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_ON_MOVE, /* someone is moving account between groups */
93 MEM_CGROUP_STAT_NSTATS,
96 enum mem_cgroup_events_index {
97 MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */
98 MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */
99 MEM_CGROUP_EVENTS_COUNT, /* # of pages paged in/out */
100 MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */
101 MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */
102 MEM_CGROUP_EVENTS_NSTATS,
105 * Per memcg event counter is incremented at every pagein/pageout. With THP,
106 * it will be incremated by the number of pages. This counter is used for
107 * for trigger some periodic events. This is straightforward and better
108 * than using jiffies etc. to handle periodic memcg event.
110 enum mem_cgroup_events_target {
111 MEM_CGROUP_TARGET_THRESH,
112 MEM_CGROUP_TARGET_SOFTLIMIT,
113 MEM_CGROUP_TARGET_NUMAINFO,
116 #define THRESHOLDS_EVENTS_TARGET (128)
117 #define SOFTLIMIT_EVENTS_TARGET (1024)
118 #define NUMAINFO_EVENTS_TARGET (1024)
120 struct mem_cgroup_stat_cpu {
121 long count[MEM_CGROUP_STAT_NSTATS];
122 unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
123 unsigned long targets[MEM_CGROUP_NTARGETS];
126 struct mem_cgroup_reclaim_iter {
127 /* css_id of the last scanned hierarchy member */
129 /* scan generation, increased every round-trip */
130 unsigned int generation;
134 * per-zone information in memory controller.
136 struct mem_cgroup_per_zone {
137 struct lruvec lruvec;
138 unsigned long count[NR_LRU_LISTS];
140 struct mem_cgroup_reclaim_iter reclaim_iter[DEF_PRIORITY + 1];
142 struct zone_reclaim_stat reclaim_stat;
143 struct rb_node tree_node; /* RB tree node */
144 unsigned long long usage_in_excess;/* Set to the value by which */
145 /* the soft limit is exceeded*/
147 struct mem_cgroup *mem; /* Back pointer, we cannot */
148 /* use container_of */
150 /* Macro for accessing counter */
151 #define MEM_CGROUP_ZSTAT(mz, idx) ((mz)->count[(idx)])
153 struct mem_cgroup_per_node {
154 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
157 struct mem_cgroup_lru_info {
158 struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
162 * Cgroups above their limits are maintained in a RB-Tree, independent of
163 * their hierarchy representation
166 struct mem_cgroup_tree_per_zone {
167 struct rb_root rb_root;
171 struct mem_cgroup_tree_per_node {
172 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
175 struct mem_cgroup_tree {
176 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
179 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
181 struct mem_cgroup_threshold {
182 struct eventfd_ctx *eventfd;
187 struct mem_cgroup_threshold_ary {
188 /* An array index points to threshold just below usage. */
189 int current_threshold;
190 /* Size of entries[] */
192 /* Array of thresholds */
193 struct mem_cgroup_threshold entries[0];
196 struct mem_cgroup_thresholds {
197 /* Primary thresholds array */
198 struct mem_cgroup_threshold_ary *primary;
200 * Spare threshold array.
201 * This is needed to make mem_cgroup_unregister_event() "never fail".
202 * It must be able to store at least primary->size - 1 entries.
204 struct mem_cgroup_threshold_ary *spare;
208 struct mem_cgroup_eventfd_list {
209 struct list_head list;
210 struct eventfd_ctx *eventfd;
213 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
214 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
217 * The memory controller data structure. The memory controller controls both
218 * page cache and RSS per cgroup. We would eventually like to provide
219 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
220 * to help the administrator determine what knobs to tune.
222 * TODO: Add a water mark for the memory controller. Reclaim will begin when
223 * we hit the water mark. May be even add a low water mark, such that
224 * no reclaim occurs from a cgroup at it's low water mark, this is
225 * a feature that will be implemented much later in the future.
228 struct cgroup_subsys_state css;
230 * the counter to account for memory usage
232 struct res_counter res;
234 * the counter to account for mem+swap usage.
236 struct res_counter memsw;
238 * the counter to account for kmem usage.
240 struct res_counter kmem;
242 * Per cgroup active and inactive list, similar to the
243 * per zone LRU lists.
245 struct mem_cgroup_lru_info info;
246 int last_scanned_node;
248 nodemask_t scan_nodes;
249 atomic_t numainfo_events;
250 atomic_t numainfo_updating;
253 * Should the accounting and control be hierarchical, per subtree?
263 /* OOM-Killer disable */
264 int oom_kill_disable;
266 /* set when res.limit == memsw.limit */
267 bool memsw_is_minimum;
269 /* protect arrays of thresholds */
270 struct mutex thresholds_lock;
272 /* thresholds for memory usage. RCU-protected */
273 struct mem_cgroup_thresholds thresholds;
275 /* thresholds for mem+swap usage. RCU-protected */
276 struct mem_cgroup_thresholds memsw_thresholds;
278 /* For oom notifier event fd */
279 struct list_head oom_notify;
282 * Should we move charges of a task when a task is moved into this
283 * mem_cgroup ? And what type of charges should we move ?
285 unsigned long move_charge_at_immigrate;
287 * Should kernel memory limits be stabilished independently
290 int kmem_independent_accounting;
294 struct mem_cgroup_stat_cpu *stat;
296 * used when a cpu is offlined or other synchronizations
297 * See mem_cgroup_read_stat().
299 struct mem_cgroup_stat_cpu nocpu_base;
300 spinlock_t pcp_counter_lock;
303 struct tcp_memcontrol tcp_mem;
307 /* Stuffs for move charges at task migration. */
309 * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
310 * left-shifted bitmap of these types.
313 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
314 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
318 /* "mc" and its members are protected by cgroup_mutex */
319 static struct move_charge_struct {
320 spinlock_t lock; /* for from, to */
321 struct mem_cgroup *from;
322 struct mem_cgroup *to;
323 unsigned long precharge;
324 unsigned long moved_charge;
325 unsigned long moved_swap;
326 struct task_struct *moving_task; /* a task moving charges */
327 wait_queue_head_t waitq; /* a waitq for other context */
329 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
330 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
333 static bool move_anon(void)
335 return test_bit(MOVE_CHARGE_TYPE_ANON,
336 &mc.to->move_charge_at_immigrate);
339 static bool move_file(void)
341 return test_bit(MOVE_CHARGE_TYPE_FILE,
342 &mc.to->move_charge_at_immigrate);
346 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
347 * limit reclaim to prevent infinite loops, if they ever occur.
349 #define MEM_CGROUP_MAX_RECLAIM_LOOPS (100)
350 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2)
353 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
354 MEM_CGROUP_CHARGE_TYPE_MAPPED,
355 MEM_CGROUP_CHARGE_TYPE_SHMEM, /* used by page migration of shmem */
356 MEM_CGROUP_CHARGE_TYPE_FORCE, /* used by force_empty */
357 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
358 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
362 /* for encoding cft->private value on file */
371 #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
372 #define MEMFILE_TYPE(val) (((val) >> 16) & 0xffff)
373 #define MEMFILE_ATTR(val) ((val) & 0xffff)
374 /* Used for OOM nofiier */
375 #define OOM_CONTROL (0)
378 * Reclaim flags for mem_cgroup_hierarchical_reclaim
380 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
381 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
382 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
383 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
385 static void mem_cgroup_get(struct mem_cgroup *memcg);
386 static void mem_cgroup_put(struct mem_cgroup *memcg);
388 /* Writing them here to avoid exposing memcg's inner layout */
389 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_KMEM
391 #include <net/sock.h>
394 static bool mem_cgroup_is_root(struct mem_cgroup *memcg);
395 void sock_update_memcg(struct sock *sk)
397 /* A socket spends its whole life in the same cgroup */
402 if (static_branch(&memcg_socket_limit_enabled)) {
403 struct mem_cgroup *memcg;
405 BUG_ON(!sk->sk_prot->proto_cgroup);
408 memcg = mem_cgroup_from_task(current);
409 if (!mem_cgroup_is_root(memcg)) {
410 mem_cgroup_get(memcg);
411 sk->sk_cgrp = sk->sk_prot->proto_cgroup(memcg);
416 EXPORT_SYMBOL(sock_update_memcg);
418 void sock_release_memcg(struct sock *sk)
420 if (static_branch(&memcg_socket_limit_enabled) && sk->sk_cgrp) {
421 struct mem_cgroup *memcg;
422 WARN_ON(!sk->sk_cgrp->memcg);
423 memcg = sk->sk_cgrp->memcg;
424 mem_cgroup_put(memcg);
428 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
430 if (!memcg || mem_cgroup_is_root(memcg))
433 return &memcg->tcp_mem.cg_proto;
435 EXPORT_SYMBOL(tcp_proto_cgroup);
436 #endif /* CONFIG_INET */
437 #endif /* CONFIG_CGROUP_MEM_RES_CTLR_KMEM */
439 static void drain_all_stock_async(struct mem_cgroup *memcg);
441 static struct mem_cgroup_per_zone *
442 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
444 return &memcg->info.nodeinfo[nid]->zoneinfo[zid];
447 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
452 static struct mem_cgroup_per_zone *
453 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
455 int nid = page_to_nid(page);
456 int zid = page_zonenum(page);
458 return mem_cgroup_zoneinfo(memcg, nid, zid);
461 static struct mem_cgroup_tree_per_zone *
462 soft_limit_tree_node_zone(int nid, int zid)
464 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
467 static struct mem_cgroup_tree_per_zone *
468 soft_limit_tree_from_page(struct page *page)
470 int nid = page_to_nid(page);
471 int zid = page_zonenum(page);
473 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
477 __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
478 struct mem_cgroup_per_zone *mz,
479 struct mem_cgroup_tree_per_zone *mctz,
480 unsigned long long new_usage_in_excess)
482 struct rb_node **p = &mctz->rb_root.rb_node;
483 struct rb_node *parent = NULL;
484 struct mem_cgroup_per_zone *mz_node;
489 mz->usage_in_excess = new_usage_in_excess;
490 if (!mz->usage_in_excess)
494 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
496 if (mz->usage_in_excess < mz_node->usage_in_excess)
499 * We can't avoid mem cgroups that are over their soft
500 * limit by the same amount
502 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
505 rb_link_node(&mz->tree_node, parent, p);
506 rb_insert_color(&mz->tree_node, &mctz->rb_root);
511 __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
512 struct mem_cgroup_per_zone *mz,
513 struct mem_cgroup_tree_per_zone *mctz)
517 rb_erase(&mz->tree_node, &mctz->rb_root);
522 mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
523 struct mem_cgroup_per_zone *mz,
524 struct mem_cgroup_tree_per_zone *mctz)
526 spin_lock(&mctz->lock);
527 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
528 spin_unlock(&mctz->lock);
532 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
534 unsigned long long excess;
535 struct mem_cgroup_per_zone *mz;
536 struct mem_cgroup_tree_per_zone *mctz;
537 int nid = page_to_nid(page);
538 int zid = page_zonenum(page);
539 mctz = soft_limit_tree_from_page(page);
542 * Necessary to update all ancestors when hierarchy is used.
543 * because their event counter is not touched.
545 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
546 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
547 excess = res_counter_soft_limit_excess(&memcg->res);
549 * We have to update the tree if mz is on RB-tree or
550 * mem is over its softlimit.
552 if (excess || mz->on_tree) {
553 spin_lock(&mctz->lock);
554 /* if on-tree, remove it */
556 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
558 * Insert again. mz->usage_in_excess will be updated.
559 * If excess is 0, no tree ops.
561 __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
562 spin_unlock(&mctz->lock);
567 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
570 struct mem_cgroup_per_zone *mz;
571 struct mem_cgroup_tree_per_zone *mctz;
573 for_each_node_state(node, N_POSSIBLE) {
574 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
575 mz = mem_cgroup_zoneinfo(memcg, node, zone);
576 mctz = soft_limit_tree_node_zone(node, zone);
577 mem_cgroup_remove_exceeded(memcg, mz, mctz);
582 static struct mem_cgroup_per_zone *
583 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
585 struct rb_node *rightmost = NULL;
586 struct mem_cgroup_per_zone *mz;
590 rightmost = rb_last(&mctz->rb_root);
592 goto done; /* Nothing to reclaim from */
594 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
596 * Remove the node now but someone else can add it back,
597 * we will to add it back at the end of reclaim to its correct
598 * position in the tree.
600 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
601 if (!res_counter_soft_limit_excess(&mz->mem->res) ||
602 !css_tryget(&mz->mem->css))
608 static struct mem_cgroup_per_zone *
609 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
611 struct mem_cgroup_per_zone *mz;
613 spin_lock(&mctz->lock);
614 mz = __mem_cgroup_largest_soft_limit_node(mctz);
615 spin_unlock(&mctz->lock);
620 * Implementation Note: reading percpu statistics for memcg.
622 * Both of vmstat[] and percpu_counter has threshold and do periodic
623 * synchronization to implement "quick" read. There are trade-off between
624 * reading cost and precision of value. Then, we may have a chance to implement
625 * a periodic synchronizion of counter in memcg's counter.
627 * But this _read() function is used for user interface now. The user accounts
628 * memory usage by memory cgroup and he _always_ requires exact value because
629 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
630 * have to visit all online cpus and make sum. So, for now, unnecessary
631 * synchronization is not implemented. (just implemented for cpu hotplug)
633 * If there are kernel internal actions which can make use of some not-exact
634 * value, and reading all cpu value can be performance bottleneck in some
635 * common workload, threashold and synchonization as vmstat[] should be
638 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
639 enum mem_cgroup_stat_index idx)
645 for_each_online_cpu(cpu)
646 val += per_cpu(memcg->stat->count[idx], cpu);
647 #ifdef CONFIG_HOTPLUG_CPU
648 spin_lock(&memcg->pcp_counter_lock);
649 val += memcg->nocpu_base.count[idx];
650 spin_unlock(&memcg->pcp_counter_lock);
656 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
659 int val = (charge) ? 1 : -1;
660 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAPOUT], val);
663 void mem_cgroup_pgfault(struct mem_cgroup *memcg, int val)
665 this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT], val);
668 void mem_cgroup_pgmajfault(struct mem_cgroup *memcg, int val)
670 this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT], val);
673 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
674 enum mem_cgroup_events_index idx)
676 unsigned long val = 0;
679 for_each_online_cpu(cpu)
680 val += per_cpu(memcg->stat->events[idx], cpu);
681 #ifdef CONFIG_HOTPLUG_CPU
682 spin_lock(&memcg->pcp_counter_lock);
683 val += memcg->nocpu_base.events[idx];
684 spin_unlock(&memcg->pcp_counter_lock);
689 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
690 bool file, int nr_pages)
695 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
698 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
701 /* pagein of a big page is an event. So, ignore page size */
703 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
705 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
706 nr_pages = -nr_pages; /* for event */
709 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT], nr_pages);
715 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
716 unsigned int lru_mask)
718 struct mem_cgroup_per_zone *mz;
720 unsigned long ret = 0;
722 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
725 if (BIT(l) & lru_mask)
726 ret += MEM_CGROUP_ZSTAT(mz, l);
732 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
733 int nid, unsigned int lru_mask)
738 for (zid = 0; zid < MAX_NR_ZONES; zid++)
739 total += mem_cgroup_zone_nr_lru_pages(memcg,
745 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
746 unsigned int lru_mask)
751 for_each_node_state(nid, N_HIGH_MEMORY)
752 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
756 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
757 enum mem_cgroup_events_target target)
759 unsigned long val, next;
761 val = __this_cpu_read(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT]);
762 next = __this_cpu_read(memcg->stat->targets[target]);
763 /* from time_after() in jiffies.h */
764 if ((long)next - (long)val < 0) {
766 case MEM_CGROUP_TARGET_THRESH:
767 next = val + THRESHOLDS_EVENTS_TARGET;
769 case MEM_CGROUP_TARGET_SOFTLIMIT:
770 next = val + SOFTLIMIT_EVENTS_TARGET;
772 case MEM_CGROUP_TARGET_NUMAINFO:
773 next = val + NUMAINFO_EVENTS_TARGET;
778 __this_cpu_write(memcg->stat->targets[target], next);
785 * Check events in order.
788 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
791 /* threshold event is triggered in finer grain than soft limit */
792 if (unlikely(mem_cgroup_event_ratelimit(memcg,
793 MEM_CGROUP_TARGET_THRESH))) {
794 bool do_softlimit, do_numainfo;
796 do_softlimit = mem_cgroup_event_ratelimit(memcg,
797 MEM_CGROUP_TARGET_SOFTLIMIT);
799 do_numainfo = mem_cgroup_event_ratelimit(memcg,
800 MEM_CGROUP_TARGET_NUMAINFO);
804 mem_cgroup_threshold(memcg);
805 if (unlikely(do_softlimit))
806 mem_cgroup_update_tree(memcg, page);
808 if (unlikely(do_numainfo))
809 atomic_inc(&memcg->numainfo_events);
815 struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
817 return container_of(cgroup_subsys_state(cont,
818 mem_cgroup_subsys_id), struct mem_cgroup,
822 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
825 * mm_update_next_owner() may clear mm->owner to NULL
826 * if it races with swapoff, page migration, etc.
827 * So this can be called with p == NULL.
832 return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
833 struct mem_cgroup, css);
836 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
838 struct mem_cgroup *memcg = NULL;
843 * Because we have no locks, mm->owner's may be being moved to other
844 * cgroup. We use css_tryget() here even if this looks
845 * pessimistic (rather than adding locks here).
849 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
850 if (unlikely(!memcg))
852 } while (!css_tryget(&memcg->css));
858 * mem_cgroup_iter - iterate over memory cgroup hierarchy
859 * @root: hierarchy root
860 * @prev: previously returned memcg, NULL on first invocation
861 * @reclaim: cookie for shared reclaim walks, NULL for full walks
863 * Returns references to children of the hierarchy below @root, or
864 * @root itself, or %NULL after a full round-trip.
866 * Caller must pass the return value in @prev on subsequent
867 * invocations for reference counting, or use mem_cgroup_iter_break()
868 * to cancel a hierarchy walk before the round-trip is complete.
870 * Reclaimers can specify a zone and a priority level in @reclaim to
871 * divide up the memcgs in the hierarchy among all concurrent
872 * reclaimers operating on the same zone and priority.
874 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
875 struct mem_cgroup *prev,
876 struct mem_cgroup_reclaim_cookie *reclaim)
878 struct mem_cgroup *memcg = NULL;
881 if (mem_cgroup_disabled())
885 root = root_mem_cgroup;
887 if (prev && !reclaim)
888 id = css_id(&prev->css);
890 if (prev && prev != root)
893 if (!root->use_hierarchy && root != root_mem_cgroup) {
900 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
901 struct cgroup_subsys_state *css;
904 int nid = zone_to_nid(reclaim->zone);
905 int zid = zone_idx(reclaim->zone);
906 struct mem_cgroup_per_zone *mz;
908 mz = mem_cgroup_zoneinfo(root, nid, zid);
909 iter = &mz->reclaim_iter[reclaim->priority];
910 if (prev && reclaim->generation != iter->generation)
916 css = css_get_next(&mem_cgroup_subsys, id + 1, &root->css, &id);
918 if (css == &root->css || css_tryget(css))
919 memcg = container_of(css,
920 struct mem_cgroup, css);
929 else if (!prev && memcg)
930 reclaim->generation = iter->generation;
940 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
941 * @root: hierarchy root
942 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
944 void mem_cgroup_iter_break(struct mem_cgroup *root,
945 struct mem_cgroup *prev)
948 root = root_mem_cgroup;
949 if (prev && prev != root)
954 * Iteration constructs for visiting all cgroups (under a tree). If
955 * loops are exited prematurely (break), mem_cgroup_iter_break() must
956 * be used for reference counting.
958 #define for_each_mem_cgroup_tree(iter, root) \
959 for (iter = mem_cgroup_iter(root, NULL, NULL); \
961 iter = mem_cgroup_iter(root, iter, NULL))
963 #define for_each_mem_cgroup(iter) \
964 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
966 iter = mem_cgroup_iter(NULL, iter, NULL))
968 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
970 return (memcg == root_mem_cgroup);
973 void mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
975 struct mem_cgroup *memcg;
981 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
982 if (unlikely(!memcg))
987 mem_cgroup_pgmajfault(memcg, 1);
990 mem_cgroup_pgfault(memcg, 1);
998 EXPORT_SYMBOL(mem_cgroup_count_vm_event);
1001 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1002 * @zone: zone of the wanted lruvec
1003 * @mem: memcg of the wanted lruvec
1005 * Returns the lru list vector holding pages for the given @zone and
1006 * @mem. This can be the global zone lruvec, if the memory controller
1009 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1010 struct mem_cgroup *memcg)
1012 struct mem_cgroup_per_zone *mz;
1014 if (mem_cgroup_disabled())
1015 return &zone->lruvec;
1017 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1022 * Following LRU functions are allowed to be used without PCG_LOCK.
1023 * Operations are called by routine of global LRU independently from memcg.
1024 * What we have to take care of here is validness of pc->mem_cgroup.
1026 * Changes to pc->mem_cgroup happens when
1029 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1030 * It is added to LRU before charge.
1031 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1032 * When moving account, the page is not on LRU. It's isolated.
1036 * mem_cgroup_lru_add_list - account for adding an lru page and return lruvec
1037 * @zone: zone of the page
1041 * This function accounts for @page being added to @lru, and returns
1042 * the lruvec for the given @zone and the memcg @page is charged to.
1044 * The callsite is then responsible for physically linking the page to
1045 * the returned lruvec->lists[@lru].
1047 struct lruvec *mem_cgroup_lru_add_list(struct zone *zone, struct page *page,
1050 struct mem_cgroup_per_zone *mz;
1051 struct mem_cgroup *memcg;
1052 struct page_cgroup *pc;
1054 if (mem_cgroup_disabled())
1055 return &zone->lruvec;
1057 pc = lookup_page_cgroup(page);
1058 VM_BUG_ON(PageCgroupAcctLRU(pc));
1061 * SetPageLRU SetPageCgroupUsed
1063 * PageCgroupUsed && add to memcg LRU PageLRU && add to memcg LRU
1065 * Ensure that one of the two sides adds the page to the memcg
1066 * LRU during a race.
1070 * If the page is uncharged, it may be freed soon, but it
1071 * could also be swap cache (readahead, swapoff) that needs to
1072 * be reclaimable in the future. root_mem_cgroup will babysit
1073 * it for the time being.
1075 if (PageCgroupUsed(pc)) {
1076 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1078 memcg = pc->mem_cgroup;
1079 SetPageCgroupAcctLRU(pc);
1081 memcg = root_mem_cgroup;
1082 mz = page_cgroup_zoneinfo(memcg, page);
1083 /* compound_order() is stabilized through lru_lock */
1084 MEM_CGROUP_ZSTAT(mz, lru) += 1 << compound_order(page);
1089 * mem_cgroup_lru_del_list - account for removing an lru page
1093 * This function accounts for @page being removed from @lru.
1095 * The callsite is then responsible for physically unlinking
1098 void mem_cgroup_lru_del_list(struct page *page, enum lru_list lru)
1100 struct mem_cgroup_per_zone *mz;
1101 struct mem_cgroup *memcg;
1102 struct page_cgroup *pc;
1104 if (mem_cgroup_disabled())
1107 pc = lookup_page_cgroup(page);
1109 * root_mem_cgroup babysits uncharged LRU pages, but
1110 * PageCgroupUsed is cleared when the page is about to get
1111 * freed. PageCgroupAcctLRU remembers whether the
1112 * LRU-accounting happened against pc->mem_cgroup or
1115 if (TestClearPageCgroupAcctLRU(pc)) {
1116 VM_BUG_ON(!pc->mem_cgroup);
1117 memcg = pc->mem_cgroup;
1119 memcg = root_mem_cgroup;
1120 mz = page_cgroup_zoneinfo(memcg, page);
1121 /* huge page split is done under lru_lock. so, we have no races. */
1122 MEM_CGROUP_ZSTAT(mz, lru) -= 1 << compound_order(page);
1125 void mem_cgroup_lru_del(struct page *page)
1127 mem_cgroup_lru_del_list(page, page_lru(page));
1131 * mem_cgroup_lru_move_lists - account for moving a page between lrus
1132 * @zone: zone of the page
1134 * @from: current lru
1137 * This function accounts for @page being moved between the lrus @from
1138 * and @to, and returns the lruvec for the given @zone and the memcg
1139 * @page is charged to.
1141 * The callsite is then responsible for physically relinking
1142 * @page->lru to the returned lruvec->lists[@to].
1144 struct lruvec *mem_cgroup_lru_move_lists(struct zone *zone,
1149 /* XXX: Optimize this, especially for @from == @to */
1150 mem_cgroup_lru_del_list(page, from);
1151 return mem_cgroup_lru_add_list(zone, page, to);
1155 * At handling SwapCache and other FUSE stuff, pc->mem_cgroup may be changed
1156 * while it's linked to lru because the page may be reused after it's fully
1157 * uncharged. To handle that, unlink page_cgroup from LRU when charge it again.
1158 * It's done under lock_page and expected that zone->lru_lock isnever held.
1160 static void mem_cgroup_lru_del_before_commit(struct page *page)
1163 unsigned long flags;
1164 struct zone *zone = page_zone(page);
1165 struct page_cgroup *pc = lookup_page_cgroup(page);
1168 * Doing this check without taking ->lru_lock seems wrong but this
1169 * is safe. Because if page_cgroup's USED bit is unset, the page
1170 * will not be added to any memcg's LRU. If page_cgroup's USED bit is
1171 * set, the commit after this will fail, anyway.
1172 * This all charge/uncharge is done under some mutual execustion.
1173 * So, we don't need to taking care of changes in USED bit.
1175 if (likely(!PageLRU(page)))
1178 spin_lock_irqsave(&zone->lru_lock, flags);
1179 lru = page_lru(page);
1181 * The uncharged page could still be registered to the LRU of
1182 * the stale pc->mem_cgroup.
1184 * As pc->mem_cgroup is about to get overwritten, the old LRU
1185 * accounting needs to be taken care of. Let root_mem_cgroup
1186 * babysit the page until the new memcg is responsible for it.
1188 * The PCG_USED bit is guarded by lock_page() as the page is
1189 * swapcache/pagecache.
1191 if (PageLRU(page) && PageCgroupAcctLRU(pc) && !PageCgroupUsed(pc)) {
1192 del_page_from_lru_list(zone, page, lru);
1193 add_page_to_lru_list(zone, page, lru);
1195 spin_unlock_irqrestore(&zone->lru_lock, flags);
1198 static void mem_cgroup_lru_add_after_commit(struct page *page)
1201 unsigned long flags;
1202 struct zone *zone = page_zone(page);
1203 struct page_cgroup *pc = lookup_page_cgroup(page);
1206 * SetPageLRU SetPageCgroupUsed
1208 * PageCgroupUsed && add to memcg LRU PageLRU && add to memcg LRU
1210 * Ensure that one of the two sides adds the page to the memcg
1211 * LRU during a race.
1214 /* taking care of that the page is added to LRU while we commit it */
1215 if (likely(!PageLRU(page)))
1217 spin_lock_irqsave(&zone->lru_lock, flags);
1218 lru = page_lru(page);
1220 * If the page is not on the LRU, someone will soon put it
1221 * there. If it is, and also already accounted for on the
1222 * memcg-side, it must be on the right lruvec as setting
1223 * pc->mem_cgroup and PageCgroupUsed is properly ordered.
1224 * Otherwise, root_mem_cgroup has been babysitting the page
1225 * during the charge. Move it to the new memcg now.
1227 if (PageLRU(page) && !PageCgroupAcctLRU(pc)) {
1228 del_page_from_lru_list(zone, page, lru);
1229 add_page_to_lru_list(zone, page, lru);
1231 spin_unlock_irqrestore(&zone->lru_lock, flags);
1235 * Checks whether given mem is same or in the root_mem_cgroup's
1238 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1239 struct mem_cgroup *memcg)
1241 if (root_memcg != memcg) {
1242 return (root_memcg->use_hierarchy &&
1243 css_is_ancestor(&memcg->css, &root_memcg->css));
1249 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *memcg)
1252 struct mem_cgroup *curr = NULL;
1253 struct task_struct *p;
1255 p = find_lock_task_mm(task);
1258 curr = try_get_mem_cgroup_from_mm(p->mm);
1263 * We should check use_hierarchy of "memcg" not "curr". Because checking
1264 * use_hierarchy of "curr" here make this function true if hierarchy is
1265 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1266 * hierarchy(even if use_hierarchy is disabled in "memcg").
1268 ret = mem_cgroup_same_or_subtree(memcg, curr);
1269 css_put(&curr->css);
1273 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg, struct zone *zone)
1275 unsigned long inactive_ratio;
1276 int nid = zone_to_nid(zone);
1277 int zid = zone_idx(zone);
1278 unsigned long inactive;
1279 unsigned long active;
1282 inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1283 BIT(LRU_INACTIVE_ANON));
1284 active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1285 BIT(LRU_ACTIVE_ANON));
1287 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1289 inactive_ratio = int_sqrt(10 * gb);
1293 return inactive * inactive_ratio < active;
1296 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg, struct zone *zone)
1298 unsigned long active;
1299 unsigned long inactive;
1300 int zid = zone_idx(zone);
1301 int nid = zone_to_nid(zone);
1303 inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1304 BIT(LRU_INACTIVE_FILE));
1305 active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1306 BIT(LRU_ACTIVE_FILE));
1308 return (active > inactive);
1311 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
1314 int nid = zone_to_nid(zone);
1315 int zid = zone_idx(zone);
1316 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
1318 return &mz->reclaim_stat;
1321 struct zone_reclaim_stat *
1322 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
1324 struct page_cgroup *pc;
1325 struct mem_cgroup_per_zone *mz;
1327 if (mem_cgroup_disabled())
1330 pc = lookup_page_cgroup(page);
1331 if (!PageCgroupUsed(pc))
1333 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1335 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1336 return &mz->reclaim_stat;
1339 #define mem_cgroup_from_res_counter(counter, member) \
1340 container_of(counter, struct mem_cgroup, member)
1343 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1344 * @mem: the memory cgroup
1346 * Returns the maximum amount of memory @mem can be charged with, in
1349 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1351 unsigned long long margin;
1353 margin = res_counter_margin(&memcg->res);
1354 if (do_swap_account)
1355 margin = min(margin, res_counter_margin(&memcg->memsw));
1356 return margin >> PAGE_SHIFT;
1359 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1361 struct cgroup *cgrp = memcg->css.cgroup;
1364 if (cgrp->parent == NULL)
1365 return vm_swappiness;
1367 return memcg->swappiness;
1370 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1375 spin_lock(&memcg->pcp_counter_lock);
1376 for_each_online_cpu(cpu)
1377 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) += 1;
1378 memcg->nocpu_base.count[MEM_CGROUP_ON_MOVE] += 1;
1379 spin_unlock(&memcg->pcp_counter_lock);
1385 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1392 spin_lock(&memcg->pcp_counter_lock);
1393 for_each_online_cpu(cpu)
1394 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) -= 1;
1395 memcg->nocpu_base.count[MEM_CGROUP_ON_MOVE] -= 1;
1396 spin_unlock(&memcg->pcp_counter_lock);
1400 * 2 routines for checking "mem" is under move_account() or not.
1402 * mem_cgroup_stealed() - checking a cgroup is mc.from or not. This is used
1403 * for avoiding race in accounting. If true,
1404 * pc->mem_cgroup may be overwritten.
1406 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1407 * under hierarchy of moving cgroups. This is for
1408 * waiting at hith-memory prressure caused by "move".
1411 static bool mem_cgroup_stealed(struct mem_cgroup *memcg)
1413 VM_BUG_ON(!rcu_read_lock_held());
1414 return this_cpu_read(memcg->stat->count[MEM_CGROUP_ON_MOVE]) > 0;
1417 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1419 struct mem_cgroup *from;
1420 struct mem_cgroup *to;
1423 * Unlike task_move routines, we access mc.to, mc.from not under
1424 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1426 spin_lock(&mc.lock);
1432 ret = mem_cgroup_same_or_subtree(memcg, from)
1433 || mem_cgroup_same_or_subtree(memcg, to);
1435 spin_unlock(&mc.lock);
1439 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1441 if (mc.moving_task && current != mc.moving_task) {
1442 if (mem_cgroup_under_move(memcg)) {
1444 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1445 /* moving charge context might have finished. */
1448 finish_wait(&mc.waitq, &wait);
1456 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1457 * @memcg: The memory cgroup that went over limit
1458 * @p: Task that is going to be killed
1460 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1463 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1465 struct cgroup *task_cgrp;
1466 struct cgroup *mem_cgrp;
1468 * Need a buffer in BSS, can't rely on allocations. The code relies
1469 * on the assumption that OOM is serialized for memory controller.
1470 * If this assumption is broken, revisit this code.
1472 static char memcg_name[PATH_MAX];
1481 mem_cgrp = memcg->css.cgroup;
1482 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1484 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1487 * Unfortunately, we are unable to convert to a useful name
1488 * But we'll still print out the usage information
1495 printk(KERN_INFO "Task in %s killed", memcg_name);
1498 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1506 * Continues from above, so we don't need an KERN_ level
1508 printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1511 printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1512 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1513 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1514 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1515 printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1517 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1518 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1519 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1523 * This function returns the number of memcg under hierarchy tree. Returns
1524 * 1(self count) if no children.
1526 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1529 struct mem_cgroup *iter;
1531 for_each_mem_cgroup_tree(iter, memcg)
1537 * Return the memory (and swap, if configured) limit for a memcg.
1539 u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1544 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1545 limit += total_swap_pages << PAGE_SHIFT;
1547 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1549 * If memsw is finite and limits the amount of swap space available
1550 * to this memcg, return that limit.
1552 return min(limit, memsw);
1555 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1557 unsigned long flags)
1559 unsigned long total = 0;
1560 bool noswap = false;
1563 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1565 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1568 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1570 drain_all_stock_async(memcg);
1571 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1573 * Allow limit shrinkers, which are triggered directly
1574 * by userspace, to catch signals and stop reclaim
1575 * after minimal progress, regardless of the margin.
1577 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1579 if (mem_cgroup_margin(memcg))
1582 * If nothing was reclaimed after two attempts, there
1583 * may be no reclaimable pages in this hierarchy.
1592 * test_mem_cgroup_node_reclaimable
1593 * @mem: the target memcg
1594 * @nid: the node ID to be checked.
1595 * @noswap : specify true here if the user wants flle only information.
1597 * This function returns whether the specified memcg contains any
1598 * reclaimable pages on a node. Returns true if there are any reclaimable
1599 * pages in the node.
1601 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1602 int nid, bool noswap)
1604 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1606 if (noswap || !total_swap_pages)
1608 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1613 #if MAX_NUMNODES > 1
1616 * Always updating the nodemask is not very good - even if we have an empty
1617 * list or the wrong list here, we can start from some node and traverse all
1618 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1621 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1625 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1626 * pagein/pageout changes since the last update.
1628 if (!atomic_read(&memcg->numainfo_events))
1630 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1633 /* make a nodemask where this memcg uses memory from */
1634 memcg->scan_nodes = node_states[N_HIGH_MEMORY];
1636 for_each_node_mask(nid, node_states[N_HIGH_MEMORY]) {
1638 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1639 node_clear(nid, memcg->scan_nodes);
1642 atomic_set(&memcg->numainfo_events, 0);
1643 atomic_set(&memcg->numainfo_updating, 0);
1647 * Selecting a node where we start reclaim from. Because what we need is just
1648 * reducing usage counter, start from anywhere is O,K. Considering
1649 * memory reclaim from current node, there are pros. and cons.
1651 * Freeing memory from current node means freeing memory from a node which
1652 * we'll use or we've used. So, it may make LRU bad. And if several threads
1653 * hit limits, it will see a contention on a node. But freeing from remote
1654 * node means more costs for memory reclaim because of memory latency.
1656 * Now, we use round-robin. Better algorithm is welcomed.
1658 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1662 mem_cgroup_may_update_nodemask(memcg);
1663 node = memcg->last_scanned_node;
1665 node = next_node(node, memcg->scan_nodes);
1666 if (node == MAX_NUMNODES)
1667 node = first_node(memcg->scan_nodes);
1669 * We call this when we hit limit, not when pages are added to LRU.
1670 * No LRU may hold pages because all pages are UNEVICTABLE or
1671 * memcg is too small and all pages are not on LRU. In that case,
1672 * we use curret node.
1674 if (unlikely(node == MAX_NUMNODES))
1675 node = numa_node_id();
1677 memcg->last_scanned_node = node;
1682 * Check all nodes whether it contains reclaimable pages or not.
1683 * For quick scan, we make use of scan_nodes. This will allow us to skip
1684 * unused nodes. But scan_nodes is lazily updated and may not cotain
1685 * enough new information. We need to do double check.
1687 bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1692 * quick check...making use of scan_node.
1693 * We can skip unused nodes.
1695 if (!nodes_empty(memcg->scan_nodes)) {
1696 for (nid = first_node(memcg->scan_nodes);
1698 nid = next_node(nid, memcg->scan_nodes)) {
1700 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1705 * Check rest of nodes.
1707 for_each_node_state(nid, N_HIGH_MEMORY) {
1708 if (node_isset(nid, memcg->scan_nodes))
1710 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1717 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1722 bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1724 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
1728 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1731 unsigned long *total_scanned)
1733 struct mem_cgroup *victim = NULL;
1736 unsigned long excess;
1737 unsigned long nr_scanned;
1738 struct mem_cgroup_reclaim_cookie reclaim = {
1743 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
1746 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1751 * If we have not been able to reclaim
1752 * anything, it might because there are
1753 * no reclaimable pages under this hierarchy
1758 * We want to do more targeted reclaim.
1759 * excess >> 2 is not to excessive so as to
1760 * reclaim too much, nor too less that we keep
1761 * coming back to reclaim from this cgroup
1763 if (total >= (excess >> 2) ||
1764 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1769 if (!mem_cgroup_reclaimable(victim, false))
1771 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1773 *total_scanned += nr_scanned;
1774 if (!res_counter_soft_limit_excess(&root_memcg->res))
1777 mem_cgroup_iter_break(root_memcg, victim);
1782 * Check OOM-Killer is already running under our hierarchy.
1783 * If someone is running, return false.
1784 * Has to be called with memcg_oom_lock
1786 static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg)
1788 struct mem_cgroup *iter, *failed = NULL;
1790 for_each_mem_cgroup_tree(iter, memcg) {
1791 if (iter->oom_lock) {
1793 * this subtree of our hierarchy is already locked
1794 * so we cannot give a lock.
1797 mem_cgroup_iter_break(memcg, iter);
1800 iter->oom_lock = true;
1807 * OK, we failed to lock the whole subtree so we have to clean up
1808 * what we set up to the failing subtree
1810 for_each_mem_cgroup_tree(iter, memcg) {
1811 if (iter == failed) {
1812 mem_cgroup_iter_break(memcg, iter);
1815 iter->oom_lock = false;
1821 * Has to be called with memcg_oom_lock
1823 static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1825 struct mem_cgroup *iter;
1827 for_each_mem_cgroup_tree(iter, memcg)
1828 iter->oom_lock = false;
1832 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1834 struct mem_cgroup *iter;
1836 for_each_mem_cgroup_tree(iter, memcg)
1837 atomic_inc(&iter->under_oom);
1840 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1842 struct mem_cgroup *iter;
1845 * When a new child is created while the hierarchy is under oom,
1846 * mem_cgroup_oom_lock() may not be called. We have to use
1847 * atomic_add_unless() here.
1849 for_each_mem_cgroup_tree(iter, memcg)
1850 atomic_add_unless(&iter->under_oom, -1, 0);
1853 static DEFINE_SPINLOCK(memcg_oom_lock);
1854 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1856 struct oom_wait_info {
1857 struct mem_cgroup *mem;
1861 static int memcg_oom_wake_function(wait_queue_t *wait,
1862 unsigned mode, int sync, void *arg)
1864 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg,
1866 struct oom_wait_info *oom_wait_info;
1868 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1869 oom_wait_memcg = oom_wait_info->mem;
1872 * Both of oom_wait_info->mem and wake_mem are stable under us.
1873 * Then we can use css_is_ancestor without taking care of RCU.
1875 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
1876 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
1878 return autoremove_wake_function(wait, mode, sync, arg);
1881 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
1883 /* for filtering, pass "memcg" as argument. */
1884 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1887 static void memcg_oom_recover(struct mem_cgroup *memcg)
1889 if (memcg && atomic_read(&memcg->under_oom))
1890 memcg_wakeup_oom(memcg);
1894 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1896 bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask)
1898 struct oom_wait_info owait;
1899 bool locked, need_to_kill;
1902 owait.wait.flags = 0;
1903 owait.wait.func = memcg_oom_wake_function;
1904 owait.wait.private = current;
1905 INIT_LIST_HEAD(&owait.wait.task_list);
1906 need_to_kill = true;
1907 mem_cgroup_mark_under_oom(memcg);
1909 /* At first, try to OOM lock hierarchy under memcg.*/
1910 spin_lock(&memcg_oom_lock);
1911 locked = mem_cgroup_oom_lock(memcg);
1913 * Even if signal_pending(), we can't quit charge() loop without
1914 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1915 * under OOM is always welcomed, use TASK_KILLABLE here.
1917 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1918 if (!locked || memcg->oom_kill_disable)
1919 need_to_kill = false;
1921 mem_cgroup_oom_notify(memcg);
1922 spin_unlock(&memcg_oom_lock);
1925 finish_wait(&memcg_oom_waitq, &owait.wait);
1926 mem_cgroup_out_of_memory(memcg, mask);
1929 finish_wait(&memcg_oom_waitq, &owait.wait);
1931 spin_lock(&memcg_oom_lock);
1933 mem_cgroup_oom_unlock(memcg);
1934 memcg_wakeup_oom(memcg);
1935 spin_unlock(&memcg_oom_lock);
1937 mem_cgroup_unmark_under_oom(memcg);
1939 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
1941 /* Give chance to dying process */
1942 schedule_timeout_uninterruptible(1);
1947 * Currently used to update mapped file statistics, but the routine can be
1948 * generalized to update other statistics as well.
1950 * Notes: Race condition
1952 * We usually use page_cgroup_lock() for accessing page_cgroup member but
1953 * it tends to be costly. But considering some conditions, we doesn't need
1954 * to do so _always_.
1956 * Considering "charge", lock_page_cgroup() is not required because all
1957 * file-stat operations happen after a page is attached to radix-tree. There
1958 * are no race with "charge".
1960 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
1961 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
1962 * if there are race with "uncharge". Statistics itself is properly handled
1965 * Considering "move", this is an only case we see a race. To make the race
1966 * small, we check MEM_CGROUP_ON_MOVE percpu value and detect there are
1967 * possibility of race condition. If there is, we take a lock.
1970 void mem_cgroup_update_page_stat(struct page *page,
1971 enum mem_cgroup_page_stat_item idx, int val)
1973 struct mem_cgroup *memcg;
1974 struct page_cgroup *pc = lookup_page_cgroup(page);
1975 bool need_unlock = false;
1976 unsigned long uninitialized_var(flags);
1982 memcg = pc->mem_cgroup;
1983 if (unlikely(!memcg || !PageCgroupUsed(pc)))
1985 /* pc->mem_cgroup is unstable ? */
1986 if (unlikely(mem_cgroup_stealed(memcg)) || PageTransHuge(page)) {
1987 /* take a lock against to access pc->mem_cgroup */
1988 move_lock_page_cgroup(pc, &flags);
1990 memcg = pc->mem_cgroup;
1991 if (!memcg || !PageCgroupUsed(pc))
1996 case MEMCG_NR_FILE_MAPPED:
1998 SetPageCgroupFileMapped(pc);
1999 else if (!page_mapped(page))
2000 ClearPageCgroupFileMapped(pc);
2001 idx = MEM_CGROUP_STAT_FILE_MAPPED;
2007 this_cpu_add(memcg->stat->count[idx], val);
2010 if (unlikely(need_unlock))
2011 move_unlock_page_cgroup(pc, &flags);
2015 EXPORT_SYMBOL(mem_cgroup_update_page_stat);
2018 * size of first charge trial. "32" comes from vmscan.c's magic value.
2019 * TODO: maybe necessary to use big numbers in big irons.
2021 #define CHARGE_BATCH 32U
2022 struct memcg_stock_pcp {
2023 struct mem_cgroup *cached; /* this never be root cgroup */
2024 unsigned int nr_pages;
2025 struct work_struct work;
2026 unsigned long flags;
2027 #define FLUSHING_CACHED_CHARGE (0)
2029 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2030 static DEFINE_MUTEX(percpu_charge_mutex);
2033 * Try to consume stocked charge on this cpu. If success, one page is consumed
2034 * from local stock and true is returned. If the stock is 0 or charges from a
2035 * cgroup which is not current target, returns false. This stock will be
2038 static bool consume_stock(struct mem_cgroup *memcg)
2040 struct memcg_stock_pcp *stock;
2043 stock = &get_cpu_var(memcg_stock);
2044 if (memcg == stock->cached && stock->nr_pages)
2046 else /* need to call res_counter_charge */
2048 put_cpu_var(memcg_stock);
2053 * Returns stocks cached in percpu to res_counter and reset cached information.
2055 static void drain_stock(struct memcg_stock_pcp *stock)
2057 struct mem_cgroup *old = stock->cached;
2059 if (stock->nr_pages) {
2060 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2062 res_counter_uncharge(&old->res, bytes);
2063 if (do_swap_account)
2064 res_counter_uncharge(&old->memsw, bytes);
2065 stock->nr_pages = 0;
2067 stock->cached = NULL;
2071 * This must be called under preempt disabled or must be called by
2072 * a thread which is pinned to local cpu.
2074 static void drain_local_stock(struct work_struct *dummy)
2076 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2078 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2082 * Cache charges(val) which is from res_counter, to local per_cpu area.
2083 * This will be consumed by consume_stock() function, later.
2085 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2087 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2089 if (stock->cached != memcg) { /* reset if necessary */
2091 stock->cached = memcg;
2093 stock->nr_pages += nr_pages;
2094 put_cpu_var(memcg_stock);
2098 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2099 * of the hierarchy under it. sync flag says whether we should block
2100 * until the work is done.
2102 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2106 /* Notify other cpus that system-wide "drain" is running */
2109 for_each_online_cpu(cpu) {
2110 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2111 struct mem_cgroup *memcg;
2113 memcg = stock->cached;
2114 if (!memcg || !stock->nr_pages)
2116 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2118 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2120 drain_local_stock(&stock->work);
2122 schedule_work_on(cpu, &stock->work);
2130 for_each_online_cpu(cpu) {
2131 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2132 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2133 flush_work(&stock->work);
2140 * Tries to drain stocked charges in other cpus. This function is asynchronous
2141 * and just put a work per cpu for draining localy on each cpu. Caller can
2142 * expects some charges will be back to res_counter later but cannot wait for
2145 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2148 * If someone calls draining, avoid adding more kworker runs.
2150 if (!mutex_trylock(&percpu_charge_mutex))
2152 drain_all_stock(root_memcg, false);
2153 mutex_unlock(&percpu_charge_mutex);
2156 /* This is a synchronous drain interface. */
2157 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2159 /* called when force_empty is called */
2160 mutex_lock(&percpu_charge_mutex);
2161 drain_all_stock(root_memcg, true);
2162 mutex_unlock(&percpu_charge_mutex);
2166 * This function drains percpu counter value from DEAD cpu and
2167 * move it to local cpu. Note that this function can be preempted.
2169 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2173 spin_lock(&memcg->pcp_counter_lock);
2174 for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) {
2175 long x = per_cpu(memcg->stat->count[i], cpu);
2177 per_cpu(memcg->stat->count[i], cpu) = 0;
2178 memcg->nocpu_base.count[i] += x;
2180 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2181 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2183 per_cpu(memcg->stat->events[i], cpu) = 0;
2184 memcg->nocpu_base.events[i] += x;
2186 /* need to clear ON_MOVE value, works as a kind of lock. */
2187 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) = 0;
2188 spin_unlock(&memcg->pcp_counter_lock);
2191 static void synchronize_mem_cgroup_on_move(struct mem_cgroup *memcg, int cpu)
2193 int idx = MEM_CGROUP_ON_MOVE;
2195 spin_lock(&memcg->pcp_counter_lock);
2196 per_cpu(memcg->stat->count[idx], cpu) = memcg->nocpu_base.count[idx];
2197 spin_unlock(&memcg->pcp_counter_lock);
2200 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2201 unsigned long action,
2204 int cpu = (unsigned long)hcpu;
2205 struct memcg_stock_pcp *stock;
2206 struct mem_cgroup *iter;
2208 if ((action == CPU_ONLINE)) {
2209 for_each_mem_cgroup(iter)
2210 synchronize_mem_cgroup_on_move(iter, cpu);
2214 if ((action != CPU_DEAD) || action != CPU_DEAD_FROZEN)
2217 for_each_mem_cgroup(iter)
2218 mem_cgroup_drain_pcp_counter(iter, cpu);
2220 stock = &per_cpu(memcg_stock, cpu);
2226 /* See __mem_cgroup_try_charge() for details */
2228 CHARGE_OK, /* success */
2229 CHARGE_RETRY, /* need to retry but retry is not bad */
2230 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2231 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2232 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2235 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2236 unsigned int nr_pages, bool oom_check)
2238 unsigned long csize = nr_pages * PAGE_SIZE;
2239 struct mem_cgroup *mem_over_limit;
2240 struct res_counter *fail_res;
2241 unsigned long flags = 0;
2244 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2247 if (!do_swap_account)
2249 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2253 res_counter_uncharge(&memcg->res, csize);
2254 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2255 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2257 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2259 * nr_pages can be either a huge page (HPAGE_PMD_NR), a batch
2260 * of regular pages (CHARGE_BATCH), or a single regular page (1).
2262 * Never reclaim on behalf of optional batching, retry with a
2263 * single page instead.
2265 if (nr_pages == CHARGE_BATCH)
2266 return CHARGE_RETRY;
2268 if (!(gfp_mask & __GFP_WAIT))
2269 return CHARGE_WOULDBLOCK;
2271 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2272 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2273 return CHARGE_RETRY;
2275 * Even though the limit is exceeded at this point, reclaim
2276 * may have been able to free some pages. Retry the charge
2277 * before killing the task.
2279 * Only for regular pages, though: huge pages are rather
2280 * unlikely to succeed so close to the limit, and we fall back
2281 * to regular pages anyway in case of failure.
2283 if (nr_pages == 1 && ret)
2284 return CHARGE_RETRY;
2287 * At task move, charge accounts can be doubly counted. So, it's
2288 * better to wait until the end of task_move if something is going on.
2290 if (mem_cgroup_wait_acct_move(mem_over_limit))
2291 return CHARGE_RETRY;
2293 /* If we don't need to call oom-killer at el, return immediately */
2295 return CHARGE_NOMEM;
2297 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask))
2298 return CHARGE_OOM_DIE;
2300 return CHARGE_RETRY;
2304 * Unlike exported interface, "oom" parameter is added. if oom==true,
2305 * oom-killer can be invoked.
2307 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2309 unsigned int nr_pages,
2310 struct mem_cgroup **ptr,
2313 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2314 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2315 struct mem_cgroup *memcg = NULL;
2319 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2320 * in system level. So, allow to go ahead dying process in addition to
2323 if (unlikely(test_thread_flag(TIF_MEMDIE)
2324 || fatal_signal_pending(current)))
2328 * We always charge the cgroup the mm_struct belongs to.
2329 * The mm_struct's mem_cgroup changes on task migration if the
2330 * thread group leader migrates. It's possible that mm is not
2331 * set, if so charge the init_mm (happens for pagecache usage).
2336 if (*ptr) { /* css should be a valid one */
2338 VM_BUG_ON(css_is_removed(&memcg->css));
2339 if (mem_cgroup_is_root(memcg))
2341 if (nr_pages == 1 && consume_stock(memcg))
2343 css_get(&memcg->css);
2345 struct task_struct *p;
2348 p = rcu_dereference(mm->owner);
2350 * Because we don't have task_lock(), "p" can exit.
2351 * In that case, "memcg" can point to root or p can be NULL with
2352 * race with swapoff. Then, we have small risk of mis-accouning.
2353 * But such kind of mis-account by race always happens because
2354 * we don't have cgroup_mutex(). It's overkill and we allo that
2356 * (*) swapoff at el will charge against mm-struct not against
2357 * task-struct. So, mm->owner can be NULL.
2359 memcg = mem_cgroup_from_task(p);
2360 if (!memcg || mem_cgroup_is_root(memcg)) {
2364 if (nr_pages == 1 && consume_stock(memcg)) {
2366 * It seems dagerous to access memcg without css_get().
2367 * But considering how consume_stok works, it's not
2368 * necessary. If consume_stock success, some charges
2369 * from this memcg are cached on this cpu. So, we
2370 * don't need to call css_get()/css_tryget() before
2371 * calling consume_stock().
2376 /* after here, we may be blocked. we need to get refcnt */
2377 if (!css_tryget(&memcg->css)) {
2387 /* If killed, bypass charge */
2388 if (fatal_signal_pending(current)) {
2389 css_put(&memcg->css);
2394 if (oom && !nr_oom_retries) {
2396 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2399 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, oom_check);
2403 case CHARGE_RETRY: /* not in OOM situation but retry */
2405 css_put(&memcg->css);
2408 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2409 css_put(&memcg->css);
2411 case CHARGE_NOMEM: /* OOM routine works */
2413 css_put(&memcg->css);
2416 /* If oom, we never return -ENOMEM */
2419 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2420 css_put(&memcg->css);
2423 } while (ret != CHARGE_OK);
2425 if (batch > nr_pages)
2426 refill_stock(memcg, batch - nr_pages);
2427 css_put(&memcg->css);
2440 * Somemtimes we have to undo a charge we got by try_charge().
2441 * This function is for that and do uncharge, put css's refcnt.
2442 * gotten by try_charge().
2444 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2445 unsigned int nr_pages)
2447 if (!mem_cgroup_is_root(memcg)) {
2448 unsigned long bytes = nr_pages * PAGE_SIZE;
2450 res_counter_uncharge(&memcg->res, bytes);
2451 if (do_swap_account)
2452 res_counter_uncharge(&memcg->memsw, bytes);
2457 * A helper function to get mem_cgroup from ID. must be called under
2458 * rcu_read_lock(). The caller must check css_is_removed() or some if
2459 * it's concern. (dropping refcnt from swap can be called against removed
2462 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2464 struct cgroup_subsys_state *css;
2466 /* ID 0 is unused ID */
2469 css = css_lookup(&mem_cgroup_subsys, id);
2472 return container_of(css, struct mem_cgroup, css);
2475 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2477 struct mem_cgroup *memcg = NULL;
2478 struct page_cgroup *pc;
2482 VM_BUG_ON(!PageLocked(page));
2484 pc = lookup_page_cgroup(page);
2485 lock_page_cgroup(pc);
2486 if (PageCgroupUsed(pc)) {
2487 memcg = pc->mem_cgroup;
2488 if (memcg && !css_tryget(&memcg->css))
2490 } else if (PageSwapCache(page)) {
2491 ent.val = page_private(page);
2492 id = lookup_swap_cgroup(ent);
2494 memcg = mem_cgroup_lookup(id);
2495 if (memcg && !css_tryget(&memcg->css))
2499 unlock_page_cgroup(pc);
2503 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2505 unsigned int nr_pages,
2506 struct page_cgroup *pc,
2507 enum charge_type ctype)
2509 lock_page_cgroup(pc);
2510 if (unlikely(PageCgroupUsed(pc))) {
2511 unlock_page_cgroup(pc);
2512 __mem_cgroup_cancel_charge(memcg, nr_pages);
2516 * we don't need page_cgroup_lock about tail pages, becase they are not
2517 * accessed by any other context at this point.
2519 pc->mem_cgroup = memcg;
2521 * We access a page_cgroup asynchronously without lock_page_cgroup().
2522 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2523 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2524 * before USED bit, we need memory barrier here.
2525 * See mem_cgroup_add_lru_list(), etc.
2529 case MEM_CGROUP_CHARGE_TYPE_CACHE:
2530 case MEM_CGROUP_CHARGE_TYPE_SHMEM:
2531 SetPageCgroupCache(pc);
2532 SetPageCgroupUsed(pc);
2534 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2535 ClearPageCgroupCache(pc);
2536 SetPageCgroupUsed(pc);
2542 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), nr_pages);
2543 unlock_page_cgroup(pc);
2545 * "charge_statistics" updated event counter. Then, check it.
2546 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2547 * if they exceeds softlimit.
2549 memcg_check_events(memcg, page);
2552 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2554 #define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MOVE_LOCK) |\
2555 (1 << PCG_ACCT_LRU) | (1 << PCG_MIGRATION))
2557 * Because tail pages are not marked as "used", set it. We're under
2558 * zone->lru_lock, 'splitting on pmd' and compound_lock.
2559 * charge/uncharge will be never happen and move_account() is done under
2560 * compound_lock(), so we don't have to take care of races.
2562 void mem_cgroup_split_huge_fixup(struct page *head)
2564 struct page_cgroup *head_pc = lookup_page_cgroup(head);
2565 struct page_cgroup *pc;
2568 if (mem_cgroup_disabled())
2570 for (i = 1; i < HPAGE_PMD_NR; i++) {
2572 pc->mem_cgroup = head_pc->mem_cgroup;
2573 smp_wmb();/* see __commit_charge() */
2575 * LRU flags cannot be copied because we need to add tail
2576 * page to LRU by generic call and our hooks will be called.
2578 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2581 if (PageCgroupAcctLRU(head_pc)) {
2583 struct mem_cgroup_per_zone *mz;
2585 * We hold lru_lock, then, reduce counter directly.
2587 lru = page_lru(head);
2588 mz = page_cgroup_zoneinfo(head_pc->mem_cgroup, head);
2589 MEM_CGROUP_ZSTAT(mz, lru) -= HPAGE_PMD_NR - 1;
2595 * mem_cgroup_move_account - move account of the page
2597 * @nr_pages: number of regular pages (>1 for huge pages)
2598 * @pc: page_cgroup of the page.
2599 * @from: mem_cgroup which the page is moved from.
2600 * @to: mem_cgroup which the page is moved to. @from != @to.
2601 * @uncharge: whether we should call uncharge and css_put against @from.
2603 * The caller must confirm following.
2604 * - page is not on LRU (isolate_page() is useful.)
2605 * - compound_lock is held when nr_pages > 1
2607 * This function doesn't do "charge" nor css_get to new cgroup. It should be
2608 * done by a caller(__mem_cgroup_try_charge would be useful). If @uncharge is
2609 * true, this function does "uncharge" from old cgroup, but it doesn't if
2610 * @uncharge is false, so a caller should do "uncharge".
2612 static int mem_cgroup_move_account(struct page *page,
2613 unsigned int nr_pages,
2614 struct page_cgroup *pc,
2615 struct mem_cgroup *from,
2616 struct mem_cgroup *to,
2619 unsigned long flags;
2622 VM_BUG_ON(from == to);
2623 VM_BUG_ON(PageLRU(page));
2625 * The page is isolated from LRU. So, collapse function
2626 * will not handle this page. But page splitting can happen.
2627 * Do this check under compound_page_lock(). The caller should
2631 if (nr_pages > 1 && !PageTransHuge(page))
2634 lock_page_cgroup(pc);
2637 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
2640 move_lock_page_cgroup(pc, &flags);
2642 if (PageCgroupFileMapped(pc)) {
2643 /* Update mapped_file data for mem_cgroup */
2645 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2646 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2649 mem_cgroup_charge_statistics(from, PageCgroupCache(pc), -nr_pages);
2651 /* This is not "cancel", but cancel_charge does all we need. */
2652 __mem_cgroup_cancel_charge(from, nr_pages);
2654 /* caller should have done css_get */
2655 pc->mem_cgroup = to;
2656 mem_cgroup_charge_statistics(to, PageCgroupCache(pc), nr_pages);
2658 * We charges against "to" which may not have any tasks. Then, "to"
2659 * can be under rmdir(). But in current implementation, caller of
2660 * this function is just force_empty() and move charge, so it's
2661 * guaranteed that "to" is never removed. So, we don't check rmdir
2664 move_unlock_page_cgroup(pc, &flags);
2667 unlock_page_cgroup(pc);
2671 memcg_check_events(to, page);
2672 memcg_check_events(from, page);
2678 * move charges to its parent.
2681 static int mem_cgroup_move_parent(struct page *page,
2682 struct page_cgroup *pc,
2683 struct mem_cgroup *child,
2686 struct cgroup *cg = child->css.cgroup;
2687 struct cgroup *pcg = cg->parent;
2688 struct mem_cgroup *parent;
2689 unsigned int nr_pages;
2690 unsigned long uninitialized_var(flags);
2698 if (!get_page_unless_zero(page))
2700 if (isolate_lru_page(page))
2703 nr_pages = hpage_nr_pages(page);
2705 parent = mem_cgroup_from_cont(pcg);
2706 ret = __mem_cgroup_try_charge(NULL, gfp_mask, nr_pages, &parent, false);
2711 flags = compound_lock_irqsave(page);
2713 ret = mem_cgroup_move_account(page, nr_pages, pc, child, parent, true);
2715 __mem_cgroup_cancel_charge(parent, nr_pages);
2718 compound_unlock_irqrestore(page, flags);
2720 putback_lru_page(page);
2728 * Charge the memory controller for page usage.
2730 * 0 if the charge was successful
2731 * < 0 if the cgroup is over its limit
2733 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2734 gfp_t gfp_mask, enum charge_type ctype)
2736 struct mem_cgroup *memcg = NULL;
2737 unsigned int nr_pages = 1;
2738 struct page_cgroup *pc;
2742 if (PageTransHuge(page)) {
2743 nr_pages <<= compound_order(page);
2744 VM_BUG_ON(!PageTransHuge(page));
2746 * Never OOM-kill a process for a huge page. The
2747 * fault handler will fall back to regular pages.
2752 pc = lookup_page_cgroup(page);
2753 BUG_ON(!pc); /* XXX: remove this and move pc lookup into commit */
2755 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
2759 __mem_cgroup_commit_charge(memcg, page, nr_pages, pc, ctype);
2763 int mem_cgroup_newpage_charge(struct page *page,
2764 struct mm_struct *mm, gfp_t gfp_mask)
2766 if (mem_cgroup_disabled())
2769 * If already mapped, we don't have to account.
2770 * If page cache, page->mapping has address_space.
2771 * But page->mapping may have out-of-use anon_vma pointer,
2772 * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
2775 if (page_mapped(page) || (page->mapping && !PageAnon(page)))
2779 return mem_cgroup_charge_common(page, mm, gfp_mask,
2780 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2784 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2785 enum charge_type ctype);
2788 __mem_cgroup_commit_charge_lrucare(struct page *page, struct mem_cgroup *memcg,
2789 enum charge_type ctype)
2791 struct page_cgroup *pc = lookup_page_cgroup(page);
2793 * In some case, SwapCache, FUSE(splice_buf->radixtree), the page
2794 * is already on LRU. It means the page may on some other page_cgroup's
2795 * LRU. Take care of it.
2797 mem_cgroup_lru_del_before_commit(page);
2798 __mem_cgroup_commit_charge(memcg, page, 1, pc, ctype);
2799 mem_cgroup_lru_add_after_commit(page);
2803 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2806 struct mem_cgroup *memcg = NULL;
2809 if (mem_cgroup_disabled())
2811 if (PageCompound(page))
2817 if (page_is_file_cache(page)) {
2818 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, &memcg, true);
2823 * FUSE reuses pages without going through the final
2824 * put that would remove them from the LRU list, make
2825 * sure that they get relinked properly.
2827 __mem_cgroup_commit_charge_lrucare(page, memcg,
2828 MEM_CGROUP_CHARGE_TYPE_CACHE);
2832 if (PageSwapCache(page)) {
2833 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &memcg);
2835 __mem_cgroup_commit_charge_swapin(page, memcg,
2836 MEM_CGROUP_CHARGE_TYPE_SHMEM);
2838 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
2839 MEM_CGROUP_CHARGE_TYPE_SHMEM);
2845 * While swap-in, try_charge -> commit or cancel, the page is locked.
2846 * And when try_charge() successfully returns, one refcnt to memcg without
2847 * struct page_cgroup is acquired. This refcnt will be consumed by
2848 * "commit()" or removed by "cancel()"
2850 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2852 gfp_t mask, struct mem_cgroup **ptr)
2854 struct mem_cgroup *memcg;
2859 if (mem_cgroup_disabled())
2862 if (!do_swap_account)
2865 * A racing thread's fault, or swapoff, may have already updated
2866 * the pte, and even removed page from swap cache: in those cases
2867 * do_swap_page()'s pte_same() test will fail; but there's also a
2868 * KSM case which does need to charge the page.
2870 if (!PageSwapCache(page))
2872 memcg = try_get_mem_cgroup_from_page(page);
2876 ret = __mem_cgroup_try_charge(NULL, mask, 1, ptr, true);
2877 css_put(&memcg->css);
2882 return __mem_cgroup_try_charge(mm, mask, 1, ptr, true);
2886 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2887 enum charge_type ctype)
2889 if (mem_cgroup_disabled())
2893 cgroup_exclude_rmdir(&ptr->css);
2895 __mem_cgroup_commit_charge_lrucare(page, ptr, ctype);
2897 * Now swap is on-memory. This means this page may be
2898 * counted both as mem and swap....double count.
2899 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2900 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2901 * may call delete_from_swap_cache() before reach here.
2903 if (do_swap_account && PageSwapCache(page)) {
2904 swp_entry_t ent = {.val = page_private(page)};
2906 struct mem_cgroup *memcg;
2908 id = swap_cgroup_record(ent, 0);
2910 memcg = mem_cgroup_lookup(id);
2913 * This recorded memcg can be obsolete one. So, avoid
2914 * calling css_tryget
2916 if (!mem_cgroup_is_root(memcg))
2917 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2918 mem_cgroup_swap_statistics(memcg, false);
2919 mem_cgroup_put(memcg);
2924 * At swapin, we may charge account against cgroup which has no tasks.
2925 * So, rmdir()->pre_destroy() can be called while we do this charge.
2926 * In that case, we need to call pre_destroy() again. check it here.
2928 cgroup_release_and_wakeup_rmdir(&ptr->css);
2931 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
2933 __mem_cgroup_commit_charge_swapin(page, ptr,
2934 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2937 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
2939 if (mem_cgroup_disabled())
2943 __mem_cgroup_cancel_charge(memcg, 1);
2946 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
2947 unsigned int nr_pages,
2948 const enum charge_type ctype)
2950 struct memcg_batch_info *batch = NULL;
2951 bool uncharge_memsw = true;
2953 /* If swapout, usage of swap doesn't decrease */
2954 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2955 uncharge_memsw = false;
2957 batch = ¤t->memcg_batch;
2959 * In usual, we do css_get() when we remember memcg pointer.
2960 * But in this case, we keep res->usage until end of a series of
2961 * uncharges. Then, it's ok to ignore memcg's refcnt.
2964 batch->memcg = memcg;
2966 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2967 * In those cases, all pages freed continuously can be expected to be in
2968 * the same cgroup and we have chance to coalesce uncharges.
2969 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2970 * because we want to do uncharge as soon as possible.
2973 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
2974 goto direct_uncharge;
2977 goto direct_uncharge;
2980 * In typical case, batch->memcg == mem. This means we can
2981 * merge a series of uncharges to an uncharge of res_counter.
2982 * If not, we uncharge res_counter ony by one.
2984 if (batch->memcg != memcg)
2985 goto direct_uncharge;
2986 /* remember freed charge and uncharge it later */
2989 batch->memsw_nr_pages++;
2992 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
2994 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
2995 if (unlikely(batch->memcg != memcg))
2996 memcg_oom_recover(memcg);
3001 * uncharge if !page_mapped(page)
3003 static struct mem_cgroup *
3004 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
3006 struct mem_cgroup *memcg = NULL;
3007 unsigned int nr_pages = 1;
3008 struct page_cgroup *pc;
3010 if (mem_cgroup_disabled())
3013 if (PageSwapCache(page))
3016 if (PageTransHuge(page)) {
3017 nr_pages <<= compound_order(page);
3018 VM_BUG_ON(!PageTransHuge(page));
3021 * Check if our page_cgroup is valid
3023 pc = lookup_page_cgroup(page);
3024 if (unlikely(!pc || !PageCgroupUsed(pc)))
3027 lock_page_cgroup(pc);
3029 memcg = pc->mem_cgroup;
3031 if (!PageCgroupUsed(pc))
3035 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
3036 case MEM_CGROUP_CHARGE_TYPE_DROP:
3037 /* See mem_cgroup_prepare_migration() */
3038 if (page_mapped(page) || PageCgroupMigration(pc))
3041 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
3042 if (!PageAnon(page)) { /* Shared memory */
3043 if (page->mapping && !page_is_file_cache(page))
3045 } else if (page_mapped(page)) /* Anon */
3052 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), -nr_pages);
3054 ClearPageCgroupUsed(pc);
3056 * pc->mem_cgroup is not cleared here. It will be accessed when it's
3057 * freed from LRU. This is safe because uncharged page is expected not
3058 * to be reused (freed soon). Exception is SwapCache, it's handled by
3059 * special functions.
3062 unlock_page_cgroup(pc);
3064 * even after unlock, we have memcg->res.usage here and this memcg
3065 * will never be freed.
3067 memcg_check_events(memcg, page);
3068 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
3069 mem_cgroup_swap_statistics(memcg, true);
3070 mem_cgroup_get(memcg);
3072 if (!mem_cgroup_is_root(memcg))
3073 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
3078 unlock_page_cgroup(pc);
3082 void mem_cgroup_uncharge_page(struct page *page)
3085 if (page_mapped(page))
3087 if (page->mapping && !PageAnon(page))
3089 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
3092 void mem_cgroup_uncharge_cache_page(struct page *page)
3094 VM_BUG_ON(page_mapped(page));
3095 VM_BUG_ON(page->mapping);
3096 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
3100 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
3101 * In that cases, pages are freed continuously and we can expect pages
3102 * are in the same memcg. All these calls itself limits the number of
3103 * pages freed at once, then uncharge_start/end() is called properly.
3104 * This may be called prural(2) times in a context,
3107 void mem_cgroup_uncharge_start(void)
3109 current->memcg_batch.do_batch++;
3110 /* We can do nest. */
3111 if (current->memcg_batch.do_batch == 1) {
3112 current->memcg_batch.memcg = NULL;
3113 current->memcg_batch.nr_pages = 0;
3114 current->memcg_batch.memsw_nr_pages = 0;
3118 void mem_cgroup_uncharge_end(void)
3120 struct memcg_batch_info *batch = ¤t->memcg_batch;
3122 if (!batch->do_batch)
3126 if (batch->do_batch) /* If stacked, do nothing. */
3132 * This "batch->memcg" is valid without any css_get/put etc...
3133 * bacause we hide charges behind us.
3135 if (batch->nr_pages)
3136 res_counter_uncharge(&batch->memcg->res,
3137 batch->nr_pages * PAGE_SIZE);
3138 if (batch->memsw_nr_pages)
3139 res_counter_uncharge(&batch->memcg->memsw,
3140 batch->memsw_nr_pages * PAGE_SIZE);
3141 memcg_oom_recover(batch->memcg);
3142 /* forget this pointer (for sanity check) */
3143 batch->memcg = NULL;
3148 * called after __delete_from_swap_cache() and drop "page" account.
3149 * memcg information is recorded to swap_cgroup of "ent"
3152 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
3154 struct mem_cgroup *memcg;
3155 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
3157 if (!swapout) /* this was a swap cache but the swap is unused ! */
3158 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
3160 memcg = __mem_cgroup_uncharge_common(page, ctype);
3163 * record memcg information, if swapout && memcg != NULL,
3164 * mem_cgroup_get() was called in uncharge().
3166 if (do_swap_account && swapout && memcg)
3167 swap_cgroup_record(ent, css_id(&memcg->css));
3171 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3173 * called from swap_entry_free(). remove record in swap_cgroup and
3174 * uncharge "memsw" account.
3176 void mem_cgroup_uncharge_swap(swp_entry_t ent)
3178 struct mem_cgroup *memcg;
3181 if (!do_swap_account)
3184 id = swap_cgroup_record(ent, 0);
3186 memcg = mem_cgroup_lookup(id);
3189 * We uncharge this because swap is freed.
3190 * This memcg can be obsolete one. We avoid calling css_tryget
3192 if (!mem_cgroup_is_root(memcg))
3193 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
3194 mem_cgroup_swap_statistics(memcg, false);
3195 mem_cgroup_put(memcg);
3201 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3202 * @entry: swap entry to be moved
3203 * @from: mem_cgroup which the entry is moved from
3204 * @to: mem_cgroup which the entry is moved to
3205 * @need_fixup: whether we should fixup res_counters and refcounts.
3207 * It succeeds only when the swap_cgroup's record for this entry is the same
3208 * as the mem_cgroup's id of @from.
3210 * Returns 0 on success, -EINVAL on failure.
3212 * The caller must have charged to @to, IOW, called res_counter_charge() about
3213 * both res and memsw, and called css_get().
3215 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3216 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3218 unsigned short old_id, new_id;
3220 old_id = css_id(&from->css);
3221 new_id = css_id(&to->css);
3223 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3224 mem_cgroup_swap_statistics(from, false);
3225 mem_cgroup_swap_statistics(to, true);
3227 * This function is only called from task migration context now.
3228 * It postpones res_counter and refcount handling till the end
3229 * of task migration(mem_cgroup_clear_mc()) for performance
3230 * improvement. But we cannot postpone mem_cgroup_get(to)
3231 * because if the process that has been moved to @to does
3232 * swap-in, the refcount of @to might be decreased to 0.
3236 if (!mem_cgroup_is_root(from))
3237 res_counter_uncharge(&from->memsw, PAGE_SIZE);
3238 mem_cgroup_put(from);
3240 * we charged both to->res and to->memsw, so we should
3243 if (!mem_cgroup_is_root(to))
3244 res_counter_uncharge(&to->res, PAGE_SIZE);
3251 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3252 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3259 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
3262 int mem_cgroup_prepare_migration(struct page *page,
3263 struct page *newpage, struct mem_cgroup **ptr, gfp_t gfp_mask)
3265 struct mem_cgroup *memcg = NULL;
3266 struct page_cgroup *pc;
3267 enum charge_type ctype;
3272 VM_BUG_ON(PageTransHuge(page));
3273 if (mem_cgroup_disabled())
3276 pc = lookup_page_cgroup(page);
3277 lock_page_cgroup(pc);
3278 if (PageCgroupUsed(pc)) {
3279 memcg = pc->mem_cgroup;
3280 css_get(&memcg->css);
3282 * At migrating an anonymous page, its mapcount goes down
3283 * to 0 and uncharge() will be called. But, even if it's fully
3284 * unmapped, migration may fail and this page has to be
3285 * charged again. We set MIGRATION flag here and delay uncharge
3286 * until end_migration() is called
3288 * Corner Case Thinking
3290 * When the old page was mapped as Anon and it's unmap-and-freed
3291 * while migration was ongoing.
3292 * If unmap finds the old page, uncharge() of it will be delayed
3293 * until end_migration(). If unmap finds a new page, it's
3294 * uncharged when it make mapcount to be 1->0. If unmap code
3295 * finds swap_migration_entry, the new page will not be mapped
3296 * and end_migration() will find it(mapcount==0).
3299 * When the old page was mapped but migraion fails, the kernel
3300 * remaps it. A charge for it is kept by MIGRATION flag even
3301 * if mapcount goes down to 0. We can do remap successfully
3302 * without charging it again.
3305 * The "old" page is under lock_page() until the end of
3306 * migration, so, the old page itself will not be swapped-out.
3307 * If the new page is swapped out before end_migraton, our
3308 * hook to usual swap-out path will catch the event.
3311 SetPageCgroupMigration(pc);
3313 unlock_page_cgroup(pc);
3315 * If the page is not charged at this point,
3322 ret = __mem_cgroup_try_charge(NULL, gfp_mask, 1, ptr, false);
3323 css_put(&memcg->css);/* drop extra refcnt */
3324 if (ret || *ptr == NULL) {
3325 if (PageAnon(page)) {
3326 lock_page_cgroup(pc);
3327 ClearPageCgroupMigration(pc);
3328 unlock_page_cgroup(pc);
3330 * The old page may be fully unmapped while we kept it.
3332 mem_cgroup_uncharge_page(page);
3337 * We charge new page before it's used/mapped. So, even if unlock_page()
3338 * is called before end_migration, we can catch all events on this new
3339 * page. In the case new page is migrated but not remapped, new page's
3340 * mapcount will be finally 0 and we call uncharge in end_migration().
3342 pc = lookup_page_cgroup(newpage);
3344 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
3345 else if (page_is_file_cache(page))
3346 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
3348 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3349 __mem_cgroup_commit_charge(memcg, page, 1, pc, ctype);
3353 /* remove redundant charge if migration failed*/
3354 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
3355 struct page *oldpage, struct page *newpage, bool migration_ok)
3357 struct page *used, *unused;
3358 struct page_cgroup *pc;
3362 /* blocks rmdir() */
3363 cgroup_exclude_rmdir(&memcg->css);
3364 if (!migration_ok) {
3372 * We disallowed uncharge of pages under migration because mapcount
3373 * of the page goes down to zero, temporarly.
3374 * Clear the flag and check the page should be charged.
3376 pc = lookup_page_cgroup(oldpage);
3377 lock_page_cgroup(pc);
3378 ClearPageCgroupMigration(pc);
3379 unlock_page_cgroup(pc);
3381 __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE);
3384 * If a page is a file cache, radix-tree replacement is very atomic
3385 * and we can skip this check. When it was an Anon page, its mapcount
3386 * goes down to 0. But because we added MIGRATION flage, it's not
3387 * uncharged yet. There are several case but page->mapcount check
3388 * and USED bit check in mem_cgroup_uncharge_page() will do enough
3389 * check. (see prepare_charge() also)
3392 mem_cgroup_uncharge_page(used);
3394 * At migration, we may charge account against cgroup which has no
3396 * So, rmdir()->pre_destroy() can be called while we do this charge.
3397 * In that case, we need to call pre_destroy() again. check it here.
3399 cgroup_release_and_wakeup_rmdir(&memcg->css);
3403 * At replace page cache, newpage is not under any memcg but it's on
3404 * LRU. So, this function doesn't touch res_counter but handles LRU
3405 * in correct way. Both pages are locked so we cannot race with uncharge.
3407 void mem_cgroup_replace_page_cache(struct page *oldpage,
3408 struct page *newpage)
3410 struct mem_cgroup *memcg;
3411 struct page_cgroup *pc;
3413 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
3414 unsigned long flags;
3416 if (mem_cgroup_disabled())
3419 pc = lookup_page_cgroup(oldpage);
3420 /* fix accounting on old pages */
3421 lock_page_cgroup(pc);
3422 memcg = pc->mem_cgroup;
3423 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), -1);
3424 ClearPageCgroupUsed(pc);
3425 unlock_page_cgroup(pc);
3427 if (PageSwapBacked(oldpage))
3428 type = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3430 zone = page_zone(newpage);
3431 pc = lookup_page_cgroup(newpage);
3433 * Even if newpage->mapping was NULL before starting replacement,
3434 * the newpage may be on LRU(or pagevec for LRU) already. We lock
3435 * LRU while we overwrite pc->mem_cgroup.
3437 spin_lock_irqsave(&zone->lru_lock, flags);
3438 if (PageLRU(newpage))
3439 del_page_from_lru_list(zone, newpage, page_lru(newpage));
3440 __mem_cgroup_commit_charge(memcg, newpage, 1, pc, type);
3441 if (PageLRU(newpage))
3442 add_page_to_lru_list(zone, newpage, page_lru(newpage));
3443 spin_unlock_irqrestore(&zone->lru_lock, flags);
3446 #ifdef CONFIG_DEBUG_VM
3447 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3449 struct page_cgroup *pc;
3451 pc = lookup_page_cgroup(page);
3452 if (likely(pc) && PageCgroupUsed(pc))
3457 bool mem_cgroup_bad_page_check(struct page *page)
3459 if (mem_cgroup_disabled())
3462 return lookup_page_cgroup_used(page) != NULL;
3465 void mem_cgroup_print_bad_page(struct page *page)
3467 struct page_cgroup *pc;
3469 pc = lookup_page_cgroup_used(page);
3474 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p",
3475 pc, pc->flags, pc->mem_cgroup);
3477 path = kmalloc(PATH_MAX, GFP_KERNEL);
3480 ret = cgroup_path(pc->mem_cgroup->css.cgroup,
3485 printk(KERN_CONT "(%s)\n",
3486 (ret < 0) ? "cannot get the path" : path);
3492 static DEFINE_MUTEX(set_limit_mutex);
3494 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3495 unsigned long long val)
3498 u64 memswlimit, memlimit;
3500 int children = mem_cgroup_count_children(memcg);
3501 u64 curusage, oldusage;
3505 * For keeping hierarchical_reclaim simple, how long we should retry
3506 * is depends on callers. We set our retry-count to be function
3507 * of # of children which we should visit in this loop.
3509 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3511 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3514 while (retry_count) {
3515 if (signal_pending(current)) {
3520 * Rather than hide all in some function, I do this in
3521 * open coded manner. You see what this really does.
3522 * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3524 mutex_lock(&set_limit_mutex);
3525 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3526 if (memswlimit < val) {
3528 mutex_unlock(&set_limit_mutex);
3532 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3536 ret = res_counter_set_limit(&memcg->res, val);
3538 if (memswlimit == val)
3539 memcg->memsw_is_minimum = true;
3541 memcg->memsw_is_minimum = false;
3543 mutex_unlock(&set_limit_mutex);
3548 mem_cgroup_reclaim(memcg, GFP_KERNEL,
3549 MEM_CGROUP_RECLAIM_SHRINK);
3550 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3551 /* Usage is reduced ? */
3552 if (curusage >= oldusage)
3555 oldusage = curusage;
3557 if (!ret && enlarge)
3558 memcg_oom_recover(memcg);
3563 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3564 unsigned long long val)
3567 u64 memlimit, memswlimit, oldusage, curusage;
3568 int children = mem_cgroup_count_children(memcg);
3572 /* see mem_cgroup_resize_res_limit */
3573 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3574 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3575 while (retry_count) {
3576 if (signal_pending(current)) {
3581 * Rather than hide all in some function, I do this in
3582 * open coded manner. You see what this really does.
3583 * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3585 mutex_lock(&set_limit_mutex);
3586 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3587 if (memlimit > val) {
3589 mutex_unlock(&set_limit_mutex);
3592 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3593 if (memswlimit < val)
3595 ret = res_counter_set_limit(&memcg->memsw, val);
3597 if (memlimit == val)
3598 memcg->memsw_is_minimum = true;
3600 memcg->memsw_is_minimum = false;
3602 mutex_unlock(&set_limit_mutex);
3607 mem_cgroup_reclaim(memcg, GFP_KERNEL,
3608 MEM_CGROUP_RECLAIM_NOSWAP |
3609 MEM_CGROUP_RECLAIM_SHRINK);
3610 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3611 /* Usage is reduced ? */
3612 if (curusage >= oldusage)
3615 oldusage = curusage;
3617 if (!ret && enlarge)
3618 memcg_oom_recover(memcg);
3622 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3624 unsigned long *total_scanned)
3626 unsigned long nr_reclaimed = 0;
3627 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3628 unsigned long reclaimed;
3630 struct mem_cgroup_tree_per_zone *mctz;
3631 unsigned long long excess;
3632 unsigned long nr_scanned;
3637 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3639 * This loop can run a while, specially if mem_cgroup's continuously
3640 * keep exceeding their soft limit and putting the system under
3647 mz = mem_cgroup_largest_soft_limit_node(mctz);
3652 reclaimed = mem_cgroup_soft_reclaim(mz->mem, zone,
3653 gfp_mask, &nr_scanned);
3654 nr_reclaimed += reclaimed;
3655 *total_scanned += nr_scanned;
3656 spin_lock(&mctz->lock);
3659 * If we failed to reclaim anything from this memory cgroup
3660 * it is time to move on to the next cgroup
3666 * Loop until we find yet another one.
3668 * By the time we get the soft_limit lock
3669 * again, someone might have aded the
3670 * group back on the RB tree. Iterate to
3671 * make sure we get a different mem.
3672 * mem_cgroup_largest_soft_limit_node returns
3673 * NULL if no other cgroup is present on
3677 __mem_cgroup_largest_soft_limit_node(mctz);
3679 css_put(&next_mz->mem->css);
3680 else /* next_mz == NULL or other memcg */
3684 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
3685 excess = res_counter_soft_limit_excess(&mz->mem->res);
3687 * One school of thought says that we should not add
3688 * back the node to the tree if reclaim returns 0.
3689 * But our reclaim could return 0, simply because due
3690 * to priority we are exposing a smaller subset of
3691 * memory to reclaim from. Consider this as a longer
3694 /* If excess == 0, no tree ops */
3695 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess);
3696 spin_unlock(&mctz->lock);
3697 css_put(&mz->mem->css);
3700 * Could not reclaim anything and there are no more
3701 * mem cgroups to try or we seem to be looping without
3702 * reclaiming anything.
3704 if (!nr_reclaimed &&
3706 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3708 } while (!nr_reclaimed);
3710 css_put(&next_mz->mem->css);
3711 return nr_reclaimed;
3715 * This routine traverse page_cgroup in given list and drop them all.
3716 * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3718 static int mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
3719 int node, int zid, enum lru_list lru)
3721 struct mem_cgroup_per_zone *mz;
3722 unsigned long flags, loop;
3723 struct list_head *list;
3728 zone = &NODE_DATA(node)->node_zones[zid];
3729 mz = mem_cgroup_zoneinfo(memcg, node, zid);
3730 list = &mz->lruvec.lists[lru];
3732 loop = MEM_CGROUP_ZSTAT(mz, lru);
3733 /* give some margin against EBUSY etc...*/
3737 struct page_cgroup *pc;
3741 spin_lock_irqsave(&zone->lru_lock, flags);
3742 if (list_empty(list)) {
3743 spin_unlock_irqrestore(&zone->lru_lock, flags);
3746 page = list_entry(list->prev, struct page, lru);
3748 list_move(&page->lru, list);
3750 spin_unlock_irqrestore(&zone->lru_lock, flags);
3753 spin_unlock_irqrestore(&zone->lru_lock, flags);
3755 pc = lookup_page_cgroup(page);
3757 ret = mem_cgroup_move_parent(page, pc, memcg, GFP_KERNEL);
3761 if (ret == -EBUSY || ret == -EINVAL) {
3762 /* found lock contention or "pc" is obsolete. */
3769 if (!ret && !list_empty(list))
3775 * make mem_cgroup's charge to be 0 if there is no task.
3776 * This enables deleting this mem_cgroup.
3778 static int mem_cgroup_force_empty(struct mem_cgroup *memcg, bool free_all)
3781 int node, zid, shrink;
3782 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3783 struct cgroup *cgrp = memcg->css.cgroup;
3785 css_get(&memcg->css);
3788 /* should free all ? */
3794 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3797 if (signal_pending(current))
3799 /* This is for making all *used* pages to be on LRU. */
3800 lru_add_drain_all();
3801 drain_all_stock_sync(memcg);
3803 mem_cgroup_start_move(memcg);
3804 for_each_node_state(node, N_HIGH_MEMORY) {
3805 for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3808 ret = mem_cgroup_force_empty_list(memcg,
3817 mem_cgroup_end_move(memcg);
3818 memcg_oom_recover(memcg);
3819 /* it seems parent cgroup doesn't have enough mem */
3823 /* "ret" should also be checked to ensure all lists are empty. */
3824 } while (memcg->res.usage > 0 || ret);
3826 css_put(&memcg->css);
3830 /* returns EBUSY if there is a task or if we come here twice. */
3831 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3835 /* we call try-to-free pages for make this cgroup empty */
3836 lru_add_drain_all();
3837 /* try to free all pages in this cgroup */
3839 while (nr_retries && memcg->res.usage > 0) {
3842 if (signal_pending(current)) {
3846 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
3850 /* maybe some writeback is necessary */
3851 congestion_wait(BLK_RW_ASYNC, HZ/10);
3856 /* try move_account...there may be some *locked* pages. */
3860 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3862 return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3866 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3868 return mem_cgroup_from_cont(cont)->use_hierarchy;
3871 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3875 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3876 struct cgroup *parent = cont->parent;
3877 struct mem_cgroup *parent_memcg = NULL;
3880 parent_memcg = mem_cgroup_from_cont(parent);
3884 * If parent's use_hierarchy is set, we can't make any modifications
3885 * in the child subtrees. If it is unset, then the change can
3886 * occur, provided the current cgroup has no children.
3888 * For the root cgroup, parent_mem is NULL, we allow value to be
3889 * set if there are no children.
3891 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3892 (val == 1 || val == 0)) {
3893 if (list_empty(&cont->children))
3894 memcg->use_hierarchy = val;
3905 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
3906 enum mem_cgroup_stat_index idx)
3908 struct mem_cgroup *iter;
3911 /* Per-cpu values can be negative, use a signed accumulator */
3912 for_each_mem_cgroup_tree(iter, memcg)
3913 val += mem_cgroup_read_stat(iter, idx);
3915 if (val < 0) /* race ? */
3920 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3924 if (!mem_cgroup_is_root(memcg)) {
3926 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_KMEM
3927 if (!memcg->kmem_independent_accounting)
3928 val = res_counter_read_u64(&memcg->kmem, RES_USAGE);
3931 val += res_counter_read_u64(&memcg->res, RES_USAGE);
3933 val += res_counter_read_u64(&memcg->memsw, RES_USAGE);
3938 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
3939 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
3942 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
3944 return val << PAGE_SHIFT;
3947 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
3949 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3953 type = MEMFILE_TYPE(cft->private);
3954 name = MEMFILE_ATTR(cft->private);
3957 if (name == RES_USAGE)
3958 val = mem_cgroup_usage(memcg, false);
3960 val = res_counter_read_u64(&memcg->res, name);
3963 if (name == RES_USAGE)
3964 val = mem_cgroup_usage(memcg, true);
3966 val = res_counter_read_u64(&memcg->memsw, name);
3968 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_KMEM
3970 val = res_counter_read_u64(&memcg->kmem, name);
3980 * The user of this function is...
3983 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
3986 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3988 unsigned long long val;
3991 type = MEMFILE_TYPE(cft->private);
3992 name = MEMFILE_ATTR(cft->private);
3995 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3999 /* This function does all necessary parse...reuse it */
4000 ret = res_counter_memparse_write_strategy(buffer, &val);
4004 ret = mem_cgroup_resize_limit(memcg, val);
4006 ret = mem_cgroup_resize_memsw_limit(memcg, val);
4008 case RES_SOFT_LIMIT:
4009 ret = res_counter_memparse_write_strategy(buffer, &val);
4013 * For memsw, soft limits are hard to implement in terms
4014 * of semantics, for now, we support soft limits for
4015 * control without swap
4018 ret = res_counter_set_soft_limit(&memcg->res, val);
4023 ret = -EINVAL; /* should be BUG() ? */
4029 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
4030 unsigned long long *mem_limit, unsigned long long *memsw_limit)
4032 struct cgroup *cgroup;
4033 unsigned long long min_limit, min_memsw_limit, tmp;
4035 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4036 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4037 cgroup = memcg->css.cgroup;
4038 if (!memcg->use_hierarchy)
4041 while (cgroup->parent) {
4042 cgroup = cgroup->parent;
4043 memcg = mem_cgroup_from_cont(cgroup);
4044 if (!memcg->use_hierarchy)
4046 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
4047 min_limit = min(min_limit, tmp);
4048 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4049 min_memsw_limit = min(min_memsw_limit, tmp);
4052 *mem_limit = min_limit;
4053 *memsw_limit = min_memsw_limit;
4057 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
4059 struct mem_cgroup *memcg;
4062 memcg = mem_cgroup_from_cont(cont);
4063 type = MEMFILE_TYPE(event);
4064 name = MEMFILE_ATTR(event);
4068 res_counter_reset_max(&memcg->res);
4070 res_counter_reset_max(&memcg->memsw);
4074 res_counter_reset_failcnt(&memcg->res);
4076 res_counter_reset_failcnt(&memcg->memsw);
4083 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
4086 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
4090 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4091 struct cftype *cft, u64 val)
4093 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4095 if (val >= (1 << NR_MOVE_TYPE))
4098 * We check this value several times in both in can_attach() and
4099 * attach(), so we need cgroup lock to prevent this value from being
4103 memcg->move_charge_at_immigrate = val;
4109 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4110 struct cftype *cft, u64 val)
4117 /* For read statistics */
4135 struct mcs_total_stat {
4136 s64 stat[NR_MCS_STAT];
4142 } memcg_stat_strings[NR_MCS_STAT] = {
4143 {"cache", "total_cache"},
4144 {"rss", "total_rss"},
4145 {"mapped_file", "total_mapped_file"},
4146 {"pgpgin", "total_pgpgin"},
4147 {"pgpgout", "total_pgpgout"},
4148 {"swap", "total_swap"},
4149 {"pgfault", "total_pgfault"},
4150 {"pgmajfault", "total_pgmajfault"},
4151 {"inactive_anon", "total_inactive_anon"},
4152 {"active_anon", "total_active_anon"},
4153 {"inactive_file", "total_inactive_file"},
4154 {"active_file", "total_active_file"},
4155 {"unevictable", "total_unevictable"}
4160 mem_cgroup_get_local_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4165 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_CACHE);
4166 s->stat[MCS_CACHE] += val * PAGE_SIZE;
4167 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_RSS);
4168 s->stat[MCS_RSS] += val * PAGE_SIZE;
4169 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_FILE_MAPPED);
4170 s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
4171 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGIN);
4172 s->stat[MCS_PGPGIN] += val;
4173 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGOUT);
4174 s->stat[MCS_PGPGOUT] += val;
4175 if (do_swap_account) {
4176 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
4177 s->stat[MCS_SWAP] += val * PAGE_SIZE;
4179 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGFAULT);
4180 s->stat[MCS_PGFAULT] += val;
4181 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGMAJFAULT);
4182 s->stat[MCS_PGMAJFAULT] += val;
4185 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_ANON));
4186 s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
4187 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_ANON));
4188 s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
4189 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_FILE));
4190 s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
4191 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_FILE));
4192 s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
4193 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
4194 s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
4198 mem_cgroup_get_total_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4200 struct mem_cgroup *iter;
4202 for_each_mem_cgroup_tree(iter, memcg)
4203 mem_cgroup_get_local_stat(iter, s);
4207 static int mem_control_numa_stat_show(struct seq_file *m, void *arg)
4210 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
4211 unsigned long node_nr;
4212 struct cgroup *cont = m->private;
4213 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4215 total_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL);
4216 seq_printf(m, "total=%lu", total_nr);
4217 for_each_node_state(nid, N_HIGH_MEMORY) {
4218 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid, LRU_ALL);
4219 seq_printf(m, " N%d=%lu", nid, node_nr);
4223 file_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_FILE);
4224 seq_printf(m, "file=%lu", file_nr);
4225 for_each_node_state(nid, N_HIGH_MEMORY) {
4226 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4228 seq_printf(m, " N%d=%lu", nid, node_nr);
4232 anon_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_ANON);
4233 seq_printf(m, "anon=%lu", anon_nr);
4234 for_each_node_state(nid, N_HIGH_MEMORY) {
4235 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4237 seq_printf(m, " N%d=%lu", nid, node_nr);
4241 unevictable_nr = mem_cgroup_nr_lru_pages(mem_cont, BIT(LRU_UNEVICTABLE));
4242 seq_printf(m, "unevictable=%lu", unevictable_nr);
4243 for_each_node_state(nid, N_HIGH_MEMORY) {
4244 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4245 BIT(LRU_UNEVICTABLE));
4246 seq_printf(m, " N%d=%lu", nid, node_nr);
4251 #endif /* CONFIG_NUMA */
4253 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
4254 struct cgroup_map_cb *cb)
4256 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4257 struct mcs_total_stat mystat;
4260 memset(&mystat, 0, sizeof(mystat));
4261 mem_cgroup_get_local_stat(mem_cont, &mystat);
4264 for (i = 0; i < NR_MCS_STAT; i++) {
4265 if (i == MCS_SWAP && !do_swap_account)
4267 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
4270 /* Hierarchical information */
4272 unsigned long long limit, memsw_limit;
4273 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
4274 cb->fill(cb, "hierarchical_memory_limit", limit);
4275 if (do_swap_account)
4276 cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
4279 memset(&mystat, 0, sizeof(mystat));
4280 mem_cgroup_get_total_stat(mem_cont, &mystat);
4281 for (i = 0; i < NR_MCS_STAT; i++) {
4282 if (i == MCS_SWAP && !do_swap_account)
4284 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
4287 #ifdef CONFIG_DEBUG_VM
4290 struct mem_cgroup_per_zone *mz;
4291 unsigned long recent_rotated[2] = {0, 0};
4292 unsigned long recent_scanned[2] = {0, 0};
4294 for_each_online_node(nid)
4295 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4296 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
4298 recent_rotated[0] +=
4299 mz->reclaim_stat.recent_rotated[0];
4300 recent_rotated[1] +=
4301 mz->reclaim_stat.recent_rotated[1];
4302 recent_scanned[0] +=
4303 mz->reclaim_stat.recent_scanned[0];
4304 recent_scanned[1] +=
4305 mz->reclaim_stat.recent_scanned[1];
4307 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
4308 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
4309 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
4310 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
4317 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
4319 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4321 return mem_cgroup_swappiness(memcg);
4324 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
4327 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4328 struct mem_cgroup *parent;
4333 if (cgrp->parent == NULL)
4336 parent = mem_cgroup_from_cont(cgrp->parent);
4340 /* If under hierarchy, only empty-root can set this value */
4341 if ((parent->use_hierarchy) ||
4342 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4347 memcg->swappiness = val;
4354 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4356 struct mem_cgroup_threshold_ary *t;
4362 t = rcu_dereference(memcg->thresholds.primary);
4364 t = rcu_dereference(memcg->memsw_thresholds.primary);
4369 usage = mem_cgroup_usage(memcg, swap);
4372 * current_threshold points to threshold just below usage.
4373 * If it's not true, a threshold was crossed after last
4374 * call of __mem_cgroup_threshold().
4376 i = t->current_threshold;
4379 * Iterate backward over array of thresholds starting from
4380 * current_threshold and check if a threshold is crossed.
4381 * If none of thresholds below usage is crossed, we read
4382 * only one element of the array here.
4384 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4385 eventfd_signal(t->entries[i].eventfd, 1);
4387 /* i = current_threshold + 1 */
4391 * Iterate forward over array of thresholds starting from
4392 * current_threshold+1 and check if a threshold is crossed.
4393 * If none of thresholds above usage is crossed, we read
4394 * only one element of the array here.
4396 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4397 eventfd_signal(t->entries[i].eventfd, 1);
4399 /* Update current_threshold */
4400 t->current_threshold = i - 1;
4405 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4408 __mem_cgroup_threshold(memcg, false);
4409 if (do_swap_account)
4410 __mem_cgroup_threshold(memcg, true);
4412 memcg = parent_mem_cgroup(memcg);
4416 static int compare_thresholds(const void *a, const void *b)
4418 const struct mem_cgroup_threshold *_a = a;
4419 const struct mem_cgroup_threshold *_b = b;
4421 return _a->threshold - _b->threshold;
4424 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4426 struct mem_cgroup_eventfd_list *ev;
4428 list_for_each_entry(ev, &memcg->oom_notify, list)
4429 eventfd_signal(ev->eventfd, 1);
4433 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4435 struct mem_cgroup *iter;
4437 for_each_mem_cgroup_tree(iter, memcg)
4438 mem_cgroup_oom_notify_cb(iter);
4441 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
4442 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4444 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4445 struct mem_cgroup_thresholds *thresholds;
4446 struct mem_cgroup_threshold_ary *new;
4447 int type = MEMFILE_TYPE(cft->private);
4448 u64 threshold, usage;
4451 ret = res_counter_memparse_write_strategy(args, &threshold);
4455 mutex_lock(&memcg->thresholds_lock);
4458 thresholds = &memcg->thresholds;
4459 else if (type == _MEMSWAP)
4460 thresholds = &memcg->memsw_thresholds;
4464 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4466 /* Check if a threshold crossed before adding a new one */
4467 if (thresholds->primary)
4468 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4470 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4472 /* Allocate memory for new array of thresholds */
4473 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4481 /* Copy thresholds (if any) to new array */
4482 if (thresholds->primary) {
4483 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4484 sizeof(struct mem_cgroup_threshold));
4487 /* Add new threshold */
4488 new->entries[size - 1].eventfd = eventfd;
4489 new->entries[size - 1].threshold = threshold;
4491 /* Sort thresholds. Registering of new threshold isn't time-critical */
4492 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4493 compare_thresholds, NULL);
4495 /* Find current threshold */
4496 new->current_threshold = -1;
4497 for (i = 0; i < size; i++) {
4498 if (new->entries[i].threshold < usage) {
4500 * new->current_threshold will not be used until
4501 * rcu_assign_pointer(), so it's safe to increment
4504 ++new->current_threshold;
4508 /* Free old spare buffer and save old primary buffer as spare */
4509 kfree(thresholds->spare);
4510 thresholds->spare = thresholds->primary;
4512 rcu_assign_pointer(thresholds->primary, new);
4514 /* To be sure that nobody uses thresholds */
4518 mutex_unlock(&memcg->thresholds_lock);
4523 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
4524 struct cftype *cft, struct eventfd_ctx *eventfd)
4526 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4527 struct mem_cgroup_thresholds *thresholds;
4528 struct mem_cgroup_threshold_ary *new;
4529 int type = MEMFILE_TYPE(cft->private);
4533 mutex_lock(&memcg->thresholds_lock);
4535 thresholds = &memcg->thresholds;
4536 else if (type == _MEMSWAP)
4537 thresholds = &memcg->memsw_thresholds;
4542 * Something went wrong if we trying to unregister a threshold
4543 * if we don't have thresholds
4545 BUG_ON(!thresholds);
4547 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4549 /* Check if a threshold crossed before removing */
4550 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4552 /* Calculate new number of threshold */
4554 for (i = 0; i < thresholds->primary->size; i++) {
4555 if (thresholds->primary->entries[i].eventfd != eventfd)
4559 new = thresholds->spare;
4561 /* Set thresholds array to NULL if we don't have thresholds */
4570 /* Copy thresholds and find current threshold */
4571 new->current_threshold = -1;
4572 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4573 if (thresholds->primary->entries[i].eventfd == eventfd)
4576 new->entries[j] = thresholds->primary->entries[i];
4577 if (new->entries[j].threshold < usage) {
4579 * new->current_threshold will not be used
4580 * until rcu_assign_pointer(), so it's safe to increment
4583 ++new->current_threshold;
4589 /* Swap primary and spare array */
4590 thresholds->spare = thresholds->primary;
4591 rcu_assign_pointer(thresholds->primary, new);
4593 /* To be sure that nobody uses thresholds */
4596 mutex_unlock(&memcg->thresholds_lock);
4599 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4600 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4602 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4603 struct mem_cgroup_eventfd_list *event;
4604 int type = MEMFILE_TYPE(cft->private);
4606 BUG_ON(type != _OOM_TYPE);
4607 event = kmalloc(sizeof(*event), GFP_KERNEL);
4611 spin_lock(&memcg_oom_lock);
4613 event->eventfd = eventfd;
4614 list_add(&event->list, &memcg->oom_notify);
4616 /* already in OOM ? */
4617 if (atomic_read(&memcg->under_oom))
4618 eventfd_signal(eventfd, 1);
4619 spin_unlock(&memcg_oom_lock);
4624 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4625 struct cftype *cft, struct eventfd_ctx *eventfd)
4627 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4628 struct mem_cgroup_eventfd_list *ev, *tmp;
4629 int type = MEMFILE_TYPE(cft->private);
4631 BUG_ON(type != _OOM_TYPE);
4633 spin_lock(&memcg_oom_lock);
4635 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4636 if (ev->eventfd == eventfd) {
4637 list_del(&ev->list);
4642 spin_unlock(&memcg_oom_lock);
4645 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4646 struct cftype *cft, struct cgroup_map_cb *cb)
4648 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4650 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
4652 if (atomic_read(&memcg->under_oom))
4653 cb->fill(cb, "under_oom", 1);
4655 cb->fill(cb, "under_oom", 0);
4659 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4660 struct cftype *cft, u64 val)
4662 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4663 struct mem_cgroup *parent;
4665 /* cannot set to root cgroup and only 0 and 1 are allowed */
4666 if (!cgrp->parent || !((val == 0) || (val == 1)))
4669 parent = mem_cgroup_from_cont(cgrp->parent);
4672 /* oom-kill-disable is a flag for subhierarchy. */
4673 if ((parent->use_hierarchy) ||
4674 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4678 memcg->oom_kill_disable = val;
4680 memcg_oom_recover(memcg);
4686 static const struct file_operations mem_control_numa_stat_file_operations = {
4688 .llseek = seq_lseek,
4689 .release = single_release,
4692 static int mem_control_numa_stat_open(struct inode *unused, struct file *file)
4694 struct cgroup *cont = file->f_dentry->d_parent->d_fsdata;
4696 file->f_op = &mem_control_numa_stat_file_operations;
4697 return single_open(file, mem_control_numa_stat_show, cont);
4699 #endif /* CONFIG_NUMA */
4701 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_KMEM
4702 static u64 kmem_limit_independent_read(struct cgroup *cgroup, struct cftype *cft)
4704 return mem_cgroup_from_cont(cgroup)->kmem_independent_accounting;
4707 static int kmem_limit_independent_write(struct cgroup *cgroup, struct cftype *cft,
4710 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup);
4711 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4716 * This follows the same hierarchy restrictions than
4717 * mem_cgroup_hierarchy_write()
4719 if (!parent || !parent->use_hierarchy) {
4720 if (list_empty(&cgroup->children))
4721 memcg->kmem_independent_accounting = val;
4730 static struct cftype kmem_cgroup_files[] = {
4732 .name = "independent_kmem_limit",
4733 .read_u64 = kmem_limit_independent_read,
4734 .write_u64 = kmem_limit_independent_write,
4737 .name = "kmem.usage_in_bytes",
4738 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4739 .read_u64 = mem_cgroup_read,
4742 .name = "kmem.limit_in_bytes",
4743 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4744 .read_u64 = mem_cgroup_read,
4748 static int register_kmem_files(struct cgroup *cont, struct cgroup_subsys *ss)
4752 ret = cgroup_add_files(cont, ss, kmem_cgroup_files,
4753 ARRAY_SIZE(kmem_cgroup_files));
4756 * Part of this would be better living in a separate allocation
4757 * function, leaving us with just the cgroup tree population work.
4758 * We, however, depend on state such as network's proto_list that
4759 * is only initialized after cgroup creation. I found the less
4760 * cumbersome way to deal with it to defer it all to populate time
4763 ret = mem_cgroup_sockets_init(cont, ss);
4767 static void kmem_cgroup_destroy(struct cgroup_subsys *ss,
4768 struct cgroup *cont)
4770 mem_cgroup_sockets_destroy(cont, ss);
4773 static int register_kmem_files(struct cgroup *cont, struct cgroup_subsys *ss)
4778 static void kmem_cgroup_destroy(struct cgroup_subsys *ss,
4779 struct cgroup *cont)
4784 static struct cftype mem_cgroup_files[] = {
4786 .name = "usage_in_bytes",
4787 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4788 .read_u64 = mem_cgroup_read,
4789 .register_event = mem_cgroup_usage_register_event,
4790 .unregister_event = mem_cgroup_usage_unregister_event,
4793 .name = "max_usage_in_bytes",
4794 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4795 .trigger = mem_cgroup_reset,
4796 .read_u64 = mem_cgroup_read,
4799 .name = "limit_in_bytes",
4800 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4801 .write_string = mem_cgroup_write,
4802 .read_u64 = mem_cgroup_read,
4805 .name = "soft_limit_in_bytes",
4806 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4807 .write_string = mem_cgroup_write,
4808 .read_u64 = mem_cgroup_read,
4812 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4813 .trigger = mem_cgroup_reset,
4814 .read_u64 = mem_cgroup_read,
4818 .read_map = mem_control_stat_show,
4821 .name = "force_empty",
4822 .trigger = mem_cgroup_force_empty_write,
4825 .name = "use_hierarchy",
4826 .write_u64 = mem_cgroup_hierarchy_write,
4827 .read_u64 = mem_cgroup_hierarchy_read,
4830 .name = "swappiness",
4831 .read_u64 = mem_cgroup_swappiness_read,
4832 .write_u64 = mem_cgroup_swappiness_write,
4835 .name = "move_charge_at_immigrate",
4836 .read_u64 = mem_cgroup_move_charge_read,
4837 .write_u64 = mem_cgroup_move_charge_write,
4840 .name = "oom_control",
4841 .read_map = mem_cgroup_oom_control_read,
4842 .write_u64 = mem_cgroup_oom_control_write,
4843 .register_event = mem_cgroup_oom_register_event,
4844 .unregister_event = mem_cgroup_oom_unregister_event,
4845 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4849 .name = "numa_stat",
4850 .open = mem_control_numa_stat_open,
4856 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4857 static struct cftype memsw_cgroup_files[] = {
4859 .name = "memsw.usage_in_bytes",
4860 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4861 .read_u64 = mem_cgroup_read,
4862 .register_event = mem_cgroup_usage_register_event,
4863 .unregister_event = mem_cgroup_usage_unregister_event,
4866 .name = "memsw.max_usage_in_bytes",
4867 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4868 .trigger = mem_cgroup_reset,
4869 .read_u64 = mem_cgroup_read,
4872 .name = "memsw.limit_in_bytes",
4873 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4874 .write_string = mem_cgroup_write,
4875 .read_u64 = mem_cgroup_read,
4878 .name = "memsw.failcnt",
4879 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4880 .trigger = mem_cgroup_reset,
4881 .read_u64 = mem_cgroup_read,
4885 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4887 if (!do_swap_account)
4889 return cgroup_add_files(cont, ss, memsw_cgroup_files,
4890 ARRAY_SIZE(memsw_cgroup_files));
4893 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4899 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4901 struct mem_cgroup_per_node *pn;
4902 struct mem_cgroup_per_zone *mz;
4904 int zone, tmp = node;
4906 * This routine is called against possible nodes.
4907 * But it's BUG to call kmalloc() against offline node.
4909 * TODO: this routine can waste much memory for nodes which will
4910 * never be onlined. It's better to use memory hotplug callback
4913 if (!node_state(node, N_NORMAL_MEMORY))
4915 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4919 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4920 mz = &pn->zoneinfo[zone];
4922 INIT_LIST_HEAD(&mz->lruvec.lists[l]);
4923 mz->usage_in_excess = 0;
4924 mz->on_tree = false;
4927 memcg->info.nodeinfo[node] = pn;
4931 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4933 kfree(memcg->info.nodeinfo[node]);
4936 static struct mem_cgroup *mem_cgroup_alloc(void)
4938 struct mem_cgroup *mem;
4939 int size = sizeof(struct mem_cgroup);
4941 /* Can be very big if MAX_NUMNODES is very big */
4942 if (size < PAGE_SIZE)
4943 mem = kzalloc(size, GFP_KERNEL);
4945 mem = vzalloc(size);
4950 mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4953 spin_lock_init(&mem->pcp_counter_lock);
4957 if (size < PAGE_SIZE)
4965 * At destroying mem_cgroup, references from swap_cgroup can remain.
4966 * (scanning all at force_empty is too costly...)
4968 * Instead of clearing all references at force_empty, we remember
4969 * the number of reference from swap_cgroup and free mem_cgroup when
4970 * it goes down to 0.
4972 * Removal of cgroup itself succeeds regardless of refs from swap.
4975 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4979 mem_cgroup_remove_from_trees(memcg);
4980 free_css_id(&mem_cgroup_subsys, &memcg->css);
4982 for_each_node_state(node, N_POSSIBLE)
4983 free_mem_cgroup_per_zone_info(memcg, node);
4985 free_percpu(memcg->stat);
4986 if (sizeof(struct mem_cgroup) < PAGE_SIZE)
4992 static void mem_cgroup_get(struct mem_cgroup *memcg)
4994 atomic_inc(&memcg->refcnt);
4997 static void __mem_cgroup_put(struct mem_cgroup *memcg, int count)
4999 if (atomic_sub_and_test(count, &memcg->refcnt)) {
5000 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5001 __mem_cgroup_free(memcg);
5003 mem_cgroup_put(parent);
5007 static void mem_cgroup_put(struct mem_cgroup *memcg)
5009 __mem_cgroup_put(memcg, 1);
5013 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
5015 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
5017 if (!memcg->res.parent)
5019 return mem_cgroup_from_res_counter(memcg->res.parent, res);
5021 EXPORT_SYMBOL(parent_mem_cgroup);
5023 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5024 static void __init enable_swap_cgroup(void)
5026 if (!mem_cgroup_disabled() && really_do_swap_account)
5027 do_swap_account = 1;
5030 static void __init enable_swap_cgroup(void)
5035 static int mem_cgroup_soft_limit_tree_init(void)
5037 struct mem_cgroup_tree_per_node *rtpn;
5038 struct mem_cgroup_tree_per_zone *rtpz;
5039 int tmp, node, zone;
5041 for_each_node_state(node, N_POSSIBLE) {
5043 if (!node_state(node, N_NORMAL_MEMORY))
5045 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
5049 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5051 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5052 rtpz = &rtpn->rb_tree_per_zone[zone];
5053 rtpz->rb_root = RB_ROOT;
5054 spin_lock_init(&rtpz->lock);
5060 static struct cgroup_subsys_state * __ref
5061 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
5063 struct mem_cgroup *memcg, *parent;
5064 long error = -ENOMEM;
5067 memcg = mem_cgroup_alloc();
5069 return ERR_PTR(error);
5071 for_each_node_state(node, N_POSSIBLE)
5072 if (alloc_mem_cgroup_per_zone_info(memcg, node))
5076 if (cont->parent == NULL) {
5078 enable_swap_cgroup();
5080 if (mem_cgroup_soft_limit_tree_init())
5082 root_mem_cgroup = memcg;
5083 for_each_possible_cpu(cpu) {
5084 struct memcg_stock_pcp *stock =
5085 &per_cpu(memcg_stock, cpu);
5086 INIT_WORK(&stock->work, drain_local_stock);
5088 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5090 parent = mem_cgroup_from_cont(cont->parent);
5091 memcg->use_hierarchy = parent->use_hierarchy;
5092 memcg->oom_kill_disable = parent->oom_kill_disable;
5095 if (parent && parent->use_hierarchy) {
5096 res_counter_init(&memcg->res, &parent->res);
5097 res_counter_init(&memcg->memsw, &parent->memsw);
5098 res_counter_init(&memcg->kmem, &parent->kmem);
5100 * We increment refcnt of the parent to ensure that we can
5101 * safely access it on res_counter_charge/uncharge.
5102 * This refcnt will be decremented when freeing this
5103 * mem_cgroup(see mem_cgroup_put).
5105 mem_cgroup_get(parent);
5107 res_counter_init(&memcg->res, NULL);
5108 res_counter_init(&memcg->memsw, NULL);
5109 res_counter_init(&memcg->kmem, NULL);
5111 memcg->last_scanned_node = MAX_NUMNODES;
5112 INIT_LIST_HEAD(&memcg->oom_notify);
5115 memcg->swappiness = mem_cgroup_swappiness(parent);
5116 atomic_set(&memcg->refcnt, 1);
5117 memcg->move_charge_at_immigrate = 0;
5118 mutex_init(&memcg->thresholds_lock);
5121 __mem_cgroup_free(memcg);
5122 return ERR_PTR(error);
5125 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
5126 struct cgroup *cont)
5128 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5130 return mem_cgroup_force_empty(memcg, false);
5133 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
5134 struct cgroup *cont)
5136 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5138 kmem_cgroup_destroy(ss, cont);
5140 mem_cgroup_put(memcg);
5143 static int mem_cgroup_populate(struct cgroup_subsys *ss,
5144 struct cgroup *cont)
5148 ret = cgroup_add_files(cont, ss, mem_cgroup_files,
5149 ARRAY_SIZE(mem_cgroup_files));
5152 ret = register_memsw_files(cont, ss);
5155 ret = register_kmem_files(cont, ss);
5161 /* Handlers for move charge at task migration. */
5162 #define PRECHARGE_COUNT_AT_ONCE 256
5163 static int mem_cgroup_do_precharge(unsigned long count)
5166 int batch_count = PRECHARGE_COUNT_AT_ONCE;
5167 struct mem_cgroup *memcg = mc.to;
5169 if (mem_cgroup_is_root(memcg)) {
5170 mc.precharge += count;
5171 /* we don't need css_get for root */
5174 /* try to charge at once */
5176 struct res_counter *dummy;
5178 * "memcg" cannot be under rmdir() because we've already checked
5179 * by cgroup_lock_live_cgroup() that it is not removed and we
5180 * are still under the same cgroup_mutex. So we can postpone
5183 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
5185 if (do_swap_account && res_counter_charge(&memcg->memsw,
5186 PAGE_SIZE * count, &dummy)) {
5187 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
5190 mc.precharge += count;
5194 /* fall back to one by one charge */
5196 if (signal_pending(current)) {
5200 if (!batch_count--) {
5201 batch_count = PRECHARGE_COUNT_AT_ONCE;
5204 ret = __mem_cgroup_try_charge(NULL,
5205 GFP_KERNEL, 1, &memcg, false);
5207 /* mem_cgroup_clear_mc() will do uncharge later */
5215 * is_target_pte_for_mc - check a pte whether it is valid for move charge
5216 * @vma: the vma the pte to be checked belongs
5217 * @addr: the address corresponding to the pte to be checked
5218 * @ptent: the pte to be checked
5219 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5222 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5223 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5224 * move charge. if @target is not NULL, the page is stored in target->page
5225 * with extra refcnt got(Callers should handle it).
5226 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5227 * target for charge migration. if @target is not NULL, the entry is stored
5230 * Called with pte lock held.
5237 enum mc_target_type {
5238 MC_TARGET_NONE, /* not used */
5243 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5244 unsigned long addr, pte_t ptent)
5246 struct page *page = vm_normal_page(vma, addr, ptent);
5248 if (!page || !page_mapped(page))
5250 if (PageAnon(page)) {
5251 /* we don't move shared anon */
5252 if (!move_anon() || page_mapcount(page) > 2)
5254 } else if (!move_file())
5255 /* we ignore mapcount for file pages */
5257 if (!get_page_unless_zero(page))
5263 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5264 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5267 struct page *page = NULL;
5268 swp_entry_t ent = pte_to_swp_entry(ptent);
5270 if (!move_anon() || non_swap_entry(ent))
5272 usage_count = mem_cgroup_count_swap_user(ent, &page);
5273 if (usage_count > 1) { /* we don't move shared anon */
5278 if (do_swap_account)
5279 entry->val = ent.val;
5284 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5285 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5287 struct page *page = NULL;
5288 struct inode *inode;
5289 struct address_space *mapping;
5292 if (!vma->vm_file) /* anonymous vma */
5297 inode = vma->vm_file->f_path.dentry->d_inode;
5298 mapping = vma->vm_file->f_mapping;
5299 if (pte_none(ptent))
5300 pgoff = linear_page_index(vma, addr);
5301 else /* pte_file(ptent) is true */
5302 pgoff = pte_to_pgoff(ptent);
5304 /* page is moved even if it's not RSS of this task(page-faulted). */
5305 page = find_get_page(mapping, pgoff);
5308 /* shmem/tmpfs may report page out on swap: account for that too. */
5309 if (radix_tree_exceptional_entry(page)) {
5310 swp_entry_t swap = radix_to_swp_entry(page);
5311 if (do_swap_account)
5313 page = find_get_page(&swapper_space, swap.val);
5319 static int is_target_pte_for_mc(struct vm_area_struct *vma,
5320 unsigned long addr, pte_t ptent, union mc_target *target)
5322 struct page *page = NULL;
5323 struct page_cgroup *pc;
5325 swp_entry_t ent = { .val = 0 };
5327 if (pte_present(ptent))
5328 page = mc_handle_present_pte(vma, addr, ptent);
5329 else if (is_swap_pte(ptent))
5330 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5331 else if (pte_none(ptent) || pte_file(ptent))
5332 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5334 if (!page && !ent.val)
5337 pc = lookup_page_cgroup(page);
5339 * Do only loose check w/o page_cgroup lock.
5340 * mem_cgroup_move_account() checks the pc is valid or not under
5343 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5344 ret = MC_TARGET_PAGE;
5346 target->page = page;
5348 if (!ret || !target)
5351 /* There is a swap entry and a page doesn't exist or isn't charged */
5352 if (ent.val && !ret &&
5353 css_id(&mc.from->css) == lookup_swap_cgroup(ent)) {
5354 ret = MC_TARGET_SWAP;
5361 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5362 unsigned long addr, unsigned long end,
5363 struct mm_walk *walk)
5365 struct vm_area_struct *vma = walk->private;
5369 split_huge_page_pmd(walk->mm, pmd);
5371 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5372 for (; addr != end; pte++, addr += PAGE_SIZE)
5373 if (is_target_pte_for_mc(vma, addr, *pte, NULL))
5374 mc.precharge++; /* increment precharge temporarily */
5375 pte_unmap_unlock(pte - 1, ptl);
5381 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5383 unsigned long precharge;
5384 struct vm_area_struct *vma;
5386 down_read(&mm->mmap_sem);
5387 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5388 struct mm_walk mem_cgroup_count_precharge_walk = {
5389 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5393 if (is_vm_hugetlb_page(vma))
5395 walk_page_range(vma->vm_start, vma->vm_end,
5396 &mem_cgroup_count_precharge_walk);
5398 up_read(&mm->mmap_sem);
5400 precharge = mc.precharge;
5406 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5408 unsigned long precharge = mem_cgroup_count_precharge(mm);
5410 VM_BUG_ON(mc.moving_task);
5411 mc.moving_task = current;
5412 return mem_cgroup_do_precharge(precharge);
5415 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5416 static void __mem_cgroup_clear_mc(void)
5418 struct mem_cgroup *from = mc.from;
5419 struct mem_cgroup *to = mc.to;
5421 /* we must uncharge all the leftover precharges from mc.to */
5423 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
5427 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5428 * we must uncharge here.
5430 if (mc.moved_charge) {
5431 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
5432 mc.moved_charge = 0;
5434 /* we must fixup refcnts and charges */
5435 if (mc.moved_swap) {
5436 /* uncharge swap account from the old cgroup */
5437 if (!mem_cgroup_is_root(mc.from))
5438 res_counter_uncharge(&mc.from->memsw,
5439 PAGE_SIZE * mc.moved_swap);
5440 __mem_cgroup_put(mc.from, mc.moved_swap);
5442 if (!mem_cgroup_is_root(mc.to)) {
5444 * we charged both to->res and to->memsw, so we should
5447 res_counter_uncharge(&mc.to->res,
5448 PAGE_SIZE * mc.moved_swap);
5450 /* we've already done mem_cgroup_get(mc.to) */
5453 memcg_oom_recover(from);
5454 memcg_oom_recover(to);
5455 wake_up_all(&mc.waitq);
5458 static void mem_cgroup_clear_mc(void)
5460 struct mem_cgroup *from = mc.from;
5463 * we must clear moving_task before waking up waiters at the end of
5466 mc.moving_task = NULL;
5467 __mem_cgroup_clear_mc();
5468 spin_lock(&mc.lock);
5471 spin_unlock(&mc.lock);
5472 mem_cgroup_end_move(from);
5475 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5476 struct cgroup *cgroup,
5477 struct cgroup_taskset *tset)
5479 struct task_struct *p = cgroup_taskset_first(tset);
5481 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup);
5483 if (memcg->move_charge_at_immigrate) {
5484 struct mm_struct *mm;
5485 struct mem_cgroup *from = mem_cgroup_from_task(p);
5487 VM_BUG_ON(from == memcg);
5489 mm = get_task_mm(p);
5492 /* We move charges only when we move a owner of the mm */
5493 if (mm->owner == p) {
5496 VM_BUG_ON(mc.precharge);
5497 VM_BUG_ON(mc.moved_charge);
5498 VM_BUG_ON(mc.moved_swap);
5499 mem_cgroup_start_move(from);
5500 spin_lock(&mc.lock);
5503 spin_unlock(&mc.lock);
5504 /* We set mc.moving_task later */
5506 ret = mem_cgroup_precharge_mc(mm);
5508 mem_cgroup_clear_mc();
5515 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5516 struct cgroup *cgroup,
5517 struct cgroup_taskset *tset)
5519 mem_cgroup_clear_mc();
5522 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5523 unsigned long addr, unsigned long end,
5524 struct mm_walk *walk)
5527 struct vm_area_struct *vma = walk->private;
5531 split_huge_page_pmd(walk->mm, pmd);
5533 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5534 for (; addr != end; addr += PAGE_SIZE) {
5535 pte_t ptent = *(pte++);
5536 union mc_target target;
5539 struct page_cgroup *pc;
5545 type = is_target_pte_for_mc(vma, addr, ptent, &target);
5547 case MC_TARGET_PAGE:
5549 if (isolate_lru_page(page))
5551 pc = lookup_page_cgroup(page);
5552 if (!mem_cgroup_move_account(page, 1, pc,
5553 mc.from, mc.to, false)) {
5555 /* we uncharge from mc.from later. */
5558 putback_lru_page(page);
5559 put: /* is_target_pte_for_mc() gets the page */
5562 case MC_TARGET_SWAP:
5564 if (!mem_cgroup_move_swap_account(ent,
5565 mc.from, mc.to, false)) {
5567 /* we fixup refcnts and charges later. */
5575 pte_unmap_unlock(pte - 1, ptl);
5580 * We have consumed all precharges we got in can_attach().
5581 * We try charge one by one, but don't do any additional
5582 * charges to mc.to if we have failed in charge once in attach()
5585 ret = mem_cgroup_do_precharge(1);
5593 static void mem_cgroup_move_charge(struct mm_struct *mm)
5595 struct vm_area_struct *vma;
5597 lru_add_drain_all();
5599 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5601 * Someone who are holding the mmap_sem might be waiting in
5602 * waitq. So we cancel all extra charges, wake up all waiters,
5603 * and retry. Because we cancel precharges, we might not be able
5604 * to move enough charges, but moving charge is a best-effort
5605 * feature anyway, so it wouldn't be a big problem.
5607 __mem_cgroup_clear_mc();
5611 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5613 struct mm_walk mem_cgroup_move_charge_walk = {
5614 .pmd_entry = mem_cgroup_move_charge_pte_range,
5618 if (is_vm_hugetlb_page(vma))
5620 ret = walk_page_range(vma->vm_start, vma->vm_end,
5621 &mem_cgroup_move_charge_walk);
5624 * means we have consumed all precharges and failed in
5625 * doing additional charge. Just abandon here.
5629 up_read(&mm->mmap_sem);
5632 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5633 struct cgroup *cont,
5634 struct cgroup_taskset *tset)
5636 struct task_struct *p = cgroup_taskset_first(tset);
5637 struct mm_struct *mm = get_task_mm(p);
5641 mem_cgroup_move_charge(mm);
5646 mem_cgroup_clear_mc();
5648 #else /* !CONFIG_MMU */
5649 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5650 struct cgroup *cgroup,
5651 struct cgroup_taskset *tset)
5655 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5656 struct cgroup *cgroup,
5657 struct cgroup_taskset *tset)
5660 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5661 struct cgroup *cont,
5662 struct cgroup_taskset *tset)
5667 struct cgroup_subsys mem_cgroup_subsys = {
5669 .subsys_id = mem_cgroup_subsys_id,
5670 .create = mem_cgroup_create,
5671 .pre_destroy = mem_cgroup_pre_destroy,
5672 .destroy = mem_cgroup_destroy,
5673 .populate = mem_cgroup_populate,
5674 .can_attach = mem_cgroup_can_attach,
5675 .cancel_attach = mem_cgroup_cancel_attach,
5676 .attach = mem_cgroup_move_task,
5681 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5682 static int __init enable_swap_account(char *s)
5684 /* consider enabled if no parameter or 1 is given */
5685 if (!strcmp(s, "1"))
5686 really_do_swap_account = 1;
5687 else if (!strcmp(s, "0"))
5688 really_do_swap_account = 0;
5691 __setup("swapaccount=", enable_swap_account);