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 <asm/uaccess.h>
56 #include <trace/events/vmscan.h>
58 struct cgroup_subsys mem_cgroup_subsys __read_mostly;
59 #define MEM_CGROUP_RECLAIM_RETRIES 5
60 struct mem_cgroup *root_mem_cgroup __read_mostly;
62 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
63 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
64 int do_swap_account __read_mostly;
66 /* for remember boot option*/
67 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP_ENABLED
68 static int really_do_swap_account __initdata = 1;
70 static int really_do_swap_account __initdata = 0;
74 #define do_swap_account (0)
79 * Statistics for memory cgroup.
81 enum mem_cgroup_stat_index {
83 * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
85 MEM_CGROUP_STAT_CACHE, /* # of pages charged as cache */
86 MEM_CGROUP_STAT_RSS, /* # of pages charged as anon rss */
87 MEM_CGROUP_STAT_FILE_MAPPED, /* # of pages charged as file rss */
88 MEM_CGROUP_STAT_SWAPOUT, /* # of pages, swapped out */
89 MEM_CGROUP_STAT_DATA, /* end of data requires synchronization */
90 MEM_CGROUP_ON_MOVE, /* someone is moving account between groups */
91 MEM_CGROUP_STAT_NSTATS,
94 enum mem_cgroup_events_index {
95 MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */
96 MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */
97 MEM_CGROUP_EVENTS_COUNT, /* # of pages paged in/out */
98 MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */
99 MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */
100 MEM_CGROUP_EVENTS_NSTATS,
103 * Per memcg event counter is incremented at every pagein/pageout. With THP,
104 * it will be incremated by the number of pages. This counter is used for
105 * for trigger some periodic events. This is straightforward and better
106 * than using jiffies etc. to handle periodic memcg event.
108 enum mem_cgroup_events_target {
109 MEM_CGROUP_TARGET_THRESH,
110 MEM_CGROUP_TARGET_SOFTLIMIT,
111 MEM_CGROUP_TARGET_NUMAINFO,
114 #define THRESHOLDS_EVENTS_TARGET (128)
115 #define SOFTLIMIT_EVENTS_TARGET (1024)
116 #define NUMAINFO_EVENTS_TARGET (1024)
118 struct mem_cgroup_stat_cpu {
119 long count[MEM_CGROUP_STAT_NSTATS];
120 unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
121 unsigned long targets[MEM_CGROUP_NTARGETS];
125 * per-zone information in memory controller.
127 struct mem_cgroup_per_zone {
129 * spin_lock to protect the per cgroup LRU
131 struct list_head lists[NR_LRU_LISTS];
132 unsigned long count[NR_LRU_LISTS];
134 struct zone_reclaim_stat reclaim_stat;
135 struct rb_node tree_node; /* RB tree node */
136 unsigned long long usage_in_excess;/* Set to the value by which */
137 /* the soft limit is exceeded*/
139 struct mem_cgroup *mem; /* Back pointer, we cannot */
140 /* use container_of */
142 /* Macro for accessing counter */
143 #define MEM_CGROUP_ZSTAT(mz, idx) ((mz)->count[(idx)])
145 struct mem_cgroup_per_node {
146 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
149 struct mem_cgroup_lru_info {
150 struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
154 * Cgroups above their limits are maintained in a RB-Tree, independent of
155 * their hierarchy representation
158 struct mem_cgroup_tree_per_zone {
159 struct rb_root rb_root;
163 struct mem_cgroup_tree_per_node {
164 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
167 struct mem_cgroup_tree {
168 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
171 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
173 struct mem_cgroup_threshold {
174 struct eventfd_ctx *eventfd;
179 struct mem_cgroup_threshold_ary {
180 /* An array index points to threshold just below usage. */
181 int current_threshold;
182 /* Size of entries[] */
184 /* Array of thresholds */
185 struct mem_cgroup_threshold entries[0];
188 struct mem_cgroup_thresholds {
189 /* Primary thresholds array */
190 struct mem_cgroup_threshold_ary *primary;
192 * Spare threshold array.
193 * This is needed to make mem_cgroup_unregister_event() "never fail".
194 * It must be able to store at least primary->size - 1 entries.
196 struct mem_cgroup_threshold_ary *spare;
200 struct mem_cgroup_eventfd_list {
201 struct list_head list;
202 struct eventfd_ctx *eventfd;
205 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
206 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
209 * The memory controller data structure. The memory controller controls both
210 * page cache and RSS per cgroup. We would eventually like to provide
211 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
212 * to help the administrator determine what knobs to tune.
214 * TODO: Add a water mark for the memory controller. Reclaim will begin when
215 * we hit the water mark. May be even add a low water mark, such that
216 * no reclaim occurs from a cgroup at it's low water mark, this is
217 * a feature that will be implemented much later in the future.
220 struct cgroup_subsys_state css;
222 * the counter to account for memory usage
224 struct res_counter res;
226 * the counter to account for mem+swap usage.
228 struct res_counter memsw;
230 * the counter to account for kmem usage.
232 struct res_counter kmem;
234 * Per cgroup active and inactive list, similar to the
235 * per zone LRU lists.
237 struct mem_cgroup_lru_info info;
239 * While reclaiming in a hierarchy, we cache the last child we
242 int last_scanned_child;
243 int last_scanned_node;
245 nodemask_t scan_nodes;
246 atomic_t numainfo_events;
247 atomic_t numainfo_updating;
250 * Should the accounting and control be hierarchical, per subtree?
260 /* OOM-Killer disable */
261 int oom_kill_disable;
263 /* set when res.limit == memsw.limit */
264 bool memsw_is_minimum;
266 /* protect arrays of thresholds */
267 struct mutex thresholds_lock;
269 /* thresholds for memory usage. RCU-protected */
270 struct mem_cgroup_thresholds thresholds;
272 /* thresholds for mem+swap usage. RCU-protected */
273 struct mem_cgroup_thresholds memsw_thresholds;
275 /* For oom notifier event fd */
276 struct list_head oom_notify;
279 * Should we move charges of a task when a task is moved into this
280 * mem_cgroup ? And what type of charges should we move ?
282 unsigned long move_charge_at_immigrate;
284 * Should kernel memory limits be stabilished independently
287 int kmem_independent_accounting;
291 struct mem_cgroup_stat_cpu *stat;
293 * used when a cpu is offlined or other synchronizations
294 * See mem_cgroup_read_stat().
296 struct mem_cgroup_stat_cpu nocpu_base;
297 spinlock_t pcp_counter_lock;
300 /* Stuffs for move charges at task migration. */
302 * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
303 * left-shifted bitmap of these types.
306 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
307 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
311 /* "mc" and its members are protected by cgroup_mutex */
312 static struct move_charge_struct {
313 spinlock_t lock; /* for from, to */
314 struct mem_cgroup *from;
315 struct mem_cgroup *to;
316 unsigned long precharge;
317 unsigned long moved_charge;
318 unsigned long moved_swap;
319 struct task_struct *moving_task; /* a task moving charges */
320 wait_queue_head_t waitq; /* a waitq for other context */
322 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
323 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
326 static bool move_anon(void)
328 return test_bit(MOVE_CHARGE_TYPE_ANON,
329 &mc.to->move_charge_at_immigrate);
332 static bool move_file(void)
334 return test_bit(MOVE_CHARGE_TYPE_FILE,
335 &mc.to->move_charge_at_immigrate);
339 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
340 * limit reclaim to prevent infinite loops, if they ever occur.
342 #define MEM_CGROUP_MAX_RECLAIM_LOOPS (100)
343 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2)
346 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
347 MEM_CGROUP_CHARGE_TYPE_MAPPED,
348 MEM_CGROUP_CHARGE_TYPE_SHMEM, /* used by page migration of shmem */
349 MEM_CGROUP_CHARGE_TYPE_FORCE, /* used by force_empty */
350 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
351 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
355 /* for encoding cft->private value on file */
364 #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
365 #define MEMFILE_TYPE(val) (((val) >> 16) & 0xffff)
366 #define MEMFILE_ATTR(val) ((val) & 0xffff)
367 /* Used for OOM nofiier */
368 #define OOM_CONTROL (0)
371 * Reclaim flags for mem_cgroup_hierarchical_reclaim
373 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
374 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
375 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
376 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
377 #define MEM_CGROUP_RECLAIM_SOFT_BIT 0x2
378 #define MEM_CGROUP_RECLAIM_SOFT (1 << MEM_CGROUP_RECLAIM_SOFT_BIT)
380 static void mem_cgroup_get(struct mem_cgroup *memcg);
381 static void mem_cgroup_put(struct mem_cgroup *memcg);
383 /* Writing them here to avoid exposing memcg's inner layout */
384 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_KMEM
386 #include <net/sock.h>
388 static bool mem_cgroup_is_root(struct mem_cgroup *memcg);
389 void sock_update_memcg(struct sock *sk)
391 /* A socket spends its whole life in the same cgroup */
396 if (static_branch(&memcg_socket_limit_enabled)) {
397 struct mem_cgroup *memcg;
399 BUG_ON(!sk->sk_prot->proto_cgroup);
402 memcg = mem_cgroup_from_task(current);
403 if (!mem_cgroup_is_root(memcg)) {
404 mem_cgroup_get(memcg);
405 sk->sk_cgrp = sk->sk_prot->proto_cgroup(memcg);
410 EXPORT_SYMBOL(sock_update_memcg);
412 void sock_release_memcg(struct sock *sk)
414 if (static_branch(&memcg_socket_limit_enabled) && sk->sk_cgrp) {
415 struct mem_cgroup *memcg;
416 WARN_ON(!sk->sk_cgrp->memcg);
417 memcg = sk->sk_cgrp->memcg;
418 mem_cgroup_put(memcg);
421 #endif /* CONFIG_INET */
422 #endif /* CONFIG_CGROUP_MEM_RES_CTLR_KMEM */
424 static void drain_all_stock_async(struct mem_cgroup *memcg);
426 static struct mem_cgroup_per_zone *
427 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
429 return &memcg->info.nodeinfo[nid]->zoneinfo[zid];
432 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
437 static struct mem_cgroup_per_zone *
438 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
440 int nid = page_to_nid(page);
441 int zid = page_zonenum(page);
443 return mem_cgroup_zoneinfo(memcg, nid, zid);
446 static struct mem_cgroup_tree_per_zone *
447 soft_limit_tree_node_zone(int nid, int zid)
449 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
452 static struct mem_cgroup_tree_per_zone *
453 soft_limit_tree_from_page(struct page *page)
455 int nid = page_to_nid(page);
456 int zid = page_zonenum(page);
458 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
462 __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
463 struct mem_cgroup_per_zone *mz,
464 struct mem_cgroup_tree_per_zone *mctz,
465 unsigned long long new_usage_in_excess)
467 struct rb_node **p = &mctz->rb_root.rb_node;
468 struct rb_node *parent = NULL;
469 struct mem_cgroup_per_zone *mz_node;
474 mz->usage_in_excess = new_usage_in_excess;
475 if (!mz->usage_in_excess)
479 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
481 if (mz->usage_in_excess < mz_node->usage_in_excess)
484 * We can't avoid mem cgroups that are over their soft
485 * limit by the same amount
487 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
490 rb_link_node(&mz->tree_node, parent, p);
491 rb_insert_color(&mz->tree_node, &mctz->rb_root);
496 __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
497 struct mem_cgroup_per_zone *mz,
498 struct mem_cgroup_tree_per_zone *mctz)
502 rb_erase(&mz->tree_node, &mctz->rb_root);
507 mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
508 struct mem_cgroup_per_zone *mz,
509 struct mem_cgroup_tree_per_zone *mctz)
511 spin_lock(&mctz->lock);
512 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
513 spin_unlock(&mctz->lock);
517 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
519 unsigned long long excess;
520 struct mem_cgroup_per_zone *mz;
521 struct mem_cgroup_tree_per_zone *mctz;
522 int nid = page_to_nid(page);
523 int zid = page_zonenum(page);
524 mctz = soft_limit_tree_from_page(page);
527 * Necessary to update all ancestors when hierarchy is used.
528 * because their event counter is not touched.
530 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
531 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
532 excess = res_counter_soft_limit_excess(&memcg->res);
534 * We have to update the tree if mz is on RB-tree or
535 * mem is over its softlimit.
537 if (excess || mz->on_tree) {
538 spin_lock(&mctz->lock);
539 /* if on-tree, remove it */
541 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
543 * Insert again. mz->usage_in_excess will be updated.
544 * If excess is 0, no tree ops.
546 __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
547 spin_unlock(&mctz->lock);
552 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
555 struct mem_cgroup_per_zone *mz;
556 struct mem_cgroup_tree_per_zone *mctz;
558 for_each_node_state(node, N_POSSIBLE) {
559 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
560 mz = mem_cgroup_zoneinfo(memcg, node, zone);
561 mctz = soft_limit_tree_node_zone(node, zone);
562 mem_cgroup_remove_exceeded(memcg, mz, mctz);
567 static struct mem_cgroup_per_zone *
568 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
570 struct rb_node *rightmost = NULL;
571 struct mem_cgroup_per_zone *mz;
575 rightmost = rb_last(&mctz->rb_root);
577 goto done; /* Nothing to reclaim from */
579 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
581 * Remove the node now but someone else can add it back,
582 * we will to add it back at the end of reclaim to its correct
583 * position in the tree.
585 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
586 if (!res_counter_soft_limit_excess(&mz->mem->res) ||
587 !css_tryget(&mz->mem->css))
593 static struct mem_cgroup_per_zone *
594 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
596 struct mem_cgroup_per_zone *mz;
598 spin_lock(&mctz->lock);
599 mz = __mem_cgroup_largest_soft_limit_node(mctz);
600 spin_unlock(&mctz->lock);
605 * Implementation Note: reading percpu statistics for memcg.
607 * Both of vmstat[] and percpu_counter has threshold and do periodic
608 * synchronization to implement "quick" read. There are trade-off between
609 * reading cost and precision of value. Then, we may have a chance to implement
610 * a periodic synchronizion of counter in memcg's counter.
612 * But this _read() function is used for user interface now. The user accounts
613 * memory usage by memory cgroup and he _always_ requires exact value because
614 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
615 * have to visit all online cpus and make sum. So, for now, unnecessary
616 * synchronization is not implemented. (just implemented for cpu hotplug)
618 * If there are kernel internal actions which can make use of some not-exact
619 * value, and reading all cpu value can be performance bottleneck in some
620 * common workload, threashold and synchonization as vmstat[] should be
623 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
624 enum mem_cgroup_stat_index idx)
630 for_each_online_cpu(cpu)
631 val += per_cpu(memcg->stat->count[idx], cpu);
632 #ifdef CONFIG_HOTPLUG_CPU
633 spin_lock(&memcg->pcp_counter_lock);
634 val += memcg->nocpu_base.count[idx];
635 spin_unlock(&memcg->pcp_counter_lock);
641 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
644 int val = (charge) ? 1 : -1;
645 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAPOUT], val);
648 void mem_cgroup_pgfault(struct mem_cgroup *memcg, int val)
650 this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT], val);
653 void mem_cgroup_pgmajfault(struct mem_cgroup *memcg, int val)
655 this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT], val);
658 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
659 enum mem_cgroup_events_index idx)
661 unsigned long val = 0;
664 for_each_online_cpu(cpu)
665 val += per_cpu(memcg->stat->events[idx], cpu);
666 #ifdef CONFIG_HOTPLUG_CPU
667 spin_lock(&memcg->pcp_counter_lock);
668 val += memcg->nocpu_base.events[idx];
669 spin_unlock(&memcg->pcp_counter_lock);
674 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
675 bool file, int nr_pages)
680 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
683 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
686 /* pagein of a big page is an event. So, ignore page size */
688 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
690 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
691 nr_pages = -nr_pages; /* for event */
694 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT], nr_pages);
700 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
701 unsigned int lru_mask)
703 struct mem_cgroup_per_zone *mz;
705 unsigned long ret = 0;
707 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
710 if (BIT(l) & lru_mask)
711 ret += MEM_CGROUP_ZSTAT(mz, l);
717 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
718 int nid, unsigned int lru_mask)
723 for (zid = 0; zid < MAX_NR_ZONES; zid++)
724 total += mem_cgroup_zone_nr_lru_pages(memcg,
730 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
731 unsigned int lru_mask)
736 for_each_node_state(nid, N_HIGH_MEMORY)
737 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
741 static bool __memcg_event_check(struct mem_cgroup *memcg, int target)
743 unsigned long val, next;
745 val = __this_cpu_read(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT]);
746 next = __this_cpu_read(memcg->stat->targets[target]);
747 /* from time_after() in jiffies.h */
748 return ((long)next - (long)val < 0);
751 static void __mem_cgroup_target_update(struct mem_cgroup *memcg, int target)
753 unsigned long val, next;
755 val = __this_cpu_read(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT]);
758 case MEM_CGROUP_TARGET_THRESH:
759 next = val + THRESHOLDS_EVENTS_TARGET;
761 case MEM_CGROUP_TARGET_SOFTLIMIT:
762 next = val + SOFTLIMIT_EVENTS_TARGET;
764 case MEM_CGROUP_TARGET_NUMAINFO:
765 next = val + NUMAINFO_EVENTS_TARGET;
771 __this_cpu_write(memcg->stat->targets[target], next);
775 * Check events in order.
778 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
781 /* threshold event is triggered in finer grain than soft limit */
782 if (unlikely(__memcg_event_check(memcg, MEM_CGROUP_TARGET_THRESH))) {
783 mem_cgroup_threshold(memcg);
784 __mem_cgroup_target_update(memcg, MEM_CGROUP_TARGET_THRESH);
785 if (unlikely(__memcg_event_check(memcg,
786 MEM_CGROUP_TARGET_SOFTLIMIT))) {
787 mem_cgroup_update_tree(memcg, page);
788 __mem_cgroup_target_update(memcg,
789 MEM_CGROUP_TARGET_SOFTLIMIT);
792 if (unlikely(__memcg_event_check(memcg,
793 MEM_CGROUP_TARGET_NUMAINFO))) {
794 atomic_inc(&memcg->numainfo_events);
795 __mem_cgroup_target_update(memcg,
796 MEM_CGROUP_TARGET_NUMAINFO);
803 static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
805 return container_of(cgroup_subsys_state(cont,
806 mem_cgroup_subsys_id), struct mem_cgroup,
810 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
813 * mm_update_next_owner() may clear mm->owner to NULL
814 * if it races with swapoff, page migration, etc.
815 * So this can be called with p == NULL.
820 return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
821 struct mem_cgroup, css);
824 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
826 struct mem_cgroup *memcg = NULL;
831 * Because we have no locks, mm->owner's may be being moved to other
832 * cgroup. We use css_tryget() here even if this looks
833 * pessimistic (rather than adding locks here).
837 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
838 if (unlikely(!memcg))
840 } while (!css_tryget(&memcg->css));
845 /* The caller has to guarantee "mem" exists before calling this */
846 static struct mem_cgroup *mem_cgroup_start_loop(struct mem_cgroup *memcg)
848 struct cgroup_subsys_state *css;
851 if (!memcg) /* ROOT cgroup has the smallest ID */
852 return root_mem_cgroup; /*css_put/get against root is ignored*/
853 if (!memcg->use_hierarchy) {
854 if (css_tryget(&memcg->css))
860 * searching a memory cgroup which has the smallest ID under given
861 * ROOT cgroup. (ID >= 1)
863 css = css_get_next(&mem_cgroup_subsys, 1, &memcg->css, &found);
864 if (css && css_tryget(css))
865 memcg = container_of(css, struct mem_cgroup, css);
872 static struct mem_cgroup *mem_cgroup_get_next(struct mem_cgroup *iter,
873 struct mem_cgroup *root,
876 int nextid = css_id(&iter->css) + 1;
879 struct cgroup_subsys_state *css;
881 hierarchy_used = iter->use_hierarchy;
884 /* If no ROOT, walk all, ignore hierarchy */
885 if (!cond || (root && !hierarchy_used))
889 root = root_mem_cgroup;
895 css = css_get_next(&mem_cgroup_subsys, nextid,
897 if (css && css_tryget(css))
898 iter = container_of(css, struct mem_cgroup, css);
900 /* If css is NULL, no more cgroups will be found */
902 } while (css && !iter);
907 * for_eacn_mem_cgroup_tree() for visiting all cgroup under tree. Please
908 * be careful that "break" loop is not allowed. We have reference count.
909 * Instead of that modify "cond" to be false and "continue" to exit the loop.
911 #define for_each_mem_cgroup_tree_cond(iter, root, cond) \
912 for (iter = mem_cgroup_start_loop(root);\
914 iter = mem_cgroup_get_next(iter, root, cond))
916 #define for_each_mem_cgroup_tree(iter, root) \
917 for_each_mem_cgroup_tree_cond(iter, root, true)
919 #define for_each_mem_cgroup_all(iter) \
920 for_each_mem_cgroup_tree_cond(iter, NULL, true)
923 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
925 return (memcg == root_mem_cgroup);
928 void mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
930 struct mem_cgroup *memcg;
936 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
937 if (unlikely(!memcg))
942 mem_cgroup_pgmajfault(memcg, 1);
945 mem_cgroup_pgfault(memcg, 1);
953 EXPORT_SYMBOL(mem_cgroup_count_vm_event);
956 * Following LRU functions are allowed to be used without PCG_LOCK.
957 * Operations are called by routine of global LRU independently from memcg.
958 * What we have to take care of here is validness of pc->mem_cgroup.
960 * Changes to pc->mem_cgroup happens when
963 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
964 * It is added to LRU before charge.
965 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
966 * When moving account, the page is not on LRU. It's isolated.
969 void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru)
971 struct page_cgroup *pc;
972 struct mem_cgroup_per_zone *mz;
974 if (mem_cgroup_disabled())
976 pc = lookup_page_cgroup(page);
977 /* can happen while we handle swapcache. */
978 if (!TestClearPageCgroupAcctLRU(pc))
980 VM_BUG_ON(!pc->mem_cgroup);
982 * We don't check PCG_USED bit. It's cleared when the "page" is finally
983 * removed from global LRU.
985 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
986 /* huge page split is done under lru_lock. so, we have no races. */
987 MEM_CGROUP_ZSTAT(mz, lru) -= 1 << compound_order(page);
988 if (mem_cgroup_is_root(pc->mem_cgroup))
990 VM_BUG_ON(list_empty(&pc->lru));
991 list_del_init(&pc->lru);
994 void mem_cgroup_del_lru(struct page *page)
996 mem_cgroup_del_lru_list(page, page_lru(page));
1000 * Writeback is about to end against a page which has been marked for immediate
1001 * reclaim. If it still appears to be reclaimable, move it to the tail of the
1004 void mem_cgroup_rotate_reclaimable_page(struct page *page)
1006 struct mem_cgroup_per_zone *mz;
1007 struct page_cgroup *pc;
1008 enum lru_list lru = page_lru(page);
1010 if (mem_cgroup_disabled())
1013 pc = lookup_page_cgroup(page);
1014 /* unused or root page is not rotated. */
1015 if (!PageCgroupUsed(pc))
1017 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1019 if (mem_cgroup_is_root(pc->mem_cgroup))
1021 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1022 list_move_tail(&pc->lru, &mz->lists[lru]);
1025 void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru)
1027 struct mem_cgroup_per_zone *mz;
1028 struct page_cgroup *pc;
1030 if (mem_cgroup_disabled())
1033 pc = lookup_page_cgroup(page);
1034 /* unused or root page is not rotated. */
1035 if (!PageCgroupUsed(pc))
1037 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1039 if (mem_cgroup_is_root(pc->mem_cgroup))
1041 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1042 list_move(&pc->lru, &mz->lists[lru]);
1045 void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
1047 struct page_cgroup *pc;
1048 struct mem_cgroup_per_zone *mz;
1050 if (mem_cgroup_disabled())
1052 pc = lookup_page_cgroup(page);
1053 VM_BUG_ON(PageCgroupAcctLRU(pc));
1056 * SetPageLRU SetPageCgroupUsed
1058 * PageCgroupUsed && add to memcg LRU PageLRU && add to memcg LRU
1060 * Ensure that one of the two sides adds the page to the memcg
1061 * LRU during a race.
1064 if (!PageCgroupUsed(pc))
1066 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1068 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1069 /* huge page split is done under lru_lock. so, we have no races. */
1070 MEM_CGROUP_ZSTAT(mz, lru) += 1 << compound_order(page);
1071 SetPageCgroupAcctLRU(pc);
1072 if (mem_cgroup_is_root(pc->mem_cgroup))
1074 list_add(&pc->lru, &mz->lists[lru]);
1078 * At handling SwapCache and other FUSE stuff, pc->mem_cgroup may be changed
1079 * while it's linked to lru because the page may be reused after it's fully
1080 * uncharged. To handle that, unlink page_cgroup from LRU when charge it again.
1081 * It's done under lock_page and expected that zone->lru_lock isnever held.
1083 static void mem_cgroup_lru_del_before_commit(struct page *page)
1085 unsigned long flags;
1086 struct zone *zone = page_zone(page);
1087 struct page_cgroup *pc = lookup_page_cgroup(page);
1090 * Doing this check without taking ->lru_lock seems wrong but this
1091 * is safe. Because if page_cgroup's USED bit is unset, the page
1092 * will not be added to any memcg's LRU. If page_cgroup's USED bit is
1093 * set, the commit after this will fail, anyway.
1094 * This all charge/uncharge is done under some mutual execustion.
1095 * So, we don't need to taking care of changes in USED bit.
1097 if (likely(!PageLRU(page)))
1100 spin_lock_irqsave(&zone->lru_lock, flags);
1102 * Forget old LRU when this page_cgroup is *not* used. This Used bit
1103 * is guarded by lock_page() because the page is SwapCache.
1105 if (!PageCgroupUsed(pc))
1106 mem_cgroup_del_lru_list(page, page_lru(page));
1107 spin_unlock_irqrestore(&zone->lru_lock, flags);
1110 static void mem_cgroup_lru_add_after_commit(struct page *page)
1112 unsigned long flags;
1113 struct zone *zone = page_zone(page);
1114 struct page_cgroup *pc = lookup_page_cgroup(page);
1117 * SetPageLRU SetPageCgroupUsed
1119 * PageCgroupUsed && add to memcg LRU PageLRU && add to memcg LRU
1121 * Ensure that one of the two sides adds the page to the memcg
1122 * LRU during a race.
1125 /* taking care of that the page is added to LRU while we commit it */
1126 if (likely(!PageLRU(page)))
1128 spin_lock_irqsave(&zone->lru_lock, flags);
1129 /* link when the page is linked to LRU but page_cgroup isn't */
1130 if (PageLRU(page) && !PageCgroupAcctLRU(pc))
1131 mem_cgroup_add_lru_list(page, page_lru(page));
1132 spin_unlock_irqrestore(&zone->lru_lock, flags);
1136 void mem_cgroup_move_lists(struct page *page,
1137 enum lru_list from, enum lru_list to)
1139 if (mem_cgroup_disabled())
1141 mem_cgroup_del_lru_list(page, from);
1142 mem_cgroup_add_lru_list(page, to);
1146 * Checks whether given mem is same or in the root_mem_cgroup's
1149 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1150 struct mem_cgroup *memcg)
1152 if (root_memcg != memcg) {
1153 return (root_memcg->use_hierarchy &&
1154 css_is_ancestor(&memcg->css, &root_memcg->css));
1160 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *memcg)
1163 struct mem_cgroup *curr = NULL;
1164 struct task_struct *p;
1166 p = find_lock_task_mm(task);
1169 curr = try_get_mem_cgroup_from_mm(p->mm);
1174 * We should check use_hierarchy of "memcg" not "curr". Because checking
1175 * use_hierarchy of "curr" here make this function true if hierarchy is
1176 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1177 * hierarchy(even if use_hierarchy is disabled in "memcg").
1179 ret = mem_cgroup_same_or_subtree(memcg, curr);
1180 css_put(&curr->css);
1184 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg, struct zone *zone)
1186 unsigned long inactive_ratio;
1187 int nid = zone_to_nid(zone);
1188 int zid = zone_idx(zone);
1189 unsigned long inactive;
1190 unsigned long active;
1193 inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1194 BIT(LRU_INACTIVE_ANON));
1195 active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1196 BIT(LRU_ACTIVE_ANON));
1198 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1200 inactive_ratio = int_sqrt(10 * gb);
1204 return inactive * inactive_ratio < active;
1207 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg, struct zone *zone)
1209 unsigned long active;
1210 unsigned long inactive;
1211 int zid = zone_idx(zone);
1212 int nid = zone_to_nid(zone);
1214 inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1215 BIT(LRU_INACTIVE_FILE));
1216 active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1217 BIT(LRU_ACTIVE_FILE));
1219 return (active > inactive);
1222 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
1225 int nid = zone_to_nid(zone);
1226 int zid = zone_idx(zone);
1227 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
1229 return &mz->reclaim_stat;
1232 struct zone_reclaim_stat *
1233 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
1235 struct page_cgroup *pc;
1236 struct mem_cgroup_per_zone *mz;
1238 if (mem_cgroup_disabled())
1241 pc = lookup_page_cgroup(page);
1242 if (!PageCgroupUsed(pc))
1244 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1246 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1247 return &mz->reclaim_stat;
1250 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
1251 struct list_head *dst,
1252 unsigned long *scanned, int order,
1253 isolate_mode_t mode,
1255 struct mem_cgroup *mem_cont,
1256 int active, int file)
1258 unsigned long nr_taken = 0;
1262 struct list_head *src;
1263 struct page_cgroup *pc, *tmp;
1264 int nid = zone_to_nid(z);
1265 int zid = zone_idx(z);
1266 struct mem_cgroup_per_zone *mz;
1267 int lru = LRU_FILE * file + active;
1271 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
1272 src = &mz->lists[lru];
1275 list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
1276 if (scan >= nr_to_scan)
1279 if (unlikely(!PageCgroupUsed(pc)))
1282 page = lookup_cgroup_page(pc);
1284 if (unlikely(!PageLRU(page)))
1288 ret = __isolate_lru_page(page, mode, file);
1291 list_move(&page->lru, dst);
1292 mem_cgroup_del_lru(page);
1293 nr_taken += hpage_nr_pages(page);
1296 /* we don't affect global LRU but rotate in our LRU */
1297 mem_cgroup_rotate_lru_list(page, page_lru(page));
1306 trace_mm_vmscan_memcg_isolate(0, nr_to_scan, scan, nr_taken,
1312 #define mem_cgroup_from_res_counter(counter, member) \
1313 container_of(counter, struct mem_cgroup, member)
1316 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1317 * @mem: the memory cgroup
1319 * Returns the maximum amount of memory @mem can be charged with, in
1322 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1324 unsigned long long margin;
1326 margin = res_counter_margin(&memcg->res);
1327 if (do_swap_account)
1328 margin = min(margin, res_counter_margin(&memcg->memsw));
1329 return margin >> PAGE_SHIFT;
1332 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1334 struct cgroup *cgrp = memcg->css.cgroup;
1337 if (cgrp->parent == NULL)
1338 return vm_swappiness;
1340 return memcg->swappiness;
1343 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1348 spin_lock(&memcg->pcp_counter_lock);
1349 for_each_online_cpu(cpu)
1350 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) += 1;
1351 memcg->nocpu_base.count[MEM_CGROUP_ON_MOVE] += 1;
1352 spin_unlock(&memcg->pcp_counter_lock);
1358 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1365 spin_lock(&memcg->pcp_counter_lock);
1366 for_each_online_cpu(cpu)
1367 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) -= 1;
1368 memcg->nocpu_base.count[MEM_CGROUP_ON_MOVE] -= 1;
1369 spin_unlock(&memcg->pcp_counter_lock);
1373 * 2 routines for checking "mem" is under move_account() or not.
1375 * mem_cgroup_stealed() - checking a cgroup is mc.from or not. This is used
1376 * for avoiding race in accounting. If true,
1377 * pc->mem_cgroup may be overwritten.
1379 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1380 * under hierarchy of moving cgroups. This is for
1381 * waiting at hith-memory prressure caused by "move".
1384 static bool mem_cgroup_stealed(struct mem_cgroup *memcg)
1386 VM_BUG_ON(!rcu_read_lock_held());
1387 return this_cpu_read(memcg->stat->count[MEM_CGROUP_ON_MOVE]) > 0;
1390 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1392 struct mem_cgroup *from;
1393 struct mem_cgroup *to;
1396 * Unlike task_move routines, we access mc.to, mc.from not under
1397 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1399 spin_lock(&mc.lock);
1405 ret = mem_cgroup_same_or_subtree(memcg, from)
1406 || mem_cgroup_same_or_subtree(memcg, to);
1408 spin_unlock(&mc.lock);
1412 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1414 if (mc.moving_task && current != mc.moving_task) {
1415 if (mem_cgroup_under_move(memcg)) {
1417 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1418 /* moving charge context might have finished. */
1421 finish_wait(&mc.waitq, &wait);
1429 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1430 * @memcg: The memory cgroup that went over limit
1431 * @p: Task that is going to be killed
1433 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1436 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1438 struct cgroup *task_cgrp;
1439 struct cgroup *mem_cgrp;
1441 * Need a buffer in BSS, can't rely on allocations. The code relies
1442 * on the assumption that OOM is serialized for memory controller.
1443 * If this assumption is broken, revisit this code.
1445 static char memcg_name[PATH_MAX];
1454 mem_cgrp = memcg->css.cgroup;
1455 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1457 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1460 * Unfortunately, we are unable to convert to a useful name
1461 * But we'll still print out the usage information
1468 printk(KERN_INFO "Task in %s killed", memcg_name);
1471 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1479 * Continues from above, so we don't need an KERN_ level
1481 printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1484 printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1485 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1486 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1487 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1488 printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1490 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1491 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1492 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1496 * This function returns the number of memcg under hierarchy tree. Returns
1497 * 1(self count) if no children.
1499 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1502 struct mem_cgroup *iter;
1504 for_each_mem_cgroup_tree(iter, memcg)
1510 * Return the memory (and swap, if configured) limit for a memcg.
1512 u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1517 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1518 limit += total_swap_pages << PAGE_SHIFT;
1520 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1522 * If memsw is finite and limits the amount of swap space available
1523 * to this memcg, return that limit.
1525 return min(limit, memsw);
1529 * Visit the first child (need not be the first child as per the ordering
1530 * of the cgroup list, since we track last_scanned_child) of @mem and use
1531 * that to reclaim free pages from.
1533 static struct mem_cgroup *
1534 mem_cgroup_select_victim(struct mem_cgroup *root_memcg)
1536 struct mem_cgroup *ret = NULL;
1537 struct cgroup_subsys_state *css;
1540 if (!root_memcg->use_hierarchy) {
1541 css_get(&root_memcg->css);
1547 nextid = root_memcg->last_scanned_child + 1;
1548 css = css_get_next(&mem_cgroup_subsys, nextid, &root_memcg->css,
1550 if (css && css_tryget(css))
1551 ret = container_of(css, struct mem_cgroup, css);
1554 /* Updates scanning parameter */
1556 /* this means start scan from ID:1 */
1557 root_memcg->last_scanned_child = 0;
1559 root_memcg->last_scanned_child = found;
1566 * test_mem_cgroup_node_reclaimable
1567 * @mem: the target memcg
1568 * @nid: the node ID to be checked.
1569 * @noswap : specify true here if the user wants flle only information.
1571 * This function returns whether the specified memcg contains any
1572 * reclaimable pages on a node. Returns true if there are any reclaimable
1573 * pages in the node.
1575 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1576 int nid, bool noswap)
1578 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1580 if (noswap || !total_swap_pages)
1582 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1587 #if MAX_NUMNODES > 1
1590 * Always updating the nodemask is not very good - even if we have an empty
1591 * list or the wrong list here, we can start from some node and traverse all
1592 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1595 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1599 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1600 * pagein/pageout changes since the last update.
1602 if (!atomic_read(&memcg->numainfo_events))
1604 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1607 /* make a nodemask where this memcg uses memory from */
1608 memcg->scan_nodes = node_states[N_HIGH_MEMORY];
1610 for_each_node_mask(nid, node_states[N_HIGH_MEMORY]) {
1612 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1613 node_clear(nid, memcg->scan_nodes);
1616 atomic_set(&memcg->numainfo_events, 0);
1617 atomic_set(&memcg->numainfo_updating, 0);
1621 * Selecting a node where we start reclaim from. Because what we need is just
1622 * reducing usage counter, start from anywhere is O,K. Considering
1623 * memory reclaim from current node, there are pros. and cons.
1625 * Freeing memory from current node means freeing memory from a node which
1626 * we'll use or we've used. So, it may make LRU bad. And if several threads
1627 * hit limits, it will see a contention on a node. But freeing from remote
1628 * node means more costs for memory reclaim because of memory latency.
1630 * Now, we use round-robin. Better algorithm is welcomed.
1632 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1636 mem_cgroup_may_update_nodemask(memcg);
1637 node = memcg->last_scanned_node;
1639 node = next_node(node, memcg->scan_nodes);
1640 if (node == MAX_NUMNODES)
1641 node = first_node(memcg->scan_nodes);
1643 * We call this when we hit limit, not when pages are added to LRU.
1644 * No LRU may hold pages because all pages are UNEVICTABLE or
1645 * memcg is too small and all pages are not on LRU. In that case,
1646 * we use curret node.
1648 if (unlikely(node == MAX_NUMNODES))
1649 node = numa_node_id();
1651 memcg->last_scanned_node = node;
1656 * Check all nodes whether it contains reclaimable pages or not.
1657 * For quick scan, we make use of scan_nodes. This will allow us to skip
1658 * unused nodes. But scan_nodes is lazily updated and may not cotain
1659 * enough new information. We need to do double check.
1661 bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1666 * quick check...making use of scan_node.
1667 * We can skip unused nodes.
1669 if (!nodes_empty(memcg->scan_nodes)) {
1670 for (nid = first_node(memcg->scan_nodes);
1672 nid = next_node(nid, memcg->scan_nodes)) {
1674 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1679 * Check rest of nodes.
1681 for_each_node_state(nid, N_HIGH_MEMORY) {
1682 if (node_isset(nid, memcg->scan_nodes))
1684 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1691 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1696 bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1698 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
1703 * Scan the hierarchy if needed to reclaim memory. We remember the last child
1704 * we reclaimed from, so that we don't end up penalizing one child extensively
1705 * based on its position in the children list.
1707 * root_memcg is the original ancestor that we've been reclaim from.
1709 * We give up and return to the caller when we visit root_memcg twice.
1710 * (other groups can be removed while we're walking....)
1712 * If shrink==true, for avoiding to free too much, this returns immedieately.
1714 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_memcg,
1717 unsigned long reclaim_options,
1718 unsigned long *total_scanned)
1720 struct mem_cgroup *victim;
1723 bool noswap = reclaim_options & MEM_CGROUP_RECLAIM_NOSWAP;
1724 bool shrink = reclaim_options & MEM_CGROUP_RECLAIM_SHRINK;
1725 bool check_soft = reclaim_options & MEM_CGROUP_RECLAIM_SOFT;
1726 unsigned long excess;
1727 unsigned long nr_scanned;
1729 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
1731 /* If memsw_is_minimum==1, swap-out is of-no-use. */
1732 if (!check_soft && !shrink && root_memcg->memsw_is_minimum)
1736 victim = mem_cgroup_select_victim(root_memcg);
1737 if (victim == root_memcg) {
1740 * We are not draining per cpu cached charges during
1741 * soft limit reclaim because global reclaim doesn't
1742 * care about charges. It tries to free some memory and
1743 * charges will not give any.
1745 if (!check_soft && loop >= 1)
1746 drain_all_stock_async(root_memcg);
1749 * If we have not been able to reclaim
1750 * anything, it might because there are
1751 * no reclaimable pages under this hierarchy
1753 if (!check_soft || !total) {
1754 css_put(&victim->css);
1758 * We want to do more targeted reclaim.
1759 * excess >> 2 is not to excessive so as to
1760 * reclaim too much, nor too less that we keep
1761 * coming back to reclaim from this cgroup
1763 if (total >= (excess >> 2) ||
1764 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS)) {
1765 css_put(&victim->css);
1770 if (!mem_cgroup_reclaimable(victim, noswap)) {
1771 /* this cgroup's local usage == 0 */
1772 css_put(&victim->css);
1775 /* we use swappiness of local cgroup */
1777 ret = mem_cgroup_shrink_node_zone(victim, gfp_mask,
1778 noswap, zone, &nr_scanned);
1779 *total_scanned += nr_scanned;
1781 ret = try_to_free_mem_cgroup_pages(victim, gfp_mask,
1783 css_put(&victim->css);
1785 * At shrinking usage, we can't check we should stop here or
1786 * reclaim more. It's depends on callers. last_scanned_child
1787 * will work enough for keeping fairness under tree.
1793 if (!res_counter_soft_limit_excess(&root_memcg->res))
1795 } else if (mem_cgroup_margin(root_memcg))
1802 * Check OOM-Killer is already running under our hierarchy.
1803 * If someone is running, return false.
1804 * Has to be called with memcg_oom_lock
1806 static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg)
1808 struct mem_cgroup *iter, *failed = NULL;
1811 for_each_mem_cgroup_tree_cond(iter, memcg, cond) {
1812 if (iter->oom_lock) {
1814 * this subtree of our hierarchy is already locked
1815 * so we cannot give a lock.
1820 iter->oom_lock = true;
1827 * OK, we failed to lock the whole subtree so we have to clean up
1828 * what we set up to the failing subtree
1831 for_each_mem_cgroup_tree_cond(iter, memcg, cond) {
1832 if (iter == failed) {
1836 iter->oom_lock = false;
1842 * Has to be called with memcg_oom_lock
1844 static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1846 struct mem_cgroup *iter;
1848 for_each_mem_cgroup_tree(iter, memcg)
1849 iter->oom_lock = false;
1853 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1855 struct mem_cgroup *iter;
1857 for_each_mem_cgroup_tree(iter, memcg)
1858 atomic_inc(&iter->under_oom);
1861 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1863 struct mem_cgroup *iter;
1866 * When a new child is created while the hierarchy is under oom,
1867 * mem_cgroup_oom_lock() may not be called. We have to use
1868 * atomic_add_unless() here.
1870 for_each_mem_cgroup_tree(iter, memcg)
1871 atomic_add_unless(&iter->under_oom, -1, 0);
1874 static DEFINE_SPINLOCK(memcg_oom_lock);
1875 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1877 struct oom_wait_info {
1878 struct mem_cgroup *mem;
1882 static int memcg_oom_wake_function(wait_queue_t *wait,
1883 unsigned mode, int sync, void *arg)
1885 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg,
1887 struct oom_wait_info *oom_wait_info;
1889 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1890 oom_wait_memcg = oom_wait_info->mem;
1893 * Both of oom_wait_info->mem and wake_mem are stable under us.
1894 * Then we can use css_is_ancestor without taking care of RCU.
1896 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
1897 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
1899 return autoremove_wake_function(wait, mode, sync, arg);
1902 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
1904 /* for filtering, pass "memcg" as argument. */
1905 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1908 static void memcg_oom_recover(struct mem_cgroup *memcg)
1910 if (memcg && atomic_read(&memcg->under_oom))
1911 memcg_wakeup_oom(memcg);
1915 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1917 bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask)
1919 struct oom_wait_info owait;
1920 bool locked, need_to_kill;
1923 owait.wait.flags = 0;
1924 owait.wait.func = memcg_oom_wake_function;
1925 owait.wait.private = current;
1926 INIT_LIST_HEAD(&owait.wait.task_list);
1927 need_to_kill = true;
1928 mem_cgroup_mark_under_oom(memcg);
1930 /* At first, try to OOM lock hierarchy under memcg.*/
1931 spin_lock(&memcg_oom_lock);
1932 locked = mem_cgroup_oom_lock(memcg);
1934 * Even if signal_pending(), we can't quit charge() loop without
1935 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1936 * under OOM is always welcomed, use TASK_KILLABLE here.
1938 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1939 if (!locked || memcg->oom_kill_disable)
1940 need_to_kill = false;
1942 mem_cgroup_oom_notify(memcg);
1943 spin_unlock(&memcg_oom_lock);
1946 finish_wait(&memcg_oom_waitq, &owait.wait);
1947 mem_cgroup_out_of_memory(memcg, mask);
1950 finish_wait(&memcg_oom_waitq, &owait.wait);
1952 spin_lock(&memcg_oom_lock);
1954 mem_cgroup_oom_unlock(memcg);
1955 memcg_wakeup_oom(memcg);
1956 spin_unlock(&memcg_oom_lock);
1958 mem_cgroup_unmark_under_oom(memcg);
1960 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
1962 /* Give chance to dying process */
1963 schedule_timeout_uninterruptible(1);
1968 * Currently used to update mapped file statistics, but the routine can be
1969 * generalized to update other statistics as well.
1971 * Notes: Race condition
1973 * We usually use page_cgroup_lock() for accessing page_cgroup member but
1974 * it tends to be costly. But considering some conditions, we doesn't need
1975 * to do so _always_.
1977 * Considering "charge", lock_page_cgroup() is not required because all
1978 * file-stat operations happen after a page is attached to radix-tree. There
1979 * are no race with "charge".
1981 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
1982 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
1983 * if there are race with "uncharge". Statistics itself is properly handled
1986 * Considering "move", this is an only case we see a race. To make the race
1987 * small, we check MEM_CGROUP_ON_MOVE percpu value and detect there are
1988 * possibility of race condition. If there is, we take a lock.
1991 void mem_cgroup_update_page_stat(struct page *page,
1992 enum mem_cgroup_page_stat_item idx, int val)
1994 struct mem_cgroup *memcg;
1995 struct page_cgroup *pc = lookup_page_cgroup(page);
1996 bool need_unlock = false;
1997 unsigned long uninitialized_var(flags);
2003 memcg = pc->mem_cgroup;
2004 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2006 /* pc->mem_cgroup is unstable ? */
2007 if (unlikely(mem_cgroup_stealed(memcg)) || PageTransHuge(page)) {
2008 /* take a lock against to access pc->mem_cgroup */
2009 move_lock_page_cgroup(pc, &flags);
2011 memcg = pc->mem_cgroup;
2012 if (!memcg || !PageCgroupUsed(pc))
2017 case MEMCG_NR_FILE_MAPPED:
2019 SetPageCgroupFileMapped(pc);
2020 else if (!page_mapped(page))
2021 ClearPageCgroupFileMapped(pc);
2022 idx = MEM_CGROUP_STAT_FILE_MAPPED;
2028 this_cpu_add(memcg->stat->count[idx], val);
2031 if (unlikely(need_unlock))
2032 move_unlock_page_cgroup(pc, &flags);
2036 EXPORT_SYMBOL(mem_cgroup_update_page_stat);
2039 * size of first charge trial. "32" comes from vmscan.c's magic value.
2040 * TODO: maybe necessary to use big numbers in big irons.
2042 #define CHARGE_BATCH 32U
2043 struct memcg_stock_pcp {
2044 struct mem_cgroup *cached; /* this never be root cgroup */
2045 unsigned int nr_pages;
2046 struct work_struct work;
2047 unsigned long flags;
2048 #define FLUSHING_CACHED_CHARGE (0)
2050 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2051 static DEFINE_MUTEX(percpu_charge_mutex);
2054 * Try to consume stocked charge on this cpu. If success, one page is consumed
2055 * from local stock and true is returned. If the stock is 0 or charges from a
2056 * cgroup which is not current target, returns false. This stock will be
2059 static bool consume_stock(struct mem_cgroup *memcg)
2061 struct memcg_stock_pcp *stock;
2064 stock = &get_cpu_var(memcg_stock);
2065 if (memcg == stock->cached && stock->nr_pages)
2067 else /* need to call res_counter_charge */
2069 put_cpu_var(memcg_stock);
2074 * Returns stocks cached in percpu to res_counter and reset cached information.
2076 static void drain_stock(struct memcg_stock_pcp *stock)
2078 struct mem_cgroup *old = stock->cached;
2080 if (stock->nr_pages) {
2081 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2083 res_counter_uncharge(&old->res, bytes);
2084 if (do_swap_account)
2085 res_counter_uncharge(&old->memsw, bytes);
2086 stock->nr_pages = 0;
2088 stock->cached = NULL;
2092 * This must be called under preempt disabled or must be called by
2093 * a thread which is pinned to local cpu.
2095 static void drain_local_stock(struct work_struct *dummy)
2097 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2099 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2103 * Cache charges(val) which is from res_counter, to local per_cpu area.
2104 * This will be consumed by consume_stock() function, later.
2106 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2108 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2110 if (stock->cached != memcg) { /* reset if necessary */
2112 stock->cached = memcg;
2114 stock->nr_pages += nr_pages;
2115 put_cpu_var(memcg_stock);
2119 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2120 * of the hierarchy under it. sync flag says whether we should block
2121 * until the work is done.
2123 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2127 /* Notify other cpus that system-wide "drain" is running */
2130 for_each_online_cpu(cpu) {
2131 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2132 struct mem_cgroup *memcg;
2134 memcg = stock->cached;
2135 if (!memcg || !stock->nr_pages)
2137 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2139 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2141 drain_local_stock(&stock->work);
2143 schedule_work_on(cpu, &stock->work);
2151 for_each_online_cpu(cpu) {
2152 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2153 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2154 flush_work(&stock->work);
2161 * Tries to drain stocked charges in other cpus. This function is asynchronous
2162 * and just put a work per cpu for draining localy on each cpu. Caller can
2163 * expects some charges will be back to res_counter later but cannot wait for
2166 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2169 * If someone calls draining, avoid adding more kworker runs.
2171 if (!mutex_trylock(&percpu_charge_mutex))
2173 drain_all_stock(root_memcg, false);
2174 mutex_unlock(&percpu_charge_mutex);
2177 /* This is a synchronous drain interface. */
2178 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2180 /* called when force_empty is called */
2181 mutex_lock(&percpu_charge_mutex);
2182 drain_all_stock(root_memcg, true);
2183 mutex_unlock(&percpu_charge_mutex);
2187 * This function drains percpu counter value from DEAD cpu and
2188 * move it to local cpu. Note that this function can be preempted.
2190 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2194 spin_lock(&memcg->pcp_counter_lock);
2195 for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) {
2196 long x = per_cpu(memcg->stat->count[i], cpu);
2198 per_cpu(memcg->stat->count[i], cpu) = 0;
2199 memcg->nocpu_base.count[i] += x;
2201 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2202 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2204 per_cpu(memcg->stat->events[i], cpu) = 0;
2205 memcg->nocpu_base.events[i] += x;
2207 /* need to clear ON_MOVE value, works as a kind of lock. */
2208 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) = 0;
2209 spin_unlock(&memcg->pcp_counter_lock);
2212 static void synchronize_mem_cgroup_on_move(struct mem_cgroup *memcg, int cpu)
2214 int idx = MEM_CGROUP_ON_MOVE;
2216 spin_lock(&memcg->pcp_counter_lock);
2217 per_cpu(memcg->stat->count[idx], cpu) = memcg->nocpu_base.count[idx];
2218 spin_unlock(&memcg->pcp_counter_lock);
2221 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2222 unsigned long action,
2225 int cpu = (unsigned long)hcpu;
2226 struct memcg_stock_pcp *stock;
2227 struct mem_cgroup *iter;
2229 if ((action == CPU_ONLINE)) {
2230 for_each_mem_cgroup_all(iter)
2231 synchronize_mem_cgroup_on_move(iter, cpu);
2235 if ((action != CPU_DEAD) || action != CPU_DEAD_FROZEN)
2238 for_each_mem_cgroup_all(iter)
2239 mem_cgroup_drain_pcp_counter(iter, cpu);
2241 stock = &per_cpu(memcg_stock, cpu);
2247 /* See __mem_cgroup_try_charge() for details */
2249 CHARGE_OK, /* success */
2250 CHARGE_RETRY, /* need to retry but retry is not bad */
2251 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2252 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2253 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2256 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2257 unsigned int nr_pages, bool oom_check)
2259 unsigned long csize = nr_pages * PAGE_SIZE;
2260 struct mem_cgroup *mem_over_limit;
2261 struct res_counter *fail_res;
2262 unsigned long flags = 0;
2265 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2268 if (!do_swap_account)
2270 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2274 res_counter_uncharge(&memcg->res, csize);
2275 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2276 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2278 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2280 * nr_pages can be either a huge page (HPAGE_PMD_NR), a batch
2281 * of regular pages (CHARGE_BATCH), or a single regular page (1).
2283 * Never reclaim on behalf of optional batching, retry with a
2284 * single page instead.
2286 if (nr_pages == CHARGE_BATCH)
2287 return CHARGE_RETRY;
2289 if (!(gfp_mask & __GFP_WAIT))
2290 return CHARGE_WOULDBLOCK;
2292 ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, NULL,
2293 gfp_mask, flags, NULL);
2294 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2295 return CHARGE_RETRY;
2297 * Even though the limit is exceeded at this point, reclaim
2298 * may have been able to free some pages. Retry the charge
2299 * before killing the task.
2301 * Only for regular pages, though: huge pages are rather
2302 * unlikely to succeed so close to the limit, and we fall back
2303 * to regular pages anyway in case of failure.
2305 if (nr_pages == 1 && ret)
2306 return CHARGE_RETRY;
2309 * At task move, charge accounts can be doubly counted. So, it's
2310 * better to wait until the end of task_move if something is going on.
2312 if (mem_cgroup_wait_acct_move(mem_over_limit))
2313 return CHARGE_RETRY;
2315 /* If we don't need to call oom-killer at el, return immediately */
2317 return CHARGE_NOMEM;
2319 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask))
2320 return CHARGE_OOM_DIE;
2322 return CHARGE_RETRY;
2326 * Unlike exported interface, "oom" parameter is added. if oom==true,
2327 * oom-killer can be invoked.
2329 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2331 unsigned int nr_pages,
2332 struct mem_cgroup **ptr,
2335 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2336 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2337 struct mem_cgroup *memcg = NULL;
2341 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2342 * in system level. So, allow to go ahead dying process in addition to
2345 if (unlikely(test_thread_flag(TIF_MEMDIE)
2346 || fatal_signal_pending(current)))
2350 * We always charge the cgroup the mm_struct belongs to.
2351 * The mm_struct's mem_cgroup changes on task migration if the
2352 * thread group leader migrates. It's possible that mm is not
2353 * set, if so charge the init_mm (happens for pagecache usage).
2358 if (*ptr) { /* css should be a valid one */
2360 VM_BUG_ON(css_is_removed(&memcg->css));
2361 if (mem_cgroup_is_root(memcg))
2363 if (nr_pages == 1 && consume_stock(memcg))
2365 css_get(&memcg->css);
2367 struct task_struct *p;
2370 p = rcu_dereference(mm->owner);
2372 * Because we don't have task_lock(), "p" can exit.
2373 * In that case, "memcg" can point to root or p can be NULL with
2374 * race with swapoff. Then, we have small risk of mis-accouning.
2375 * But such kind of mis-account by race always happens because
2376 * we don't have cgroup_mutex(). It's overkill and we allo that
2378 * (*) swapoff at el will charge against mm-struct not against
2379 * task-struct. So, mm->owner can be NULL.
2381 memcg = mem_cgroup_from_task(p);
2382 if (!memcg || mem_cgroup_is_root(memcg)) {
2386 if (nr_pages == 1 && consume_stock(memcg)) {
2388 * It seems dagerous to access memcg without css_get().
2389 * But considering how consume_stok works, it's not
2390 * necessary. If consume_stock success, some charges
2391 * from this memcg are cached on this cpu. So, we
2392 * don't need to call css_get()/css_tryget() before
2393 * calling consume_stock().
2398 /* after here, we may be blocked. we need to get refcnt */
2399 if (!css_tryget(&memcg->css)) {
2409 /* If killed, bypass charge */
2410 if (fatal_signal_pending(current)) {
2411 css_put(&memcg->css);
2416 if (oom && !nr_oom_retries) {
2418 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2421 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, oom_check);
2425 case CHARGE_RETRY: /* not in OOM situation but retry */
2427 css_put(&memcg->css);
2430 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2431 css_put(&memcg->css);
2433 case CHARGE_NOMEM: /* OOM routine works */
2435 css_put(&memcg->css);
2438 /* If oom, we never return -ENOMEM */
2441 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2442 css_put(&memcg->css);
2445 } while (ret != CHARGE_OK);
2447 if (batch > nr_pages)
2448 refill_stock(memcg, batch - nr_pages);
2449 css_put(&memcg->css);
2462 * Somemtimes we have to undo a charge we got by try_charge().
2463 * This function is for that and do uncharge, put css's refcnt.
2464 * gotten by try_charge().
2466 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2467 unsigned int nr_pages)
2469 if (!mem_cgroup_is_root(memcg)) {
2470 unsigned long bytes = nr_pages * PAGE_SIZE;
2472 res_counter_uncharge(&memcg->res, bytes);
2473 if (do_swap_account)
2474 res_counter_uncharge(&memcg->memsw, bytes);
2479 * A helper function to get mem_cgroup from ID. must be called under
2480 * rcu_read_lock(). The caller must check css_is_removed() or some if
2481 * it's concern. (dropping refcnt from swap can be called against removed
2484 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2486 struct cgroup_subsys_state *css;
2488 /* ID 0 is unused ID */
2491 css = css_lookup(&mem_cgroup_subsys, id);
2494 return container_of(css, struct mem_cgroup, css);
2497 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2499 struct mem_cgroup *memcg = NULL;
2500 struct page_cgroup *pc;
2504 VM_BUG_ON(!PageLocked(page));
2506 pc = lookup_page_cgroup(page);
2507 lock_page_cgroup(pc);
2508 if (PageCgroupUsed(pc)) {
2509 memcg = pc->mem_cgroup;
2510 if (memcg && !css_tryget(&memcg->css))
2512 } else if (PageSwapCache(page)) {
2513 ent.val = page_private(page);
2514 id = lookup_swap_cgroup(ent);
2516 memcg = mem_cgroup_lookup(id);
2517 if (memcg && !css_tryget(&memcg->css))
2521 unlock_page_cgroup(pc);
2525 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2527 unsigned int nr_pages,
2528 struct page_cgroup *pc,
2529 enum charge_type ctype)
2531 lock_page_cgroup(pc);
2532 if (unlikely(PageCgroupUsed(pc))) {
2533 unlock_page_cgroup(pc);
2534 __mem_cgroup_cancel_charge(memcg, nr_pages);
2538 * we don't need page_cgroup_lock about tail pages, becase they are not
2539 * accessed by any other context at this point.
2541 pc->mem_cgroup = memcg;
2543 * We access a page_cgroup asynchronously without lock_page_cgroup().
2544 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2545 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2546 * before USED bit, we need memory barrier here.
2547 * See mem_cgroup_add_lru_list(), etc.
2551 case MEM_CGROUP_CHARGE_TYPE_CACHE:
2552 case MEM_CGROUP_CHARGE_TYPE_SHMEM:
2553 SetPageCgroupCache(pc);
2554 SetPageCgroupUsed(pc);
2556 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2557 ClearPageCgroupCache(pc);
2558 SetPageCgroupUsed(pc);
2564 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), nr_pages);
2565 unlock_page_cgroup(pc);
2567 * "charge_statistics" updated event counter. Then, check it.
2568 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2569 * if they exceeds softlimit.
2571 memcg_check_events(memcg, page);
2574 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2576 #define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MOVE_LOCK) |\
2577 (1 << PCG_ACCT_LRU) | (1 << PCG_MIGRATION))
2579 * Because tail pages are not marked as "used", set it. We're under
2580 * zone->lru_lock, 'splitting on pmd' and compund_lock.
2582 void mem_cgroup_split_huge_fixup(struct page *head, struct page *tail)
2584 struct page_cgroup *head_pc = lookup_page_cgroup(head);
2585 struct page_cgroup *tail_pc = lookup_page_cgroup(tail);
2586 unsigned long flags;
2588 if (mem_cgroup_disabled())
2591 * We have no races with charge/uncharge but will have races with
2592 * page state accounting.
2594 move_lock_page_cgroup(head_pc, &flags);
2596 tail_pc->mem_cgroup = head_pc->mem_cgroup;
2597 smp_wmb(); /* see __commit_charge() */
2598 if (PageCgroupAcctLRU(head_pc)) {
2600 struct mem_cgroup_per_zone *mz;
2603 * LRU flags cannot be copied because we need to add tail
2604 *.page to LRU by generic call and our hook will be called.
2605 * We hold lru_lock, then, reduce counter directly.
2607 lru = page_lru(head);
2608 mz = page_cgroup_zoneinfo(head_pc->mem_cgroup, head);
2609 MEM_CGROUP_ZSTAT(mz, lru) -= 1;
2611 tail_pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2612 move_unlock_page_cgroup(head_pc, &flags);
2617 * mem_cgroup_move_account - move account of the page
2619 * @nr_pages: number of regular pages (>1 for huge pages)
2620 * @pc: page_cgroup of the page.
2621 * @from: mem_cgroup which the page is moved from.
2622 * @to: mem_cgroup which the page is moved to. @from != @to.
2623 * @uncharge: whether we should call uncharge and css_put against @from.
2625 * The caller must confirm following.
2626 * - page is not on LRU (isolate_page() is useful.)
2627 * - compound_lock is held when nr_pages > 1
2629 * This function doesn't do "charge" nor css_get to new cgroup. It should be
2630 * done by a caller(__mem_cgroup_try_charge would be useful). If @uncharge is
2631 * true, this function does "uncharge" from old cgroup, but it doesn't if
2632 * @uncharge is false, so a caller should do "uncharge".
2634 static int mem_cgroup_move_account(struct page *page,
2635 unsigned int nr_pages,
2636 struct page_cgroup *pc,
2637 struct mem_cgroup *from,
2638 struct mem_cgroup *to,
2641 unsigned long flags;
2644 VM_BUG_ON(from == to);
2645 VM_BUG_ON(PageLRU(page));
2647 * The page is isolated from LRU. So, collapse function
2648 * will not handle this page. But page splitting can happen.
2649 * Do this check under compound_page_lock(). The caller should
2653 if (nr_pages > 1 && !PageTransHuge(page))
2656 lock_page_cgroup(pc);
2659 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
2662 move_lock_page_cgroup(pc, &flags);
2664 if (PageCgroupFileMapped(pc)) {
2665 /* Update mapped_file data for mem_cgroup */
2667 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2668 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2671 mem_cgroup_charge_statistics(from, PageCgroupCache(pc), -nr_pages);
2673 /* This is not "cancel", but cancel_charge does all we need. */
2674 __mem_cgroup_cancel_charge(from, nr_pages);
2676 /* caller should have done css_get */
2677 pc->mem_cgroup = to;
2678 mem_cgroup_charge_statistics(to, PageCgroupCache(pc), nr_pages);
2680 * We charges against "to" which may not have any tasks. Then, "to"
2681 * can be under rmdir(). But in current implementation, caller of
2682 * this function is just force_empty() and move charge, so it's
2683 * guaranteed that "to" is never removed. So, we don't check rmdir
2686 move_unlock_page_cgroup(pc, &flags);
2689 unlock_page_cgroup(pc);
2693 memcg_check_events(to, page);
2694 memcg_check_events(from, page);
2700 * move charges to its parent.
2703 static int mem_cgroup_move_parent(struct page *page,
2704 struct page_cgroup *pc,
2705 struct mem_cgroup *child,
2708 struct cgroup *cg = child->css.cgroup;
2709 struct cgroup *pcg = cg->parent;
2710 struct mem_cgroup *parent;
2711 unsigned int nr_pages;
2712 unsigned long uninitialized_var(flags);
2720 if (!get_page_unless_zero(page))
2722 if (isolate_lru_page(page))
2725 nr_pages = hpage_nr_pages(page);
2727 parent = mem_cgroup_from_cont(pcg);
2728 ret = __mem_cgroup_try_charge(NULL, gfp_mask, nr_pages, &parent, false);
2733 flags = compound_lock_irqsave(page);
2735 ret = mem_cgroup_move_account(page, nr_pages, pc, child, parent, true);
2737 __mem_cgroup_cancel_charge(parent, nr_pages);
2740 compound_unlock_irqrestore(page, flags);
2742 putback_lru_page(page);
2750 * Charge the memory controller for page usage.
2752 * 0 if the charge was successful
2753 * < 0 if the cgroup is over its limit
2755 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2756 gfp_t gfp_mask, enum charge_type ctype)
2758 struct mem_cgroup *memcg = NULL;
2759 unsigned int nr_pages = 1;
2760 struct page_cgroup *pc;
2764 if (PageTransHuge(page)) {
2765 nr_pages <<= compound_order(page);
2766 VM_BUG_ON(!PageTransHuge(page));
2768 * Never OOM-kill a process for a huge page. The
2769 * fault handler will fall back to regular pages.
2774 pc = lookup_page_cgroup(page);
2775 BUG_ON(!pc); /* XXX: remove this and move pc lookup into commit */
2777 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
2781 __mem_cgroup_commit_charge(memcg, page, nr_pages, pc, ctype);
2785 int mem_cgroup_newpage_charge(struct page *page,
2786 struct mm_struct *mm, gfp_t gfp_mask)
2788 if (mem_cgroup_disabled())
2791 * If already mapped, we don't have to account.
2792 * If page cache, page->mapping has address_space.
2793 * But page->mapping may have out-of-use anon_vma pointer,
2794 * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
2797 if (page_mapped(page) || (page->mapping && !PageAnon(page)))
2801 return mem_cgroup_charge_common(page, mm, gfp_mask,
2802 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2806 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2807 enum charge_type ctype);
2810 __mem_cgroup_commit_charge_lrucare(struct page *page, struct mem_cgroup *memcg,
2811 enum charge_type ctype)
2813 struct page_cgroup *pc = lookup_page_cgroup(page);
2815 * In some case, SwapCache, FUSE(splice_buf->radixtree), the page
2816 * is already on LRU. It means the page may on some other page_cgroup's
2817 * LRU. Take care of it.
2819 mem_cgroup_lru_del_before_commit(page);
2820 __mem_cgroup_commit_charge(memcg, page, 1, pc, ctype);
2821 mem_cgroup_lru_add_after_commit(page);
2825 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2828 struct mem_cgroup *memcg = NULL;
2831 if (mem_cgroup_disabled())
2833 if (PageCompound(page))
2839 if (page_is_file_cache(page)) {
2840 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, &memcg, true);
2845 * FUSE reuses pages without going through the final
2846 * put that would remove them from the LRU list, make
2847 * sure that they get relinked properly.
2849 __mem_cgroup_commit_charge_lrucare(page, memcg,
2850 MEM_CGROUP_CHARGE_TYPE_CACHE);
2854 if (PageSwapCache(page)) {
2855 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &memcg);
2857 __mem_cgroup_commit_charge_swapin(page, memcg,
2858 MEM_CGROUP_CHARGE_TYPE_SHMEM);
2860 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
2861 MEM_CGROUP_CHARGE_TYPE_SHMEM);
2867 * While swap-in, try_charge -> commit or cancel, the page is locked.
2868 * And when try_charge() successfully returns, one refcnt to memcg without
2869 * struct page_cgroup is acquired. This refcnt will be consumed by
2870 * "commit()" or removed by "cancel()"
2872 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2874 gfp_t mask, struct mem_cgroup **ptr)
2876 struct mem_cgroup *memcg;
2881 if (mem_cgroup_disabled())
2884 if (!do_swap_account)
2887 * A racing thread's fault, or swapoff, may have already updated
2888 * the pte, and even removed page from swap cache: in those cases
2889 * do_swap_page()'s pte_same() test will fail; but there's also a
2890 * KSM case which does need to charge the page.
2892 if (!PageSwapCache(page))
2894 memcg = try_get_mem_cgroup_from_page(page);
2898 ret = __mem_cgroup_try_charge(NULL, mask, 1, ptr, true);
2899 css_put(&memcg->css);
2904 return __mem_cgroup_try_charge(mm, mask, 1, ptr, true);
2908 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2909 enum charge_type ctype)
2911 if (mem_cgroup_disabled())
2915 cgroup_exclude_rmdir(&ptr->css);
2917 __mem_cgroup_commit_charge_lrucare(page, ptr, ctype);
2919 * Now swap is on-memory. This means this page may be
2920 * counted both as mem and swap....double count.
2921 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2922 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2923 * may call delete_from_swap_cache() before reach here.
2925 if (do_swap_account && PageSwapCache(page)) {
2926 swp_entry_t ent = {.val = page_private(page)};
2928 struct mem_cgroup *memcg;
2930 id = swap_cgroup_record(ent, 0);
2932 memcg = mem_cgroup_lookup(id);
2935 * This recorded memcg can be obsolete one. So, avoid
2936 * calling css_tryget
2938 if (!mem_cgroup_is_root(memcg))
2939 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2940 mem_cgroup_swap_statistics(memcg, false);
2941 mem_cgroup_put(memcg);
2946 * At swapin, we may charge account against cgroup which has no tasks.
2947 * So, rmdir()->pre_destroy() can be called while we do this charge.
2948 * In that case, we need to call pre_destroy() again. check it here.
2950 cgroup_release_and_wakeup_rmdir(&ptr->css);
2953 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
2955 __mem_cgroup_commit_charge_swapin(page, ptr,
2956 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2959 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
2961 if (mem_cgroup_disabled())
2965 __mem_cgroup_cancel_charge(memcg, 1);
2968 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
2969 unsigned int nr_pages,
2970 const enum charge_type ctype)
2972 struct memcg_batch_info *batch = NULL;
2973 bool uncharge_memsw = true;
2975 /* If swapout, usage of swap doesn't decrease */
2976 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2977 uncharge_memsw = false;
2979 batch = ¤t->memcg_batch;
2981 * In usual, we do css_get() when we remember memcg pointer.
2982 * But in this case, we keep res->usage until end of a series of
2983 * uncharges. Then, it's ok to ignore memcg's refcnt.
2986 batch->memcg = memcg;
2988 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2989 * In those cases, all pages freed continuously can be expected to be in
2990 * the same cgroup and we have chance to coalesce uncharges.
2991 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2992 * because we want to do uncharge as soon as possible.
2995 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
2996 goto direct_uncharge;
2999 goto direct_uncharge;
3002 * In typical case, batch->memcg == mem. This means we can
3003 * merge a series of uncharges to an uncharge of res_counter.
3004 * If not, we uncharge res_counter ony by one.
3006 if (batch->memcg != memcg)
3007 goto direct_uncharge;
3008 /* remember freed charge and uncharge it later */
3011 batch->memsw_nr_pages++;
3014 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
3016 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
3017 if (unlikely(batch->memcg != memcg))
3018 memcg_oom_recover(memcg);
3023 * uncharge if !page_mapped(page)
3025 static struct mem_cgroup *
3026 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
3028 struct mem_cgroup *memcg = NULL;
3029 unsigned int nr_pages = 1;
3030 struct page_cgroup *pc;
3032 if (mem_cgroup_disabled())
3035 if (PageSwapCache(page))
3038 if (PageTransHuge(page)) {
3039 nr_pages <<= compound_order(page);
3040 VM_BUG_ON(!PageTransHuge(page));
3043 * Check if our page_cgroup is valid
3045 pc = lookup_page_cgroup(page);
3046 if (unlikely(!pc || !PageCgroupUsed(pc)))
3049 lock_page_cgroup(pc);
3051 memcg = pc->mem_cgroup;
3053 if (!PageCgroupUsed(pc))
3057 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
3058 case MEM_CGROUP_CHARGE_TYPE_DROP:
3059 /* See mem_cgroup_prepare_migration() */
3060 if (page_mapped(page) || PageCgroupMigration(pc))
3063 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
3064 if (!PageAnon(page)) { /* Shared memory */
3065 if (page->mapping && !page_is_file_cache(page))
3067 } else if (page_mapped(page)) /* Anon */
3074 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), -nr_pages);
3076 ClearPageCgroupUsed(pc);
3078 * pc->mem_cgroup is not cleared here. It will be accessed when it's
3079 * freed from LRU. This is safe because uncharged page is expected not
3080 * to be reused (freed soon). Exception is SwapCache, it's handled by
3081 * special functions.
3084 unlock_page_cgroup(pc);
3086 * even after unlock, we have memcg->res.usage here and this memcg
3087 * will never be freed.
3089 memcg_check_events(memcg, page);
3090 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
3091 mem_cgroup_swap_statistics(memcg, true);
3092 mem_cgroup_get(memcg);
3094 if (!mem_cgroup_is_root(memcg))
3095 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
3100 unlock_page_cgroup(pc);
3104 void mem_cgroup_uncharge_page(struct page *page)
3107 if (page_mapped(page))
3109 if (page->mapping && !PageAnon(page))
3111 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
3114 void mem_cgroup_uncharge_cache_page(struct page *page)
3116 VM_BUG_ON(page_mapped(page));
3117 VM_BUG_ON(page->mapping);
3118 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
3122 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
3123 * In that cases, pages are freed continuously and we can expect pages
3124 * are in the same memcg. All these calls itself limits the number of
3125 * pages freed at once, then uncharge_start/end() is called properly.
3126 * This may be called prural(2) times in a context,
3129 void mem_cgroup_uncharge_start(void)
3131 current->memcg_batch.do_batch++;
3132 /* We can do nest. */
3133 if (current->memcg_batch.do_batch == 1) {
3134 current->memcg_batch.memcg = NULL;
3135 current->memcg_batch.nr_pages = 0;
3136 current->memcg_batch.memsw_nr_pages = 0;
3140 void mem_cgroup_uncharge_end(void)
3142 struct memcg_batch_info *batch = ¤t->memcg_batch;
3144 if (!batch->do_batch)
3148 if (batch->do_batch) /* If stacked, do nothing. */
3154 * This "batch->memcg" is valid without any css_get/put etc...
3155 * bacause we hide charges behind us.
3157 if (batch->nr_pages)
3158 res_counter_uncharge(&batch->memcg->res,
3159 batch->nr_pages * PAGE_SIZE);
3160 if (batch->memsw_nr_pages)
3161 res_counter_uncharge(&batch->memcg->memsw,
3162 batch->memsw_nr_pages * PAGE_SIZE);
3163 memcg_oom_recover(batch->memcg);
3164 /* forget this pointer (for sanity check) */
3165 batch->memcg = NULL;
3170 * called after __delete_from_swap_cache() and drop "page" account.
3171 * memcg information is recorded to swap_cgroup of "ent"
3174 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
3176 struct mem_cgroup *memcg;
3177 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
3179 if (!swapout) /* this was a swap cache but the swap is unused ! */
3180 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
3182 memcg = __mem_cgroup_uncharge_common(page, ctype);
3185 * record memcg information, if swapout && memcg != NULL,
3186 * mem_cgroup_get() was called in uncharge().
3188 if (do_swap_account && swapout && memcg)
3189 swap_cgroup_record(ent, css_id(&memcg->css));
3193 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3195 * called from swap_entry_free(). remove record in swap_cgroup and
3196 * uncharge "memsw" account.
3198 void mem_cgroup_uncharge_swap(swp_entry_t ent)
3200 struct mem_cgroup *memcg;
3203 if (!do_swap_account)
3206 id = swap_cgroup_record(ent, 0);
3208 memcg = mem_cgroup_lookup(id);
3211 * We uncharge this because swap is freed.
3212 * This memcg can be obsolete one. We avoid calling css_tryget
3214 if (!mem_cgroup_is_root(memcg))
3215 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
3216 mem_cgroup_swap_statistics(memcg, false);
3217 mem_cgroup_put(memcg);
3223 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3224 * @entry: swap entry to be moved
3225 * @from: mem_cgroup which the entry is moved from
3226 * @to: mem_cgroup which the entry is moved to
3227 * @need_fixup: whether we should fixup res_counters and refcounts.
3229 * It succeeds only when the swap_cgroup's record for this entry is the same
3230 * as the mem_cgroup's id of @from.
3232 * Returns 0 on success, -EINVAL on failure.
3234 * The caller must have charged to @to, IOW, called res_counter_charge() about
3235 * both res and memsw, and called css_get().
3237 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3238 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3240 unsigned short old_id, new_id;
3242 old_id = css_id(&from->css);
3243 new_id = css_id(&to->css);
3245 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3246 mem_cgroup_swap_statistics(from, false);
3247 mem_cgroup_swap_statistics(to, true);
3249 * This function is only called from task migration context now.
3250 * It postpones res_counter and refcount handling till the end
3251 * of task migration(mem_cgroup_clear_mc()) for performance
3252 * improvement. But we cannot postpone mem_cgroup_get(to)
3253 * because if the process that has been moved to @to does
3254 * swap-in, the refcount of @to might be decreased to 0.
3258 if (!mem_cgroup_is_root(from))
3259 res_counter_uncharge(&from->memsw, PAGE_SIZE);
3260 mem_cgroup_put(from);
3262 * we charged both to->res and to->memsw, so we should
3265 if (!mem_cgroup_is_root(to))
3266 res_counter_uncharge(&to->res, PAGE_SIZE);
3273 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3274 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3281 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
3284 int mem_cgroup_prepare_migration(struct page *page,
3285 struct page *newpage, struct mem_cgroup **ptr, gfp_t gfp_mask)
3287 struct mem_cgroup *memcg = NULL;
3288 struct page_cgroup *pc;
3289 enum charge_type ctype;
3294 VM_BUG_ON(PageTransHuge(page));
3295 if (mem_cgroup_disabled())
3298 pc = lookup_page_cgroup(page);
3299 lock_page_cgroup(pc);
3300 if (PageCgroupUsed(pc)) {
3301 memcg = pc->mem_cgroup;
3302 css_get(&memcg->css);
3304 * At migrating an anonymous page, its mapcount goes down
3305 * to 0 and uncharge() will be called. But, even if it's fully
3306 * unmapped, migration may fail and this page has to be
3307 * charged again. We set MIGRATION flag here and delay uncharge
3308 * until end_migration() is called
3310 * Corner Case Thinking
3312 * When the old page was mapped as Anon and it's unmap-and-freed
3313 * while migration was ongoing.
3314 * If unmap finds the old page, uncharge() of it will be delayed
3315 * until end_migration(). If unmap finds a new page, it's
3316 * uncharged when it make mapcount to be 1->0. If unmap code
3317 * finds swap_migration_entry, the new page will not be mapped
3318 * and end_migration() will find it(mapcount==0).
3321 * When the old page was mapped but migraion fails, the kernel
3322 * remaps it. A charge for it is kept by MIGRATION flag even
3323 * if mapcount goes down to 0. We can do remap successfully
3324 * without charging it again.
3327 * The "old" page is under lock_page() until the end of
3328 * migration, so, the old page itself will not be swapped-out.
3329 * If the new page is swapped out before end_migraton, our
3330 * hook to usual swap-out path will catch the event.
3333 SetPageCgroupMigration(pc);
3335 unlock_page_cgroup(pc);
3337 * If the page is not charged at this point,
3344 ret = __mem_cgroup_try_charge(NULL, gfp_mask, 1, ptr, false);
3345 css_put(&memcg->css);/* drop extra refcnt */
3346 if (ret || *ptr == NULL) {
3347 if (PageAnon(page)) {
3348 lock_page_cgroup(pc);
3349 ClearPageCgroupMigration(pc);
3350 unlock_page_cgroup(pc);
3352 * The old page may be fully unmapped while we kept it.
3354 mem_cgroup_uncharge_page(page);
3359 * We charge new page before it's used/mapped. So, even if unlock_page()
3360 * is called before end_migration, we can catch all events on this new
3361 * page. In the case new page is migrated but not remapped, new page's
3362 * mapcount will be finally 0 and we call uncharge in end_migration().
3364 pc = lookup_page_cgroup(newpage);
3366 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
3367 else if (page_is_file_cache(page))
3368 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
3370 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3371 __mem_cgroup_commit_charge(memcg, page, 1, pc, ctype);
3375 /* remove redundant charge if migration failed*/
3376 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
3377 struct page *oldpage, struct page *newpage, bool migration_ok)
3379 struct page *used, *unused;
3380 struct page_cgroup *pc;
3384 /* blocks rmdir() */
3385 cgroup_exclude_rmdir(&memcg->css);
3386 if (!migration_ok) {
3394 * We disallowed uncharge of pages under migration because mapcount
3395 * of the page goes down to zero, temporarly.
3396 * Clear the flag and check the page should be charged.
3398 pc = lookup_page_cgroup(oldpage);
3399 lock_page_cgroup(pc);
3400 ClearPageCgroupMigration(pc);
3401 unlock_page_cgroup(pc);
3403 __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE);
3406 * If a page is a file cache, radix-tree replacement is very atomic
3407 * and we can skip this check. When it was an Anon page, its mapcount
3408 * goes down to 0. But because we added MIGRATION flage, it's not
3409 * uncharged yet. There are several case but page->mapcount check
3410 * and USED bit check in mem_cgroup_uncharge_page() will do enough
3411 * check. (see prepare_charge() also)
3414 mem_cgroup_uncharge_page(used);
3416 * At migration, we may charge account against cgroup which has no
3418 * So, rmdir()->pre_destroy() can be called while we do this charge.
3419 * In that case, we need to call pre_destroy() again. check it here.
3421 cgroup_release_and_wakeup_rmdir(&memcg->css);
3424 #ifdef CONFIG_DEBUG_VM
3425 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3427 struct page_cgroup *pc;
3429 pc = lookup_page_cgroup(page);
3430 if (likely(pc) && PageCgroupUsed(pc))
3435 bool mem_cgroup_bad_page_check(struct page *page)
3437 if (mem_cgroup_disabled())
3440 return lookup_page_cgroup_used(page) != NULL;
3443 void mem_cgroup_print_bad_page(struct page *page)
3445 struct page_cgroup *pc;
3447 pc = lookup_page_cgroup_used(page);
3452 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p",
3453 pc, pc->flags, pc->mem_cgroup);
3455 path = kmalloc(PATH_MAX, GFP_KERNEL);
3458 ret = cgroup_path(pc->mem_cgroup->css.cgroup,
3463 printk(KERN_CONT "(%s)\n",
3464 (ret < 0) ? "cannot get the path" : path);
3470 static DEFINE_MUTEX(set_limit_mutex);
3472 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3473 unsigned long long val)
3476 u64 memswlimit, memlimit;
3478 int children = mem_cgroup_count_children(memcg);
3479 u64 curusage, oldusage;
3483 * For keeping hierarchical_reclaim simple, how long we should retry
3484 * is depends on callers. We set our retry-count to be function
3485 * of # of children which we should visit in this loop.
3487 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3489 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3492 while (retry_count) {
3493 if (signal_pending(current)) {
3498 * Rather than hide all in some function, I do this in
3499 * open coded manner. You see what this really does.
3500 * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3502 mutex_lock(&set_limit_mutex);
3503 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3504 if (memswlimit < val) {
3506 mutex_unlock(&set_limit_mutex);
3510 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3514 ret = res_counter_set_limit(&memcg->res, val);
3516 if (memswlimit == val)
3517 memcg->memsw_is_minimum = true;
3519 memcg->memsw_is_minimum = false;
3521 mutex_unlock(&set_limit_mutex);
3526 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3527 MEM_CGROUP_RECLAIM_SHRINK,
3529 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3530 /* Usage is reduced ? */
3531 if (curusage >= oldusage)
3534 oldusage = curusage;
3536 if (!ret && enlarge)
3537 memcg_oom_recover(memcg);
3542 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3543 unsigned long long val)
3546 u64 memlimit, memswlimit, oldusage, curusage;
3547 int children = mem_cgroup_count_children(memcg);
3551 /* see mem_cgroup_resize_res_limit */
3552 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3553 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3554 while (retry_count) {
3555 if (signal_pending(current)) {
3560 * Rather than hide all in some function, I do this in
3561 * open coded manner. You see what this really does.
3562 * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3564 mutex_lock(&set_limit_mutex);
3565 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3566 if (memlimit > val) {
3568 mutex_unlock(&set_limit_mutex);
3571 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3572 if (memswlimit < val)
3574 ret = res_counter_set_limit(&memcg->memsw, val);
3576 if (memlimit == val)
3577 memcg->memsw_is_minimum = true;
3579 memcg->memsw_is_minimum = false;
3581 mutex_unlock(&set_limit_mutex);
3586 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3587 MEM_CGROUP_RECLAIM_NOSWAP |
3588 MEM_CGROUP_RECLAIM_SHRINK,
3590 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3591 /* Usage is reduced ? */
3592 if (curusage >= oldusage)
3595 oldusage = curusage;
3597 if (!ret && enlarge)
3598 memcg_oom_recover(memcg);
3602 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3604 unsigned long *total_scanned)
3606 unsigned long nr_reclaimed = 0;
3607 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3608 unsigned long reclaimed;
3610 struct mem_cgroup_tree_per_zone *mctz;
3611 unsigned long long excess;
3612 unsigned long nr_scanned;
3617 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3619 * This loop can run a while, specially if mem_cgroup's continuously
3620 * keep exceeding their soft limit and putting the system under
3627 mz = mem_cgroup_largest_soft_limit_node(mctz);
3632 reclaimed = mem_cgroup_hierarchical_reclaim(mz->mem, zone,
3634 MEM_CGROUP_RECLAIM_SOFT,
3636 nr_reclaimed += reclaimed;
3637 *total_scanned += nr_scanned;
3638 spin_lock(&mctz->lock);
3641 * If we failed to reclaim anything from this memory cgroup
3642 * it is time to move on to the next cgroup
3648 * Loop until we find yet another one.
3650 * By the time we get the soft_limit lock
3651 * again, someone might have aded the
3652 * group back on the RB tree. Iterate to
3653 * make sure we get a different mem.
3654 * mem_cgroup_largest_soft_limit_node returns
3655 * NULL if no other cgroup is present on
3659 __mem_cgroup_largest_soft_limit_node(mctz);
3661 css_put(&next_mz->mem->css);
3662 else /* next_mz == NULL or other memcg */
3666 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
3667 excess = res_counter_soft_limit_excess(&mz->mem->res);
3669 * One school of thought says that we should not add
3670 * back the node to the tree if reclaim returns 0.
3671 * But our reclaim could return 0, simply because due
3672 * to priority we are exposing a smaller subset of
3673 * memory to reclaim from. Consider this as a longer
3676 /* If excess == 0, no tree ops */
3677 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess);
3678 spin_unlock(&mctz->lock);
3679 css_put(&mz->mem->css);
3682 * Could not reclaim anything and there are no more
3683 * mem cgroups to try or we seem to be looping without
3684 * reclaiming anything.
3686 if (!nr_reclaimed &&
3688 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3690 } while (!nr_reclaimed);
3692 css_put(&next_mz->mem->css);
3693 return nr_reclaimed;
3697 * This routine traverse page_cgroup in given list and drop them all.
3698 * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3700 static int mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
3701 int node, int zid, enum lru_list lru)
3704 struct mem_cgroup_per_zone *mz;
3705 struct page_cgroup *pc, *busy;
3706 unsigned long flags, loop;
3707 struct list_head *list;
3710 zone = &NODE_DATA(node)->node_zones[zid];
3711 mz = mem_cgroup_zoneinfo(memcg, node, zid);
3712 list = &mz->lists[lru];
3714 loop = MEM_CGROUP_ZSTAT(mz, lru);
3715 /* give some margin against EBUSY etc...*/
3722 spin_lock_irqsave(&zone->lru_lock, flags);
3723 if (list_empty(list)) {
3724 spin_unlock_irqrestore(&zone->lru_lock, flags);
3727 pc = list_entry(list->prev, struct page_cgroup, lru);
3729 list_move(&pc->lru, list);
3731 spin_unlock_irqrestore(&zone->lru_lock, flags);
3734 spin_unlock_irqrestore(&zone->lru_lock, flags);
3736 page = lookup_cgroup_page(pc);
3738 ret = mem_cgroup_move_parent(page, pc, memcg, GFP_KERNEL);
3742 if (ret == -EBUSY || ret == -EINVAL) {
3743 /* found lock contention or "pc" is obsolete. */
3750 if (!ret && !list_empty(list))
3756 * make mem_cgroup's charge to be 0 if there is no task.
3757 * This enables deleting this mem_cgroup.
3759 static int mem_cgroup_force_empty(struct mem_cgroup *memcg, bool free_all)
3762 int node, zid, shrink;
3763 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3764 struct cgroup *cgrp = memcg->css.cgroup;
3766 css_get(&memcg->css);
3769 /* should free all ? */
3775 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3778 if (signal_pending(current))
3780 /* This is for making all *used* pages to be on LRU. */
3781 lru_add_drain_all();
3782 drain_all_stock_sync(memcg);
3784 mem_cgroup_start_move(memcg);
3785 for_each_node_state(node, N_HIGH_MEMORY) {
3786 for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3789 ret = mem_cgroup_force_empty_list(memcg,
3798 mem_cgroup_end_move(memcg);
3799 memcg_oom_recover(memcg);
3800 /* it seems parent cgroup doesn't have enough mem */
3804 /* "ret" should also be checked to ensure all lists are empty. */
3805 } while (memcg->res.usage > 0 || ret);
3807 css_put(&memcg->css);
3811 /* returns EBUSY if there is a task or if we come here twice. */
3812 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3816 /* we call try-to-free pages for make this cgroup empty */
3817 lru_add_drain_all();
3818 /* try to free all pages in this cgroup */
3820 while (nr_retries && memcg->res.usage > 0) {
3823 if (signal_pending(current)) {
3827 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
3831 /* maybe some writeback is necessary */
3832 congestion_wait(BLK_RW_ASYNC, HZ/10);
3837 /* try move_account...there may be some *locked* pages. */
3841 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3843 return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3847 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3849 return mem_cgroup_from_cont(cont)->use_hierarchy;
3852 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3856 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3857 struct cgroup *parent = cont->parent;
3858 struct mem_cgroup *parent_memcg = NULL;
3861 parent_memcg = mem_cgroup_from_cont(parent);
3865 * If parent's use_hierarchy is set, we can't make any modifications
3866 * in the child subtrees. If it is unset, then the change can
3867 * occur, provided the current cgroup has no children.
3869 * For the root cgroup, parent_mem is NULL, we allow value to be
3870 * set if there are no children.
3872 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3873 (val == 1 || val == 0)) {
3874 if (list_empty(&cont->children))
3875 memcg->use_hierarchy = val;
3886 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
3887 enum mem_cgroup_stat_index idx)
3889 struct mem_cgroup *iter;
3892 /* Per-cpu values can be negative, use a signed accumulator */
3893 for_each_mem_cgroup_tree(iter, memcg)
3894 val += mem_cgroup_read_stat(iter, idx);
3896 if (val < 0) /* race ? */
3901 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3905 if (!mem_cgroup_is_root(memcg)) {
3907 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_KMEM
3908 if (!memcg->kmem_independent_accounting)
3909 val = res_counter_read_u64(&memcg->kmem, RES_USAGE);
3912 val += res_counter_read_u64(&memcg->res, RES_USAGE);
3914 val += res_counter_read_u64(&memcg->memsw, RES_USAGE);
3919 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
3920 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
3923 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
3925 return val << PAGE_SHIFT;
3928 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
3930 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3934 type = MEMFILE_TYPE(cft->private);
3935 name = MEMFILE_ATTR(cft->private);
3938 if (name == RES_USAGE)
3939 val = mem_cgroup_usage(memcg, false);
3941 val = res_counter_read_u64(&memcg->res, name);
3944 if (name == RES_USAGE)
3945 val = mem_cgroup_usage(memcg, true);
3947 val = res_counter_read_u64(&memcg->memsw, name);
3949 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_KMEM
3951 val = res_counter_read_u64(&memcg->kmem, name);
3961 * The user of this function is...
3964 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
3967 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3969 unsigned long long val;
3972 type = MEMFILE_TYPE(cft->private);
3973 name = MEMFILE_ATTR(cft->private);
3976 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3980 /* This function does all necessary parse...reuse it */
3981 ret = res_counter_memparse_write_strategy(buffer, &val);
3985 ret = mem_cgroup_resize_limit(memcg, val);
3987 ret = mem_cgroup_resize_memsw_limit(memcg, val);
3989 case RES_SOFT_LIMIT:
3990 ret = res_counter_memparse_write_strategy(buffer, &val);
3994 * For memsw, soft limits are hard to implement in terms
3995 * of semantics, for now, we support soft limits for
3996 * control without swap
3999 ret = res_counter_set_soft_limit(&memcg->res, val);
4004 ret = -EINVAL; /* should be BUG() ? */
4010 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
4011 unsigned long long *mem_limit, unsigned long long *memsw_limit)
4013 struct cgroup *cgroup;
4014 unsigned long long min_limit, min_memsw_limit, tmp;
4016 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4017 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4018 cgroup = memcg->css.cgroup;
4019 if (!memcg->use_hierarchy)
4022 while (cgroup->parent) {
4023 cgroup = cgroup->parent;
4024 memcg = mem_cgroup_from_cont(cgroup);
4025 if (!memcg->use_hierarchy)
4027 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
4028 min_limit = min(min_limit, tmp);
4029 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4030 min_memsw_limit = min(min_memsw_limit, tmp);
4033 *mem_limit = min_limit;
4034 *memsw_limit = min_memsw_limit;
4038 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
4040 struct mem_cgroup *memcg;
4043 memcg = mem_cgroup_from_cont(cont);
4044 type = MEMFILE_TYPE(event);
4045 name = MEMFILE_ATTR(event);
4049 res_counter_reset_max(&memcg->res);
4051 res_counter_reset_max(&memcg->memsw);
4055 res_counter_reset_failcnt(&memcg->res);
4057 res_counter_reset_failcnt(&memcg->memsw);
4064 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
4067 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
4071 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4072 struct cftype *cft, u64 val)
4074 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4076 if (val >= (1 << NR_MOVE_TYPE))
4079 * We check this value several times in both in can_attach() and
4080 * attach(), so we need cgroup lock to prevent this value from being
4084 memcg->move_charge_at_immigrate = val;
4090 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4091 struct cftype *cft, u64 val)
4098 /* For read statistics */
4116 struct mcs_total_stat {
4117 s64 stat[NR_MCS_STAT];
4123 } memcg_stat_strings[NR_MCS_STAT] = {
4124 {"cache", "total_cache"},
4125 {"rss", "total_rss"},
4126 {"mapped_file", "total_mapped_file"},
4127 {"pgpgin", "total_pgpgin"},
4128 {"pgpgout", "total_pgpgout"},
4129 {"swap", "total_swap"},
4130 {"pgfault", "total_pgfault"},
4131 {"pgmajfault", "total_pgmajfault"},
4132 {"inactive_anon", "total_inactive_anon"},
4133 {"active_anon", "total_active_anon"},
4134 {"inactive_file", "total_inactive_file"},
4135 {"active_file", "total_active_file"},
4136 {"unevictable", "total_unevictable"}
4141 mem_cgroup_get_local_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4146 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_CACHE);
4147 s->stat[MCS_CACHE] += val * PAGE_SIZE;
4148 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_RSS);
4149 s->stat[MCS_RSS] += val * PAGE_SIZE;
4150 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_FILE_MAPPED);
4151 s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
4152 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGIN);
4153 s->stat[MCS_PGPGIN] += val;
4154 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGOUT);
4155 s->stat[MCS_PGPGOUT] += val;
4156 if (do_swap_account) {
4157 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
4158 s->stat[MCS_SWAP] += val * PAGE_SIZE;
4160 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGFAULT);
4161 s->stat[MCS_PGFAULT] += val;
4162 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGMAJFAULT);
4163 s->stat[MCS_PGMAJFAULT] += val;
4166 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_ANON));
4167 s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
4168 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_ANON));
4169 s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
4170 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_FILE));
4171 s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
4172 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_FILE));
4173 s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
4174 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
4175 s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
4179 mem_cgroup_get_total_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4181 struct mem_cgroup *iter;
4183 for_each_mem_cgroup_tree(iter, memcg)
4184 mem_cgroup_get_local_stat(iter, s);
4188 static int mem_control_numa_stat_show(struct seq_file *m, void *arg)
4191 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
4192 unsigned long node_nr;
4193 struct cgroup *cont = m->private;
4194 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4196 total_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL);
4197 seq_printf(m, "total=%lu", total_nr);
4198 for_each_node_state(nid, N_HIGH_MEMORY) {
4199 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid, LRU_ALL);
4200 seq_printf(m, " N%d=%lu", nid, node_nr);
4204 file_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_FILE);
4205 seq_printf(m, "file=%lu", file_nr);
4206 for_each_node_state(nid, N_HIGH_MEMORY) {
4207 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4209 seq_printf(m, " N%d=%lu", nid, node_nr);
4213 anon_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_ANON);
4214 seq_printf(m, "anon=%lu", anon_nr);
4215 for_each_node_state(nid, N_HIGH_MEMORY) {
4216 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4218 seq_printf(m, " N%d=%lu", nid, node_nr);
4222 unevictable_nr = mem_cgroup_nr_lru_pages(mem_cont, BIT(LRU_UNEVICTABLE));
4223 seq_printf(m, "unevictable=%lu", unevictable_nr);
4224 for_each_node_state(nid, N_HIGH_MEMORY) {
4225 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4226 BIT(LRU_UNEVICTABLE));
4227 seq_printf(m, " N%d=%lu", nid, node_nr);
4232 #endif /* CONFIG_NUMA */
4234 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
4235 struct cgroup_map_cb *cb)
4237 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4238 struct mcs_total_stat mystat;
4241 memset(&mystat, 0, sizeof(mystat));
4242 mem_cgroup_get_local_stat(mem_cont, &mystat);
4245 for (i = 0; i < NR_MCS_STAT; i++) {
4246 if (i == MCS_SWAP && !do_swap_account)
4248 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
4251 /* Hierarchical information */
4253 unsigned long long limit, memsw_limit;
4254 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
4255 cb->fill(cb, "hierarchical_memory_limit", limit);
4256 if (do_swap_account)
4257 cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
4260 memset(&mystat, 0, sizeof(mystat));
4261 mem_cgroup_get_total_stat(mem_cont, &mystat);
4262 for (i = 0; i < NR_MCS_STAT; i++) {
4263 if (i == MCS_SWAP && !do_swap_account)
4265 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
4268 #ifdef CONFIG_DEBUG_VM
4271 struct mem_cgroup_per_zone *mz;
4272 unsigned long recent_rotated[2] = {0, 0};
4273 unsigned long recent_scanned[2] = {0, 0};
4275 for_each_online_node(nid)
4276 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4277 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
4279 recent_rotated[0] +=
4280 mz->reclaim_stat.recent_rotated[0];
4281 recent_rotated[1] +=
4282 mz->reclaim_stat.recent_rotated[1];
4283 recent_scanned[0] +=
4284 mz->reclaim_stat.recent_scanned[0];
4285 recent_scanned[1] +=
4286 mz->reclaim_stat.recent_scanned[1];
4288 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
4289 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
4290 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
4291 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
4298 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
4300 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4302 return mem_cgroup_swappiness(memcg);
4305 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
4308 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4309 struct mem_cgroup *parent;
4314 if (cgrp->parent == NULL)
4317 parent = mem_cgroup_from_cont(cgrp->parent);
4321 /* If under hierarchy, only empty-root can set this value */
4322 if ((parent->use_hierarchy) ||
4323 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4328 memcg->swappiness = val;
4335 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4337 struct mem_cgroup_threshold_ary *t;
4343 t = rcu_dereference(memcg->thresholds.primary);
4345 t = rcu_dereference(memcg->memsw_thresholds.primary);
4350 usage = mem_cgroup_usage(memcg, swap);
4353 * current_threshold points to threshold just below usage.
4354 * If it's not true, a threshold was crossed after last
4355 * call of __mem_cgroup_threshold().
4357 i = t->current_threshold;
4360 * Iterate backward over array of thresholds starting from
4361 * current_threshold and check if a threshold is crossed.
4362 * If none of thresholds below usage is crossed, we read
4363 * only one element of the array here.
4365 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4366 eventfd_signal(t->entries[i].eventfd, 1);
4368 /* i = current_threshold + 1 */
4372 * Iterate forward over array of thresholds starting from
4373 * current_threshold+1 and check if a threshold is crossed.
4374 * If none of thresholds above usage is crossed, we read
4375 * only one element of the array here.
4377 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4378 eventfd_signal(t->entries[i].eventfd, 1);
4380 /* Update current_threshold */
4381 t->current_threshold = i - 1;
4386 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4389 __mem_cgroup_threshold(memcg, false);
4390 if (do_swap_account)
4391 __mem_cgroup_threshold(memcg, true);
4393 memcg = parent_mem_cgroup(memcg);
4397 static int compare_thresholds(const void *a, const void *b)
4399 const struct mem_cgroup_threshold *_a = a;
4400 const struct mem_cgroup_threshold *_b = b;
4402 return _a->threshold - _b->threshold;
4405 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4407 struct mem_cgroup_eventfd_list *ev;
4409 list_for_each_entry(ev, &memcg->oom_notify, list)
4410 eventfd_signal(ev->eventfd, 1);
4414 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4416 struct mem_cgroup *iter;
4418 for_each_mem_cgroup_tree(iter, memcg)
4419 mem_cgroup_oom_notify_cb(iter);
4422 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
4423 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4425 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4426 struct mem_cgroup_thresholds *thresholds;
4427 struct mem_cgroup_threshold_ary *new;
4428 int type = MEMFILE_TYPE(cft->private);
4429 u64 threshold, usage;
4432 ret = res_counter_memparse_write_strategy(args, &threshold);
4436 mutex_lock(&memcg->thresholds_lock);
4439 thresholds = &memcg->thresholds;
4440 else if (type == _MEMSWAP)
4441 thresholds = &memcg->memsw_thresholds;
4445 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4447 /* Check if a threshold crossed before adding a new one */
4448 if (thresholds->primary)
4449 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4451 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4453 /* Allocate memory for new array of thresholds */
4454 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4462 /* Copy thresholds (if any) to new array */
4463 if (thresholds->primary) {
4464 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4465 sizeof(struct mem_cgroup_threshold));
4468 /* Add new threshold */
4469 new->entries[size - 1].eventfd = eventfd;
4470 new->entries[size - 1].threshold = threshold;
4472 /* Sort thresholds. Registering of new threshold isn't time-critical */
4473 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4474 compare_thresholds, NULL);
4476 /* Find current threshold */
4477 new->current_threshold = -1;
4478 for (i = 0; i < size; i++) {
4479 if (new->entries[i].threshold < usage) {
4481 * new->current_threshold will not be used until
4482 * rcu_assign_pointer(), so it's safe to increment
4485 ++new->current_threshold;
4489 /* Free old spare buffer and save old primary buffer as spare */
4490 kfree(thresholds->spare);
4491 thresholds->spare = thresholds->primary;
4493 rcu_assign_pointer(thresholds->primary, new);
4495 /* To be sure that nobody uses thresholds */
4499 mutex_unlock(&memcg->thresholds_lock);
4504 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
4505 struct cftype *cft, struct eventfd_ctx *eventfd)
4507 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4508 struct mem_cgroup_thresholds *thresholds;
4509 struct mem_cgroup_threshold_ary *new;
4510 int type = MEMFILE_TYPE(cft->private);
4514 mutex_lock(&memcg->thresholds_lock);
4516 thresholds = &memcg->thresholds;
4517 else if (type == _MEMSWAP)
4518 thresholds = &memcg->memsw_thresholds;
4523 * Something went wrong if we trying to unregister a threshold
4524 * if we don't have thresholds
4526 BUG_ON(!thresholds);
4528 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4530 /* Check if a threshold crossed before removing */
4531 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4533 /* Calculate new number of threshold */
4535 for (i = 0; i < thresholds->primary->size; i++) {
4536 if (thresholds->primary->entries[i].eventfd != eventfd)
4540 new = thresholds->spare;
4542 /* Set thresholds array to NULL if we don't have thresholds */
4551 /* Copy thresholds and find current threshold */
4552 new->current_threshold = -1;
4553 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4554 if (thresholds->primary->entries[i].eventfd == eventfd)
4557 new->entries[j] = thresholds->primary->entries[i];
4558 if (new->entries[j].threshold < usage) {
4560 * new->current_threshold will not be used
4561 * until rcu_assign_pointer(), so it's safe to increment
4564 ++new->current_threshold;
4570 /* Swap primary and spare array */
4571 thresholds->spare = thresholds->primary;
4572 rcu_assign_pointer(thresholds->primary, new);
4574 /* To be sure that nobody uses thresholds */
4577 mutex_unlock(&memcg->thresholds_lock);
4580 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4581 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4583 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4584 struct mem_cgroup_eventfd_list *event;
4585 int type = MEMFILE_TYPE(cft->private);
4587 BUG_ON(type != _OOM_TYPE);
4588 event = kmalloc(sizeof(*event), GFP_KERNEL);
4592 spin_lock(&memcg_oom_lock);
4594 event->eventfd = eventfd;
4595 list_add(&event->list, &memcg->oom_notify);
4597 /* already in OOM ? */
4598 if (atomic_read(&memcg->under_oom))
4599 eventfd_signal(eventfd, 1);
4600 spin_unlock(&memcg_oom_lock);
4605 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4606 struct cftype *cft, struct eventfd_ctx *eventfd)
4608 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4609 struct mem_cgroup_eventfd_list *ev, *tmp;
4610 int type = MEMFILE_TYPE(cft->private);
4612 BUG_ON(type != _OOM_TYPE);
4614 spin_lock(&memcg_oom_lock);
4616 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4617 if (ev->eventfd == eventfd) {
4618 list_del(&ev->list);
4623 spin_unlock(&memcg_oom_lock);
4626 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4627 struct cftype *cft, struct cgroup_map_cb *cb)
4629 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4631 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
4633 if (atomic_read(&memcg->under_oom))
4634 cb->fill(cb, "under_oom", 1);
4636 cb->fill(cb, "under_oom", 0);
4640 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4641 struct cftype *cft, u64 val)
4643 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4644 struct mem_cgroup *parent;
4646 /* cannot set to root cgroup and only 0 and 1 are allowed */
4647 if (!cgrp->parent || !((val == 0) || (val == 1)))
4650 parent = mem_cgroup_from_cont(cgrp->parent);
4653 /* oom-kill-disable is a flag for subhierarchy. */
4654 if ((parent->use_hierarchy) ||
4655 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4659 memcg->oom_kill_disable = val;
4661 memcg_oom_recover(memcg);
4667 static const struct file_operations mem_control_numa_stat_file_operations = {
4669 .llseek = seq_lseek,
4670 .release = single_release,
4673 static int mem_control_numa_stat_open(struct inode *unused, struct file *file)
4675 struct cgroup *cont = file->f_dentry->d_parent->d_fsdata;
4677 file->f_op = &mem_control_numa_stat_file_operations;
4678 return single_open(file, mem_control_numa_stat_show, cont);
4680 #endif /* CONFIG_NUMA */
4682 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_KMEM
4683 static u64 kmem_limit_independent_read(struct cgroup *cgroup, struct cftype *cft)
4685 return mem_cgroup_from_cont(cgroup)->kmem_independent_accounting;
4688 static int kmem_limit_independent_write(struct cgroup *cgroup, struct cftype *cft,
4691 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup);
4692 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4697 * This follows the same hierarchy restrictions than
4698 * mem_cgroup_hierarchy_write()
4700 if (!parent || !parent->use_hierarchy) {
4701 if (list_empty(&cgroup->children))
4702 memcg->kmem_independent_accounting = val;
4711 static struct cftype kmem_cgroup_files[] = {
4713 .name = "independent_kmem_limit",
4714 .read_u64 = kmem_limit_independent_read,
4715 .write_u64 = kmem_limit_independent_write,
4718 .name = "kmem.usage_in_bytes",
4719 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4720 .read_u64 = mem_cgroup_read,
4723 .name = "kmem.limit_in_bytes",
4724 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4725 .read_u64 = mem_cgroup_read,
4729 static int register_kmem_files(struct cgroup *cont, struct cgroup_subsys *ss)
4733 ret = cgroup_add_files(cont, ss, kmem_cgroup_files,
4734 ARRAY_SIZE(kmem_cgroup_files));
4739 static int register_kmem_files(struct cgroup *cont, struct cgroup_subsys *ss)
4745 static struct cftype mem_cgroup_files[] = {
4747 .name = "usage_in_bytes",
4748 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4749 .read_u64 = mem_cgroup_read,
4750 .register_event = mem_cgroup_usage_register_event,
4751 .unregister_event = mem_cgroup_usage_unregister_event,
4754 .name = "max_usage_in_bytes",
4755 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4756 .trigger = mem_cgroup_reset,
4757 .read_u64 = mem_cgroup_read,
4760 .name = "limit_in_bytes",
4761 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4762 .write_string = mem_cgroup_write,
4763 .read_u64 = mem_cgroup_read,
4766 .name = "soft_limit_in_bytes",
4767 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4768 .write_string = mem_cgroup_write,
4769 .read_u64 = mem_cgroup_read,
4773 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4774 .trigger = mem_cgroup_reset,
4775 .read_u64 = mem_cgroup_read,
4779 .read_map = mem_control_stat_show,
4782 .name = "force_empty",
4783 .trigger = mem_cgroup_force_empty_write,
4786 .name = "use_hierarchy",
4787 .write_u64 = mem_cgroup_hierarchy_write,
4788 .read_u64 = mem_cgroup_hierarchy_read,
4791 .name = "swappiness",
4792 .read_u64 = mem_cgroup_swappiness_read,
4793 .write_u64 = mem_cgroup_swappiness_write,
4796 .name = "move_charge_at_immigrate",
4797 .read_u64 = mem_cgroup_move_charge_read,
4798 .write_u64 = mem_cgroup_move_charge_write,
4801 .name = "oom_control",
4802 .read_map = mem_cgroup_oom_control_read,
4803 .write_u64 = mem_cgroup_oom_control_write,
4804 .register_event = mem_cgroup_oom_register_event,
4805 .unregister_event = mem_cgroup_oom_unregister_event,
4806 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4810 .name = "numa_stat",
4811 .open = mem_control_numa_stat_open,
4817 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4818 static struct cftype memsw_cgroup_files[] = {
4820 .name = "memsw.usage_in_bytes",
4821 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4822 .read_u64 = mem_cgroup_read,
4823 .register_event = mem_cgroup_usage_register_event,
4824 .unregister_event = mem_cgroup_usage_unregister_event,
4827 .name = "memsw.max_usage_in_bytes",
4828 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4829 .trigger = mem_cgroup_reset,
4830 .read_u64 = mem_cgroup_read,
4833 .name = "memsw.limit_in_bytes",
4834 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4835 .write_string = mem_cgroup_write,
4836 .read_u64 = mem_cgroup_read,
4839 .name = "memsw.failcnt",
4840 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4841 .trigger = mem_cgroup_reset,
4842 .read_u64 = mem_cgroup_read,
4846 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4848 if (!do_swap_account)
4850 return cgroup_add_files(cont, ss, memsw_cgroup_files,
4851 ARRAY_SIZE(memsw_cgroup_files));
4854 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4860 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4862 struct mem_cgroup_per_node *pn;
4863 struct mem_cgroup_per_zone *mz;
4865 int zone, tmp = node;
4867 * This routine is called against possible nodes.
4868 * But it's BUG to call kmalloc() against offline node.
4870 * TODO: this routine can waste much memory for nodes which will
4871 * never be onlined. It's better to use memory hotplug callback
4874 if (!node_state(node, N_NORMAL_MEMORY))
4876 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4880 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4881 mz = &pn->zoneinfo[zone];
4883 INIT_LIST_HEAD(&mz->lists[l]);
4884 mz->usage_in_excess = 0;
4885 mz->on_tree = false;
4888 memcg->info.nodeinfo[node] = pn;
4892 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4894 kfree(memcg->info.nodeinfo[node]);
4897 static struct mem_cgroup *mem_cgroup_alloc(void)
4899 struct mem_cgroup *mem;
4900 int size = sizeof(struct mem_cgroup);
4902 /* Can be very big if MAX_NUMNODES is very big */
4903 if (size < PAGE_SIZE)
4904 mem = kzalloc(size, GFP_KERNEL);
4906 mem = vzalloc(size);
4911 mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4914 spin_lock_init(&mem->pcp_counter_lock);
4918 if (size < PAGE_SIZE)
4926 * At destroying mem_cgroup, references from swap_cgroup can remain.
4927 * (scanning all at force_empty is too costly...)
4929 * Instead of clearing all references at force_empty, we remember
4930 * the number of reference from swap_cgroup and free mem_cgroup when
4931 * it goes down to 0.
4933 * Removal of cgroup itself succeeds regardless of refs from swap.
4936 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4940 mem_cgroup_remove_from_trees(memcg);
4941 free_css_id(&mem_cgroup_subsys, &memcg->css);
4943 for_each_node_state(node, N_POSSIBLE)
4944 free_mem_cgroup_per_zone_info(memcg, node);
4946 free_percpu(memcg->stat);
4947 if (sizeof(struct mem_cgroup) < PAGE_SIZE)
4953 static void mem_cgroup_get(struct mem_cgroup *memcg)
4955 atomic_inc(&memcg->refcnt);
4958 static void __mem_cgroup_put(struct mem_cgroup *memcg, int count)
4960 if (atomic_sub_and_test(count, &memcg->refcnt)) {
4961 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4962 __mem_cgroup_free(memcg);
4964 mem_cgroup_put(parent);
4968 static void mem_cgroup_put(struct mem_cgroup *memcg)
4970 __mem_cgroup_put(memcg, 1);
4974 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4976 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4978 if (!memcg->res.parent)
4980 return mem_cgroup_from_res_counter(memcg->res.parent, res);
4982 EXPORT_SYMBOL(parent_mem_cgroup);
4984 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4985 static void __init enable_swap_cgroup(void)
4987 if (!mem_cgroup_disabled() && really_do_swap_account)
4988 do_swap_account = 1;
4991 static void __init enable_swap_cgroup(void)
4996 static int mem_cgroup_soft_limit_tree_init(void)
4998 struct mem_cgroup_tree_per_node *rtpn;
4999 struct mem_cgroup_tree_per_zone *rtpz;
5000 int tmp, node, zone;
5002 for_each_node_state(node, N_POSSIBLE) {
5004 if (!node_state(node, N_NORMAL_MEMORY))
5006 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
5010 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5012 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5013 rtpz = &rtpn->rb_tree_per_zone[zone];
5014 rtpz->rb_root = RB_ROOT;
5015 spin_lock_init(&rtpz->lock);
5021 static struct cgroup_subsys_state * __ref
5022 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
5024 struct mem_cgroup *memcg, *parent;
5025 long error = -ENOMEM;
5028 memcg = mem_cgroup_alloc();
5030 return ERR_PTR(error);
5032 for_each_node_state(node, N_POSSIBLE)
5033 if (alloc_mem_cgroup_per_zone_info(memcg, node))
5037 if (cont->parent == NULL) {
5039 enable_swap_cgroup();
5041 root_mem_cgroup = memcg;
5042 if (mem_cgroup_soft_limit_tree_init())
5044 for_each_possible_cpu(cpu) {
5045 struct memcg_stock_pcp *stock =
5046 &per_cpu(memcg_stock, cpu);
5047 INIT_WORK(&stock->work, drain_local_stock);
5049 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5051 parent = mem_cgroup_from_cont(cont->parent);
5052 memcg->use_hierarchy = parent->use_hierarchy;
5053 memcg->oom_kill_disable = parent->oom_kill_disable;
5056 if (parent && parent->use_hierarchy) {
5057 res_counter_init(&memcg->res, &parent->res);
5058 res_counter_init(&memcg->memsw, &parent->memsw);
5059 res_counter_init(&memcg->kmem, &parent->kmem);
5061 * We increment refcnt of the parent to ensure that we can
5062 * safely access it on res_counter_charge/uncharge.
5063 * This refcnt will be decremented when freeing this
5064 * mem_cgroup(see mem_cgroup_put).
5066 mem_cgroup_get(parent);
5068 res_counter_init(&memcg->res, NULL);
5069 res_counter_init(&memcg->memsw, NULL);
5070 res_counter_init(&memcg->kmem, NULL);
5072 memcg->last_scanned_child = 0;
5073 memcg->last_scanned_node = MAX_NUMNODES;
5074 INIT_LIST_HEAD(&memcg->oom_notify);
5077 memcg->swappiness = mem_cgroup_swappiness(parent);
5078 atomic_set(&memcg->refcnt, 1);
5079 memcg->move_charge_at_immigrate = 0;
5080 mutex_init(&memcg->thresholds_lock);
5083 __mem_cgroup_free(memcg);
5084 root_mem_cgroup = NULL;
5085 return ERR_PTR(error);
5088 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
5089 struct cgroup *cont)
5091 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5093 return mem_cgroup_force_empty(memcg, false);
5096 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
5097 struct cgroup *cont)
5099 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5101 mem_cgroup_put(memcg);
5104 static int mem_cgroup_populate(struct cgroup_subsys *ss,
5105 struct cgroup *cont)
5109 ret = cgroup_add_files(cont, ss, mem_cgroup_files,
5110 ARRAY_SIZE(mem_cgroup_files));
5113 ret = register_memsw_files(cont, ss);
5116 ret = register_kmem_files(cont, ss);
5122 /* Handlers for move charge at task migration. */
5123 #define PRECHARGE_COUNT_AT_ONCE 256
5124 static int mem_cgroup_do_precharge(unsigned long count)
5127 int batch_count = PRECHARGE_COUNT_AT_ONCE;
5128 struct mem_cgroup *memcg = mc.to;
5130 if (mem_cgroup_is_root(memcg)) {
5131 mc.precharge += count;
5132 /* we don't need css_get for root */
5135 /* try to charge at once */
5137 struct res_counter *dummy;
5139 * "memcg" cannot be under rmdir() because we've already checked
5140 * by cgroup_lock_live_cgroup() that it is not removed and we
5141 * are still under the same cgroup_mutex. So we can postpone
5144 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
5146 if (do_swap_account && res_counter_charge(&memcg->memsw,
5147 PAGE_SIZE * count, &dummy)) {
5148 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
5151 mc.precharge += count;
5155 /* fall back to one by one charge */
5157 if (signal_pending(current)) {
5161 if (!batch_count--) {
5162 batch_count = PRECHARGE_COUNT_AT_ONCE;
5165 ret = __mem_cgroup_try_charge(NULL,
5166 GFP_KERNEL, 1, &memcg, false);
5168 /* mem_cgroup_clear_mc() will do uncharge later */
5176 * is_target_pte_for_mc - check a pte whether it is valid for move charge
5177 * @vma: the vma the pte to be checked belongs
5178 * @addr: the address corresponding to the pte to be checked
5179 * @ptent: the pte to be checked
5180 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5183 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5184 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5185 * move charge. if @target is not NULL, the page is stored in target->page
5186 * with extra refcnt got(Callers should handle it).
5187 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5188 * target for charge migration. if @target is not NULL, the entry is stored
5191 * Called with pte lock held.
5198 enum mc_target_type {
5199 MC_TARGET_NONE, /* not used */
5204 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5205 unsigned long addr, pte_t ptent)
5207 struct page *page = vm_normal_page(vma, addr, ptent);
5209 if (!page || !page_mapped(page))
5211 if (PageAnon(page)) {
5212 /* we don't move shared anon */
5213 if (!move_anon() || page_mapcount(page) > 2)
5215 } else if (!move_file())
5216 /* we ignore mapcount for file pages */
5218 if (!get_page_unless_zero(page))
5224 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5225 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5228 struct page *page = NULL;
5229 swp_entry_t ent = pte_to_swp_entry(ptent);
5231 if (!move_anon() || non_swap_entry(ent))
5233 usage_count = mem_cgroup_count_swap_user(ent, &page);
5234 if (usage_count > 1) { /* we don't move shared anon */
5239 if (do_swap_account)
5240 entry->val = ent.val;
5245 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5246 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5248 struct page *page = NULL;
5249 struct inode *inode;
5250 struct address_space *mapping;
5253 if (!vma->vm_file) /* anonymous vma */
5258 inode = vma->vm_file->f_path.dentry->d_inode;
5259 mapping = vma->vm_file->f_mapping;
5260 if (pte_none(ptent))
5261 pgoff = linear_page_index(vma, addr);
5262 else /* pte_file(ptent) is true */
5263 pgoff = pte_to_pgoff(ptent);
5265 /* page is moved even if it's not RSS of this task(page-faulted). */
5266 page = find_get_page(mapping, pgoff);
5269 /* shmem/tmpfs may report page out on swap: account for that too. */
5270 if (radix_tree_exceptional_entry(page)) {
5271 swp_entry_t swap = radix_to_swp_entry(page);
5272 if (do_swap_account)
5274 page = find_get_page(&swapper_space, swap.val);
5280 static int is_target_pte_for_mc(struct vm_area_struct *vma,
5281 unsigned long addr, pte_t ptent, union mc_target *target)
5283 struct page *page = NULL;
5284 struct page_cgroup *pc;
5286 swp_entry_t ent = { .val = 0 };
5288 if (pte_present(ptent))
5289 page = mc_handle_present_pte(vma, addr, ptent);
5290 else if (is_swap_pte(ptent))
5291 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5292 else if (pte_none(ptent) || pte_file(ptent))
5293 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5295 if (!page && !ent.val)
5298 pc = lookup_page_cgroup(page);
5300 * Do only loose check w/o page_cgroup lock.
5301 * mem_cgroup_move_account() checks the pc is valid or not under
5304 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5305 ret = MC_TARGET_PAGE;
5307 target->page = page;
5309 if (!ret || !target)
5312 /* There is a swap entry and a page doesn't exist or isn't charged */
5313 if (ent.val && !ret &&
5314 css_id(&mc.from->css) == lookup_swap_cgroup(ent)) {
5315 ret = MC_TARGET_SWAP;
5322 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5323 unsigned long addr, unsigned long end,
5324 struct mm_walk *walk)
5326 struct vm_area_struct *vma = walk->private;
5330 split_huge_page_pmd(walk->mm, pmd);
5332 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5333 for (; addr != end; pte++, addr += PAGE_SIZE)
5334 if (is_target_pte_for_mc(vma, addr, *pte, NULL))
5335 mc.precharge++; /* increment precharge temporarily */
5336 pte_unmap_unlock(pte - 1, ptl);
5342 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5344 unsigned long precharge;
5345 struct vm_area_struct *vma;
5347 down_read(&mm->mmap_sem);
5348 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5349 struct mm_walk mem_cgroup_count_precharge_walk = {
5350 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5354 if (is_vm_hugetlb_page(vma))
5356 walk_page_range(vma->vm_start, vma->vm_end,
5357 &mem_cgroup_count_precharge_walk);
5359 up_read(&mm->mmap_sem);
5361 precharge = mc.precharge;
5367 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5369 unsigned long precharge = mem_cgroup_count_precharge(mm);
5371 VM_BUG_ON(mc.moving_task);
5372 mc.moving_task = current;
5373 return mem_cgroup_do_precharge(precharge);
5376 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5377 static void __mem_cgroup_clear_mc(void)
5379 struct mem_cgroup *from = mc.from;
5380 struct mem_cgroup *to = mc.to;
5382 /* we must uncharge all the leftover precharges from mc.to */
5384 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
5388 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5389 * we must uncharge here.
5391 if (mc.moved_charge) {
5392 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
5393 mc.moved_charge = 0;
5395 /* we must fixup refcnts and charges */
5396 if (mc.moved_swap) {
5397 /* uncharge swap account from the old cgroup */
5398 if (!mem_cgroup_is_root(mc.from))
5399 res_counter_uncharge(&mc.from->memsw,
5400 PAGE_SIZE * mc.moved_swap);
5401 __mem_cgroup_put(mc.from, mc.moved_swap);
5403 if (!mem_cgroup_is_root(mc.to)) {
5405 * we charged both to->res and to->memsw, so we should
5408 res_counter_uncharge(&mc.to->res,
5409 PAGE_SIZE * mc.moved_swap);
5411 /* we've already done mem_cgroup_get(mc.to) */
5414 memcg_oom_recover(from);
5415 memcg_oom_recover(to);
5416 wake_up_all(&mc.waitq);
5419 static void mem_cgroup_clear_mc(void)
5421 struct mem_cgroup *from = mc.from;
5424 * we must clear moving_task before waking up waiters at the end of
5427 mc.moving_task = NULL;
5428 __mem_cgroup_clear_mc();
5429 spin_lock(&mc.lock);
5432 spin_unlock(&mc.lock);
5433 mem_cgroup_end_move(from);
5436 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5437 struct cgroup *cgroup,
5438 struct task_struct *p)
5441 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup);
5443 if (memcg->move_charge_at_immigrate) {
5444 struct mm_struct *mm;
5445 struct mem_cgroup *from = mem_cgroup_from_task(p);
5447 VM_BUG_ON(from == memcg);
5449 mm = get_task_mm(p);
5452 /* We move charges only when we move a owner of the mm */
5453 if (mm->owner == p) {
5456 VM_BUG_ON(mc.precharge);
5457 VM_BUG_ON(mc.moved_charge);
5458 VM_BUG_ON(mc.moved_swap);
5459 mem_cgroup_start_move(from);
5460 spin_lock(&mc.lock);
5463 spin_unlock(&mc.lock);
5464 /* We set mc.moving_task later */
5466 ret = mem_cgroup_precharge_mc(mm);
5468 mem_cgroup_clear_mc();
5475 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5476 struct cgroup *cgroup,
5477 struct task_struct *p)
5479 mem_cgroup_clear_mc();
5482 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5483 unsigned long addr, unsigned long end,
5484 struct mm_walk *walk)
5487 struct vm_area_struct *vma = walk->private;
5491 split_huge_page_pmd(walk->mm, pmd);
5493 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5494 for (; addr != end; addr += PAGE_SIZE) {
5495 pte_t ptent = *(pte++);
5496 union mc_target target;
5499 struct page_cgroup *pc;
5505 type = is_target_pte_for_mc(vma, addr, ptent, &target);
5507 case MC_TARGET_PAGE:
5509 if (isolate_lru_page(page))
5511 pc = lookup_page_cgroup(page);
5512 if (!mem_cgroup_move_account(page, 1, pc,
5513 mc.from, mc.to, false)) {
5515 /* we uncharge from mc.from later. */
5518 putback_lru_page(page);
5519 put: /* is_target_pte_for_mc() gets the page */
5522 case MC_TARGET_SWAP:
5524 if (!mem_cgroup_move_swap_account(ent,
5525 mc.from, mc.to, false)) {
5527 /* we fixup refcnts and charges later. */
5535 pte_unmap_unlock(pte - 1, ptl);
5540 * We have consumed all precharges we got in can_attach().
5541 * We try charge one by one, but don't do any additional
5542 * charges to mc.to if we have failed in charge once in attach()
5545 ret = mem_cgroup_do_precharge(1);
5553 static void mem_cgroup_move_charge(struct mm_struct *mm)
5555 struct vm_area_struct *vma;
5557 lru_add_drain_all();
5559 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5561 * Someone who are holding the mmap_sem might be waiting in
5562 * waitq. So we cancel all extra charges, wake up all waiters,
5563 * and retry. Because we cancel precharges, we might not be able
5564 * to move enough charges, but moving charge is a best-effort
5565 * feature anyway, so it wouldn't be a big problem.
5567 __mem_cgroup_clear_mc();
5571 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5573 struct mm_walk mem_cgroup_move_charge_walk = {
5574 .pmd_entry = mem_cgroup_move_charge_pte_range,
5578 if (is_vm_hugetlb_page(vma))
5580 ret = walk_page_range(vma->vm_start, vma->vm_end,
5581 &mem_cgroup_move_charge_walk);
5584 * means we have consumed all precharges and failed in
5585 * doing additional charge. Just abandon here.
5589 up_read(&mm->mmap_sem);
5592 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5593 struct cgroup *cont,
5594 struct cgroup *old_cont,
5595 struct task_struct *p)
5597 struct mm_struct *mm = get_task_mm(p);
5601 mem_cgroup_move_charge(mm);
5606 mem_cgroup_clear_mc();
5608 #else /* !CONFIG_MMU */
5609 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5610 struct cgroup *cgroup,
5611 struct task_struct *p)
5615 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5616 struct cgroup *cgroup,
5617 struct task_struct *p)
5620 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5621 struct cgroup *cont,
5622 struct cgroup *old_cont,
5623 struct task_struct *p)
5628 struct cgroup_subsys mem_cgroup_subsys = {
5630 .subsys_id = mem_cgroup_subsys_id,
5631 .create = mem_cgroup_create,
5632 .pre_destroy = mem_cgroup_pre_destroy,
5633 .destroy = mem_cgroup_destroy,
5634 .populate = mem_cgroup_populate,
5635 .can_attach = mem_cgroup_can_attach,
5636 .cancel_attach = mem_cgroup_cancel_attach,
5637 .attach = mem_cgroup_move_task,
5642 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5643 static int __init enable_swap_account(char *s)
5645 /* consider enabled if no parameter or 1 is given */
5646 if (!strcmp(s, "1"))
5647 really_do_swap_account = 1;
5648 else if (!strcmp(s, "0"))
5649 really_do_swap_account = 0;
5652 __setup("swapaccount=", enable_swap_account);