree. Iterate to
3606 * make sure we get a different mem.
3607 * mem_cgroup_largest_soft_limit_node returns
3608 * NULL if no other cgroup is present on
3612 __mem_cgroup_largest_soft_limit_node(mctz);
3614 css_put(&next_mz->memcg->css);
3615 else /* next_mz == NULL or other memcg */
3619 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
3620 excess = res_counter_soft_limit_excess(&mz->memcg->res);
3622 * One school of thought says that we should not add
3623 * back the node to the tree if reclaim returns 0.
3624 * But our reclaim could return 0, simply because due
3625 * to priority we are exposing a smaller subset of
3626 * memory to reclaim from. Consider this as a longer
3629 /* If excess == 0, no tree ops */
3630 __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess);
3631 spin_unlock(&mctz->lock);
3632 css_put(&mz->memcg->css);
3635 * Could not reclaim anything and there are no more
3636 * mem cgroups to try or we seem to be looping without
3637 * reclaiming anything.
3639 if (!nr_reclaimed &&
3641 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3643 } while (!nr_reclaimed);
3645 css_put(&next_mz->memcg->css);
3646 return nr_reclaimed;
3650 * This routine traverse page_cgroup in given list and drop them all.
3651 * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3653 static int mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
3654 int node, int zid, enum lru_list lru)
3656 struct mem_cgroup_per_zone *mz;
3657 unsigned long flags, loop;
3658 struct list_head *list;
3663 zone = &NODE_DATA(node)->node_zones[zid];
3664 mz = mem_cgroup_zoneinfo(memcg, node, zid);
3665 list = &mz->lruvec.lists[lru];
3667 loop = mz->lru_size[lru];
3668 /* give some margin against EBUSY etc...*/
3672 struct page_cgroup *pc;
3676 spin_lock_irqsave(&zone->lru_lock, flags);
3677 if (list_empty(list)) {
3678 spin_unlock_irqrestore(&zone->lru_lock, flags);
3681 page = list_entry(list->prev, struct page, lru);
3683 list_move(&page->lru, list);
3685 spin_unlock_irqrestore(&zone->lru_lock, flags);
3688 spin_unlock_irqrestore(&zone->lru_lock, flags);
3690 pc = lookup_page_cgroup(page);
3692 ret = mem_cgroup_move_parent(page, pc, memcg, GFP_KERNEL);
3693 if (ret == -ENOMEM || ret == -EINTR)
3696 if (ret == -EBUSY || ret == -EINVAL) {
3697 /* found lock contention or "pc" is obsolete. */
3704 if (!ret && !list_empty(list))
3710 * make mem_cgroup's charge to be 0 if there is no task.
3711 * This enables deleting this mem_cgroup.
3713 static int mem_cgroup_force_empty(struct mem_cgroup *memcg, bool free_all)
3716 int node, zid, shrink;
3717 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3718 struct cgroup *cgrp = memcg->css.cgroup;
3720 css_get(&memcg->css);
3723 /* should free all ? */
3729 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3732 if (signal_pending(current))
3734 /* This is for making all *used* pages to be on LRU. */
3735 lru_add_drain_all();
3736 drain_all_stock_sync(memcg);
3738 mem_cgroup_start_move(memcg);
3739 for_each_node_state(node, N_HIGH_MEMORY) {
3740 for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3743 ret = mem_cgroup_force_empty_list(memcg,
3752 mem_cgroup_end_move(memcg);
3753 memcg_oom_recover(memcg);
3754 /* it seems parent cgroup doesn't have enough mem */
3758 /* "ret" should also be checked to ensure all lists are empty. */
3759 } while (memcg->res.usage > 0 || ret);
3761 css_put(&memcg->css);
3765 /* returns EBUSY if there is a task or if we come here twice. */
3766 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3770 /* we call try-to-free pages for make this cgroup empty */
3771 lru_add_drain_all();
3772 /* try to free all pages in this cgroup */
3774 while (nr_retries && memcg->res.usage > 0) {
3777 if (signal_pending(current)) {
3781 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
3785 /* maybe some writeback is necessary */
3786 congestion_wait(BLK_RW_ASYNC, HZ/10);
3791 /* try move_account...there may be some *locked* pages. */
3795 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3797 return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3801 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3803 return mem_cgroup_from_cont(cont)->use_hierarchy;
3806 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3810 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3811 struct cgroup *parent = cont->parent;
3812 struct mem_cgroup *parent_memcg = NULL;
3815 parent_memcg = mem_cgroup_from_cont(parent);
3819 * If parent's use_hierarchy is set, we can't make any modifications
3820 * in the child subtrees. If it is unset, then the change can
3821 * occur, provided the current cgroup has no children.
3823 * For the root cgroup, parent_mem is NULL, we allow value to be
3824 * set if there are no children.
3826 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3827 (val == 1 || val == 0)) {
3828 if (list_empty(&cont->children))
3829 memcg->use_hierarchy = val;
3840 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
3841 enum mem_cgroup_stat_index idx)
3843 struct mem_cgroup *iter;
3846 /* Per-cpu values can be negative, use a signed accumulator */
3847 for_each_mem_cgroup_tree(iter, memcg)
3848 val += mem_cgroup_read_stat(iter, idx);
3850 if (val < 0) /* race ? */
3855 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3859 if (!mem_cgroup_is_root(memcg)) {
3861 return res_counter_read_u64(&memcg->res, RES_USAGE);
3863 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
3866 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
3867 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
3870 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
3872 return val << PAGE_SHIFT;
3875 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
3877 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3881 type = MEMFILE_TYPE(cft->private);
3882 name = MEMFILE_ATTR(cft->private);
3885 if (name == RES_USAGE)
3886 val = mem_cgroup_usage(memcg, false);
3888 val = res_counter_read_u64(&memcg->res, name);
3891 if (name == RES_USAGE)
3892 val = mem_cgroup_usage(memcg, true);
3894 val = res_counter_read_u64(&memcg->memsw, name);
3903 * The user of this function is...
3906 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
3909 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3911 unsigned long long val;
3914 type = MEMFILE_TYPE(cft->private);
3915 name = MEMFILE_ATTR(cft->private);
3918 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3922 /* This function does all necessary parse...reuse it */
3923 ret = res_counter_memparse_write_strategy(buffer, &val);
3927 ret = mem_cgroup_resize_limit(memcg, val);
3929 ret = mem_cgroup_resize_memsw_limit(memcg, val);
3931 case RES_SOFT_LIMIT:
3932 ret = res_counter_memparse_write_strategy(buffer, &val);
3936 * For memsw, soft limits are hard to implement in terms
3937 * of semantics, for now, we support soft limits for
3938 * control without swap
3941 ret = res_counter_set_soft_limit(&memcg->res, val);
3946 ret = -EINVAL; /* should be BUG() ? */
3952 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
3953 unsigned long long *mem_limit, unsigned long long *memsw_limit)
3955 struct cgroup *cgroup;
3956 unsigned long long min_limit, min_memsw_limit, tmp;
3958 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3959 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3960 cgroup = memcg->css.cgroup;
3961 if (!memcg->use_hierarchy)
3964 while (cgroup->parent) {
3965 cgroup = cgroup->parent;
3966 memcg = mem_cgroup_from_cont(cgroup);
3967 if (!memcg->use_hierarchy)
3969 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
3970 min_limit = min(min_limit, tmp);
3971 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3972 min_memsw_limit = min(min_memsw_limit, tmp);
3975 *mem_limit = min_limit;
3976 *memsw_limit = min_memsw_limit;
3979 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
3981 struct mem_cgroup *memcg;
3984 memcg = mem_cgroup_from_cont(cont);
3985 type = MEMFILE_TYPE(event);
3986 name = MEMFILE_ATTR(event);
3990 res_counter_reset_max(&memcg->res);
3992 res_counter_reset_max(&memcg->memsw);
3996 res_counter_reset_failcnt(&memcg->res);
3998 res_counter_reset_failcnt(&memcg->memsw);
4005 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
4008 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
4012 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4013 struct cftype *cft, u64 val)
4015 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4017 if (val >= (1 << NR_MOVE_TYPE))
4020 * We check this value several times in both in can_attach() and
4021 * attach(), so we need cgroup lock to prevent this value from being
4025 memcg->move_charge_at_immigrate = val;
4031 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4032 struct cftype *cft, u64 val)
4039 /* For read statistics */
4057 struct mcs_total_stat {
4058 s64 stat[NR_MCS_STAT];
4064 } memcg_stat_strings[NR_MCS_STAT] = {
4065 {"cache", "total_cache"},
4066 {"rss", "total_rss"},
4067 {"mapped_file", "total_mapped_file"},
4068 {"pgpgin", "total_pgpgin"},
4069 {"pgpgout", "total_pgpgout"},
4070 {"swap", "total_swap"},
4071 {"pgfault", "total_pgfault"},
4072 {"pgmajfault", "total_pgmajfault"},
4073 {"inactive_anon", "total_inactive_anon"},
4074 {"active_anon", "total_active_anon"},
4075 {"inactive_file", "total_inactive_file"},
4076 {"active_file", "total_active_file"},
4077 {"unevictable", "total_unevictable"}
4082 mem_cgroup_get_local_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4087 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_CACHE);
4088 s->stat[MCS_CACHE] += val * PAGE_SIZE;
4089 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_RSS);
4090 s->stat[MCS_RSS] += val * PAGE_SIZE;
4091 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_FILE_MAPPED);
4092 s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
4093 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGIN);
4094 s->stat[MCS_PGPGIN] += val;
4095 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGOUT);
4096 s->stat[MCS_PGPGOUT] += val;
4097 if (do_swap_account) {
4098 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
4099 s->stat[MCS_SWAP] += val * PAGE_SIZE;
4101 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGFAULT);
4102 s->stat[MCS_PGFAULT] += val;
4103 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGMAJFAULT);
4104 s->stat[MCS_PGMAJFAULT] += val;
4107 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_ANON));
4108 s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
4109 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_ANON));
4110 s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
4111 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_FILE));
4112 s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
4113 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_FILE));
4114 s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
4115 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
4116 s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
4120 mem_cgroup_get_total_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4122 struct mem_cgroup *iter;
4124 for_each_mem_cgroup_tree(iter, memcg)
4125 mem_cgroup_get_local_stat(iter, s);
4129 static int mem_control_numa_stat_show(struct seq_file *m, void *arg)
4132 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
4133 unsigned long node_nr;
4134 struct cgroup *cont = m->private;
4135 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4137 total_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL);
4138 seq_printf(m, "total=%lu", total_nr);
4139 for_each_node_state(nid, N_HIGH_MEMORY) {
4140 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL);
4141 seq_printf(m, " N%d=%lu", nid, node_nr);
4145 file_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_FILE);
4146 seq_printf(m, "file=%lu", file_nr);
4147 for_each_node_state(nid, N_HIGH_MEMORY) {
4148 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4150 seq_printf(m, " N%d=%lu", nid, node_nr);
4154 anon_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_ANON);
4155 seq_printf(m, "anon=%lu", anon_nr);
4156 for_each_node_state(nid, N_HIGH_MEMORY) {
4157 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4159 seq_printf(m, " N%d=%lu", nid, node_nr);
4163 unevictable_nr = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
4164 seq_printf(m, "unevictable=%lu", unevictable_nr);
4165 for_each_node_state(nid, N_HIGH_MEMORY) {
4166 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4167 BIT(LRU_UNEVICTABLE));
4168 seq_printf(m, " N%d=%lu", nid, node_nr);
4173 #endif /* CONFIG_NUMA */
4175 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
4176 struct cgroup_map_cb *cb)
4178 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4179 struct mcs_total_stat mystat;
4182 memset(&mystat, 0, sizeof(mystat));
4183 mem_cgroup_get_local_stat(memcg, &mystat);
4186 for (i = 0; i < NR_MCS_STAT; i++) {
4187 if (i == MCS_SWAP && !do_swap_account)
4189 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
4192 /* Hierarchical information */
4194 unsigned long long limit, memsw_limit;
4195 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
4196 cb->fill(cb, "hierarchical_memory_limit", limit);
4197 if (do_swap_account)
4198 cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
4201 memset(&mystat, 0, sizeof(mystat));
4202 mem_cgroup_get_total_stat(memcg, &mystat);
4203 for (i = 0; i < NR_MCS_STAT; i++) {
4204 if (i == MCS_SWAP && !do_swap_account)
4206 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
4209 #ifdef CONFIG_DEBUG_VM
4212 struct mem_cgroup_per_zone *mz;
4213 unsigned long recent_rotated[2] = {0, 0};
4214 unsigned long recent_scanned[2] = {0, 0};
4216 for_each_online_node(nid)
4217 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4218 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
4220 recent_rotated[0] +=
4221 mz->reclaim_stat.recent_rotated[0];
4222 recent_rotated[1] +=
4223 mz->reclaim_stat.recent_rotated[1];
4224 recent_scanned[0] +=
4225 mz->reclaim_stat.recent_scanned[0];
4226 recent_scanned[1] +=
4227 mz->reclaim_stat.recent_scanned[1];
4229 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
4230 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
4231 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
4232 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
4239 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
4241 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4243 return mem_cgroup_swappiness(memcg);
4246 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
4249 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4250 struct mem_cgroup *parent;
4255 if (cgrp->parent == NULL)
4258 parent = mem_cgroup_from_cont(cgrp->parent);
4262 /* If under hierarchy, only empty-root can set this value */
4263 if ((parent->use_hierarchy) ||
4264 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4269 memcg->swappiness = val;
4276 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4278 struct mem_cgroup_threshold_ary *t;
4284 t = rcu_dereference(memcg->thresholds.primary);
4286 t = rcu_dereference(memcg->memsw_thresholds.primary);
4291 usage = mem_cgroup_usage(memcg, swap);
4294 * current_threshold points to threshold just below usage.
4295 * If it's not true, a threshold was crossed after last
4296 * call of __mem_cgroup_threshold().
4298 i = t->current_threshold;
4301 * Iterate backward over array of thresholds starting from
4302 * current_threshold and check if a threshold is crossed.
4303 * If none of thresholds below usage is crossed, we read
4304 * only one element of the array here.
4306 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4307 eventfd_signal(t->entries[i].eventfd, 1);
4309 /* i = current_threshold + 1 */
4313 * Iterate forward over array of thresholds starting from
4314 * current_threshold+1 and check if a threshold is crossed.
4315 * If none of thresholds above usage is crossed, we read
4316 * only one element of the array here.
4318 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4319 eventfd_signal(t->entries[i].eventfd, 1);
4321 /* Update current_threshold */
4322 t->current_threshold = i - 1;
4327 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4330 __mem_cgroup_threshold(memcg, false);
4331 if (do_swap_account)
4332 __mem_cgroup_threshold(memcg, true);
4334 memcg = parent_mem_cgroup(memcg);
4338 static int compare_thresholds(const void *a, const void *b)
4340 const struct mem_cgroup_threshold *_a = a;
4341 const struct mem_cgroup_threshold *_b = b;
4343 return _a->threshold - _b->threshold;
4346 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4348 struct mem_cgroup_eventfd_list *ev;
4350 list_for_each_entry(ev, &memcg->oom_notify, list)
4351 eventfd_signal(ev->eventfd, 1);
4355 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4357 struct mem_cgroup *iter;
4359 for_each_mem_cgroup_tree(iter, memcg)
4360 mem_cgroup_oom_notify_cb(iter);
4363 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
4364 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4366 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4367 struct mem_cgroup_thresholds *thresholds;
4368 struct mem_cgroup_threshold_ary *new;
4369 int type = MEMFILE_TYPE(cft->private);
4370 u64 threshold, usage;
4373 ret = res_counter_memparse_write_strategy(args, &threshold);
4377 mutex_lock(&memcg->thresholds_lock);
4380 thresholds = &memcg->thresholds;
4381 else if (type == _MEMSWAP)
4382 thresholds = &memcg->memsw_thresholds;
4386 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4388 /* Check if a threshold crossed before adding a new one */
4389 if (thresholds->primary)
4390 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4392 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4394 /* Allocate memory for new array of thresholds */
4395 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4403 /* Copy thresholds (if any) to new array */
4404 if (thresholds->primary) {
4405 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4406 sizeof(struct mem_cgroup_threshold));
4409 /* Add new threshold */
4410 new->entries[size - 1].eventfd = eventfd;
4411 new->entries[size - 1].threshold = threshold;
4413 /* Sort thresholds. Registering of new threshold isn't time-critical */
4414 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4415 compare_thresholds, NULL);
4417 /* Find current threshold */
4418 new->current_threshold = -1;
4419 for (i = 0; i < size; i++) {
4420 if (new->entries[i].threshold < usage) {
4422 * new->current_threshold will not be used until
4423 * rcu_assign_pointer(), so it's safe to increment
4426 ++new->current_threshold;
4430 /* Free old spare buffer and save old primary buffer as spare */
4431 kfree(thresholds->spare);
4432 thresholds->spare = thresholds->primary;
4434 rcu_assign_pointer(thresholds->primary, new);
4436 /* To be sure that nobody uses thresholds */
4440 mutex_unlock(&memcg->thresholds_lock);
4445 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
4446 struct cftype *cft, struct eventfd_ctx *eventfd)
4448 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4449 struct mem_cgroup_thresholds *thresholds;
4450 struct mem_cgroup_threshold_ary *new;
4451 int type = MEMFILE_TYPE(cft->private);
4455 mutex_lock(&memcg->thresholds_lock);
4457 thresholds = &memcg->thresholds;
4458 else if (type == _MEMSWAP)
4459 thresholds = &memcg->memsw_thresholds;
4464 * Something went wrong if we trying to unregister a threshold
4465 * if we don't have thresholds
4467 BUG_ON(!thresholds);
4469 if (!thresholds->primary)
4472 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4474 /* Check if a threshold crossed before removing */
4475 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4477 /* Calculate new number of threshold */
4479 for (i = 0; i < thresholds->primary->size; i++) {
4480 if (thresholds->primary->entries[i].eventfd != eventfd)
4484 new = thresholds->spare;
4486 /* Set thresholds array to NULL if we don't have thresholds */
4495 /* Copy thresholds and find current threshold */
4496 new->current_threshold = -1;
4497 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4498 if (thresholds->primary->entries[i].eventfd == eventfd)
4501 new->entries[j] = thresholds->primary->entries[i];
4502 if (new->entries[j].threshold < usage) {
4504 * new->current_threshold will not be used
4505 * until rcu_assign_pointer(), so it's safe to increment
4508 ++new->current_threshold;
4514 /* Swap primary and spare array */
4515 thresholds->spare = thresholds->primary;
4516 rcu_assign_pointer(thresholds->primary, new);
4518 /* To be sure that nobody uses thresholds */
4521 mutex_unlock(&memcg->thresholds_lock);
4524 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4525 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4527 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4528 struct mem_cgroup_eventfd_list *event;
4529 int type = MEMFILE_TYPE(cft->private);
4531 BUG_ON(type != _OOM_TYPE);
4532 event = kmalloc(sizeof(*event), GFP_KERNEL);
4536 spin_lock(&memcg_oom_lock);
4538 event->eventfd = eventfd;
4539 list_add(&event->list, &memcg->oom_notify);
4541 /* already in OOM ? */
4542 if (atomic_read(&memcg->under_oom))
4543 eventfd_signal(eventfd, 1);
4544 spin_unlock(&memcg_oom_lock);
4549 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4550 struct cftype *cft, struct eventfd_ctx *eventfd)
4552 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4553 struct mem_cgroup_eventfd_list *ev, *tmp;
4554 int type = MEMFILE_TYPE(cft->private);
4556 BUG_ON(type != _OOM_TYPE);
4558 spin_lock(&memcg_oom_lock);
4560 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4561 if (ev->eventfd == eventfd) {
4562 list_del(&ev->list);
4567 spin_unlock(&memcg_oom_lock);
4570 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4571 struct cftype *cft, struct cgroup_map_cb *cb)
4573 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4575 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
4577 if (atomic_read(&memcg->under_oom))
4578 cb->fill(cb, "under_oom", 1);
4580 cb->fill(cb, "under_oom", 0);
4584 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4585 struct cftype *cft, u64 val)
4587 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4588 struct mem_cgroup *parent;
4590 /* cannot set to root cgroup and only 0 and 1 are allowed */
4591 if (!cgrp->parent || !((val == 0) || (val == 1)))
4594 parent = mem_cgroup_from_cont(cgrp->parent);
4597 /* oom-kill-disable is a flag for subhierarchy. */
4598 if ((parent->use_hierarchy) ||
4599 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4603 memcg->oom_kill_disable = val;
4605 memcg_oom_recover(memcg);
4611 static const struct file_operations mem_control_numa_stat_file_operations = {
4613 .llseek = seq_lseek,
4614 .release = single_release,
4617 static int mem_control_numa_stat_open(struct inode *unused, struct file *file)
4619 struct cgroup *cont = file->f_dentry->d_parent->d_fsdata;
4621 file->f_op = &mem_control_numa_stat_file_operations;
4622 return single_open(file, mem_control_numa_stat_show, cont);
4624 #endif /* CONFIG_NUMA */
4626 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_KMEM
4627 static int register_kmem_files(struct cgroup *cont, struct cgroup_subsys *ss)
4630 * Part of this would be better living in a separate allocation
4631 * function, leaving us with just the cgroup tree population work.
4632 * We, however, depend on state such as network's proto_list that
4633 * is only initialized after cgroup creation. I found the less
4634 * cumbersome way to deal with it to defer it all to populate time
4636 return mem_cgroup_sockets_init(cont, ss);
4639 static void kmem_cgroup_destroy(struct cgroup *cont)
4641 mem_cgroup_sockets_destroy(cont);
4644 static int register_kmem_files(struct cgroup *cont, struct cgroup_subsys *ss)
4649 static void kmem_cgroup_destroy(struct cgroup *cont)
4654 static struct cftype mem_cgroup_files[] = {
4656 .name = "usage_in_bytes",
4657 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4658 .read_u64 = mem_cgroup_read,
4659 .register_event = mem_cgroup_usage_register_event,
4660 .unregister_event = mem_cgroup_usage_unregister_event,
4663 .name = "max_usage_in_bytes",
4664 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4665 .trigger = mem_cgroup_reset,
4666 .read_u64 = mem_cgroup_read,
4669 .name = "limit_in_bytes",
4670 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4671 .write_string = mem_cgroup_write,
4672 .read_u64 = mem_cgroup_read,
4675 .name = "soft_limit_in_bytes",
4676 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4677 .write_string = mem_cgroup_write,
4678 .read_u64 = mem_cgroup_read,
4682 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4683 .trigger = mem_cgroup_reset,
4684 .read_u64 = mem_cgroup_read,
4688 .read_map = mem_control_stat_show,
4691 .name = "force_empty",
4692 .trigger = mem_cgroup_force_empty_write,
4695 .name = "use_hierarchy",
4696 .write_u64 = mem_cgroup_hierarchy_write,
4697 .read_u64 = mem_cgroup_hierarchy_read,
4700 .name = "swappiness",
4701 .read_u64 = mem_cgroup_swappiness_read,
4702 .write_u64 = mem_cgroup_swappiness_write,
4705 .name = "move_charge_at_immigrate",
4706 .read_u64 = mem_cgroup_move_charge_read,
4707 .write_u64 = mem_cgroup_move_charge_write,
4710 .name = "oom_control",
4711 .read_map = mem_cgroup_oom_control_read,
4712 .write_u64 = mem_cgroup_oom_control_write,
4713 .register_event = mem_cgroup_oom_register_event,
4714 .unregister_event = mem_cgroup_oom_unregister_event,
4715 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4719 .name = "numa_stat",
4720 .open = mem_control_numa_stat_open,
4726 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4727 static struct cftype memsw_cgroup_files[] = {
4729 .name = "memsw.usage_in_bytes",
4730 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4731 .read_u64 = mem_cgroup_read,
4732 .register_event = mem_cgroup_usage_register_event,
4733 .unregister_event = mem_cgroup_usage_unregister_event,
4736 .name = "memsw.max_usage_in_bytes",
4737 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4738 .trigger = mem_cgroup_reset,
4739 .read_u64 = mem_cgroup_read,
4742 .name = "memsw.limit_in_bytes",
4743 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4744 .write_string = mem_cgroup_write,
4745 .read_u64 = mem_cgroup_read,
4748 .name = "memsw.failcnt",
4749 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4750 .trigger = mem_cgroup_reset,
4751 .read_u64 = mem_cgroup_read,
4755 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4757 if (!do_swap_account)
4759 return cgroup_add_files(cont, ss, memsw_cgroup_files,
4760 ARRAY_SIZE(memsw_cgroup_files));
4763 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4769 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4771 struct mem_cgroup_per_node *pn;
4772 struct mem_cgroup_per_zone *mz;
4774 int zone, tmp = node;
4776 * This routine is called against possible nodes.
4777 * But it's BUG to call kmalloc() against offline node.
4779 * TODO: this routine can waste much memory for nodes which will
4780 * never be onlined. It's better to use memory hotplug callback
4783 if (!node_state(node, N_NORMAL_MEMORY))
4785 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4789 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4790 mz = &pn->zoneinfo[zone];
4792 INIT_LIST_HEAD(&mz->lruvec.lists[lru]);
4793 mz->usage_in_excess = 0;
4794 mz->on_tree = false;
4797 memcg->info.nodeinfo[node] = pn;
4801 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4803 kfree(memcg->info.nodeinfo[node]);
4806 static struct mem_cgroup *mem_cgroup_alloc(void)
4808 struct mem_cgroup *memcg;
4809 int size = sizeof(struct mem_cgroup);
4811 /* Can be very big if MAX_NUMNODES is very big */
4812 if (size < PAGE_SIZE)
4813 memcg = kzalloc(size, GFP_KERNEL);
4815 memcg = vzalloc(size);
4820 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4823 spin_lock_init(&memcg->pcp_counter_lock);
4827 if (size < PAGE_SIZE)
4835 * Helpers for freeing a vzalloc()ed mem_cgroup by RCU,
4836 * but in process context. The work_freeing structure is overlaid
4837 * on the rcu_freeing structure, which itself is overlaid on memsw.
4839 static void vfree_work(struct work_struct *work)
4841 struct mem_cgroup *memcg;
4843 memcg = container_of(work, struct mem_cgroup, work_freeing);
4846 static void vfree_rcu(struct rcu_head *rcu_head)
4848 struct mem_cgroup *memcg;
4850 memcg = container_of(rcu_head, struct mem_cgroup, rcu_freeing);
4851 INIT_WORK(&memcg->work_freeing, vfree_work);
4852 schedule_work(&memcg->work_freeing);
4856 * At destroying mem_cgroup, references from swap_cgroup can remain.
4857 * (scanning all at force_empty is too costly...)
4859 * Instead of clearing all references at force_empty, we remember
4860 * the number of reference from swap_cgroup and free mem_cgroup when
4861 * it goes down to 0.
4863 * Removal of cgroup itself succeeds regardless of refs from swap.
4866 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4870 mem_cgroup_remove_from_trees(memcg);
4871 free_css_id(&mem_cgroup_subsys, &memcg->css);
4874 free_mem_cgroup_per_zone_info(memcg, node);
4876 free_percpu(memcg->stat);
4877 if (sizeof(struct mem_cgroup) < PAGE_SIZE)
4878 kfree_rcu(memcg, rcu_freeing);
4880 call_rcu(&memcg->rcu_freeing, vfree_rcu);
4883 static void mem_cgroup_get(struct mem_cgroup *memcg)
4885 atomic_inc(&memcg->refcnt);
4888 static void __mem_cgroup_put(struct mem_cgroup *memcg, int count)
4890 if (atomic_sub_and_test(count, &memcg->refcnt)) {
4891 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4892 __mem_cgroup_free(memcg);
4894 mem_cgroup_put(parent);
4898 static void mem_cgroup_put(struct mem_cgroup *memcg)
4900 __mem_cgroup_put(memcg, 1);
4904 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4906 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4908 if (!memcg->res.parent)
4910 return mem_cgroup_from_res_counter(memcg->res.parent, res);
4912 EXPORT_SYMBOL(parent_mem_cgroup);
4914 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4915 static void __init enable_swap_cgroup(void)
4917 if (!mem_cgroup_disabled() && really_do_swap_account)
4918 do_swap_account = 1;
4921 static void __init enable_swap_cgroup(void)
4926 static int mem_cgroup_soft_limit_tree_init(void)
4928 struct mem_cgroup_tree_per_node *rtpn;
4929 struct mem_cgroup_tree_per_zone *rtpz;
4930 int tmp, node, zone;
4932 for_each_node(node) {
4934 if (!node_state(node, N_NORMAL_MEMORY))
4936 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
4940 soft_limit_tree.rb_tree_per_node[node] = rtpn;
4942 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4943 rtpz = &rtpn->rb_tree_per_zone[zone];
4944 rtpz->rb_root = RB_ROOT;
4945 spin_lock_init(&rtpz->lock);
4951 for_each_node(node) {
4952 if (!soft_limit_tree.rb_tree_per_node[node])
4954 kfree(soft_limit_tree.rb_tree_per_node[node]);
4955 soft_limit_tree.rb_tree_per_node[node] = NULL;
4961 static struct cgroup_subsys_state * __ref
4962 mem_cgroup_create(struct cgroup *cont)
4964 struct mem_cgroup *memcg, *parent;
4965 long error = -ENOMEM;
4968 memcg = mem_cgroup_alloc();
4970 return ERR_PTR(error);
4973 if (alloc_mem_cgroup_per_zone_info(memcg, node))
4977 if (cont->parent == NULL) {
4979 enable_swap_cgroup();
4981 if (mem_cgroup_soft_limit_tree_init())
4983 root_mem_cgroup = memcg;
4984 for_each_possible_cpu(cpu) {
4985 struct memcg_stock_pcp *stock =
4986 &per_cpu(memcg_stock, cpu);
4987 INIT_WORK(&stock->work, drain_local_stock);
4989 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
4991 parent = mem_cgroup_from_cont(cont->parent);
4992 memcg->use_hierarchy = parent->use_hierarchy;
4993 memcg->oom_kill_disable = parent->oom_kill_disable;
4996 if (parent && parent->use_hierarchy) {
4997 res_counter_init(&memcg->res, &parent->res);
4998 res_counter_init(&memcg->memsw, &parent->memsw);
5000 * We increment refcnt of the parent to ensure that we can
5001 * safely access it on res_counter_charge/uncharge.
5002 * This refcnt will be decremented when freeing this
5003 * mem_cgroup(see mem_cgroup_put).
5005 mem_cgroup_get(parent);
5007 res_counter_init(&memcg->res, NULL);
5008 res_counter_init(&memcg->memsw, NULL);
5010 memcg->last_scanned_node = MAX_NUMNODES;
5011 INIT_LIST_HEAD(&memcg->oom_notify);
5014 memcg->swappiness = mem_cgroup_swappiness(parent);
5015 atomic_set(&memcg->refcnt, 1);
5016 memcg->move_charge_at_immigrate = 0;
5017 mutex_init(&memcg->thresholds_lock);
5018 spin_lock_init(&memcg->move_lock);
5021 __mem_cgroup_free(memcg);
5022 return ERR_PTR(error);
5025 static int mem_cgroup_pre_destroy(struct cgroup *cont)
5027 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5029 return mem_cgroup_force_empty(memcg, false);
5032 static void mem_cgroup_destroy(struct cgroup *cont)
5034 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5036 kmem_cgroup_destroy(cont);
5038 mem_cgroup_put(memcg);
5041 static int mem_cgroup_populate(struct cgroup_subsys *ss,
5042 struct cgroup *cont)
5046 ret = cgroup_add_files(cont, ss, mem_cgroup_files,
5047 ARRAY_SIZE(mem_cgroup_files));
5050 ret = register_memsw_files(cont, ss);
5053 ret = register_kmem_files(cont, ss);
5059 /* Handlers for move charge at task migration. */
5060 #define PRECHARGE_COUNT_AT_ONCE 256
5061 static int mem_cgroup_do_precharge(unsigned long count)
5064 int batch_count = PRECHARGE_COUNT_AT_ONCE;
5065 struct mem_cgroup *memcg = mc.to;
5067 if (mem_cgroup_is_root(memcg)) {
5068 mc.precharge += count;
5069 /* we don't need css_get for root */
5072 /* try to charge at once */
5074 struct res_counter *dummy;
5076 * "memcg" cannot be under rmdir() because we've already checked
5077 * by cgroup_lock_live_cgroup() that it is not removed and we
5078 * are still under the same cgroup_mutex. So we can postpone
5081 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
5083 if (do_swap_account && res_counter_charge(&memcg->memsw,
5084 PAGE_SIZE * count, &dummy)) {
5085 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
5088 mc.precharge += count;
5092 /* fall back to one by one charge */
5094 if (signal_pending(current)) {
5098 if (!batch_count--) {
5099 batch_count = PRECHARGE_COUNT_AT_ONCE;
5102 ret = __mem_cgroup_try_charge(NULL,
5103 GFP_KERNEL, 1, &memcg, false);
5105 /* mem_cgroup_clear_mc() will do uncharge later */
5113 * is_target_pte_for_mc - check a pte whether it is valid for move charge
5114 * @vma: the vma the pte to be checked belongs
5115 * @addr: the address corresponding to the pte to be checked
5116 * @ptent: the pte to be checked
5117 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5120 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5121 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5122 * move charge. if @target is not NULL, the page is stored in target->page
5123 * with extra refcnt got(Callers should handle it).
5124 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5125 * target for charge migration. if @target is not NULL, the entry is stored
5128 * Called with pte lock held.
5135 enum mc_target_type {
5136 MC_TARGET_NONE, /* not used */
5141 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5142 unsigned long addr, pte_t ptent)
5144 struct page *page = vm_normal_page(vma, addr, ptent);
5146 if (!page || !page_mapped(page))
5148 if (PageAnon(page)) {
5149 /* we don't move shared anon */
5150 if (!move_anon() || page_mapcount(page) > 2)
5152 } else if (!move_file())
5153 /* we ignore mapcount for file pages */
5155 if (!get_page_unless_zero(page))
5161 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5162 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5165 struct page *page = NULL;
5166 swp_entry_t ent = pte_to_swp_entry(ptent);
5168 if (!move_anon() || non_swap_entry(ent))
5170 usage_count = mem_cgroup_count_swap_user(ent, &page);
5171 if (usage_count > 1) { /* we don't move shared anon */
5176 if (do_swap_account)
5177 entry->val = ent.val;
5182 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5183 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5185 struct page *page = NULL;
5186 struct inode *inode;
5187 struct address_space *mapping;
5190 if (!vma->vm_file) /* anonymous vma */
5195 inode = vma->vm_file->f_path.dentry->d_inode;
5196 mapping = vma->vm_file->f_mapping;
5197 if (pte_none(ptent))
5198 pgoff = linear_page_index(vma, addr);
5199 else /* pte_file(ptent) is true */
5200 pgoff = pte_to_pgoff(ptent);
5202 /* page is moved even if it's not RSS of this task(page-faulted). */
5203 page = find_get_page(mapping, pgoff);
5206 /* shmem/tmpfs may report page out on swap: account for that too. */
5207 if (radix_tree_exceptional_entry(page)) {
5208 swp_entry_t swap = radix_to_swp_entry(page);
5209 if (do_swap_account)
5211 page = find_get_page(&swapper_space, swap.val);
5217 static int is_target_pte_for_mc(struct vm_area_struct *vma,
5218 unsigned long addr, pte_t ptent, union mc_target *target)
5220 struct page *page = NULL;
5221 struct page_cgroup *pc;
5223 swp_entry_t ent = { .val = 0 };
5225 if (pte_present(ptent))
5226 page = mc_handle_present_pte(vma, addr, ptent);
5227 else if (is_swap_pte(ptent))
5228 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5229 else if (pte_none(ptent) || pte_file(ptent))
5230 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5232 if (!page && !ent.val)
5235 pc = lookup_page_cgroup(page);
5237 * Do only loose check w/o page_cgroup lock.
5238 * mem_cgroup_move_account() checks the pc is valid or not under
5241 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5242 ret = MC_TARGET_PAGE;
5244 target->page = page;
5246 if (!ret || !target)
5249 /* There is a swap entry and a page doesn't exist or isn't charged */
5250 if (ent.val && !ret &&
5251 css_id(&mc.from->css) == lookup_swap_cgroup_id(ent)) {
5252 ret = MC_TARGET_SWAP;
5259 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5260 unsigned long addr, unsigned long end,
5261 struct mm_walk *walk)
5263 struct vm_area_struct *vma = walk->private;
5267 split_huge_page_pmd(walk->mm, pmd);
5268 if (pmd_trans_unstable(pmd))
5271 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5272 for (; addr != end; pte++, addr += PAGE_SIZE)
5273 if (is_target_pte_for_mc(vma, addr, *pte, NULL))
5274 mc.precharge++; /* increment precharge temporarily */
5275 pte_unmap_unlock(pte - 1, ptl);
5281 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5283 unsigned long precharge;
5284 struct vm_area_struct *vma;
5286 down_read(&mm->mmap_sem);
5287 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5288 struct mm_walk mem_cgroup_count_precharge_walk = {
5289 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5293 if (is_vm_hugetlb_page(vma))
5295 walk_page_range(vma->vm_start, vma->vm_end,
5296 &mem_cgroup_count_precharge_walk);
5298 up_read(&mm->mmap_sem);
5300 precharge = mc.precharge;
5306 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5308 unsigned long precharge = mem_cgroup_count_precharge(mm);
5310 VM_BUG_ON(mc.moving_task);
5311 mc.moving_task = current;
5312 return mem_cgroup_do_precharge(precharge);
5315 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5316 static void __mem_cgroup_clear_mc(void)
5318 struct mem_cgroup *from = mc.from;
5319 struct mem_cgroup *to = mc.to;
5321 /* we must uncharge all the leftover precharges from mc.to */
5323 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
5327 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5328 * we must uncharge here.
5330 if (mc.moved_charge) {
5331 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
5332 mc.moved_charge = 0;
5334 /* we must fixup refcnts and charges */
5335 if (mc.moved_swap) {
5336 /* uncharge swap account from the old cgroup */
5337 if (!mem_cgroup_is_root(mc.from))
5338 res_counter_uncharge(&mc.from->memsw,
5339 PAGE_SIZE * mc.moved_swap);
5340 __mem_cgroup_put(mc.from, mc.moved_swap);
5342 if (!mem_cgroup_is_root(mc.to)) {
5344 * we charged both to->res and to->memsw, so we should
5347 res_counter_uncharge(&mc.to->res,
5348 PAGE_SIZE * mc.moved_swap);
5350 /* we've already done mem_cgroup_get(mc.to) */
5353 memcg_oom_recover(from);
5354 memcg_oom_recover(to);
5355 wake_up_all(&mc.waitq);
5358 static void mem_cgroup_clear_mc(void)
5360 struct mem_cgroup *from = mc.from;
5363 * we must clear moving_task before waking up waiters at the end of
5366 mc.moving_task = NULL;
5367 __mem_cgroup_clear_mc();
5368 spin_lock(&mc.lock);
5371 spin_unlock(&mc.lock);
5372 mem_cgroup_end_move(from);
5375 static int mem_cgroup_can_attach(struct cgroup *cgroup,
5376 struct cgroup_taskset *tset)
5378 struct task_struct *p = cgroup_taskset_first(tset);
5380 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup);
5382 if (memcg->move_charge_at_immigrate) {
5383 struct mm_struct *mm;
5384 struct mem_cgroup *from = mem_cgroup_from_task(p);
5386 VM_BUG_ON(from == memcg);
5388 mm = get_task_mm(p);
5391 /* We move charges only when we move a owner of the mm */
5392 if (mm->owner == p) {
5395 VM_BUG_ON(mc.precharge);
5396 VM_BUG_ON(mc.moved_charge);
5397 VM_BUG_ON(mc.moved_swap);
5398 mem_cgroup_start_move(from);
5399 spin_lock(&mc.lock);
5402 spin_unlock(&mc.lock);
5403 /* We set mc.moving_task later */
5405 ret = mem_cgroup_precharge_mc(mm);
5407 mem_cgroup_clear_mc();
5414 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
5415 struct cgroup_taskset *tset)
5417 mem_cgroup_clear_mc();
5420 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5421 unsigned long addr, unsigned long end,
5422 struct mm_walk *walk)
5425 struct vm_area_struct *vma = walk->private;
5429 split_huge_page_pmd(walk->mm, pmd);
5430 if (pmd_trans_unstable(pmd))
5433 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5434 for (; addr != end; addr += PAGE_SIZE) {
5435 pte_t ptent = *(pte++);
5436 union mc_target target;
5439 struct page_cgroup *pc;
5445 type = is_target_pte_for_mc(vma, addr, ptent, &target);
5447 case MC_TARGET_PAGE:
5449 if (isolate_lru_page(page))
5451 pc = lookup_page_cgroup(page);
5452 if (!mem_cgroup_move_account(page, 1, pc,
5453 mc.from, mc.to, false)) {
5455 /* we uncharge from mc.from later. */
5458 putback_lru_page(page);
5459 put: /* is_target_pte_for_mc() gets the page */
5462 case MC_TARGET_SWAP:
5464 if (!mem_cgroup_move_swap_account(ent,
5465 mc.from, mc.to, false)) {
5467 /* we fixup refcnts and charges later. */
5475 pte_unmap_unlock(pte - 1, ptl);
5480 * We have consumed all precharges we got in can_attach().
5481 * We try charge one by one, but don't do any additional
5482 * charges to mc.to if we have failed in charge once in attach()
5485 ret = mem_cgroup_do_precharge(1);
5493 static void mem_cgroup_move_charge(struct mm_struct *mm)
5495 struct vm_area_struct *vma;
5497 lru_add_drain_all();
5499 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5501 * Someone who are holding the mmap_sem might be waiting in
5502 * waitq. So we cancel all extra charges, wake up all waiters,
5503 * and retry. Because we cancel precharges, we might not be able
5504 * to move enough charges, but moving charge is a best-effort
5505 * feature anyway, so it wouldn't be a big problem.
5507 __mem_cgroup_clear_mc();
5511 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5513 struct mm_walk mem_cgroup_move_charge_walk = {
5514 .pmd_entry = mem_cgroup_move_charge_pte_range,
5518 if (is_vm_hugetlb_page(vma))
5520 ret = walk_page_range(vma->vm_start, vma->vm_end,
5521 &mem_cgroup_move_charge_walk);
5524 * means we have consumed all precharges and failed in
5525 * doing additional charge. Just abandon here.
5529 up_read(&mm->mmap_sem);
5532 static void mem_cgroup_move_task(struct cgroup *cont,
5533 struct cgroup_taskset *tset)
5535 struct task_struct *p = cgroup_taskset_first(tset);
5536 struct mm_struct *mm = get_task_mm(p);
5540 mem_cgroup_move_charge(mm);
5545 mem_cgroup_clear_mc();
5547 #else /* !CONFIG_MMU */
5548 static int mem_cgroup_can_attach(struct cgroup *cgroup,
5549 struct cgroup_taskset *tset)
5553 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
5554 struct cgroup_taskset *tset)
5557 static void mem_cgroup_move_task(struct cgroup *cont,
5558 struct cgroup_taskset *tset)
5563 struct cgroup_subsys mem_cgroup_subsys = {
5565 .subsys_id = mem_cgroup_subsys_id,
5566 .create = mem_cgroup_create,
5567 .pre_destroy = mem_cgroup_pre_destroy,
5568 .destroy = mem_cgroup_destroy,
5569 .populate = mem_cgroup_populate,
5570 .can_attach = mem_cgroup_can_attach,
5571 .cancel_attach = mem_cgroup_cancel_attach,
5572 .attach = mem_cgroup_move_task,
5577 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5578 static int __init enable_swap_account(char *s)
5580 /* consider enabled if no parameter or 1 is given */
5581 if (!strcmp(s, "1"))
5582 really_do_swap_account = 1;
5583 else if (!strcmp(s, "0"))
5584 really_do_swap_account = 0;
5587 __setup("swapaccount=", enable_swap_account);