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:
1970 SetPageCgroupFileMapped(pc);
1971 else if (!page_mapped(page))
1972 ClearPageCgroupFileMapped(pc);
1973 idx = MEM_CGROUP_STAT_FILE_MAPPED;
1979 this_cpu_add(memcg->stat->count[idx], val);
1983 * size of first charge trial. "32" comes from vmscan.c's magic value.
1984 * TODO: maybe necessary to use big numbers in big irons.
1986 #define CHARGE_BATCH 32U
1987 struct memcg_stock_pcp {
1988 struct mem_cgroup *cached; /* this never be root cgroup */
1989 unsigned int nr_pages;
1990 struct work_struct work;
1991 unsigned long flags;
1992 #define FLUSHING_CACHED_CHARGE (0)
1994 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1995 static DEFINE_MUTEX(percpu_charge_mutex);
1998 * Try to consume stocked charge on this cpu. If success, one page is consumed
1999 * from local stock and true is returned. If the stock is 0 or charges from a
2000 * cgroup which is not current target, returns false. This stock will be
2003 static bool consume_stock(struct mem_cgroup *memcg)
2005 struct memcg_stock_pcp *stock;
2008 stock = &get_cpu_var(memcg_stock);
2009 if (memcg == stock->cached && stock->nr_pages)
2011 else /* need to call res_counter_charge */
2013 put_cpu_var(memcg_stock);
2018 * Returns stocks cached in percpu to res_counter and reset cached information.
2020 static void drain_stock(struct memcg_stock_pcp *stock)
2022 struct mem_cgroup *old = stock->cached;
2024 if (stock->nr_pages) {
2025 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2027 res_counter_uncharge(&old->res, bytes);
2028 if (do_swap_account)
2029 res_counter_uncharge(&old->memsw, bytes);
2030 stock->nr_pages = 0;
2032 stock->cached = NULL;
2036 * This must be called under preempt disabled or must be called by
2037 * a thread which is pinned to local cpu.
2039 static void drain_local_stock(struct work_struct *dummy)
2041 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2043 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2047 * Cache charges(val) which is from res_counter, to local per_cpu area.
2048 * This will be consumed by consume_stock() function, later.
2050 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2052 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2054 if (stock->cached != memcg) { /* reset if necessary */
2056 stock->cached = memcg;
2058 stock->nr_pages += nr_pages;
2059 put_cpu_var(memcg_stock);
2063 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2064 * of the hierarchy under it. sync flag says whether we should block
2065 * until the work is done.
2067 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2071 /* Notify other cpus that system-wide "drain" is running */
2074 for_each_online_cpu(cpu) {
2075 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2076 struct mem_cgroup *memcg;
2078 memcg = stock->cached;
2079 if (!memcg || !stock->nr_pages)
2081 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2083 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2085 drain_local_stock(&stock->work);
2087 schedule_work_on(cpu, &stock->work);
2095 for_each_online_cpu(cpu) {
2096 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2097 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2098 flush_work(&stock->work);
2105 * Tries to drain stocked charges in other cpus. This function is asynchronous
2106 * and just put a work per cpu for draining localy on each cpu. Caller can
2107 * expects some charges will be back to res_counter later but cannot wait for
2110 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2113 * If someone calls draining, avoid adding more kworker runs.
2115 if (!mutex_trylock(&percpu_charge_mutex))
2117 drain_all_stock(root_memcg, false);
2118 mutex_unlock(&percpu_charge_mutex);
2121 /* This is a synchronous drain interface. */
2122 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2124 /* called when force_empty is called */
2125 mutex_lock(&percpu_charge_mutex);
2126 drain_all_stock(root_memcg, true);
2127 mutex_unlock(&percpu_charge_mutex);
2131 * This function drains percpu counter value from DEAD cpu and
2132 * move it to local cpu. Note that this function can be preempted.
2134 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2138 spin_lock(&memcg->pcp_counter_lock);
2139 for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) {
2140 long x = per_cpu(memcg->stat->count[i], cpu);
2142 per_cpu(memcg->stat->count[i], cpu) = 0;
2143 memcg->nocpu_base.count[i] += x;
2145 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2146 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2148 per_cpu(memcg->stat->events[i], cpu) = 0;
2149 memcg->nocpu_base.events[i] += x;
2151 spin_unlock(&memcg->pcp_counter_lock);
2154 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2155 unsigned long action,
2158 int cpu = (unsigned long)hcpu;
2159 struct memcg_stock_pcp *stock;
2160 struct mem_cgroup *iter;
2162 if (action == CPU_ONLINE)
2165 if ((action != CPU_DEAD) || action != CPU_DEAD_FROZEN)
2168 for_each_mem_cgroup(iter)
2169 mem_cgroup_drain_pcp_counter(iter, cpu);
2171 stock = &per_cpu(memcg_stock, cpu);
2177 /* See __mem_cgroup_try_charge() for details */
2179 CHARGE_OK, /* success */
2180 CHARGE_RETRY, /* need to retry but retry is not bad */
2181 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2182 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2183 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2186 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2187 unsigned int nr_pages, bool oom_check)
2189 unsigned long csize = nr_pages * PAGE_SIZE;
2190 struct mem_cgroup *mem_over_limit;
2191 struct res_counter *fail_res;
2192 unsigned long flags = 0;
2195 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2198 if (!do_swap_account)
2200 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2204 res_counter_uncharge(&memcg->res, csize);
2205 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2206 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2208 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2210 * nr_pages can be either a huge page (HPAGE_PMD_NR), a batch
2211 * of regular pages (CHARGE_BATCH), or a single regular page (1).
2213 * Never reclaim on behalf of optional batching, retry with a
2214 * single page instead.
2216 if (nr_pages == CHARGE_BATCH)
2217 return CHARGE_RETRY;
2219 if (!(gfp_mask & __GFP_WAIT))
2220 return CHARGE_WOULDBLOCK;
2222 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2223 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2224 return CHARGE_RETRY;
2226 * Even though the limit is exceeded at this point, reclaim
2227 * may have been able to free some pages. Retry the charge
2228 * before killing the task.
2230 * Only for regular pages, though: huge pages are rather
2231 * unlikely to succeed so close to the limit, and we fall back
2232 * to regular pages anyway in case of failure.
2234 if (nr_pages == 1 && ret)
2235 return CHARGE_RETRY;
2238 * At task move, charge accounts can be doubly counted. So, it's
2239 * better to wait until the end of task_move if something is going on.
2241 if (mem_cgroup_wait_acct_move(mem_over_limit))
2242 return CHARGE_RETRY;
2244 /* If we don't need to call oom-killer at el, return immediately */
2246 return CHARGE_NOMEM;
2248 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask, get_order(csize)))
2249 return CHARGE_OOM_DIE;
2251 return CHARGE_RETRY;
2255 * __mem_cgroup_try_charge() does
2256 * 1. detect memcg to be charged against from passed *mm and *ptr,
2257 * 2. update res_counter
2258 * 3. call memory reclaim if necessary.
2260 * In some special case, if the task is fatal, fatal_signal_pending() or
2261 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2262 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2263 * as possible without any hazards. 2: all pages should have a valid
2264 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2265 * pointer, that is treated as a charge to root_mem_cgroup.
2267 * So __mem_cgroup_try_charge() will return
2268 * 0 ... on success, filling *ptr with a valid memcg pointer.
2269 * -ENOMEM ... charge failure because of resource limits.
2270 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2272 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2273 * the oom-killer can be invoked.
2275 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2277 unsigned int nr_pages,
2278 struct mem_cgroup **ptr,
2281 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2282 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2283 struct mem_cgroup *memcg = NULL;
2287 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2288 * in system level. So, allow to go ahead dying process in addition to
2291 if (unlikely(test_thread_flag(TIF_MEMDIE)
2292 || fatal_signal_pending(current)))
2296 * We always charge the cgroup the mm_struct belongs to.
2297 * The mm_struct's mem_cgroup changes on task migration if the
2298 * thread group leader migrates. It's possible that mm is not
2299 * set, if so charge the init_mm (happens for pagecache usage).
2302 *ptr = root_mem_cgroup;
2304 if (*ptr) { /* css should be a valid one */
2306 VM_BUG_ON(css_is_removed(&memcg->css));
2307 if (mem_cgroup_is_root(memcg))
2309 if (nr_pages == 1 && consume_stock(memcg))
2311 css_get(&memcg->css);
2313 struct task_struct *p;
2316 p = rcu_dereference(mm->owner);
2318 * Because we don't have task_lock(), "p" can exit.
2319 * In that case, "memcg" can point to root or p can be NULL with
2320 * race with swapoff. Then, we have small risk of mis-accouning.
2321 * But such kind of mis-account by race always happens because
2322 * we don't have cgroup_mutex(). It's overkill and we allo that
2324 * (*) swapoff at el will charge against mm-struct not against
2325 * task-struct. So, mm->owner can be NULL.
2327 memcg = mem_cgroup_from_task(p);
2329 memcg = root_mem_cgroup;
2330 if (mem_cgroup_is_root(memcg)) {
2334 if (nr_pages == 1 && consume_stock(memcg)) {
2336 * It seems dagerous to access memcg without css_get().
2337 * But considering how consume_stok works, it's not
2338 * necessary. If consume_stock success, some charges
2339 * from this memcg are cached on this cpu. So, we
2340 * don't need to call css_get()/css_tryget() before
2341 * calling consume_stock().
2346 /* after here, we may be blocked. we need to get refcnt */
2347 if (!css_tryget(&memcg->css)) {
2357 /* If killed, bypass charge */
2358 if (fatal_signal_pending(current)) {
2359 css_put(&memcg->css);
2364 if (oom && !nr_oom_retries) {
2366 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2369 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, oom_check);
2373 case CHARGE_RETRY: /* not in OOM situation but retry */
2375 css_put(&memcg->css);
2378 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2379 css_put(&memcg->css);
2381 case CHARGE_NOMEM: /* OOM routine works */
2383 css_put(&memcg->css);
2386 /* If oom, we never return -ENOMEM */
2389 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2390 css_put(&memcg->css);
2393 } while (ret != CHARGE_OK);
2395 if (batch > nr_pages)
2396 refill_stock(memcg, batch - nr_pages);
2397 css_put(&memcg->css);
2405 *ptr = root_mem_cgroup;
2410 * Somemtimes we have to undo a charge we got by try_charge().
2411 * This function is for that and do uncharge, put css's refcnt.
2412 * gotten by try_charge().
2414 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2415 unsigned int nr_pages)
2417 if (!mem_cgroup_is_root(memcg)) {
2418 unsigned long bytes = nr_pages * PAGE_SIZE;
2420 res_counter_uncharge(&memcg->res, bytes);
2421 if (do_swap_account)
2422 res_counter_uncharge(&memcg->memsw, bytes);
2427 * A helper function to get mem_cgroup from ID. must be called under
2428 * rcu_read_lock(). The caller must check css_is_removed() or some if
2429 * it's concern. (dropping refcnt from swap can be called against removed
2432 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2434 struct cgroup_subsys_state *css;
2436 /* ID 0 is unused ID */
2439 css = css_lookup(&mem_cgroup_subsys, id);
2442 return container_of(css, struct mem_cgroup, css);
2445 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2447 struct mem_cgroup *memcg = NULL;
2448 struct page_cgroup *pc;
2452 VM_BUG_ON(!PageLocked(page));
2454 pc = lookup_page_cgroup(page);
2455 lock_page_cgroup(pc);
2456 if (PageCgroupUsed(pc)) {
2457 memcg = pc->mem_cgroup;
2458 if (memcg && !css_tryget(&memcg->css))
2460 } else if (PageSwapCache(page)) {
2461 ent.val = page_private(page);
2462 id = lookup_swap_cgroup_id(ent);
2464 memcg = mem_cgroup_lookup(id);
2465 if (memcg && !css_tryget(&memcg->css))
2469 unlock_page_cgroup(pc);
2473 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2475 unsigned int nr_pages,
2476 struct page_cgroup *pc,
2477 enum charge_type ctype,
2480 struct zone *uninitialized_var(zone);
2481 bool was_on_lru = false;
2484 lock_page_cgroup(pc);
2485 if (unlikely(PageCgroupUsed(pc))) {
2486 unlock_page_cgroup(pc);
2487 __mem_cgroup_cancel_charge(memcg, nr_pages);
2491 * we don't need page_cgroup_lock about tail pages, becase they are not
2492 * accessed by any other context at this point.
2496 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2497 * may already be on some other mem_cgroup's LRU. Take care of it.
2500 zone = page_zone(page);
2501 spin_lock_irq(&zone->lru_lock);
2502 if (PageLRU(page)) {
2504 del_page_from_lru_list(zone, page, page_lru(page));
2509 pc->mem_cgroup = memcg;
2511 * We access a page_cgroup asynchronously without lock_page_cgroup().
2512 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2513 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2514 * before USED bit, we need memory barrier here.
2515 * See mem_cgroup_add_lru_list(), etc.
2518 SetPageCgroupUsed(pc);
2522 VM_BUG_ON(PageLRU(page));
2524 add_page_to_lru_list(zone, page, page_lru(page));
2526 spin_unlock_irq(&zone->lru_lock);
2529 if (ctype == MEM_CGROUP_CHARGE_TYPE_MAPPED)
2534 mem_cgroup_charge_statistics(memcg, anon, nr_pages);
2535 unlock_page_cgroup(pc);
2538 * "charge_statistics" updated event counter. Then, check it.
2539 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2540 * if they exceeds softlimit.
2542 memcg_check_events(memcg, page);
2545 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2547 #define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MIGRATION))
2549 * Because tail pages are not marked as "used", set it. We're under
2550 * zone->lru_lock, 'splitting on pmd' and compound_lock.
2551 * charge/uncharge will be never happen and move_account() is done under
2552 * compound_lock(), so we don't have to take care of races.
2554 void mem_cgroup_split_huge_fixup(struct page *head)
2556 struct page_cgroup *head_pc = lookup_page_cgroup(head);
2557 struct page_cgroup *pc;
2560 if (mem_cgroup_disabled())
2562 for (i = 1; i < HPAGE_PMD_NR; i++) {
2564 pc->mem_cgroup = head_pc->mem_cgroup;
2565 smp_wmb();/* see __commit_charge() */
2566 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2569 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2572 * mem_cgroup_move_account - move account of the page
2574 * @nr_pages: number of regular pages (>1 for huge pages)
2575 * @pc: page_cgroup of the page.
2576 * @from: mem_cgroup which the page is moved from.
2577 * @to: mem_cgroup which the page is moved to. @from != @to.
2578 * @uncharge: whether we should call uncharge and css_put against @from.
2580 * The caller must confirm following.
2581 * - page is not on LRU (isolate_page() is useful.)
2582 * - compound_lock is held when nr_pages > 1
2584 * This function doesn't do "charge" nor css_get to new cgroup. It should be
2585 * done by a caller(__mem_cgroup_try_charge would be useful). If @uncharge is
2586 * true, this function does "uncharge" from old cgroup, but it doesn't if
2587 * @uncharge is false, so a caller should do "uncharge".
2589 static int mem_cgroup_move_account(struct page *page,
2590 unsigned int nr_pages,
2591 struct page_cgroup *pc,
2592 struct mem_cgroup *from,
2593 struct mem_cgroup *to,
2596 unsigned long flags;
2598 bool anon = PageAnon(page);
2600 VM_BUG_ON(from == to);
2601 VM_BUG_ON(PageLRU(page));
2603 * The page is isolated from LRU. So, collapse function
2604 * will not handle this page. But page splitting can happen.
2605 * Do this check under compound_page_lock(). The caller should
2609 if (nr_pages > 1 && !PageTransHuge(page))
2612 lock_page_cgroup(pc);
2615 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
2618 move_lock_mem_cgroup(from, &flags);
2620 if (PageCgroupFileMapped(pc)) {
2621 /* Update mapped_file data for mem_cgroup */
2623 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2624 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2627 mem_cgroup_charge_statistics(from, anon, -nr_pages);
2629 /* This is not "cancel", but cancel_charge does all we need. */
2630 __mem_cgroup_cancel_charge(from, nr_pages);
2632 /* caller should have done css_get */
2633 pc->mem_cgroup = to;
2634 mem_cgroup_charge_statistics(to, anon, nr_pages);
2636 * We charges against "to" which may not have any tasks. Then, "to"
2637 * can be under rmdir(). But in current implementation, caller of
2638 * this function is just force_empty() and move charge, so it's
2639 * guaranteed that "to" is never removed. So, we don't check rmdir
2642 move_unlock_mem_cgroup(from, &flags);
2645 unlock_page_cgroup(pc);
2649 memcg_check_events(to, page);
2650 memcg_check_events(from, page);
2656 * move charges to its parent.
2659 static int mem_cgroup_move_parent(struct page *page,
2660 struct page_cgroup *pc,
2661 struct mem_cgroup *child,
2664 struct cgroup *cg = child->css.cgroup;
2665 struct cgroup *pcg = cg->parent;
2666 struct mem_cgroup *parent;
2667 unsigned int nr_pages;
2668 unsigned long uninitialized_var(flags);
2676 if (!get_page_unless_zero(page))
2678 if (isolate_lru_page(page))
2681 nr_pages = hpage_nr_pages(page);
2683 parent = mem_cgroup_from_cont(pcg);
2684 ret = __mem_cgroup_try_charge(NULL, gfp_mask, nr_pages, &parent, false);
2689 flags = compound_lock_irqsave(page);
2691 ret = mem_cgroup_move_account(page, nr_pages, pc, child, parent, true);
2693 __mem_cgroup_cancel_charge(parent, nr_pages);
2696 compound_unlock_irqrestore(page, flags);
2698 putback_lru_page(page);
2706 * Charge the memory controller for page usage.
2708 * 0 if the charge was successful
2709 * < 0 if the cgroup is over its limit
2711 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2712 gfp_t gfp_mask, enum charge_type ctype)
2714 struct mem_cgroup *memcg = NULL;
2715 unsigned int nr_pages = 1;
2716 struct page_cgroup *pc;
2720 if (PageTransHuge(page)) {
2721 nr_pages <<= compound_order(page);
2722 VM_BUG_ON(!PageTransHuge(page));
2724 * Never OOM-kill a process for a huge page. The
2725 * fault handler will fall back to regular pages.
2730 pc = lookup_page_cgroup(page);
2731 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
2734 __mem_cgroup_commit_charge(memcg, page, nr_pages, pc, ctype, false);
2738 int mem_cgroup_newpage_charge(struct page *page,
2739 struct mm_struct *mm, gfp_t gfp_mask)
2741 if (mem_cgroup_disabled())
2743 VM_BUG_ON(page_mapped(page));
2744 VM_BUG_ON(page->mapping && !PageAnon(page));
2746 return mem_cgroup_charge_common(page, mm, gfp_mask,
2747 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2751 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2752 enum charge_type ctype);
2754 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2757 struct mem_cgroup *memcg = NULL;
2758 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
2761 if (mem_cgroup_disabled())
2763 if (PageCompound(page))
2768 if (!page_is_file_cache(page))
2769 type = MEM_CGROUP_CHARGE_TYPE_SHMEM;
2771 if (!PageSwapCache(page))
2772 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
2773 else { /* page is swapcache/shmem */
2774 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &memcg);
2776 __mem_cgroup_commit_charge_swapin(page, memcg, type);
2782 * While swap-in, try_charge -> commit or cancel, the page is locked.
2783 * And when try_charge() successfully returns, one refcnt to memcg without
2784 * struct page_cgroup is acquired. This refcnt will be consumed by
2785 * "commit()" or removed by "cancel()"
2787 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2789 gfp_t mask, struct mem_cgroup **memcgp)
2791 struct mem_cgroup *memcg;
2796 if (mem_cgroup_disabled())
2799 if (!do_swap_account)
2802 * A racing thread's fault, or swapoff, may have already updated
2803 * the pte, and even removed page from swap cache: in those cases
2804 * do_swap_page()'s pte_same() test will fail; but there's also a
2805 * KSM case which does need to charge the page.
2807 if (!PageSwapCache(page))
2809 memcg = try_get_mem_cgroup_from_page(page);
2813 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
2814 css_put(&memcg->css);
2821 ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
2828 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
2829 enum charge_type ctype)
2831 struct page_cgroup *pc;
2833 if (mem_cgroup_disabled())
2837 cgroup_exclude_rmdir(&memcg->css);
2839 pc = lookup_page_cgroup(page);
2840 __mem_cgroup_commit_charge(memcg, page, 1, pc, ctype, true);
2842 * Now swap is on-memory. This means this page may be
2843 * counted both as mem and swap....double count.
2844 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2845 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2846 * may call delete_from_swap_cache() before reach here.
2848 if (do_swap_account && PageSwapCache(page)) {
2849 swp_entry_t ent = {.val = page_private(page)};
2850 struct mem_cgroup *swap_memcg;
2853 id = swap_cgroup_record(ent, 0);
2855 swap_memcg = mem_cgroup_lookup(id);
2858 * This recorded memcg can be obsolete one. So, avoid
2859 * calling css_tryget
2861 if (!mem_cgroup_is_root(swap_memcg))
2862 res_counter_uncharge(&swap_memcg->memsw,
2864 mem_cgroup_swap_statistics(swap_memcg, false);
2865 mem_cgroup_put(swap_memcg);
2870 * At swapin, we may charge account against cgroup which has no tasks.
2871 * So, rmdir()->pre_destroy() can be called while we do this charge.
2872 * In that case, we need to call pre_destroy() again. check it here.
2874 cgroup_release_and_wakeup_rmdir(&memcg->css);
2877 void mem_cgroup_commit_charge_swapin(struct page *page,
2878 struct mem_cgroup *memcg)
2880 __mem_cgroup_commit_charge_swapin(page, memcg,
2881 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2884 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
2886 if (mem_cgroup_disabled())
2890 __mem_cgroup_cancel_charge(memcg, 1);
2893 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
2894 unsigned int nr_pages,
2895 const enum charge_type ctype)
2897 struct memcg_batch_info *batch = NULL;
2898 bool uncharge_memsw = true;
2900 /* If swapout, usage of swap doesn't decrease */
2901 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2902 uncharge_memsw = false;
2904 batch = ¤t->memcg_batch;
2906 * In usual, we do css_get() when we remember memcg pointer.
2907 * But in this case, we keep res->usage until end of a series of
2908 * uncharges. Then, it's ok to ignore memcg's refcnt.
2911 batch->memcg = memcg;
2913 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2914 * In those cases, all pages freed continuously can be expected to be in
2915 * the same cgroup and we have chance to coalesce uncharges.
2916 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2917 * because we want to do uncharge as soon as possible.
2920 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
2921 goto direct_uncharge;
2924 goto direct_uncharge;
2927 * In typical case, batch->memcg == mem. This means we can
2928 * merge a series of uncharges to an uncharge of res_counter.
2929 * If not, we uncharge res_counter ony by one.
2931 if (batch->memcg != memcg)
2932 goto direct_uncharge;
2933 /* remember freed charge and uncharge it later */
2936 batch->memsw_nr_pages++;
2939 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
2941 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
2942 if (unlikely(batch->memcg != memcg))
2943 memcg_oom_recover(memcg);
2947 * uncharge if !page_mapped(page)
2949 static struct mem_cgroup *
2950 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
2952 struct mem_cgroup *memcg = NULL;
2953 unsigned int nr_pages = 1;
2954 struct page_cgroup *pc;
2957 if (mem_cgroup_disabled())
2960 if (PageSwapCache(page))
2963 if (PageTransHuge(page)) {
2964 nr_pages <<= compound_order(page);
2965 VM_BUG_ON(!PageTransHuge(page));
2968 * Check if our page_cgroup is valid
2970 pc = lookup_page_cgroup(page);
2971 if (unlikely(!PageCgroupUsed(pc)))
2974 lock_page_cgroup(pc);
2976 memcg = pc->mem_cgroup;
2978 if (!PageCgroupUsed(pc))
2981 anon = PageAnon(page);
2984 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2987 case MEM_CGROUP_CHARGE_TYPE_DROP:
2988 /* See mem_cgroup_prepare_migration() */
2989 if (page_mapped(page) || PageCgroupMigration(pc))
2992 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
2993 if (!PageAnon(page)) { /* Shared memory */
2994 if (page->mapping && !page_is_file_cache(page))
2996 } else if (page_mapped(page)) /* Anon */
3003 mem_cgroup_charge_statistics(memcg, anon, -nr_pages);
3005 ClearPageCgroupUsed(pc);
3007 * pc->mem_cgroup is not cleared here. It will be accessed when it's
3008 * freed from LRU. This is safe because uncharged page is expected not
3009 * to be reused (freed soon). Exception is SwapCache, it's handled by
3010 * special functions.
3013 unlock_page_cgroup(pc);
3015 * even after unlock, we have memcg->res.usage here and this memcg
3016 * will never be freed.
3018 memcg_check_events(memcg, page);
3019 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
3020 mem_cgroup_swap_statistics(memcg, true);
3021 mem_cgroup_get(memcg);
3023 if (!mem_cgroup_is_root(memcg))
3024 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
3029 unlock_page_cgroup(pc);
3033 void mem_cgroup_uncharge_page(struct page *page)
3036 if (page_mapped(page))
3038 VM_BUG_ON(page->mapping && !PageAnon(page));
3039 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
3042 void mem_cgroup_uncharge_cache_page(struct page *page)
3044 VM_BUG_ON(page_mapped(page));
3045 VM_BUG_ON(page->mapping);
3046 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
3050 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
3051 * In that cases, pages are freed continuously and we can expect pages
3052 * are in the same memcg. All these calls itself limits the number of
3053 * pages freed at once, then uncharge_start/end() is called properly.
3054 * This may be called prural(2) times in a context,
3057 void mem_cgroup_uncharge_start(void)
3059 current->memcg_batch.do_batch++;
3060 /* We can do nest. */
3061 if (current->memcg_batch.do_batch == 1) {
3062 current->memcg_batch.memcg = NULL;
3063 current->memcg_batch.nr_pages = 0;
3064 current->memcg_batch.memsw_nr_pages = 0;
3068 void mem_cgroup_uncharge_end(void)
3070 struct memcg_batch_info *batch = ¤t->memcg_batch;
3072 if (!batch->do_batch)
3076 if (batch->do_batch) /* If stacked, do nothing. */
3082 * This "batch->memcg" is valid without any css_get/put etc...
3083 * bacause we hide charges behind us.
3085 if (batch->nr_pages)
3086 res_counter_uncharge(&batch->memcg->res,
3087 batch->nr_pages * PAGE_SIZE);
3088 if (batch->memsw_nr_pages)
3089 res_counter_uncharge(&batch->memcg->memsw,
3090 batch->memsw_nr_pages * PAGE_SIZE);
3091 memcg_oom_recover(batch->memcg);
3092 /* forget this pointer (for sanity check) */
3093 batch->memcg = NULL;
3098 * called after __delete_from_swap_cache() and drop "page" account.
3099 * memcg information is recorded to swap_cgroup of "ent"
3102 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
3104 struct mem_cgroup *memcg;
3105 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
3107 if (!swapout) /* this was a swap cache but the swap is unused ! */
3108 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
3110 memcg = __mem_cgroup_uncharge_common(page, ctype);
3113 * record memcg information, if swapout && memcg != NULL,
3114 * mem_cgroup_get() was called in uncharge().
3116 if (do_swap_account && swapout && memcg)
3117 swap_cgroup_record(ent, css_id(&memcg->css));
3121 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3123 * called from swap_entry_free(). remove record in swap_cgroup and
3124 * uncharge "memsw" account.
3126 void mem_cgroup_uncharge_swap(swp_entry_t ent)
3128 struct mem_cgroup *memcg;
3131 if (!do_swap_account)
3134 id = swap_cgroup_record(ent, 0);
3136 memcg = mem_cgroup_lookup(id);
3139 * We uncharge this because swap is freed.
3140 * This memcg can be obsolete one. We avoid calling css_tryget
3142 if (!mem_cgroup_is_root(memcg))
3143 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
3144 mem_cgroup_swap_statistics(memcg, false);
3145 mem_cgroup_put(memcg);
3151 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3152 * @entry: swap entry to be moved
3153 * @from: mem_cgroup which the entry is moved from
3154 * @to: mem_cgroup which the entry is moved to
3155 * @need_fixup: whether we should fixup res_counters and refcounts.
3157 * It succeeds only when the swap_cgroup's record for this entry is the same
3158 * as the mem_cgroup's id of @from.
3160 * Returns 0 on success, -EINVAL on failure.
3162 * The caller must have charged to @to, IOW, called res_counter_charge() about
3163 * both res and memsw, and called css_get().
3165 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3166 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3168 unsigned short old_id, new_id;
3170 old_id = css_id(&from->css);
3171 new_id = css_id(&to->css);
3173 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3174 mem_cgroup_swap_statistics(from, false);
3175 mem_cgroup_swap_statistics(to, true);
3177 * This function is only called from task migration context now.
3178 * It postpones res_counter and refcount handling till the end
3179 * of task migration(mem_cgroup_clear_mc()) for performance
3180 * improvement. But we cannot postpone mem_cgroup_get(to)
3181 * because if the process that has been moved to @to does
3182 * swap-in, the refcount of @to might be decreased to 0.
3186 if (!mem_cgroup_is_root(from))
3187 res_counter_uncharge(&from->memsw, PAGE_SIZE);
3188 mem_cgroup_put(from);
3190 * we charged both to->res and to->memsw, so we should
3193 if (!mem_cgroup_is_root(to))
3194 res_counter_uncharge(&to->res, PAGE_SIZE);
3201 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3202 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3209 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
3212 int mem_cgroup_prepare_migration(struct page *page,
3213 struct page *newpage, struct mem_cgroup **memcgp, gfp_t gfp_mask)
3215 struct mem_cgroup *memcg = NULL;
3216 struct page_cgroup *pc;
3217 enum charge_type ctype;
3222 VM_BUG_ON(PageTransHuge(page));
3223 if (mem_cgroup_disabled())
3226 pc = lookup_page_cgroup(page);
3227 lock_page_cgroup(pc);
3228 if (PageCgroupUsed(pc)) {
3229 memcg = pc->mem_cgroup;
3230 css_get(&memcg->css);
3232 * At migrating an anonymous page, its mapcount goes down
3233 * to 0 and uncharge() will be called. But, even if it's fully
3234 * unmapped, migration may fail and this page has to be
3235 * charged again. We set MIGRATION flag here and delay uncharge
3236 * until end_migration() is called
3238 * Corner Case Thinking
3240 * When the old page was mapped as Anon and it's unmap-and-freed
3241 * while migration was ongoing.
3242 * If unmap finds the old page, uncharge() of it will be delayed
3243 * until end_migration(). If unmap finds a new page, it's
3244 * uncharged when it make mapcount to be 1->0. If unmap code
3245 * finds swap_migration_entry, the new page will not be mapped
3246 * and end_migration() will find it(mapcount==0).
3249 * When the old page was mapped but migraion fails, the kernel
3250 * remaps it. A charge for it is kept by MIGRATION flag even
3251 * if mapcount goes down to 0. We can do remap successfully
3252 * without charging it again.
3255 * The "old" page is under lock_page() until the end of
3256 * migration, so, the old page itself will not be swapped-out.
3257 * If the new page is swapped out before end_migraton, our
3258 * hook to usual swap-out path will catch the event.
3261 SetPageCgroupMigration(pc);
3263 unlock_page_cgroup(pc);
3265 * If the page is not charged at this point,
3272 ret = __mem_cgroup_try_charge(NULL, gfp_mask, 1, memcgp, false);
3273 css_put(&memcg->css);/* drop extra refcnt */
3275 if (PageAnon(page)) {
3276 lock_page_cgroup(pc);
3277 ClearPageCgroupMigration(pc);
3278 unlock_page_cgroup(pc);
3280 * The old page may be fully unmapped while we kept it.
3282 mem_cgroup_uncharge_page(page);
3284 /* we'll need to revisit this error code (we have -EINTR) */
3288 * We charge new page before it's used/mapped. So, even if unlock_page()
3289 * is called before end_migration, we can catch all events on this new
3290 * page. In the case new page is migrated but not remapped, new page's
3291 * mapcount will be finally 0 and we call uncharge in end_migration().
3293 pc = lookup_page_cgroup(newpage);
3295 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
3296 else if (page_is_file_cache(page))
3297 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
3299 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3300 __mem_cgroup_commit_charge(memcg, newpage, 1, pc, ctype, false);
3304 /* remove redundant charge if migration failed*/
3305 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
3306 struct page *oldpage, struct page *newpage, bool migration_ok)
3308 struct page *used, *unused;
3309 struct page_cgroup *pc;
3314 /* blocks rmdir() */
3315 cgroup_exclude_rmdir(&memcg->css);
3316 if (!migration_ok) {
3324 * We disallowed uncharge of pages under migration because mapcount
3325 * of the page goes down to zero, temporarly.
3326 * Clear the flag and check the page should be charged.
3328 pc = lookup_page_cgroup(oldpage);
3329 lock_page_cgroup(pc);
3330 ClearPageCgroupMigration(pc);
3331 unlock_page_cgroup(pc);
3332 anon = PageAnon(used);
3333 __mem_cgroup_uncharge_common(unused,
3334 anon ? MEM_CGROUP_CHARGE_TYPE_MAPPED
3335 : MEM_CGROUP_CHARGE_TYPE_CACHE);
3338 * If a page is a file cache, radix-tree replacement is very atomic
3339 * and we can skip this check. When it was an Anon page, its mapcount
3340 * goes down to 0. But because we added MIGRATION flage, it's not
3341 * uncharged yet. There are several case but page->mapcount check
3342 * and USED bit check in mem_cgroup_uncharge_page() will do enough
3343 * check. (see prepare_charge() also)
3346 mem_cgroup_uncharge_page(used);
3348 * At migration, we may charge account against cgroup which has no
3350 * So, rmdir()->pre_destroy() can be called while we do this charge.
3351 * In that case, we need to call pre_destroy() again. check it here.
3353 cgroup_release_and_wakeup_rmdir(&memcg->css);
3357 * At replace page cache, newpage is not under any memcg but it's on
3358 * LRU. So, this function doesn't touch res_counter but handles LRU
3359 * in correct way. Both pages are locked so we cannot race with uncharge.
3361 void mem_cgroup_replace_page_cache(struct page *oldpage,
3362 struct page *newpage)
3364 struct mem_cgroup *memcg;
3365 struct page_cgroup *pc;
3366 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
3368 if (mem_cgroup_disabled())
3371 pc = lookup_page_cgroup(oldpage);
3372 /* fix accounting on old pages */
3373 lock_page_cgroup(pc);
3374 memcg = pc->mem_cgroup;
3375 mem_cgroup_charge_statistics(memcg, false, -1);
3376 ClearPageCgroupUsed(pc);
3377 unlock_page_cgroup(pc);
3379 if (PageSwapBacked(oldpage))
3380 type = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3383 * Even if newpage->mapping was NULL before starting replacement,
3384 * the newpage may be on LRU(or pagevec for LRU) already. We lock
3385 * LRU while we overwrite pc->mem_cgroup.
3387 __mem_cgroup_commit_charge(memcg, newpage, 1, pc, type, true);
3390 #ifdef CONFIG_DEBUG_VM
3391 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3393 struct page_cgroup *pc;
3395 pc = lookup_page_cgroup(page);
3397 * Can be NULL while feeding pages into the page allocator for
3398 * the first time, i.e. during boot or memory hotplug;
3399 * or when mem_cgroup_disabled().
3401 if (likely(pc) && PageCgroupUsed(pc))
3406 bool mem_cgroup_bad_page_check(struct page *page)
3408 if (mem_cgroup_disabled())
3411 return lookup_page_cgroup_used(page) != NULL;
3414 void mem_cgroup_print_bad_page(struct page *page)
3416 struct page_cgroup *pc;
3418 pc = lookup_page_cgroup_used(page);
3420 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
3421 pc, pc->flags, pc->mem_cgroup);
3426 static DEFINE_MUTEX(set_limit_mutex);
3428 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3429 unsigned long long val)
3432 u64 memswlimit, memlimit;
3434 int children = mem_cgroup_count_children(memcg);
3435 u64 curusage, oldusage;
3439 * For keeping hierarchical_reclaim simple, how long we should retry
3440 * is depends on callers. We set our retry-count to be function
3441 * of # of children which we should visit in this loop.
3443 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3445 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3448 while (retry_count) {
3449 if (signal_pending(current)) {
3454 * Rather than hide all in some function, I do this in
3455 * open coded manner. You see what this really does.
3456 * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3458 mutex_lock(&set_limit_mutex);
3459 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3460 if (memswlimit < val) {
3462 mutex_unlock(&set_limit_mutex);
3466 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3470 ret = res_counter_set_limit(&memcg->res, val);
3472 if (memswlimit == val)
3473 memcg->memsw_is_minimum = true;
3475 memcg->memsw_is_minimum = false;
3477 mutex_unlock(&set_limit_mutex);
3482 mem_cgroup_reclaim(memcg, GFP_KERNEL,
3483 MEM_CGROUP_RECLAIM_SHRINK);
3484 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3485 /* Usage is reduced ? */
3486 if (curusage >= oldusage)
3489 oldusage = curusage;
3491 if (!ret && enlarge)
3492 memcg_oom_recover(memcg);
3497 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3498 unsigned long long val)
3501 u64 memlimit, memswlimit, oldusage, curusage;
3502 int children = mem_cgroup_count_children(memcg);
3506 /* see mem_cgroup_resize_res_limit */
3507 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3508 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3509 while (retry_count) {
3510 if (signal_pending(current)) {
3515 * Rather than hide all in some function, I do this in
3516 * open coded manner. You see what this really does.
3517 * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3519 mutex_lock(&set_limit_mutex);
3520 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3521 if (memlimit > val) {
3523 mutex_unlock(&set_limit_mutex);
3526 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3527 if (memswlimit < val)
3529 ret = res_counter_set_limit(&memcg->memsw, val);
3531 if (memlimit == val)
3532 memcg->memsw_is_minimum = true;
3534 memcg->memsw_is_minimum = false;
3536 mutex_unlock(&set_limit_mutex);
3541 mem_cgroup_reclaim(memcg, GFP_KERNEL,
3542 MEM_CGROUP_RECLAIM_NOSWAP |
3543 MEM_CGROUP_RECLAIM_SHRINK);
3544 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3545 /* Usage is reduced ? */
3546 if (curusage >= oldusage)
3549 oldusage = curusage;
3551 if (!ret && enlarge)
3552 memcg_oom_recover(memcg);
3556 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3558 unsigned long *total_scanned)
3560 unsigned long nr_reclaimed = 0;
3561 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3562 unsigned long reclaimed;
3564 struct mem_cgroup_tree_per_zone *mctz;
3565 unsigned long long excess;
3566 unsigned long nr_scanned;
3571 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3573 * This loop can run a while, specially if mem_cgroup's continuously
3574 * keep exceeding their soft limit and putting the system under
3581 mz = mem_cgroup_largest_soft_limit_node(mctz);
3586 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
3587 gfp_mask, &nr_scanned);
3588 nr_reclaimed += reclaimed;
3589 *total_scanned += nr_scanned;
3590 spin_lock(&mctz->lock);
3593 * If we failed to reclaim anything from this memory cgroup
3594 * it is time to move on to the next cgroup
3600 * Loop until we find yet another one.
3602 * By the time we get the soft_limit lock
3603 * again, someone might have aded the
3604 * group back on the RB tree. Iterate to
3605 * make sure we get a different mem.
3606 * mem_cgroup_largest_soft_limit_node returns
3607 * NULL if no other cgroup is present on
3611 __mem_cgroup_largest_soft_limit_node(mctz);
3613 css_put(&next_mz->memcg->css);
3614 else /* next_mz == NULL or other memcg */
3618 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
3619 excess = res_counter_soft_limit_excess(&mz->memcg->res);
3621 * One school of thought says that we should not add
3622 * back the node to the tree if reclaim returns 0.
3623 * But our reclaim could return 0, simply because due
3624 * to priority we are exposing a smaller subset of
3625 * memory to reclaim from. Consider this as a longer
3628 /* If excess == 0, no tree ops */
3629 __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess);
3630 spin_unlock(&mctz->lock);
3631 css_put(&mz->memcg->css);
3634 * Could not reclaim anything and there are no more
3635 * mem cgroups to try or we seem to be looping without
3636 * reclaiming anything.
3638 if (!nr_reclaimed &&
3640 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3642 } while (!nr_reclaimed);
3644 css_put(&next_mz->memcg->css);
3645 return nr_reclaimed;
3649 * This routine traverse page_cgroup in given list and drop them all.
3650 * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3652 static int mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
3653 int node, int zid, enum lru_list lru)
3655 struct mem_cgroup_per_zone *mz;
3656 unsigned long flags, loop;
3657 struct list_head *list;
3662 zone = &NODE_DATA(node)->node_zones[zid];
3663 mz = mem_cgroup_zoneinfo(memcg, node, zid);
3664 list = &mz->lruvec.lists[lru];
3666 loop = mz->lru_size[lru];
3667 /* give some margin against EBUSY etc...*/
3671 struct page_cgroup *pc;
3675 spin_lock_irqsave(&zone->lru_lock, flags);
3676 if (list_empty(list)) {
3677 spin_unlock_irqrestore(&zone->lru_lock, flags);
3680 page = list_entry(list->prev, struct page, lru);
3682 list_move(&page->lru, list);
3684 spin_unlock_irqrestore(&zone->lru_lock, flags);
3687 spin_unlock_irqrestore(&zone->lru_lock, flags);
3689 pc = lookup_page_cgroup(page);
3691 ret = mem_cgroup_move_parent(page, pc, memcg, GFP_KERNEL);
3692 if (ret == -ENOMEM || ret == -EINTR)
3695 if (ret == -EBUSY || ret == -EINVAL) {
3696 /* found lock contention or "pc" is obsolete. */
3703 if (!ret && !list_empty(list))
3709 * make mem_cgroup's charge to be 0 if there is no task.
3710 * This enables deleting this mem_cgroup.
3712 static int mem_cgroup_force_empty(struct mem_cgroup *memcg, bool free_all)
3715 int node, zid, shrink;
3716 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3717 struct cgroup *cgrp = memcg->css.cgroup;
3719 css_get(&memcg->css);
3722 /* should free all ? */
3728 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3731 if (signal_pending(current))
3733 /* This is for making all *used* pages to be on LRU. */
3734 lru_add_drain_all();
3735 drain_all_stock_sync(memcg);
3737 mem_cgroup_start_move(memcg);
3738 for_each_node_state(node, N_HIGH_MEMORY) {
3739 for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3742 ret = mem_cgroup_force_empty_list(memcg,
3751 mem_cgroup_end_move(memcg);
3752 memcg_oom_recover(memcg);
3753 /* it seems parent cgroup doesn't have enough mem */
3757 /* "ret" should also be checked to ensure all lists are empty. */
3758 } while (memcg->res.usage > 0 || ret);
3760 css_put(&memcg->css);
3764 /* returns EBUSY if there is a task or if we come here twice. */
3765 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3769 /* we call try-to-free pages for make this cgroup empty */
3770 lru_add_drain_all();
3771 /* try to free all pages in this cgroup */
3773 while (nr_retries && memcg->res.usage > 0) {
3776 if (signal_pending(current)) {
3780 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
3784 /* maybe some writeback is necessary */
3785 congestion_wait(BLK_RW_ASYNC, HZ/10);
3790 /* try move_account...there may be some *locked* pages. */
3794 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3796 return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3800 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3802 return mem_cgroup_from_cont(cont)->use_hierarchy;
3805 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3809 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3810 struct cgroup *parent = cont->parent;
3811 struct mem_cgroup *parent_memcg = NULL;
3814 parent_memcg = mem_cgroup_from_cont(parent);
3818 * If parent's use_hierarchy is set, we can't make any modifications
3819 * in the child subtrees. If it is unset, then the change can
3820 * occur, provided the current cgroup has no children.
3822 * For the root cgroup, parent_mem is NULL, we allow value to be
3823 * set if there are no children.
3825 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3826 (val == 1 || val == 0)) {
3827 if (list_empty(&cont->children))
3828 memcg->use_hierarchy = val;
3839 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
3840 enum mem_cgroup_stat_index idx)
3842 struct mem_cgroup *iter;
3845 /* Per-cpu values can be negative, use a signed accumulator */
3846 for_each_mem_cgroup_tree(iter, memcg)
3847 val += mem_cgroup_read_stat(iter, idx);
3849 if (val < 0) /* race ? */
3854 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3858 if (!mem_cgroup_is_root(memcg)) {
3860 return res_counter_read_u64(&memcg->res, RES_USAGE);
3862 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
3865 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
3866 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
3869 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
3871 return val << PAGE_SHIFT;
3874 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
3876 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3880 type = MEMFILE_TYPE(cft->private);
3881 name = MEMFILE_ATTR(cft->private);
3884 if (name == RES_USAGE)
3885 val = mem_cgroup_usage(memcg, false);
3887 val = res_counter_read_u64(&memcg->res, name);
3890 if (name == RES_USAGE)
3891 val = mem_cgroup_usage(memcg, true);
3893 val = res_counter_read_u64(&memcg->memsw, name);
3902 * The user of this function is...
3905 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
3908 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3910 unsigned long long val;
3913 type = MEMFILE_TYPE(cft->private);
3914 name = MEMFILE_ATTR(cft->private);
3917 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3921 /* This function does all necessary parse...reuse it */
3922 ret = res_counter_memparse_write_strategy(buffer, &val);
3926 ret = mem_cgroup_resize_limit(memcg, val);
3928 ret = mem_cgroup_resize_memsw_limit(memcg, val);
3930 case RES_SOFT_LIMIT:
3931 ret = res_counter_memparse_write_strategy(buffer, &val);
3935 * For memsw, soft limits are hard to implement in terms
3936 * of semantics, for now, we support soft limits for
3937 * control without swap
3940 ret = res_counter_set_soft_limit(&memcg->res, val);
3945 ret = -EINVAL; /* should be BUG() ? */
3951 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
3952 unsigned long long *mem_limit, unsigned long long *memsw_limit)
3954 struct cgroup *cgroup;
3955 unsigned long long min_limit, min_memsw_limit, tmp;
3957 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3958 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3959 cgroup = memcg->css.cgroup;
3960 if (!memcg->use_hierarchy)
3963 while (cgroup->parent) {
3964 cgroup = cgroup->parent;
3965 memcg = mem_cgroup_from_cont(cgroup);
3966 if (!memcg->use_hierarchy)
3968 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
3969 min_limit = min(min_limit, tmp);
3970 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3971 min_memsw_limit = min(min_memsw_limit, tmp);
3974 *mem_limit = min_limit;
3975 *memsw_limit = min_memsw_limit;
3978 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
3980 struct mem_cgroup *memcg;
3983 memcg = mem_cgroup_from_cont(cont);
3984 type = MEMFILE_TYPE(event);
3985 name = MEMFILE_ATTR(event);
3989 res_counter_reset_max(&memcg->res);
3991 res_counter_reset_max(&memcg->memsw);
3995 res_counter_reset_failcnt(&memcg->res);
3997 res_counter_reset_failcnt(&memcg->memsw);
4004 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
4007 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
4011 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4012 struct cftype *cft, u64 val)
4014 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4016 if (val >= (1 << NR_MOVE_TYPE))
4019 * We check this value several times in both in can_attach() and
4020 * attach(), so we need cgroup lock to prevent this value from being
4024 memcg->move_charge_at_immigrate = val;
4030 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4031 struct cftype *cft, u64 val)
4038 /* For read statistics */
4056 struct mcs_total_stat {
4057 s64 stat[NR_MCS_STAT];
4063 } memcg_stat_strings[NR_MCS_STAT] = {
4064 {"cache", "total_cache"},
4065 {"rss", "total_rss"},
4066 {"mapped_file", "total_mapped_file"},
4067 {"pgpgin", "total_pgpgin"},
4068 {"pgpgout", "total_pgpgout"},
4069 {"swap", "total_swap"},
4070 {"pgfault", "total_pgfault"},
4071 {"pgmajfault", "total_pgmajfault"},
4072 {"inactive_anon", "total_inactive_anon"},
4073 {"active_anon", "total_active_anon"},
4074 {"inactive_file", "total_inactive_file"},
4075 {"active_file", "total_active_file"},
4076 {"unevictable", "total_unevictable"}
4081 mem_cgroup_get_local_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4086 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_CACHE);
4087 s->stat[MCS_CACHE] += val * PAGE_SIZE;
4088 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_RSS);
4089 s->stat[MCS_RSS] += val * PAGE_SIZE;
4090 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_FILE_MAPPED);
4091 s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
4092 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGIN);
4093 s->stat[MCS_PGPGIN] += val;
4094 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGOUT);
4095 s->stat[MCS_PGPGOUT] += val;
4096 if (do_swap_account) {
4097 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
4098 s->stat[MCS_SWAP] += val * PAGE_SIZE;
4100 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGFAULT);
4101 s->stat[MCS_PGFAULT] += val;
4102 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGMAJFAULT);
4103 s->stat[MCS_PGMAJFAULT] += val;
4106 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_ANON));
4107 s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
4108 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_ANON));
4109 s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
4110 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_FILE));
4111 s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
4112 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_FILE));
4113 s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
4114 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
4115 s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
4119 mem_cgroup_get_total_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4121 struct mem_cgroup *iter;
4123 for_each_mem_cgroup_tree(iter, memcg)
4124 mem_cgroup_get_local_stat(iter, s);
4128 static int mem_control_numa_stat_show(struct seq_file *m, void *arg)
4131 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
4132 unsigned long node_nr;
4133 struct cgroup *cont = m->private;
4134 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4136 total_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL);
4137 seq_printf(m, "total=%lu", total_nr);
4138 for_each_node_state(nid, N_HIGH_MEMORY) {
4139 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL);
4140 seq_printf(m, " N%d=%lu", nid, node_nr);
4144 file_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_FILE);
4145 seq_printf(m, "file=%lu", file_nr);
4146 for_each_node_state(nid, N_HIGH_MEMORY) {
4147 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4149 seq_printf(m, " N%d=%lu", nid, node_nr);
4153 anon_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_ANON);
4154 seq_printf(m, "anon=%lu", anon_nr);
4155 for_each_node_state(nid, N_HIGH_MEMORY) {
4156 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4158 seq_printf(m, " N%d=%lu", nid, node_nr);
4162 unevictable_nr = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
4163 seq_printf(m, "unevictable=%lu", unevictable_nr);
4164 for_each_node_state(nid, N_HIGH_MEMORY) {
4165 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4166 BIT(LRU_UNEVICTABLE));
4167 seq_printf(m, " N%d=%lu", nid, node_nr);
4172 #endif /* CONFIG_NUMA */
4174 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
4175 struct cgroup_map_cb *cb)
4177 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4178 struct mcs_total_stat mystat;
4181 memset(&mystat, 0, sizeof(mystat));
4182 mem_cgroup_get_local_stat(memcg, &mystat);
4185 for (i = 0; i < NR_MCS_STAT; i++) {
4186 if (i == MCS_SWAP && !do_swap_account)
4188 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
4191 /* Hierarchical information */
4193 unsigned long long limit, memsw_limit;
4194 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
4195 cb->fill(cb, "hierarchical_memory_limit", limit);
4196 if (do_swap_account)
4197 cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
4200 memset(&mystat, 0, sizeof(mystat));
4201 mem_cgroup_get_total_stat(memcg, &mystat);
4202 for (i = 0; i < NR_MCS_STAT; i++) {
4203 if (i == MCS_SWAP && !do_swap_account)
4205 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
4208 #ifdef CONFIG_DEBUG_VM
4211 struct mem_cgroup_per_zone *mz;
4212 unsigned long recent_rotated[2] = {0, 0};
4213 unsigned long recent_scanned[2] = {0, 0};
4215 for_each_online_node(nid)
4216 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4217 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
4219 recent_rotated[0] +=
4220 mz->reclaim_stat.recent_rotated[0];
4221 recent_rotated[1] +=
4222 mz->reclaim_stat.recent_rotated[1];
4223 recent_scanned[0] +=
4224 mz->reclaim_stat.recent_scanned[0];
4225 recent_scanned[1] +=
4226 mz->reclaim_stat.recent_scanned[1];
4228 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
4229 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
4230 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
4231 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
4238 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
4240 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4242 return mem_cgroup_swappiness(memcg);
4245 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
4248 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4249 struct mem_cgroup *parent;
4254 if (cgrp->parent == NULL)
4257 parent = mem_cgroup_from_cont(cgrp->parent);
4261 /* If under hierarchy, only empty-root can set this value */
4262 if ((parent->use_hierarchy) ||
4263 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4268 memcg->swappiness = val;
4275 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4277 struct mem_cgroup_threshold_ary *t;
4283 t = rcu_dereference(memcg->thresholds.primary);
4285 t = rcu_dereference(memcg->memsw_thresholds.primary);
4290 usage = mem_cgroup_usage(memcg, swap);
4293 * current_threshold points to threshold just below usage.
4294 * If it's not true, a threshold was crossed after last
4295 * call of __mem_cgroup_threshold().
4297 i = t->current_threshold;
4300 * Iterate backward over array of thresholds starting from
4301 * current_threshold and check if a threshold is crossed.
4302 * If none of thresholds below usage is crossed, we read
4303 * only one element of the array here.
4305 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4306 eventfd_signal(t->entries[i].eventfd, 1);
4308 /* i = current_threshold + 1 */
4312 * Iterate forward over array of thresholds starting from
4313 * current_threshold+1 and check if a threshold is crossed.
4314 * If none of thresholds above usage is crossed, we read
4315 * only one element of the array here.
4317 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4318 eventfd_signal(t->entries[i].eventfd, 1);
4320 /* Update current_threshold */
4321 t->current_threshold = i - 1;
4326 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4329 __mem_cgroup_threshold(memcg, false);
4330 if (do_swap_account)
4331 __mem_cgroup_threshold(memcg, true);
4333 memcg = parent_mem_cgroup(memcg);
4337 static int compare_thresholds(const void *a, const void *b)
4339 const struct mem_cgroup_threshold *_a = a;
4340 const struct mem_cgroup_threshold *_b = b;
4342 return _a->threshold - _b->threshold;
4345 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4347 struct mem_cgroup_eventfd_list *ev;
4349 list_for_each_entry(ev, &memcg->oom_notify, list)
4350 eventfd_signal(ev->eventfd, 1);
4354 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4356 struct mem_cgroup *iter;
4358 for_each_mem_cgroup_tree(iter, memcg)
4359 mem_cgroup_oom_notify_cb(iter);
4362 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
4363 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4365 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4366 struct mem_cgroup_thresholds *thresholds;
4367 struct mem_cgroup_threshold_ary *new;
4368 int type = MEMFILE_TYPE(cft->private);
4369 u64 threshold, usage;
4372 ret = res_counter_memparse_write_strategy(args, &threshold);
4376 mutex_lock(&memcg->thresholds_lock);
4379 thresholds = &memcg->thresholds;
4380 else if (type == _MEMSWAP)
4381 thresholds = &memcg->memsw_thresholds;
4385 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4387 /* Check if a threshold crossed before adding a new one */
4388 if (thresholds->primary)
4389 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4391 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4393 /* Allocate memory for new array of thresholds */
4394 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4402 /* Copy thresholds (if any) to new array */
4403 if (thresholds->primary) {
4404 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4405 sizeof(struct mem_cgroup_threshold));
4408 /* Add new threshold */
4409 new->entries[size - 1].eventfd = eventfd;
4410 new->entries[size - 1].threshold = threshold;
4412 /* Sort thresholds. Registering of new threshold isn't time-critical */
4413 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4414 compare_thresholds, NULL);
4416 /* Find current threshold */
4417 new->current_threshold = -1;
4418 for (i = 0; i < size; i++) {
4419 if (new->entries[i].threshold < usage) {
4421 * new->current_threshold will not be used until
4422 * rcu_assign_pointer(), so it's safe to increment
4425 ++new->current_threshold;
4429 /* Free old spare buffer and save old primary buffer as spare */
4430 kfree(thresholds->spare);
4431 thresholds->spare = thresholds->primary;
4433 rcu_assign_pointer(thresholds->primary, new);
4435 /* To be sure that nobody uses thresholds */
4439 mutex_unlock(&memcg->thresholds_lock);
4444 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
4445 struct cftype *cft, struct eventfd_ctx *eventfd)
4447 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4448 struct mem_cgroup_thresholds *thresholds;
4449 struct mem_cgroup_threshold_ary *new;
4450 int type = MEMFILE_TYPE(cft->private);
4454 mutex_lock(&memcg->thresholds_lock);
4456 thresholds = &memcg->thresholds;
4457 else if (type == _MEMSWAP)
4458 thresholds = &memcg->memsw_thresholds;
4463 * Something went wrong if we trying to unregister a threshold
4464 * if we don't have thresholds
4466 BUG_ON(!thresholds);
4468 if (!thresholds->primary)
4471 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4473 /* Check if a threshold crossed before removing */
4474 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4476 /* Calculate new number of threshold */
4478 for (i = 0; i < thresholds->primary->size; i++) {
4479 if (thresholds->primary->entries[i].eventfd != eventfd)
4483 new = thresholds->spare;
4485 /* Set thresholds array to NULL if we don't have thresholds */
4494 /* Copy thresholds and find current threshold */
4495 new->current_threshold = -1;
4496 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4497 if (thresholds->primary->entries[i].eventfd == eventfd)
4500 new->entries[j] = thresholds->primary->entries[i];
4501 if (new->entries[j].threshold < usage) {
4503 * new->current_threshold will not be used
4504 * until rcu_assign_pointer(), so it's safe to increment
4507 ++new->current_threshold;
4513 /* Swap primary and spare array */
4514 thresholds->spare = thresholds->primary;
4515 rcu_assign_pointer(thresholds->primary, new);
4517 /* To be sure that nobody uses thresholds */
4520 mutex_unlock(&memcg->thresholds_lock);
4523 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4524 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4526 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4527 struct mem_cgroup_eventfd_list *event;
4528 int type = MEMFILE_TYPE(cft->private);
4530 BUG_ON(type != _OOM_TYPE);
4531 event = kmalloc(sizeof(*event), GFP_KERNEL);
4535 spin_lock(&memcg_oom_lock);
4537 event->eventfd = eventfd;
4538 list_add(&event->list, &memcg->oom_notify);
4540 /* already in OOM ? */
4541 if (atomic_read(&memcg->under_oom))
4542 eventfd_signal(eventfd, 1);
4543 spin_unlock(&memcg_oom_lock);
4548 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4549 struct cftype *cft, struct eventfd_ctx *eventfd)
4551 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4552 struct mem_cgroup_eventfd_list *ev, *tmp;
4553 int type = MEMFILE_TYPE(cft->private);
4555 BUG_ON(type != _OOM_TYPE);
4557 spin_lock(&memcg_oom_lock);
4559 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4560 if (ev->eventfd == eventfd) {
4561 list_del(&ev->list);
4566 spin_unlock(&memcg_oom_lock);
4569 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4570 struct cftype *cft, struct cgroup_map_cb *cb)
4572 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4574 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
4576 if (atomic_read(&memcg->under_oom))
4577 cb->fill(cb, "under_oom", 1);
4579 cb->fill(cb, "under_oom", 0);
4583 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4584 struct cftype *cft, u64 val)
4586 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4587 struct mem_cgroup *parent;
4589 /* cannot set to root cgroup and only 0 and 1 are allowed */
4590 if (!cgrp->parent || !((val == 0) || (val == 1)))
4593 parent = mem_cgroup_from_cont(cgrp->parent);
4596 /* oom-kill-disable is a flag for subhierarchy. */
4597 if ((parent->use_hierarchy) ||
4598 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4602 memcg->oom_kill_disable = val;
4604 memcg_oom_recover(memcg);
4610 static const struct file_operations mem_control_numa_stat_file_operations = {
4612 .llseek = seq_lseek,
4613 .release = single_release,
4616 static int mem_control_numa_stat_open(struct inode *unused, struct file *file)
4618 struct cgroup *cont = file->f_dentry->d_parent->d_fsdata;
4620 file->f_op = &mem_control_numa_stat_file_operations;
4621 return single_open(file, mem_control_numa_stat_show, cont);
4623 #endif /* CONFIG_NUMA */
4625 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_KMEM
4626 static int register_kmem_files(struct cgroup *cont, struct cgroup_subsys *ss)
4629 * Part of this would be better living in a separate allocation
4630 * function, leaving us with just the cgroup tree population work.
4631 * We, however, depend on state such as network's proto_list that
4632 * is only initialized after cgroup creation. I found the less
4633 * cumbersome way to deal with it to defer it all to populate time
4635 return mem_cgroup_sockets_init(cont, ss);
4638 static void kmem_cgroup_destroy(struct cgroup *cont)
4640 mem_cgroup_sockets_destroy(cont);
4643 static int register_kmem_files(struct cgroup *cont, struct cgroup_subsys *ss)
4648 static void kmem_cgroup_destroy(struct cgroup *cont)
4653 static struct cftype mem_cgroup_files[] = {
4655 .name = "usage_in_bytes",
4656 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4657 .read_u64 = mem_cgroup_read,
4658 .register_event = mem_cgroup_usage_register_event,
4659 .unregister_event = mem_cgroup_usage_unregister_event,
4662 .name = "max_usage_in_bytes",
4663 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4664 .trigger = mem_cgroup_reset,
4665 .read_u64 = mem_cgroup_read,
4668 .name = "limit_in_bytes",
4669 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4670 .write_string = mem_cgroup_write,
4671 .read_u64 = mem_cgroup_read,
4674 .name = "soft_limit_in_bytes",
4675 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4676 .write_string = mem_cgroup_write,
4677 .read_u64 = mem_cgroup_read,
4681 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4682 .trigger = mem_cgroup_reset,
4683 .read_u64 = mem_cgroup_read,
4687 .read_map = mem_control_stat_show,
4690 .name = "force_empty",
4691 .trigger = mem_cgroup_force_empty_write,
4694 .name = "use_hierarchy",
4695 .write_u64 = mem_cgroup_hierarchy_write,
4696 .read_u64 = mem_cgroup_hierarchy_read,
4699 .name = "swappiness",
4700 .read_u64 = mem_cgroup_swappiness_read,
4701 .write_u64 = mem_cgroup_swappiness_write,
4704 .name = "move_charge_at_immigrate",
4705 .read_u64 = mem_cgroup_move_charge_read,
4706 .write_u64 = mem_cgroup_move_charge_write,
4709 .name = "oom_control",
4710 .read_map = mem_cgroup_oom_control_read,
4711 .write_u64 = mem_cgroup_oom_control_write,
4712 .register_event = mem_cgroup_oom_register_event,
4713 .unregister_event = mem_cgroup_oom_unregister_event,
4714 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4718 .name = "numa_stat",
4719 .open = mem_control_numa_stat_open,
4725 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4726 static struct cftype memsw_cgroup_files[] = {
4728 .name = "memsw.usage_in_bytes",
4729 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4730 .read_u64 = mem_cgroup_read,
4731 .register_event = mem_cgroup_usage_register_event,
4732 .unregister_event = mem_cgroup_usage_unregister_event,
4735 .name = "memsw.max_usage_in_bytes",
4736 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4737 .trigger = mem_cgroup_reset,
4738 .read_u64 = mem_cgroup_read,
4741 .name = "memsw.limit_in_bytes",
4742 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4743 .write_string = mem_cgroup_write,
4744 .read_u64 = mem_cgroup_read,
4747 .name = "memsw.failcnt",
4748 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4749 .trigger = mem_cgroup_reset,
4750 .read_u64 = mem_cgroup_read,
4754 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4756 if (!do_swap_account)
4758 return cgroup_add_files(cont, ss, memsw_cgroup_files,
4759 ARRAY_SIZE(memsw_cgroup_files));
4762 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4768 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4770 struct mem_cgroup_per_node *pn;
4771 struct mem_cgroup_per_zone *mz;
4773 int zone, tmp = node;
4775 * This routine is called against possible nodes.
4776 * But it's BUG to call kmalloc() against offline node.
4778 * TODO: this routine can waste much memory for nodes which will
4779 * never be onlined. It's better to use memory hotplug callback
4782 if (!node_state(node, N_NORMAL_MEMORY))
4784 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4788 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4789 mz = &pn->zoneinfo[zone];
4791 INIT_LIST_HEAD(&mz->lruvec.lists[lru]);
4792 mz->usage_in_excess = 0;
4793 mz->on_tree = false;
4796 memcg->info.nodeinfo[node] = pn;
4800 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4802 kfree(memcg->info.nodeinfo[node]);
4805 static struct mem_cgroup *mem_cgroup_alloc(void)
4807 struct mem_cgroup *memcg;
4808 int size = sizeof(struct mem_cgroup);
4810 /* Can be very big if MAX_NUMNODES is very big */
4811 if (size < PAGE_SIZE)
4812 memcg = kzalloc(size, GFP_KERNEL);
4814 memcg = vzalloc(size);
4819 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4822 spin_lock_init(&memcg->pcp_counter_lock);
4826 if (size < PAGE_SIZE)
4834 * Helpers for freeing a vzalloc()ed mem_cgroup by RCU,
4835 * but in process context. The work_freeing structure is overlaid
4836 * on the rcu_freeing structure, which itself is overlaid on memsw.
4838 static void vfree_work(struct work_struct *work)
4840 struct mem_cgroup *memcg;
4842 memcg = container_of(work, struct mem_cgroup, work_freeing);
4845 static void vfree_rcu(struct rcu_head *rcu_head)
4847 struct mem_cgroup *memcg;
4849 memcg = container_of(rcu_head, struct mem_cgroup, rcu_freeing);
4850 INIT_WORK(&memcg->work_freeing, vfree_work);
4851 schedule_work(&memcg->work_freeing);
4855 * At destroying mem_cgroup, references from swap_cgroup can remain.
4856 * (scanning all at force_empty is too costly...)
4858 * Instead of clearing all references at force_empty, we remember
4859 * the number of reference from swap_cgroup and free mem_cgroup when
4860 * it goes down to 0.
4862 * Removal of cgroup itself succeeds regardless of refs from swap.
4865 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4869 mem_cgroup_remove_from_trees(memcg);
4870 free_css_id(&mem_cgroup_subsys, &memcg->css);
4873 free_mem_cgroup_per_zone_info(memcg, node);
4875 free_percpu(memcg->stat);
4876 if (sizeof(struct mem_cgroup) < PAGE_SIZE)
4877 kfree_rcu(memcg, rcu_freeing);
4879 call_rcu(&memcg->rcu_freeing, vfree_rcu);
4882 static void mem_cgroup_get(struct mem_cgroup *memcg)
4884 atomic_inc(&memcg->refcnt);
4887 static void __mem_cgroup_put(struct mem_cgroup *memcg, int count)
4889 if (atomic_sub_and_test(count, &memcg->refcnt)) {
4890 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4891 __mem_cgroup_free(memcg);
4893 mem_cgroup_put(parent);
4897 static void mem_cgroup_put(struct mem_cgroup *memcg)
4899 __mem_cgroup_put(memcg, 1);
4903 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4905 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4907 if (!memcg->res.parent)
4909 return mem_cgroup_from_res_counter(memcg->res.parent, res);
4911 EXPORT_SYMBOL(parent_mem_cgroup);
4913 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4914 static void __init enable_swap_cgroup(void)
4916 if (!mem_cgroup_disabled() && really_do_swap_account)
4917 do_swap_account = 1;
4920 static void __init enable_swap_cgroup(void)
4925 static int mem_cgroup_soft_limit_tree_init(void)
4927 struct mem_cgroup_tree_per_node *rtpn;
4928 struct mem_cgroup_tree_per_zone *rtpz;
4929 int tmp, node, zone;
4931 for_each_node(node) {
4933 if (!node_state(node, N_NORMAL_MEMORY))
4935 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
4939 soft_limit_tree.rb_tree_per_node[node] = rtpn;
4941 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4942 rtpz = &rtpn->rb_tree_per_zone[zone];
4943 rtpz->rb_root = RB_ROOT;
4944 spin_lock_init(&rtpz->lock);
4950 for_each_node(node) {
4951 if (!soft_limit_tree.rb_tree_per_node[node])
4953 kfree(soft_limit_tree.rb_tree_per_node[node]);
4954 soft_limit_tree.rb_tree_per_node[node] = NULL;
4960 static struct cgroup_subsys_state * __ref
4961 mem_cgroup_create(struct cgroup *cont)
4963 struct mem_cgroup *memcg, *parent;
4964 long error = -ENOMEM;
4967 memcg = mem_cgroup_alloc();
4969 return ERR_PTR(error);
4972 if (alloc_mem_cgroup_per_zone_info(memcg, node))
4976 if (cont->parent == NULL) {
4978 enable_swap_cgroup();
4980 if (mem_cgroup_soft_limit_tree_init())
4982 root_mem_cgroup = memcg;
4983 for_each_possible_cpu(cpu) {
4984 struct memcg_stock_pcp *stock =
4985 &per_cpu(memcg_stock, cpu);
4986 INIT_WORK(&stock->work, drain_local_stock);
4988 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
4990 parent = mem_cgroup_from_cont(cont->parent);
4991 memcg->use_hierarchy = parent->use_hierarchy;
4992 memcg->oom_kill_disable = parent->oom_kill_disable;
4995 if (parent && parent->use_hierarchy) {
4996 res_counter_init(&memcg->res, &parent->res);
4997 res_counter_init(&memcg->memsw, &parent->memsw);
4999 * We increment refcnt of the parent to ensure that we can
5000 * safely access it on res_counter_charge/uncharge.
5001 * This refcnt will be decremented when freeing this
5002 * mem_cgroup(see mem_cgroup_put).
5004 mem_cgroup_get(parent);
5006 res_counter_init(&memcg->res, NULL);
5007 res_counter_init(&memcg->memsw, NULL);
5009 memcg->last_scanned_node = MAX_NUMNODES;
5010 INIT_LIST_HEAD(&memcg->oom_notify);
5013 memcg->swappiness = mem_cgroup_swappiness(parent);
5014 atomic_set(&memcg->refcnt, 1);
5015 memcg->move_charge_at_immigrate = 0;
5016 mutex_init(&memcg->thresholds_lock);
5017 spin_lock_init(&memcg->move_lock);
5020 __mem_cgroup_free(memcg);
5021 return ERR_PTR(error);
5024 static int mem_cgroup_pre_destroy(struct cgroup *cont)
5026 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5028 return mem_cgroup_force_empty(memcg, false);
5031 static void mem_cgroup_destroy(struct cgroup *cont)
5033 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5035 kmem_cgroup_destroy(cont);
5037 mem_cgroup_put(memcg);
5040 static int mem_cgroup_populate(struct cgroup_subsys *ss,
5041 struct cgroup *cont)
5045 ret = cgroup_add_files(cont, ss, mem_cgroup_files,
5046 ARRAY_SIZE(mem_cgroup_files));
5049 ret = register_memsw_files(cont, ss);
5052 ret = register_kmem_files(cont, ss);
5058 /* Handlers for move charge at task migration. */
5059 #define PRECHARGE_COUNT_AT_ONCE 256
5060 static int mem_cgroup_do_precharge(unsigned long count)
5063 int batch_count = PRECHARGE_COUNT_AT_ONCE;
5064 struct mem_cgroup *memcg = mc.to;
5066 if (mem_cgroup_is_root(memcg)) {
5067 mc.precharge += count;
5068 /* we don't need css_get for root */
5071 /* try to charge at once */
5073 struct res_counter *dummy;
5075 * "memcg" cannot be under rmdir() because we've already checked
5076 * by cgroup_lock_live_cgroup() that it is not removed and we
5077 * are still under the same cgroup_mutex. So we can postpone
5080 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
5082 if (do_swap_account && res_counter_charge(&memcg->memsw,
5083 PAGE_SIZE * count, &dummy)) {
5084 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
5087 mc.precharge += count;
5091 /* fall back to one by one charge */
5093 if (signal_pending(current)) {
5097 if (!batch_count--) {
5098 batch_count = PRECHARGE_COUNT_AT_ONCE;
5101 ret = __mem_cgroup_try_charge(NULL,
5102 GFP_KERNEL, 1, &memcg, false);
5104 /* mem_cgroup_clear_mc() will do uncharge later */
5112 * is_target_pte_for_mc - check a pte whether it is valid for move charge
5113 * @vma: the vma the pte to be checked belongs
5114 * @addr: the address corresponding to the pte to be checked
5115 * @ptent: the pte to be checked
5116 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5119 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5120 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5121 * move charge. if @target is not NULL, the page is stored in target->page
5122 * with extra refcnt got(Callers should handle it).
5123 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5124 * target for charge migration. if @target is not NULL, the entry is stored
5127 * Called with pte lock held.
5134 enum mc_target_type {
5135 MC_TARGET_NONE, /* not used */
5140 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5141 unsigned long addr, pte_t ptent)
5143 struct page *page = vm_normal_page(vma, addr, ptent);
5145 if (!page || !page_mapped(page))
5147 if (PageAnon(page)) {
5148 /* we don't move shared anon */
5149 if (!move_anon() || page_mapcount(page) > 2)
5151 } else if (!move_file())
5152 /* we ignore mapcount for file pages */
5154 if (!get_page_unless_zero(page))
5160 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5161 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5164 struct page *page = NULL;
5165 swp_entry_t ent = pte_to_swp_entry(ptent);
5167 if (!move_anon() || non_swap_entry(ent))
5169 usage_count = mem_cgroup_count_swap_user(ent, &page);
5170 if (usage_count > 1) { /* we don't move shared anon */
5175 if (do_swap_account)
5176 entry->val = ent.val;
5181 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5182 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5184 struct page *page = NULL;
5185 struct inode *inode;
5186 struct address_space *mapping;
5189 if (!vma->vm_file) /* anonymous vma */
5194 inode = vma->vm_file->f_path.dentry->d_inode;
5195 mapping = vma->vm_file->f_mapping;
5196 if (pte_none(ptent))
5197 pgoff = linear_page_index(vma, addr);
5198 else /* pte_file(ptent) is true */
5199 pgoff = pte_to_pgoff(ptent);
5201 /* page is moved even if it's not RSS of this task(page-faulted). */
5202 page = find_get_page(mapping, pgoff);
5205 /* shmem/tmpfs may report page out on swap: account for that too. */
5206 if (radix_tree_exceptional_entry(page)) {
5207 swp_entry_t swap = radix_to_swp_entry(page);
5208 if (do_swap_account)
5210 page = find_get_page(&swapper_space, swap.val);
5216 static int is_target_pte_for_mc(struct vm_area_struct *vma,
5217 unsigned long addr, pte_t ptent, union mc_target *target)
5219 struct page *page = NULL;
5220 struct page_cgroup *pc;
5222 swp_entry_t ent = { .val = 0 };
5224 if (pte_present(ptent))
5225 page = mc_handle_present_pte(vma, addr, ptent);
5226 else if (is_swap_pte(ptent))
5227 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5228 else if (pte_none(ptent) || pte_file(ptent))
5229 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5231 if (!page && !ent.val)
5234 pc = lookup_page_cgroup(page);
5236 * Do only loose check w/o page_cgroup lock.
5237 * mem_cgroup_move_account() checks the pc is valid or not under
5240 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5241 ret = MC_TARGET_PAGE;
5243 target->page = page;
5245 if (!ret || !target)
5248 /* There is a swap entry and a page doesn't exist or isn't charged */
5249 if (ent.val && !ret &&
5250 css_id(&mc.from->css) == lookup_swap_cgroup_id(ent)) {
5251 ret = MC_TARGET_SWAP;
5258 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5259 unsigned long addr, unsigned long end,
5260 struct mm_walk *walk)
5262 struct vm_area_struct *vma = walk->private;
5266 split_huge_page_pmd(walk->mm, pmd);
5267 if (pmd_trans_unstable(pmd))
5270 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5271 for (; addr != end; pte++, addr += PAGE_SIZE)
5272 if (is_target_pte_for_mc(vma, addr, *pte, NULL))
5273 mc.precharge++; /* increment precharge temporarily */
5274 pte_unmap_unlock(pte - 1, ptl);
5280 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5282 unsigned long precharge;
5283 struct vm_area_struct *vma;
5285 down_read(&mm->mmap_sem);
5286 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5287 struct mm_walk mem_cgroup_count_precharge_walk = {
5288 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5292 if (is_vm_hugetlb_page(vma))
5294 walk_page_range(vma->vm_start, vma->vm_end,
5295 &mem_cgroup_count_precharge_walk);
5297 up_read(&mm->mmap_sem);
5299 precharge = mc.precharge;
5305 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5307 unsigned long precharge = mem_cgroup_count_precharge(mm);
5309 VM_BUG_ON(mc.moving_task);
5310 mc.moving_task = current;
5311 return mem_cgroup_do_precharge(precharge);
5314 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5315 static void __mem_cgroup_clear_mc(void)
5317 struct mem_cgroup *from = mc.from;
5318 struct mem_cgroup *to = mc.to;
5320 /* we must uncharge all the leftover precharges from mc.to */
5322 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
5326 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5327 * we must uncharge here.
5329 if (mc.moved_charge) {
5330 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
5331 mc.moved_charge = 0;
5333 /* we must fixup refcnts and charges */
5334 if (mc.moved_swap) {
5335 /* uncharge swap account from the old cgroup */
5336 if (!mem_cgroup_is_root(mc.from))
5337 res_counter_uncharge(&mc.from->memsw,
5338 PAGE_SIZE * mc.moved_swap);
5339 __mem_cgroup_put(mc.from, mc.moved_swap);
5341 if (!mem_cgroup_is_root(mc.to)) {
5343 * we charged both to->res and to->memsw, so we should
5346 res_counter_uncharge(&mc.to->res,
5347 PAGE_SIZE * mc.moved_swap);
5349 /* we've already done mem_cgroup_get(mc.to) */
5352 memcg_oom_recover(from);
5353 memcg_oom_recover(to);
5354 wake_up_all(&mc.waitq);
5357 static void mem_cgroup_clear_mc(void)
5359 struct mem_cgroup *from = mc.from;
5362 * we must clear moving_task before waking up waiters at the end of
5365 mc.moving_task = NULL;
5366 __mem_cgroup_clear_mc();
5367 spin_lock(&mc.lock);
5370 spin_unlock(&mc.lock);
5371 mem_cgroup_end_move(from);
5374 static int mem_cgroup_can_attach(struct cgroup *cgroup,
5375 struct cgroup_taskset *tset)
5377 struct task_struct *p = cgroup_taskset_first(tset);
5379 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup);
5381 if (memcg->move_charge_at_immigrate) {
5382 struct mm_struct *mm;
5383 struct mem_cgroup *from = mem_cgroup_from_task(p);
5385 VM_BUG_ON(from == memcg);
5387 mm = get_task_mm(p);
5390 /* We move charges only when we move a owner of the mm */
5391 if (mm->owner == p) {
5394 VM_BUG_ON(mc.precharge);
5395 VM_BUG_ON(mc.moved_charge);
5396 VM_BUG_ON(mc.moved_swap);
5397 mem_cgroup_start_move(from);
5398 spin_lock(&mc.lock);
5401 spin_unlock(&mc.lock);
5402 /* We set mc.moving_task later */
5404 ret = mem_cgroup_precharge_mc(mm);
5406 mem_cgroup_clear_mc();
5413 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
5414 struct cgroup_taskset *tset)
5416 mem_cgroup_clear_mc();
5419 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5420 unsigned long addr, unsigned long end,
5421 struct mm_walk *walk)
5424 struct vm_area_struct *vma = walk->private;
5428 split_huge_page_pmd(walk->mm, pmd);
5429 if (pmd_trans_unstable(pmd))
5432 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5433 for (; addr != end; addr += PAGE_SIZE) {
5434 pte_t ptent = *(pte++);
5435 union mc_target target;
5438 struct page_cgroup *pc;
5444 type = is_target_pte_for_mc(vma, addr, ptent, &target);
5446 case MC_TARGET_PAGE:
5448 if (isolate_lru_page(page))
5450 pc = lookup_page_cgroup(page);
5451 if (!mem_cgroup_move_account(page, 1, pc,
5452 mc.from, mc.to, false)) {
5454 /* we uncharge from mc.from later. */
5457 putback_lru_page(page);
5458 put: /* is_target_pte_for_mc() gets the page */
5461 case MC_TARGET_SWAP:
5463 if (!mem_cgroup_move_swap_account(ent,
5464 mc.from, mc.to, false)) {
5466 /* we fixup refcnts and charges later. */
5474 pte_unmap_unlock(pte - 1, ptl);
5479 * We have consumed all precharges we got in can_attach().
5480 * We try charge one by one, but don't do any additional
5481 * charges to mc.to if we have failed in charge once in attach()
5484 ret = mem_cgroup_do_precharge(1);
5492 static void mem_cgroup_move_charge(struct mm_struct *mm)
5494 struct vm_area_struct *vma;
5496 lru_add_drain_all();
5498 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5500 * Someone who are holding the mmap_sem might be waiting in
5501 * waitq. So we cancel all extra charges, wake up all waiters,
5502 * and retry. Because we cancel precharges, we might not be able
5503 * to move enough charges, but moving charge is a best-effort
5504 * feature anyway, so it wouldn't be a big problem.
5506 __mem_cgroup_clear_mc();
5510 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5512 struct mm_walk mem_cgroup_move_charge_walk = {
5513 .pmd_entry = mem_cgroup_move_charge_pte_range,
5517 if (is_vm_hugetlb_page(vma))
5519 ret = walk_page_range(vma->vm_start, vma->vm_end,
5520 &mem_cgroup_move_charge_walk);
5523 * means we have consumed all precharges and failed in
5524 * doing additional charge. Just abandon here.
5528 up_read(&mm->mmap_sem);
5531 static void mem_cgroup_move_task(struct cgroup *cont,
5532 struct cgroup_taskset *tset)
5534 struct task_struct *p = cgroup_taskset_first(tset);
5535 struct mm_struct *mm = get_task_mm(p);
5539 mem_cgroup_move_charge(mm);
5544 mem_cgroup_clear_mc();
5546 #else /* !CONFIG_MMU */
5547 static int mem_cgroup_can_attach(struct cgroup *cgroup,
5548 struct cgroup_taskset *tset)
5552 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
5553 struct cgroup_taskset *tset)
5556 static void mem_cgroup_move_task(struct cgroup *cont,
5557 struct cgroup_taskset *tset)
5562 struct cgroup_subsys mem_cgroup_subsys = {
5564 .subsys_id = mem_cgroup_subsys_id,
5565 .create = mem_cgroup_create,
5566 .pre_destroy = mem_cgroup_pre_destroy,
5567 .destroy = mem_cgroup_destroy,
5568 .populate = mem_cgroup_populate,
5569 .can_attach = mem_cgroup_can_attach,
5570 .cancel_attach = mem_cgroup_cancel_attach,
5571 .attach = mem_cgroup_move_task,
5576 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5577 static int __init enable_swap_account(char *s)
5579 /* consider enabled if no parameter or 1 is given */
5580 if (!strcmp(s, "1"))
5581 really_do_swap_account = 1;
5582 else if (!strcmp(s, "0"))
5583 really_do_swap_account = 0;
5586 __setup("swapaccount=", enable_swap_account);