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 * Kernel Memory Controller
14 * Copyright (C) 2012 Parallels Inc. and Google Inc.
15 * Authors: Glauber Costa and Suleiman Souhlal
17 * This program is free software; you can redistribute it and/or modify
18 * it under the terms of the GNU General Public License as published by
19 * the Free Software Foundation; either version 2 of the License, or
20 * (at your option) any later version.
22 * This program is distributed in the hope that it will be useful,
23 * but WITHOUT ANY WARRANTY; without even the implied warranty of
24 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
25 * GNU General Public License for more details.
28 #include <linux/res_counter.h>
29 #include <linux/memcontrol.h>
30 #include <linux/cgroup.h>
32 #include <linux/hugetlb.h>
33 #include <linux/pagemap.h>
34 #include <linux/smp.h>
35 #include <linux/page-flags.h>
36 #include <linux/backing-dev.h>
37 #include <linux/bit_spinlock.h>
38 #include <linux/rcupdate.h>
39 #include <linux/limits.h>
40 #include <linux/export.h>
41 #include <linux/mutex.h>
42 #include <linux/rbtree.h>
43 #include <linux/slab.h>
44 #include <linux/swap.h>
45 #include <linux/swapops.h>
46 #include <linux/spinlock.h>
47 #include <linux/eventfd.h>
48 #include <linux/sort.h>
50 #include <linux/seq_file.h>
51 #include <linux/vmalloc.h>
52 #include <linux/mm_inline.h>
53 #include <linux/page_cgroup.h>
54 #include <linux/cpu.h>
55 #include <linux/oom.h>
59 #include <net/tcp_memcontrol.h>
61 #include <asm/uaccess.h>
63 #include <trace/events/vmscan.h>
65 struct cgroup_subsys mem_cgroup_subsys __read_mostly;
66 EXPORT_SYMBOL(mem_cgroup_subsys);
68 #define MEM_CGROUP_RECLAIM_RETRIES 5
69 static struct mem_cgroup *root_mem_cgroup __read_mostly;
71 #ifdef CONFIG_MEMCG_SWAP
72 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
73 int do_swap_account __read_mostly;
75 /* for remember boot option*/
76 #ifdef CONFIG_MEMCG_SWAP_ENABLED
77 static int really_do_swap_account __initdata = 1;
79 static int really_do_swap_account __initdata = 0;
83 #define do_swap_account 0
88 * Statistics for memory cgroup.
90 enum mem_cgroup_stat_index {
92 * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
94 MEM_CGROUP_STAT_CACHE, /* # of pages charged as cache */
95 MEM_CGROUP_STAT_RSS, /* # of pages charged as anon rss */
96 MEM_CGROUP_STAT_FILE_MAPPED, /* # of pages charged as file rss */
97 MEM_CGROUP_STAT_SWAP, /* # of pages, swapped out */
98 MEM_CGROUP_STAT_NSTATS,
101 static const char * const mem_cgroup_stat_names[] = {
108 enum mem_cgroup_events_index {
109 MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */
110 MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */
111 MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */
112 MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */
113 MEM_CGROUP_EVENTS_NSTATS,
116 static const char * const mem_cgroup_events_names[] = {
124 * Per memcg event counter is incremented at every pagein/pageout. With THP,
125 * it will be incremated by the number of pages. This counter is used for
126 * for trigger some periodic events. This is straightforward and better
127 * than using jiffies etc. to handle periodic memcg event.
129 enum mem_cgroup_events_target {
130 MEM_CGROUP_TARGET_THRESH,
131 MEM_CGROUP_TARGET_SOFTLIMIT,
132 MEM_CGROUP_TARGET_NUMAINFO,
135 #define THRESHOLDS_EVENTS_TARGET 128
136 #define SOFTLIMIT_EVENTS_TARGET 1024
137 #define NUMAINFO_EVENTS_TARGET 1024
139 struct mem_cgroup_stat_cpu {
140 long count[MEM_CGROUP_STAT_NSTATS];
141 unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
142 unsigned long nr_page_events;
143 unsigned long targets[MEM_CGROUP_NTARGETS];
146 struct mem_cgroup_reclaim_iter {
147 /* css_id of the last scanned hierarchy member */
149 /* scan generation, increased every round-trip */
150 unsigned int generation;
154 * per-zone information in memory controller.
156 struct mem_cgroup_per_zone {
157 struct lruvec lruvec;
158 unsigned long lru_size[NR_LRU_LISTS];
160 struct mem_cgroup_reclaim_iter reclaim_iter[DEF_PRIORITY + 1];
162 struct rb_node tree_node; /* RB tree node */
163 unsigned long long usage_in_excess;/* Set to the value by which */
164 /* the soft limit is exceeded*/
166 struct mem_cgroup *memcg; /* Back pointer, we cannot */
167 /* use container_of */
170 struct mem_cgroup_per_node {
171 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
174 struct mem_cgroup_lru_info {
175 struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
179 * Cgroups above their limits are maintained in a RB-Tree, independent of
180 * their hierarchy representation
183 struct mem_cgroup_tree_per_zone {
184 struct rb_root rb_root;
188 struct mem_cgroup_tree_per_node {
189 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
192 struct mem_cgroup_tree {
193 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
196 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
198 struct mem_cgroup_threshold {
199 struct eventfd_ctx *eventfd;
204 struct mem_cgroup_threshold_ary {
205 /* An array index points to threshold just below or equal to usage. */
206 int current_threshold;
207 /* Size of entries[] */
209 /* Array of thresholds */
210 struct mem_cgroup_threshold entries[0];
213 struct mem_cgroup_thresholds {
214 /* Primary thresholds array */
215 struct mem_cgroup_threshold_ary *primary;
217 * Spare threshold array.
218 * This is needed to make mem_cgroup_unregister_event() "never fail".
219 * It must be able to store at least primary->size - 1 entries.
221 struct mem_cgroup_threshold_ary *spare;
225 struct mem_cgroup_eventfd_list {
226 struct list_head list;
227 struct eventfd_ctx *eventfd;
230 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
231 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
234 * The memory controller data structure. The memory controller controls both
235 * page cache and RSS per cgroup. We would eventually like to provide
236 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
237 * to help the administrator determine what knobs to tune.
239 * TODO: Add a water mark for the memory controller. Reclaim will begin when
240 * we hit the water mark. May be even add a low water mark, such that
241 * no reclaim occurs from a cgroup at it's low water mark, this is
242 * a feature that will be implemented much later in the future.
245 struct cgroup_subsys_state css;
247 * the counter to account for memory usage
249 struct res_counter res;
253 * the counter to account for mem+swap usage.
255 struct res_counter memsw;
258 * rcu_freeing is used only when freeing struct mem_cgroup,
259 * so put it into a union to avoid wasting more memory.
260 * It must be disjoint from the css field. It could be
261 * in a union with the res field, but res plays a much
262 * larger part in mem_cgroup life than memsw, and might
263 * be of interest, even at time of free, when debugging.
264 * So share rcu_head with the less interesting memsw.
266 struct rcu_head rcu_freeing;
268 * We also need some space for a worker in deferred freeing.
269 * By the time we call it, rcu_freeing is no longer in use.
271 struct work_struct work_freeing;
275 * the counter to account for kernel memory usage.
277 struct res_counter kmem;
279 * Per cgroup active and inactive list, similar to the
280 * per zone LRU lists.
282 struct mem_cgroup_lru_info info;
283 int last_scanned_node;
285 nodemask_t scan_nodes;
286 atomic_t numainfo_events;
287 atomic_t numainfo_updating;
290 * Should the accounting and control be hierarchical, per subtree?
293 unsigned long kmem_account_flags; /* See KMEM_ACCOUNTED_*, below */
301 /* OOM-Killer disable */
302 int oom_kill_disable;
304 /* set when res.limit == memsw.limit */
305 bool memsw_is_minimum;
307 /* protect arrays of thresholds */
308 struct mutex thresholds_lock;
310 /* thresholds for memory usage. RCU-protected */
311 struct mem_cgroup_thresholds thresholds;
313 /* thresholds for mem+swap usage. RCU-protected */
314 struct mem_cgroup_thresholds memsw_thresholds;
316 /* For oom notifier event fd */
317 struct list_head oom_notify;
320 * Should we move charges of a task when a task is moved into this
321 * mem_cgroup ? And what type of charges should we move ?
323 unsigned long move_charge_at_immigrate;
325 * set > 0 if pages under this cgroup are moving to other cgroup.
327 atomic_t moving_account;
328 /* taken only while moving_account > 0 */
329 spinlock_t move_lock;
333 struct mem_cgroup_stat_cpu __percpu *stat;
335 * used when a cpu is offlined or other synchronizations
336 * See mem_cgroup_read_stat().
338 struct mem_cgroup_stat_cpu nocpu_base;
339 spinlock_t pcp_counter_lock;
341 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
342 struct tcp_memcontrol tcp_mem;
346 /* internal only representation about the status of kmem accounting. */
348 KMEM_ACCOUNTED_ACTIVE = 0, /* accounted by this cgroup itself */
351 #define KMEM_ACCOUNTED_MASK (1 << KMEM_ACCOUNTED_ACTIVE)
353 #ifdef CONFIG_MEMCG_KMEM
354 static inline void memcg_kmem_set_active(struct mem_cgroup *memcg)
356 set_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
360 /* Stuffs for move charges at task migration. */
362 * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
363 * left-shifted bitmap of these types.
366 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
367 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
371 /* "mc" and its members are protected by cgroup_mutex */
372 static struct move_charge_struct {
373 spinlock_t lock; /* for from, to */
374 struct mem_cgroup *from;
375 struct mem_cgroup *to;
376 unsigned long precharge;
377 unsigned long moved_charge;
378 unsigned long moved_swap;
379 struct task_struct *moving_task; /* a task moving charges */
380 wait_queue_head_t waitq; /* a waitq for other context */
382 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
383 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
386 static bool move_anon(void)
388 return test_bit(MOVE_CHARGE_TYPE_ANON,
389 &mc.to->move_charge_at_immigrate);
392 static bool move_file(void)
394 return test_bit(MOVE_CHARGE_TYPE_FILE,
395 &mc.to->move_charge_at_immigrate);
399 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
400 * limit reclaim to prevent infinite loops, if they ever occur.
402 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
403 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
406 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
407 MEM_CGROUP_CHARGE_TYPE_ANON,
408 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
409 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
413 /* for encoding cft->private value on file */
421 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
422 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
423 #define MEMFILE_ATTR(val) ((val) & 0xffff)
424 /* Used for OOM nofiier */
425 #define OOM_CONTROL (0)
428 * Reclaim flags for mem_cgroup_hierarchical_reclaim
430 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
431 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
432 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
433 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
435 static void mem_cgroup_get(struct mem_cgroup *memcg);
436 static void mem_cgroup_put(struct mem_cgroup *memcg);
439 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
441 return container_of(s, struct mem_cgroup, css);
444 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
446 return (memcg == root_mem_cgroup);
449 /* Writing them here to avoid exposing memcg's inner layout */
450 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
452 void sock_update_memcg(struct sock *sk)
454 if (mem_cgroup_sockets_enabled) {
455 struct mem_cgroup *memcg;
456 struct cg_proto *cg_proto;
458 BUG_ON(!sk->sk_prot->proto_cgroup);
460 /* Socket cloning can throw us here with sk_cgrp already
461 * filled. It won't however, necessarily happen from
462 * process context. So the test for root memcg given
463 * the current task's memcg won't help us in this case.
465 * Respecting the original socket's memcg is a better
466 * decision in this case.
469 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
470 mem_cgroup_get(sk->sk_cgrp->memcg);
475 memcg = mem_cgroup_from_task(current);
476 cg_proto = sk->sk_prot->proto_cgroup(memcg);
477 if (!mem_cgroup_is_root(memcg) && memcg_proto_active(cg_proto)) {
478 mem_cgroup_get(memcg);
479 sk->sk_cgrp = cg_proto;
484 EXPORT_SYMBOL(sock_update_memcg);
486 void sock_release_memcg(struct sock *sk)
488 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
489 struct mem_cgroup *memcg;
490 WARN_ON(!sk->sk_cgrp->memcg);
491 memcg = sk->sk_cgrp->memcg;
492 mem_cgroup_put(memcg);
496 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
498 if (!memcg || mem_cgroup_is_root(memcg))
501 return &memcg->tcp_mem.cg_proto;
503 EXPORT_SYMBOL(tcp_proto_cgroup);
505 static void disarm_sock_keys(struct mem_cgroup *memcg)
507 if (!memcg_proto_activated(&memcg->tcp_mem.cg_proto))
509 static_key_slow_dec(&memcg_socket_limit_enabled);
512 static void disarm_sock_keys(struct mem_cgroup *memcg)
517 static void drain_all_stock_async(struct mem_cgroup *memcg);
519 static struct mem_cgroup_per_zone *
520 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
522 return &memcg->info.nodeinfo[nid]->zoneinfo[zid];
525 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
530 static struct mem_cgroup_per_zone *
531 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
533 int nid = page_to_nid(page);
534 int zid = page_zonenum(page);
536 return mem_cgroup_zoneinfo(memcg, nid, zid);
539 static struct mem_cgroup_tree_per_zone *
540 soft_limit_tree_node_zone(int nid, int zid)
542 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
545 static struct mem_cgroup_tree_per_zone *
546 soft_limit_tree_from_page(struct page *page)
548 int nid = page_to_nid(page);
549 int zid = page_zonenum(page);
551 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
555 __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
556 struct mem_cgroup_per_zone *mz,
557 struct mem_cgroup_tree_per_zone *mctz,
558 unsigned long long new_usage_in_excess)
560 struct rb_node **p = &mctz->rb_root.rb_node;
561 struct rb_node *parent = NULL;
562 struct mem_cgroup_per_zone *mz_node;
567 mz->usage_in_excess = new_usage_in_excess;
568 if (!mz->usage_in_excess)
572 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
574 if (mz->usage_in_excess < mz_node->usage_in_excess)
577 * We can't avoid mem cgroups that are over their soft
578 * limit by the same amount
580 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
583 rb_link_node(&mz->tree_node, parent, p);
584 rb_insert_color(&mz->tree_node, &mctz->rb_root);
589 __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
590 struct mem_cgroup_per_zone *mz,
591 struct mem_cgroup_tree_per_zone *mctz)
595 rb_erase(&mz->tree_node, &mctz->rb_root);
600 mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
601 struct mem_cgroup_per_zone *mz,
602 struct mem_cgroup_tree_per_zone *mctz)
604 spin_lock(&mctz->lock);
605 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
606 spin_unlock(&mctz->lock);
610 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
612 unsigned long long excess;
613 struct mem_cgroup_per_zone *mz;
614 struct mem_cgroup_tree_per_zone *mctz;
615 int nid = page_to_nid(page);
616 int zid = page_zonenum(page);
617 mctz = soft_limit_tree_from_page(page);
620 * Necessary to update all ancestors when hierarchy is used.
621 * because their event counter is not touched.
623 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
624 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
625 excess = res_counter_soft_limit_excess(&memcg->res);
627 * We have to update the tree if mz is on RB-tree or
628 * mem is over its softlimit.
630 if (excess || mz->on_tree) {
631 spin_lock(&mctz->lock);
632 /* if on-tree, remove it */
634 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
636 * Insert again. mz->usage_in_excess will be updated.
637 * If excess is 0, no tree ops.
639 __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
640 spin_unlock(&mctz->lock);
645 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
648 struct mem_cgroup_per_zone *mz;
649 struct mem_cgroup_tree_per_zone *mctz;
651 for_each_node(node) {
652 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
653 mz = mem_cgroup_zoneinfo(memcg, node, zone);
654 mctz = soft_limit_tree_node_zone(node, zone);
655 mem_cgroup_remove_exceeded(memcg, mz, mctz);
660 static struct mem_cgroup_per_zone *
661 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
663 struct rb_node *rightmost = NULL;
664 struct mem_cgroup_per_zone *mz;
668 rightmost = rb_last(&mctz->rb_root);
670 goto done; /* Nothing to reclaim from */
672 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
674 * Remove the node now but someone else can add it back,
675 * we will to add it back at the end of reclaim to its correct
676 * position in the tree.
678 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
679 if (!res_counter_soft_limit_excess(&mz->memcg->res) ||
680 !css_tryget(&mz->memcg->css))
686 static struct mem_cgroup_per_zone *
687 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
689 struct mem_cgroup_per_zone *mz;
691 spin_lock(&mctz->lock);
692 mz = __mem_cgroup_largest_soft_limit_node(mctz);
693 spin_unlock(&mctz->lock);
698 * Implementation Note: reading percpu statistics for memcg.
700 * Both of vmstat[] and percpu_counter has threshold and do periodic
701 * synchronization to implement "quick" read. There are trade-off between
702 * reading cost and precision of value. Then, we may have a chance to implement
703 * a periodic synchronizion of counter in memcg's counter.
705 * But this _read() function is used for user interface now. The user accounts
706 * memory usage by memory cgroup and he _always_ requires exact value because
707 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
708 * have to visit all online cpus and make sum. So, for now, unnecessary
709 * synchronization is not implemented. (just implemented for cpu hotplug)
711 * If there are kernel internal actions which can make use of some not-exact
712 * value, and reading all cpu value can be performance bottleneck in some
713 * common workload, threashold and synchonization as vmstat[] should be
716 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
717 enum mem_cgroup_stat_index idx)
723 for_each_online_cpu(cpu)
724 val += per_cpu(memcg->stat->count[idx], cpu);
725 #ifdef CONFIG_HOTPLUG_CPU
726 spin_lock(&memcg->pcp_counter_lock);
727 val += memcg->nocpu_base.count[idx];
728 spin_unlock(&memcg->pcp_counter_lock);
734 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
737 int val = (charge) ? 1 : -1;
738 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
741 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
742 enum mem_cgroup_events_index idx)
744 unsigned long val = 0;
747 for_each_online_cpu(cpu)
748 val += per_cpu(memcg->stat->events[idx], cpu);
749 #ifdef CONFIG_HOTPLUG_CPU
750 spin_lock(&memcg->pcp_counter_lock);
751 val += memcg->nocpu_base.events[idx];
752 spin_unlock(&memcg->pcp_counter_lock);
757 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
758 bool anon, int nr_pages)
763 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
764 * counted as CACHE even if it's on ANON LRU.
767 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
770 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
773 /* pagein of a big page is an event. So, ignore page size */
775 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
777 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
778 nr_pages = -nr_pages; /* for event */
781 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
787 mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
789 struct mem_cgroup_per_zone *mz;
791 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
792 return mz->lru_size[lru];
796 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
797 unsigned int lru_mask)
799 struct mem_cgroup_per_zone *mz;
801 unsigned long ret = 0;
803 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
806 if (BIT(lru) & lru_mask)
807 ret += mz->lru_size[lru];
813 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
814 int nid, unsigned int lru_mask)
819 for (zid = 0; zid < MAX_NR_ZONES; zid++)
820 total += mem_cgroup_zone_nr_lru_pages(memcg,
826 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
827 unsigned int lru_mask)
832 for_each_node_state(nid, N_MEMORY)
833 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
837 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
838 enum mem_cgroup_events_target target)
840 unsigned long val, next;
842 val = __this_cpu_read(memcg->stat->nr_page_events);
843 next = __this_cpu_read(memcg->stat->targets[target]);
844 /* from time_after() in jiffies.h */
845 if ((long)next - (long)val < 0) {
847 case MEM_CGROUP_TARGET_THRESH:
848 next = val + THRESHOLDS_EVENTS_TARGET;
850 case MEM_CGROUP_TARGET_SOFTLIMIT:
851 next = val + SOFTLIMIT_EVENTS_TARGET;
853 case MEM_CGROUP_TARGET_NUMAINFO:
854 next = val + NUMAINFO_EVENTS_TARGET;
859 __this_cpu_write(memcg->stat->targets[target], next);
866 * Check events in order.
869 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
872 /* threshold event is triggered in finer grain than soft limit */
873 if (unlikely(mem_cgroup_event_ratelimit(memcg,
874 MEM_CGROUP_TARGET_THRESH))) {
876 bool do_numainfo __maybe_unused;
878 do_softlimit = mem_cgroup_event_ratelimit(memcg,
879 MEM_CGROUP_TARGET_SOFTLIMIT);
881 do_numainfo = mem_cgroup_event_ratelimit(memcg,
882 MEM_CGROUP_TARGET_NUMAINFO);
886 mem_cgroup_threshold(memcg);
887 if (unlikely(do_softlimit))
888 mem_cgroup_update_tree(memcg, page);
890 if (unlikely(do_numainfo))
891 atomic_inc(&memcg->numainfo_events);
897 struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
899 return mem_cgroup_from_css(
900 cgroup_subsys_state(cont, mem_cgroup_subsys_id));
903 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
906 * mm_update_next_owner() may clear mm->owner to NULL
907 * if it races with swapoff, page migration, etc.
908 * So this can be called with p == NULL.
913 return mem_cgroup_from_css(task_subsys_state(p, mem_cgroup_subsys_id));
916 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
918 struct mem_cgroup *memcg = NULL;
923 * Because we have no locks, mm->owner's may be being moved to other
924 * cgroup. We use css_tryget() here even if this looks
925 * pessimistic (rather than adding locks here).
929 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
930 if (unlikely(!memcg))
932 } while (!css_tryget(&memcg->css));
938 * mem_cgroup_iter - iterate over memory cgroup hierarchy
939 * @root: hierarchy root
940 * @prev: previously returned memcg, NULL on first invocation
941 * @reclaim: cookie for shared reclaim walks, NULL for full walks
943 * Returns references to children of the hierarchy below @root, or
944 * @root itself, or %NULL after a full round-trip.
946 * Caller must pass the return value in @prev on subsequent
947 * invocations for reference counting, or use mem_cgroup_iter_break()
948 * to cancel a hierarchy walk before the round-trip is complete.
950 * Reclaimers can specify a zone and a priority level in @reclaim to
951 * divide up the memcgs in the hierarchy among all concurrent
952 * reclaimers operating on the same zone and priority.
954 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
955 struct mem_cgroup *prev,
956 struct mem_cgroup_reclaim_cookie *reclaim)
958 struct mem_cgroup *memcg = NULL;
961 if (mem_cgroup_disabled())
965 root = root_mem_cgroup;
967 if (prev && !reclaim)
968 id = css_id(&prev->css);
970 if (prev && prev != root)
973 if (!root->use_hierarchy && root != root_mem_cgroup) {
980 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
981 struct cgroup_subsys_state *css;
984 int nid = zone_to_nid(reclaim->zone);
985 int zid = zone_idx(reclaim->zone);
986 struct mem_cgroup_per_zone *mz;
988 mz = mem_cgroup_zoneinfo(root, nid, zid);
989 iter = &mz->reclaim_iter[reclaim->priority];
990 if (prev && reclaim->generation != iter->generation)
996 css = css_get_next(&mem_cgroup_subsys, id + 1, &root->css, &id);
998 if (css == &root->css || css_tryget(css))
999 memcg = mem_cgroup_from_css(css);
1005 iter->position = id;
1008 else if (!prev && memcg)
1009 reclaim->generation = iter->generation;
1019 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1020 * @root: hierarchy root
1021 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1023 void mem_cgroup_iter_break(struct mem_cgroup *root,
1024 struct mem_cgroup *prev)
1027 root = root_mem_cgroup;
1028 if (prev && prev != root)
1029 css_put(&prev->css);
1033 * Iteration constructs for visiting all cgroups (under a tree). If
1034 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1035 * be used for reference counting.
1037 #define for_each_mem_cgroup_tree(iter, root) \
1038 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1040 iter = mem_cgroup_iter(root, iter, NULL))
1042 #define for_each_mem_cgroup(iter) \
1043 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1045 iter = mem_cgroup_iter(NULL, iter, NULL))
1047 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1049 struct mem_cgroup *memcg;
1052 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1053 if (unlikely(!memcg))
1058 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1061 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1069 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1072 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1073 * @zone: zone of the wanted lruvec
1074 * @memcg: memcg of the wanted lruvec
1076 * Returns the lru list vector holding pages for the given @zone and
1077 * @mem. This can be the global zone lruvec, if the memory controller
1080 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1081 struct mem_cgroup *memcg)
1083 struct mem_cgroup_per_zone *mz;
1084 struct lruvec *lruvec;
1086 if (mem_cgroup_disabled()) {
1087 lruvec = &zone->lruvec;
1091 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1092 lruvec = &mz->lruvec;
1095 * Since a node can be onlined after the mem_cgroup was created,
1096 * we have to be prepared to initialize lruvec->zone here;
1097 * and if offlined then reonlined, we need to reinitialize it.
1099 if (unlikely(lruvec->zone != zone))
1100 lruvec->zone = zone;
1105 * Following LRU functions are allowed to be used without PCG_LOCK.
1106 * Operations are called by routine of global LRU independently from memcg.
1107 * What we have to take care of here is validness of pc->mem_cgroup.
1109 * Changes to pc->mem_cgroup happens when
1112 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1113 * It is added to LRU before charge.
1114 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1115 * When moving account, the page is not on LRU. It's isolated.
1119 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1121 * @zone: zone of the page
1123 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1125 struct mem_cgroup_per_zone *mz;
1126 struct mem_cgroup *memcg;
1127 struct page_cgroup *pc;
1128 struct lruvec *lruvec;
1130 if (mem_cgroup_disabled()) {
1131 lruvec = &zone->lruvec;
1135 pc = lookup_page_cgroup(page);
1136 memcg = pc->mem_cgroup;
1139 * Surreptitiously switch any uncharged offlist page to root:
1140 * an uncharged page off lru does nothing to secure
1141 * its former mem_cgroup from sudden removal.
1143 * Our caller holds lru_lock, and PageCgroupUsed is updated
1144 * under page_cgroup lock: between them, they make all uses
1145 * of pc->mem_cgroup safe.
1147 if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1148 pc->mem_cgroup = memcg = root_mem_cgroup;
1150 mz = page_cgroup_zoneinfo(memcg, page);
1151 lruvec = &mz->lruvec;
1154 * Since a node can be onlined after the mem_cgroup was created,
1155 * we have to be prepared to initialize lruvec->zone here;
1156 * and if offlined then reonlined, we need to reinitialize it.
1158 if (unlikely(lruvec->zone != zone))
1159 lruvec->zone = zone;
1164 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1165 * @lruvec: mem_cgroup per zone lru vector
1166 * @lru: index of lru list the page is sitting on
1167 * @nr_pages: positive when adding or negative when removing
1169 * This function must be called when a page is added to or removed from an
1172 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1175 struct mem_cgroup_per_zone *mz;
1176 unsigned long *lru_size;
1178 if (mem_cgroup_disabled())
1181 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1182 lru_size = mz->lru_size + lru;
1183 *lru_size += nr_pages;
1184 VM_BUG_ON((long)(*lru_size) < 0);
1188 * Checks whether given mem is same or in the root_mem_cgroup's
1191 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1192 struct mem_cgroup *memcg)
1194 if (root_memcg == memcg)
1196 if (!root_memcg->use_hierarchy || !memcg)
1198 return css_is_ancestor(&memcg->css, &root_memcg->css);
1201 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1202 struct mem_cgroup *memcg)
1207 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1212 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *memcg)
1215 struct mem_cgroup *curr = NULL;
1216 struct task_struct *p;
1218 p = find_lock_task_mm(task);
1220 curr = try_get_mem_cgroup_from_mm(p->mm);
1224 * All threads may have already detached their mm's, but the oom
1225 * killer still needs to detect if they have already been oom
1226 * killed to prevent needlessly killing additional tasks.
1229 curr = mem_cgroup_from_task(task);
1231 css_get(&curr->css);
1237 * We should check use_hierarchy of "memcg" not "curr". Because checking
1238 * use_hierarchy of "curr" here make this function true if hierarchy is
1239 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1240 * hierarchy(even if use_hierarchy is disabled in "memcg").
1242 ret = mem_cgroup_same_or_subtree(memcg, curr);
1243 css_put(&curr->css);
1247 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1249 unsigned long inactive_ratio;
1250 unsigned long inactive;
1251 unsigned long active;
1254 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1255 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1257 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1259 inactive_ratio = int_sqrt(10 * gb);
1263 return inactive * inactive_ratio < active;
1266 int mem_cgroup_inactive_file_is_low(struct lruvec *lruvec)
1268 unsigned long active;
1269 unsigned long inactive;
1271 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_FILE);
1272 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_FILE);
1274 return (active > inactive);
1277 #define mem_cgroup_from_res_counter(counter, member) \
1278 container_of(counter, struct mem_cgroup, member)
1281 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1282 * @memcg: the memory cgroup
1284 * Returns the maximum amount of memory @mem can be charged with, in
1287 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1289 unsigned long long margin;
1291 margin = res_counter_margin(&memcg->res);
1292 if (do_swap_account)
1293 margin = min(margin, res_counter_margin(&memcg->memsw));
1294 return margin >> PAGE_SHIFT;
1297 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1299 struct cgroup *cgrp = memcg->css.cgroup;
1302 if (cgrp->parent == NULL)
1303 return vm_swappiness;
1305 return memcg->swappiness;
1309 * memcg->moving_account is used for checking possibility that some thread is
1310 * calling move_account(). When a thread on CPU-A starts moving pages under
1311 * a memcg, other threads should check memcg->moving_account under
1312 * rcu_read_lock(), like this:
1316 * memcg->moving_account+1 if (memcg->mocing_account)
1318 * synchronize_rcu() update something.
1323 /* for quick checking without looking up memcg */
1324 atomic_t memcg_moving __read_mostly;
1326 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1328 atomic_inc(&memcg_moving);
1329 atomic_inc(&memcg->moving_account);
1333 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1336 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1337 * We check NULL in callee rather than caller.
1340 atomic_dec(&memcg_moving);
1341 atomic_dec(&memcg->moving_account);
1346 * 2 routines for checking "mem" is under move_account() or not.
1348 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1349 * is used for avoiding races in accounting. If true,
1350 * pc->mem_cgroup may be overwritten.
1352 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1353 * under hierarchy of moving cgroups. This is for
1354 * waiting at hith-memory prressure caused by "move".
1357 static bool mem_cgroup_stolen(struct mem_cgroup *memcg)
1359 VM_BUG_ON(!rcu_read_lock_held());
1360 return atomic_read(&memcg->moving_account) > 0;
1363 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1365 struct mem_cgroup *from;
1366 struct mem_cgroup *to;
1369 * Unlike task_move routines, we access mc.to, mc.from not under
1370 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1372 spin_lock(&mc.lock);
1378 ret = mem_cgroup_same_or_subtree(memcg, from)
1379 || mem_cgroup_same_or_subtree(memcg, to);
1381 spin_unlock(&mc.lock);
1385 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1387 if (mc.moving_task && current != mc.moving_task) {
1388 if (mem_cgroup_under_move(memcg)) {
1390 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1391 /* moving charge context might have finished. */
1394 finish_wait(&mc.waitq, &wait);
1402 * Take this lock when
1403 * - a code tries to modify page's memcg while it's USED.
1404 * - a code tries to modify page state accounting in a memcg.
1405 * see mem_cgroup_stolen(), too.
1407 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1408 unsigned long *flags)
1410 spin_lock_irqsave(&memcg->move_lock, *flags);
1413 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1414 unsigned long *flags)
1416 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1420 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1421 * @memcg: The memory cgroup that went over limit
1422 * @p: Task that is going to be killed
1424 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1427 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1429 struct cgroup *task_cgrp;
1430 struct cgroup *mem_cgrp;
1432 * Need a buffer in BSS, can't rely on allocations. The code relies
1433 * on the assumption that OOM is serialized for memory controller.
1434 * If this assumption is broken, revisit this code.
1436 static char memcg_name[PATH_MAX];
1444 mem_cgrp = memcg->css.cgroup;
1445 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1447 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1450 * Unfortunately, we are unable to convert to a useful name
1451 * But we'll still print out the usage information
1458 printk(KERN_INFO "Task in %s killed", memcg_name);
1461 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1469 * Continues from above, so we don't need an KERN_ level
1471 printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1474 printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1475 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1476 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1477 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1478 printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1480 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1481 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1482 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1483 printk(KERN_INFO "kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1484 res_counter_read_u64(&memcg->kmem, RES_USAGE) >> 10,
1485 res_counter_read_u64(&memcg->kmem, RES_LIMIT) >> 10,
1486 res_counter_read_u64(&memcg->kmem, RES_FAILCNT));
1490 * This function returns the number of memcg under hierarchy tree. Returns
1491 * 1(self count) if no children.
1493 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1496 struct mem_cgroup *iter;
1498 for_each_mem_cgroup_tree(iter, memcg)
1504 * Return the memory (and swap, if configured) limit for a memcg.
1506 static u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1510 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1513 * Do not consider swap space if we cannot swap due to swappiness
1515 if (mem_cgroup_swappiness(memcg)) {
1518 limit += total_swap_pages << PAGE_SHIFT;
1519 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1522 * If memsw is finite and limits the amount of swap space
1523 * available to this memcg, return that limit.
1525 limit = min(limit, memsw);
1531 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1534 struct mem_cgroup *iter;
1535 unsigned long chosen_points = 0;
1536 unsigned long totalpages;
1537 unsigned int points = 0;
1538 struct task_struct *chosen = NULL;
1541 * If current has a pending SIGKILL, then automatically select it. The
1542 * goal is to allow it to allocate so that it may quickly exit and free
1545 if (fatal_signal_pending(current)) {
1546 set_thread_flag(TIF_MEMDIE);
1550 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1551 totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
1552 for_each_mem_cgroup_tree(iter, memcg) {
1553 struct cgroup *cgroup = iter->css.cgroup;
1554 struct cgroup_iter it;
1555 struct task_struct *task;
1557 cgroup_iter_start(cgroup, &it);
1558 while ((task = cgroup_iter_next(cgroup, &it))) {
1559 switch (oom_scan_process_thread(task, totalpages, NULL,
1561 case OOM_SCAN_SELECT:
1563 put_task_struct(chosen);
1565 chosen_points = ULONG_MAX;
1566 get_task_struct(chosen);
1568 case OOM_SCAN_CONTINUE:
1570 case OOM_SCAN_ABORT:
1571 cgroup_iter_end(cgroup, &it);
1572 mem_cgroup_iter_break(memcg, iter);
1574 put_task_struct(chosen);
1579 points = oom_badness(task, memcg, NULL, totalpages);
1580 if (points > chosen_points) {
1582 put_task_struct(chosen);
1584 chosen_points = points;
1585 get_task_struct(chosen);
1588 cgroup_iter_end(cgroup, &it);
1593 points = chosen_points * 1000 / totalpages;
1594 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1595 NULL, "Memory cgroup out of memory");
1598 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1600 unsigned long flags)
1602 unsigned long total = 0;
1603 bool noswap = false;
1606 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1608 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1611 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1613 drain_all_stock_async(memcg);
1614 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1616 * Allow limit shrinkers, which are triggered directly
1617 * by userspace, to catch signals and stop reclaim
1618 * after minimal progress, regardless of the margin.
1620 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1622 if (mem_cgroup_margin(memcg))
1625 * If nothing was reclaimed after two attempts, there
1626 * may be no reclaimable pages in this hierarchy.
1635 * test_mem_cgroup_node_reclaimable
1636 * @memcg: the target memcg
1637 * @nid: the node ID to be checked.
1638 * @noswap : specify true here if the user wants flle only information.
1640 * This function returns whether the specified memcg contains any
1641 * reclaimable pages on a node. Returns true if there are any reclaimable
1642 * pages in the node.
1644 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1645 int nid, bool noswap)
1647 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1649 if (noswap || !total_swap_pages)
1651 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1656 #if MAX_NUMNODES > 1
1659 * Always updating the nodemask is not very good - even if we have an empty
1660 * list or the wrong list here, we can start from some node and traverse all
1661 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1664 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1668 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1669 * pagein/pageout changes since the last update.
1671 if (!atomic_read(&memcg->numainfo_events))
1673 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1676 /* make a nodemask where this memcg uses memory from */
1677 memcg->scan_nodes = node_states[N_MEMORY];
1679 for_each_node_mask(nid, node_states[N_MEMORY]) {
1681 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1682 node_clear(nid, memcg->scan_nodes);
1685 atomic_set(&memcg->numainfo_events, 0);
1686 atomic_set(&memcg->numainfo_updating, 0);
1690 * Selecting a node where we start reclaim from. Because what we need is just
1691 * reducing usage counter, start from anywhere is O,K. Considering
1692 * memory reclaim from current node, there are pros. and cons.
1694 * Freeing memory from current node means freeing memory from a node which
1695 * we'll use or we've used. So, it may make LRU bad. And if several threads
1696 * hit limits, it will see a contention on a node. But freeing from remote
1697 * node means more costs for memory reclaim because of memory latency.
1699 * Now, we use round-robin. Better algorithm is welcomed.
1701 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1705 mem_cgroup_may_update_nodemask(memcg);
1706 node = memcg->last_scanned_node;
1708 node = next_node(node, memcg->scan_nodes);
1709 if (node == MAX_NUMNODES)
1710 node = first_node(memcg->scan_nodes);
1712 * We call this when we hit limit, not when pages are added to LRU.
1713 * No LRU may hold pages because all pages are UNEVICTABLE or
1714 * memcg is too small and all pages are not on LRU. In that case,
1715 * we use curret node.
1717 if (unlikely(node == MAX_NUMNODES))
1718 node = numa_node_id();
1720 memcg->last_scanned_node = node;
1725 * Check all nodes whether it contains reclaimable pages or not.
1726 * For quick scan, we make use of scan_nodes. This will allow us to skip
1727 * unused nodes. But scan_nodes is lazily updated and may not cotain
1728 * enough new information. We need to do double check.
1730 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1735 * quick check...making use of scan_node.
1736 * We can skip unused nodes.
1738 if (!nodes_empty(memcg->scan_nodes)) {
1739 for (nid = first_node(memcg->scan_nodes);
1741 nid = next_node(nid, memcg->scan_nodes)) {
1743 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1748 * Check rest of nodes.
1750 for_each_node_state(nid, N_MEMORY) {
1751 if (node_isset(nid, memcg->scan_nodes))
1753 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1760 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1765 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1767 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
1771 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1774 unsigned long *total_scanned)
1776 struct mem_cgroup *victim = NULL;
1779 unsigned long excess;
1780 unsigned long nr_scanned;
1781 struct mem_cgroup_reclaim_cookie reclaim = {
1786 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
1789 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1794 * If we have not been able to reclaim
1795 * anything, it might because there are
1796 * no reclaimable pages under this hierarchy
1801 * We want to do more targeted reclaim.
1802 * excess >> 2 is not to excessive so as to
1803 * reclaim too much, nor too less that we keep
1804 * coming back to reclaim from this cgroup
1806 if (total >= (excess >> 2) ||
1807 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1812 if (!mem_cgroup_reclaimable(victim, false))
1814 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1816 *total_scanned += nr_scanned;
1817 if (!res_counter_soft_limit_excess(&root_memcg->res))
1820 mem_cgroup_iter_break(root_memcg, victim);
1825 * Check OOM-Killer is already running under our hierarchy.
1826 * If someone is running, return false.
1827 * Has to be called with memcg_oom_lock
1829 static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg)
1831 struct mem_cgroup *iter, *failed = NULL;
1833 for_each_mem_cgroup_tree(iter, memcg) {
1834 if (iter->oom_lock) {
1836 * this subtree of our hierarchy is already locked
1837 * so we cannot give a lock.
1840 mem_cgroup_iter_break(memcg, iter);
1843 iter->oom_lock = true;
1850 * OK, we failed to lock the whole subtree so we have to clean up
1851 * what we set up to the failing subtree
1853 for_each_mem_cgroup_tree(iter, memcg) {
1854 if (iter == failed) {
1855 mem_cgroup_iter_break(memcg, iter);
1858 iter->oom_lock = false;
1864 * Has to be called with memcg_oom_lock
1866 static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1868 struct mem_cgroup *iter;
1870 for_each_mem_cgroup_tree(iter, memcg)
1871 iter->oom_lock = false;
1875 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1877 struct mem_cgroup *iter;
1879 for_each_mem_cgroup_tree(iter, memcg)
1880 atomic_inc(&iter->under_oom);
1883 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1885 struct mem_cgroup *iter;
1888 * When a new child is created while the hierarchy is under oom,
1889 * mem_cgroup_oom_lock() may not be called. We have to use
1890 * atomic_add_unless() here.
1892 for_each_mem_cgroup_tree(iter, memcg)
1893 atomic_add_unless(&iter->under_oom, -1, 0);
1896 static DEFINE_SPINLOCK(memcg_oom_lock);
1897 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1899 struct oom_wait_info {
1900 struct mem_cgroup *memcg;
1904 static int memcg_oom_wake_function(wait_queue_t *wait,
1905 unsigned mode, int sync, void *arg)
1907 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1908 struct mem_cgroup *oom_wait_memcg;
1909 struct oom_wait_info *oom_wait_info;
1911 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1912 oom_wait_memcg = oom_wait_info->memcg;
1915 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
1916 * Then we can use css_is_ancestor without taking care of RCU.
1918 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
1919 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
1921 return autoremove_wake_function(wait, mode, sync, arg);
1924 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
1926 /* for filtering, pass "memcg" as argument. */
1927 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1930 static void memcg_oom_recover(struct mem_cgroup *memcg)
1932 if (memcg && atomic_read(&memcg->under_oom))
1933 memcg_wakeup_oom(memcg);
1937 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1939 static bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask,
1942 struct oom_wait_info owait;
1943 bool locked, need_to_kill;
1945 owait.memcg = memcg;
1946 owait.wait.flags = 0;
1947 owait.wait.func = memcg_oom_wake_function;
1948 owait.wait.private = current;
1949 INIT_LIST_HEAD(&owait.wait.task_list);
1950 need_to_kill = true;
1951 mem_cgroup_mark_under_oom(memcg);
1953 /* At first, try to OOM lock hierarchy under memcg.*/
1954 spin_lock(&memcg_oom_lock);
1955 locked = mem_cgroup_oom_lock(memcg);
1957 * Even if signal_pending(), we can't quit charge() loop without
1958 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1959 * under OOM is always welcomed, use TASK_KILLABLE here.
1961 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1962 if (!locked || memcg->oom_kill_disable)
1963 need_to_kill = false;
1965 mem_cgroup_oom_notify(memcg);
1966 spin_unlock(&memcg_oom_lock);
1969 finish_wait(&memcg_oom_waitq, &owait.wait);
1970 mem_cgroup_out_of_memory(memcg, mask, order);
1973 finish_wait(&memcg_oom_waitq, &owait.wait);
1975 spin_lock(&memcg_oom_lock);
1977 mem_cgroup_oom_unlock(memcg);
1978 memcg_wakeup_oom(memcg);
1979 spin_unlock(&memcg_oom_lock);
1981 mem_cgroup_unmark_under_oom(memcg);
1983 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
1985 /* Give chance to dying process */
1986 schedule_timeout_uninterruptible(1);
1991 * Currently used to update mapped file statistics, but the routine can be
1992 * generalized to update other statistics as well.
1994 * Notes: Race condition
1996 * We usually use page_cgroup_lock() for accessing page_cgroup member but
1997 * it tends to be costly. But considering some conditions, we doesn't need
1998 * to do so _always_.
2000 * Considering "charge", lock_page_cgroup() is not required because all
2001 * file-stat operations happen after a page is attached to radix-tree. There
2002 * are no race with "charge".
2004 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2005 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2006 * if there are race with "uncharge". Statistics itself is properly handled
2009 * Considering "move", this is an only case we see a race. To make the race
2010 * small, we check mm->moving_account and detect there are possibility of race
2011 * If there is, we take a lock.
2014 void __mem_cgroup_begin_update_page_stat(struct page *page,
2015 bool *locked, unsigned long *flags)
2017 struct mem_cgroup *memcg;
2018 struct page_cgroup *pc;
2020 pc = lookup_page_cgroup(page);
2022 memcg = pc->mem_cgroup;
2023 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2026 * If this memory cgroup is not under account moving, we don't
2027 * need to take move_lock_mem_cgroup(). Because we already hold
2028 * rcu_read_lock(), any calls to move_account will be delayed until
2029 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2031 if (!mem_cgroup_stolen(memcg))
2034 move_lock_mem_cgroup(memcg, flags);
2035 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2036 move_unlock_mem_cgroup(memcg, flags);
2042 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
2044 struct page_cgroup *pc = lookup_page_cgroup(page);
2047 * It's guaranteed that pc->mem_cgroup never changes while
2048 * lock is held because a routine modifies pc->mem_cgroup
2049 * should take move_lock_mem_cgroup().
2051 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
2054 void mem_cgroup_update_page_stat(struct page *page,
2055 enum mem_cgroup_page_stat_item idx, int val)
2057 struct mem_cgroup *memcg;
2058 struct page_cgroup *pc = lookup_page_cgroup(page);
2059 unsigned long uninitialized_var(flags);
2061 if (mem_cgroup_disabled())
2064 memcg = pc->mem_cgroup;
2065 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2069 case MEMCG_NR_FILE_MAPPED:
2070 idx = MEM_CGROUP_STAT_FILE_MAPPED;
2076 this_cpu_add(memcg->stat->count[idx], val);
2080 * size of first charge trial. "32" comes from vmscan.c's magic value.
2081 * TODO: maybe necessary to use big numbers in big irons.
2083 #define CHARGE_BATCH 32U
2084 struct memcg_stock_pcp {
2085 struct mem_cgroup *cached; /* this never be root cgroup */
2086 unsigned int nr_pages;
2087 struct work_struct work;
2088 unsigned long flags;
2089 #define FLUSHING_CACHED_CHARGE 0
2091 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2092 static DEFINE_MUTEX(percpu_charge_mutex);
2095 * consume_stock: Try to consume stocked charge on this cpu.
2096 * @memcg: memcg to consume from.
2097 * @nr_pages: how many pages to charge.
2099 * The charges will only happen if @memcg matches the current cpu's memcg
2100 * stock, and at least @nr_pages are available in that stock. Failure to
2101 * service an allocation will refill the stock.
2103 * returns true if successful, false otherwise.
2105 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2107 struct memcg_stock_pcp *stock;
2110 if (nr_pages > CHARGE_BATCH)
2113 stock = &get_cpu_var(memcg_stock);
2114 if (memcg == stock->cached && stock->nr_pages >= nr_pages)
2115 stock->nr_pages -= nr_pages;
2116 else /* need to call res_counter_charge */
2118 put_cpu_var(memcg_stock);
2123 * Returns stocks cached in percpu to res_counter and reset cached information.
2125 static void drain_stock(struct memcg_stock_pcp *stock)
2127 struct mem_cgroup *old = stock->cached;
2129 if (stock->nr_pages) {
2130 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2132 res_counter_uncharge(&old->res, bytes);
2133 if (do_swap_account)
2134 res_counter_uncharge(&old->memsw, bytes);
2135 stock->nr_pages = 0;
2137 stock->cached = NULL;
2141 * This must be called under preempt disabled or must be called by
2142 * a thread which is pinned to local cpu.
2144 static void drain_local_stock(struct work_struct *dummy)
2146 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2148 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2152 * Cache charges(val) which is from res_counter, to local per_cpu area.
2153 * This will be consumed by consume_stock() function, later.
2155 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2157 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2159 if (stock->cached != memcg) { /* reset if necessary */
2161 stock->cached = memcg;
2163 stock->nr_pages += nr_pages;
2164 put_cpu_var(memcg_stock);
2168 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2169 * of the hierarchy under it. sync flag says whether we should block
2170 * until the work is done.
2172 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2176 /* Notify other cpus that system-wide "drain" is running */
2179 for_each_online_cpu(cpu) {
2180 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2181 struct mem_cgroup *memcg;
2183 memcg = stock->cached;
2184 if (!memcg || !stock->nr_pages)
2186 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2188 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2190 drain_local_stock(&stock->work);
2192 schedule_work_on(cpu, &stock->work);
2200 for_each_online_cpu(cpu) {
2201 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2202 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2203 flush_work(&stock->work);
2210 * Tries to drain stocked charges in other cpus. This function is asynchronous
2211 * and just put a work per cpu for draining localy on each cpu. Caller can
2212 * expects some charges will be back to res_counter later but cannot wait for
2215 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2218 * If someone calls draining, avoid adding more kworker runs.
2220 if (!mutex_trylock(&percpu_charge_mutex))
2222 drain_all_stock(root_memcg, false);
2223 mutex_unlock(&percpu_charge_mutex);
2226 /* This is a synchronous drain interface. */
2227 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2229 /* called when force_empty is called */
2230 mutex_lock(&percpu_charge_mutex);
2231 drain_all_stock(root_memcg, true);
2232 mutex_unlock(&percpu_charge_mutex);
2236 * This function drains percpu counter value from DEAD cpu and
2237 * move it to local cpu. Note that this function can be preempted.
2239 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2243 spin_lock(&memcg->pcp_counter_lock);
2244 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2245 long x = per_cpu(memcg->stat->count[i], cpu);
2247 per_cpu(memcg->stat->count[i], cpu) = 0;
2248 memcg->nocpu_base.count[i] += x;
2250 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2251 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2253 per_cpu(memcg->stat->events[i], cpu) = 0;
2254 memcg->nocpu_base.events[i] += x;
2256 spin_unlock(&memcg->pcp_counter_lock);
2259 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2260 unsigned long action,
2263 int cpu = (unsigned long)hcpu;
2264 struct memcg_stock_pcp *stock;
2265 struct mem_cgroup *iter;
2267 if (action == CPU_ONLINE)
2270 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2273 for_each_mem_cgroup(iter)
2274 mem_cgroup_drain_pcp_counter(iter, cpu);
2276 stock = &per_cpu(memcg_stock, cpu);
2282 /* See __mem_cgroup_try_charge() for details */
2284 CHARGE_OK, /* success */
2285 CHARGE_RETRY, /* need to retry but retry is not bad */
2286 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2287 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2288 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2291 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2292 unsigned int nr_pages, unsigned int min_pages,
2295 unsigned long csize = nr_pages * PAGE_SIZE;
2296 struct mem_cgroup *mem_over_limit;
2297 struct res_counter *fail_res;
2298 unsigned long flags = 0;
2301 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2304 if (!do_swap_account)
2306 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2310 res_counter_uncharge(&memcg->res, csize);
2311 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2312 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2314 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2316 * Never reclaim on behalf of optional batching, retry with a
2317 * single page instead.
2319 if (nr_pages > min_pages)
2320 return CHARGE_RETRY;
2322 if (!(gfp_mask & __GFP_WAIT))
2323 return CHARGE_WOULDBLOCK;
2325 if (gfp_mask & __GFP_NORETRY)
2326 return CHARGE_NOMEM;
2328 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2329 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2330 return CHARGE_RETRY;
2332 * Even though the limit is exceeded at this point, reclaim
2333 * may have been able to free some pages. Retry the charge
2334 * before killing the task.
2336 * Only for regular pages, though: huge pages are rather
2337 * unlikely to succeed so close to the limit, and we fall back
2338 * to regular pages anyway in case of failure.
2340 if (nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER) && ret)
2341 return CHARGE_RETRY;
2344 * At task move, charge accounts can be doubly counted. So, it's
2345 * better to wait until the end of task_move if something is going on.
2347 if (mem_cgroup_wait_acct_move(mem_over_limit))
2348 return CHARGE_RETRY;
2350 /* If we don't need to call oom-killer at el, return immediately */
2352 return CHARGE_NOMEM;
2354 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask, get_order(csize)))
2355 return CHARGE_OOM_DIE;
2357 return CHARGE_RETRY;
2361 * __mem_cgroup_try_charge() does
2362 * 1. detect memcg to be charged against from passed *mm and *ptr,
2363 * 2. update res_counter
2364 * 3. call memory reclaim if necessary.
2366 * In some special case, if the task is fatal, fatal_signal_pending() or
2367 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2368 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2369 * as possible without any hazards. 2: all pages should have a valid
2370 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2371 * pointer, that is treated as a charge to root_mem_cgroup.
2373 * So __mem_cgroup_try_charge() will return
2374 * 0 ... on success, filling *ptr with a valid memcg pointer.
2375 * -ENOMEM ... charge failure because of resource limits.
2376 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2378 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2379 * the oom-killer can be invoked.
2381 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2383 unsigned int nr_pages,
2384 struct mem_cgroup **ptr,
2387 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2388 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2389 struct mem_cgroup *memcg = NULL;
2393 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2394 * in system level. So, allow to go ahead dying process in addition to
2397 if (unlikely(test_thread_flag(TIF_MEMDIE)
2398 || fatal_signal_pending(current)))
2402 * We always charge the cgroup the mm_struct belongs to.
2403 * The mm_struct's mem_cgroup changes on task migration if the
2404 * thread group leader migrates. It's possible that mm is not
2405 * set, if so charge the root memcg (happens for pagecache usage).
2408 *ptr = root_mem_cgroup;
2410 if (*ptr) { /* css should be a valid one */
2412 if (mem_cgroup_is_root(memcg))
2414 if (consume_stock(memcg, nr_pages))
2416 css_get(&memcg->css);
2418 struct task_struct *p;
2421 p = rcu_dereference(mm->owner);
2423 * Because we don't have task_lock(), "p" can exit.
2424 * In that case, "memcg" can point to root or p can be NULL with
2425 * race with swapoff. Then, we have small risk of mis-accouning.
2426 * But such kind of mis-account by race always happens because
2427 * we don't have cgroup_mutex(). It's overkill and we allo that
2429 * (*) swapoff at el will charge against mm-struct not against
2430 * task-struct. So, mm->owner can be NULL.
2432 memcg = mem_cgroup_from_task(p);
2434 memcg = root_mem_cgroup;
2435 if (mem_cgroup_is_root(memcg)) {
2439 if (consume_stock(memcg, nr_pages)) {
2441 * It seems dagerous to access memcg without css_get().
2442 * But considering how consume_stok works, it's not
2443 * necessary. If consume_stock success, some charges
2444 * from this memcg are cached on this cpu. So, we
2445 * don't need to call css_get()/css_tryget() before
2446 * calling consume_stock().
2451 /* after here, we may be blocked. we need to get refcnt */
2452 if (!css_tryget(&memcg->css)) {
2462 /* If killed, bypass charge */
2463 if (fatal_signal_pending(current)) {
2464 css_put(&memcg->css);
2469 if (oom && !nr_oom_retries) {
2471 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2474 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, nr_pages,
2479 case CHARGE_RETRY: /* not in OOM situation but retry */
2481 css_put(&memcg->css);
2484 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2485 css_put(&memcg->css);
2487 case CHARGE_NOMEM: /* OOM routine works */
2489 css_put(&memcg->css);
2492 /* If oom, we never return -ENOMEM */
2495 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2496 css_put(&memcg->css);
2499 } while (ret != CHARGE_OK);
2501 if (batch > nr_pages)
2502 refill_stock(memcg, batch - nr_pages);
2503 css_put(&memcg->css);
2511 *ptr = root_mem_cgroup;
2516 * Somemtimes we have to undo a charge we got by try_charge().
2517 * This function is for that and do uncharge, put css's refcnt.
2518 * gotten by try_charge().
2520 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2521 unsigned int nr_pages)
2523 if (!mem_cgroup_is_root(memcg)) {
2524 unsigned long bytes = nr_pages * PAGE_SIZE;
2526 res_counter_uncharge(&memcg->res, bytes);
2527 if (do_swap_account)
2528 res_counter_uncharge(&memcg->memsw, bytes);
2533 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2534 * This is useful when moving usage to parent cgroup.
2536 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2537 unsigned int nr_pages)
2539 unsigned long bytes = nr_pages * PAGE_SIZE;
2541 if (mem_cgroup_is_root(memcg))
2544 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2545 if (do_swap_account)
2546 res_counter_uncharge_until(&memcg->memsw,
2547 memcg->memsw.parent, bytes);
2551 * A helper function to get mem_cgroup from ID. must be called under
2552 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2553 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2554 * called against removed memcg.)
2556 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2558 struct cgroup_subsys_state *css;
2560 /* ID 0 is unused ID */
2563 css = css_lookup(&mem_cgroup_subsys, id);
2566 return mem_cgroup_from_css(css);
2569 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2571 struct mem_cgroup *memcg = NULL;
2572 struct page_cgroup *pc;
2576 VM_BUG_ON(!PageLocked(page));
2578 pc = lookup_page_cgroup(page);
2579 lock_page_cgroup(pc);
2580 if (PageCgroupUsed(pc)) {
2581 memcg = pc->mem_cgroup;
2582 if (memcg && !css_tryget(&memcg->css))
2584 } else if (PageSwapCache(page)) {
2585 ent.val = page_private(page);
2586 id = lookup_swap_cgroup_id(ent);
2588 memcg = mem_cgroup_lookup(id);
2589 if (memcg && !css_tryget(&memcg->css))
2593 unlock_page_cgroup(pc);
2597 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2599 unsigned int nr_pages,
2600 enum charge_type ctype,
2603 struct page_cgroup *pc = lookup_page_cgroup(page);
2604 struct zone *uninitialized_var(zone);
2605 struct lruvec *lruvec;
2606 bool was_on_lru = false;
2609 lock_page_cgroup(pc);
2610 VM_BUG_ON(PageCgroupUsed(pc));
2612 * we don't need page_cgroup_lock about tail pages, becase they are not
2613 * accessed by any other context at this point.
2617 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2618 * may already be on some other mem_cgroup's LRU. Take care of it.
2621 zone = page_zone(page);
2622 spin_lock_irq(&zone->lru_lock);
2623 if (PageLRU(page)) {
2624 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2626 del_page_from_lru_list(page, lruvec, page_lru(page));
2631 pc->mem_cgroup = memcg;
2633 * We access a page_cgroup asynchronously without lock_page_cgroup().
2634 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2635 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2636 * before USED bit, we need memory barrier here.
2637 * See mem_cgroup_add_lru_list(), etc.
2640 SetPageCgroupUsed(pc);
2644 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2645 VM_BUG_ON(PageLRU(page));
2647 add_page_to_lru_list(page, lruvec, page_lru(page));
2649 spin_unlock_irq(&zone->lru_lock);
2652 if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
2657 mem_cgroup_charge_statistics(memcg, anon, nr_pages);
2658 unlock_page_cgroup(pc);
2661 * "charge_statistics" updated event counter. Then, check it.
2662 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2663 * if they exceeds softlimit.
2665 memcg_check_events(memcg, page);
2668 #ifdef CONFIG_MEMCG_KMEM
2669 static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg)
2671 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) &&
2672 (memcg->kmem_account_flags & KMEM_ACCOUNTED_MASK);
2675 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size)
2677 struct res_counter *fail_res;
2678 struct mem_cgroup *_memcg;
2682 ret = res_counter_charge(&memcg->kmem, size, &fail_res);
2687 * Conditions under which we can wait for the oom_killer. Those are
2688 * the same conditions tested by the core page allocator
2690 may_oom = (gfp & __GFP_FS) && !(gfp & __GFP_NORETRY);
2693 ret = __mem_cgroup_try_charge(NULL, gfp, size >> PAGE_SHIFT,
2696 if (ret == -EINTR) {
2698 * __mem_cgroup_try_charge() chosed to bypass to root due to
2699 * OOM kill or fatal signal. Since our only options are to
2700 * either fail the allocation or charge it to this cgroup, do
2701 * it as a temporary condition. But we can't fail. From a
2702 * kmem/slab perspective, the cache has already been selected,
2703 * by mem_cgroup_kmem_get_cache(), so it is too late to change
2706 * This condition will only trigger if the task entered
2707 * memcg_charge_kmem in a sane state, but was OOM-killed during
2708 * __mem_cgroup_try_charge() above. Tasks that were already
2709 * dying when the allocation triggers should have been already
2710 * directed to the root cgroup in memcontrol.h
2712 res_counter_charge_nofail(&memcg->res, size, &fail_res);
2713 if (do_swap_account)
2714 res_counter_charge_nofail(&memcg->memsw, size,
2718 res_counter_uncharge(&memcg->kmem, size);
2723 static void memcg_uncharge_kmem(struct mem_cgroup *memcg, u64 size)
2725 res_counter_uncharge(&memcg->kmem, size);
2726 res_counter_uncharge(&memcg->res, size);
2727 if (do_swap_account)
2728 res_counter_uncharge(&memcg->memsw, size);
2732 * We need to verify if the allocation against current->mm->owner's memcg is
2733 * possible for the given order. But the page is not allocated yet, so we'll
2734 * need a further commit step to do the final arrangements.
2736 * It is possible for the task to switch cgroups in this mean time, so at
2737 * commit time, we can't rely on task conversion any longer. We'll then use
2738 * the handle argument to return to the caller which cgroup we should commit
2739 * against. We could also return the memcg directly and avoid the pointer
2740 * passing, but a boolean return value gives better semantics considering
2741 * the compiled-out case as well.
2743 * Returning true means the allocation is possible.
2746 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
2748 struct mem_cgroup *memcg;
2752 memcg = try_get_mem_cgroup_from_mm(current->mm);
2755 * very rare case described in mem_cgroup_from_task. Unfortunately there
2756 * isn't much we can do without complicating this too much, and it would
2757 * be gfp-dependent anyway. Just let it go
2759 if (unlikely(!memcg))
2762 if (!memcg_can_account_kmem(memcg)) {
2763 css_put(&memcg->css);
2767 mem_cgroup_get(memcg);
2769 ret = memcg_charge_kmem(memcg, gfp, PAGE_SIZE << order);
2773 mem_cgroup_put(memcg);
2775 css_put(&memcg->css);
2779 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
2782 struct page_cgroup *pc;
2784 VM_BUG_ON(mem_cgroup_is_root(memcg));
2786 /* The page allocation failed. Revert */
2788 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
2789 mem_cgroup_put(memcg);
2793 pc = lookup_page_cgroup(page);
2794 lock_page_cgroup(pc);
2795 pc->mem_cgroup = memcg;
2796 SetPageCgroupUsed(pc);
2797 unlock_page_cgroup(pc);
2800 void __memcg_kmem_uncharge_pages(struct page *page, int order)
2802 struct mem_cgroup *memcg = NULL;
2803 struct page_cgroup *pc;
2806 pc = lookup_page_cgroup(page);
2808 * Fast unlocked return. Theoretically might have changed, have to
2809 * check again after locking.
2811 if (!PageCgroupUsed(pc))
2814 lock_page_cgroup(pc);
2815 if (PageCgroupUsed(pc)) {
2816 memcg = pc->mem_cgroup;
2817 ClearPageCgroupUsed(pc);
2819 unlock_page_cgroup(pc);
2822 * We trust that only if there is a memcg associated with the page, it
2823 * is a valid allocation
2828 VM_BUG_ON(mem_cgroup_is_root(memcg));
2829 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
2830 mem_cgroup_put(memcg);
2832 #endif /* CONFIG_MEMCG_KMEM */
2834 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2836 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
2838 * Because tail pages are not marked as "used", set it. We're under
2839 * zone->lru_lock, 'splitting on pmd' and compound_lock.
2840 * charge/uncharge will be never happen and move_account() is done under
2841 * compound_lock(), so we don't have to take care of races.
2843 void mem_cgroup_split_huge_fixup(struct page *head)
2845 struct page_cgroup *head_pc = lookup_page_cgroup(head);
2846 struct page_cgroup *pc;
2849 if (mem_cgroup_disabled())
2851 for (i = 1; i < HPAGE_PMD_NR; i++) {
2853 pc->mem_cgroup = head_pc->mem_cgroup;
2854 smp_wmb();/* see __commit_charge() */
2855 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2858 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2861 * mem_cgroup_move_account - move account of the page
2863 * @nr_pages: number of regular pages (>1 for huge pages)
2864 * @pc: page_cgroup of the page.
2865 * @from: mem_cgroup which the page is moved from.
2866 * @to: mem_cgroup which the page is moved to. @from != @to.
2868 * The caller must confirm following.
2869 * - page is not on LRU (isolate_page() is useful.)
2870 * - compound_lock is held when nr_pages > 1
2872 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
2875 static int mem_cgroup_move_account(struct page *page,
2876 unsigned int nr_pages,
2877 struct page_cgroup *pc,
2878 struct mem_cgroup *from,
2879 struct mem_cgroup *to)
2881 unsigned long flags;
2883 bool anon = PageAnon(page);
2885 VM_BUG_ON(from == to);
2886 VM_BUG_ON(PageLRU(page));
2888 * The page is isolated from LRU. So, collapse function
2889 * will not handle this page. But page splitting can happen.
2890 * Do this check under compound_page_lock(). The caller should
2894 if (nr_pages > 1 && !PageTransHuge(page))
2897 lock_page_cgroup(pc);
2900 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
2903 move_lock_mem_cgroup(from, &flags);
2905 if (!anon && page_mapped(page)) {
2906 /* Update mapped_file data for mem_cgroup */
2908 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2909 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2912 mem_cgroup_charge_statistics(from, anon, -nr_pages);
2914 /* caller should have done css_get */
2915 pc->mem_cgroup = to;
2916 mem_cgroup_charge_statistics(to, anon, nr_pages);
2917 move_unlock_mem_cgroup(from, &flags);
2920 unlock_page_cgroup(pc);
2924 memcg_check_events(to, page);
2925 memcg_check_events(from, page);
2931 * mem_cgroup_move_parent - moves page to the parent group
2932 * @page: the page to move
2933 * @pc: page_cgroup of the page
2934 * @child: page's cgroup
2936 * move charges to its parent or the root cgroup if the group has no
2937 * parent (aka use_hierarchy==0).
2938 * Although this might fail (get_page_unless_zero, isolate_lru_page or
2939 * mem_cgroup_move_account fails) the failure is always temporary and
2940 * it signals a race with a page removal/uncharge or migration. In the
2941 * first case the page is on the way out and it will vanish from the LRU
2942 * on the next attempt and the call should be retried later.
2943 * Isolation from the LRU fails only if page has been isolated from
2944 * the LRU since we looked at it and that usually means either global
2945 * reclaim or migration going on. The page will either get back to the
2947 * Finaly mem_cgroup_move_account fails only if the page got uncharged
2948 * (!PageCgroupUsed) or moved to a different group. The page will
2949 * disappear in the next attempt.
2951 static int mem_cgroup_move_parent(struct page *page,
2952 struct page_cgroup *pc,
2953 struct mem_cgroup *child)
2955 struct mem_cgroup *parent;
2956 unsigned int nr_pages;
2957 unsigned long uninitialized_var(flags);
2960 VM_BUG_ON(mem_cgroup_is_root(child));
2963 if (!get_page_unless_zero(page))
2965 if (isolate_lru_page(page))
2968 nr_pages = hpage_nr_pages(page);
2970 parent = parent_mem_cgroup(child);
2972 * If no parent, move charges to root cgroup.
2975 parent = root_mem_cgroup;
2978 VM_BUG_ON(!PageTransHuge(page));
2979 flags = compound_lock_irqsave(page);
2982 ret = mem_cgroup_move_account(page, nr_pages,
2985 __mem_cgroup_cancel_local_charge(child, nr_pages);
2988 compound_unlock_irqrestore(page, flags);
2989 putback_lru_page(page);
2997 * Charge the memory controller for page usage.
2999 * 0 if the charge was successful
3000 * < 0 if the cgroup is over its limit
3002 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
3003 gfp_t gfp_mask, enum charge_type ctype)
3005 struct mem_cgroup *memcg = NULL;
3006 unsigned int nr_pages = 1;
3010 if (PageTransHuge(page)) {
3011 nr_pages <<= compound_order(page);
3012 VM_BUG_ON(!PageTransHuge(page));
3014 * Never OOM-kill a process for a huge page. The
3015 * fault handler will fall back to regular pages.
3020 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
3023 __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false);
3027 int mem_cgroup_newpage_charge(struct page *page,
3028 struct mm_struct *mm, gfp_t gfp_mask)
3030 if (mem_cgroup_disabled())
3032 VM_BUG_ON(page_mapped(page));
3033 VM_BUG_ON(page->mapping && !PageAnon(page));
3035 return mem_cgroup_charge_common(page, mm, gfp_mask,
3036 MEM_CGROUP_CHARGE_TYPE_ANON);
3040 * While swap-in, try_charge -> commit or cancel, the page is locked.
3041 * And when try_charge() successfully returns, one refcnt to memcg without
3042 * struct page_cgroup is acquired. This refcnt will be consumed by
3043 * "commit()" or removed by "cancel()"
3045 static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm,
3048 struct mem_cgroup **memcgp)
3050 struct mem_cgroup *memcg;
3051 struct page_cgroup *pc;
3054 pc = lookup_page_cgroup(page);
3056 * Every swap fault against a single page tries to charge the
3057 * page, bail as early as possible. shmem_unuse() encounters
3058 * already charged pages, too. The USED bit is protected by
3059 * the page lock, which serializes swap cache removal, which
3060 * in turn serializes uncharging.
3062 if (PageCgroupUsed(pc))
3064 if (!do_swap_account)
3066 memcg = try_get_mem_cgroup_from_page(page);
3070 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
3071 css_put(&memcg->css);
3076 ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
3082 int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page,
3083 gfp_t gfp_mask, struct mem_cgroup **memcgp)
3086 if (mem_cgroup_disabled())
3089 * A racing thread's fault, or swapoff, may have already
3090 * updated the pte, and even removed page from swap cache: in
3091 * those cases unuse_pte()'s pte_same() test will fail; but
3092 * there's also a KSM case which does need to charge the page.
3094 if (!PageSwapCache(page)) {
3097 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, memcgp, true);
3102 return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp);
3105 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
3107 if (mem_cgroup_disabled())
3111 __mem_cgroup_cancel_charge(memcg, 1);
3115 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
3116 enum charge_type ctype)
3118 if (mem_cgroup_disabled())
3123 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
3125 * Now swap is on-memory. This means this page may be
3126 * counted both as mem and swap....double count.
3127 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
3128 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
3129 * may call delete_from_swap_cache() before reach here.
3131 if (do_swap_account && PageSwapCache(page)) {
3132 swp_entry_t ent = {.val = page_private(page)};
3133 mem_cgroup_uncharge_swap(ent);
3137 void mem_cgroup_commit_charge_swapin(struct page *page,
3138 struct mem_cgroup *memcg)
3140 __mem_cgroup_commit_charge_swapin(page, memcg,
3141 MEM_CGROUP_CHARGE_TYPE_ANON);
3144 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
3147 struct mem_cgroup *memcg = NULL;
3148 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
3151 if (mem_cgroup_disabled())
3153 if (PageCompound(page))
3156 if (!PageSwapCache(page))
3157 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
3158 else { /* page is swapcache/shmem */
3159 ret = __mem_cgroup_try_charge_swapin(mm, page,
3162 __mem_cgroup_commit_charge_swapin(page, memcg, type);
3167 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
3168 unsigned int nr_pages,
3169 const enum charge_type ctype)
3171 struct memcg_batch_info *batch = NULL;
3172 bool uncharge_memsw = true;
3174 /* If swapout, usage of swap doesn't decrease */
3175 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
3176 uncharge_memsw = false;
3178 batch = ¤t->memcg_batch;
3180 * In usual, we do css_get() when we remember memcg pointer.
3181 * But in this case, we keep res->usage until end of a series of
3182 * uncharges. Then, it's ok to ignore memcg's refcnt.
3185 batch->memcg = memcg;
3187 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
3188 * In those cases, all pages freed continuously can be expected to be in
3189 * the same cgroup and we have chance to coalesce uncharges.
3190 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
3191 * because we want to do uncharge as soon as possible.
3194 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
3195 goto direct_uncharge;
3198 goto direct_uncharge;
3201 * In typical case, batch->memcg == mem. This means we can
3202 * merge a series of uncharges to an uncharge of res_counter.
3203 * If not, we uncharge res_counter ony by one.
3205 if (batch->memcg != memcg)
3206 goto direct_uncharge;
3207 /* remember freed charge and uncharge it later */
3210 batch->memsw_nr_pages++;
3213 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
3215 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
3216 if (unlikely(batch->memcg != memcg))
3217 memcg_oom_recover(memcg);
3221 * uncharge if !page_mapped(page)
3223 static struct mem_cgroup *
3224 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
3227 struct mem_cgroup *memcg = NULL;
3228 unsigned int nr_pages = 1;
3229 struct page_cgroup *pc;
3232 if (mem_cgroup_disabled())
3235 VM_BUG_ON(PageSwapCache(page));
3237 if (PageTransHuge(page)) {
3238 nr_pages <<= compound_order(page);
3239 VM_BUG_ON(!PageTransHuge(page));
3242 * Check if our page_cgroup is valid
3244 pc = lookup_page_cgroup(page);
3245 if (unlikely(!PageCgroupUsed(pc)))
3248 lock_page_cgroup(pc);
3250 memcg = pc->mem_cgroup;
3252 if (!PageCgroupUsed(pc))
3255 anon = PageAnon(page);
3258 case MEM_CGROUP_CHARGE_TYPE_ANON:
3260 * Generally PageAnon tells if it's the anon statistics to be
3261 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
3262 * used before page reached the stage of being marked PageAnon.
3266 case MEM_CGROUP_CHARGE_TYPE_DROP:
3267 /* See mem_cgroup_prepare_migration() */
3268 if (page_mapped(page))
3271 * Pages under migration may not be uncharged. But
3272 * end_migration() /must/ be the one uncharging the
3273 * unused post-migration page and so it has to call
3274 * here with the migration bit still set. See the
3275 * res_counter handling below.
3277 if (!end_migration && PageCgroupMigration(pc))
3280 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
3281 if (!PageAnon(page)) { /* Shared memory */
3282 if (page->mapping && !page_is_file_cache(page))
3284 } else if (page_mapped(page)) /* Anon */
3291 mem_cgroup_charge_statistics(memcg, anon, -nr_pages);
3293 ClearPageCgroupUsed(pc);
3295 * pc->mem_cgroup is not cleared here. It will be accessed when it's
3296 * freed from LRU. This is safe because uncharged page is expected not
3297 * to be reused (freed soon). Exception is SwapCache, it's handled by
3298 * special functions.
3301 unlock_page_cgroup(pc);
3303 * even after unlock, we have memcg->res.usage here and this memcg
3304 * will never be freed.
3306 memcg_check_events(memcg, page);
3307 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
3308 mem_cgroup_swap_statistics(memcg, true);
3309 mem_cgroup_get(memcg);
3312 * Migration does not charge the res_counter for the
3313 * replacement page, so leave it alone when phasing out the
3314 * page that is unused after the migration.
3316 if (!end_migration && !mem_cgroup_is_root(memcg))
3317 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
3322 unlock_page_cgroup(pc);
3326 void mem_cgroup_uncharge_page(struct page *page)
3329 if (page_mapped(page))
3331 VM_BUG_ON(page->mapping && !PageAnon(page));
3332 if (PageSwapCache(page))
3334 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
3337 void mem_cgroup_uncharge_cache_page(struct page *page)
3339 VM_BUG_ON(page_mapped(page));
3340 VM_BUG_ON(page->mapping);
3341 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
3345 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
3346 * In that cases, pages are freed continuously and we can expect pages
3347 * are in the same memcg. All these calls itself limits the number of
3348 * pages freed at once, then uncharge_start/end() is called properly.
3349 * This may be called prural(2) times in a context,
3352 void mem_cgroup_uncharge_start(void)
3354 current->memcg_batch.do_batch++;
3355 /* We can do nest. */
3356 if (current->memcg_batch.do_batch == 1) {
3357 current->memcg_batch.memcg = NULL;
3358 current->memcg_batch.nr_pages = 0;
3359 current->memcg_batch.memsw_nr_pages = 0;
3363 void mem_cgroup_uncharge_end(void)
3365 struct memcg_batch_info *batch = ¤t->memcg_batch;
3367 if (!batch->do_batch)
3371 if (batch->do_batch) /* If stacked, do nothing. */
3377 * This "batch->memcg" is valid without any css_get/put etc...
3378 * bacause we hide charges behind us.
3380 if (batch->nr_pages)
3381 res_counter_uncharge(&batch->memcg->res,
3382 batch->nr_pages * PAGE_SIZE);
3383 if (batch->memsw_nr_pages)
3384 res_counter_uncharge(&batch->memcg->memsw,
3385 batch->memsw_nr_pages * PAGE_SIZE);
3386 memcg_oom_recover(batch->memcg);
3387 /* forget this pointer (for sanity check) */
3388 batch->memcg = NULL;
3393 * called after __delete_from_swap_cache() and drop "page" account.
3394 * memcg information is recorded to swap_cgroup of "ent"
3397 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
3399 struct mem_cgroup *memcg;
3400 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
3402 if (!swapout) /* this was a swap cache but the swap is unused ! */
3403 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
3405 memcg = __mem_cgroup_uncharge_common(page, ctype, false);
3408 * record memcg information, if swapout && memcg != NULL,
3409 * mem_cgroup_get() was called in uncharge().
3411 if (do_swap_account && swapout && memcg)
3412 swap_cgroup_record(ent, css_id(&memcg->css));
3416 #ifdef CONFIG_MEMCG_SWAP
3418 * called from swap_entry_free(). remove record in swap_cgroup and
3419 * uncharge "memsw" account.
3421 void mem_cgroup_uncharge_swap(swp_entry_t ent)
3423 struct mem_cgroup *memcg;
3426 if (!do_swap_account)
3429 id = swap_cgroup_record(ent, 0);
3431 memcg = mem_cgroup_lookup(id);
3434 * We uncharge this because swap is freed.
3435 * This memcg can be obsolete one. We avoid calling css_tryget
3437 if (!mem_cgroup_is_root(memcg))
3438 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
3439 mem_cgroup_swap_statistics(memcg, false);
3440 mem_cgroup_put(memcg);
3446 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3447 * @entry: swap entry to be moved
3448 * @from: mem_cgroup which the entry is moved from
3449 * @to: mem_cgroup which the entry is moved to
3451 * It succeeds only when the swap_cgroup's record for this entry is the same
3452 * as the mem_cgroup's id of @from.
3454 * Returns 0 on success, -EINVAL on failure.
3456 * The caller must have charged to @to, IOW, called res_counter_charge() about
3457 * both res and memsw, and called css_get().
3459 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3460 struct mem_cgroup *from, struct mem_cgroup *to)
3462 unsigned short old_id, new_id;
3464 old_id = css_id(&from->css);
3465 new_id = css_id(&to->css);
3467 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3468 mem_cgroup_swap_statistics(from, false);
3469 mem_cgroup_swap_statistics(to, true);
3471 * This function is only called from task migration context now.
3472 * It postpones res_counter and refcount handling till the end
3473 * of task migration(mem_cgroup_clear_mc()) for performance
3474 * improvement. But we cannot postpone mem_cgroup_get(to)
3475 * because if the process that has been moved to @to does
3476 * swap-in, the refcount of @to might be decreased to 0.
3484 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3485 struct mem_cgroup *from, struct mem_cgroup *to)
3492 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
3495 void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
3496 struct mem_cgroup **memcgp)
3498 struct mem_cgroup *memcg = NULL;
3499 unsigned int nr_pages = 1;
3500 struct page_cgroup *pc;
3501 enum charge_type ctype;
3505 if (mem_cgroup_disabled())
3508 if (PageTransHuge(page))
3509 nr_pages <<= compound_order(page);
3511 pc = lookup_page_cgroup(page);
3512 lock_page_cgroup(pc);
3513 if (PageCgroupUsed(pc)) {
3514 memcg = pc->mem_cgroup;
3515 css_get(&memcg->css);
3517 * At migrating an anonymous page, its mapcount goes down
3518 * to 0 and uncharge() will be called. But, even if it's fully
3519 * unmapped, migration may fail and this page has to be
3520 * charged again. We set MIGRATION flag here and delay uncharge
3521 * until end_migration() is called
3523 * Corner Case Thinking
3525 * When the old page was mapped as Anon and it's unmap-and-freed
3526 * while migration was ongoing.
3527 * If unmap finds the old page, uncharge() of it will be delayed
3528 * until end_migration(). If unmap finds a new page, it's
3529 * uncharged when it make mapcount to be 1->0. If unmap code
3530 * finds swap_migration_entry, the new page will not be mapped
3531 * and end_migration() will find it(mapcount==0).
3534 * When the old page was mapped but migraion fails, the kernel
3535 * remaps it. A charge for it is kept by MIGRATION flag even
3536 * if mapcount goes down to 0. We can do remap successfully
3537 * without charging it again.
3540 * The "old" page is under lock_page() until the end of
3541 * migration, so, the old page itself will not be swapped-out.
3542 * If the new page is swapped out before end_migraton, our
3543 * hook to usual swap-out path will catch the event.
3546 SetPageCgroupMigration(pc);
3548 unlock_page_cgroup(pc);
3550 * If the page is not charged at this point,
3558 * We charge new page before it's used/mapped. So, even if unlock_page()
3559 * is called before end_migration, we can catch all events on this new
3560 * page. In the case new page is migrated but not remapped, new page's
3561 * mapcount will be finally 0 and we call uncharge in end_migration().
3564 ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
3566 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
3568 * The page is committed to the memcg, but it's not actually
3569 * charged to the res_counter since we plan on replacing the
3570 * old one and only one page is going to be left afterwards.
3572 __mem_cgroup_commit_charge(memcg, newpage, nr_pages, ctype, false);
3575 /* remove redundant charge if migration failed*/
3576 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
3577 struct page *oldpage, struct page *newpage, bool migration_ok)
3579 struct page *used, *unused;
3580 struct page_cgroup *pc;
3586 if (!migration_ok) {
3593 anon = PageAnon(used);
3594 __mem_cgroup_uncharge_common(unused,
3595 anon ? MEM_CGROUP_CHARGE_TYPE_ANON
3596 : MEM_CGROUP_CHARGE_TYPE_CACHE,
3598 css_put(&memcg->css);
3600 * We disallowed uncharge of pages under migration because mapcount
3601 * of the page goes down to zero, temporarly.
3602 * Clear the flag and check the page should be charged.
3604 pc = lookup_page_cgroup(oldpage);
3605 lock_page_cgroup(pc);
3606 ClearPageCgroupMigration(pc);
3607 unlock_page_cgroup(pc);
3610 * If a page is a file cache, radix-tree replacement is very atomic
3611 * and we can skip this check. When it was an Anon page, its mapcount
3612 * goes down to 0. But because we added MIGRATION flage, it's not
3613 * uncharged yet. There are several case but page->mapcount check
3614 * and USED bit check in mem_cgroup_uncharge_page() will do enough
3615 * check. (see prepare_charge() also)
3618 mem_cgroup_uncharge_page(used);
3622 * At replace page cache, newpage is not under any memcg but it's on
3623 * LRU. So, this function doesn't touch res_counter but handles LRU
3624 * in correct way. Both pages are locked so we cannot race with uncharge.
3626 void mem_cgroup_replace_page_cache(struct page *oldpage,
3627 struct page *newpage)
3629 struct mem_cgroup *memcg = NULL;
3630 struct page_cgroup *pc;
3631 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
3633 if (mem_cgroup_disabled())
3636 pc = lookup_page_cgroup(oldpage);
3637 /* fix accounting on old pages */
3638 lock_page_cgroup(pc);
3639 if (PageCgroupUsed(pc)) {
3640 memcg = pc->mem_cgroup;
3641 mem_cgroup_charge_statistics(memcg, false, -1);
3642 ClearPageCgroupUsed(pc);
3644 unlock_page_cgroup(pc);
3647 * When called from shmem_replace_page(), in some cases the
3648 * oldpage has already been charged, and in some cases not.
3653 * Even if newpage->mapping was NULL before starting replacement,
3654 * the newpage may be on LRU(or pagevec for LRU) already. We lock
3655 * LRU while we overwrite pc->mem_cgroup.
3657 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
3660 #ifdef CONFIG_DEBUG_VM
3661 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3663 struct page_cgroup *pc;
3665 pc = lookup_page_cgroup(page);
3667 * Can be NULL while feeding pages into the page allocator for
3668 * the first time, i.e. during boot or memory hotplug;
3669 * or when mem_cgroup_disabled().
3671 if (likely(pc) && PageCgroupUsed(pc))
3676 bool mem_cgroup_bad_page_check(struct page *page)
3678 if (mem_cgroup_disabled())
3681 return lookup_page_cgroup_used(page) != NULL;
3684 void mem_cgroup_print_bad_page(struct page *page)
3686 struct page_cgroup *pc;
3688 pc = lookup_page_cgroup_used(page);
3690 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
3691 pc, pc->flags, pc->mem_cgroup);
3696 static DEFINE_MUTEX(set_limit_mutex);
3698 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3699 unsigned long long val)
3702 u64 memswlimit, memlimit;
3704 int children = mem_cgroup_count_children(memcg);
3705 u64 curusage, oldusage;
3709 * For keeping hierarchical_reclaim simple, how long we should retry
3710 * is depends on callers. We set our retry-count to be function
3711 * of # of children which we should visit in this loop.
3713 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3715 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3718 while (retry_count) {
3719 if (signal_pending(current)) {
3724 * Rather than hide all in some function, I do this in
3725 * open coded manner. You see what this really does.
3726 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
3728 mutex_lock(&set_limit_mutex);
3729 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3730 if (memswlimit < val) {
3732 mutex_unlock(&set_limit_mutex);
3736 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3740 ret = res_counter_set_limit(&memcg->res, val);
3742 if (memswlimit == val)
3743 memcg->memsw_is_minimum = true;
3745 memcg->memsw_is_minimum = false;
3747 mutex_unlock(&set_limit_mutex);
3752 mem_cgroup_reclaim(memcg, GFP_KERNEL,
3753 MEM_CGROUP_RECLAIM_SHRINK);
3754 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3755 /* Usage is reduced ? */
3756 if (curusage >= oldusage)
3759 oldusage = curusage;
3761 if (!ret && enlarge)
3762 memcg_oom_recover(memcg);
3767 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3768 unsigned long long val)
3771 u64 memlimit, memswlimit, oldusage, curusage;
3772 int children = mem_cgroup_count_children(memcg);
3776 /* see mem_cgroup_resize_res_limit */
3777 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3778 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3779 while (retry_count) {
3780 if (signal_pending(current)) {
3785 * Rather than hide all in some function, I do this in
3786 * open coded manner. You see what this really does.
3787 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
3789 mutex_lock(&set_limit_mutex);
3790 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3791 if (memlimit > val) {
3793 mutex_unlock(&set_limit_mutex);
3796 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3797 if (memswlimit < val)
3799 ret = res_counter_set_limit(&memcg->memsw, val);
3801 if (memlimit == val)
3802 memcg->memsw_is_minimum = true;
3804 memcg->memsw_is_minimum = false;
3806 mutex_unlock(&set_limit_mutex);
3811 mem_cgroup_reclaim(memcg, GFP_KERNEL,
3812 MEM_CGROUP_RECLAIM_NOSWAP |
3813 MEM_CGROUP_RECLAIM_SHRINK);
3814 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3815 /* Usage is reduced ? */
3816 if (curusage >= oldusage)
3819 oldusage = curusage;
3821 if (!ret && enlarge)
3822 memcg_oom_recover(memcg);
3826 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3828 unsigned long *total_scanned)
3830 unsigned long nr_reclaimed = 0;
3831 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3832 unsigned long reclaimed;
3834 struct mem_cgroup_tree_per_zone *mctz;
3835 unsigned long long excess;
3836 unsigned long nr_scanned;
3841 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3843 * This loop can run a while, specially if mem_cgroup's continuously
3844 * keep exceeding their soft limit and putting the system under
3851 mz = mem_cgroup_largest_soft_limit_node(mctz);
3856 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
3857 gfp_mask, &nr_scanned);
3858 nr_reclaimed += reclaimed;
3859 *total_scanned += nr_scanned;
3860 spin_lock(&mctz->lock);
3863 * If we failed to reclaim anything from this memory cgroup
3864 * it is time to move on to the next cgroup
3870 * Loop until we find yet another one.
3872 * By the time we get the soft_limit lock
3873 * again, someone might have aded the
3874 * group back on the RB tree. Iterate to
3875 * make sure we get a different mem.
3876 * mem_cgroup_largest_soft_limit_node returns
3877 * NULL if no other cgroup is present on
3881 __mem_cgroup_largest_soft_limit_node(mctz);
3883 css_put(&next_mz->memcg->css);
3884 else /* next_mz == NULL or other memcg */
3888 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
3889 excess = res_counter_soft_limit_excess(&mz->memcg->res);
3891 * One school of thought says that we should not add
3892 * back the node to the tree if reclaim returns 0.
3893 * But our reclaim could return 0, simply because due
3894 * to priority we are exposing a smaller subset of
3895 * memory to reclaim from. Consider this as a longer
3898 /* If excess == 0, no tree ops */
3899 __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess);
3900 spin_unlock(&mctz->lock);
3901 css_put(&mz->memcg->css);
3904 * Could not reclaim anything and there are no more
3905 * mem cgroups to try or we seem to be looping without
3906 * reclaiming anything.
3908 if (!nr_reclaimed &&
3910 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3912 } while (!nr_reclaimed);
3914 css_put(&next_mz->memcg->css);
3915 return nr_reclaimed;
3919 * mem_cgroup_force_empty_list - clears LRU of a group
3920 * @memcg: group to clear
3923 * @lru: lru to to clear
3925 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
3926 * reclaim the pages page themselves - pages are moved to the parent (or root)
3929 static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
3930 int node, int zid, enum lru_list lru)
3932 struct lruvec *lruvec;
3933 unsigned long flags;
3934 struct list_head *list;
3938 zone = &NODE_DATA(node)->node_zones[zid];
3939 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
3940 list = &lruvec->lists[lru];
3944 struct page_cgroup *pc;
3947 spin_lock_irqsave(&zone->lru_lock, flags);
3948 if (list_empty(list)) {
3949 spin_unlock_irqrestore(&zone->lru_lock, flags);
3952 page = list_entry(list->prev, struct page, lru);
3954 list_move(&page->lru, list);
3956 spin_unlock_irqrestore(&zone->lru_lock, flags);
3959 spin_unlock_irqrestore(&zone->lru_lock, flags);
3961 pc = lookup_page_cgroup(page);
3963 if (mem_cgroup_move_parent(page, pc, memcg)) {
3964 /* found lock contention or "pc" is obsolete. */
3969 } while (!list_empty(list));
3973 * make mem_cgroup's charge to be 0 if there is no task by moving
3974 * all the charges and pages to the parent.
3975 * This enables deleting this mem_cgroup.
3977 * Caller is responsible for holding css reference on the memcg.
3979 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
3984 /* This is for making all *used* pages to be on LRU. */
3985 lru_add_drain_all();
3986 drain_all_stock_sync(memcg);
3987 mem_cgroup_start_move(memcg);
3988 for_each_node_state(node, N_MEMORY) {
3989 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3992 mem_cgroup_force_empty_list(memcg,
3997 mem_cgroup_end_move(memcg);
3998 memcg_oom_recover(memcg);
4002 * This is a safety check because mem_cgroup_force_empty_list
4003 * could have raced with mem_cgroup_replace_page_cache callers
4004 * so the lru seemed empty but the page could have been added
4005 * right after the check. RES_USAGE should be safe as we always
4006 * charge before adding to the LRU.
4008 } while (res_counter_read_u64(&memcg->res, RES_USAGE) > 0);
4012 * Reclaims as many pages from the given memcg as possible and moves
4013 * the rest to the parent.
4015 * Caller is responsible for holding css reference for memcg.
4017 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
4019 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
4020 struct cgroup *cgrp = memcg->css.cgroup;
4022 /* returns EBUSY if there is a task or if we come here twice. */
4023 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
4026 /* we call try-to-free pages for make this cgroup empty */
4027 lru_add_drain_all();
4028 /* try to free all pages in this cgroup */
4029 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
4032 if (signal_pending(current))
4035 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
4039 /* maybe some writeback is necessary */
4040 congestion_wait(BLK_RW_ASYNC, HZ/10);
4045 mem_cgroup_reparent_charges(memcg);
4050 static int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
4052 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4055 if (mem_cgroup_is_root(memcg))
4057 css_get(&memcg->css);
4058 ret = mem_cgroup_force_empty(memcg);
4059 css_put(&memcg->css);
4065 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
4067 return mem_cgroup_from_cont(cont)->use_hierarchy;
4070 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
4074 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4075 struct cgroup *parent = cont->parent;
4076 struct mem_cgroup *parent_memcg = NULL;
4079 parent_memcg = mem_cgroup_from_cont(parent);
4083 if (memcg->use_hierarchy == val)
4087 * If parent's use_hierarchy is set, we can't make any modifications
4088 * in the child subtrees. If it is unset, then the change can
4089 * occur, provided the current cgroup has no children.
4091 * For the root cgroup, parent_mem is NULL, we allow value to be
4092 * set if there are no children.
4094 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
4095 (val == 1 || val == 0)) {
4096 if (list_empty(&cont->children))
4097 memcg->use_hierarchy = val;
4110 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
4111 enum mem_cgroup_stat_index idx)
4113 struct mem_cgroup *iter;
4116 /* Per-cpu values can be negative, use a signed accumulator */
4117 for_each_mem_cgroup_tree(iter, memcg)
4118 val += mem_cgroup_read_stat(iter, idx);
4120 if (val < 0) /* race ? */
4125 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
4129 if (!mem_cgroup_is_root(memcg)) {
4131 return res_counter_read_u64(&memcg->res, RES_USAGE);
4133 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
4136 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
4137 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
4140 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
4142 return val << PAGE_SHIFT;
4145 static ssize_t mem_cgroup_read(struct cgroup *cont, struct cftype *cft,
4146 struct file *file, char __user *buf,
4147 size_t nbytes, loff_t *ppos)
4149 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4155 type = MEMFILE_TYPE(cft->private);
4156 name = MEMFILE_ATTR(cft->private);
4158 if (!do_swap_account && type == _MEMSWAP)
4163 if (name == RES_USAGE)
4164 val = mem_cgroup_usage(memcg, false);
4166 val = res_counter_read_u64(&memcg->res, name);
4169 if (name == RES_USAGE)
4170 val = mem_cgroup_usage(memcg, true);
4172 val = res_counter_read_u64(&memcg->memsw, name);
4175 val = res_counter_read_u64(&memcg->kmem, name);
4181 len = scnprintf(str, sizeof(str), "%llu\n", (unsigned long long)val);
4182 return simple_read_from_buffer(buf, nbytes, ppos, str, len);
4185 static int memcg_update_kmem_limit(struct cgroup *cont, u64 val)
4188 #ifdef CONFIG_MEMCG_KMEM
4189 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4191 * For simplicity, we won't allow this to be disabled. It also can't
4192 * be changed if the cgroup has children already, or if tasks had
4195 * If tasks join before we set the limit, a person looking at
4196 * kmem.usage_in_bytes will have no way to determine when it took
4197 * place, which makes the value quite meaningless.
4199 * After it first became limited, changes in the value of the limit are
4200 * of course permitted.
4202 * Taking the cgroup_lock is really offensive, but it is so far the only
4203 * way to guarantee that no children will appear. There are plenty of
4204 * other offenders, and they should all go away. Fine grained locking
4205 * is probably the way to go here. When we are fully hierarchical, we
4206 * can also get rid of the use_hierarchy check.
4209 mutex_lock(&set_limit_mutex);
4210 if (!memcg->kmem_account_flags && val != RESOURCE_MAX) {
4211 if (cgroup_task_count(cont) || (memcg->use_hierarchy &&
4212 !list_empty(&cont->children))) {
4216 ret = res_counter_set_limit(&memcg->kmem, val);
4219 memcg_kmem_set_active(memcg);
4221 ret = res_counter_set_limit(&memcg->kmem, val);
4223 mutex_unlock(&set_limit_mutex);
4229 static void memcg_propagate_kmem(struct mem_cgroup *memcg)
4231 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4234 memcg->kmem_account_flags = parent->kmem_account_flags;
4238 * The user of this function is...
4241 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
4244 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4247 unsigned long long val;
4250 type = MEMFILE_TYPE(cft->private);
4251 name = MEMFILE_ATTR(cft->private);
4253 if (!do_swap_account && type == _MEMSWAP)
4258 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
4262 /* This function does all necessary parse...reuse it */
4263 ret = res_counter_memparse_write_strategy(buffer, &val);
4267 ret = mem_cgroup_resize_limit(memcg, val);
4268 else if (type == _MEMSWAP)
4269 ret = mem_cgroup_resize_memsw_limit(memcg, val);
4270 else if (type == _KMEM)
4271 ret = memcg_update_kmem_limit(cont, val);
4275 case RES_SOFT_LIMIT:
4276 ret = res_counter_memparse_write_strategy(buffer, &val);
4280 * For memsw, soft limits are hard to implement in terms
4281 * of semantics, for now, we support soft limits for
4282 * control without swap
4285 ret = res_counter_set_soft_limit(&memcg->res, val);
4290 ret = -EINVAL; /* should be BUG() ? */
4296 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
4297 unsigned long long *mem_limit, unsigned long long *memsw_limit)
4299 struct cgroup *cgroup;
4300 unsigned long long min_limit, min_memsw_limit, tmp;
4302 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4303 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4304 cgroup = memcg->css.cgroup;
4305 if (!memcg->use_hierarchy)
4308 while (cgroup->parent) {
4309 cgroup = cgroup->parent;
4310 memcg = mem_cgroup_from_cont(cgroup);
4311 if (!memcg->use_hierarchy)
4313 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
4314 min_limit = min(min_limit, tmp);
4315 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4316 min_memsw_limit = min(min_memsw_limit, tmp);
4319 *mem_limit = min_limit;
4320 *memsw_limit = min_memsw_limit;
4323 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
4325 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4329 type = MEMFILE_TYPE(event);
4330 name = MEMFILE_ATTR(event);
4332 if (!do_swap_account && type == _MEMSWAP)
4338 res_counter_reset_max(&memcg->res);
4339 else if (type == _MEMSWAP)
4340 res_counter_reset_max(&memcg->memsw);
4341 else if (type == _KMEM)
4342 res_counter_reset_max(&memcg->kmem);
4348 res_counter_reset_failcnt(&memcg->res);
4349 else if (type == _MEMSWAP)
4350 res_counter_reset_failcnt(&memcg->memsw);
4351 else if (type == _KMEM)
4352 res_counter_reset_failcnt(&memcg->kmem);
4361 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
4364 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
4368 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4369 struct cftype *cft, u64 val)
4371 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4373 if (val >= (1 << NR_MOVE_TYPE))
4376 * We check this value several times in both in can_attach() and
4377 * attach(), so we need cgroup lock to prevent this value from being
4381 memcg->move_charge_at_immigrate = val;
4387 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4388 struct cftype *cft, u64 val)
4395 static int memcg_numa_stat_show(struct cgroup *cont, struct cftype *cft,
4399 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
4400 unsigned long node_nr;
4401 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4403 total_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL);
4404 seq_printf(m, "total=%lu", total_nr);
4405 for_each_node_state(nid, N_MEMORY) {
4406 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL);
4407 seq_printf(m, " N%d=%lu", nid, node_nr);
4411 file_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_FILE);
4412 seq_printf(m, "file=%lu", file_nr);
4413 for_each_node_state(nid, N_MEMORY) {
4414 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4416 seq_printf(m, " N%d=%lu", nid, node_nr);
4420 anon_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_ANON);
4421 seq_printf(m, "anon=%lu", anon_nr);
4422 for_each_node_state(nid, N_MEMORY) {
4423 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4425 seq_printf(m, " N%d=%lu", nid, node_nr);
4429 unevictable_nr = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
4430 seq_printf(m, "unevictable=%lu", unevictable_nr);
4431 for_each_node_state(nid, N_MEMORY) {
4432 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4433 BIT(LRU_UNEVICTABLE));
4434 seq_printf(m, " N%d=%lu", nid, node_nr);
4439 #endif /* CONFIG_NUMA */
4441 static const char * const mem_cgroup_lru_names[] = {
4449 static inline void mem_cgroup_lru_names_not_uptodate(void)
4451 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
4454 static int memcg_stat_show(struct cgroup *cont, struct cftype *cft,
4457 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4458 struct mem_cgroup *mi;
4461 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
4462 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
4464 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
4465 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
4468 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
4469 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
4470 mem_cgroup_read_events(memcg, i));
4472 for (i = 0; i < NR_LRU_LISTS; i++)
4473 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
4474 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
4476 /* Hierarchical information */
4478 unsigned long long limit, memsw_limit;
4479 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
4480 seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
4481 if (do_swap_account)
4482 seq_printf(m, "hierarchical_memsw_limit %llu\n",
4486 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
4489 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
4491 for_each_mem_cgroup_tree(mi, memcg)
4492 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
4493 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
4496 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
4497 unsigned long long val = 0;
4499 for_each_mem_cgroup_tree(mi, memcg)
4500 val += mem_cgroup_read_events(mi, i);
4501 seq_printf(m, "total_%s %llu\n",
4502 mem_cgroup_events_names[i], val);
4505 for (i = 0; i < NR_LRU_LISTS; i++) {
4506 unsigned long long val = 0;
4508 for_each_mem_cgroup_tree(mi, memcg)
4509 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
4510 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
4513 #ifdef CONFIG_DEBUG_VM
4516 struct mem_cgroup_per_zone *mz;
4517 struct zone_reclaim_stat *rstat;
4518 unsigned long recent_rotated[2] = {0, 0};
4519 unsigned long recent_scanned[2] = {0, 0};
4521 for_each_online_node(nid)
4522 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4523 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
4524 rstat = &mz->lruvec.reclaim_stat;
4526 recent_rotated[0] += rstat->recent_rotated[0];
4527 recent_rotated[1] += rstat->recent_rotated[1];
4528 recent_scanned[0] += rstat->recent_scanned[0];
4529 recent_scanned[1] += rstat->recent_scanned[1];
4531 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
4532 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
4533 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
4534 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
4541 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
4543 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4545 return mem_cgroup_swappiness(memcg);
4548 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
4551 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4552 struct mem_cgroup *parent;
4557 if (cgrp->parent == NULL)
4560 parent = mem_cgroup_from_cont(cgrp->parent);
4564 /* If under hierarchy, only empty-root can set this value */
4565 if ((parent->use_hierarchy) ||
4566 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4571 memcg->swappiness = val;
4578 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4580 struct mem_cgroup_threshold_ary *t;
4586 t = rcu_dereference(memcg->thresholds.primary);
4588 t = rcu_dereference(memcg->memsw_thresholds.primary);
4593 usage = mem_cgroup_usage(memcg, swap);
4596 * current_threshold points to threshold just below or equal to usage.
4597 * If it's not true, a threshold was crossed after last
4598 * call of __mem_cgroup_threshold().
4600 i = t->current_threshold;
4603 * Iterate backward over array of thresholds starting from
4604 * current_threshold and check if a threshold is crossed.
4605 * If none of thresholds below usage is crossed, we read
4606 * only one element of the array here.
4608 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4609 eventfd_signal(t->entries[i].eventfd, 1);
4611 /* i = current_threshold + 1 */
4615 * Iterate forward over array of thresholds starting from
4616 * current_threshold+1 and check if a threshold is crossed.
4617 * If none of thresholds above usage is crossed, we read
4618 * only one element of the array here.
4620 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4621 eventfd_signal(t->entries[i].eventfd, 1);
4623 /* Update current_threshold */
4624 t->current_threshold = i - 1;
4629 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4632 __mem_cgroup_threshold(memcg, false);
4633 if (do_swap_account)
4634 __mem_cgroup_threshold(memcg, true);
4636 memcg = parent_mem_cgroup(memcg);
4640 static int compare_thresholds(const void *a, const void *b)
4642 const struct mem_cgroup_threshold *_a = a;
4643 const struct mem_cgroup_threshold *_b = b;
4645 return _a->threshold - _b->threshold;
4648 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4650 struct mem_cgroup_eventfd_list *ev;
4652 list_for_each_entry(ev, &memcg->oom_notify, list)
4653 eventfd_signal(ev->eventfd, 1);
4657 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4659 struct mem_cgroup *iter;
4661 for_each_mem_cgroup_tree(iter, memcg)
4662 mem_cgroup_oom_notify_cb(iter);
4665 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
4666 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4668 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4669 struct mem_cgroup_thresholds *thresholds;
4670 struct mem_cgroup_threshold_ary *new;
4671 enum res_type type = MEMFILE_TYPE(cft->private);
4672 u64 threshold, usage;
4675 ret = res_counter_memparse_write_strategy(args, &threshold);
4679 mutex_lock(&memcg->thresholds_lock);
4682 thresholds = &memcg->thresholds;
4683 else if (type == _MEMSWAP)
4684 thresholds = &memcg->memsw_thresholds;
4688 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4690 /* Check if a threshold crossed before adding a new one */
4691 if (thresholds->primary)
4692 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4694 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4696 /* Allocate memory for new array of thresholds */
4697 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4705 /* Copy thresholds (if any) to new array */
4706 if (thresholds->primary) {
4707 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4708 sizeof(struct mem_cgroup_threshold));
4711 /* Add new threshold */
4712 new->entries[size - 1].eventfd = eventfd;
4713 new->entries[size - 1].threshold = threshold;
4715 /* Sort thresholds. Registering of new threshold isn't time-critical */
4716 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4717 compare_thresholds, NULL);
4719 /* Find current threshold */
4720 new->current_threshold = -1;
4721 for (i = 0; i < size; i++) {
4722 if (new->entries[i].threshold <= usage) {
4724 * new->current_threshold will not be used until
4725 * rcu_assign_pointer(), so it's safe to increment
4728 ++new->current_threshold;
4733 /* Free old spare buffer and save old primary buffer as spare */
4734 kfree(thresholds->spare);
4735 thresholds->spare = thresholds->primary;
4737 rcu_assign_pointer(thresholds->primary, new);
4739 /* To be sure that nobody uses thresholds */
4743 mutex_unlock(&memcg->thresholds_lock);
4748 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
4749 struct cftype *cft, struct eventfd_ctx *eventfd)
4751 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4752 struct mem_cgroup_thresholds *thresholds;
4753 struct mem_cgroup_threshold_ary *new;
4754 enum res_type type = MEMFILE_TYPE(cft->private);
4758 mutex_lock(&memcg->thresholds_lock);
4760 thresholds = &memcg->thresholds;
4761 else if (type == _MEMSWAP)
4762 thresholds = &memcg->memsw_thresholds;
4766 if (!thresholds->primary)
4769 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4771 /* Check if a threshold crossed before removing */
4772 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4774 /* Calculate new number of threshold */
4776 for (i = 0; i < thresholds->primary->size; i++) {
4777 if (thresholds->primary->entries[i].eventfd != eventfd)
4781 new = thresholds->spare;
4783 /* Set thresholds array to NULL if we don't have thresholds */
4792 /* Copy thresholds and find current threshold */
4793 new->current_threshold = -1;
4794 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4795 if (thresholds->primary->entries[i].eventfd == eventfd)
4798 new->entries[j] = thresholds->primary->entries[i];
4799 if (new->entries[j].threshold <= usage) {
4801 * new->current_threshold will not be used
4802 * until rcu_assign_pointer(), so it's safe to increment
4805 ++new->current_threshold;
4811 /* Swap primary and spare array */
4812 thresholds->spare = thresholds->primary;
4813 /* If all events are unregistered, free the spare array */
4815 kfree(thresholds->spare);
4816 thresholds->spare = NULL;
4819 rcu_assign_pointer(thresholds->primary, new);
4821 /* To be sure that nobody uses thresholds */
4824 mutex_unlock(&memcg->thresholds_lock);
4827 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4828 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4830 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4831 struct mem_cgroup_eventfd_list *event;
4832 enum res_type type = MEMFILE_TYPE(cft->private);
4834 BUG_ON(type != _OOM_TYPE);
4835 event = kmalloc(sizeof(*event), GFP_KERNEL);
4839 spin_lock(&memcg_oom_lock);
4841 event->eventfd = eventfd;
4842 list_add(&event->list, &memcg->oom_notify);
4844 /* already in OOM ? */
4845 if (atomic_read(&memcg->under_oom))
4846 eventfd_signal(eventfd, 1);
4847 spin_unlock(&memcg_oom_lock);
4852 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4853 struct cftype *cft, struct eventfd_ctx *eventfd)
4855 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4856 struct mem_cgroup_eventfd_list *ev, *tmp;
4857 enum res_type type = MEMFILE_TYPE(cft->private);
4859 BUG_ON(type != _OOM_TYPE);
4861 spin_lock(&memcg_oom_lock);
4863 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4864 if (ev->eventfd == eventfd) {
4865 list_del(&ev->list);
4870 spin_unlock(&memcg_oom_lock);
4873 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4874 struct cftype *cft, struct cgroup_map_cb *cb)
4876 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4878 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
4880 if (atomic_read(&memcg->under_oom))
4881 cb->fill(cb, "under_oom", 1);
4883 cb->fill(cb, "under_oom", 0);
4887 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4888 struct cftype *cft, u64 val)
4890 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4891 struct mem_cgroup *parent;
4893 /* cannot set to root cgroup and only 0 and 1 are allowed */
4894 if (!cgrp->parent || !((val == 0) || (val == 1)))
4897 parent = mem_cgroup_from_cont(cgrp->parent);
4900 /* oom-kill-disable is a flag for subhierarchy. */
4901 if ((parent->use_hierarchy) ||
4902 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4906 memcg->oom_kill_disable = val;
4908 memcg_oom_recover(memcg);
4913 #ifdef CONFIG_MEMCG_KMEM
4914 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
4916 memcg_propagate_kmem(memcg);
4917 return mem_cgroup_sockets_init(memcg, ss);
4920 static void kmem_cgroup_destroy(struct mem_cgroup *memcg)
4922 mem_cgroup_sockets_destroy(memcg);
4925 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
4930 static void kmem_cgroup_destroy(struct mem_cgroup *memcg)
4935 static struct cftype mem_cgroup_files[] = {
4937 .name = "usage_in_bytes",
4938 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4939 .read = mem_cgroup_read,
4940 .register_event = mem_cgroup_usage_register_event,
4941 .unregister_event = mem_cgroup_usage_unregister_event,
4944 .name = "max_usage_in_bytes",
4945 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4946 .trigger = mem_cgroup_reset,
4947 .read = mem_cgroup_read,
4950 .name = "limit_in_bytes",
4951 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4952 .write_string = mem_cgroup_write,
4953 .read = mem_cgroup_read,
4956 .name = "soft_limit_in_bytes",
4957 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4958 .write_string = mem_cgroup_write,
4959 .read = mem_cgroup_read,
4963 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4964 .trigger = mem_cgroup_reset,
4965 .read = mem_cgroup_read,
4969 .read_seq_string = memcg_stat_show,
4972 .name = "force_empty",
4973 .trigger = mem_cgroup_force_empty_write,
4976 .name = "use_hierarchy",
4977 .write_u64 = mem_cgroup_hierarchy_write,
4978 .read_u64 = mem_cgroup_hierarchy_read,
4981 .name = "swappiness",
4982 .read_u64 = mem_cgroup_swappiness_read,
4983 .write_u64 = mem_cgroup_swappiness_write,
4986 .name = "move_charge_at_immigrate",
4987 .read_u64 = mem_cgroup_move_charge_read,
4988 .write_u64 = mem_cgroup_move_charge_write,
4991 .name = "oom_control",
4992 .read_map = mem_cgroup_oom_control_read,
4993 .write_u64 = mem_cgroup_oom_control_write,
4994 .register_event = mem_cgroup_oom_register_event,
4995 .unregister_event = mem_cgroup_oom_unregister_event,
4996 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
5000 .name = "numa_stat",
5001 .read_seq_string = memcg_numa_stat_show,
5004 #ifdef CONFIG_MEMCG_SWAP
5006 .name = "memsw.usage_in_bytes",
5007 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
5008 .read = mem_cgroup_read,
5009 .register_event = mem_cgroup_usage_register_event,
5010 .unregister_event = mem_cgroup_usage_unregister_event,
5013 .name = "memsw.max_usage_in_bytes",
5014 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
5015 .trigger = mem_cgroup_reset,
5016 .read = mem_cgroup_read,
5019 .name = "memsw.limit_in_bytes",
5020 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
5021 .write_string = mem_cgroup_write,
5022 .read = mem_cgroup_read,
5025 .name = "memsw.failcnt",
5026 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
5027 .trigger = mem_cgroup_reset,
5028 .read = mem_cgroup_read,
5031 #ifdef CONFIG_MEMCG_KMEM
5033 .name = "kmem.limit_in_bytes",
5034 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
5035 .write_string = mem_cgroup_write,
5036 .read = mem_cgroup_read,
5039 .name = "kmem.usage_in_bytes",
5040 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5041 .read = mem_cgroup_read,
5044 .name = "kmem.failcnt",
5045 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5046 .trigger = mem_cgroup_reset,
5047 .read = mem_cgroup_read,
5050 .name = "kmem.max_usage_in_bytes",
5051 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5052 .trigger = mem_cgroup_reset,
5053 .read = mem_cgroup_read,
5056 { }, /* terminate */
5059 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
5061 struct mem_cgroup_per_node *pn;
5062 struct mem_cgroup_per_zone *mz;
5063 int zone, tmp = node;
5065 * This routine is called against possible nodes.
5066 * But it's BUG to call kmalloc() against offline node.
5068 * TODO: this routine can waste much memory for nodes which will
5069 * never be onlined. It's better to use memory hotplug callback
5072 if (!node_state(node, N_NORMAL_MEMORY))
5074 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5078 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5079 mz = &pn->zoneinfo[zone];
5080 lruvec_init(&mz->lruvec);
5081 mz->usage_in_excess = 0;
5082 mz->on_tree = false;
5085 memcg->info.nodeinfo[node] = pn;
5089 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
5091 kfree(memcg->info.nodeinfo[node]);
5094 static struct mem_cgroup *mem_cgroup_alloc(void)
5096 struct mem_cgroup *memcg;
5097 int size = sizeof(struct mem_cgroup);
5099 /* Can be very big if MAX_NUMNODES is very big */
5100 if (size < PAGE_SIZE)
5101 memcg = kzalloc(size, GFP_KERNEL);
5103 memcg = vzalloc(size);
5108 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
5111 spin_lock_init(&memcg->pcp_counter_lock);
5115 if (size < PAGE_SIZE)
5123 * Helpers for freeing a kmalloc()ed/vzalloc()ed mem_cgroup by RCU,
5124 * but in process context. The work_freeing structure is overlaid
5125 * on the rcu_freeing structure, which itself is overlaid on memsw.
5127 static void free_work(struct work_struct *work)
5129 struct mem_cgroup *memcg;
5130 int size = sizeof(struct mem_cgroup);
5132 memcg = container_of(work, struct mem_cgroup, work_freeing);
5134 * We need to make sure that (at least for now), the jump label
5135 * destruction code runs outside of the cgroup lock. This is because
5136 * get_online_cpus(), which is called from the static_branch update,
5137 * can't be called inside the cgroup_lock. cpusets are the ones
5138 * enforcing this dependency, so if they ever change, we might as well.
5140 * schedule_work() will guarantee this happens. Be careful if you need
5141 * to move this code around, and make sure it is outside
5144 disarm_sock_keys(memcg);
5145 if (size < PAGE_SIZE)
5151 static void free_rcu(struct rcu_head *rcu_head)
5153 struct mem_cgroup *memcg;
5155 memcg = container_of(rcu_head, struct mem_cgroup, rcu_freeing);
5156 INIT_WORK(&memcg->work_freeing, free_work);
5157 schedule_work(&memcg->work_freeing);
5161 * At destroying mem_cgroup, references from swap_cgroup can remain.
5162 * (scanning all at force_empty is too costly...)
5164 * Instead of clearing all references at force_empty, we remember
5165 * the number of reference from swap_cgroup and free mem_cgroup when
5166 * it goes down to 0.
5168 * Removal of cgroup itself succeeds regardless of refs from swap.
5171 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5175 mem_cgroup_remove_from_trees(memcg);
5176 free_css_id(&mem_cgroup_subsys, &memcg->css);
5179 free_mem_cgroup_per_zone_info(memcg, node);
5181 free_percpu(memcg->stat);
5182 call_rcu(&memcg->rcu_freeing, free_rcu);
5185 static void mem_cgroup_get(struct mem_cgroup *memcg)
5187 atomic_inc(&memcg->refcnt);
5190 static void __mem_cgroup_put(struct mem_cgroup *memcg, int count)
5192 if (atomic_sub_and_test(count, &memcg->refcnt)) {
5193 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5194 __mem_cgroup_free(memcg);
5196 mem_cgroup_put(parent);
5200 static void mem_cgroup_put(struct mem_cgroup *memcg)
5202 __mem_cgroup_put(memcg, 1);
5206 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
5208 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
5210 if (!memcg->res.parent)
5212 return mem_cgroup_from_res_counter(memcg->res.parent, res);
5214 EXPORT_SYMBOL(parent_mem_cgroup);
5216 #ifdef CONFIG_MEMCG_SWAP
5217 static void __init enable_swap_cgroup(void)
5219 if (!mem_cgroup_disabled() && really_do_swap_account)
5220 do_swap_account = 1;
5223 static void __init enable_swap_cgroup(void)
5228 static int mem_cgroup_soft_limit_tree_init(void)
5230 struct mem_cgroup_tree_per_node *rtpn;
5231 struct mem_cgroup_tree_per_zone *rtpz;
5232 int tmp, node, zone;
5234 for_each_node(node) {
5236 if (!node_state(node, N_NORMAL_MEMORY))
5238 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
5242 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5244 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5245 rtpz = &rtpn->rb_tree_per_zone[zone];
5246 rtpz->rb_root = RB_ROOT;
5247 spin_lock_init(&rtpz->lock);
5253 for_each_node(node) {
5254 if (!soft_limit_tree.rb_tree_per_node[node])
5256 kfree(soft_limit_tree.rb_tree_per_node[node]);
5257 soft_limit_tree.rb_tree_per_node[node] = NULL;
5263 static struct cgroup_subsys_state * __ref
5264 mem_cgroup_css_alloc(struct cgroup *cont)
5266 struct mem_cgroup *memcg, *parent;
5267 long error = -ENOMEM;
5270 memcg = mem_cgroup_alloc();
5272 return ERR_PTR(error);
5275 if (alloc_mem_cgroup_per_zone_info(memcg, node))
5279 if (cont->parent == NULL) {
5281 enable_swap_cgroup();
5283 if (mem_cgroup_soft_limit_tree_init())
5285 root_mem_cgroup = memcg;
5286 for_each_possible_cpu(cpu) {
5287 struct memcg_stock_pcp *stock =
5288 &per_cpu(memcg_stock, cpu);
5289 INIT_WORK(&stock->work, drain_local_stock);
5291 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5293 parent = mem_cgroup_from_cont(cont->parent);
5294 memcg->use_hierarchy = parent->use_hierarchy;
5295 memcg->oom_kill_disable = parent->oom_kill_disable;
5298 if (parent && parent->use_hierarchy) {
5299 res_counter_init(&memcg->res, &parent->res);
5300 res_counter_init(&memcg->memsw, &parent->memsw);
5301 res_counter_init(&memcg->kmem, &parent->kmem);
5303 * We increment refcnt of the parent to ensure that we can
5304 * safely access it on res_counter_charge/uncharge.
5305 * This refcnt will be decremented when freeing this
5306 * mem_cgroup(see mem_cgroup_put).
5308 mem_cgroup_get(parent);
5310 res_counter_init(&memcg->res, NULL);
5311 res_counter_init(&memcg->memsw, NULL);
5312 res_counter_init(&memcg->kmem, NULL);
5314 * Deeper hierachy with use_hierarchy == false doesn't make
5315 * much sense so let cgroup subsystem know about this
5316 * unfortunate state in our controller.
5318 if (parent && parent != root_mem_cgroup)
5319 mem_cgroup_subsys.broken_hierarchy = true;
5321 memcg->last_scanned_node = MAX_NUMNODES;
5322 INIT_LIST_HEAD(&memcg->oom_notify);
5325 memcg->swappiness = mem_cgroup_swappiness(parent);
5326 atomic_set(&memcg->refcnt, 1);
5327 memcg->move_charge_at_immigrate = 0;
5328 mutex_init(&memcg->thresholds_lock);
5329 spin_lock_init(&memcg->move_lock);
5331 error = memcg_init_kmem(memcg, &mem_cgroup_subsys);
5334 * We call put now because our (and parent's) refcnts
5335 * are already in place. mem_cgroup_put() will internally
5336 * call __mem_cgroup_free, so return directly
5338 mem_cgroup_put(memcg);
5339 return ERR_PTR(error);
5343 __mem_cgroup_free(memcg);
5344 return ERR_PTR(error);
5347 static void mem_cgroup_css_offline(struct cgroup *cont)
5349 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5351 mem_cgroup_reparent_charges(memcg);
5354 static void mem_cgroup_css_free(struct cgroup *cont)
5356 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5358 kmem_cgroup_destroy(memcg);
5360 mem_cgroup_put(memcg);
5364 /* Handlers for move charge at task migration. */
5365 #define PRECHARGE_COUNT_AT_ONCE 256
5366 static int mem_cgroup_do_precharge(unsigned long count)
5369 int batch_count = PRECHARGE_COUNT_AT_ONCE;
5370 struct mem_cgroup *memcg = mc.to;
5372 if (mem_cgroup_is_root(memcg)) {
5373 mc.precharge += count;
5374 /* we don't need css_get for root */
5377 /* try to charge at once */
5379 struct res_counter *dummy;
5381 * "memcg" cannot be under rmdir() because we've already checked
5382 * by cgroup_lock_live_cgroup() that it is not removed and we
5383 * are still under the same cgroup_mutex. So we can postpone
5386 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
5388 if (do_swap_account && res_counter_charge(&memcg->memsw,
5389 PAGE_SIZE * count, &dummy)) {
5390 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
5393 mc.precharge += count;
5397 /* fall back to one by one charge */
5399 if (signal_pending(current)) {
5403 if (!batch_count--) {
5404 batch_count = PRECHARGE_COUNT_AT_ONCE;
5407 ret = __mem_cgroup_try_charge(NULL,
5408 GFP_KERNEL, 1, &memcg, false);
5410 /* mem_cgroup_clear_mc() will do uncharge later */
5418 * get_mctgt_type - get target type of moving charge
5419 * @vma: the vma the pte to be checked belongs
5420 * @addr: the address corresponding to the pte to be checked
5421 * @ptent: the pte to be checked
5422 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5425 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5426 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5427 * move charge. if @target is not NULL, the page is stored in target->page
5428 * with extra refcnt got(Callers should handle it).
5429 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5430 * target for charge migration. if @target is not NULL, the entry is stored
5433 * Called with pte lock held.
5440 enum mc_target_type {
5446 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5447 unsigned long addr, pte_t ptent)
5449 struct page *page = vm_normal_page(vma, addr, ptent);
5451 if (!page || !page_mapped(page))
5453 if (PageAnon(page)) {
5454 /* we don't move shared anon */
5457 } else if (!move_file())
5458 /* we ignore mapcount for file pages */
5460 if (!get_page_unless_zero(page))
5467 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5468 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5470 struct page *page = NULL;
5471 swp_entry_t ent = pte_to_swp_entry(ptent);
5473 if (!move_anon() || non_swap_entry(ent))
5476 * Because lookup_swap_cache() updates some statistics counter,
5477 * we call find_get_page() with swapper_space directly.
5479 page = find_get_page(&swapper_space, ent.val);
5480 if (do_swap_account)
5481 entry->val = ent.val;
5486 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5487 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5493 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5494 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5496 struct page *page = NULL;
5497 struct address_space *mapping;
5500 if (!vma->vm_file) /* anonymous vma */
5505 mapping = vma->vm_file->f_mapping;
5506 if (pte_none(ptent))
5507 pgoff = linear_page_index(vma, addr);
5508 else /* pte_file(ptent) is true */
5509 pgoff = pte_to_pgoff(ptent);
5511 /* page is moved even if it's not RSS of this task(page-faulted). */
5512 page = find_get_page(mapping, pgoff);
5515 /* shmem/tmpfs may report page out on swap: account for that too. */
5516 if (radix_tree_exceptional_entry(page)) {
5517 swp_entry_t swap = radix_to_swp_entry(page);
5518 if (do_swap_account)
5520 page = find_get_page(&swapper_space, swap.val);
5526 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5527 unsigned long addr, pte_t ptent, union mc_target *target)
5529 struct page *page = NULL;
5530 struct page_cgroup *pc;
5531 enum mc_target_type ret = MC_TARGET_NONE;
5532 swp_entry_t ent = { .val = 0 };
5534 if (pte_present(ptent))
5535 page = mc_handle_present_pte(vma, addr, ptent);
5536 else if (is_swap_pte(ptent))
5537 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5538 else if (pte_none(ptent) || pte_file(ptent))
5539 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5541 if (!page && !ent.val)
5544 pc = lookup_page_cgroup(page);
5546 * Do only loose check w/o page_cgroup lock.
5547 * mem_cgroup_move_account() checks the pc is valid or not under
5550 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5551 ret = MC_TARGET_PAGE;
5553 target->page = page;
5555 if (!ret || !target)
5558 /* There is a swap entry and a page doesn't exist or isn't charged */
5559 if (ent.val && !ret &&
5560 css_id(&mc.from->css) == lookup_swap_cgroup_id(ent)) {
5561 ret = MC_TARGET_SWAP;
5568 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5570 * We don't consider swapping or file mapped pages because THP does not
5571 * support them for now.
5572 * Caller should make sure that pmd_trans_huge(pmd) is true.
5574 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5575 unsigned long addr, pmd_t pmd, union mc_target *target)
5577 struct page *page = NULL;
5578 struct page_cgroup *pc;
5579 enum mc_target_type ret = MC_TARGET_NONE;
5581 page = pmd_page(pmd);
5582 VM_BUG_ON(!page || !PageHead(page));
5585 pc = lookup_page_cgroup(page);
5586 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5587 ret = MC_TARGET_PAGE;
5590 target->page = page;
5596 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5597 unsigned long addr, pmd_t pmd, union mc_target *target)
5599 return MC_TARGET_NONE;
5603 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5604 unsigned long addr, unsigned long end,
5605 struct mm_walk *walk)
5607 struct vm_area_struct *vma = walk->private;
5611 if (pmd_trans_huge_lock(pmd, vma) == 1) {
5612 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5613 mc.precharge += HPAGE_PMD_NR;
5614 spin_unlock(&vma->vm_mm->page_table_lock);
5618 if (pmd_trans_unstable(pmd))
5620 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5621 for (; addr != end; pte++, addr += PAGE_SIZE)
5622 if (get_mctgt_type(vma, addr, *pte, NULL))
5623 mc.precharge++; /* increment precharge temporarily */
5624 pte_unmap_unlock(pte - 1, ptl);
5630 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5632 unsigned long precharge;
5633 struct vm_area_struct *vma;
5635 down_read(&mm->mmap_sem);
5636 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5637 struct mm_walk mem_cgroup_count_precharge_walk = {
5638 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5642 if (is_vm_hugetlb_page(vma))
5644 walk_page_range(vma->vm_start, vma->vm_end,
5645 &mem_cgroup_count_precharge_walk);
5647 up_read(&mm->mmap_sem);
5649 precharge = mc.precharge;
5655 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5657 unsigned long precharge = mem_cgroup_count_precharge(mm);
5659 VM_BUG_ON(mc.moving_task);
5660 mc.moving_task = current;
5661 return mem_cgroup_do_precharge(precharge);
5664 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5665 static void __mem_cgroup_clear_mc(void)
5667 struct mem_cgroup *from = mc.from;
5668 struct mem_cgroup *to = mc.to;
5670 /* we must uncharge all the leftover precharges from mc.to */
5672 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
5676 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5677 * we must uncharge here.
5679 if (mc.moved_charge) {
5680 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
5681 mc.moved_charge = 0;
5683 /* we must fixup refcnts and charges */
5684 if (mc.moved_swap) {
5685 /* uncharge swap account from the old cgroup */
5686 if (!mem_cgroup_is_root(mc.from))
5687 res_counter_uncharge(&mc.from->memsw,
5688 PAGE_SIZE * mc.moved_swap);
5689 __mem_cgroup_put(mc.from, mc.moved_swap);
5691 if (!mem_cgroup_is_root(mc.to)) {
5693 * we charged both to->res and to->memsw, so we should
5696 res_counter_uncharge(&mc.to->res,
5697 PAGE_SIZE * mc.moved_swap);
5699 /* we've already done mem_cgroup_get(mc.to) */
5702 memcg_oom_recover(from);
5703 memcg_oom_recover(to);
5704 wake_up_all(&mc.waitq);
5707 static void mem_cgroup_clear_mc(void)
5709 struct mem_cgroup *from = mc.from;
5712 * we must clear moving_task before waking up waiters at the end of
5715 mc.moving_task = NULL;
5716 __mem_cgroup_clear_mc();
5717 spin_lock(&mc.lock);
5720 spin_unlock(&mc.lock);
5721 mem_cgroup_end_move(from);
5724 static int mem_cgroup_can_attach(struct cgroup *cgroup,
5725 struct cgroup_taskset *tset)
5727 struct task_struct *p = cgroup_taskset_first(tset);
5729 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup);
5731 if (memcg->move_charge_at_immigrate) {
5732 struct mm_struct *mm;
5733 struct mem_cgroup *from = mem_cgroup_from_task(p);
5735 VM_BUG_ON(from == memcg);
5737 mm = get_task_mm(p);
5740 /* We move charges only when we move a owner of the mm */
5741 if (mm->owner == p) {
5744 VM_BUG_ON(mc.precharge);
5745 VM_BUG_ON(mc.moved_charge);
5746 VM_BUG_ON(mc.moved_swap);
5747 mem_cgroup_start_move(from);
5748 spin_lock(&mc.lock);
5751 spin_unlock(&mc.lock);
5752 /* We set mc.moving_task later */
5754 ret = mem_cgroup_precharge_mc(mm);
5756 mem_cgroup_clear_mc();
5763 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
5764 struct cgroup_taskset *tset)
5766 mem_cgroup_clear_mc();
5769 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5770 unsigned long addr, unsigned long end,
5771 struct mm_walk *walk)
5774 struct vm_area_struct *vma = walk->private;
5777 enum mc_target_type target_type;
5778 union mc_target target;
5780 struct page_cgroup *pc;
5783 * We don't take compound_lock() here but no race with splitting thp
5785 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
5786 * under splitting, which means there's no concurrent thp split,
5787 * - if another thread runs into split_huge_page() just after we
5788 * entered this if-block, the thread must wait for page table lock
5789 * to be unlocked in __split_huge_page_splitting(), where the main
5790 * part of thp split is not executed yet.
5792 if (pmd_trans_huge_lock(pmd, vma) == 1) {
5793 if (mc.precharge < HPAGE_PMD_NR) {
5794 spin_unlock(&vma->vm_mm->page_table_lock);
5797 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5798 if (target_type == MC_TARGET_PAGE) {
5800 if (!isolate_lru_page(page)) {
5801 pc = lookup_page_cgroup(page);
5802 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
5803 pc, mc.from, mc.to)) {
5804 mc.precharge -= HPAGE_PMD_NR;
5805 mc.moved_charge += HPAGE_PMD_NR;
5807 putback_lru_page(page);
5811 spin_unlock(&vma->vm_mm->page_table_lock);
5815 if (pmd_trans_unstable(pmd))
5818 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5819 for (; addr != end; addr += PAGE_SIZE) {
5820 pte_t ptent = *(pte++);
5826 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5827 case MC_TARGET_PAGE:
5829 if (isolate_lru_page(page))
5831 pc = lookup_page_cgroup(page);
5832 if (!mem_cgroup_move_account(page, 1, pc,
5835 /* we uncharge from mc.from later. */
5838 putback_lru_page(page);
5839 put: /* get_mctgt_type() gets the page */
5842 case MC_TARGET_SWAP:
5844 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5846 /* we fixup refcnts and charges later. */
5854 pte_unmap_unlock(pte - 1, ptl);
5859 * We have consumed all precharges we got in can_attach().
5860 * We try charge one by one, but don't do any additional
5861 * charges to mc.to if we have failed in charge once in attach()
5864 ret = mem_cgroup_do_precharge(1);
5872 static void mem_cgroup_move_charge(struct mm_struct *mm)
5874 struct vm_area_struct *vma;
5876 lru_add_drain_all();
5878 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5880 * Someone who are holding the mmap_sem might be waiting in
5881 * waitq. So we cancel all extra charges, wake up all waiters,
5882 * and retry. Because we cancel precharges, we might not be able
5883 * to move enough charges, but moving charge is a best-effort
5884 * feature anyway, so it wouldn't be a big problem.
5886 __mem_cgroup_clear_mc();
5890 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5892 struct mm_walk mem_cgroup_move_charge_walk = {
5893 .pmd_entry = mem_cgroup_move_charge_pte_range,
5897 if (is_vm_hugetlb_page(vma))
5899 ret = walk_page_range(vma->vm_start, vma->vm_end,
5900 &mem_cgroup_move_charge_walk);
5903 * means we have consumed all precharges and failed in
5904 * doing additional charge. Just abandon here.
5908 up_read(&mm->mmap_sem);
5911 static void mem_cgroup_move_task(struct cgroup *cont,
5912 struct cgroup_taskset *tset)
5914 struct task_struct *p = cgroup_taskset_first(tset);
5915 struct mm_struct *mm = get_task_mm(p);
5919 mem_cgroup_move_charge(mm);
5923 mem_cgroup_clear_mc();
5925 #else /* !CONFIG_MMU */
5926 static int mem_cgroup_can_attach(struct cgroup *cgroup,
5927 struct cgroup_taskset *tset)
5931 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
5932 struct cgroup_taskset *tset)
5935 static void mem_cgroup_move_task(struct cgroup *cont,
5936 struct cgroup_taskset *tset)
5941 struct cgroup_subsys mem_cgroup_subsys = {
5943 .subsys_id = mem_cgroup_subsys_id,
5944 .css_alloc = mem_cgroup_css_alloc,
5945 .css_offline = mem_cgroup_css_offline,
5946 .css_free = mem_cgroup_css_free,
5947 .can_attach = mem_cgroup_can_attach,
5948 .cancel_attach = mem_cgroup_cancel_attach,
5949 .attach = mem_cgroup_move_task,
5950 .base_cftypes = mem_cgroup_files,
5955 #ifdef CONFIG_MEMCG_SWAP
5956 static int __init enable_swap_account(char *s)
5958 /* consider enabled if no parameter or 1 is given */
5959 if (!strcmp(s, "1"))
5960 really_do_swap_account = 1;
5961 else if (!strcmp(s, "0"))
5962 really_do_swap_account = 0;
5965 __setup("swapaccount=", enable_swap_account);