1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * This program is free software; you can redistribute it and/or modify
14 * it under the terms of the GNU General Public License as published by
15 * the Free Software Foundation; either version 2 of the License, or
16 * (at your option) any later version.
18 * This program is distributed in the hope that it will be useful,
19 * but WITHOUT ANY WARRANTY; without even the implied warranty of
20 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
21 * GNU General Public License for more details.
24 #include <linux/res_counter.h>
25 #include <linux/memcontrol.h>
26 #include <linux/cgroup.h>
28 #include <linux/hugetlb.h>
29 #include <linux/pagemap.h>
30 #include <linux/smp.h>
31 #include <linux/page-flags.h>
32 #include <linux/backing-dev.h>
33 #include <linux/bit_spinlock.h>
34 #include <linux/rcupdate.h>
35 #include <linux/limits.h>
36 #include <linux/export.h>
37 #include <linux/mutex.h>
38 #include <linux/rbtree.h>
39 #include <linux/slab.h>
40 #include <linux/swap.h>
41 #include <linux/swapops.h>
42 #include <linux/spinlock.h>
43 #include <linux/eventfd.h>
44 #include <linux/sort.h>
46 #include <linux/seq_file.h>
47 #include <linux/vmalloc.h>
48 #include <linux/mm_inline.h>
49 #include <linux/page_cgroup.h>
50 #include <linux/cpu.h>
51 #include <linux/oom.h>
54 #include <net/tcp_memcontrol.h>
56 #include <asm/uaccess.h>
58 #include <trace/events/vmscan.h>
60 struct cgroup_subsys mem_cgroup_subsys __read_mostly;
61 #define MEM_CGROUP_RECLAIM_RETRIES 5
62 static struct mem_cgroup *root_mem_cgroup __read_mostly;
64 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
65 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
66 int do_swap_account __read_mostly;
68 /* for remember boot option*/
69 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP_ENABLED
70 static int really_do_swap_account __initdata = 1;
72 static int really_do_swap_account __initdata = 0;
76 #define do_swap_account (0)
81 * Statistics for memory cgroup.
83 enum mem_cgroup_stat_index {
85 * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
87 MEM_CGROUP_STAT_CACHE, /* # of pages charged as cache */
88 MEM_CGROUP_STAT_RSS, /* # of pages charged as anon rss */
89 MEM_CGROUP_STAT_FILE_MAPPED, /* # of pages charged as file rss */
90 MEM_CGROUP_STAT_SWAPOUT, /* # of pages, swapped out */
91 MEM_CGROUP_STAT_DATA, /* end of data requires synchronization */
92 MEM_CGROUP_STAT_NSTATS,
95 enum mem_cgroup_events_index {
96 MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */
97 MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */
98 MEM_CGROUP_EVENTS_COUNT, /* # of pages paged in/out */
99 MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */
100 MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */
101 MEM_CGROUP_EVENTS_NSTATS,
104 * Per memcg event counter is incremented at every pagein/pageout. With THP,
105 * it will be incremated by the number of pages. This counter is used for
106 * for trigger some periodic events. This is straightforward and better
107 * than using jiffies etc. to handle periodic memcg event.
109 enum mem_cgroup_events_target {
110 MEM_CGROUP_TARGET_THRESH,
111 MEM_CGROUP_TARGET_SOFTLIMIT,
112 MEM_CGROUP_TARGET_NUMAINFO,
115 #define THRESHOLDS_EVENTS_TARGET (128)
116 #define SOFTLIMIT_EVENTS_TARGET (1024)
117 #define NUMAINFO_EVENTS_TARGET (1024)
119 struct mem_cgroup_stat_cpu {
120 long count[MEM_CGROUP_STAT_NSTATS];
121 unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
122 unsigned long targets[MEM_CGROUP_NTARGETS];
125 struct mem_cgroup_reclaim_iter {
126 /* css_id of the last scanned hierarchy member */
128 /* scan generation, increased every round-trip */
129 unsigned int generation;
133 * per-zone information in memory controller.
135 struct mem_cgroup_per_zone {
136 struct lruvec lruvec;
137 unsigned long lru_size[NR_LRU_LISTS];
139 struct mem_cgroup_reclaim_iter reclaim_iter[DEF_PRIORITY + 1];
141 struct rb_node tree_node; /* RB tree node */
142 unsigned long long usage_in_excess;/* Set to the value by which */
143 /* the soft limit is exceeded*/
145 struct mem_cgroup *memcg; /* Back pointer, we cannot */
146 /* use container_of */
149 struct mem_cgroup_per_node {
150 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
153 struct mem_cgroup_lru_info {
154 struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
158 * Cgroups above their limits are maintained in a RB-Tree, independent of
159 * their hierarchy representation
162 struct mem_cgroup_tree_per_zone {
163 struct rb_root rb_root;
167 struct mem_cgroup_tree_per_node {
168 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
171 struct mem_cgroup_tree {
172 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
175 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
177 struct mem_cgroup_threshold {
178 struct eventfd_ctx *eventfd;
183 struct mem_cgroup_threshold_ary {
184 /* An array index points to threshold just below usage. */
185 int current_threshold;
186 /* Size of entries[] */
188 /* Array of thresholds */
189 struct mem_cgroup_threshold entries[0];
192 struct mem_cgroup_thresholds {
193 /* Primary thresholds array */
194 struct mem_cgroup_threshold_ary *primary;
196 * Spare threshold array.
197 * This is needed to make mem_cgroup_unregister_event() "never fail".
198 * It must be able to store at least primary->size - 1 entries.
200 struct mem_cgroup_threshold_ary *spare;
204 struct mem_cgroup_eventfd_list {
205 struct list_head list;
206 struct eventfd_ctx *eventfd;
209 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
210 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
213 * The memory controller data structure. The memory controller controls both
214 * page cache and RSS per cgroup. We would eventually like to provide
215 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
216 * to help the administrator determine what knobs to tune.
218 * TODO: Add a water mark for the memory controller. Reclaim will begin when
219 * we hit the water mark. May be even add a low water mark, such that
220 * no reclaim occurs from a cgroup at it's low water mark, this is
221 * a feature that will be implemented much later in the future.
224 struct cgroup_subsys_state css;
226 * the counter to account for memory usage
228 struct res_counter res;
232 * the counter to account for mem+swap usage.
234 struct res_counter memsw;
237 * rcu_freeing is used only when freeing struct mem_cgroup,
238 * so put it into a union to avoid wasting more memory.
239 * It must be disjoint from the css field. It could be
240 * in a union with the res field, but res plays a much
241 * larger part in mem_cgroup life than memsw, and might
242 * be of interest, even at time of free, when debugging.
243 * So share rcu_head with the less interesting memsw.
245 struct rcu_head rcu_freeing;
247 * But when using vfree(), that cannot be done at
248 * interrupt time, so we must then queue the work.
250 struct work_struct work_freeing;
254 * Per cgroup active and inactive list, similar to the
255 * per zone LRU lists.
257 struct mem_cgroup_lru_info info;
258 int last_scanned_node;
260 nodemask_t scan_nodes;
261 atomic_t numainfo_events;
262 atomic_t numainfo_updating;
265 * Should the accounting and control be hierarchical, per subtree?
275 /* OOM-Killer disable */
276 int oom_kill_disable;
278 /* set when res.limit == memsw.limit */
279 bool memsw_is_minimum;
281 /* protect arrays of thresholds */
282 struct mutex thresholds_lock;
284 /* thresholds for memory usage. RCU-protected */
285 struct mem_cgroup_thresholds thresholds;
287 /* thresholds for mem+swap usage. RCU-protected */
288 struct mem_cgroup_thresholds memsw_thresholds;
290 /* For oom notifier event fd */
291 struct list_head oom_notify;
294 * Should we move charges of a task when a task is moved into this
295 * mem_cgroup ? And what type of charges should we move ?
297 unsigned long move_charge_at_immigrate;
299 * set > 0 if pages under this cgroup are moving to other cgroup.
301 atomic_t moving_account;
302 /* taken only while moving_account > 0 */
303 spinlock_t move_lock;
307 struct mem_cgroup_stat_cpu __percpu *stat;
309 * used when a cpu is offlined or other synchronizations
310 * See mem_cgroup_read_stat().
312 struct mem_cgroup_stat_cpu nocpu_base;
313 spinlock_t pcp_counter_lock;
316 struct tcp_memcontrol tcp_mem;
320 /* Stuffs for move charges at task migration. */
322 * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
323 * left-shifted bitmap of these types.
326 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
327 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
331 /* "mc" and its members are protected by cgroup_mutex */
332 static struct move_charge_struct {
333 spinlock_t lock; /* for from, to */
334 struct mem_cgroup *from;
335 struct mem_cgroup *to;
336 unsigned long precharge;
337 unsigned long moved_charge;
338 unsigned long moved_swap;
339 struct task_struct *moving_task; /* a task moving charges */
340 wait_queue_head_t waitq; /* a waitq for other context */
342 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
343 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
346 static bool move_anon(void)
348 return test_bit(MOVE_CHARGE_TYPE_ANON,
349 &mc.to->move_charge_at_immigrate);
352 static bool move_file(void)
354 return test_bit(MOVE_CHARGE_TYPE_FILE,
355 &mc.to->move_charge_at_immigrate);
359 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
360 * limit reclaim to prevent infinite loops, if they ever occur.
362 #define MEM_CGROUP_MAX_RECLAIM_LOOPS (100)
363 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2)
366 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
367 MEM_CGROUP_CHARGE_TYPE_MAPPED,
368 MEM_CGROUP_CHARGE_TYPE_SHMEM, /* used by page migration of shmem */
369 MEM_CGROUP_CHARGE_TYPE_FORCE, /* used by force_empty */
370 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
371 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
375 /* for encoding cft->private value on file */
378 #define _OOM_TYPE (2)
379 #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
380 #define MEMFILE_TYPE(val) (((val) >> 16) & 0xffff)
381 #define MEMFILE_ATTR(val) ((val) & 0xffff)
382 /* Used for OOM nofiier */
383 #define OOM_CONTROL (0)
386 * Reclaim flags for mem_cgroup_hierarchical_reclaim
388 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
389 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
390 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
391 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
393 static void mem_cgroup_get(struct mem_cgroup *memcg);
394 static void mem_cgroup_put(struct mem_cgroup *memcg);
396 /* Writing them here to avoid exposing memcg's inner layout */
397 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_KMEM
398 #include <net/sock.h>
401 static bool mem_cgroup_is_root(struct mem_cgroup *memcg);
402 void sock_update_memcg(struct sock *sk)
404 if (mem_cgroup_sockets_enabled) {
405 struct mem_cgroup *memcg;
407 BUG_ON(!sk->sk_prot->proto_cgroup);
409 /* Socket cloning can throw us here with sk_cgrp already
410 * filled. It won't however, necessarily happen from
411 * process context. So the test for root memcg given
412 * the current task's memcg won't help us in this case.
414 * Respecting the original socket's memcg is a better
415 * decision in this case.
418 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
419 mem_cgroup_get(sk->sk_cgrp->memcg);
424 memcg = mem_cgroup_from_task(current);
425 if (!mem_cgroup_is_root(memcg)) {
426 mem_cgroup_get(memcg);
427 sk->sk_cgrp = sk->sk_prot->proto_cgroup(memcg);
432 EXPORT_SYMBOL(sock_update_memcg);
434 void sock_release_memcg(struct sock *sk)
436 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
437 struct mem_cgroup *memcg;
438 WARN_ON(!sk->sk_cgrp->memcg);
439 memcg = sk->sk_cgrp->memcg;
440 mem_cgroup_put(memcg);
445 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
447 if (!memcg || mem_cgroup_is_root(memcg))
450 return &memcg->tcp_mem.cg_proto;
452 EXPORT_SYMBOL(tcp_proto_cgroup);
453 #endif /* CONFIG_INET */
454 #endif /* CONFIG_CGROUP_MEM_RES_CTLR_KMEM */
456 static void drain_all_stock_async(struct mem_cgroup *memcg);
458 static struct mem_cgroup_per_zone *
459 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
461 return &memcg->info.nodeinfo[nid]->zoneinfo[zid];
464 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
469 static struct mem_cgroup_per_zone *
470 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
472 int nid = page_to_nid(page);
473 int zid = page_zonenum(page);
475 return mem_cgroup_zoneinfo(memcg, nid, zid);
478 static struct mem_cgroup_tree_per_zone *
479 soft_limit_tree_node_zone(int nid, int zid)
481 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
484 static struct mem_cgroup_tree_per_zone *
485 soft_limit_tree_from_page(struct page *page)
487 int nid = page_to_nid(page);
488 int zid = page_zonenum(page);
490 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
494 __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
495 struct mem_cgroup_per_zone *mz,
496 struct mem_cgroup_tree_per_zone *mctz,
497 unsigned long long new_usage_in_excess)
499 struct rb_node **p = &mctz->rb_root.rb_node;
500 struct rb_node *parent = NULL;
501 struct mem_cgroup_per_zone *mz_node;
506 mz->usage_in_excess = new_usage_in_excess;
507 if (!mz->usage_in_excess)
511 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
513 if (mz->usage_in_excess < mz_node->usage_in_excess)
516 * We can't avoid mem cgroups that are over their soft
517 * limit by the same amount
519 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
522 rb_link_node(&mz->tree_node, parent, p);
523 rb_insert_color(&mz->tree_node, &mctz->rb_root);
528 __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
529 struct mem_cgroup_per_zone *mz,
530 struct mem_cgroup_tree_per_zone *mctz)
534 rb_erase(&mz->tree_node, &mctz->rb_root);
539 mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
540 struct mem_cgroup_per_zone *mz,
541 struct mem_cgroup_tree_per_zone *mctz)
543 spin_lock(&mctz->lock);
544 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
545 spin_unlock(&mctz->lock);
549 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
551 unsigned long long excess;
552 struct mem_cgroup_per_zone *mz;
553 struct mem_cgroup_tree_per_zone *mctz;
554 int nid = page_to_nid(page);
555 int zid = page_zonenum(page);
556 mctz = soft_limit_tree_from_page(page);
559 * Necessary to update all ancestors when hierarchy is used.
560 * because their event counter is not touched.
562 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
563 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
564 excess = res_counter_soft_limit_excess(&memcg->res);
566 * We have to update the tree if mz is on RB-tree or
567 * mem is over its softlimit.
569 if (excess || mz->on_tree) {
570 spin_lock(&mctz->lock);
571 /* if on-tree, remove it */
573 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
575 * Insert again. mz->usage_in_excess will be updated.
576 * If excess is 0, no tree ops.
578 __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
579 spin_unlock(&mctz->lock);
584 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
587 struct mem_cgroup_per_zone *mz;
588 struct mem_cgroup_tree_per_zone *mctz;
590 for_each_node(node) {
591 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
592 mz = mem_cgroup_zoneinfo(memcg, node, zone);
593 mctz = soft_limit_tree_node_zone(node, zone);
594 mem_cgroup_remove_exceeded(memcg, mz, mctz);
599 static struct mem_cgroup_per_zone *
600 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
602 struct rb_node *rightmost = NULL;
603 struct mem_cgroup_per_zone *mz;
607 rightmost = rb_last(&mctz->rb_root);
609 goto done; /* Nothing to reclaim from */
611 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
613 * Remove the node now but someone else can add it back,
614 * we will to add it back at the end of reclaim to its correct
615 * position in the tree.
617 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
618 if (!res_counter_soft_limit_excess(&mz->memcg->res) ||
619 !css_tryget(&mz->memcg->css))
625 static struct mem_cgroup_per_zone *
626 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
628 struct mem_cgroup_per_zone *mz;
630 spin_lock(&mctz->lock);
631 mz = __mem_cgroup_largest_soft_limit_node(mctz);
632 spin_unlock(&mctz->lock);
637 * Implementation Note: reading percpu statistics for memcg.
639 * Both of vmstat[] and percpu_counter has threshold and do periodic
640 * synchronization to implement "quick" read. There are trade-off between
641 * reading cost and precision of value. Then, we may have a chance to implement
642 * a periodic synchronizion of counter in memcg's counter.
644 * But this _read() function is used for user interface now. The user accounts
645 * memory usage by memory cgroup and he _always_ requires exact value because
646 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
647 * have to visit all online cpus and make sum. So, for now, unnecessary
648 * synchronization is not implemented. (just implemented for cpu hotplug)
650 * If there are kernel internal actions which can make use of some not-exact
651 * value, and reading all cpu value can be performance bottleneck in some
652 * common workload, threashold and synchonization as vmstat[] should be
655 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
656 enum mem_cgroup_stat_index idx)
662 for_each_online_cpu(cpu)
663 val += per_cpu(memcg->stat->count[idx], cpu);
664 #ifdef CONFIG_HOTPLUG_CPU
665 spin_lock(&memcg->pcp_counter_lock);
666 val += memcg->nocpu_base.count[idx];
667 spin_unlock(&memcg->pcp_counter_lock);
673 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
676 int val = (charge) ? 1 : -1;
677 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAPOUT], val);
680 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
681 enum mem_cgroup_events_index idx)
683 unsigned long val = 0;
686 for_each_online_cpu(cpu)
687 val += per_cpu(memcg->stat->events[idx], cpu);
688 #ifdef CONFIG_HOTPLUG_CPU
689 spin_lock(&memcg->pcp_counter_lock);
690 val += memcg->nocpu_base.events[idx];
691 spin_unlock(&memcg->pcp_counter_lock);
696 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
697 bool anon, int nr_pages)
702 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
703 * counted as CACHE even if it's on ANON LRU.
706 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
709 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
712 /* pagein of a big page is an event. So, ignore page size */
714 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
716 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
717 nr_pages = -nr_pages; /* for event */
720 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT], nr_pages);
726 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
727 unsigned int lru_mask)
729 struct mem_cgroup_per_zone *mz;
731 unsigned long ret = 0;
733 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
736 if (BIT(lru) & lru_mask)
737 ret += mz->lru_size[lru];
743 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
744 int nid, unsigned int lru_mask)
749 for (zid = 0; zid < MAX_NR_ZONES; zid++)
750 total += mem_cgroup_zone_nr_lru_pages(memcg,
756 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
757 unsigned int lru_mask)
762 for_each_node_state(nid, N_HIGH_MEMORY)
763 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
767 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
768 enum mem_cgroup_events_target target)
770 unsigned long val, next;
772 val = __this_cpu_read(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT]);
773 next = __this_cpu_read(memcg->stat->targets[target]);
774 /* from time_after() in jiffies.h */
775 if ((long)next - (long)val < 0) {
777 case MEM_CGROUP_TARGET_THRESH:
778 next = val + THRESHOLDS_EVENTS_TARGET;
780 case MEM_CGROUP_TARGET_SOFTLIMIT:
781 next = val + SOFTLIMIT_EVENTS_TARGET;
783 case MEM_CGROUP_TARGET_NUMAINFO:
784 next = val + NUMAINFO_EVENTS_TARGET;
789 __this_cpu_write(memcg->stat->targets[target], next);
796 * Check events in order.
799 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
802 /* threshold event is triggered in finer grain than soft limit */
803 if (unlikely(mem_cgroup_event_ratelimit(memcg,
804 MEM_CGROUP_TARGET_THRESH))) {
806 bool do_numainfo __maybe_unused;
808 do_softlimit = mem_cgroup_event_ratelimit(memcg,
809 MEM_CGROUP_TARGET_SOFTLIMIT);
811 do_numainfo = mem_cgroup_event_ratelimit(memcg,
812 MEM_CGROUP_TARGET_NUMAINFO);
816 mem_cgroup_threshold(memcg);
817 if (unlikely(do_softlimit))
818 mem_cgroup_update_tree(memcg, page);
820 if (unlikely(do_numainfo))
821 atomic_inc(&memcg->numainfo_events);
827 struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
829 return container_of(cgroup_subsys_state(cont,
830 mem_cgroup_subsys_id), struct mem_cgroup,
834 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
837 * mm_update_next_owner() may clear mm->owner to NULL
838 * if it races with swapoff, page migration, etc.
839 * So this can be called with p == NULL.
844 return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
845 struct mem_cgroup, css);
848 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
850 struct mem_cgroup *memcg = NULL;
855 * Because we have no locks, mm->owner's may be being moved to other
856 * cgroup. We use css_tryget() here even if this looks
857 * pessimistic (rather than adding locks here).
861 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
862 if (unlikely(!memcg))
864 } while (!css_tryget(&memcg->css));
870 * mem_cgroup_iter - iterate over memory cgroup hierarchy
871 * @root: hierarchy root
872 * @prev: previously returned memcg, NULL on first invocation
873 * @reclaim: cookie for shared reclaim walks, NULL for full walks
875 * Returns references to children of the hierarchy below @root, or
876 * @root itself, or %NULL after a full round-trip.
878 * Caller must pass the return value in @prev on subsequent
879 * invocations for reference counting, or use mem_cgroup_iter_break()
880 * to cancel a hierarchy walk before the round-trip is complete.
882 * Reclaimers can specify a zone and a priority level in @reclaim to
883 * divide up the memcgs in the hierarchy among all concurrent
884 * reclaimers operating on the same zone and priority.
886 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
887 struct mem_cgroup *prev,
888 struct mem_cgroup_reclaim_cookie *reclaim)
890 struct mem_cgroup *memcg = NULL;
893 if (mem_cgroup_disabled())
897 root = root_mem_cgroup;
899 if (prev && !reclaim)
900 id = css_id(&prev->css);
902 if (prev && prev != root)
905 if (!root->use_hierarchy && root != root_mem_cgroup) {
912 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
913 struct cgroup_subsys_state *css;
916 int nid = zone_to_nid(reclaim->zone);
917 int zid = zone_idx(reclaim->zone);
918 struct mem_cgroup_per_zone *mz;
920 mz = mem_cgroup_zoneinfo(root, nid, zid);
921 iter = &mz->reclaim_iter[reclaim->priority];
922 if (prev && reclaim->generation != iter->generation)
928 css = css_get_next(&mem_cgroup_subsys, id + 1, &root->css, &id);
930 if (css == &root->css || css_tryget(css))
931 memcg = container_of(css,
932 struct mem_cgroup, css);
941 else if (!prev && memcg)
942 reclaim->generation = iter->generation;
952 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
953 * @root: hierarchy root
954 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
956 void mem_cgroup_iter_break(struct mem_cgroup *root,
957 struct mem_cgroup *prev)
960 root = root_mem_cgroup;
961 if (prev && prev != root)
966 * Iteration constructs for visiting all cgroups (under a tree). If
967 * loops are exited prematurely (break), mem_cgroup_iter_break() must
968 * be used for reference counting.
970 #define for_each_mem_cgroup_tree(iter, root) \
971 for (iter = mem_cgroup_iter(root, NULL, NULL); \
973 iter = mem_cgroup_iter(root, iter, NULL))
975 #define for_each_mem_cgroup(iter) \
976 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
978 iter = mem_cgroup_iter(NULL, iter, NULL))
980 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
982 return (memcg == root_mem_cgroup);
985 void mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
987 struct mem_cgroup *memcg;
993 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
994 if (unlikely(!memcg))
999 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1002 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1010 EXPORT_SYMBOL(mem_cgroup_count_vm_event);
1013 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1014 * @zone: zone of the wanted lruvec
1015 * @mem: memcg of the wanted lruvec
1017 * Returns the lru list vector holding pages for the given @zone and
1018 * @mem. This can be the global zone lruvec, if the memory controller
1021 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1022 struct mem_cgroup *memcg)
1024 struct mem_cgroup_per_zone *mz;
1026 if (mem_cgroup_disabled())
1027 return &zone->lruvec;
1029 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1034 * Following LRU functions are allowed to be used without PCG_LOCK.
1035 * Operations are called by routine of global LRU independently from memcg.
1036 * What we have to take care of here is validness of pc->mem_cgroup.
1038 * Changes to pc->mem_cgroup happens when
1041 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1042 * It is added to LRU before charge.
1043 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1044 * When moving account, the page is not on LRU. It's isolated.
1048 * mem_cgroup_lru_add_list - account for adding an lru page and return lruvec
1049 * @zone: zone of the page
1053 * This function accounts for @page being added to @lru, and returns
1054 * the lruvec for the given @zone and the memcg @page is charged to.
1056 * The callsite is then responsible for physically linking the page to
1057 * the returned lruvec->lists[@lru].
1059 struct lruvec *mem_cgroup_lru_add_list(struct zone *zone, struct page *page,
1062 struct mem_cgroup_per_zone *mz;
1063 struct mem_cgroup *memcg;
1064 struct page_cgroup *pc;
1066 if (mem_cgroup_disabled())
1067 return &zone->lruvec;
1069 pc = lookup_page_cgroup(page);
1070 memcg = pc->mem_cgroup;
1073 * Surreptitiously switch any uncharged page to root:
1074 * an uncharged page off lru does nothing to secure
1075 * its former mem_cgroup from sudden removal.
1077 * Our caller holds lru_lock, and PageCgroupUsed is updated
1078 * under page_cgroup lock: between them, they make all uses
1079 * of pc->mem_cgroup safe.
1081 if (!PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1082 pc->mem_cgroup = memcg = root_mem_cgroup;
1084 mz = page_cgroup_zoneinfo(memcg, page);
1085 /* compound_order() is stabilized through lru_lock */
1086 mz->lru_size[lru] += 1 << compound_order(page);
1091 * mem_cgroup_lru_del_list - account for removing an lru page
1095 * This function accounts for @page being removed from @lru.
1097 * The callsite is then responsible for physically unlinking
1100 void mem_cgroup_lru_del_list(struct page *page, enum lru_list lru)
1102 struct mem_cgroup_per_zone *mz;
1103 struct mem_cgroup *memcg;
1104 struct page_cgroup *pc;
1106 if (mem_cgroup_disabled())
1109 pc = lookup_page_cgroup(page);
1110 memcg = pc->mem_cgroup;
1112 mz = page_cgroup_zoneinfo(memcg, page);
1113 /* huge page split is done under lru_lock. so, we have no races. */
1114 VM_BUG_ON(mz->lru_size[lru] < (1 << compound_order(page)));
1115 mz->lru_size[lru] -= 1 << compound_order(page);
1119 * mem_cgroup_lru_move_lists - account for moving a page between lrus
1120 * @zone: zone of the page
1122 * @from: current lru
1125 * This function accounts for @page being moved between the lrus @from
1126 * and @to, and returns the lruvec for the given @zone and the memcg
1127 * @page is charged to.
1129 * The callsite is then responsible for physically relinking
1130 * @page->lru to the returned lruvec->lists[@to].
1132 struct lruvec *mem_cgroup_lru_move_lists(struct zone *zone,
1137 /* XXX: Optimize this, especially for @from == @to */
1138 mem_cgroup_lru_del_list(page, from);
1139 return mem_cgroup_lru_add_list(zone, page, to);
1143 * Checks whether given mem is same or in the root_mem_cgroup's
1146 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1147 struct mem_cgroup *memcg)
1149 if (root_memcg == memcg)
1151 if (!root_memcg->use_hierarchy)
1153 return css_is_ancestor(&memcg->css, &root_memcg->css);
1156 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1157 struct mem_cgroup *memcg)
1162 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1167 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *memcg)
1170 struct mem_cgroup *curr = NULL;
1171 struct task_struct *p;
1173 p = find_lock_task_mm(task);
1175 curr = try_get_mem_cgroup_from_mm(p->mm);
1179 * All threads may have already detached their mm's, but the oom
1180 * killer still needs to detect if they have already been oom
1181 * killed to prevent needlessly killing additional tasks.
1184 curr = mem_cgroup_from_task(task);
1186 css_get(&curr->css);
1192 * We should check use_hierarchy of "memcg" not "curr". Because checking
1193 * use_hierarchy of "curr" here make this function true if hierarchy is
1194 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1195 * hierarchy(even if use_hierarchy is disabled in "memcg").
1197 ret = mem_cgroup_same_or_subtree(memcg, curr);
1198 css_put(&curr->css);
1202 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg, struct zone *zone)
1204 unsigned long inactive_ratio;
1205 int nid = zone_to_nid(zone);
1206 int zid = zone_idx(zone);
1207 unsigned long inactive;
1208 unsigned long active;
1211 inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1212 BIT(LRU_INACTIVE_ANON));
1213 active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1214 BIT(LRU_ACTIVE_ANON));
1216 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1218 inactive_ratio = int_sqrt(10 * gb);
1222 return inactive * inactive_ratio < active;
1225 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg, struct zone *zone)
1227 unsigned long active;
1228 unsigned long inactive;
1229 int zid = zone_idx(zone);
1230 int nid = zone_to_nid(zone);
1232 inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1233 BIT(LRU_INACTIVE_FILE));
1234 active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1235 BIT(LRU_ACTIVE_FILE));
1237 return (active > inactive);
1240 struct zone_reclaim_stat *
1241 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
1243 struct page_cgroup *pc;
1244 struct mem_cgroup_per_zone *mz;
1246 if (mem_cgroup_disabled())
1249 pc = lookup_page_cgroup(page);
1250 if (!PageCgroupUsed(pc))
1252 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1254 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1255 return &mz->lruvec.reclaim_stat;
1258 #define mem_cgroup_from_res_counter(counter, member) \
1259 container_of(counter, struct mem_cgroup, member)
1262 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1263 * @mem: the memory cgroup
1265 * Returns the maximum amount of memory @mem can be charged with, in
1268 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1270 unsigned long long margin;
1272 margin = res_counter_margin(&memcg->res);
1273 if (do_swap_account)
1274 margin = min(margin, res_counter_margin(&memcg->memsw));
1275 return margin >> PAGE_SHIFT;
1278 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1280 struct cgroup *cgrp = memcg->css.cgroup;
1283 if (cgrp->parent == NULL)
1284 return vm_swappiness;
1286 return memcg->swappiness;
1290 * memcg->moving_account is used for checking possibility that some thread is
1291 * calling move_account(). When a thread on CPU-A starts moving pages under
1292 * a memcg, other threads should check memcg->moving_account under
1293 * rcu_read_lock(), like this:
1297 * memcg->moving_account+1 if (memcg->mocing_account)
1299 * synchronize_rcu() update something.
1304 /* for quick checking without looking up memcg */
1305 atomic_t memcg_moving __read_mostly;
1307 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1309 atomic_inc(&memcg_moving);
1310 atomic_inc(&memcg->moving_account);
1314 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1317 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1318 * We check NULL in callee rather than caller.
1321 atomic_dec(&memcg_moving);
1322 atomic_dec(&memcg->moving_account);
1327 * 2 routines for checking "mem" is under move_account() or not.
1329 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1330 * is used for avoiding races in accounting. If true,
1331 * pc->mem_cgroup may be overwritten.
1333 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1334 * under hierarchy of moving cgroups. This is for
1335 * waiting at hith-memory prressure caused by "move".
1338 static bool mem_cgroup_stolen(struct mem_cgroup *memcg)
1340 VM_BUG_ON(!rcu_read_lock_held());
1341 return atomic_read(&memcg->moving_account) > 0;
1344 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1346 struct mem_cgroup *from;
1347 struct mem_cgroup *to;
1350 * Unlike task_move routines, we access mc.to, mc.from not under
1351 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1353 spin_lock(&mc.lock);
1359 ret = mem_cgroup_same_or_subtree(memcg, from)
1360 || mem_cgroup_same_or_subtree(memcg, to);
1362 spin_unlock(&mc.lock);
1366 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1368 if (mc.moving_task && current != mc.moving_task) {
1369 if (mem_cgroup_under_move(memcg)) {
1371 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1372 /* moving charge context might have finished. */
1375 finish_wait(&mc.waitq, &wait);
1383 * Take this lock when
1384 * - a code tries to modify page's memcg while it's USED.
1385 * - a code tries to modify page state accounting in a memcg.
1386 * see mem_cgroup_stolen(), too.
1388 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1389 unsigned long *flags)
1391 spin_lock_irqsave(&memcg->move_lock, *flags);
1394 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1395 unsigned long *flags)
1397 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1401 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1402 * @memcg: The memory cgroup that went over limit
1403 * @p: Task that is going to be killed
1405 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1408 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1410 struct cgroup *task_cgrp;
1411 struct cgroup *mem_cgrp;
1413 * Need a buffer in BSS, can't rely on allocations. The code relies
1414 * on the assumption that OOM is serialized for memory controller.
1415 * If this assumption is broken, revisit this code.
1417 static char memcg_name[PATH_MAX];
1425 mem_cgrp = memcg->css.cgroup;
1426 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1428 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1431 * Unfortunately, we are unable to convert to a useful name
1432 * But we'll still print out the usage information
1439 printk(KERN_INFO "Task in %s killed", memcg_name);
1442 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1450 * Continues from above, so we don't need an KERN_ level
1452 printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1455 printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1456 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1457 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1458 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1459 printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1461 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1462 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1463 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1467 * This function returns the number of memcg under hierarchy tree. Returns
1468 * 1(self count) if no children.
1470 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1473 struct mem_cgroup *iter;
1475 for_each_mem_cgroup_tree(iter, memcg)
1481 * Return the memory (and swap, if configured) limit for a memcg.
1483 u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1488 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1489 limit += total_swap_pages << PAGE_SHIFT;
1491 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1493 * If memsw is finite and limits the amount of swap space available
1494 * to this memcg, return that limit.
1496 return min(limit, memsw);
1499 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1501 unsigned long flags)
1503 unsigned long total = 0;
1504 bool noswap = false;
1507 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1509 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1512 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1514 drain_all_stock_async(memcg);
1515 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1517 * Allow limit shrinkers, which are triggered directly
1518 * by userspace, to catch signals and stop reclaim
1519 * after minimal progress, regardless of the margin.
1521 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1523 if (mem_cgroup_margin(memcg))
1526 * If nothing was reclaimed after two attempts, there
1527 * may be no reclaimable pages in this hierarchy.
1536 * test_mem_cgroup_node_reclaimable
1537 * @mem: the target memcg
1538 * @nid: the node ID to be checked.
1539 * @noswap : specify true here if the user wants flle only information.
1541 * This function returns whether the specified memcg contains any
1542 * reclaimable pages on a node. Returns true if there are any reclaimable
1543 * pages in the node.
1545 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1546 int nid, bool noswap)
1548 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1550 if (noswap || !total_swap_pages)
1552 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1557 #if MAX_NUMNODES > 1
1560 * Always updating the nodemask is not very good - even if we have an empty
1561 * list or the wrong list here, we can start from some node and traverse all
1562 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1565 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1569 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1570 * pagein/pageout changes since the last update.
1572 if (!atomic_read(&memcg->numainfo_events))
1574 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1577 /* make a nodemask where this memcg uses memory from */
1578 memcg->scan_nodes = node_states[N_HIGH_MEMORY];
1580 for_each_node_mask(nid, node_states[N_HIGH_MEMORY]) {
1582 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1583 node_clear(nid, memcg->scan_nodes);
1586 atomic_set(&memcg->numainfo_events, 0);
1587 atomic_set(&memcg->numainfo_updating, 0);
1591 * Selecting a node where we start reclaim from. Because what we need is just
1592 * reducing usage counter, start from anywhere is O,K. Considering
1593 * memory reclaim from current node, there are pros. and cons.
1595 * Freeing memory from current node means freeing memory from a node which
1596 * we'll use or we've used. So, it may make LRU bad. And if several threads
1597 * hit limits, it will see a contention on a node. But freeing from remote
1598 * node means more costs for memory reclaim because of memory latency.
1600 * Now, we use round-robin. Better algorithm is welcomed.
1602 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1606 mem_cgroup_may_update_nodemask(memcg);
1607 node = memcg->last_scanned_node;
1609 node = next_node(node, memcg->scan_nodes);
1610 if (node == MAX_NUMNODES)
1611 node = first_node(memcg->scan_nodes);
1613 * We call this when we hit limit, not when pages are added to LRU.
1614 * No LRU may hold pages because all pages are UNEVICTABLE or
1615 * memcg is too small and all pages are not on LRU. In that case,
1616 * we use curret node.
1618 if (unlikely(node == MAX_NUMNODES))
1619 node = numa_node_id();
1621 memcg->last_scanned_node = node;
1626 * Check all nodes whether it contains reclaimable pages or not.
1627 * For quick scan, we make use of scan_nodes. This will allow us to skip
1628 * unused nodes. But scan_nodes is lazily updated and may not cotain
1629 * enough new information. We need to do double check.
1631 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1636 * quick check...making use of scan_node.
1637 * We can skip unused nodes.
1639 if (!nodes_empty(memcg->scan_nodes)) {
1640 for (nid = first_node(memcg->scan_nodes);
1642 nid = next_node(nid, memcg->scan_nodes)) {
1644 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1649 * Check rest of nodes.
1651 for_each_node_state(nid, N_HIGH_MEMORY) {
1652 if (node_isset(nid, memcg->scan_nodes))
1654 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1661 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1666 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1668 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
1672 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1675 unsigned long *total_scanned)
1677 struct mem_cgroup *victim = NULL;
1680 unsigned long excess;
1681 unsigned long nr_scanned;
1682 struct mem_cgroup_reclaim_cookie reclaim = {
1687 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
1690 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1695 * If we have not been able to reclaim
1696 * anything, it might because there are
1697 * no reclaimable pages under this hierarchy
1702 * We want to do more targeted reclaim.
1703 * excess >> 2 is not to excessive so as to
1704 * reclaim too much, nor too less that we keep
1705 * coming back to reclaim from this cgroup
1707 if (total >= (excess >> 2) ||
1708 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1713 if (!mem_cgroup_reclaimable(victim, false))
1715 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1717 *total_scanned += nr_scanned;
1718 if (!res_counter_soft_limit_excess(&root_memcg->res))
1721 mem_cgroup_iter_break(root_memcg, victim);
1726 * Check OOM-Killer is already running under our hierarchy.
1727 * If someone is running, return false.
1728 * Has to be called with memcg_oom_lock
1730 static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg)
1732 struct mem_cgroup *iter, *failed = NULL;
1734 for_each_mem_cgroup_tree(iter, memcg) {
1735 if (iter->oom_lock) {
1737 * this subtree of our hierarchy is already locked
1738 * so we cannot give a lock.
1741 mem_cgroup_iter_break(memcg, iter);
1744 iter->oom_lock = true;
1751 * OK, we failed to lock the whole subtree so we have to clean up
1752 * what we set up to the failing subtree
1754 for_each_mem_cgroup_tree(iter, memcg) {
1755 if (iter == failed) {
1756 mem_cgroup_iter_break(memcg, iter);
1759 iter->oom_lock = false;
1765 * Has to be called with memcg_oom_lock
1767 static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1769 struct mem_cgroup *iter;
1771 for_each_mem_cgroup_tree(iter, memcg)
1772 iter->oom_lock = false;
1776 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1778 struct mem_cgroup *iter;
1780 for_each_mem_cgroup_tree(iter, memcg)
1781 atomic_inc(&iter->under_oom);
1784 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1786 struct mem_cgroup *iter;
1789 * When a new child is created while the hierarchy is under oom,
1790 * mem_cgroup_oom_lock() may not be called. We have to use
1791 * atomic_add_unless() here.
1793 for_each_mem_cgroup_tree(iter, memcg)
1794 atomic_add_unless(&iter->under_oom, -1, 0);
1797 static DEFINE_SPINLOCK(memcg_oom_lock);
1798 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1800 struct oom_wait_info {
1801 struct mem_cgroup *memcg;
1805 static int memcg_oom_wake_function(wait_queue_t *wait,
1806 unsigned mode, int sync, void *arg)
1808 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1809 struct mem_cgroup *oom_wait_memcg;
1810 struct oom_wait_info *oom_wait_info;
1812 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1813 oom_wait_memcg = oom_wait_info->memcg;
1816 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
1817 * Then we can use css_is_ancestor without taking care of RCU.
1819 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
1820 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
1822 return autoremove_wake_function(wait, mode, sync, arg);
1825 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
1827 /* for filtering, pass "memcg" as argument. */
1828 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1831 static void memcg_oom_recover(struct mem_cgroup *memcg)
1833 if (memcg && atomic_read(&memcg->under_oom))
1834 memcg_wakeup_oom(memcg);
1838 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1840 static bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask,
1843 struct oom_wait_info owait;
1844 bool locked, need_to_kill;
1846 owait.memcg = memcg;
1847 owait.wait.flags = 0;
1848 owait.wait.func = memcg_oom_wake_function;
1849 owait.wait.private = current;
1850 INIT_LIST_HEAD(&owait.wait.task_list);
1851 need_to_kill = true;
1852 mem_cgroup_mark_under_oom(memcg);
1854 /* At first, try to OOM lock hierarchy under memcg.*/
1855 spin_lock(&memcg_oom_lock);
1856 locked = mem_cgroup_oom_lock(memcg);
1858 * Even if signal_pending(), we can't quit charge() loop without
1859 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1860 * under OOM is always welcomed, use TASK_KILLABLE here.
1862 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1863 if (!locked || memcg->oom_kill_disable)
1864 need_to_kill = false;
1866 mem_cgroup_oom_notify(memcg);
1867 spin_unlock(&memcg_oom_lock);
1870 finish_wait(&memcg_oom_waitq, &owait.wait);
1871 mem_cgroup_out_of_memory(memcg, mask, order);
1874 finish_wait(&memcg_oom_waitq, &owait.wait);
1876 spin_lock(&memcg_oom_lock);
1878 mem_cgroup_oom_unlock(memcg);
1879 memcg_wakeup_oom(memcg);
1880 spin_unlock(&memcg_oom_lock);
1882 mem_cgroup_unmark_under_oom(memcg);
1884 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
1886 /* Give chance to dying process */
1887 schedule_timeout_uninterruptible(1);
1892 * Currently used to update mapped file statistics, but the routine can be
1893 * generalized to update other statistics as well.
1895 * Notes: Race condition
1897 * We usually use page_cgroup_lock() for accessing page_cgroup member but
1898 * it tends to be costly. But considering some conditions, we doesn't need
1899 * to do so _always_.
1901 * Considering "charge", lock_page_cgroup() is not required because all
1902 * file-stat operations happen after a page is attached to radix-tree. There
1903 * are no race with "charge".
1905 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
1906 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
1907 * if there are race with "uncharge". Statistics itself is properly handled
1910 * Considering "move", this is an only case we see a race. To make the race
1911 * small, we check mm->moving_account and detect there are possibility of race
1912 * If there is, we take a lock.
1915 void __mem_cgroup_begin_update_page_stat(struct page *page,
1916 bool *locked, unsigned long *flags)
1918 struct mem_cgroup *memcg;
1919 struct page_cgroup *pc;
1921 pc = lookup_page_cgroup(page);
1923 memcg = pc->mem_cgroup;
1924 if (unlikely(!memcg || !PageCgroupUsed(pc)))
1927 * If this memory cgroup is not under account moving, we don't
1928 * need to take move_lock_page_cgroup(). Because we already hold
1929 * rcu_read_lock(), any calls to move_account will be delayed until
1930 * rcu_read_unlock() if mem_cgroup_stolen() == true.
1932 if (!mem_cgroup_stolen(memcg))
1935 move_lock_mem_cgroup(memcg, flags);
1936 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
1937 move_unlock_mem_cgroup(memcg, flags);
1943 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
1945 struct page_cgroup *pc = lookup_page_cgroup(page);
1948 * It's guaranteed that pc->mem_cgroup never changes while
1949 * lock is held because a routine modifies pc->mem_cgroup
1950 * should take move_lock_page_cgroup().
1952 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
1955 void mem_cgroup_update_page_stat(struct page *page,
1956 enum mem_cgroup_page_stat_item idx, int val)
1958 struct mem_cgroup *memcg;
1959 struct page_cgroup *pc = lookup_page_cgroup(page);
1960 unsigned long uninitialized_var(flags);
1962 if (mem_cgroup_disabled())
1965 memcg = pc->mem_cgroup;
1966 if (unlikely(!memcg || !PageCgroupUsed(pc)))
1970 case MEMCG_NR_FILE_MAPPED:
1971 idx = MEM_CGROUP_STAT_FILE_MAPPED;
1977 this_cpu_add(memcg->stat->count[idx], val);
1981 * size of first charge trial. "32" comes from vmscan.c's magic value.
1982 * TODO: maybe necessary to use big numbers in big irons.
1984 #define CHARGE_BATCH 32U
1985 struct memcg_stock_pcp {
1986 struct mem_cgroup *cached; /* this never be root cgroup */
1987 unsigned int nr_pages;
1988 struct work_struct work;
1989 unsigned long flags;
1990 #define FLUSHING_CACHED_CHARGE (0)
1992 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1993 static DEFINE_MUTEX(percpu_charge_mutex);
1996 * Try to consume stocked charge on this cpu. If success, one page is consumed
1997 * from local stock and true is returned. If the stock is 0 or charges from a
1998 * cgroup which is not current target, returns false. This stock will be
2001 static bool consume_stock(struct mem_cgroup *memcg)
2003 struct memcg_stock_pcp *stock;
2006 stock = &get_cpu_var(memcg_stock);
2007 if (memcg == stock->cached && stock->nr_pages)
2009 else /* need to call res_counter_charge */
2011 put_cpu_var(memcg_stock);
2016 * Returns stocks cached in percpu to res_counter and reset cached information.
2018 static void drain_stock(struct memcg_stock_pcp *stock)
2020 struct mem_cgroup *old = stock->cached;
2022 if (stock->nr_pages) {
2023 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2025 res_counter_uncharge(&old->res, bytes);
2026 if (do_swap_account)
2027 res_counter_uncharge(&old->memsw, bytes);
2028 stock->nr_pages = 0;
2030 stock->cached = NULL;
2034 * This must be called under preempt disabled or must be called by
2035 * a thread which is pinned to local cpu.
2037 static void drain_local_stock(struct work_struct *dummy)
2039 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2041 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2045 * Cache charges(val) which is from res_counter, to local per_cpu area.
2046 * This will be consumed by consume_stock() function, later.
2048 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2050 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2052 if (stock->cached != memcg) { /* reset if necessary */
2054 stock->cached = memcg;
2056 stock->nr_pages += nr_pages;
2057 put_cpu_var(memcg_stock);
2061 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2062 * of the hierarchy under it. sync flag says whether we should block
2063 * until the work is done.
2065 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2069 /* Notify other cpus that system-wide "drain" is running */
2072 for_each_online_cpu(cpu) {
2073 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2074 struct mem_cgroup *memcg;
2076 memcg = stock->cached;
2077 if (!memcg || !stock->nr_pages)
2079 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2081 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2083 drain_local_stock(&stock->work);
2085 schedule_work_on(cpu, &stock->work);
2093 for_each_online_cpu(cpu) {
2094 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2095 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2096 flush_work(&stock->work);
2103 * Tries to drain stocked charges in other cpus. This function is asynchronous
2104 * and just put a work per cpu for draining localy on each cpu. Caller can
2105 * expects some charges will be back to res_counter later but cannot wait for
2108 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2111 * If someone calls draining, avoid adding more kworker runs.
2113 if (!mutex_trylock(&percpu_charge_mutex))
2115 drain_all_stock(root_memcg, false);
2116 mutex_unlock(&percpu_charge_mutex);
2119 /* This is a synchronous drain interface. */
2120 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2122 /* called when force_empty is called */
2123 mutex_lock(&percpu_charge_mutex);
2124 drain_all_stock(root_memcg, true);
2125 mutex_unlock(&percpu_charge_mutex);
2129 * This function drains percpu counter value from DEAD cpu and
2130 * move it to local cpu. Note that this function can be preempted.
2132 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2136 spin_lock(&memcg->pcp_counter_lock);
2137 for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) {
2138 long x = per_cpu(memcg->stat->count[i], cpu);
2140 per_cpu(memcg->stat->count[i], cpu) = 0;
2141 memcg->nocpu_base.count[i] += x;
2143 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2144 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2146 per_cpu(memcg->stat->events[i], cpu) = 0;
2147 memcg->nocpu_base.events[i] += x;
2149 spin_unlock(&memcg->pcp_counter_lock);
2152 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2153 unsigned long action,
2156 int cpu = (unsigned long)hcpu;
2157 struct memcg_stock_pcp *stock;
2158 struct mem_cgroup *iter;
2160 if (action == CPU_ONLINE)
2163 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2166 for_each_mem_cgroup(iter)
2167 mem_cgroup_drain_pcp_counter(iter, cpu);
2169 stock = &per_cpu(memcg_stock, cpu);
2175 /* See __mem_cgroup_try_charge() for details */
2177 CHARGE_OK, /* success */
2178 CHARGE_RETRY, /* need to retry but retry is not bad */
2179 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2180 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2181 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2184 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2185 unsigned int nr_pages, bool oom_check)
2187 unsigned long csize = nr_pages * PAGE_SIZE;
2188 struct mem_cgroup *mem_over_limit;
2189 struct res_counter *fail_res;
2190 unsigned long flags = 0;
2193 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2196 if (!do_swap_account)
2198 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2202 res_counter_uncharge(&memcg->res, csize);
2203 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2204 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2206 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2208 * nr_pages can be either a huge page (HPAGE_PMD_NR), a batch
2209 * of regular pages (CHARGE_BATCH), or a single regular page (1).
2211 * Never reclaim on behalf of optional batching, retry with a
2212 * single page instead.
2214 if (nr_pages == CHARGE_BATCH)
2215 return CHARGE_RETRY;
2217 if (!(gfp_mask & __GFP_WAIT))
2218 return CHARGE_WOULDBLOCK;
2220 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2221 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2222 return CHARGE_RETRY;
2224 * Even though the limit is exceeded at this point, reclaim
2225 * may have been able to free some pages. Retry the charge
2226 * before killing the task.
2228 * Only for regular pages, though: huge pages are rather
2229 * unlikely to succeed so close to the limit, and we fall back
2230 * to regular pages anyway in case of failure.
2232 if (nr_pages == 1 && ret)
2233 return CHARGE_RETRY;
2236 * At task move, charge accounts can be doubly counted. So, it's
2237 * better to wait until the end of task_move if something is going on.
2239 if (mem_cgroup_wait_acct_move(mem_over_limit))
2240 return CHARGE_RETRY;
2242 /* If we don't need to call oom-killer at el, return immediately */
2244 return CHARGE_NOMEM;
2246 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask, get_order(csize)))
2247 return CHARGE_OOM_DIE;
2249 return CHARGE_RETRY;
2253 * __mem_cgroup_try_charge() does
2254 * 1. detect memcg to be charged against from passed *mm and *ptr,
2255 * 2. update res_counter
2256 * 3. call memory reclaim if necessary.
2258 * In some special case, if the task is fatal, fatal_signal_pending() or
2259 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2260 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2261 * as possible without any hazards. 2: all pages should have a valid
2262 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2263 * pointer, that is treated as a charge to root_mem_cgroup.
2265 * So __mem_cgroup_try_charge() will return
2266 * 0 ... on success, filling *ptr with a valid memcg pointer.
2267 * -ENOMEM ... charge failure because of resource limits.
2268 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2270 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2271 * the oom-killer can be invoked.
2273 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2275 unsigned int nr_pages,
2276 struct mem_cgroup **ptr,
2279 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2280 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2281 struct mem_cgroup *memcg = NULL;
2285 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2286 * in system level. So, allow to go ahead dying process in addition to
2289 if (unlikely(test_thread_flag(TIF_MEMDIE)
2290 || fatal_signal_pending(current)))
2294 * We always charge the cgroup the mm_struct belongs to.
2295 * The mm_struct's mem_cgroup changes on task migration if the
2296 * thread group leader migrates. It's possible that mm is not
2297 * set, if so charge the init_mm (happens for pagecache usage).
2300 *ptr = root_mem_cgroup;
2302 if (*ptr) { /* css should be a valid one */
2304 VM_BUG_ON(css_is_removed(&memcg->css));
2305 if (mem_cgroup_is_root(memcg))
2307 if (nr_pages == 1 && consume_stock(memcg))
2309 css_get(&memcg->css);
2311 struct task_struct *p;
2314 p = rcu_dereference(mm->owner);
2316 * Because we don't have task_lock(), "p" can exit.
2317 * In that case, "memcg" can point to root or p can be NULL with
2318 * race with swapoff. Then, we have small risk of mis-accouning.
2319 * But such kind of mis-account by race always happens because
2320 * we don't have cgroup_mutex(). It's overkill and we allo that
2322 * (*) swapoff at el will charge against mm-struct not against
2323 * task-struct. So, mm->owner can be NULL.
2325 memcg = mem_cgroup_from_task(p);
2327 memcg = root_mem_cgroup;
2328 if (mem_cgroup_is_root(memcg)) {
2332 if (nr_pages == 1 && consume_stock(memcg)) {
2334 * It seems dagerous to access memcg without css_get().
2335 * But considering how consume_stok works, it's not
2336 * necessary. If consume_stock success, some charges
2337 * from this memcg are cached on this cpu. So, we
2338 * don't need to call css_get()/css_tryget() before
2339 * calling consume_stock().
2344 /* after here, we may be blocked. we need to get refcnt */
2345 if (!css_tryget(&memcg->css)) {
2355 /* If killed, bypass charge */
2356 if (fatal_signal_pending(current)) {
2357 css_put(&memcg->css);
2362 if (oom && !nr_oom_retries) {
2364 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2367 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, oom_check);
2371 case CHARGE_RETRY: /* not in OOM situation but retry */
2373 css_put(&memcg->css);
2376 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2377 css_put(&memcg->css);
2379 case CHARGE_NOMEM: /* OOM routine works */
2381 css_put(&memcg->css);
2384 /* If oom, we never return -ENOMEM */
2387 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2388 css_put(&memcg->css);
2391 } while (ret != CHARGE_OK);
2393 if (batch > nr_pages)
2394 refill_stock(memcg, batch - nr_pages);
2395 css_put(&memcg->css);
2403 *ptr = root_mem_cgroup;
2408 * Somemtimes we have to undo a charge we got by try_charge().
2409 * This function is for that and do uncharge, put css's refcnt.
2410 * gotten by try_charge().
2412 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2413 unsigned int nr_pages)
2415 if (!mem_cgroup_is_root(memcg)) {
2416 unsigned long bytes = nr_pages * PAGE_SIZE;
2418 res_counter_uncharge(&memcg->res, bytes);
2419 if (do_swap_account)
2420 res_counter_uncharge(&memcg->memsw, bytes);
2425 * A helper function to get mem_cgroup from ID. must be called under
2426 * rcu_read_lock(). The caller must check css_is_removed() or some if
2427 * it's concern. (dropping refcnt from swap can be called against removed
2430 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2432 struct cgroup_subsys_state *css;
2434 /* ID 0 is unused ID */
2437 css = css_lookup(&mem_cgroup_subsys, id);
2440 return container_of(css, struct mem_cgroup, css);
2443 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2445 struct mem_cgroup *memcg = NULL;
2446 struct page_cgroup *pc;
2450 VM_BUG_ON(!PageLocked(page));
2452 pc = lookup_page_cgroup(page);
2453 lock_page_cgroup(pc);
2454 if (PageCgroupUsed(pc)) {
2455 memcg = pc->mem_cgroup;
2456 if (memcg && !css_tryget(&memcg->css))
2458 } else if (PageSwapCache(page)) {
2459 ent.val = page_private(page);
2460 id = lookup_swap_cgroup_id(ent);
2462 memcg = mem_cgroup_lookup(id);
2463 if (memcg && !css_tryget(&memcg->css))
2467 unlock_page_cgroup(pc);
2471 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2473 unsigned int nr_pages,
2474 enum charge_type ctype,
2477 struct page_cgroup *pc = lookup_page_cgroup(page);
2478 struct zone *uninitialized_var(zone);
2479 bool was_on_lru = false;
2482 lock_page_cgroup(pc);
2483 if (unlikely(PageCgroupUsed(pc))) {
2484 unlock_page_cgroup(pc);
2485 __mem_cgroup_cancel_charge(memcg, nr_pages);
2489 * we don't need page_cgroup_lock about tail pages, becase they are not
2490 * accessed by any other context at this point.
2494 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2495 * may already be on some other mem_cgroup's LRU. Take care of it.
2498 zone = page_zone(page);
2499 spin_lock_irq(&zone->lru_lock);
2500 if (PageLRU(page)) {
2502 del_page_from_lru_list(zone, page, page_lru(page));
2507 pc->mem_cgroup = memcg;
2509 * We access a page_cgroup asynchronously without lock_page_cgroup().
2510 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2511 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2512 * before USED bit, we need memory barrier here.
2513 * See mem_cgroup_add_lru_list(), etc.
2516 SetPageCgroupUsed(pc);
2520 VM_BUG_ON(PageLRU(page));
2522 add_page_to_lru_list(zone, page, page_lru(page));
2524 spin_unlock_irq(&zone->lru_lock);
2527 if (ctype == MEM_CGROUP_CHARGE_TYPE_MAPPED)
2532 mem_cgroup_charge_statistics(memcg, anon, nr_pages);
2533 unlock_page_cgroup(pc);
2536 * "charge_statistics" updated event counter. Then, check it.
2537 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2538 * if they exceeds softlimit.
2540 memcg_check_events(memcg, page);
2543 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2545 #define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MIGRATION))
2547 * Because tail pages are not marked as "used", set it. We're under
2548 * zone->lru_lock, 'splitting on pmd' and compound_lock.
2549 * charge/uncharge will be never happen and move_account() is done under
2550 * compound_lock(), so we don't have to take care of races.
2552 void mem_cgroup_split_huge_fixup(struct page *head)
2554 struct page_cgroup *head_pc = lookup_page_cgroup(head);
2555 struct page_cgroup *pc;
2558 if (mem_cgroup_disabled())
2560 for (i = 1; i < HPAGE_PMD_NR; i++) {
2562 pc->mem_cgroup = head_pc->mem_cgroup;
2563 smp_wmb();/* see __commit_charge() */
2564 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2567 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2570 * mem_cgroup_move_account - move account of the page
2572 * @nr_pages: number of regular pages (>1 for huge pages)
2573 * @pc: page_cgroup of the page.
2574 * @from: mem_cgroup which the page is moved from.
2575 * @to: mem_cgroup which the page is moved to. @from != @to.
2576 * @uncharge: whether we should call uncharge and css_put against @from.
2578 * The caller must confirm following.
2579 * - page is not on LRU (isolate_page() is useful.)
2580 * - compound_lock is held when nr_pages > 1
2582 * This function doesn't do "charge" nor css_get to new cgroup. It should be
2583 * done by a caller(__mem_cgroup_try_charge would be useful). If @uncharge is
2584 * true, this function does "uncharge" from old cgroup, but it doesn't if
2585 * @uncharge is false, so a caller should do "uncharge".
2587 static int mem_cgroup_move_account(struct page *page,
2588 unsigned int nr_pages,
2589 struct page_cgroup *pc,
2590 struct mem_cgroup *from,
2591 struct mem_cgroup *to,
2594 unsigned long flags;
2596 bool anon = PageAnon(page);
2598 VM_BUG_ON(from == to);
2599 VM_BUG_ON(PageLRU(page));
2601 * The page is isolated from LRU. So, collapse function
2602 * will not handle this page. But page splitting can happen.
2603 * Do this check under compound_page_lock(). The caller should
2607 if (nr_pages > 1 && !PageTransHuge(page))
2610 lock_page_cgroup(pc);
2613 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
2616 move_lock_mem_cgroup(from, &flags);
2618 if (!anon && page_mapped(page)) {
2619 /* Update mapped_file data for mem_cgroup */
2621 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2622 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2625 mem_cgroup_charge_statistics(from, anon, -nr_pages);
2627 /* This is not "cancel", but cancel_charge does all we need. */
2628 __mem_cgroup_cancel_charge(from, nr_pages);
2630 /* caller should have done css_get */
2631 pc->mem_cgroup = to;
2632 mem_cgroup_charge_statistics(to, anon, nr_pages);
2634 * We charges against "to" which may not have any tasks. Then, "to"
2635 * can be under rmdir(). But in current implementation, caller of
2636 * this function is just force_empty() and move charge, so it's
2637 * guaranteed that "to" is never removed. So, we don't check rmdir
2640 move_unlock_mem_cgroup(from, &flags);
2643 unlock_page_cgroup(pc);
2647 memcg_check_events(to, page);
2648 memcg_check_events(from, page);
2654 * move charges to its parent.
2657 static int mem_cgroup_move_parent(struct page *page,
2658 struct page_cgroup *pc,
2659 struct mem_cgroup *child,
2662 struct cgroup *cg = child->css.cgroup;
2663 struct cgroup *pcg = cg->parent;
2664 struct mem_cgroup *parent;
2665 unsigned int nr_pages;
2666 unsigned long uninitialized_var(flags);
2674 if (!get_page_unless_zero(page))
2676 if (isolate_lru_page(page))
2679 nr_pages = hpage_nr_pages(page);
2681 parent = mem_cgroup_from_cont(pcg);
2682 ret = __mem_cgroup_try_charge(NULL, gfp_mask, nr_pages, &parent, false);
2687 flags = compound_lock_irqsave(page);
2689 ret = mem_cgroup_move_account(page, nr_pages, pc, child, parent, true);
2691 __mem_cgroup_cancel_charge(parent, nr_pages);
2694 compound_unlock_irqrestore(page, flags);
2696 putback_lru_page(page);
2704 * Charge the memory controller for page usage.
2706 * 0 if the charge was successful
2707 * < 0 if the cgroup is over its limit
2709 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2710 gfp_t gfp_mask, enum charge_type ctype)
2712 struct mem_cgroup *memcg = NULL;
2713 unsigned int nr_pages = 1;
2717 if (PageTransHuge(page)) {
2718 nr_pages <<= compound_order(page);
2719 VM_BUG_ON(!PageTransHuge(page));
2721 * Never OOM-kill a process for a huge page. The
2722 * fault handler will fall back to regular pages.
2727 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
2730 __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false);
2734 int mem_cgroup_newpage_charge(struct page *page,
2735 struct mm_struct *mm, gfp_t gfp_mask)
2737 if (mem_cgroup_disabled())
2739 VM_BUG_ON(page_mapped(page));
2740 VM_BUG_ON(page->mapping && !PageAnon(page));
2742 return mem_cgroup_charge_common(page, mm, gfp_mask,
2743 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2747 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2748 enum charge_type ctype);
2750 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2753 struct mem_cgroup *memcg = NULL;
2754 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
2757 if (mem_cgroup_disabled())
2759 if (PageCompound(page))
2764 if (!page_is_file_cache(page))
2765 type = MEM_CGROUP_CHARGE_TYPE_SHMEM;
2767 if (!PageSwapCache(page))
2768 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
2769 else { /* page is swapcache/shmem */
2770 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &memcg);
2772 __mem_cgroup_commit_charge_swapin(page, memcg, type);
2778 * While swap-in, try_charge -> commit or cancel, the page is locked.
2779 * And when try_charge() successfully returns, one refcnt to memcg without
2780 * struct page_cgroup is acquired. This refcnt will be consumed by
2781 * "commit()" or removed by "cancel()"
2783 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2785 gfp_t mask, struct mem_cgroup **memcgp)
2787 struct mem_cgroup *memcg;
2792 if (mem_cgroup_disabled())
2795 if (!do_swap_account)
2798 * A racing thread's fault, or swapoff, may have already updated
2799 * the pte, and even removed page from swap cache: in those cases
2800 * do_swap_page()'s pte_same() test will fail; but there's also a
2801 * KSM case which does need to charge the page.
2803 if (!PageSwapCache(page))
2805 memcg = try_get_mem_cgroup_from_page(page);
2809 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
2810 css_put(&memcg->css);
2817 ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
2824 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
2825 enum charge_type ctype)
2827 if (mem_cgroup_disabled())
2831 cgroup_exclude_rmdir(&memcg->css);
2833 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
2835 * Now swap is on-memory. This means this page may be
2836 * counted both as mem and swap....double count.
2837 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2838 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2839 * may call delete_from_swap_cache() before reach here.
2841 if (do_swap_account && PageSwapCache(page)) {
2842 swp_entry_t ent = {.val = page_private(page)};
2843 mem_cgroup_uncharge_swap(ent);
2846 * At swapin, we may charge account against cgroup which has no tasks.
2847 * So, rmdir()->pre_destroy() can be called while we do this charge.
2848 * In that case, we need to call pre_destroy() again. check it here.
2850 cgroup_release_and_wakeup_rmdir(&memcg->css);
2853 void mem_cgroup_commit_charge_swapin(struct page *page,
2854 struct mem_cgroup *memcg)
2856 __mem_cgroup_commit_charge_swapin(page, memcg,
2857 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2860 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
2862 if (mem_cgroup_disabled())
2866 __mem_cgroup_cancel_charge(memcg, 1);
2869 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
2870 unsigned int nr_pages,
2871 const enum charge_type ctype)
2873 struct memcg_batch_info *batch = NULL;
2874 bool uncharge_memsw = true;
2876 /* If swapout, usage of swap doesn't decrease */
2877 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2878 uncharge_memsw = false;
2880 batch = ¤t->memcg_batch;
2882 * In usual, we do css_get() when we remember memcg pointer.
2883 * But in this case, we keep res->usage until end of a series of
2884 * uncharges. Then, it's ok to ignore memcg's refcnt.
2887 batch->memcg = memcg;
2889 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2890 * In those cases, all pages freed continuously can be expected to be in
2891 * the same cgroup and we have chance to coalesce uncharges.
2892 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2893 * because we want to do uncharge as soon as possible.
2896 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
2897 goto direct_uncharge;
2900 goto direct_uncharge;
2903 * In typical case, batch->memcg == mem. This means we can
2904 * merge a series of uncharges to an uncharge of res_counter.
2905 * If not, we uncharge res_counter ony by one.
2907 if (batch->memcg != memcg)
2908 goto direct_uncharge;
2909 /* remember freed charge and uncharge it later */
2912 batch->memsw_nr_pages++;
2915 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
2917 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
2918 if (unlikely(batch->memcg != memcg))
2919 memcg_oom_recover(memcg);
2923 * uncharge if !page_mapped(page)
2925 static struct mem_cgroup *
2926 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
2928 struct mem_cgroup *memcg = NULL;
2929 unsigned int nr_pages = 1;
2930 struct page_cgroup *pc;
2933 if (mem_cgroup_disabled())
2936 if (PageSwapCache(page))
2939 if (PageTransHuge(page)) {
2940 nr_pages <<= compound_order(page);
2941 VM_BUG_ON(!PageTransHuge(page));
2944 * Check if our page_cgroup is valid
2946 pc = lookup_page_cgroup(page);
2947 if (unlikely(!PageCgroupUsed(pc)))
2950 lock_page_cgroup(pc);
2952 memcg = pc->mem_cgroup;
2954 if (!PageCgroupUsed(pc))
2957 anon = PageAnon(page);
2960 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2962 * Generally PageAnon tells if it's the anon statistics to be
2963 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
2964 * used before page reached the stage of being marked PageAnon.
2968 case MEM_CGROUP_CHARGE_TYPE_DROP:
2969 /* See mem_cgroup_prepare_migration() */
2970 if (page_mapped(page) || PageCgroupMigration(pc))
2973 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
2974 if (!PageAnon(page)) { /* Shared memory */
2975 if (page->mapping && !page_is_file_cache(page))
2977 } else if (page_mapped(page)) /* Anon */
2984 mem_cgroup_charge_statistics(memcg, anon, -nr_pages);
2986 ClearPageCgroupUsed(pc);
2988 * pc->mem_cgroup is not cleared here. It will be accessed when it's
2989 * freed from LRU. This is safe because uncharged page is expected not
2990 * to be reused (freed soon). Exception is SwapCache, it's handled by
2991 * special functions.
2994 unlock_page_cgroup(pc);
2996 * even after unlock, we have memcg->res.usage here and this memcg
2997 * will never be freed.
2999 memcg_check_events(memcg, page);
3000 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
3001 mem_cgroup_swap_statistics(memcg, true);
3002 mem_cgroup_get(memcg);
3004 if (!mem_cgroup_is_root(memcg))
3005 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
3010 unlock_page_cgroup(pc);
3014 void mem_cgroup_uncharge_page(struct page *page)
3017 if (page_mapped(page))
3019 VM_BUG_ON(page->mapping && !PageAnon(page));
3020 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
3023 void mem_cgroup_uncharge_cache_page(struct page *page)
3025 VM_BUG_ON(page_mapped(page));
3026 VM_BUG_ON(page->mapping);
3027 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
3031 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
3032 * In that cases, pages are freed continuously and we can expect pages
3033 * are in the same memcg. All these calls itself limits the number of
3034 * pages freed at once, then uncharge_start/end() is called properly.
3035 * This may be called prural(2) times in a context,
3038 void mem_cgroup_uncharge_start(void)
3040 current->memcg_batch.do_batch++;
3041 /* We can do nest. */
3042 if (current->memcg_batch.do_batch == 1) {
3043 current->memcg_batch.memcg = NULL;
3044 current->memcg_batch.nr_pages = 0;
3045 current->memcg_batch.memsw_nr_pages = 0;
3049 void mem_cgroup_uncharge_end(void)
3051 struct memcg_batch_info *batch = ¤t->memcg_batch;
3053 if (!batch->do_batch)
3057 if (batch->do_batch) /* If stacked, do nothing. */
3063 * This "batch->memcg" is valid without any css_get/put etc...
3064 * bacause we hide charges behind us.
3066 if (batch->nr_pages)
3067 res_counter_uncharge(&batch->memcg->res,
3068 batch->nr_pages * PAGE_SIZE);
3069 if (batch->memsw_nr_pages)
3070 res_counter_uncharge(&batch->memcg->memsw,
3071 batch->memsw_nr_pages * PAGE_SIZE);
3072 memcg_oom_recover(batch->memcg);
3073 /* forget this pointer (for sanity check) */
3074 batch->memcg = NULL;
3079 * called after __delete_from_swap_cache() and drop "page" account.
3080 * memcg information is recorded to swap_cgroup of "ent"
3083 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
3085 struct mem_cgroup *memcg;
3086 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
3088 if (!swapout) /* this was a swap cache but the swap is unused ! */
3089 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
3091 memcg = __mem_cgroup_uncharge_common(page, ctype);
3094 * record memcg information, if swapout && memcg != NULL,
3095 * mem_cgroup_get() was called in uncharge().
3097 if (do_swap_account && swapout && memcg)
3098 swap_cgroup_record(ent, css_id(&memcg->css));
3102 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3104 * called from swap_entry_free(). remove record in swap_cgroup and
3105 * uncharge "memsw" account.
3107 void mem_cgroup_uncharge_swap(swp_entry_t ent)
3109 struct mem_cgroup *memcg;
3112 if (!do_swap_account)
3115 id = swap_cgroup_record(ent, 0);
3117 memcg = mem_cgroup_lookup(id);
3120 * We uncharge this because swap is freed.
3121 * This memcg can be obsolete one. We avoid calling css_tryget
3123 if (!mem_cgroup_is_root(memcg))
3124 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
3125 mem_cgroup_swap_statistics(memcg, false);
3126 mem_cgroup_put(memcg);
3132 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3133 * @entry: swap entry to be moved
3134 * @from: mem_cgroup which the entry is moved from
3135 * @to: mem_cgroup which the entry is moved to
3137 * It succeeds only when the swap_cgroup's record for this entry is the same
3138 * as the mem_cgroup's id of @from.
3140 * Returns 0 on success, -EINVAL on failure.
3142 * The caller must have charged to @to, IOW, called res_counter_charge() about
3143 * both res and memsw, and called css_get().
3145 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3146 struct mem_cgroup *from, struct mem_cgroup *to)
3148 unsigned short old_id, new_id;
3150 old_id = css_id(&from->css);
3151 new_id = css_id(&to->css);
3153 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3154 mem_cgroup_swap_statistics(from, false);
3155 mem_cgroup_swap_statistics(to, true);
3157 * This function is only called from task migration context now.
3158 * It postpones res_counter and refcount handling till the end
3159 * of task migration(mem_cgroup_clear_mc()) for performance
3160 * improvement. But we cannot postpone mem_cgroup_get(to)
3161 * because if the process that has been moved to @to does
3162 * swap-in, the refcount of @to might be decreased to 0.
3170 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3171 struct mem_cgroup *from, struct mem_cgroup *to)
3178 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
3181 int mem_cgroup_prepare_migration(struct page *page,
3182 struct page *newpage, struct mem_cgroup **memcgp, gfp_t gfp_mask)
3184 struct mem_cgroup *memcg = NULL;
3185 struct page_cgroup *pc;
3186 enum charge_type ctype;
3191 VM_BUG_ON(PageTransHuge(page));
3192 if (mem_cgroup_disabled())
3195 pc = lookup_page_cgroup(page);
3196 lock_page_cgroup(pc);
3197 if (PageCgroupUsed(pc)) {
3198 memcg = pc->mem_cgroup;
3199 css_get(&memcg->css);
3201 * At migrating an anonymous page, its mapcount goes down
3202 * to 0 and uncharge() will be called. But, even if it's fully
3203 * unmapped, migration may fail and this page has to be
3204 * charged again. We set MIGRATION flag here and delay uncharge
3205 * until end_migration() is called
3207 * Corner Case Thinking
3209 * When the old page was mapped as Anon and it's unmap-and-freed
3210 * while migration was ongoing.
3211 * If unmap finds the old page, uncharge() of it will be delayed
3212 * until end_migration(). If unmap finds a new page, it's
3213 * uncharged when it make mapcount to be 1->0. If unmap code
3214 * finds swap_migration_entry, the new page will not be mapped
3215 * and end_migration() will find it(mapcount==0).
3218 * When the old page was mapped but migraion fails, the kernel
3219 * remaps it. A charge for it is kept by MIGRATION flag even
3220 * if mapcount goes down to 0. We can do remap successfully
3221 * without charging it again.
3224 * The "old" page is under lock_page() until the end of
3225 * migration, so, the old page itself will not be swapped-out.
3226 * If the new page is swapped out before end_migraton, our
3227 * hook to usual swap-out path will catch the event.
3230 SetPageCgroupMigration(pc);
3232 unlock_page_cgroup(pc);
3234 * If the page is not charged at this point,
3241 ret = __mem_cgroup_try_charge(NULL, gfp_mask, 1, memcgp, false);
3242 css_put(&memcg->css);/* drop extra refcnt */
3244 if (PageAnon(page)) {
3245 lock_page_cgroup(pc);
3246 ClearPageCgroupMigration(pc);
3247 unlock_page_cgroup(pc);
3249 * The old page may be fully unmapped while we kept it.
3251 mem_cgroup_uncharge_page(page);
3253 /* we'll need to revisit this error code (we have -EINTR) */
3257 * We charge new page before it's used/mapped. So, even if unlock_page()
3258 * is called before end_migration, we can catch all events on this new
3259 * page. In the case new page is migrated but not remapped, new page's
3260 * mapcount will be finally 0 and we call uncharge in end_migration().
3263 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
3264 else if (page_is_file_cache(page))
3265 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
3267 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3268 __mem_cgroup_commit_charge(memcg, newpage, 1, ctype, false);
3272 /* remove redundant charge if migration failed*/
3273 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
3274 struct page *oldpage, struct page *newpage, bool migration_ok)
3276 struct page *used, *unused;
3277 struct page_cgroup *pc;
3282 /* blocks rmdir() */
3283 cgroup_exclude_rmdir(&memcg->css);
3284 if (!migration_ok) {
3292 * We disallowed uncharge of pages under migration because mapcount
3293 * of the page goes down to zero, temporarly.
3294 * Clear the flag and check the page should be charged.
3296 pc = lookup_page_cgroup(oldpage);
3297 lock_page_cgroup(pc);
3298 ClearPageCgroupMigration(pc);
3299 unlock_page_cgroup(pc);
3300 anon = PageAnon(used);
3301 __mem_cgroup_uncharge_common(unused,
3302 anon ? MEM_CGROUP_CHARGE_TYPE_MAPPED
3303 : MEM_CGROUP_CHARGE_TYPE_CACHE);
3306 * If a page is a file cache, radix-tree replacement is very atomic
3307 * and we can skip this check. When it was an Anon page, its mapcount
3308 * goes down to 0. But because we added MIGRATION flage, it's not
3309 * uncharged yet. There are several case but page->mapcount check
3310 * and USED bit check in mem_cgroup_uncharge_page() will do enough
3311 * check. (see prepare_charge() also)
3314 mem_cgroup_uncharge_page(used);
3316 * At migration, we may charge account against cgroup which has no
3318 * So, rmdir()->pre_destroy() can be called while we do this charge.
3319 * In that case, we need to call pre_destroy() again. check it here.
3321 cgroup_release_and_wakeup_rmdir(&memcg->css);
3325 * At replace page cache, newpage is not under any memcg but it's on
3326 * LRU. So, this function doesn't touch res_counter but handles LRU
3327 * in correct way. Both pages are locked so we cannot race with uncharge.
3329 void mem_cgroup_replace_page_cache(struct page *oldpage,
3330 struct page *newpage)
3332 struct mem_cgroup *memcg = NULL;
3333 struct page_cgroup *pc;
3334 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
3336 if (mem_cgroup_disabled())
3339 pc = lookup_page_cgroup(oldpage);
3340 /* fix accounting on old pages */
3341 lock_page_cgroup(pc);
3342 if (PageCgroupUsed(pc)) {
3343 memcg = pc->mem_cgroup;
3344 mem_cgroup_charge_statistics(memcg, false, -1);
3345 ClearPageCgroupUsed(pc);
3347 unlock_page_cgroup(pc);
3350 * When called from shmem_replace_page(), in some cases the
3351 * oldpage has already been charged, and in some cases not.
3356 if (PageSwapBacked(oldpage))
3357 type = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3360 * Even if newpage->mapping was NULL before starting replacement,
3361 * the newpage may be on LRU(or pagevec for LRU) already. We lock
3362 * LRU while we overwrite pc->mem_cgroup.
3364 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
3367 #ifdef CONFIG_DEBUG_VM
3368 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3370 struct page_cgroup *pc;
3372 pc = lookup_page_cgroup(page);
3374 * Can be NULL while feeding pages into the page allocator for
3375 * the first time, i.e. during boot or memory hotplug;
3376 * or when mem_cgroup_disabled().
3378 if (likely(pc) && PageCgroupUsed(pc))
3383 bool mem_cgroup_bad_page_check(struct page *page)
3385 if (mem_cgroup_disabled())
3388 return lookup_page_cgroup_used(page) != NULL;
3391 void mem_cgroup_print_bad_page(struct page *page)
3393 struct page_cgroup *pc;
3395 pc = lookup_page_cgroup_used(page);
3397 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
3398 pc, pc->flags, pc->mem_cgroup);
3403 static DEFINE_MUTEX(set_limit_mutex);
3405 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3406 unsigned long long val)
3409 u64 memswlimit, memlimit;
3411 int children = mem_cgroup_count_children(memcg);
3412 u64 curusage, oldusage;
3416 * For keeping hierarchical_reclaim simple, how long we should retry
3417 * is depends on callers. We set our retry-count to be function
3418 * of # of children which we should visit in this loop.
3420 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3422 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3425 while (retry_count) {
3426 if (signal_pending(current)) {
3431 * Rather than hide all in some function, I do this in
3432 * open coded manner. You see what this really does.
3433 * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3435 mutex_lock(&set_limit_mutex);
3436 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3437 if (memswlimit < val) {
3439 mutex_unlock(&set_limit_mutex);
3443 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3447 ret = res_counter_set_limit(&memcg->res, val);
3449 if (memswlimit == val)
3450 memcg->memsw_is_minimum = true;
3452 memcg->memsw_is_minimum = false;
3454 mutex_unlock(&set_limit_mutex);
3459 mem_cgroup_reclaim(memcg, GFP_KERNEL,
3460 MEM_CGROUP_RECLAIM_SHRINK);
3461 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3462 /* Usage is reduced ? */
3463 if (curusage >= oldusage)
3466 oldusage = curusage;
3468 if (!ret && enlarge)
3469 memcg_oom_recover(memcg);
3474 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3475 unsigned long long val)
3478 u64 memlimit, memswlimit, oldusage, curusage;
3479 int children = mem_cgroup_count_children(memcg);
3483 /* see mem_cgroup_resize_res_limit */
3484 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3485 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3486 while (retry_count) {
3487 if (signal_pending(current)) {
3492 * Rather than hide all in some function, I do this in
3493 * open coded manner. You see what this really does.
3494 * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3496 mutex_lock(&set_limit_mutex);
3497 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3498 if (memlimit > val) {
3500 mutex_unlock(&set_limit_mutex);
3503 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3504 if (memswlimit < val)
3506 ret = res_counter_set_limit(&memcg->memsw, val);
3508 if (memlimit == val)
3509 memcg->memsw_is_minimum = true;
3511 memcg->memsw_is_minimum = false;
3513 mutex_unlock(&set_limit_mutex);
3518 mem_cgroup_reclaim(memcg, GFP_KERNEL,
3519 MEM_CGROUP_RECLAIM_NOSWAP |
3520 MEM_CGROUP_RECLAIM_SHRINK);
3521 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3522 /* Usage is reduced ? */
3523 if (curusage >= oldusage)
3526 oldusage = curusage;
3528 if (!ret && enlarge)
3529 memcg_oom_recover(memcg);
3533 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3535 unsigned long *total_scanned)
3537 unsigned long nr_reclaimed = 0;
3538 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3539 unsigned long reclaimed;
3541 struct mem_cgroup_tree_per_zone *mctz;
3542 unsigned long long excess;
3543 unsigned long nr_scanned;
3548 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3550 * This loop can run a while, specially if mem_cgroup's continuously
3551 * keep exceeding their soft limit and putting the system under
3558 mz = mem_cgroup_largest_soft_limit_node(mctz);
3563 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
3564 gfp_mask, &nr_scanned);
3565 nr_reclaimed += reclaimed;
3566 *total_scanned += nr_scanned;
3567 spin_lock(&mctz->lock);
3570 * If we failed to reclaim anything from this memory cgroup
3571 * it is time to move on to the next cgroup
3577 * Loop until we find yet another one.
3579 * By the time we get the soft_limit lock
3580 * again, someone might have aded the
3581 * group back on the RB tree. Iterate to
3582 * make sure we get a different mem.
3583 * mem_cgroup_largest_soft_limit_node returns
3584 * NULL if no other cgroup is present on
3588 __mem_cgroup_largest_soft_limit_node(mctz);
3590 css_put(&next_mz->memcg->css);
3591 else /* next_mz == NULL or other memcg */
3595 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
3596 excess = res_counter_soft_limit_excess(&mz->memcg->res);
3598 * One school of thought says that we should not add
3599 * back the node to the tree if reclaim returns 0.
3600 * But our reclaim could return 0, simply because due
3601 * to priority we are exposing a smaller subset of
3602 * memory to reclaim from. Consider this as a longer
3605 /* If excess == 0, no tree ops */
3606 __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess);
3607 spin_unlock(&mctz->lock);
3608 css_put(&mz->memcg->css);
3611 * Could not reclaim anything and there are no more
3612 * mem cgroups to try or we seem to be looping without
3613 * reclaiming anything.
3615 if (!nr_reclaimed &&
3617 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3619 } while (!nr_reclaimed);
3621 css_put(&next_mz->memcg->css);
3622 return nr_reclaimed;
3626 * This routine traverse page_cgroup in given list and drop them all.
3627 * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3629 static int mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
3630 int node, int zid, enum lru_list lru)
3632 struct mem_cgroup_per_zone *mz;
3633 unsigned long flags, loop;
3634 struct list_head *list;
3639 zone = &NODE_DATA(node)->node_zones[zid];
3640 mz = mem_cgroup_zoneinfo(memcg, node, zid);
3641 list = &mz->lruvec.lists[lru];
3643 loop = mz->lru_size[lru];
3644 /* give some margin against EBUSY etc...*/
3648 struct page_cgroup *pc;
3652 spin_lock_irqsave(&zone->lru_lock, flags);
3653 if (list_empty(list)) {
3654 spin_unlock_irqrestore(&zone->lru_lock, flags);
3657 page = list_entry(list->prev, struct page, lru);
3659 list_move(&page->lru, list);
3661 spin_unlock_irqrestore(&zone->lru_lock, flags);
3664 spin_unlock_irqrestore(&zone->lru_lock, flags);
3666 pc = lookup_page_cgroup(page);
3668 ret = mem_cgroup_move_parent(page, pc, memcg, GFP_KERNEL);
3669 if (ret == -ENOMEM || ret == -EINTR)
3672 if (ret == -EBUSY || ret == -EINVAL) {
3673 /* found lock contention or "pc" is obsolete. */
3680 if (!ret && !list_empty(list))
3686 * make mem_cgroup's charge to be 0 if there is no task.
3687 * This enables deleting this mem_cgroup.
3689 static int mem_cgroup_force_empty(struct mem_cgroup *memcg, bool free_all)
3692 int node, zid, shrink;
3693 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3694 struct cgroup *cgrp = memcg->css.cgroup;
3696 css_get(&memcg->css);
3699 /* should free all ? */
3705 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3708 if (signal_pending(current))
3710 /* This is for making all *used* pages to be on LRU. */
3711 lru_add_drain_all();
3712 drain_all_stock_sync(memcg);
3714 mem_cgroup_start_move(memcg);
3715 for_each_node_state(node, N_HIGH_MEMORY) {
3716 for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3719 ret = mem_cgroup_force_empty_list(memcg,
3728 mem_cgroup_end_move(memcg);
3729 memcg_oom_recover(memcg);
3730 /* it seems parent cgroup doesn't have enough mem */
3734 /* "ret" should also be checked to ensure all lists are empty. */
3735 } while (res_counter_read_u64(&memcg->res, RES_USAGE) > 0 || ret);
3737 css_put(&memcg->css);
3741 /* returns EBUSY if there is a task or if we come here twice. */
3742 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3746 /* we call try-to-free pages for make this cgroup empty */
3747 lru_add_drain_all();
3748 /* try to free all pages in this cgroup */
3750 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
3753 if (signal_pending(current)) {
3757 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
3761 /* maybe some writeback is necessary */
3762 congestion_wait(BLK_RW_ASYNC, HZ/10);
3767 /* try move_account...there may be some *locked* pages. */
3771 static int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3773 return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3777 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3779 return mem_cgroup_from_cont(cont)->use_hierarchy;
3782 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3786 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3787 struct cgroup *parent = cont->parent;
3788 struct mem_cgroup *parent_memcg = NULL;
3791 parent_memcg = mem_cgroup_from_cont(parent);
3795 * If parent's use_hierarchy is set, we can't make any modifications
3796 * in the child subtrees. If it is unset, then the change can
3797 * occur, provided the current cgroup has no children.
3799 * For the root cgroup, parent_mem is NULL, we allow value to be
3800 * set if there are no children.
3802 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3803 (val == 1 || val == 0)) {
3804 if (list_empty(&cont->children))
3805 memcg->use_hierarchy = val;
3816 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
3817 enum mem_cgroup_stat_index idx)
3819 struct mem_cgroup *iter;
3822 /* Per-cpu values can be negative, use a signed accumulator */
3823 for_each_mem_cgroup_tree(iter, memcg)
3824 val += mem_cgroup_read_stat(iter, idx);
3826 if (val < 0) /* race ? */
3831 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3835 if (!mem_cgroup_is_root(memcg)) {
3837 return res_counter_read_u64(&memcg->res, RES_USAGE);
3839 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
3842 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
3843 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
3846 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
3848 return val << PAGE_SHIFT;
3851 static ssize_t mem_cgroup_read(struct cgroup *cont, struct cftype *cft,
3852 struct file *file, char __user *buf,
3853 size_t nbytes, loff_t *ppos)
3855 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3858 int type, name, len;
3860 type = MEMFILE_TYPE(cft->private);
3861 name = MEMFILE_ATTR(cft->private);
3863 if (!do_swap_account && type == _MEMSWAP)
3868 if (name == RES_USAGE)
3869 val = mem_cgroup_usage(memcg, false);
3871 val = res_counter_read_u64(&memcg->res, name);
3874 if (name == RES_USAGE)
3875 val = mem_cgroup_usage(memcg, true);
3877 val = res_counter_read_u64(&memcg->memsw, name);
3883 len = scnprintf(str, sizeof(str), "%llu\n", (unsigned long long)val);
3884 return simple_read_from_buffer(buf, nbytes, ppos, str, len);
3887 * The user of this function is...
3890 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
3893 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3895 unsigned long long val;
3898 type = MEMFILE_TYPE(cft->private);
3899 name = MEMFILE_ATTR(cft->private);
3901 if (!do_swap_account && type == _MEMSWAP)
3906 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3910 /* This function does all necessary parse...reuse it */
3911 ret = res_counter_memparse_write_strategy(buffer, &val);
3915 ret = mem_cgroup_resize_limit(memcg, val);
3917 ret = mem_cgroup_resize_memsw_limit(memcg, val);
3919 case RES_SOFT_LIMIT:
3920 ret = res_counter_memparse_write_strategy(buffer, &val);
3924 * For memsw, soft limits are hard to implement in terms
3925 * of semantics, for now, we support soft limits for
3926 * control without swap
3929 ret = res_counter_set_soft_limit(&memcg->res, val);
3934 ret = -EINVAL; /* should be BUG() ? */
3940 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
3941 unsigned long long *mem_limit, unsigned long long *memsw_limit)
3943 struct cgroup *cgroup;
3944 unsigned long long min_limit, min_memsw_limit, tmp;
3946 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3947 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3948 cgroup = memcg->css.cgroup;
3949 if (!memcg->use_hierarchy)
3952 while (cgroup->parent) {
3953 cgroup = cgroup->parent;
3954 memcg = mem_cgroup_from_cont(cgroup);
3955 if (!memcg->use_hierarchy)
3957 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
3958 min_limit = min(min_limit, tmp);
3959 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3960 min_memsw_limit = min(min_memsw_limit, tmp);
3963 *mem_limit = min_limit;
3964 *memsw_limit = min_memsw_limit;
3967 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
3969 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3972 type = MEMFILE_TYPE(event);
3973 name = MEMFILE_ATTR(event);
3975 if (!do_swap_account && type == _MEMSWAP)
3981 res_counter_reset_max(&memcg->res);
3983 res_counter_reset_max(&memcg->memsw);
3987 res_counter_reset_failcnt(&memcg->res);
3989 res_counter_reset_failcnt(&memcg->memsw);
3996 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
3999 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
4003 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4004 struct cftype *cft, u64 val)
4006 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4008 if (val >= (1 << NR_MOVE_TYPE))
4011 * We check this value several times in both in can_attach() and
4012 * attach(), so we need cgroup lock to prevent this value from being
4016 memcg->move_charge_at_immigrate = val;
4022 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4023 struct cftype *cft, u64 val)
4030 /* For read statistics */
4048 struct mcs_total_stat {
4049 s64 stat[NR_MCS_STAT];
4055 } memcg_stat_strings[NR_MCS_STAT] = {
4056 {"cache", "total_cache"},
4057 {"rss", "total_rss"},
4058 {"mapped_file", "total_mapped_file"},
4059 {"pgpgin", "total_pgpgin"},
4060 {"pgpgout", "total_pgpgout"},
4061 {"swap", "total_swap"},
4062 {"pgfault", "total_pgfault"},
4063 {"pgmajfault", "total_pgmajfault"},
4064 {"inactive_anon", "total_inactive_anon"},
4065 {"active_anon", "total_active_anon"},
4066 {"inactive_file", "total_inactive_file"},
4067 {"active_file", "total_active_file"},
4068 {"unevictable", "total_unevictable"}
4073 mem_cgroup_get_local_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4078 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_CACHE);
4079 s->stat[MCS_CACHE] += val * PAGE_SIZE;
4080 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_RSS);
4081 s->stat[MCS_RSS] += val * PAGE_SIZE;
4082 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_FILE_MAPPED);
4083 s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
4084 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGIN);
4085 s->stat[MCS_PGPGIN] += val;
4086 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGOUT);
4087 s->stat[MCS_PGPGOUT] += val;
4088 if (do_swap_account) {
4089 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
4090 s->stat[MCS_SWAP] += val * PAGE_SIZE;
4092 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGFAULT);
4093 s->stat[MCS_PGFAULT] += val;
4094 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGMAJFAULT);
4095 s->stat[MCS_PGMAJFAULT] += val;
4098 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_ANON));
4099 s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
4100 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_ANON));
4101 s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
4102 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_FILE));
4103 s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
4104 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_FILE));
4105 s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
4106 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
4107 s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
4111 mem_cgroup_get_total_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4113 struct mem_cgroup *iter;
4115 for_each_mem_cgroup_tree(iter, memcg)
4116 mem_cgroup_get_local_stat(iter, s);
4120 static int mem_control_numa_stat_show(struct seq_file *m, void *arg)
4123 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
4124 unsigned long node_nr;
4125 struct cgroup *cont = m->private;
4126 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4128 total_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL);
4129 seq_printf(m, "total=%lu", total_nr);
4130 for_each_node_state(nid, N_HIGH_MEMORY) {
4131 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL);
4132 seq_printf(m, " N%d=%lu", nid, node_nr);
4136 file_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_FILE);
4137 seq_printf(m, "file=%lu", file_nr);
4138 for_each_node_state(nid, N_HIGH_MEMORY) {
4139 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4141 seq_printf(m, " N%d=%lu", nid, node_nr);
4145 anon_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_ANON);
4146 seq_printf(m, "anon=%lu", anon_nr);
4147 for_each_node_state(nid, N_HIGH_MEMORY) {
4148 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4150 seq_printf(m, " N%d=%lu", nid, node_nr);
4154 unevictable_nr = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
4155 seq_printf(m, "unevictable=%lu", unevictable_nr);
4156 for_each_node_state(nid, N_HIGH_MEMORY) {
4157 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4158 BIT(LRU_UNEVICTABLE));
4159 seq_printf(m, " N%d=%lu", nid, node_nr);
4164 #endif /* CONFIG_NUMA */
4166 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
4167 struct cgroup_map_cb *cb)
4169 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4170 struct mcs_total_stat mystat;
4173 memset(&mystat, 0, sizeof(mystat));
4174 mem_cgroup_get_local_stat(memcg, &mystat);
4177 for (i = 0; i < NR_MCS_STAT; i++) {
4178 if (i == MCS_SWAP && !do_swap_account)
4180 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
4183 /* Hierarchical information */
4185 unsigned long long limit, memsw_limit;
4186 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
4187 cb->fill(cb, "hierarchical_memory_limit", limit);
4188 if (do_swap_account)
4189 cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
4192 memset(&mystat, 0, sizeof(mystat));
4193 mem_cgroup_get_total_stat(memcg, &mystat);
4194 for (i = 0; i < NR_MCS_STAT; i++) {
4195 if (i == MCS_SWAP && !do_swap_account)
4197 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
4200 #ifdef CONFIG_DEBUG_VM
4203 struct mem_cgroup_per_zone *mz;
4204 struct zone_reclaim_stat *rstat;
4205 unsigned long recent_rotated[2] = {0, 0};
4206 unsigned long recent_scanned[2] = {0, 0};
4208 for_each_online_node(nid)
4209 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4210 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
4211 rstat = &mz->lruvec.reclaim_stat;
4213 recent_rotated[0] += rstat->recent_rotated[0];
4214 recent_rotated[1] += rstat->recent_rotated[1];
4215 recent_scanned[0] += rstat->recent_scanned[0];
4216 recent_scanned[1] += rstat->recent_scanned[1];
4218 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
4219 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
4220 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
4221 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
4228 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
4230 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4232 return mem_cgroup_swappiness(memcg);
4235 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
4238 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4239 struct mem_cgroup *parent;
4244 if (cgrp->parent == NULL)
4247 parent = mem_cgroup_from_cont(cgrp->parent);
4251 /* If under hierarchy, only empty-root can set this value */
4252 if ((parent->use_hierarchy) ||
4253 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4258 memcg->swappiness = val;
4265 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4267 struct mem_cgroup_threshold_ary *t;
4273 t = rcu_dereference(memcg->thresholds.primary);
4275 t = rcu_dereference(memcg->memsw_thresholds.primary);
4280 usage = mem_cgroup_usage(memcg, swap);
4283 * current_threshold points to threshold just below usage.
4284 * If it's not true, a threshold was crossed after last
4285 * call of __mem_cgroup_threshold().
4287 i = t->current_threshold;
4290 * Iterate backward over array of thresholds starting from
4291 * current_threshold and check if a threshold is crossed.
4292 * If none of thresholds below usage is crossed, we read
4293 * only one element of the array here.
4295 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4296 eventfd_signal(t->entries[i].eventfd, 1);
4298 /* i = current_threshold + 1 */
4302 * Iterate forward over array of thresholds starting from
4303 * current_threshold+1 and check if a threshold is crossed.
4304 * If none of thresholds above usage is crossed, we read
4305 * only one element of the array here.
4307 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4308 eventfd_signal(t->entries[i].eventfd, 1);
4310 /* Update current_threshold */
4311 t->current_threshold = i - 1;
4316 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4319 __mem_cgroup_threshold(memcg, false);
4320 if (do_swap_account)
4321 __mem_cgroup_threshold(memcg, true);
4323 memcg = parent_mem_cgroup(memcg);
4327 static int compare_thresholds(const void *a, const void *b)
4329 const struct mem_cgroup_threshold *_a = a;
4330 const struct mem_cgroup_threshold *_b = b;
4332 return _a->threshold - _b->threshold;
4335 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4337 struct mem_cgroup_eventfd_list *ev;
4339 list_for_each_entry(ev, &memcg->oom_notify, list)
4340 eventfd_signal(ev->eventfd, 1);
4344 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4346 struct mem_cgroup *iter;
4348 for_each_mem_cgroup_tree(iter, memcg)
4349 mem_cgroup_oom_notify_cb(iter);
4352 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
4353 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4355 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4356 struct mem_cgroup_thresholds *thresholds;
4357 struct mem_cgroup_threshold_ary *new;
4358 int type = MEMFILE_TYPE(cft->private);
4359 u64 threshold, usage;
4362 ret = res_counter_memparse_write_strategy(args, &threshold);
4366 mutex_lock(&memcg->thresholds_lock);
4369 thresholds = &memcg->thresholds;
4370 else if (type == _MEMSWAP)
4371 thresholds = &memcg->memsw_thresholds;
4375 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4377 /* Check if a threshold crossed before adding a new one */
4378 if (thresholds->primary)
4379 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4381 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4383 /* Allocate memory for new array of thresholds */
4384 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4392 /* Copy thresholds (if any) to new array */
4393 if (thresholds->primary) {
4394 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4395 sizeof(struct mem_cgroup_threshold));
4398 /* Add new threshold */
4399 new->entries[size - 1].eventfd = eventfd;
4400 new->entries[size - 1].threshold = threshold;
4402 /* Sort thresholds. Registering of new threshold isn't time-critical */
4403 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4404 compare_thresholds, NULL);
4406 /* Find current threshold */
4407 new->current_threshold = -1;
4408 for (i = 0; i < size; i++) {
4409 if (new->entries[i].threshold < usage) {
4411 * new->current_threshold will not be used until
4412 * rcu_assign_pointer(), so it's safe to increment
4415 ++new->current_threshold;
4419 /* Free old spare buffer and save old primary buffer as spare */
4420 kfree(thresholds->spare);
4421 thresholds->spare = thresholds->primary;
4423 rcu_assign_pointer(thresholds->primary, new);
4425 /* To be sure that nobody uses thresholds */
4429 mutex_unlock(&memcg->thresholds_lock);
4434 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
4435 struct cftype *cft, struct eventfd_ctx *eventfd)
4437 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4438 struct mem_cgroup_thresholds *thresholds;
4439 struct mem_cgroup_threshold_ary *new;
4440 int type = MEMFILE_TYPE(cft->private);
4444 mutex_lock(&memcg->thresholds_lock);
4446 thresholds = &memcg->thresholds;
4447 else if (type == _MEMSWAP)
4448 thresholds = &memcg->memsw_thresholds;
4452 if (!thresholds->primary)
4455 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4457 /* Check if a threshold crossed before removing */
4458 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4460 /* Calculate new number of threshold */
4462 for (i = 0; i < thresholds->primary->size; i++) {
4463 if (thresholds->primary->entries[i].eventfd != eventfd)
4467 new = thresholds->spare;
4469 /* Set thresholds array to NULL if we don't have thresholds */
4478 /* Copy thresholds and find current threshold */
4479 new->current_threshold = -1;
4480 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4481 if (thresholds->primary->entries[i].eventfd == eventfd)
4484 new->entries[j] = thresholds->primary->entries[i];
4485 if (new->entries[j].threshold < usage) {
4487 * new->current_threshold will not be used
4488 * until rcu_assign_pointer(), so it's safe to increment
4491 ++new->current_threshold;
4497 /* Swap primary and spare array */
4498 thresholds->spare = thresholds->primary;
4499 /* If all events are unregistered, free the spare array */
4501 kfree(thresholds->spare);
4502 thresholds->spare = NULL;
4505 rcu_assign_pointer(thresholds->primary, new);
4507 /* To be sure that nobody uses thresholds */
4510 mutex_unlock(&memcg->thresholds_lock);
4513 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4514 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4516 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4517 struct mem_cgroup_eventfd_list *event;
4518 int type = MEMFILE_TYPE(cft->private);
4520 BUG_ON(type != _OOM_TYPE);
4521 event = kmalloc(sizeof(*event), GFP_KERNEL);
4525 spin_lock(&memcg_oom_lock);
4527 event->eventfd = eventfd;
4528 list_add(&event->list, &memcg->oom_notify);
4530 /* already in OOM ? */
4531 if (atomic_read(&memcg->under_oom))
4532 eventfd_signal(eventfd, 1);
4533 spin_unlock(&memcg_oom_lock);
4538 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4539 struct cftype *cft, struct eventfd_ctx *eventfd)
4541 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4542 struct mem_cgroup_eventfd_list *ev, *tmp;
4543 int type = MEMFILE_TYPE(cft->private);
4545 BUG_ON(type != _OOM_TYPE);
4547 spin_lock(&memcg_oom_lock);
4549 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4550 if (ev->eventfd == eventfd) {
4551 list_del(&ev->list);
4556 spin_unlock(&memcg_oom_lock);
4559 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4560 struct cftype *cft, struct cgroup_map_cb *cb)
4562 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4564 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
4566 if (atomic_read(&memcg->under_oom))
4567 cb->fill(cb, "under_oom", 1);
4569 cb->fill(cb, "under_oom", 0);
4573 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4574 struct cftype *cft, u64 val)
4576 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4577 struct mem_cgroup *parent;
4579 /* cannot set to root cgroup and only 0 and 1 are allowed */
4580 if (!cgrp->parent || !((val == 0) || (val == 1)))
4583 parent = mem_cgroup_from_cont(cgrp->parent);
4586 /* oom-kill-disable is a flag for subhierarchy. */
4587 if ((parent->use_hierarchy) ||
4588 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4592 memcg->oom_kill_disable = val;
4594 memcg_oom_recover(memcg);
4600 static const struct file_operations mem_control_numa_stat_file_operations = {
4602 .llseek = seq_lseek,
4603 .release = single_release,
4606 static int mem_control_numa_stat_open(struct inode *unused, struct file *file)
4608 struct cgroup *cont = file->f_dentry->d_parent->d_fsdata;
4610 file->f_op = &mem_control_numa_stat_file_operations;
4611 return single_open(file, mem_control_numa_stat_show, cont);
4613 #endif /* CONFIG_NUMA */
4615 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_KMEM
4616 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
4618 return mem_cgroup_sockets_init(memcg, ss);
4621 static void kmem_cgroup_destroy(struct mem_cgroup *memcg)
4623 mem_cgroup_sockets_destroy(memcg);
4626 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
4631 static void kmem_cgroup_destroy(struct mem_cgroup *memcg)
4636 static struct cftype mem_cgroup_files[] = {
4638 .name = "usage_in_bytes",
4639 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4640 .read = mem_cgroup_read,
4641 .register_event = mem_cgroup_usage_register_event,
4642 .unregister_event = mem_cgroup_usage_unregister_event,
4645 .name = "max_usage_in_bytes",
4646 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4647 .trigger = mem_cgroup_reset,
4648 .read = mem_cgroup_read,
4651 .name = "limit_in_bytes",
4652 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4653 .write_string = mem_cgroup_write,
4654 .read = mem_cgroup_read,
4657 .name = "soft_limit_in_bytes",
4658 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4659 .write_string = mem_cgroup_write,
4660 .read = mem_cgroup_read,
4664 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4665 .trigger = mem_cgroup_reset,
4666 .read = mem_cgroup_read,
4670 .read_map = mem_control_stat_show,
4673 .name = "force_empty",
4674 .trigger = mem_cgroup_force_empty_write,
4677 .name = "use_hierarchy",
4678 .write_u64 = mem_cgroup_hierarchy_write,
4679 .read_u64 = mem_cgroup_hierarchy_read,
4682 .name = "swappiness",
4683 .read_u64 = mem_cgroup_swappiness_read,
4684 .write_u64 = mem_cgroup_swappiness_write,
4687 .name = "move_charge_at_immigrate",
4688 .read_u64 = mem_cgroup_move_charge_read,
4689 .write_u64 = mem_cgroup_move_charge_write,
4692 .name = "oom_control",
4693 .read_map = mem_cgroup_oom_control_read,
4694 .write_u64 = mem_cgroup_oom_control_write,
4695 .register_event = mem_cgroup_oom_register_event,
4696 .unregister_event = mem_cgroup_oom_unregister_event,
4697 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4701 .name = "numa_stat",
4702 .open = mem_control_numa_stat_open,
4706 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4708 .name = "memsw.usage_in_bytes",
4709 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4710 .read = mem_cgroup_read,
4711 .register_event = mem_cgroup_usage_register_event,
4712 .unregister_event = mem_cgroup_usage_unregister_event,
4715 .name = "memsw.max_usage_in_bytes",
4716 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4717 .trigger = mem_cgroup_reset,
4718 .read = mem_cgroup_read,
4721 .name = "memsw.limit_in_bytes",
4722 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4723 .write_string = mem_cgroup_write,
4724 .read = mem_cgroup_read,
4727 .name = "memsw.failcnt",
4728 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4729 .trigger = mem_cgroup_reset,
4730 .read = mem_cgroup_read,
4733 { }, /* terminate */
4736 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4738 struct mem_cgroup_per_node *pn;
4739 struct mem_cgroup_per_zone *mz;
4741 int zone, tmp = node;
4743 * This routine is called against possible nodes.
4744 * But it's BUG to call kmalloc() against offline node.
4746 * TODO: this routine can waste much memory for nodes which will
4747 * never be onlined. It's better to use memory hotplug callback
4750 if (!node_state(node, N_NORMAL_MEMORY))
4752 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4756 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4757 mz = &pn->zoneinfo[zone];
4759 INIT_LIST_HEAD(&mz->lruvec.lists[lru]);
4760 mz->usage_in_excess = 0;
4761 mz->on_tree = false;
4764 memcg->info.nodeinfo[node] = pn;
4768 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4770 kfree(memcg->info.nodeinfo[node]);
4773 static struct mem_cgroup *mem_cgroup_alloc(void)
4775 struct mem_cgroup *memcg;
4776 int size = sizeof(struct mem_cgroup);
4778 /* Can be very big if MAX_NUMNODES is very big */
4779 if (size < PAGE_SIZE)
4780 memcg = kzalloc(size, GFP_KERNEL);
4782 memcg = vzalloc(size);
4787 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4790 spin_lock_init(&memcg->pcp_counter_lock);
4794 if (size < PAGE_SIZE)
4802 * Helpers for freeing a vzalloc()ed mem_cgroup by RCU,
4803 * but in process context. The work_freeing structure is overlaid
4804 * on the rcu_freeing structure, which itself is overlaid on memsw.
4806 static void vfree_work(struct work_struct *work)
4808 struct mem_cgroup *memcg;
4810 memcg = container_of(work, struct mem_cgroup, work_freeing);
4813 static void vfree_rcu(struct rcu_head *rcu_head)
4815 struct mem_cgroup *memcg;
4817 memcg = container_of(rcu_head, struct mem_cgroup, rcu_freeing);
4818 INIT_WORK(&memcg->work_freeing, vfree_work);
4819 schedule_work(&memcg->work_freeing);
4823 * At destroying mem_cgroup, references from swap_cgroup can remain.
4824 * (scanning all at force_empty is too costly...)
4826 * Instead of clearing all references at force_empty, we remember
4827 * the number of reference from swap_cgroup and free mem_cgroup when
4828 * it goes down to 0.
4830 * Removal of cgroup itself succeeds regardless of refs from swap.
4833 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4837 mem_cgroup_remove_from_trees(memcg);
4838 free_css_id(&mem_cgroup_subsys, &memcg->css);
4841 free_mem_cgroup_per_zone_info(memcg, node);
4843 free_percpu(memcg->stat);
4844 if (sizeof(struct mem_cgroup) < PAGE_SIZE)
4845 kfree_rcu(memcg, rcu_freeing);
4847 call_rcu(&memcg->rcu_freeing, vfree_rcu);
4850 static void mem_cgroup_get(struct mem_cgroup *memcg)
4852 atomic_inc(&memcg->refcnt);
4855 static void __mem_cgroup_put(struct mem_cgroup *memcg, int count)
4857 if (atomic_sub_and_test(count, &memcg->refcnt)) {
4858 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4859 __mem_cgroup_free(memcg);
4861 mem_cgroup_put(parent);
4865 static void mem_cgroup_put(struct mem_cgroup *memcg)
4867 __mem_cgroup_put(memcg, 1);
4871 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4873 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4875 if (!memcg->res.parent)
4877 return mem_cgroup_from_res_counter(memcg->res.parent, res);
4879 EXPORT_SYMBOL(parent_mem_cgroup);
4881 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4882 static void __init enable_swap_cgroup(void)
4884 if (!mem_cgroup_disabled() && really_do_swap_account)
4885 do_swap_account = 1;
4888 static void __init enable_swap_cgroup(void)
4893 static int mem_cgroup_soft_limit_tree_init(void)
4895 struct mem_cgroup_tree_per_node *rtpn;
4896 struct mem_cgroup_tree_per_zone *rtpz;
4897 int tmp, node, zone;
4899 for_each_node(node) {
4901 if (!node_state(node, N_NORMAL_MEMORY))
4903 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
4907 soft_limit_tree.rb_tree_per_node[node] = rtpn;
4909 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4910 rtpz = &rtpn->rb_tree_per_zone[zone];
4911 rtpz->rb_root = RB_ROOT;
4912 spin_lock_init(&rtpz->lock);
4918 for_each_node(node) {
4919 if (!soft_limit_tree.rb_tree_per_node[node])
4921 kfree(soft_limit_tree.rb_tree_per_node[node]);
4922 soft_limit_tree.rb_tree_per_node[node] = NULL;
4928 static struct cgroup_subsys_state * __ref
4929 mem_cgroup_create(struct cgroup *cont)
4931 struct mem_cgroup *memcg, *parent;
4932 long error = -ENOMEM;
4935 memcg = mem_cgroup_alloc();
4937 return ERR_PTR(error);
4940 if (alloc_mem_cgroup_per_zone_info(memcg, node))
4944 if (cont->parent == NULL) {
4946 enable_swap_cgroup();
4948 if (mem_cgroup_soft_limit_tree_init())
4950 root_mem_cgroup = memcg;
4951 for_each_possible_cpu(cpu) {
4952 struct memcg_stock_pcp *stock =
4953 &per_cpu(memcg_stock, cpu);
4954 INIT_WORK(&stock->work, drain_local_stock);
4956 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
4958 parent = mem_cgroup_from_cont(cont->parent);
4959 memcg->use_hierarchy = parent->use_hierarchy;
4960 memcg->oom_kill_disable = parent->oom_kill_disable;
4963 if (parent && parent->use_hierarchy) {
4964 res_counter_init(&memcg->res, &parent->res);
4965 res_counter_init(&memcg->memsw, &parent->memsw);
4967 * We increment refcnt of the parent to ensure that we can
4968 * safely access it on res_counter_charge/uncharge.
4969 * This refcnt will be decremented when freeing this
4970 * mem_cgroup(see mem_cgroup_put).
4972 mem_cgroup_get(parent);
4974 res_counter_init(&memcg->res, NULL);
4975 res_counter_init(&memcg->memsw, NULL);
4977 memcg->last_scanned_node = MAX_NUMNODES;
4978 INIT_LIST_HEAD(&memcg->oom_notify);
4981 memcg->swappiness = mem_cgroup_swappiness(parent);
4982 atomic_set(&memcg->refcnt, 1);
4983 memcg->move_charge_at_immigrate = 0;
4984 mutex_init(&memcg->thresholds_lock);
4985 spin_lock_init(&memcg->move_lock);
4987 error = memcg_init_kmem(memcg, &mem_cgroup_subsys);
4990 * We call put now because our (and parent's) refcnts
4991 * are already in place. mem_cgroup_put() will internally
4992 * call __mem_cgroup_free, so return directly
4994 mem_cgroup_put(memcg);
4995 return ERR_PTR(error);
4999 __mem_cgroup_free(memcg);
5000 return ERR_PTR(error);
5003 static int mem_cgroup_pre_destroy(struct cgroup *cont)
5005 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5007 return mem_cgroup_force_empty(memcg, false);
5010 static void mem_cgroup_destroy(struct cgroup *cont)
5012 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5014 kmem_cgroup_destroy(memcg);
5016 mem_cgroup_put(memcg);
5020 /* Handlers for move charge at task migration. */
5021 #define PRECHARGE_COUNT_AT_ONCE 256
5022 static int mem_cgroup_do_precharge(unsigned long count)
5025 int batch_count = PRECHARGE_COUNT_AT_ONCE;
5026 struct mem_cgroup *memcg = mc.to;
5028 if (mem_cgroup_is_root(memcg)) {
5029 mc.precharge += count;
5030 /* we don't need css_get for root */
5033 /* try to charge at once */
5035 struct res_counter *dummy;
5037 * "memcg" cannot be under rmdir() because we've already checked
5038 * by cgroup_lock_live_cgroup() that it is not removed and we
5039 * are still under the same cgroup_mutex. So we can postpone
5042 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
5044 if (do_swap_account && res_counter_charge(&memcg->memsw,
5045 PAGE_SIZE * count, &dummy)) {
5046 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
5049 mc.precharge += count;
5053 /* fall back to one by one charge */
5055 if (signal_pending(current)) {
5059 if (!batch_count--) {
5060 batch_count = PRECHARGE_COUNT_AT_ONCE;
5063 ret = __mem_cgroup_try_charge(NULL,
5064 GFP_KERNEL, 1, &memcg, false);
5066 /* mem_cgroup_clear_mc() will do uncharge later */
5074 * get_mctgt_type - get target type of moving charge
5075 * @vma: the vma the pte to be checked belongs
5076 * @addr: the address corresponding to the pte to be checked
5077 * @ptent: the pte to be checked
5078 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5081 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5082 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5083 * move charge. if @target is not NULL, the page is stored in target->page
5084 * with extra refcnt got(Callers should handle it).
5085 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5086 * target for charge migration. if @target is not NULL, the entry is stored
5089 * Called with pte lock held.
5096 enum mc_target_type {
5102 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5103 unsigned long addr, pte_t ptent)
5105 struct page *page = vm_normal_page(vma, addr, ptent);
5107 if (!page || !page_mapped(page))
5109 if (PageAnon(page)) {
5110 /* we don't move shared anon */
5113 } else if (!move_file())
5114 /* we ignore mapcount for file pages */
5116 if (!get_page_unless_zero(page))
5123 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5124 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5126 struct page *page = NULL;
5127 swp_entry_t ent = pte_to_swp_entry(ptent);
5129 if (!move_anon() || non_swap_entry(ent))
5132 * Because lookup_swap_cache() updates some statistics counter,
5133 * we call find_get_page() with swapper_space directly.
5135 page = find_get_page(&swapper_space, ent.val);
5136 if (do_swap_account)
5137 entry->val = ent.val;
5142 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5143 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5149 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5150 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5152 struct page *page = NULL;
5153 struct address_space *mapping;
5156 if (!vma->vm_file) /* anonymous vma */
5161 mapping = vma->vm_file->f_mapping;
5162 if (pte_none(ptent))
5163 pgoff = linear_page_index(vma, addr);
5164 else /* pte_file(ptent) is true */
5165 pgoff = pte_to_pgoff(ptent);
5167 /* page is moved even if it's not RSS of this task(page-faulted). */
5168 page = find_get_page(mapping, pgoff);
5171 /* shmem/tmpfs may report page out on swap: account for that too. */
5172 if (radix_tree_exceptional_entry(page)) {
5173 swp_entry_t swap = radix_to_swp_entry(page);
5174 if (do_swap_account)
5176 page = find_get_page(&swapper_space, swap.val);
5182 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5183 unsigned long addr, pte_t ptent, union mc_target *target)
5185 struct page *page = NULL;
5186 struct page_cgroup *pc;
5187 enum mc_target_type ret = MC_TARGET_NONE;
5188 swp_entry_t ent = { .val = 0 };
5190 if (pte_present(ptent))
5191 page = mc_handle_present_pte(vma, addr, ptent);
5192 else if (is_swap_pte(ptent))
5193 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5194 else if (pte_none(ptent) || pte_file(ptent))
5195 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5197 if (!page && !ent.val)
5200 pc = lookup_page_cgroup(page);
5202 * Do only loose check w/o page_cgroup lock.
5203 * mem_cgroup_move_account() checks the pc is valid or not under
5206 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5207 ret = MC_TARGET_PAGE;
5209 target->page = page;
5211 if (!ret || !target)
5214 /* There is a swap entry and a page doesn't exist or isn't charged */
5215 if (ent.val && !ret &&
5216 css_id(&mc.from->css) == lookup_swap_cgroup_id(ent)) {
5217 ret = MC_TARGET_SWAP;
5224 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5226 * We don't consider swapping or file mapped pages because THP does not
5227 * support them for now.
5228 * Caller should make sure that pmd_trans_huge(pmd) is true.
5230 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5231 unsigned long addr, pmd_t pmd, union mc_target *target)
5233 struct page *page = NULL;
5234 struct page_cgroup *pc;
5235 enum mc_target_type ret = MC_TARGET_NONE;
5237 page = pmd_page(pmd);
5238 VM_BUG_ON(!page || !PageHead(page));
5241 pc = lookup_page_cgroup(page);
5242 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5243 ret = MC_TARGET_PAGE;
5246 target->page = page;
5252 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5253 unsigned long addr, pmd_t pmd, union mc_target *target)
5255 return MC_TARGET_NONE;
5259 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5260 unsigned long addr, unsigned long end,
5261 struct mm_walk *walk)
5263 struct vm_area_struct *vma = walk->private;
5267 if (pmd_trans_huge_lock(pmd, vma) == 1) {
5268 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5269 mc.precharge += HPAGE_PMD_NR;
5270 spin_unlock(&vma->vm_mm->page_table_lock);
5274 if (pmd_trans_unstable(pmd))
5276 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5277 for (; addr != end; pte++, addr += PAGE_SIZE)
5278 if (get_mctgt_type(vma, addr, *pte, NULL))
5279 mc.precharge++; /* increment precharge temporarily */
5280 pte_unmap_unlock(pte - 1, ptl);
5286 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5288 unsigned long precharge;
5289 struct vm_area_struct *vma;
5291 down_read(&mm->mmap_sem);
5292 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5293 struct mm_walk mem_cgroup_count_precharge_walk = {
5294 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5298 if (is_vm_hugetlb_page(vma))
5300 walk_page_range(vma->vm_start, vma->vm_end,
5301 &mem_cgroup_count_precharge_walk);
5303 up_read(&mm->mmap_sem);
5305 precharge = mc.precharge;
5311 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5313 unsigned long precharge = mem_cgroup_count_precharge(mm);
5315 VM_BUG_ON(mc.moving_task);
5316 mc.moving_task = current;
5317 return mem_cgroup_do_precharge(precharge);
5320 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5321 static void __mem_cgroup_clear_mc(void)
5323 struct mem_cgroup *from = mc.from;
5324 struct mem_cgroup *to = mc.to;
5326 /* we must uncharge all the leftover precharges from mc.to */
5328 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
5332 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5333 * we must uncharge here.
5335 if (mc.moved_charge) {
5336 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
5337 mc.moved_charge = 0;
5339 /* we must fixup refcnts and charges */
5340 if (mc.moved_swap) {
5341 /* uncharge swap account from the old cgroup */
5342 if (!mem_cgroup_is_root(mc.from))
5343 res_counter_uncharge(&mc.from->memsw,
5344 PAGE_SIZE * mc.moved_swap);
5345 __mem_cgroup_put(mc.from, mc.moved_swap);
5347 if (!mem_cgroup_is_root(mc.to)) {
5349 * we charged both to->res and to->memsw, so we should
5352 res_counter_uncharge(&mc.to->res,
5353 PAGE_SIZE * mc.moved_swap);
5355 /* we've already done mem_cgroup_get(mc.to) */
5358 memcg_oom_recover(from);
5359 memcg_oom_recover(to);
5360 wake_up_all(&mc.waitq);
5363 static void mem_cgroup_clear_mc(void)
5365 struct mem_cgroup *from = mc.from;
5368 * we must clear moving_task before waking up waiters at the end of
5371 mc.moving_task = NULL;
5372 __mem_cgroup_clear_mc();
5373 spin_lock(&mc.lock);
5376 spin_unlock(&mc.lock);
5377 mem_cgroup_end_move(from);
5380 static int mem_cgroup_can_attach(struct cgroup *cgroup,
5381 struct cgroup_taskset *tset)
5383 struct task_struct *p = cgroup_taskset_first(tset);
5385 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup);
5387 if (memcg->move_charge_at_immigrate) {
5388 struct mm_struct *mm;
5389 struct mem_cgroup *from = mem_cgroup_from_task(p);
5391 VM_BUG_ON(from == memcg);
5393 mm = get_task_mm(p);
5396 /* We move charges only when we move a owner of the mm */
5397 if (mm->owner == p) {
5400 VM_BUG_ON(mc.precharge);
5401 VM_BUG_ON(mc.moved_charge);
5402 VM_BUG_ON(mc.moved_swap);
5403 mem_cgroup_start_move(from);
5404 spin_lock(&mc.lock);
5407 spin_unlock(&mc.lock);
5408 /* We set mc.moving_task later */
5410 ret = mem_cgroup_precharge_mc(mm);
5412 mem_cgroup_clear_mc();
5419 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
5420 struct cgroup_taskset *tset)
5422 mem_cgroup_clear_mc();
5425 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5426 unsigned long addr, unsigned long end,
5427 struct mm_walk *walk)
5430 struct vm_area_struct *vma = walk->private;
5433 enum mc_target_type target_type;
5434 union mc_target target;
5436 struct page_cgroup *pc;
5439 * We don't take compound_lock() here but no race with splitting thp
5441 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
5442 * under splitting, which means there's no concurrent thp split,
5443 * - if another thread runs into split_huge_page() just after we
5444 * entered this if-block, the thread must wait for page table lock
5445 * to be unlocked in __split_huge_page_splitting(), where the main
5446 * part of thp split is not executed yet.
5448 if (pmd_trans_huge_lock(pmd, vma) == 1) {
5449 if (mc.precharge < HPAGE_PMD_NR) {
5450 spin_unlock(&vma->vm_mm->page_table_lock);
5453 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5454 if (target_type == MC_TARGET_PAGE) {
5456 if (!isolate_lru_page(page)) {
5457 pc = lookup_page_cgroup(page);
5458 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
5461 mc.precharge -= HPAGE_PMD_NR;
5462 mc.moved_charge += HPAGE_PMD_NR;
5464 putback_lru_page(page);
5468 spin_unlock(&vma->vm_mm->page_table_lock);
5472 if (pmd_trans_unstable(pmd))
5475 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5476 for (; addr != end; addr += PAGE_SIZE) {
5477 pte_t ptent = *(pte++);
5483 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5484 case MC_TARGET_PAGE:
5486 if (isolate_lru_page(page))
5488 pc = lookup_page_cgroup(page);
5489 if (!mem_cgroup_move_account(page, 1, pc,
5490 mc.from, mc.to, false)) {
5492 /* we uncharge from mc.from later. */
5495 putback_lru_page(page);
5496 put: /* get_mctgt_type() gets the page */
5499 case MC_TARGET_SWAP:
5501 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5503 /* we fixup refcnts and charges later. */
5511 pte_unmap_unlock(pte - 1, ptl);
5516 * We have consumed all precharges we got in can_attach().
5517 * We try charge one by one, but don't do any additional
5518 * charges to mc.to if we have failed in charge once in attach()
5521 ret = mem_cgroup_do_precharge(1);
5529 static void mem_cgroup_move_charge(struct mm_struct *mm)
5531 struct vm_area_struct *vma;
5533 lru_add_drain_all();
5535 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5537 * Someone who are holding the mmap_sem might be waiting in
5538 * waitq. So we cancel all extra charges, wake up all waiters,
5539 * and retry. Because we cancel precharges, we might not be able
5540 * to move enough charges, but moving charge is a best-effort
5541 * feature anyway, so it wouldn't be a big problem.
5543 __mem_cgroup_clear_mc();
5547 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5549 struct mm_walk mem_cgroup_move_charge_walk = {
5550 .pmd_entry = mem_cgroup_move_charge_pte_range,
5554 if (is_vm_hugetlb_page(vma))
5556 ret = walk_page_range(vma->vm_start, vma->vm_end,
5557 &mem_cgroup_move_charge_walk);
5560 * means we have consumed all precharges and failed in
5561 * doing additional charge. Just abandon here.
5565 up_read(&mm->mmap_sem);
5568 static void mem_cgroup_move_task(struct cgroup *cont,
5569 struct cgroup_taskset *tset)
5571 struct task_struct *p = cgroup_taskset_first(tset);
5572 struct mm_struct *mm = get_task_mm(p);
5576 mem_cgroup_move_charge(mm);
5580 mem_cgroup_clear_mc();
5582 #else /* !CONFIG_MMU */
5583 static int mem_cgroup_can_attach(struct cgroup *cgroup,
5584 struct cgroup_taskset *tset)
5588 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
5589 struct cgroup_taskset *tset)
5592 static void mem_cgroup_move_task(struct cgroup *cont,
5593 struct cgroup_taskset *tset)
5598 struct cgroup_subsys mem_cgroup_subsys = {
5600 .subsys_id = mem_cgroup_subsys_id,
5601 .create = mem_cgroup_create,
5602 .pre_destroy = mem_cgroup_pre_destroy,
5603 .destroy = mem_cgroup_destroy,
5604 .can_attach = mem_cgroup_can_attach,
5605 .cancel_attach = mem_cgroup_cancel_attach,
5606 .attach = mem_cgroup_move_task,
5607 .base_cftypes = mem_cgroup_files,
5610 .__DEPRECATED_clear_css_refs = true,
5613 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5614 static int __init enable_swap_account(char *s)
5616 /* consider enabled if no parameter or 1 is given */
5617 if (!strcmp(s, "1"))
5618 really_do_swap_account = 1;
5619 else if (!strcmp(s, "0"))
5620 really_do_swap_account = 0;
5623 __setup("swapaccount=", enable_swap_account);