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 __memcg_event_check(struct mem_cgroup *memcg, int target)
758 unsigned long val, next;
760 val = __this_cpu_read(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT]);
761 next = __this_cpu_read(memcg->stat->targets[target]);
762 /* from time_after() in jiffies.h */
763 return ((long)next - (long)val < 0);
766 static void __mem_cgroup_target_update(struct mem_cgroup *memcg, int target)
768 unsigned long val, next;
770 val = __this_cpu_read(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT]);
773 case MEM_CGROUP_TARGET_THRESH:
774 next = val + THRESHOLDS_EVENTS_TARGET;
776 case MEM_CGROUP_TARGET_SOFTLIMIT:
777 next = val + SOFTLIMIT_EVENTS_TARGET;
779 case MEM_CGROUP_TARGET_NUMAINFO:
780 next = val + NUMAINFO_EVENTS_TARGET;
786 __this_cpu_write(memcg->stat->targets[target], next);
790 * Check events in order.
793 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
796 /* threshold event is triggered in finer grain than soft limit */
797 if (unlikely(__memcg_event_check(memcg, MEM_CGROUP_TARGET_THRESH))) {
798 mem_cgroup_threshold(memcg);
799 __mem_cgroup_target_update(memcg, MEM_CGROUP_TARGET_THRESH);
800 if (unlikely(__memcg_event_check(memcg,
801 MEM_CGROUP_TARGET_SOFTLIMIT))) {
802 mem_cgroup_update_tree(memcg, page);
803 __mem_cgroup_target_update(memcg,
804 MEM_CGROUP_TARGET_SOFTLIMIT);
807 if (unlikely(__memcg_event_check(memcg,
808 MEM_CGROUP_TARGET_NUMAINFO))) {
809 atomic_inc(&memcg->numainfo_events);
810 __mem_cgroup_target_update(memcg,
811 MEM_CGROUP_TARGET_NUMAINFO);
818 struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
820 return container_of(cgroup_subsys_state(cont,
821 mem_cgroup_subsys_id), struct mem_cgroup,
825 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
828 * mm_update_next_owner() may clear mm->owner to NULL
829 * if it races with swapoff, page migration, etc.
830 * So this can be called with p == NULL.
835 return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
836 struct mem_cgroup, css);
839 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
841 struct mem_cgroup *memcg = NULL;
846 * Because we have no locks, mm->owner's may be being moved to other
847 * cgroup. We use css_tryget() here even if this looks
848 * pessimistic (rather than adding locks here).
852 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
853 if (unlikely(!memcg))
855 } while (!css_tryget(&memcg->css));
861 * mem_cgroup_iter - iterate over memory cgroup hierarchy
862 * @root: hierarchy root
863 * @prev: previously returned memcg, NULL on first invocation
864 * @reclaim: cookie for shared reclaim walks, NULL for full walks
866 * Returns references to children of the hierarchy below @root, or
867 * @root itself, or %NULL after a full round-trip.
869 * Caller must pass the return value in @prev on subsequent
870 * invocations for reference counting, or use mem_cgroup_iter_break()
871 * to cancel a hierarchy walk before the round-trip is complete.
873 * Reclaimers can specify a zone and a priority level in @reclaim to
874 * divide up the memcgs in the hierarchy among all concurrent
875 * reclaimers operating on the same zone and priority.
877 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
878 struct mem_cgroup *prev,
879 struct mem_cgroup_reclaim_cookie *reclaim)
881 struct mem_cgroup *memcg = NULL;
884 if (mem_cgroup_disabled())
888 root = root_mem_cgroup;
890 if (prev && !reclaim)
891 id = css_id(&prev->css);
893 if (prev && prev != root)
896 if (!root->use_hierarchy && root != root_mem_cgroup) {
903 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
904 struct cgroup_subsys_state *css;
907 int nid = zone_to_nid(reclaim->zone);
908 int zid = zone_idx(reclaim->zone);
909 struct mem_cgroup_per_zone *mz;
911 mz = mem_cgroup_zoneinfo(root, nid, zid);
912 iter = &mz->reclaim_iter[reclaim->priority];
913 if (prev && reclaim->generation != iter->generation)
919 css = css_get_next(&mem_cgroup_subsys, id + 1, &root->css, &id);
921 if (css == &root->css || css_tryget(css))
922 memcg = container_of(css,
923 struct mem_cgroup, css);
932 else if (!prev && memcg)
933 reclaim->generation = iter->generation;
943 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
944 * @root: hierarchy root
945 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
947 void mem_cgroup_iter_break(struct mem_cgroup *root,
948 struct mem_cgroup *prev)
951 root = root_mem_cgroup;
952 if (prev && prev != root)
957 * Iteration constructs for visiting all cgroups (under a tree). If
958 * loops are exited prematurely (break), mem_cgroup_iter_break() must
959 * be used for reference counting.
961 #define for_each_mem_cgroup_tree(iter, root) \
962 for (iter = mem_cgroup_iter(root, NULL, NULL); \
964 iter = mem_cgroup_iter(root, iter, NULL))
966 #define for_each_mem_cgroup(iter) \
967 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
969 iter = mem_cgroup_iter(NULL, iter, NULL))
971 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
973 return (memcg == root_mem_cgroup);
976 void mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
978 struct mem_cgroup *memcg;
984 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
985 if (unlikely(!memcg))
990 mem_cgroup_pgmajfault(memcg, 1);
993 mem_cgroup_pgfault(memcg, 1);
1001 EXPORT_SYMBOL(mem_cgroup_count_vm_event);
1004 * Following LRU functions are allowed to be used without PCG_LOCK.
1005 * Operations are called by routine of global LRU independently from memcg.
1006 * What we have to take care of here is validness of pc->mem_cgroup.
1008 * Changes to pc->mem_cgroup happens when
1011 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1012 * It is added to LRU before charge.
1013 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1014 * When moving account, the page is not on LRU. It's isolated.
1017 void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru)
1019 struct page_cgroup *pc;
1020 struct mem_cgroup_per_zone *mz;
1022 if (mem_cgroup_disabled())
1024 pc = lookup_page_cgroup(page);
1025 /* can happen while we handle swapcache. */
1026 if (!TestClearPageCgroupAcctLRU(pc))
1028 VM_BUG_ON(!pc->mem_cgroup);
1030 * We don't check PCG_USED bit. It's cleared when the "page" is finally
1031 * removed from global LRU.
1033 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1034 /* huge page split is done under lru_lock. so, we have no races. */
1035 MEM_CGROUP_ZSTAT(mz, lru) -= 1 << compound_order(page);
1036 VM_BUG_ON(list_empty(&pc->lru));
1037 list_del_init(&pc->lru);
1040 void mem_cgroup_del_lru(struct page *page)
1042 mem_cgroup_del_lru_list(page, page_lru(page));
1046 * Writeback is about to end against a page which has been marked for immediate
1047 * reclaim. If it still appears to be reclaimable, move it to the tail of the
1050 void mem_cgroup_rotate_reclaimable_page(struct page *page)
1052 struct mem_cgroup_per_zone *mz;
1053 struct page_cgroup *pc;
1054 enum lru_list lru = page_lru(page);
1056 if (mem_cgroup_disabled())
1059 pc = lookup_page_cgroup(page);
1060 /* unused page is not rotated. */
1061 if (!PageCgroupUsed(pc))
1063 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1065 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1066 list_move_tail(&pc->lru, &mz->lruvec.lists[lru]);
1069 void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru)
1071 struct mem_cgroup_per_zone *mz;
1072 struct page_cgroup *pc;
1074 if (mem_cgroup_disabled())
1077 pc = lookup_page_cgroup(page);
1078 /* unused page is not rotated. */
1079 if (!PageCgroupUsed(pc))
1081 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1083 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1084 list_move(&pc->lru, &mz->lruvec.lists[lru]);
1087 void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
1089 struct page_cgroup *pc;
1090 struct mem_cgroup_per_zone *mz;
1092 if (mem_cgroup_disabled())
1094 pc = lookup_page_cgroup(page);
1095 VM_BUG_ON(PageCgroupAcctLRU(pc));
1098 * SetPageLRU SetPageCgroupUsed
1100 * PageCgroupUsed && add to memcg LRU PageLRU && add to memcg LRU
1102 * Ensure that one of the two sides adds the page to the memcg
1103 * LRU during a race.
1106 if (!PageCgroupUsed(pc))
1108 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1110 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1111 /* huge page split is done under lru_lock. so, we have no races. */
1112 MEM_CGROUP_ZSTAT(mz, lru) += 1 << compound_order(page);
1113 SetPageCgroupAcctLRU(pc);
1114 list_add(&pc->lru, &mz->lruvec.lists[lru]);
1118 * At handling SwapCache and other FUSE stuff, pc->mem_cgroup may be changed
1119 * while it's linked to lru because the page may be reused after it's fully
1120 * uncharged. To handle that, unlink page_cgroup from LRU when charge it again.
1121 * It's done under lock_page and expected that zone->lru_lock isnever held.
1123 static void mem_cgroup_lru_del_before_commit(struct page *page)
1125 unsigned long flags;
1126 struct zone *zone = page_zone(page);
1127 struct page_cgroup *pc = lookup_page_cgroup(page);
1130 * Doing this check without taking ->lru_lock seems wrong but this
1131 * is safe. Because if page_cgroup's USED bit is unset, the page
1132 * will not be added to any memcg's LRU. If page_cgroup's USED bit is
1133 * set, the commit after this will fail, anyway.
1134 * This all charge/uncharge is done under some mutual execustion.
1135 * So, we don't need to taking care of changes in USED bit.
1137 if (likely(!PageLRU(page)))
1140 spin_lock_irqsave(&zone->lru_lock, flags);
1142 * Forget old LRU when this page_cgroup is *not* used. This Used bit
1143 * is guarded by lock_page() because the page is SwapCache.
1145 if (!PageCgroupUsed(pc))
1146 mem_cgroup_del_lru_list(page, page_lru(page));
1147 spin_unlock_irqrestore(&zone->lru_lock, flags);
1150 static void mem_cgroup_lru_add_after_commit(struct page *page)
1152 unsigned long flags;
1153 struct zone *zone = page_zone(page);
1154 struct page_cgroup *pc = lookup_page_cgroup(page);
1157 * SetPageLRU SetPageCgroupUsed
1159 * PageCgroupUsed && add to memcg LRU PageLRU && add to memcg LRU
1161 * Ensure that one of the two sides adds the page to the memcg
1162 * LRU during a race.
1165 /* taking care of that the page is added to LRU while we commit it */
1166 if (likely(!PageLRU(page)))
1168 spin_lock_irqsave(&zone->lru_lock, flags);
1169 /* link when the page is linked to LRU but page_cgroup isn't */
1170 if (PageLRU(page) && !PageCgroupAcctLRU(pc))
1171 mem_cgroup_add_lru_list(page, page_lru(page));
1172 spin_unlock_irqrestore(&zone->lru_lock, flags);
1176 void mem_cgroup_move_lists(struct page *page,
1177 enum lru_list from, enum lru_list to)
1179 if (mem_cgroup_disabled())
1181 mem_cgroup_del_lru_list(page, from);
1182 mem_cgroup_add_lru_list(page, to);
1186 * Checks whether given mem is same or in the root_mem_cgroup's
1189 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1190 struct mem_cgroup *memcg)
1192 if (root_memcg != memcg) {
1193 return (root_memcg->use_hierarchy &&
1194 css_is_ancestor(&memcg->css, &root_memcg->css));
1200 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *memcg)
1203 struct mem_cgroup *curr = NULL;
1204 struct task_struct *p;
1206 p = find_lock_task_mm(task);
1209 curr = try_get_mem_cgroup_from_mm(p->mm);
1214 * We should check use_hierarchy of "memcg" not "curr". Because checking
1215 * use_hierarchy of "curr" here make this function true if hierarchy is
1216 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1217 * hierarchy(even if use_hierarchy is disabled in "memcg").
1219 ret = mem_cgroup_same_or_subtree(memcg, curr);
1220 css_put(&curr->css);
1224 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg, struct zone *zone)
1226 unsigned long inactive_ratio;
1227 int nid = zone_to_nid(zone);
1228 int zid = zone_idx(zone);
1229 unsigned long inactive;
1230 unsigned long active;
1233 inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1234 BIT(LRU_INACTIVE_ANON));
1235 active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1236 BIT(LRU_ACTIVE_ANON));
1238 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1240 inactive_ratio = int_sqrt(10 * gb);
1244 return inactive * inactive_ratio < active;
1247 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg, struct zone *zone)
1249 unsigned long active;
1250 unsigned long inactive;
1251 int zid = zone_idx(zone);
1252 int nid = zone_to_nid(zone);
1254 inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1255 BIT(LRU_INACTIVE_FILE));
1256 active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1257 BIT(LRU_ACTIVE_FILE));
1259 return (active > inactive);
1262 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
1265 int nid = zone_to_nid(zone);
1266 int zid = zone_idx(zone);
1267 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
1269 return &mz->reclaim_stat;
1272 struct zone_reclaim_stat *
1273 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
1275 struct page_cgroup *pc;
1276 struct mem_cgroup_per_zone *mz;
1278 if (mem_cgroup_disabled())
1281 pc = lookup_page_cgroup(page);
1282 if (!PageCgroupUsed(pc))
1284 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1286 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1287 return &mz->reclaim_stat;
1290 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
1291 struct list_head *dst,
1292 unsigned long *scanned, int order,
1293 isolate_mode_t mode,
1295 struct mem_cgroup *mem_cont,
1296 int active, int file)
1298 unsigned long nr_taken = 0;
1302 struct list_head *src;
1303 struct page_cgroup *pc, *tmp;
1304 int nid = zone_to_nid(z);
1305 int zid = zone_idx(z);
1306 struct mem_cgroup_per_zone *mz;
1307 int lru = LRU_FILE * file + active;
1311 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
1312 src = &mz->lruvec.lists[lru];
1315 list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
1316 if (scan >= nr_to_scan)
1319 if (unlikely(!PageCgroupUsed(pc)))
1322 page = lookup_cgroup_page(pc);
1324 if (unlikely(!PageLRU(page)))
1328 ret = __isolate_lru_page(page, mode, file);
1331 list_move(&page->lru, dst);
1332 mem_cgroup_del_lru(page);
1333 nr_taken += hpage_nr_pages(page);
1336 /* we don't affect global LRU but rotate in our LRU */
1337 mem_cgroup_rotate_lru_list(page, page_lru(page));
1346 trace_mm_vmscan_memcg_isolate(0, nr_to_scan, scan, nr_taken,
1352 #define mem_cgroup_from_res_counter(counter, member) \
1353 container_of(counter, struct mem_cgroup, member)
1356 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1357 * @mem: the memory cgroup
1359 * Returns the maximum amount of memory @mem can be charged with, in
1362 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1364 unsigned long long margin;
1366 margin = res_counter_margin(&memcg->res);
1367 if (do_swap_account)
1368 margin = min(margin, res_counter_margin(&memcg->memsw));
1369 return margin >> PAGE_SHIFT;
1372 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1374 struct cgroup *cgrp = memcg->css.cgroup;
1377 if (cgrp->parent == NULL)
1378 return vm_swappiness;
1380 return memcg->swappiness;
1383 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1388 spin_lock(&memcg->pcp_counter_lock);
1389 for_each_online_cpu(cpu)
1390 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) += 1;
1391 memcg->nocpu_base.count[MEM_CGROUP_ON_MOVE] += 1;
1392 spin_unlock(&memcg->pcp_counter_lock);
1398 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1405 spin_lock(&memcg->pcp_counter_lock);
1406 for_each_online_cpu(cpu)
1407 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) -= 1;
1408 memcg->nocpu_base.count[MEM_CGROUP_ON_MOVE] -= 1;
1409 spin_unlock(&memcg->pcp_counter_lock);
1413 * 2 routines for checking "mem" is under move_account() or not.
1415 * mem_cgroup_stealed() - checking a cgroup is mc.from or not. This is used
1416 * for avoiding race in accounting. If true,
1417 * pc->mem_cgroup may be overwritten.
1419 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1420 * under hierarchy of moving cgroups. This is for
1421 * waiting at hith-memory prressure caused by "move".
1424 static bool mem_cgroup_stealed(struct mem_cgroup *memcg)
1426 VM_BUG_ON(!rcu_read_lock_held());
1427 return this_cpu_read(memcg->stat->count[MEM_CGROUP_ON_MOVE]) > 0;
1430 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1432 struct mem_cgroup *from;
1433 struct mem_cgroup *to;
1436 * Unlike task_move routines, we access mc.to, mc.from not under
1437 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1439 spin_lock(&mc.lock);
1445 ret = mem_cgroup_same_or_subtree(memcg, from)
1446 || mem_cgroup_same_or_subtree(memcg, to);
1448 spin_unlock(&mc.lock);
1452 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1454 if (mc.moving_task && current != mc.moving_task) {
1455 if (mem_cgroup_under_move(memcg)) {
1457 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1458 /* moving charge context might have finished. */
1461 finish_wait(&mc.waitq, &wait);
1469 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1470 * @memcg: The memory cgroup that went over limit
1471 * @p: Task that is going to be killed
1473 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1476 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1478 struct cgroup *task_cgrp;
1479 struct cgroup *mem_cgrp;
1481 * Need a buffer in BSS, can't rely on allocations. The code relies
1482 * on the assumption that OOM is serialized for memory controller.
1483 * If this assumption is broken, revisit this code.
1485 static char memcg_name[PATH_MAX];
1494 mem_cgrp = memcg->css.cgroup;
1495 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1497 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1500 * Unfortunately, we are unable to convert to a useful name
1501 * But we'll still print out the usage information
1508 printk(KERN_INFO "Task in %s killed", memcg_name);
1511 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1519 * Continues from above, so we don't need an KERN_ level
1521 printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1524 printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1525 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1526 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1527 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1528 printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1530 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1531 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1532 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1536 * This function returns the number of memcg under hierarchy tree. Returns
1537 * 1(self count) if no children.
1539 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1542 struct mem_cgroup *iter;
1544 for_each_mem_cgroup_tree(iter, memcg)
1550 * Return the memory (and swap, if configured) limit for a memcg.
1552 u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1557 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1558 limit += total_swap_pages << PAGE_SHIFT;
1560 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1562 * If memsw is finite and limits the amount of swap space available
1563 * to this memcg, return that limit.
1565 return min(limit, memsw);
1568 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1570 unsigned long flags)
1572 unsigned long total = 0;
1573 bool noswap = false;
1576 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1578 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1581 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1583 drain_all_stock_async(memcg);
1584 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1586 * Allow limit shrinkers, which are triggered directly
1587 * by userspace, to catch signals and stop reclaim
1588 * after minimal progress, regardless of the margin.
1590 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1592 if (mem_cgroup_margin(memcg))
1595 * If nothing was reclaimed after two attempts, there
1596 * may be no reclaimable pages in this hierarchy.
1605 * test_mem_cgroup_node_reclaimable
1606 * @mem: the target memcg
1607 * @nid: the node ID to be checked.
1608 * @noswap : specify true here if the user wants flle only information.
1610 * This function returns whether the specified memcg contains any
1611 * reclaimable pages on a node. Returns true if there are any reclaimable
1612 * pages in the node.
1614 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1615 int nid, bool noswap)
1617 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1619 if (noswap || !total_swap_pages)
1621 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1626 #if MAX_NUMNODES > 1
1629 * Always updating the nodemask is not very good - even if we have an empty
1630 * list or the wrong list here, we can start from some node and traverse all
1631 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1634 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1638 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1639 * pagein/pageout changes since the last update.
1641 if (!atomic_read(&memcg->numainfo_events))
1643 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1646 /* make a nodemask where this memcg uses memory from */
1647 memcg->scan_nodes = node_states[N_HIGH_MEMORY];
1649 for_each_node_mask(nid, node_states[N_HIGH_MEMORY]) {
1651 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1652 node_clear(nid, memcg->scan_nodes);
1655 atomic_set(&memcg->numainfo_events, 0);
1656 atomic_set(&memcg->numainfo_updating, 0);
1660 * Selecting a node where we start reclaim from. Because what we need is just
1661 * reducing usage counter, start from anywhere is O,K. Considering
1662 * memory reclaim from current node, there are pros. and cons.
1664 * Freeing memory from current node means freeing memory from a node which
1665 * we'll use or we've used. So, it may make LRU bad. And if several threads
1666 * hit limits, it will see a contention on a node. But freeing from remote
1667 * node means more costs for memory reclaim because of memory latency.
1669 * Now, we use round-robin. Better algorithm is welcomed.
1671 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1675 mem_cgroup_may_update_nodemask(memcg);
1676 node = memcg->last_scanned_node;
1678 node = next_node(node, memcg->scan_nodes);
1679 if (node == MAX_NUMNODES)
1680 node = first_node(memcg->scan_nodes);
1682 * We call this when we hit limit, not when pages are added to LRU.
1683 * No LRU may hold pages because all pages are UNEVICTABLE or
1684 * memcg is too small and all pages are not on LRU. In that case,
1685 * we use curret node.
1687 if (unlikely(node == MAX_NUMNODES))
1688 node = numa_node_id();
1690 memcg->last_scanned_node = node;
1695 * Check all nodes whether it contains reclaimable pages or not.
1696 * For quick scan, we make use of scan_nodes. This will allow us to skip
1697 * unused nodes. But scan_nodes is lazily updated and may not cotain
1698 * enough new information. We need to do double check.
1700 bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1705 * quick check...making use of scan_node.
1706 * We can skip unused nodes.
1708 if (!nodes_empty(memcg->scan_nodes)) {
1709 for (nid = first_node(memcg->scan_nodes);
1711 nid = next_node(nid, memcg->scan_nodes)) {
1713 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1718 * Check rest of nodes.
1720 for_each_node_state(nid, N_HIGH_MEMORY) {
1721 if (node_isset(nid, memcg->scan_nodes))
1723 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1730 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1735 bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1737 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
1741 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1744 unsigned long *total_scanned)
1746 struct mem_cgroup *victim = NULL;
1749 unsigned long excess;
1750 unsigned long nr_scanned;
1751 struct mem_cgroup_reclaim_cookie reclaim = {
1756 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
1759 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1764 * If we have not been able to reclaim
1765 * anything, it might because there are
1766 * no reclaimable pages under this hierarchy
1771 * We want to do more targeted reclaim.
1772 * excess >> 2 is not to excessive so as to
1773 * reclaim too much, nor too less that we keep
1774 * coming back to reclaim from this cgroup
1776 if (total >= (excess >> 2) ||
1777 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1782 if (!mem_cgroup_reclaimable(victim, false))
1784 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1786 *total_scanned += nr_scanned;
1787 if (!res_counter_soft_limit_excess(&root_memcg->res))
1790 mem_cgroup_iter_break(root_memcg, victim);
1795 * Check OOM-Killer is already running under our hierarchy.
1796 * If someone is running, return false.
1797 * Has to be called with memcg_oom_lock
1799 static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg)
1801 struct mem_cgroup *iter, *failed = NULL;
1803 for_each_mem_cgroup_tree(iter, memcg) {
1804 if (iter->oom_lock) {
1806 * this subtree of our hierarchy is already locked
1807 * so we cannot give a lock.
1810 mem_cgroup_iter_break(memcg, iter);
1813 iter->oom_lock = true;
1820 * OK, we failed to lock the whole subtree so we have to clean up
1821 * what we set up to the failing subtree
1823 for_each_mem_cgroup_tree(iter, memcg) {
1824 if (iter == failed) {
1825 mem_cgroup_iter_break(memcg, iter);
1828 iter->oom_lock = false;
1834 * Has to be called with memcg_oom_lock
1836 static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1838 struct mem_cgroup *iter;
1840 for_each_mem_cgroup_tree(iter, memcg)
1841 iter->oom_lock = false;
1845 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1847 struct mem_cgroup *iter;
1849 for_each_mem_cgroup_tree(iter, memcg)
1850 atomic_inc(&iter->under_oom);
1853 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1855 struct mem_cgroup *iter;
1858 * When a new child is created while the hierarchy is under oom,
1859 * mem_cgroup_oom_lock() may not be called. We have to use
1860 * atomic_add_unless() here.
1862 for_each_mem_cgroup_tree(iter, memcg)
1863 atomic_add_unless(&iter->under_oom, -1, 0);
1866 static DEFINE_SPINLOCK(memcg_oom_lock);
1867 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1869 struct oom_wait_info {
1870 struct mem_cgroup *mem;
1874 static int memcg_oom_wake_function(wait_queue_t *wait,
1875 unsigned mode, int sync, void *arg)
1877 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg,
1879 struct oom_wait_info *oom_wait_info;
1881 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1882 oom_wait_memcg = oom_wait_info->mem;
1885 * Both of oom_wait_info->mem and wake_mem are stable under us.
1886 * Then we can use css_is_ancestor without taking care of RCU.
1888 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
1889 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
1891 return autoremove_wake_function(wait, mode, sync, arg);
1894 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
1896 /* for filtering, pass "memcg" as argument. */
1897 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1900 static void memcg_oom_recover(struct mem_cgroup *memcg)
1902 if (memcg && atomic_read(&memcg->under_oom))
1903 memcg_wakeup_oom(memcg);
1907 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1909 bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask)
1911 struct oom_wait_info owait;
1912 bool locked, need_to_kill;
1915 owait.wait.flags = 0;
1916 owait.wait.func = memcg_oom_wake_function;
1917 owait.wait.private = current;
1918 INIT_LIST_HEAD(&owait.wait.task_list);
1919 need_to_kill = true;
1920 mem_cgroup_mark_under_oom(memcg);
1922 /* At first, try to OOM lock hierarchy under memcg.*/
1923 spin_lock(&memcg_oom_lock);
1924 locked = mem_cgroup_oom_lock(memcg);
1926 * Even if signal_pending(), we can't quit charge() loop without
1927 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1928 * under OOM is always welcomed, use TASK_KILLABLE here.
1930 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1931 if (!locked || memcg->oom_kill_disable)
1932 need_to_kill = false;
1934 mem_cgroup_oom_notify(memcg);
1935 spin_unlock(&memcg_oom_lock);
1938 finish_wait(&memcg_oom_waitq, &owait.wait);
1939 mem_cgroup_out_of_memory(memcg, mask);
1942 finish_wait(&memcg_oom_waitq, &owait.wait);
1944 spin_lock(&memcg_oom_lock);
1946 mem_cgroup_oom_unlock(memcg);
1947 memcg_wakeup_oom(memcg);
1948 spin_unlock(&memcg_oom_lock);
1950 mem_cgroup_unmark_under_oom(memcg);
1952 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
1954 /* Give chance to dying process */
1955 schedule_timeout_uninterruptible(1);
1960 * Currently used to update mapped file statistics, but the routine can be
1961 * generalized to update other statistics as well.
1963 * Notes: Race condition
1965 * We usually use page_cgroup_lock() for accessing page_cgroup member but
1966 * it tends to be costly. But considering some conditions, we doesn't need
1967 * to do so _always_.
1969 * Considering "charge", lock_page_cgroup() is not required because all
1970 * file-stat operations happen after a page is attached to radix-tree. There
1971 * are no race with "charge".
1973 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
1974 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
1975 * if there are race with "uncharge". Statistics itself is properly handled
1978 * Considering "move", this is an only case we see a race. To make the race
1979 * small, we check MEM_CGROUP_ON_MOVE percpu value and detect there are
1980 * possibility of race condition. If there is, we take a lock.
1983 void mem_cgroup_update_page_stat(struct page *page,
1984 enum mem_cgroup_page_stat_item idx, int val)
1986 struct mem_cgroup *memcg;
1987 struct page_cgroup *pc = lookup_page_cgroup(page);
1988 bool need_unlock = false;
1989 unsigned long uninitialized_var(flags);
1995 memcg = pc->mem_cgroup;
1996 if (unlikely(!memcg || !PageCgroupUsed(pc)))
1998 /* pc->mem_cgroup is unstable ? */
1999 if (unlikely(mem_cgroup_stealed(memcg)) || PageTransHuge(page)) {
2000 /* take a lock against to access pc->mem_cgroup */
2001 move_lock_page_cgroup(pc, &flags);
2003 memcg = pc->mem_cgroup;
2004 if (!memcg || !PageCgroupUsed(pc))
2009 case MEMCG_NR_FILE_MAPPED:
2011 SetPageCgroupFileMapped(pc);
2012 else if (!page_mapped(page))
2013 ClearPageCgroupFileMapped(pc);
2014 idx = MEM_CGROUP_STAT_FILE_MAPPED;
2020 this_cpu_add(memcg->stat->count[idx], val);
2023 if (unlikely(need_unlock))
2024 move_unlock_page_cgroup(pc, &flags);
2028 EXPORT_SYMBOL(mem_cgroup_update_page_stat);
2031 * size of first charge trial. "32" comes from vmscan.c's magic value.
2032 * TODO: maybe necessary to use big numbers in big irons.
2034 #define CHARGE_BATCH 32U
2035 struct memcg_stock_pcp {
2036 struct mem_cgroup *cached; /* this never be root cgroup */
2037 unsigned int nr_pages;
2038 struct work_struct work;
2039 unsigned long flags;
2040 #define FLUSHING_CACHED_CHARGE (0)
2042 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2043 static DEFINE_MUTEX(percpu_charge_mutex);
2046 * Try to consume stocked charge on this cpu. If success, one page is consumed
2047 * from local stock and true is returned. If the stock is 0 or charges from a
2048 * cgroup which is not current target, returns false. This stock will be
2051 static bool consume_stock(struct mem_cgroup *memcg)
2053 struct memcg_stock_pcp *stock;
2056 stock = &get_cpu_var(memcg_stock);
2057 if (memcg == stock->cached && stock->nr_pages)
2059 else /* need to call res_counter_charge */
2061 put_cpu_var(memcg_stock);
2066 * Returns stocks cached in percpu to res_counter and reset cached information.
2068 static void drain_stock(struct memcg_stock_pcp *stock)
2070 struct mem_cgroup *old = stock->cached;
2072 if (stock->nr_pages) {
2073 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2075 res_counter_uncharge(&old->res, bytes);
2076 if (do_swap_account)
2077 res_counter_uncharge(&old->memsw, bytes);
2078 stock->nr_pages = 0;
2080 stock->cached = NULL;
2084 * This must be called under preempt disabled or must be called by
2085 * a thread which is pinned to local cpu.
2087 static void drain_local_stock(struct work_struct *dummy)
2089 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2091 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2095 * Cache charges(val) which is from res_counter, to local per_cpu area.
2096 * This will be consumed by consume_stock() function, later.
2098 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2100 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2102 if (stock->cached != memcg) { /* reset if necessary */
2104 stock->cached = memcg;
2106 stock->nr_pages += nr_pages;
2107 put_cpu_var(memcg_stock);
2111 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2112 * of the hierarchy under it. sync flag says whether we should block
2113 * until the work is done.
2115 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2119 /* Notify other cpus that system-wide "drain" is running */
2122 for_each_online_cpu(cpu) {
2123 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2124 struct mem_cgroup *memcg;
2126 memcg = stock->cached;
2127 if (!memcg || !stock->nr_pages)
2129 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2131 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2133 drain_local_stock(&stock->work);
2135 schedule_work_on(cpu, &stock->work);
2143 for_each_online_cpu(cpu) {
2144 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2145 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2146 flush_work(&stock->work);
2153 * Tries to drain stocked charges in other cpus. This function is asynchronous
2154 * and just put a work per cpu for draining localy on each cpu. Caller can
2155 * expects some charges will be back to res_counter later but cannot wait for
2158 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2161 * If someone calls draining, avoid adding more kworker runs.
2163 if (!mutex_trylock(&percpu_charge_mutex))
2165 drain_all_stock(root_memcg, false);
2166 mutex_unlock(&percpu_charge_mutex);
2169 /* This is a synchronous drain interface. */
2170 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2172 /* called when force_empty is called */
2173 mutex_lock(&percpu_charge_mutex);
2174 drain_all_stock(root_memcg, true);
2175 mutex_unlock(&percpu_charge_mutex);
2179 * This function drains percpu counter value from DEAD cpu and
2180 * move it to local cpu. Note that this function can be preempted.
2182 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2186 spin_lock(&memcg->pcp_counter_lock);
2187 for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) {
2188 long x = per_cpu(memcg->stat->count[i], cpu);
2190 per_cpu(memcg->stat->count[i], cpu) = 0;
2191 memcg->nocpu_base.count[i] += x;
2193 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2194 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2196 per_cpu(memcg->stat->events[i], cpu) = 0;
2197 memcg->nocpu_base.events[i] += x;
2199 /* need to clear ON_MOVE value, works as a kind of lock. */
2200 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) = 0;
2201 spin_unlock(&memcg->pcp_counter_lock);
2204 static void synchronize_mem_cgroup_on_move(struct mem_cgroup *memcg, int cpu)
2206 int idx = MEM_CGROUP_ON_MOVE;
2208 spin_lock(&memcg->pcp_counter_lock);
2209 per_cpu(memcg->stat->count[idx], cpu) = memcg->nocpu_base.count[idx];
2210 spin_unlock(&memcg->pcp_counter_lock);
2213 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2214 unsigned long action,
2217 int cpu = (unsigned long)hcpu;
2218 struct memcg_stock_pcp *stock;
2219 struct mem_cgroup *iter;
2221 if ((action == CPU_ONLINE)) {
2222 for_each_mem_cgroup(iter)
2223 synchronize_mem_cgroup_on_move(iter, cpu);
2227 if ((action != CPU_DEAD) || action != CPU_DEAD_FROZEN)
2230 for_each_mem_cgroup(iter)
2231 mem_cgroup_drain_pcp_counter(iter, cpu);
2233 stock = &per_cpu(memcg_stock, cpu);
2239 /* See __mem_cgroup_try_charge() for details */
2241 CHARGE_OK, /* success */
2242 CHARGE_RETRY, /* need to retry but retry is not bad */
2243 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2244 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2245 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2248 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2249 unsigned int nr_pages, bool oom_check)
2251 unsigned long csize = nr_pages * PAGE_SIZE;
2252 struct mem_cgroup *mem_over_limit;
2253 struct res_counter *fail_res;
2254 unsigned long flags = 0;
2257 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2260 if (!do_swap_account)
2262 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2266 res_counter_uncharge(&memcg->res, csize);
2267 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2268 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2270 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2272 * nr_pages can be either a huge page (HPAGE_PMD_NR), a batch
2273 * of regular pages (CHARGE_BATCH), or a single regular page (1).
2275 * Never reclaim on behalf of optional batching, retry with a
2276 * single page instead.
2278 if (nr_pages == CHARGE_BATCH)
2279 return CHARGE_RETRY;
2281 if (!(gfp_mask & __GFP_WAIT))
2282 return CHARGE_WOULDBLOCK;
2284 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2285 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2286 return CHARGE_RETRY;
2288 * Even though the limit is exceeded at this point, reclaim
2289 * may have been able to free some pages. Retry the charge
2290 * before killing the task.
2292 * Only for regular pages, though: huge pages are rather
2293 * unlikely to succeed so close to the limit, and we fall back
2294 * to regular pages anyway in case of failure.
2296 if (nr_pages == 1 && ret)
2297 return CHARGE_RETRY;
2300 * At task move, charge accounts can be doubly counted. So, it's
2301 * better to wait until the end of task_move if something is going on.
2303 if (mem_cgroup_wait_acct_move(mem_over_limit))
2304 return CHARGE_RETRY;
2306 /* If we don't need to call oom-killer at el, return immediately */
2308 return CHARGE_NOMEM;
2310 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask))
2311 return CHARGE_OOM_DIE;
2313 return CHARGE_RETRY;
2317 * Unlike exported interface, "oom" parameter is added. if oom==true,
2318 * oom-killer can be invoked.
2320 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2322 unsigned int nr_pages,
2323 struct mem_cgroup **ptr,
2326 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2327 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2328 struct mem_cgroup *memcg = NULL;
2332 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2333 * in system level. So, allow to go ahead dying process in addition to
2336 if (unlikely(test_thread_flag(TIF_MEMDIE)
2337 || fatal_signal_pending(current)))
2341 * We always charge the cgroup the mm_struct belongs to.
2342 * The mm_struct's mem_cgroup changes on task migration if the
2343 * thread group leader migrates. It's possible that mm is not
2344 * set, if so charge the init_mm (happens for pagecache usage).
2349 if (*ptr) { /* css should be a valid one */
2351 VM_BUG_ON(css_is_removed(&memcg->css));
2352 if (mem_cgroup_is_root(memcg))
2354 if (nr_pages == 1 && consume_stock(memcg))
2356 css_get(&memcg->css);
2358 struct task_struct *p;
2361 p = rcu_dereference(mm->owner);
2363 * Because we don't have task_lock(), "p" can exit.
2364 * In that case, "memcg" can point to root or p can be NULL with
2365 * race with swapoff. Then, we have small risk of mis-accouning.
2366 * But such kind of mis-account by race always happens because
2367 * we don't have cgroup_mutex(). It's overkill and we allo that
2369 * (*) swapoff at el will charge against mm-struct not against
2370 * task-struct. So, mm->owner can be NULL.
2372 memcg = mem_cgroup_from_task(p);
2373 if (!memcg || mem_cgroup_is_root(memcg)) {
2377 if (nr_pages == 1 && consume_stock(memcg)) {
2379 * It seems dagerous to access memcg without css_get().
2380 * But considering how consume_stok works, it's not
2381 * necessary. If consume_stock success, some charges
2382 * from this memcg are cached on this cpu. So, we
2383 * don't need to call css_get()/css_tryget() before
2384 * calling consume_stock().
2389 /* after here, we may be blocked. we need to get refcnt */
2390 if (!css_tryget(&memcg->css)) {
2400 /* If killed, bypass charge */
2401 if (fatal_signal_pending(current)) {
2402 css_put(&memcg->css);
2407 if (oom && !nr_oom_retries) {
2409 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2412 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, oom_check);
2416 case CHARGE_RETRY: /* not in OOM situation but retry */
2418 css_put(&memcg->css);
2421 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2422 css_put(&memcg->css);
2424 case CHARGE_NOMEM: /* OOM routine works */
2426 css_put(&memcg->css);
2429 /* If oom, we never return -ENOMEM */
2432 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2433 css_put(&memcg->css);
2436 } while (ret != CHARGE_OK);
2438 if (batch > nr_pages)
2439 refill_stock(memcg, batch - nr_pages);
2440 css_put(&memcg->css);
2453 * Somemtimes we have to undo a charge we got by try_charge().
2454 * This function is for that and do uncharge, put css's refcnt.
2455 * gotten by try_charge().
2457 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2458 unsigned int nr_pages)
2460 if (!mem_cgroup_is_root(memcg)) {
2461 unsigned long bytes = nr_pages * PAGE_SIZE;
2463 res_counter_uncharge(&memcg->res, bytes);
2464 if (do_swap_account)
2465 res_counter_uncharge(&memcg->memsw, bytes);
2470 * A helper function to get mem_cgroup from ID. must be called under
2471 * rcu_read_lock(). The caller must check css_is_removed() or some if
2472 * it's concern. (dropping refcnt from swap can be called against removed
2475 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2477 struct cgroup_subsys_state *css;
2479 /* ID 0 is unused ID */
2482 css = css_lookup(&mem_cgroup_subsys, id);
2485 return container_of(css, struct mem_cgroup, css);
2488 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2490 struct mem_cgroup *memcg = NULL;
2491 struct page_cgroup *pc;
2495 VM_BUG_ON(!PageLocked(page));
2497 pc = lookup_page_cgroup(page);
2498 lock_page_cgroup(pc);
2499 if (PageCgroupUsed(pc)) {
2500 memcg = pc->mem_cgroup;
2501 if (memcg && !css_tryget(&memcg->css))
2503 } else if (PageSwapCache(page)) {
2504 ent.val = page_private(page);
2505 id = lookup_swap_cgroup(ent);
2507 memcg = mem_cgroup_lookup(id);
2508 if (memcg && !css_tryget(&memcg->css))
2512 unlock_page_cgroup(pc);
2516 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2518 unsigned int nr_pages,
2519 struct page_cgroup *pc,
2520 enum charge_type ctype)
2522 lock_page_cgroup(pc);
2523 if (unlikely(PageCgroupUsed(pc))) {
2524 unlock_page_cgroup(pc);
2525 __mem_cgroup_cancel_charge(memcg, nr_pages);
2529 * we don't need page_cgroup_lock about tail pages, becase they are not
2530 * accessed by any other context at this point.
2532 pc->mem_cgroup = memcg;
2534 * We access a page_cgroup asynchronously without lock_page_cgroup().
2535 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2536 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2537 * before USED bit, we need memory barrier here.
2538 * See mem_cgroup_add_lru_list(), etc.
2542 case MEM_CGROUP_CHARGE_TYPE_CACHE:
2543 case MEM_CGROUP_CHARGE_TYPE_SHMEM:
2544 SetPageCgroupCache(pc);
2545 SetPageCgroupUsed(pc);
2547 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2548 ClearPageCgroupCache(pc);
2549 SetPageCgroupUsed(pc);
2555 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), nr_pages);
2556 unlock_page_cgroup(pc);
2558 * "charge_statistics" updated event counter. Then, check it.
2559 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2560 * if they exceeds softlimit.
2562 memcg_check_events(memcg, page);
2565 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2567 #define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MOVE_LOCK) |\
2568 (1 << PCG_ACCT_LRU) | (1 << PCG_MIGRATION))
2570 * Because tail pages are not marked as "used", set it. We're under
2571 * zone->lru_lock, 'splitting on pmd' and compund_lock.
2573 void mem_cgroup_split_huge_fixup(struct page *head, struct page *tail)
2575 struct page_cgroup *head_pc = lookup_page_cgroup(head);
2576 struct page_cgroup *tail_pc = lookup_page_cgroup(tail);
2577 unsigned long flags;
2579 if (mem_cgroup_disabled())
2582 * We have no races with charge/uncharge but will have races with
2583 * page state accounting.
2585 move_lock_page_cgroup(head_pc, &flags);
2587 tail_pc->mem_cgroup = head_pc->mem_cgroup;
2588 smp_wmb(); /* see __commit_charge() */
2589 if (PageCgroupAcctLRU(head_pc)) {
2591 struct mem_cgroup_per_zone *mz;
2594 * LRU flags cannot be copied because we need to add tail
2595 *.page to LRU by generic call and our hook will be called.
2596 * We hold lru_lock, then, reduce counter directly.
2598 lru = page_lru(head);
2599 mz = page_cgroup_zoneinfo(head_pc->mem_cgroup, head);
2600 MEM_CGROUP_ZSTAT(mz, lru) -= 1;
2602 tail_pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2603 move_unlock_page_cgroup(head_pc, &flags);
2608 * mem_cgroup_move_account - move account of the page
2610 * @nr_pages: number of regular pages (>1 for huge pages)
2611 * @pc: page_cgroup of the page.
2612 * @from: mem_cgroup which the page is moved from.
2613 * @to: mem_cgroup which the page is moved to. @from != @to.
2614 * @uncharge: whether we should call uncharge and css_put against @from.
2616 * The caller must confirm following.
2617 * - page is not on LRU (isolate_page() is useful.)
2618 * - compound_lock is held when nr_pages > 1
2620 * This function doesn't do "charge" nor css_get to new cgroup. It should be
2621 * done by a caller(__mem_cgroup_try_charge would be useful). If @uncharge is
2622 * true, this function does "uncharge" from old cgroup, but it doesn't if
2623 * @uncharge is false, so a caller should do "uncharge".
2625 static int mem_cgroup_move_account(struct page *page,
2626 unsigned int nr_pages,
2627 struct page_cgroup *pc,
2628 struct mem_cgroup *from,
2629 struct mem_cgroup *to,
2632 unsigned long flags;
2635 VM_BUG_ON(from == to);
2636 VM_BUG_ON(PageLRU(page));
2638 * The page is isolated from LRU. So, collapse function
2639 * will not handle this page. But page splitting can happen.
2640 * Do this check under compound_page_lock(). The caller should
2644 if (nr_pages > 1 && !PageTransHuge(page))
2647 lock_page_cgroup(pc);
2650 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
2653 move_lock_page_cgroup(pc, &flags);
2655 if (PageCgroupFileMapped(pc)) {
2656 /* Update mapped_file data for mem_cgroup */
2658 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2659 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2662 mem_cgroup_charge_statistics(from, PageCgroupCache(pc), -nr_pages);
2664 /* This is not "cancel", but cancel_charge does all we need. */
2665 __mem_cgroup_cancel_charge(from, nr_pages);
2667 /* caller should have done css_get */
2668 pc->mem_cgroup = to;
2669 mem_cgroup_charge_statistics(to, PageCgroupCache(pc), nr_pages);
2671 * We charges against "to" which may not have any tasks. Then, "to"
2672 * can be under rmdir(). But in current implementation, caller of
2673 * this function is just force_empty() and move charge, so it's
2674 * guaranteed that "to" is never removed. So, we don't check rmdir
2677 move_unlock_page_cgroup(pc, &flags);
2680 unlock_page_cgroup(pc);
2684 memcg_check_events(to, page);
2685 memcg_check_events(from, page);
2691 * move charges to its parent.
2694 static int mem_cgroup_move_parent(struct page *page,
2695 struct page_cgroup *pc,
2696 struct mem_cgroup *child,
2699 struct cgroup *cg = child->css.cgroup;
2700 struct cgroup *pcg = cg->parent;
2701 struct mem_cgroup *parent;
2702 unsigned int nr_pages;
2703 unsigned long uninitialized_var(flags);
2711 if (!get_page_unless_zero(page))
2713 if (isolate_lru_page(page))
2716 nr_pages = hpage_nr_pages(page);
2718 parent = mem_cgroup_from_cont(pcg);
2719 ret = __mem_cgroup_try_charge(NULL, gfp_mask, nr_pages, &parent, false);
2724 flags = compound_lock_irqsave(page);
2726 ret = mem_cgroup_move_account(page, nr_pages, pc, child, parent, true);
2728 __mem_cgroup_cancel_charge(parent, nr_pages);
2731 compound_unlock_irqrestore(page, flags);
2733 putback_lru_page(page);
2741 * Charge the memory controller for page usage.
2743 * 0 if the charge was successful
2744 * < 0 if the cgroup is over its limit
2746 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2747 gfp_t gfp_mask, enum charge_type ctype)
2749 struct mem_cgroup *memcg = NULL;
2750 unsigned int nr_pages = 1;
2751 struct page_cgroup *pc;
2755 if (PageTransHuge(page)) {
2756 nr_pages <<= compound_order(page);
2757 VM_BUG_ON(!PageTransHuge(page));
2759 * Never OOM-kill a process for a huge page. The
2760 * fault handler will fall back to regular pages.
2765 pc = lookup_page_cgroup(page);
2766 BUG_ON(!pc); /* XXX: remove this and move pc lookup into commit */
2768 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
2772 __mem_cgroup_commit_charge(memcg, page, nr_pages, pc, ctype);
2776 int mem_cgroup_newpage_charge(struct page *page,
2777 struct mm_struct *mm, gfp_t gfp_mask)
2779 if (mem_cgroup_disabled())
2782 * If already mapped, we don't have to account.
2783 * If page cache, page->mapping has address_space.
2784 * But page->mapping may have out-of-use anon_vma pointer,
2785 * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
2788 if (page_mapped(page) || (page->mapping && !PageAnon(page)))
2792 return mem_cgroup_charge_common(page, mm, gfp_mask,
2793 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2797 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2798 enum charge_type ctype);
2801 __mem_cgroup_commit_charge_lrucare(struct page *page, struct mem_cgroup *memcg,
2802 enum charge_type ctype)
2804 struct page_cgroup *pc = lookup_page_cgroup(page);
2806 * In some case, SwapCache, FUSE(splice_buf->radixtree), the page
2807 * is already on LRU. It means the page may on some other page_cgroup's
2808 * LRU. Take care of it.
2810 mem_cgroup_lru_del_before_commit(page);
2811 __mem_cgroup_commit_charge(memcg, page, 1, pc, ctype);
2812 mem_cgroup_lru_add_after_commit(page);
2816 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2819 struct mem_cgroup *memcg = NULL;
2822 if (mem_cgroup_disabled())
2824 if (PageCompound(page))
2830 if (page_is_file_cache(page)) {
2831 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, &memcg, true);
2836 * FUSE reuses pages without going through the final
2837 * put that would remove them from the LRU list, make
2838 * sure that they get relinked properly.
2840 __mem_cgroup_commit_charge_lrucare(page, memcg,
2841 MEM_CGROUP_CHARGE_TYPE_CACHE);
2845 if (PageSwapCache(page)) {
2846 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &memcg);
2848 __mem_cgroup_commit_charge_swapin(page, memcg,
2849 MEM_CGROUP_CHARGE_TYPE_SHMEM);
2851 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
2852 MEM_CGROUP_CHARGE_TYPE_SHMEM);
2858 * While swap-in, try_charge -> commit or cancel, the page is locked.
2859 * And when try_charge() successfully returns, one refcnt to memcg without
2860 * struct page_cgroup is acquired. This refcnt will be consumed by
2861 * "commit()" or removed by "cancel()"
2863 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2865 gfp_t mask, struct mem_cgroup **ptr)
2867 struct mem_cgroup *memcg;
2872 if (mem_cgroup_disabled())
2875 if (!do_swap_account)
2878 * A racing thread's fault, or swapoff, may have already updated
2879 * the pte, and even removed page from swap cache: in those cases
2880 * do_swap_page()'s pte_same() test will fail; but there's also a
2881 * KSM case which does need to charge the page.
2883 if (!PageSwapCache(page))
2885 memcg = try_get_mem_cgroup_from_page(page);
2889 ret = __mem_cgroup_try_charge(NULL, mask, 1, ptr, true);
2890 css_put(&memcg->css);
2895 return __mem_cgroup_try_charge(mm, mask, 1, ptr, true);
2899 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2900 enum charge_type ctype)
2902 if (mem_cgroup_disabled())
2906 cgroup_exclude_rmdir(&ptr->css);
2908 __mem_cgroup_commit_charge_lrucare(page, ptr, ctype);
2910 * Now swap is on-memory. This means this page may be
2911 * counted both as mem and swap....double count.
2912 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2913 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2914 * may call delete_from_swap_cache() before reach here.
2916 if (do_swap_account && PageSwapCache(page)) {
2917 swp_entry_t ent = {.val = page_private(page)};
2919 struct mem_cgroup *memcg;
2921 id = swap_cgroup_record(ent, 0);
2923 memcg = mem_cgroup_lookup(id);
2926 * This recorded memcg can be obsolete one. So, avoid
2927 * calling css_tryget
2929 if (!mem_cgroup_is_root(memcg))
2930 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2931 mem_cgroup_swap_statistics(memcg, false);
2932 mem_cgroup_put(memcg);
2937 * At swapin, we may charge account against cgroup which has no tasks.
2938 * So, rmdir()->pre_destroy() can be called while we do this charge.
2939 * In that case, we need to call pre_destroy() again. check it here.
2941 cgroup_release_and_wakeup_rmdir(&ptr->css);
2944 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
2946 __mem_cgroup_commit_charge_swapin(page, ptr,
2947 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2950 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
2952 if (mem_cgroup_disabled())
2956 __mem_cgroup_cancel_charge(memcg, 1);
2959 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
2960 unsigned int nr_pages,
2961 const enum charge_type ctype)
2963 struct memcg_batch_info *batch = NULL;
2964 bool uncharge_memsw = true;
2966 /* If swapout, usage of swap doesn't decrease */
2967 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2968 uncharge_memsw = false;
2970 batch = ¤t->memcg_batch;
2972 * In usual, we do css_get() when we remember memcg pointer.
2973 * But in this case, we keep res->usage until end of a series of
2974 * uncharges. Then, it's ok to ignore memcg's refcnt.
2977 batch->memcg = memcg;
2979 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2980 * In those cases, all pages freed continuously can be expected to be in
2981 * the same cgroup and we have chance to coalesce uncharges.
2982 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2983 * because we want to do uncharge as soon as possible.
2986 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
2987 goto direct_uncharge;
2990 goto direct_uncharge;
2993 * In typical case, batch->memcg == mem. This means we can
2994 * merge a series of uncharges to an uncharge of res_counter.
2995 * If not, we uncharge res_counter ony by one.
2997 if (batch->memcg != memcg)
2998 goto direct_uncharge;
2999 /* remember freed charge and uncharge it later */
3002 batch->memsw_nr_pages++;
3005 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
3007 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
3008 if (unlikely(batch->memcg != memcg))
3009 memcg_oom_recover(memcg);
3014 * uncharge if !page_mapped(page)
3016 static struct mem_cgroup *
3017 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
3019 struct mem_cgroup *memcg = NULL;
3020 unsigned int nr_pages = 1;
3021 struct page_cgroup *pc;
3023 if (mem_cgroup_disabled())
3026 if (PageSwapCache(page))
3029 if (PageTransHuge(page)) {
3030 nr_pages <<= compound_order(page);
3031 VM_BUG_ON(!PageTransHuge(page));
3034 * Check if our page_cgroup is valid
3036 pc = lookup_page_cgroup(page);
3037 if (unlikely(!pc || !PageCgroupUsed(pc)))
3040 lock_page_cgroup(pc);
3042 memcg = pc->mem_cgroup;
3044 if (!PageCgroupUsed(pc))
3048 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
3049 case MEM_CGROUP_CHARGE_TYPE_DROP:
3050 /* See mem_cgroup_prepare_migration() */
3051 if (page_mapped(page) || PageCgroupMigration(pc))
3054 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
3055 if (!PageAnon(page)) { /* Shared memory */
3056 if (page->mapping && !page_is_file_cache(page))
3058 } else if (page_mapped(page)) /* Anon */
3065 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), -nr_pages);
3067 ClearPageCgroupUsed(pc);
3069 * pc->mem_cgroup is not cleared here. It will be accessed when it's
3070 * freed from LRU. This is safe because uncharged page is expected not
3071 * to be reused (freed soon). Exception is SwapCache, it's handled by
3072 * special functions.
3075 unlock_page_cgroup(pc);
3077 * even after unlock, we have memcg->res.usage here and this memcg
3078 * will never be freed.
3080 memcg_check_events(memcg, page);
3081 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
3082 mem_cgroup_swap_statistics(memcg, true);
3083 mem_cgroup_get(memcg);
3085 if (!mem_cgroup_is_root(memcg))
3086 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
3091 unlock_page_cgroup(pc);
3095 void mem_cgroup_uncharge_page(struct page *page)
3098 if (page_mapped(page))
3100 if (page->mapping && !PageAnon(page))
3102 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
3105 void mem_cgroup_uncharge_cache_page(struct page *page)
3107 VM_BUG_ON(page_mapped(page));
3108 VM_BUG_ON(page->mapping);
3109 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
3113 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
3114 * In that cases, pages are freed continuously and we can expect pages
3115 * are in the same memcg. All these calls itself limits the number of
3116 * pages freed at once, then uncharge_start/end() is called properly.
3117 * This may be called prural(2) times in a context,
3120 void mem_cgroup_uncharge_start(void)
3122 current->memcg_batch.do_batch++;
3123 /* We can do nest. */
3124 if (current->memcg_batch.do_batch == 1) {
3125 current->memcg_batch.memcg = NULL;
3126 current->memcg_batch.nr_pages = 0;
3127 current->memcg_batch.memsw_nr_pages = 0;
3131 void mem_cgroup_uncharge_end(void)
3133 struct memcg_batch_info *batch = ¤t->memcg_batch;
3135 if (!batch->do_batch)
3139 if (batch->do_batch) /* If stacked, do nothing. */
3145 * This "batch->memcg" is valid without any css_get/put etc...
3146 * bacause we hide charges behind us.
3148 if (batch->nr_pages)
3149 res_counter_uncharge(&batch->memcg->res,
3150 batch->nr_pages * PAGE_SIZE);
3151 if (batch->memsw_nr_pages)
3152 res_counter_uncharge(&batch->memcg->memsw,
3153 batch->memsw_nr_pages * PAGE_SIZE);
3154 memcg_oom_recover(batch->memcg);
3155 /* forget this pointer (for sanity check) */
3156 batch->memcg = NULL;
3161 * called after __delete_from_swap_cache() and drop "page" account.
3162 * memcg information is recorded to swap_cgroup of "ent"
3165 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
3167 struct mem_cgroup *memcg;
3168 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
3170 if (!swapout) /* this was a swap cache but the swap is unused ! */
3171 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
3173 memcg = __mem_cgroup_uncharge_common(page, ctype);
3176 * record memcg information, if swapout && memcg != NULL,
3177 * mem_cgroup_get() was called in uncharge().
3179 if (do_swap_account && swapout && memcg)
3180 swap_cgroup_record(ent, css_id(&memcg->css));
3184 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3186 * called from swap_entry_free(). remove record in swap_cgroup and
3187 * uncharge "memsw" account.
3189 void mem_cgroup_uncharge_swap(swp_entry_t ent)
3191 struct mem_cgroup *memcg;
3194 if (!do_swap_account)
3197 id = swap_cgroup_record(ent, 0);
3199 memcg = mem_cgroup_lookup(id);
3202 * We uncharge this because swap is freed.
3203 * This memcg can be obsolete one. We avoid calling css_tryget
3205 if (!mem_cgroup_is_root(memcg))
3206 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
3207 mem_cgroup_swap_statistics(memcg, false);
3208 mem_cgroup_put(memcg);
3214 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3215 * @entry: swap entry to be moved
3216 * @from: mem_cgroup which the entry is moved from
3217 * @to: mem_cgroup which the entry is moved to
3218 * @need_fixup: whether we should fixup res_counters and refcounts.
3220 * It succeeds only when the swap_cgroup's record for this entry is the same
3221 * as the mem_cgroup's id of @from.
3223 * Returns 0 on success, -EINVAL on failure.
3225 * The caller must have charged to @to, IOW, called res_counter_charge() about
3226 * both res and memsw, and called css_get().
3228 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3229 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3231 unsigned short old_id, new_id;
3233 old_id = css_id(&from->css);
3234 new_id = css_id(&to->css);
3236 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3237 mem_cgroup_swap_statistics(from, false);
3238 mem_cgroup_swap_statistics(to, true);
3240 * This function is only called from task migration context now.
3241 * It postpones res_counter and refcount handling till the end
3242 * of task migration(mem_cgroup_clear_mc()) for performance
3243 * improvement. But we cannot postpone mem_cgroup_get(to)
3244 * because if the process that has been moved to @to does
3245 * swap-in, the refcount of @to might be decreased to 0.
3249 if (!mem_cgroup_is_root(from))
3250 res_counter_uncharge(&from->memsw, PAGE_SIZE);
3251 mem_cgroup_put(from);
3253 * we charged both to->res and to->memsw, so we should
3256 if (!mem_cgroup_is_root(to))
3257 res_counter_uncharge(&to->res, PAGE_SIZE);
3264 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3265 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3272 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
3275 int mem_cgroup_prepare_migration(struct page *page,
3276 struct page *newpage, struct mem_cgroup **ptr, gfp_t gfp_mask)
3278 struct mem_cgroup *memcg = NULL;
3279 struct page_cgroup *pc;
3280 enum charge_type ctype;
3285 VM_BUG_ON(PageTransHuge(page));
3286 if (mem_cgroup_disabled())
3289 pc = lookup_page_cgroup(page);
3290 lock_page_cgroup(pc);
3291 if (PageCgroupUsed(pc)) {
3292 memcg = pc->mem_cgroup;
3293 css_get(&memcg->css);
3295 * At migrating an anonymous page, its mapcount goes down
3296 * to 0 and uncharge() will be called. But, even if it's fully
3297 * unmapped, migration may fail and this page has to be
3298 * charged again. We set MIGRATION flag here and delay uncharge
3299 * until end_migration() is called
3301 * Corner Case Thinking
3303 * When the old page was mapped as Anon and it's unmap-and-freed
3304 * while migration was ongoing.
3305 * If unmap finds the old page, uncharge() of it will be delayed
3306 * until end_migration(). If unmap finds a new page, it's
3307 * uncharged when it make mapcount to be 1->0. If unmap code
3308 * finds swap_migration_entry, the new page will not be mapped
3309 * and end_migration() will find it(mapcount==0).
3312 * When the old page was mapped but migraion fails, the kernel
3313 * remaps it. A charge for it is kept by MIGRATION flag even
3314 * if mapcount goes down to 0. We can do remap successfully
3315 * without charging it again.
3318 * The "old" page is under lock_page() until the end of
3319 * migration, so, the old page itself will not be swapped-out.
3320 * If the new page is swapped out before end_migraton, our
3321 * hook to usual swap-out path will catch the event.
3324 SetPageCgroupMigration(pc);
3326 unlock_page_cgroup(pc);
3328 * If the page is not charged at this point,
3335 ret = __mem_cgroup_try_charge(NULL, gfp_mask, 1, ptr, false);
3336 css_put(&memcg->css);/* drop extra refcnt */
3337 if (ret || *ptr == NULL) {
3338 if (PageAnon(page)) {
3339 lock_page_cgroup(pc);
3340 ClearPageCgroupMigration(pc);
3341 unlock_page_cgroup(pc);
3343 * The old page may be fully unmapped while we kept it.
3345 mem_cgroup_uncharge_page(page);
3350 * We charge new page before it's used/mapped. So, even if unlock_page()
3351 * is called before end_migration, we can catch all events on this new
3352 * page. In the case new page is migrated but not remapped, new page's
3353 * mapcount will be finally 0 and we call uncharge in end_migration().
3355 pc = lookup_page_cgroup(newpage);
3357 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
3358 else if (page_is_file_cache(page))
3359 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
3361 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3362 __mem_cgroup_commit_charge(memcg, page, 1, pc, ctype);
3366 /* remove redundant charge if migration failed*/
3367 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
3368 struct page *oldpage, struct page *newpage, bool migration_ok)
3370 struct page *used, *unused;
3371 struct page_cgroup *pc;
3375 /* blocks rmdir() */
3376 cgroup_exclude_rmdir(&memcg->css);
3377 if (!migration_ok) {
3385 * We disallowed uncharge of pages under migration because mapcount
3386 * of the page goes down to zero, temporarly.
3387 * Clear the flag and check the page should be charged.
3389 pc = lookup_page_cgroup(oldpage);
3390 lock_page_cgroup(pc);
3391 ClearPageCgroupMigration(pc);
3392 unlock_page_cgroup(pc);
3394 __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE);
3397 * If a page is a file cache, radix-tree replacement is very atomic
3398 * and we can skip this check. When it was an Anon page, its mapcount
3399 * goes down to 0. But because we added MIGRATION flage, it's not
3400 * uncharged yet. There are several case but page->mapcount check
3401 * and USED bit check in mem_cgroup_uncharge_page() will do enough
3402 * check. (see prepare_charge() also)
3405 mem_cgroup_uncharge_page(used);
3407 * At migration, we may charge account against cgroup which has no
3409 * So, rmdir()->pre_destroy() can be called while we do this charge.
3410 * In that case, we need to call pre_destroy() again. check it here.
3412 cgroup_release_and_wakeup_rmdir(&memcg->css);
3415 #ifdef CONFIG_DEBUG_VM
3416 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3418 struct page_cgroup *pc;
3420 pc = lookup_page_cgroup(page);
3421 if (likely(pc) && PageCgroupUsed(pc))
3426 bool mem_cgroup_bad_page_check(struct page *page)
3428 if (mem_cgroup_disabled())
3431 return lookup_page_cgroup_used(page) != NULL;
3434 void mem_cgroup_print_bad_page(struct page *page)
3436 struct page_cgroup *pc;
3438 pc = lookup_page_cgroup_used(page);
3443 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p",
3444 pc, pc->flags, pc->mem_cgroup);
3446 path = kmalloc(PATH_MAX, GFP_KERNEL);
3449 ret = cgroup_path(pc->mem_cgroup->css.cgroup,
3454 printk(KERN_CONT "(%s)\n",
3455 (ret < 0) ? "cannot get the path" : path);
3461 static DEFINE_MUTEX(set_limit_mutex);
3463 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3464 unsigned long long val)
3467 u64 memswlimit, memlimit;
3469 int children = mem_cgroup_count_children(memcg);
3470 u64 curusage, oldusage;
3474 * For keeping hierarchical_reclaim simple, how long we should retry
3475 * is depends on callers. We set our retry-count to be function
3476 * of # of children which we should visit in this loop.
3478 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3480 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3483 while (retry_count) {
3484 if (signal_pending(current)) {
3489 * Rather than hide all in some function, I do this in
3490 * open coded manner. You see what this really does.
3491 * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3493 mutex_lock(&set_limit_mutex);
3494 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3495 if (memswlimit < val) {
3497 mutex_unlock(&set_limit_mutex);
3501 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3505 ret = res_counter_set_limit(&memcg->res, val);
3507 if (memswlimit == val)
3508 memcg->memsw_is_minimum = true;
3510 memcg->memsw_is_minimum = false;
3512 mutex_unlock(&set_limit_mutex);
3517 mem_cgroup_reclaim(memcg, GFP_KERNEL,
3518 MEM_CGROUP_RECLAIM_SHRINK);
3519 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3520 /* Usage is reduced ? */
3521 if (curusage >= oldusage)
3524 oldusage = curusage;
3526 if (!ret && enlarge)
3527 memcg_oom_recover(memcg);
3532 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3533 unsigned long long val)
3536 u64 memlimit, memswlimit, oldusage, curusage;
3537 int children = mem_cgroup_count_children(memcg);
3541 /* see mem_cgroup_resize_res_limit */
3542 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3543 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3544 while (retry_count) {
3545 if (signal_pending(current)) {
3550 * Rather than hide all in some function, I do this in
3551 * open coded manner. You see what this really does.
3552 * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3554 mutex_lock(&set_limit_mutex);
3555 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3556 if (memlimit > val) {
3558 mutex_unlock(&set_limit_mutex);
3561 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3562 if (memswlimit < val)
3564 ret = res_counter_set_limit(&memcg->memsw, val);
3566 if (memlimit == val)
3567 memcg->memsw_is_minimum = true;
3569 memcg->memsw_is_minimum = false;
3571 mutex_unlock(&set_limit_mutex);
3576 mem_cgroup_reclaim(memcg, GFP_KERNEL,
3577 MEM_CGROUP_RECLAIM_NOSWAP |
3578 MEM_CGROUP_RECLAIM_SHRINK);
3579 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3580 /* Usage is reduced ? */
3581 if (curusage >= oldusage)
3584 oldusage = curusage;
3586 if (!ret && enlarge)
3587 memcg_oom_recover(memcg);
3591 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3593 unsigned long *total_scanned)
3595 unsigned long nr_reclaimed = 0;
3596 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3597 unsigned long reclaimed;
3599 struct mem_cgroup_tree_per_zone *mctz;
3600 unsigned long long excess;
3601 unsigned long nr_scanned;
3606 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3608 * This loop can run a while, specially if mem_cgroup's continuously
3609 * keep exceeding their soft limit and putting the system under
3616 mz = mem_cgroup_largest_soft_limit_node(mctz);
3621 reclaimed = mem_cgroup_soft_reclaim(mz->mem, zone,
3622 gfp_mask, &nr_scanned);
3623 nr_reclaimed += reclaimed;
3624 *total_scanned += nr_scanned;
3625 spin_lock(&mctz->lock);
3628 * If we failed to reclaim anything from this memory cgroup
3629 * it is time to move on to the next cgroup
3635 * Loop until we find yet another one.
3637 * By the time we get the soft_limit lock
3638 * again, someone might have aded the
3639 * group back on the RB tree. Iterate to
3640 * make sure we get a different mem.
3641 * mem_cgroup_largest_soft_limit_node returns
3642 * NULL if no other cgroup is present on
3646 __mem_cgroup_largest_soft_limit_node(mctz);
3648 css_put(&next_mz->mem->css);
3649 else /* next_mz == NULL or other memcg */
3653 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
3654 excess = res_counter_soft_limit_excess(&mz->mem->res);
3656 * One school of thought says that we should not add
3657 * back the node to the tree if reclaim returns 0.
3658 * But our reclaim could return 0, simply because due
3659 * to priority we are exposing a smaller subset of
3660 * memory to reclaim from. Consider this as a longer
3663 /* If excess == 0, no tree ops */
3664 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess);
3665 spin_unlock(&mctz->lock);
3666 css_put(&mz->mem->css);
3669 * Could not reclaim anything and there are no more
3670 * mem cgroups to try or we seem to be looping without
3671 * reclaiming anything.
3673 if (!nr_reclaimed &&
3675 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3677 } while (!nr_reclaimed);
3679 css_put(&next_mz->mem->css);
3680 return nr_reclaimed;
3684 * This routine traverse page_cgroup in given list and drop them all.
3685 * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3687 static int mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
3688 int node, int zid, enum lru_list lru)
3691 struct mem_cgroup_per_zone *mz;
3692 struct page_cgroup *pc, *busy;
3693 unsigned long flags, loop;
3694 struct list_head *list;
3697 zone = &NODE_DATA(node)->node_zones[zid];
3698 mz = mem_cgroup_zoneinfo(memcg, node, zid);
3699 list = &mz->lruvec.lists[lru];
3701 loop = MEM_CGROUP_ZSTAT(mz, lru);
3702 /* give some margin against EBUSY etc...*/
3709 spin_lock_irqsave(&zone->lru_lock, flags);
3710 if (list_empty(list)) {
3711 spin_unlock_irqrestore(&zone->lru_lock, flags);
3714 pc = list_entry(list->prev, struct page_cgroup, lru);
3716 list_move(&pc->lru, list);
3718 spin_unlock_irqrestore(&zone->lru_lock, flags);
3721 spin_unlock_irqrestore(&zone->lru_lock, flags);
3723 page = lookup_cgroup_page(pc);
3725 ret = mem_cgroup_move_parent(page, pc, memcg, GFP_KERNEL);
3729 if (ret == -EBUSY || ret == -EINVAL) {
3730 /* found lock contention or "pc" is obsolete. */
3737 if (!ret && !list_empty(list))
3743 * make mem_cgroup's charge to be 0 if there is no task.
3744 * This enables deleting this mem_cgroup.
3746 static int mem_cgroup_force_empty(struct mem_cgroup *memcg, bool free_all)
3749 int node, zid, shrink;
3750 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3751 struct cgroup *cgrp = memcg->css.cgroup;
3753 css_get(&memcg->css);
3756 /* should free all ? */
3762 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3765 if (signal_pending(current))
3767 /* This is for making all *used* pages to be on LRU. */
3768 lru_add_drain_all();
3769 drain_all_stock_sync(memcg);
3771 mem_cgroup_start_move(memcg);
3772 for_each_node_state(node, N_HIGH_MEMORY) {
3773 for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3776 ret = mem_cgroup_force_empty_list(memcg,
3785 mem_cgroup_end_move(memcg);
3786 memcg_oom_recover(memcg);
3787 /* it seems parent cgroup doesn't have enough mem */
3791 /* "ret" should also be checked to ensure all lists are empty. */
3792 } while (memcg->res.usage > 0 || ret);
3794 css_put(&memcg->css);
3798 /* returns EBUSY if there is a task or if we come here twice. */
3799 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3803 /* we call try-to-free pages for make this cgroup empty */
3804 lru_add_drain_all();
3805 /* try to free all pages in this cgroup */
3807 while (nr_retries && memcg->res.usage > 0) {
3810 if (signal_pending(current)) {
3814 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
3818 /* maybe some writeback is necessary */
3819 congestion_wait(BLK_RW_ASYNC, HZ/10);
3824 /* try move_account...there may be some *locked* pages. */
3828 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3830 return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3834 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3836 return mem_cgroup_from_cont(cont)->use_hierarchy;
3839 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3843 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3844 struct cgroup *parent = cont->parent;
3845 struct mem_cgroup *parent_memcg = NULL;
3848 parent_memcg = mem_cgroup_from_cont(parent);
3852 * If parent's use_hierarchy is set, we can't make any modifications
3853 * in the child subtrees. If it is unset, then the change can
3854 * occur, provided the current cgroup has no children.
3856 * For the root cgroup, parent_mem is NULL, we allow value to be
3857 * set if there are no children.
3859 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3860 (val == 1 || val == 0)) {
3861 if (list_empty(&cont->children))
3862 memcg->use_hierarchy = val;
3873 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
3874 enum mem_cgroup_stat_index idx)
3876 struct mem_cgroup *iter;
3879 /* Per-cpu values can be negative, use a signed accumulator */
3880 for_each_mem_cgroup_tree(iter, memcg)
3881 val += mem_cgroup_read_stat(iter, idx);
3883 if (val < 0) /* race ? */
3888 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3892 if (!mem_cgroup_is_root(memcg)) {
3894 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_KMEM
3895 if (!memcg->kmem_independent_accounting)
3896 val = res_counter_read_u64(&memcg->kmem, RES_USAGE);
3899 val += res_counter_read_u64(&memcg->res, RES_USAGE);
3901 val += res_counter_read_u64(&memcg->memsw, RES_USAGE);
3906 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
3907 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
3910 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
3912 return val << PAGE_SHIFT;
3915 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
3917 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3921 type = MEMFILE_TYPE(cft->private);
3922 name = MEMFILE_ATTR(cft->private);
3925 if (name == RES_USAGE)
3926 val = mem_cgroup_usage(memcg, false);
3928 val = res_counter_read_u64(&memcg->res, name);
3931 if (name == RES_USAGE)
3932 val = mem_cgroup_usage(memcg, true);
3934 val = res_counter_read_u64(&memcg->memsw, name);
3936 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_KMEM
3938 val = res_counter_read_u64(&memcg->kmem, name);
3948 * The user of this function is...
3951 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
3954 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3956 unsigned long long val;
3959 type = MEMFILE_TYPE(cft->private);
3960 name = MEMFILE_ATTR(cft->private);
3963 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3967 /* This function does all necessary parse...reuse it */
3968 ret = res_counter_memparse_write_strategy(buffer, &val);
3972 ret = mem_cgroup_resize_limit(memcg, val);
3974 ret = mem_cgroup_resize_memsw_limit(memcg, val);
3976 case RES_SOFT_LIMIT:
3977 ret = res_counter_memparse_write_strategy(buffer, &val);
3981 * For memsw, soft limits are hard to implement in terms
3982 * of semantics, for now, we support soft limits for
3983 * control without swap
3986 ret = res_counter_set_soft_limit(&memcg->res, val);
3991 ret = -EINVAL; /* should be BUG() ? */
3997 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
3998 unsigned long long *mem_limit, unsigned long long *memsw_limit)
4000 struct cgroup *cgroup;
4001 unsigned long long min_limit, min_memsw_limit, tmp;
4003 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4004 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4005 cgroup = memcg->css.cgroup;
4006 if (!memcg->use_hierarchy)
4009 while (cgroup->parent) {
4010 cgroup = cgroup->parent;
4011 memcg = mem_cgroup_from_cont(cgroup);
4012 if (!memcg->use_hierarchy)
4014 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
4015 min_limit = min(min_limit, tmp);
4016 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4017 min_memsw_limit = min(min_memsw_limit, tmp);
4020 *mem_limit = min_limit;
4021 *memsw_limit = min_memsw_limit;
4025 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
4027 struct mem_cgroup *memcg;
4030 memcg = mem_cgroup_from_cont(cont);
4031 type = MEMFILE_TYPE(event);
4032 name = MEMFILE_ATTR(event);
4036 res_counter_reset_max(&memcg->res);
4038 res_counter_reset_max(&memcg->memsw);
4042 res_counter_reset_failcnt(&memcg->res);
4044 res_counter_reset_failcnt(&memcg->memsw);
4051 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
4054 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
4058 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4059 struct cftype *cft, u64 val)
4061 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4063 if (val >= (1 << NR_MOVE_TYPE))
4066 * We check this value several times in both in can_attach() and
4067 * attach(), so we need cgroup lock to prevent this value from being
4071 memcg->move_charge_at_immigrate = val;
4077 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4078 struct cftype *cft, u64 val)
4085 /* For read statistics */
4103 struct mcs_total_stat {
4104 s64 stat[NR_MCS_STAT];
4110 } memcg_stat_strings[NR_MCS_STAT] = {
4111 {"cache", "total_cache"},
4112 {"rss", "total_rss"},
4113 {"mapped_file", "total_mapped_file"},
4114 {"pgpgin", "total_pgpgin"},
4115 {"pgpgout", "total_pgpgout"},
4116 {"swap", "total_swap"},
4117 {"pgfault", "total_pgfault"},
4118 {"pgmajfault", "total_pgmajfault"},
4119 {"inactive_anon", "total_inactive_anon"},
4120 {"active_anon", "total_active_anon"},
4121 {"inactive_file", "total_inactive_file"},
4122 {"active_file", "total_active_file"},
4123 {"unevictable", "total_unevictable"}
4128 mem_cgroup_get_local_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4133 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_CACHE);
4134 s->stat[MCS_CACHE] += val * PAGE_SIZE;
4135 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_RSS);
4136 s->stat[MCS_RSS] += val * PAGE_SIZE;
4137 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_FILE_MAPPED);
4138 s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
4139 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGIN);
4140 s->stat[MCS_PGPGIN] += val;
4141 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGOUT);
4142 s->stat[MCS_PGPGOUT] += val;
4143 if (do_swap_account) {
4144 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
4145 s->stat[MCS_SWAP] += val * PAGE_SIZE;
4147 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGFAULT);
4148 s->stat[MCS_PGFAULT] += val;
4149 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGMAJFAULT);
4150 s->stat[MCS_PGMAJFAULT] += val;
4153 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_ANON));
4154 s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
4155 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_ANON));
4156 s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
4157 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_FILE));
4158 s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
4159 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_FILE));
4160 s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
4161 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
4162 s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
4166 mem_cgroup_get_total_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4168 struct mem_cgroup *iter;
4170 for_each_mem_cgroup_tree(iter, memcg)
4171 mem_cgroup_get_local_stat(iter, s);
4175 static int mem_control_numa_stat_show(struct seq_file *m, void *arg)
4178 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
4179 unsigned long node_nr;
4180 struct cgroup *cont = m->private;
4181 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4183 total_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL);
4184 seq_printf(m, "total=%lu", total_nr);
4185 for_each_node_state(nid, N_HIGH_MEMORY) {
4186 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid, LRU_ALL);
4187 seq_printf(m, " N%d=%lu", nid, node_nr);
4191 file_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_FILE);
4192 seq_printf(m, "file=%lu", file_nr);
4193 for_each_node_state(nid, N_HIGH_MEMORY) {
4194 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4196 seq_printf(m, " N%d=%lu", nid, node_nr);
4200 anon_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_ANON);
4201 seq_printf(m, "anon=%lu", anon_nr);
4202 for_each_node_state(nid, N_HIGH_MEMORY) {
4203 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4205 seq_printf(m, " N%d=%lu", nid, node_nr);
4209 unevictable_nr = mem_cgroup_nr_lru_pages(mem_cont, BIT(LRU_UNEVICTABLE));
4210 seq_printf(m, "unevictable=%lu", unevictable_nr);
4211 for_each_node_state(nid, N_HIGH_MEMORY) {
4212 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4213 BIT(LRU_UNEVICTABLE));
4214 seq_printf(m, " N%d=%lu", nid, node_nr);
4219 #endif /* CONFIG_NUMA */
4221 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
4222 struct cgroup_map_cb *cb)
4224 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4225 struct mcs_total_stat mystat;
4228 memset(&mystat, 0, sizeof(mystat));
4229 mem_cgroup_get_local_stat(mem_cont, &mystat);
4232 for (i = 0; i < NR_MCS_STAT; i++) {
4233 if (i == MCS_SWAP && !do_swap_account)
4235 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
4238 /* Hierarchical information */
4240 unsigned long long limit, memsw_limit;
4241 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
4242 cb->fill(cb, "hierarchical_memory_limit", limit);
4243 if (do_swap_account)
4244 cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
4247 memset(&mystat, 0, sizeof(mystat));
4248 mem_cgroup_get_total_stat(mem_cont, &mystat);
4249 for (i = 0; i < NR_MCS_STAT; i++) {
4250 if (i == MCS_SWAP && !do_swap_account)
4252 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
4255 #ifdef CONFIG_DEBUG_VM
4258 struct mem_cgroup_per_zone *mz;
4259 unsigned long recent_rotated[2] = {0, 0};
4260 unsigned long recent_scanned[2] = {0, 0};
4262 for_each_online_node(nid)
4263 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4264 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
4266 recent_rotated[0] +=
4267 mz->reclaim_stat.recent_rotated[0];
4268 recent_rotated[1] +=
4269 mz->reclaim_stat.recent_rotated[1];
4270 recent_scanned[0] +=
4271 mz->reclaim_stat.recent_scanned[0];
4272 recent_scanned[1] +=
4273 mz->reclaim_stat.recent_scanned[1];
4275 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
4276 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
4277 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
4278 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
4285 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
4287 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4289 return mem_cgroup_swappiness(memcg);
4292 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
4295 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4296 struct mem_cgroup *parent;
4301 if (cgrp->parent == NULL)
4304 parent = mem_cgroup_from_cont(cgrp->parent);
4308 /* If under hierarchy, only empty-root can set this value */
4309 if ((parent->use_hierarchy) ||
4310 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4315 memcg->swappiness = val;
4322 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4324 struct mem_cgroup_threshold_ary *t;
4330 t = rcu_dereference(memcg->thresholds.primary);
4332 t = rcu_dereference(memcg->memsw_thresholds.primary);
4337 usage = mem_cgroup_usage(memcg, swap);
4340 * current_threshold points to threshold just below usage.
4341 * If it's not true, a threshold was crossed after last
4342 * call of __mem_cgroup_threshold().
4344 i = t->current_threshold;
4347 * Iterate backward over array of thresholds starting from
4348 * current_threshold and check if a threshold is crossed.
4349 * If none of thresholds below usage is crossed, we read
4350 * only one element of the array here.
4352 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4353 eventfd_signal(t->entries[i].eventfd, 1);
4355 /* i = current_threshold + 1 */
4359 * Iterate forward over array of thresholds starting from
4360 * current_threshold+1 and check if a threshold is crossed.
4361 * If none of thresholds above usage is crossed, we read
4362 * only one element of the array here.
4364 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4365 eventfd_signal(t->entries[i].eventfd, 1);
4367 /* Update current_threshold */
4368 t->current_threshold = i - 1;
4373 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4376 __mem_cgroup_threshold(memcg, false);
4377 if (do_swap_account)
4378 __mem_cgroup_threshold(memcg, true);
4380 memcg = parent_mem_cgroup(memcg);
4384 static int compare_thresholds(const void *a, const void *b)
4386 const struct mem_cgroup_threshold *_a = a;
4387 const struct mem_cgroup_threshold *_b = b;
4389 return _a->threshold - _b->threshold;
4392 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4394 struct mem_cgroup_eventfd_list *ev;
4396 list_for_each_entry(ev, &memcg->oom_notify, list)
4397 eventfd_signal(ev->eventfd, 1);
4401 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4403 struct mem_cgroup *iter;
4405 for_each_mem_cgroup_tree(iter, memcg)
4406 mem_cgroup_oom_notify_cb(iter);
4409 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
4410 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4412 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4413 struct mem_cgroup_thresholds *thresholds;
4414 struct mem_cgroup_threshold_ary *new;
4415 int type = MEMFILE_TYPE(cft->private);
4416 u64 threshold, usage;
4419 ret = res_counter_memparse_write_strategy(args, &threshold);
4423 mutex_lock(&memcg->thresholds_lock);
4426 thresholds = &memcg->thresholds;
4427 else if (type == _MEMSWAP)
4428 thresholds = &memcg->memsw_thresholds;
4432 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4434 /* Check if a threshold crossed before adding a new one */
4435 if (thresholds->primary)
4436 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4438 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4440 /* Allocate memory for new array of thresholds */
4441 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4449 /* Copy thresholds (if any) to new array */
4450 if (thresholds->primary) {
4451 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4452 sizeof(struct mem_cgroup_threshold));
4455 /* Add new threshold */
4456 new->entries[size - 1].eventfd = eventfd;
4457 new->entries[size - 1].threshold = threshold;
4459 /* Sort thresholds. Registering of new threshold isn't time-critical */
4460 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4461 compare_thresholds, NULL);
4463 /* Find current threshold */
4464 new->current_threshold = -1;
4465 for (i = 0; i < size; i++) {
4466 if (new->entries[i].threshold < usage) {
4468 * new->current_threshold will not be used until
4469 * rcu_assign_pointer(), so it's safe to increment
4472 ++new->current_threshold;
4476 /* Free old spare buffer and save old primary buffer as spare */
4477 kfree(thresholds->spare);
4478 thresholds->spare = thresholds->primary;
4480 rcu_assign_pointer(thresholds->primary, new);
4482 /* To be sure that nobody uses thresholds */
4486 mutex_unlock(&memcg->thresholds_lock);
4491 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
4492 struct cftype *cft, struct eventfd_ctx *eventfd)
4494 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4495 struct mem_cgroup_thresholds *thresholds;
4496 struct mem_cgroup_threshold_ary *new;
4497 int type = MEMFILE_TYPE(cft->private);
4501 mutex_lock(&memcg->thresholds_lock);
4503 thresholds = &memcg->thresholds;
4504 else if (type == _MEMSWAP)
4505 thresholds = &memcg->memsw_thresholds;
4510 * Something went wrong if we trying to unregister a threshold
4511 * if we don't have thresholds
4513 BUG_ON(!thresholds);
4515 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4517 /* Check if a threshold crossed before removing */
4518 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4520 /* Calculate new number of threshold */
4522 for (i = 0; i < thresholds->primary->size; i++) {
4523 if (thresholds->primary->entries[i].eventfd != eventfd)
4527 new = thresholds->spare;
4529 /* Set thresholds array to NULL if we don't have thresholds */
4538 /* Copy thresholds and find current threshold */
4539 new->current_threshold = -1;
4540 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4541 if (thresholds->primary->entries[i].eventfd == eventfd)
4544 new->entries[j] = thresholds->primary->entries[i];
4545 if (new->entries[j].threshold < usage) {
4547 * new->current_threshold will not be used
4548 * until rcu_assign_pointer(), so it's safe to increment
4551 ++new->current_threshold;
4557 /* Swap primary and spare array */
4558 thresholds->spare = thresholds->primary;
4559 rcu_assign_pointer(thresholds->primary, new);
4561 /* To be sure that nobody uses thresholds */
4564 mutex_unlock(&memcg->thresholds_lock);
4567 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4568 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4570 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4571 struct mem_cgroup_eventfd_list *event;
4572 int type = MEMFILE_TYPE(cft->private);
4574 BUG_ON(type != _OOM_TYPE);
4575 event = kmalloc(sizeof(*event), GFP_KERNEL);
4579 spin_lock(&memcg_oom_lock);
4581 event->eventfd = eventfd;
4582 list_add(&event->list, &memcg->oom_notify);
4584 /* already in OOM ? */
4585 if (atomic_read(&memcg->under_oom))
4586 eventfd_signal(eventfd, 1);
4587 spin_unlock(&memcg_oom_lock);
4592 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4593 struct cftype *cft, struct eventfd_ctx *eventfd)
4595 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4596 struct mem_cgroup_eventfd_list *ev, *tmp;
4597 int type = MEMFILE_TYPE(cft->private);
4599 BUG_ON(type != _OOM_TYPE);
4601 spin_lock(&memcg_oom_lock);
4603 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4604 if (ev->eventfd == eventfd) {
4605 list_del(&ev->list);
4610 spin_unlock(&memcg_oom_lock);
4613 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4614 struct cftype *cft, struct cgroup_map_cb *cb)
4616 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4618 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
4620 if (atomic_read(&memcg->under_oom))
4621 cb->fill(cb, "under_oom", 1);
4623 cb->fill(cb, "under_oom", 0);
4627 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4628 struct cftype *cft, u64 val)
4630 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4631 struct mem_cgroup *parent;
4633 /* cannot set to root cgroup and only 0 and 1 are allowed */
4634 if (!cgrp->parent || !((val == 0) || (val == 1)))
4637 parent = mem_cgroup_from_cont(cgrp->parent);
4640 /* oom-kill-disable is a flag for subhierarchy. */
4641 if ((parent->use_hierarchy) ||
4642 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4646 memcg->oom_kill_disable = val;
4648 memcg_oom_recover(memcg);
4654 static const struct file_operations mem_control_numa_stat_file_operations = {
4656 .llseek = seq_lseek,
4657 .release = single_release,
4660 static int mem_control_numa_stat_open(struct inode *unused, struct file *file)
4662 struct cgroup *cont = file->f_dentry->d_parent->d_fsdata;
4664 file->f_op = &mem_control_numa_stat_file_operations;
4665 return single_open(file, mem_control_numa_stat_show, cont);
4667 #endif /* CONFIG_NUMA */
4669 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_KMEM
4670 static u64 kmem_limit_independent_read(struct cgroup *cgroup, struct cftype *cft)
4672 return mem_cgroup_from_cont(cgroup)->kmem_independent_accounting;
4675 static int kmem_limit_independent_write(struct cgroup *cgroup, struct cftype *cft,
4678 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup);
4679 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4684 * This follows the same hierarchy restrictions than
4685 * mem_cgroup_hierarchy_write()
4687 if (!parent || !parent->use_hierarchy) {
4688 if (list_empty(&cgroup->children))
4689 memcg->kmem_independent_accounting = val;
4698 static struct cftype kmem_cgroup_files[] = {
4700 .name = "independent_kmem_limit",
4701 .read_u64 = kmem_limit_independent_read,
4702 .write_u64 = kmem_limit_independent_write,
4705 .name = "kmem.usage_in_bytes",
4706 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4707 .read_u64 = mem_cgroup_read,
4710 .name = "kmem.limit_in_bytes",
4711 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4712 .read_u64 = mem_cgroup_read,
4716 static int register_kmem_files(struct cgroup *cont, struct cgroup_subsys *ss)
4720 ret = cgroup_add_files(cont, ss, kmem_cgroup_files,
4721 ARRAY_SIZE(kmem_cgroup_files));
4724 * Part of this would be better living in a separate allocation
4725 * function, leaving us with just the cgroup tree population work.
4726 * We, however, depend on state such as network's proto_list that
4727 * is only initialized after cgroup creation. I found the less
4728 * cumbersome way to deal with it to defer it all to populate time
4731 ret = mem_cgroup_sockets_init(cont, ss);
4735 static void kmem_cgroup_destroy(struct cgroup_subsys *ss,
4736 struct cgroup *cont)
4738 mem_cgroup_sockets_destroy(cont, ss);
4741 static int register_kmem_files(struct cgroup *cont, struct cgroup_subsys *ss)
4746 static void kmem_cgroup_destroy(struct cgroup_subsys *ss,
4747 struct cgroup *cont)
4752 static struct cftype mem_cgroup_files[] = {
4754 .name = "usage_in_bytes",
4755 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4756 .read_u64 = mem_cgroup_read,
4757 .register_event = mem_cgroup_usage_register_event,
4758 .unregister_event = mem_cgroup_usage_unregister_event,
4761 .name = "max_usage_in_bytes",
4762 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4763 .trigger = mem_cgroup_reset,
4764 .read_u64 = mem_cgroup_read,
4767 .name = "limit_in_bytes",
4768 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4769 .write_string = mem_cgroup_write,
4770 .read_u64 = mem_cgroup_read,
4773 .name = "soft_limit_in_bytes",
4774 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4775 .write_string = mem_cgroup_write,
4776 .read_u64 = mem_cgroup_read,
4780 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4781 .trigger = mem_cgroup_reset,
4782 .read_u64 = mem_cgroup_read,
4786 .read_map = mem_control_stat_show,
4789 .name = "force_empty",
4790 .trigger = mem_cgroup_force_empty_write,
4793 .name = "use_hierarchy",
4794 .write_u64 = mem_cgroup_hierarchy_write,
4795 .read_u64 = mem_cgroup_hierarchy_read,
4798 .name = "swappiness",
4799 .read_u64 = mem_cgroup_swappiness_read,
4800 .write_u64 = mem_cgroup_swappiness_write,
4803 .name = "move_charge_at_immigrate",
4804 .read_u64 = mem_cgroup_move_charge_read,
4805 .write_u64 = mem_cgroup_move_charge_write,
4808 .name = "oom_control",
4809 .read_map = mem_cgroup_oom_control_read,
4810 .write_u64 = mem_cgroup_oom_control_write,
4811 .register_event = mem_cgroup_oom_register_event,
4812 .unregister_event = mem_cgroup_oom_unregister_event,
4813 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4817 .name = "numa_stat",
4818 .open = mem_control_numa_stat_open,
4824 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4825 static struct cftype memsw_cgroup_files[] = {
4827 .name = "memsw.usage_in_bytes",
4828 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4829 .read_u64 = mem_cgroup_read,
4830 .register_event = mem_cgroup_usage_register_event,
4831 .unregister_event = mem_cgroup_usage_unregister_event,
4834 .name = "memsw.max_usage_in_bytes",
4835 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4836 .trigger = mem_cgroup_reset,
4837 .read_u64 = mem_cgroup_read,
4840 .name = "memsw.limit_in_bytes",
4841 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4842 .write_string = mem_cgroup_write,
4843 .read_u64 = mem_cgroup_read,
4846 .name = "memsw.failcnt",
4847 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4848 .trigger = mem_cgroup_reset,
4849 .read_u64 = mem_cgroup_read,
4853 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4855 if (!do_swap_account)
4857 return cgroup_add_files(cont, ss, memsw_cgroup_files,
4858 ARRAY_SIZE(memsw_cgroup_files));
4861 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4867 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4869 struct mem_cgroup_per_node *pn;
4870 struct mem_cgroup_per_zone *mz;
4872 int zone, tmp = node;
4874 * This routine is called against possible nodes.
4875 * But it's BUG to call kmalloc() against offline node.
4877 * TODO: this routine can waste much memory for nodes which will
4878 * never be onlined. It's better to use memory hotplug callback
4881 if (!node_state(node, N_NORMAL_MEMORY))
4883 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4887 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4888 mz = &pn->zoneinfo[zone];
4890 INIT_LIST_HEAD(&mz->lruvec.lists[l]);
4891 mz->usage_in_excess = 0;
4892 mz->on_tree = false;
4895 memcg->info.nodeinfo[node] = pn;
4899 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4901 kfree(memcg->info.nodeinfo[node]);
4904 static struct mem_cgroup *mem_cgroup_alloc(void)
4906 struct mem_cgroup *mem;
4907 int size = sizeof(struct mem_cgroup);
4909 /* Can be very big if MAX_NUMNODES is very big */
4910 if (size < PAGE_SIZE)
4911 mem = kzalloc(size, GFP_KERNEL);
4913 mem = vzalloc(size);
4918 mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4921 spin_lock_init(&mem->pcp_counter_lock);
4925 if (size < PAGE_SIZE)
4933 * At destroying mem_cgroup, references from swap_cgroup can remain.
4934 * (scanning all at force_empty is too costly...)
4936 * Instead of clearing all references at force_empty, we remember
4937 * the number of reference from swap_cgroup and free mem_cgroup when
4938 * it goes down to 0.
4940 * Removal of cgroup itself succeeds regardless of refs from swap.
4943 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4947 mem_cgroup_remove_from_trees(memcg);
4948 free_css_id(&mem_cgroup_subsys, &memcg->css);
4950 for_each_node_state(node, N_POSSIBLE)
4951 free_mem_cgroup_per_zone_info(memcg, node);
4953 free_percpu(memcg->stat);
4954 if (sizeof(struct mem_cgroup) < PAGE_SIZE)
4960 static void mem_cgroup_get(struct mem_cgroup *memcg)
4962 atomic_inc(&memcg->refcnt);
4965 static void __mem_cgroup_put(struct mem_cgroup *memcg, int count)
4967 if (atomic_sub_and_test(count, &memcg->refcnt)) {
4968 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4969 __mem_cgroup_free(memcg);
4971 mem_cgroup_put(parent);
4975 static void mem_cgroup_put(struct mem_cgroup *memcg)
4977 __mem_cgroup_put(memcg, 1);
4981 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4983 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4985 if (!memcg->res.parent)
4987 return mem_cgroup_from_res_counter(memcg->res.parent, res);
4989 EXPORT_SYMBOL(parent_mem_cgroup);
4991 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4992 static void __init enable_swap_cgroup(void)
4994 if (!mem_cgroup_disabled() && really_do_swap_account)
4995 do_swap_account = 1;
4998 static void __init enable_swap_cgroup(void)
5003 static int mem_cgroup_soft_limit_tree_init(void)
5005 struct mem_cgroup_tree_per_node *rtpn;
5006 struct mem_cgroup_tree_per_zone *rtpz;
5007 int tmp, node, zone;
5009 for_each_node_state(node, N_POSSIBLE) {
5011 if (!node_state(node, N_NORMAL_MEMORY))
5013 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
5017 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5019 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5020 rtpz = &rtpn->rb_tree_per_zone[zone];
5021 rtpz->rb_root = RB_ROOT;
5022 spin_lock_init(&rtpz->lock);
5028 static struct cgroup_subsys_state * __ref
5029 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
5031 struct mem_cgroup *memcg, *parent;
5032 long error = -ENOMEM;
5035 memcg = mem_cgroup_alloc();
5037 return ERR_PTR(error);
5039 for_each_node_state(node, N_POSSIBLE)
5040 if (alloc_mem_cgroup_per_zone_info(memcg, node))
5044 if (cont->parent == NULL) {
5046 enable_swap_cgroup();
5048 root_mem_cgroup = memcg;
5049 if (mem_cgroup_soft_limit_tree_init())
5051 for_each_possible_cpu(cpu) {
5052 struct memcg_stock_pcp *stock =
5053 &per_cpu(memcg_stock, cpu);
5054 INIT_WORK(&stock->work, drain_local_stock);
5056 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5058 parent = mem_cgroup_from_cont(cont->parent);
5059 memcg->use_hierarchy = parent->use_hierarchy;
5060 memcg->oom_kill_disable = parent->oom_kill_disable;
5063 if (parent && parent->use_hierarchy) {
5064 res_counter_init(&memcg->res, &parent->res);
5065 res_counter_init(&memcg->memsw, &parent->memsw);
5066 res_counter_init(&memcg->kmem, &parent->kmem);
5068 * We increment refcnt of the parent to ensure that we can
5069 * safely access it on res_counter_charge/uncharge.
5070 * This refcnt will be decremented when freeing this
5071 * mem_cgroup(see mem_cgroup_put).
5073 mem_cgroup_get(parent);
5075 res_counter_init(&memcg->res, NULL);
5076 res_counter_init(&memcg->memsw, NULL);
5077 res_counter_init(&memcg->kmem, NULL);
5079 memcg->last_scanned_node = MAX_NUMNODES;
5080 INIT_LIST_HEAD(&memcg->oom_notify);
5083 memcg->swappiness = mem_cgroup_swappiness(parent);
5084 atomic_set(&memcg->refcnt, 1);
5085 memcg->move_charge_at_immigrate = 0;
5086 mutex_init(&memcg->thresholds_lock);
5089 __mem_cgroup_free(memcg);
5090 root_mem_cgroup = NULL;
5091 return ERR_PTR(error);
5094 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
5095 struct cgroup *cont)
5097 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5099 return mem_cgroup_force_empty(memcg, false);
5102 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
5103 struct cgroup *cont)
5105 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5107 kmem_cgroup_destroy(ss, cont);
5109 mem_cgroup_put(memcg);
5112 static int mem_cgroup_populate(struct cgroup_subsys *ss,
5113 struct cgroup *cont)
5117 ret = cgroup_add_files(cont, ss, mem_cgroup_files,
5118 ARRAY_SIZE(mem_cgroup_files));
5121 ret = register_memsw_files(cont, ss);
5124 ret = register_kmem_files(cont, ss);
5130 /* Handlers for move charge at task migration. */
5131 #define PRECHARGE_COUNT_AT_ONCE 256
5132 static int mem_cgroup_do_precharge(unsigned long count)
5135 int batch_count = PRECHARGE_COUNT_AT_ONCE;
5136 struct mem_cgroup *memcg = mc.to;
5138 if (mem_cgroup_is_root(memcg)) {
5139 mc.precharge += count;
5140 /* we don't need css_get for root */
5143 /* try to charge at once */
5145 struct res_counter *dummy;
5147 * "memcg" cannot be under rmdir() because we've already checked
5148 * by cgroup_lock_live_cgroup() that it is not removed and we
5149 * are still under the same cgroup_mutex. So we can postpone
5152 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
5154 if (do_swap_account && res_counter_charge(&memcg->memsw,
5155 PAGE_SIZE * count, &dummy)) {
5156 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
5159 mc.precharge += count;
5163 /* fall back to one by one charge */
5165 if (signal_pending(current)) {
5169 if (!batch_count--) {
5170 batch_count = PRECHARGE_COUNT_AT_ONCE;
5173 ret = __mem_cgroup_try_charge(NULL,
5174 GFP_KERNEL, 1, &memcg, false);
5176 /* mem_cgroup_clear_mc() will do uncharge later */
5184 * is_target_pte_for_mc - check a pte whether it is valid for move charge
5185 * @vma: the vma the pte to be checked belongs
5186 * @addr: the address corresponding to the pte to be checked
5187 * @ptent: the pte to be checked
5188 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5191 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5192 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5193 * move charge. if @target is not NULL, the page is stored in target->page
5194 * with extra refcnt got(Callers should handle it).
5195 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5196 * target for charge migration. if @target is not NULL, the entry is stored
5199 * Called with pte lock held.
5206 enum mc_target_type {
5207 MC_TARGET_NONE, /* not used */
5212 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5213 unsigned long addr, pte_t ptent)
5215 struct page *page = vm_normal_page(vma, addr, ptent);
5217 if (!page || !page_mapped(page))
5219 if (PageAnon(page)) {
5220 /* we don't move shared anon */
5221 if (!move_anon() || page_mapcount(page) > 2)
5223 } else if (!move_file())
5224 /* we ignore mapcount for file pages */
5226 if (!get_page_unless_zero(page))
5232 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5233 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5236 struct page *page = NULL;
5237 swp_entry_t ent = pte_to_swp_entry(ptent);
5239 if (!move_anon() || non_swap_entry(ent))
5241 usage_count = mem_cgroup_count_swap_user(ent, &page);
5242 if (usage_count > 1) { /* we don't move shared anon */
5247 if (do_swap_account)
5248 entry->val = ent.val;
5253 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5254 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5256 struct page *page = NULL;
5257 struct inode *inode;
5258 struct address_space *mapping;
5261 if (!vma->vm_file) /* anonymous vma */
5266 inode = vma->vm_file->f_path.dentry->d_inode;
5267 mapping = vma->vm_file->f_mapping;
5268 if (pte_none(ptent))
5269 pgoff = linear_page_index(vma, addr);
5270 else /* pte_file(ptent) is true */
5271 pgoff = pte_to_pgoff(ptent);
5273 /* page is moved even if it's not RSS of this task(page-faulted). */
5274 page = find_get_page(mapping, pgoff);
5277 /* shmem/tmpfs may report page out on swap: account for that too. */
5278 if (radix_tree_exceptional_entry(page)) {
5279 swp_entry_t swap = radix_to_swp_entry(page);
5280 if (do_swap_account)
5282 page = find_get_page(&swapper_space, swap.val);
5288 static int is_target_pte_for_mc(struct vm_area_struct *vma,
5289 unsigned long addr, pte_t ptent, union mc_target *target)
5291 struct page *page = NULL;
5292 struct page_cgroup *pc;
5294 swp_entry_t ent = { .val = 0 };
5296 if (pte_present(ptent))
5297 page = mc_handle_present_pte(vma, addr, ptent);
5298 else if (is_swap_pte(ptent))
5299 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5300 else if (pte_none(ptent) || pte_file(ptent))
5301 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5303 if (!page && !ent.val)
5306 pc = lookup_page_cgroup(page);
5308 * Do only loose check w/o page_cgroup lock.
5309 * mem_cgroup_move_account() checks the pc is valid or not under
5312 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5313 ret = MC_TARGET_PAGE;
5315 target->page = page;
5317 if (!ret || !target)
5320 /* There is a swap entry and a page doesn't exist or isn't charged */
5321 if (ent.val && !ret &&
5322 css_id(&mc.from->css) == lookup_swap_cgroup(ent)) {
5323 ret = MC_TARGET_SWAP;
5330 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5331 unsigned long addr, unsigned long end,
5332 struct mm_walk *walk)
5334 struct vm_area_struct *vma = walk->private;
5338 split_huge_page_pmd(walk->mm, pmd);
5340 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5341 for (; addr != end; pte++, addr += PAGE_SIZE)
5342 if (is_target_pte_for_mc(vma, addr, *pte, NULL))
5343 mc.precharge++; /* increment precharge temporarily */
5344 pte_unmap_unlock(pte - 1, ptl);
5350 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5352 unsigned long precharge;
5353 struct vm_area_struct *vma;
5355 down_read(&mm->mmap_sem);
5356 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5357 struct mm_walk mem_cgroup_count_precharge_walk = {
5358 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5362 if (is_vm_hugetlb_page(vma))
5364 walk_page_range(vma->vm_start, vma->vm_end,
5365 &mem_cgroup_count_precharge_walk);
5367 up_read(&mm->mmap_sem);
5369 precharge = mc.precharge;
5375 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5377 unsigned long precharge = mem_cgroup_count_precharge(mm);
5379 VM_BUG_ON(mc.moving_task);
5380 mc.moving_task = current;
5381 return mem_cgroup_do_precharge(precharge);
5384 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5385 static void __mem_cgroup_clear_mc(void)
5387 struct mem_cgroup *from = mc.from;
5388 struct mem_cgroup *to = mc.to;
5390 /* we must uncharge all the leftover precharges from mc.to */
5392 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
5396 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5397 * we must uncharge here.
5399 if (mc.moved_charge) {
5400 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
5401 mc.moved_charge = 0;
5403 /* we must fixup refcnts and charges */
5404 if (mc.moved_swap) {
5405 /* uncharge swap account from the old cgroup */
5406 if (!mem_cgroup_is_root(mc.from))
5407 res_counter_uncharge(&mc.from->memsw,
5408 PAGE_SIZE * mc.moved_swap);
5409 __mem_cgroup_put(mc.from, mc.moved_swap);
5411 if (!mem_cgroup_is_root(mc.to)) {
5413 * we charged both to->res and to->memsw, so we should
5416 res_counter_uncharge(&mc.to->res,
5417 PAGE_SIZE * mc.moved_swap);
5419 /* we've already done mem_cgroup_get(mc.to) */
5422 memcg_oom_recover(from);
5423 memcg_oom_recover(to);
5424 wake_up_all(&mc.waitq);
5427 static void mem_cgroup_clear_mc(void)
5429 struct mem_cgroup *from = mc.from;
5432 * we must clear moving_task before waking up waiters at the end of
5435 mc.moving_task = NULL;
5436 __mem_cgroup_clear_mc();
5437 spin_lock(&mc.lock);
5440 spin_unlock(&mc.lock);
5441 mem_cgroup_end_move(from);
5444 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5445 struct cgroup *cgroup,
5446 struct cgroup_taskset *tset)
5448 struct task_struct *p = cgroup_taskset_first(tset);
5450 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup);
5452 if (memcg->move_charge_at_immigrate) {
5453 struct mm_struct *mm;
5454 struct mem_cgroup *from = mem_cgroup_from_task(p);
5456 VM_BUG_ON(from == memcg);
5458 mm = get_task_mm(p);
5461 /* We move charges only when we move a owner of the mm */
5462 if (mm->owner == p) {
5465 VM_BUG_ON(mc.precharge);
5466 VM_BUG_ON(mc.moved_charge);
5467 VM_BUG_ON(mc.moved_swap);
5468 mem_cgroup_start_move(from);
5469 spin_lock(&mc.lock);
5472 spin_unlock(&mc.lock);
5473 /* We set mc.moving_task later */
5475 ret = mem_cgroup_precharge_mc(mm);
5477 mem_cgroup_clear_mc();
5484 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5485 struct cgroup *cgroup,
5486 struct cgroup_taskset *tset)
5488 mem_cgroup_clear_mc();
5491 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5492 unsigned long addr, unsigned long end,
5493 struct mm_walk *walk)
5496 struct vm_area_struct *vma = walk->private;
5500 split_huge_page_pmd(walk->mm, pmd);
5502 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5503 for (; addr != end; addr += PAGE_SIZE) {
5504 pte_t ptent = *(pte++);
5505 union mc_target target;
5508 struct page_cgroup *pc;
5514 type = is_target_pte_for_mc(vma, addr, ptent, &target);
5516 case MC_TARGET_PAGE:
5518 if (isolate_lru_page(page))
5520 pc = lookup_page_cgroup(page);
5521 if (!mem_cgroup_move_account(page, 1, pc,
5522 mc.from, mc.to, false)) {
5524 /* we uncharge from mc.from later. */
5527 putback_lru_page(page);
5528 put: /* is_target_pte_for_mc() gets the page */
5531 case MC_TARGET_SWAP:
5533 if (!mem_cgroup_move_swap_account(ent,
5534 mc.from, mc.to, false)) {
5536 /* we fixup refcnts and charges later. */
5544 pte_unmap_unlock(pte - 1, ptl);
5549 * We have consumed all precharges we got in can_attach().
5550 * We try charge one by one, but don't do any additional
5551 * charges to mc.to if we have failed in charge once in attach()
5554 ret = mem_cgroup_do_precharge(1);
5562 static void mem_cgroup_move_charge(struct mm_struct *mm)
5564 struct vm_area_struct *vma;
5566 lru_add_drain_all();
5568 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5570 * Someone who are holding the mmap_sem might be waiting in
5571 * waitq. So we cancel all extra charges, wake up all waiters,
5572 * and retry. Because we cancel precharges, we might not be able
5573 * to move enough charges, but moving charge is a best-effort
5574 * feature anyway, so it wouldn't be a big problem.
5576 __mem_cgroup_clear_mc();
5580 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5582 struct mm_walk mem_cgroup_move_charge_walk = {
5583 .pmd_entry = mem_cgroup_move_charge_pte_range,
5587 if (is_vm_hugetlb_page(vma))
5589 ret = walk_page_range(vma->vm_start, vma->vm_end,
5590 &mem_cgroup_move_charge_walk);
5593 * means we have consumed all precharges and failed in
5594 * doing additional charge. Just abandon here.
5598 up_read(&mm->mmap_sem);
5601 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5602 struct cgroup *cont,
5603 struct cgroup_taskset *tset)
5605 struct task_struct *p = cgroup_taskset_first(tset);
5606 struct mm_struct *mm = get_task_mm(p);
5610 mem_cgroup_move_charge(mm);
5615 mem_cgroup_clear_mc();
5617 #else /* !CONFIG_MMU */
5618 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5619 struct cgroup *cgroup,
5620 struct cgroup_taskset *tset)
5624 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5625 struct cgroup *cgroup,
5626 struct cgroup_taskset *tset)
5629 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5630 struct cgroup *cont,
5631 struct cgroup_taskset *tset)
5636 struct cgroup_subsys mem_cgroup_subsys = {
5638 .subsys_id = mem_cgroup_subsys_id,
5639 .create = mem_cgroup_create,
5640 .pre_destroy = mem_cgroup_pre_destroy,
5641 .destroy = mem_cgroup_destroy,
5642 .populate = mem_cgroup_populate,
5643 .can_attach = mem_cgroup_can_attach,
5644 .cancel_attach = mem_cgroup_cancel_attach,
5645 .attach = mem_cgroup_move_task,
5650 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5651 static int __init enable_swap_account(char *s)
5653 /* consider enabled if no parameter or 1 is given */
5654 if (!strcmp(s, "1"))
5655 really_do_swap_account = 1;
5656 else if (!strcmp(s, "0"))
5657 really_do_swap_account = 0;
5660 __setup("swapaccount=", enable_swap_account);