1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * This program is free software; you can redistribute it and/or modify
14 * it under the terms of the GNU General Public License as published by
15 * the Free Software Foundation; either version 2 of the License, or
16 * (at your option) any later version.
18 * This program is distributed in the hope that it will be useful,
19 * but WITHOUT ANY WARRANTY; without even the implied warranty of
20 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
21 * GNU General Public License for more details.
24 #include <linux/res_counter.h>
25 #include <linux/memcontrol.h>
26 #include <linux/cgroup.h>
28 #include <linux/hugetlb.h>
29 #include <linux/pagemap.h>
30 #include <linux/smp.h>
31 #include <linux/page-flags.h>
32 #include <linux/backing-dev.h>
33 #include <linux/bit_spinlock.h>
34 #include <linux/rcupdate.h>
35 #include <linux/limits.h>
36 #include <linux/export.h>
37 #include <linux/mutex.h>
38 #include <linux/rbtree.h>
39 #include <linux/slab.h>
40 #include <linux/swap.h>
41 #include <linux/swapops.h>
42 #include <linux/spinlock.h>
43 #include <linux/eventfd.h>
44 #include <linux/sort.h>
46 #include <linux/seq_file.h>
47 #include <linux/vmalloc.h>
48 #include <linux/mm_inline.h>
49 #include <linux/page_cgroup.h>
50 #include <linux/cpu.h>
51 #include <linux/oom.h>
54 #include <net/tcp_memcontrol.h>
56 #include <asm/uaccess.h>
58 #include <trace/events/vmscan.h>
60 struct cgroup_subsys mem_cgroup_subsys __read_mostly;
61 #define MEM_CGROUP_RECLAIM_RETRIES 5
62 struct mem_cgroup *root_mem_cgroup __read_mostly;
64 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
65 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
66 int do_swap_account __read_mostly;
68 /* for remember boot option*/
69 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP_ENABLED
70 static int really_do_swap_account __initdata = 1;
72 static int really_do_swap_account __initdata = 0;
76 #define do_swap_account (0)
81 * Statistics for memory cgroup.
83 enum mem_cgroup_stat_index {
85 * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
87 MEM_CGROUP_STAT_CACHE, /* # of pages charged as cache */
88 MEM_CGROUP_STAT_RSS, /* # of pages charged as anon rss */
89 MEM_CGROUP_STAT_FILE_MAPPED, /* # of pages charged as file rss */
90 MEM_CGROUP_STAT_SWAPOUT, /* # of pages, swapped out */
91 MEM_CGROUP_STAT_DATA, /* end of data requires synchronization */
92 MEM_CGROUP_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 *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 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 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 bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1842 struct oom_wait_info owait;
1843 bool locked, need_to_kill;
1845 owait.memcg = memcg;
1846 owait.wait.flags = 0;
1847 owait.wait.func = memcg_oom_wake_function;
1848 owait.wait.private = current;
1849 INIT_LIST_HEAD(&owait.wait.task_list);
1850 need_to_kill = true;
1851 mem_cgroup_mark_under_oom(memcg);
1853 /* At first, try to OOM lock hierarchy under memcg.*/
1854 spin_lock(&memcg_oom_lock);
1855 locked = mem_cgroup_oom_lock(memcg);
1857 * Even if signal_pending(), we can't quit charge() loop without
1858 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1859 * under OOM is always welcomed, use TASK_KILLABLE here.
1861 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1862 if (!locked || memcg->oom_kill_disable)
1863 need_to_kill = false;
1865 mem_cgroup_oom_notify(memcg);
1866 spin_unlock(&memcg_oom_lock);
1869 finish_wait(&memcg_oom_waitq, &owait.wait);
1870 mem_cgroup_out_of_memory(memcg, mask, order);
1873 finish_wait(&memcg_oom_waitq, &owait.wait);
1875 spin_lock(&memcg_oom_lock);
1877 mem_cgroup_oom_unlock(memcg);
1878 memcg_wakeup_oom(memcg);
1879 spin_unlock(&memcg_oom_lock);
1881 mem_cgroup_unmark_under_oom(memcg);
1883 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
1885 /* Give chance to dying process */
1886 schedule_timeout_uninterruptible(1);
1891 * Currently used to update mapped file statistics, but the routine can be
1892 * generalized to update other statistics as well.
1894 * Notes: Race condition
1896 * We usually use page_cgroup_lock() for accessing page_cgroup member but
1897 * it tends to be costly. But considering some conditions, we doesn't need
1898 * to do so _always_.
1900 * Considering "charge", lock_page_cgroup() is not required because all
1901 * file-stat operations happen after a page is attached to radix-tree. There
1902 * are no race with "charge".
1904 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
1905 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
1906 * if there are race with "uncharge". Statistics itself is properly handled
1909 * Considering "move", this is an only case we see a race. To make the race
1910 * small, we check mm->moving_account and detect there are possibility of race
1911 * If there is, we take a lock.
1914 void __mem_cgroup_begin_update_page_stat(struct page *page,
1915 bool *locked, unsigned long *flags)
1917 struct mem_cgroup *memcg;
1918 struct page_cgroup *pc;
1920 pc = lookup_page_cgroup(page);
1922 memcg = pc->mem_cgroup;
1923 if (unlikely(!memcg || !PageCgroupUsed(pc)))
1926 * If this memory cgroup is not under account moving, we don't
1927 * need to take move_lock_page_cgroup(). Because we already hold
1928 * rcu_read_lock(), any calls to move_account will be delayed until
1929 * rcu_read_unlock() if mem_cgroup_stolen() == true.
1931 if (!mem_cgroup_stolen(memcg))
1934 move_lock_mem_cgroup(memcg, flags);
1935 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
1936 move_unlock_mem_cgroup(memcg, flags);
1942 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
1944 struct page_cgroup *pc = lookup_page_cgroup(page);
1947 * It's guaranteed that pc->mem_cgroup never changes while
1948 * lock is held because a routine modifies pc->mem_cgroup
1949 * should take move_lock_page_cgroup().
1951 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
1954 void mem_cgroup_update_page_stat(struct page *page,
1955 enum mem_cgroup_page_stat_item idx, int val)
1957 struct mem_cgroup *memcg;
1958 struct page_cgroup *pc = lookup_page_cgroup(page);
1959 unsigned long uninitialized_var(flags);
1961 if (mem_cgroup_disabled())
1964 memcg = pc->mem_cgroup;
1965 if (unlikely(!memcg || !PageCgroupUsed(pc)))
1969 case MEMCG_NR_FILE_MAPPED:
1970 idx = MEM_CGROUP_STAT_FILE_MAPPED;
1976 this_cpu_add(memcg->stat->count[idx], val);
1980 * size of first charge trial. "32" comes from vmscan.c's magic value.
1981 * TODO: maybe necessary to use big numbers in big irons.
1983 #define CHARGE_BATCH 32U
1984 struct memcg_stock_pcp {
1985 struct mem_cgroup *cached; /* this never be root cgroup */
1986 unsigned int nr_pages;
1987 struct work_struct work;
1988 unsigned long flags;
1989 #define FLUSHING_CACHED_CHARGE (0)
1991 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1992 static DEFINE_MUTEX(percpu_charge_mutex);
1995 * Try to consume stocked charge on this cpu. If success, one page is consumed
1996 * from local stock and true is returned. If the stock is 0 or charges from a
1997 * cgroup which is not current target, returns false. This stock will be
2000 static bool consume_stock(struct mem_cgroup *memcg)
2002 struct memcg_stock_pcp *stock;
2005 stock = &get_cpu_var(memcg_stock);
2006 if (memcg == stock->cached && stock->nr_pages)
2008 else /* need to call res_counter_charge */
2010 put_cpu_var(memcg_stock);
2015 * Returns stocks cached in percpu to res_counter and reset cached information.
2017 static void drain_stock(struct memcg_stock_pcp *stock)
2019 struct mem_cgroup *old = stock->cached;
2021 if (stock->nr_pages) {
2022 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2024 res_counter_uncharge(&old->res, bytes);
2025 if (do_swap_account)
2026 res_counter_uncharge(&old->memsw, bytes);
2027 stock->nr_pages = 0;
2029 stock->cached = NULL;
2033 * This must be called under preempt disabled or must be called by
2034 * a thread which is pinned to local cpu.
2036 static void drain_local_stock(struct work_struct *dummy)
2038 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2040 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2044 * Cache charges(val) which is from res_counter, to local per_cpu area.
2045 * This will be consumed by consume_stock() function, later.
2047 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2049 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2051 if (stock->cached != memcg) { /* reset if necessary */
2053 stock->cached = memcg;
2055 stock->nr_pages += nr_pages;
2056 put_cpu_var(memcg_stock);
2060 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2061 * of the hierarchy under it. sync flag says whether we should block
2062 * until the work is done.
2064 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2068 /* Notify other cpus that system-wide "drain" is running */
2071 for_each_online_cpu(cpu) {
2072 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2073 struct mem_cgroup *memcg;
2075 memcg = stock->cached;
2076 if (!memcg || !stock->nr_pages)
2078 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2080 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2082 drain_local_stock(&stock->work);
2084 schedule_work_on(cpu, &stock->work);
2092 for_each_online_cpu(cpu) {
2093 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2094 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2095 flush_work(&stock->work);
2102 * Tries to drain stocked charges in other cpus. This function is asynchronous
2103 * and just put a work per cpu for draining localy on each cpu. Caller can
2104 * expects some charges will be back to res_counter later but cannot wait for
2107 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2110 * If someone calls draining, avoid adding more kworker runs.
2112 if (!mutex_trylock(&percpu_charge_mutex))
2114 drain_all_stock(root_memcg, false);
2115 mutex_unlock(&percpu_charge_mutex);
2118 /* This is a synchronous drain interface. */
2119 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2121 /* called when force_empty is called */
2122 mutex_lock(&percpu_charge_mutex);
2123 drain_all_stock(root_memcg, true);
2124 mutex_unlock(&percpu_charge_mutex);
2128 * This function drains percpu counter value from DEAD cpu and
2129 * move it to local cpu. Note that this function can be preempted.
2131 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2135 spin_lock(&memcg->pcp_counter_lock);
2136 for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) {
2137 long x = per_cpu(memcg->stat->count[i], cpu);
2139 per_cpu(memcg->stat->count[i], cpu) = 0;
2140 memcg->nocpu_base.count[i] += x;
2142 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2143 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2145 per_cpu(memcg->stat->events[i], cpu) = 0;
2146 memcg->nocpu_base.events[i] += x;
2148 spin_unlock(&memcg->pcp_counter_lock);
2151 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2152 unsigned long action,
2155 int cpu = (unsigned long)hcpu;
2156 struct memcg_stock_pcp *stock;
2157 struct mem_cgroup *iter;
2159 if (action == CPU_ONLINE)
2162 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2165 for_each_mem_cgroup(iter)
2166 mem_cgroup_drain_pcp_counter(iter, cpu);
2168 stock = &per_cpu(memcg_stock, cpu);
2174 /* See __mem_cgroup_try_charge() for details */
2176 CHARGE_OK, /* success */
2177 CHARGE_RETRY, /* need to retry but retry is not bad */
2178 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2179 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2180 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2183 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2184 unsigned int nr_pages, bool oom_check)
2186 unsigned long csize = nr_pages * PAGE_SIZE;
2187 struct mem_cgroup *mem_over_limit;
2188 struct res_counter *fail_res;
2189 unsigned long flags = 0;
2192 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2195 if (!do_swap_account)
2197 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2201 res_counter_uncharge(&memcg->res, csize);
2202 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2203 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2205 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2207 * nr_pages can be either a huge page (HPAGE_PMD_NR), a batch
2208 * of regular pages (CHARGE_BATCH), or a single regular page (1).
2210 * Never reclaim on behalf of optional batching, retry with a
2211 * single page instead.
2213 if (nr_pages == CHARGE_BATCH)
2214 return CHARGE_RETRY;
2216 if (!(gfp_mask & __GFP_WAIT))
2217 return CHARGE_WOULDBLOCK;
2219 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2220 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2221 return CHARGE_RETRY;
2223 * Even though the limit is exceeded at this point, reclaim
2224 * may have been able to free some pages. Retry the charge
2225 * before killing the task.
2227 * Only for regular pages, though: huge pages are rather
2228 * unlikely to succeed so close to the limit, and we fall back
2229 * to regular pages anyway in case of failure.
2231 if (nr_pages == 1 && ret)
2232 return CHARGE_RETRY;
2235 * At task move, charge accounts can be doubly counted. So, it's
2236 * better to wait until the end of task_move if something is going on.
2238 if (mem_cgroup_wait_acct_move(mem_over_limit))
2239 return CHARGE_RETRY;
2241 /* If we don't need to call oom-killer at el, return immediately */
2243 return CHARGE_NOMEM;
2245 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask, get_order(csize)))
2246 return CHARGE_OOM_DIE;
2248 return CHARGE_RETRY;
2252 * __mem_cgroup_try_charge() does
2253 * 1. detect memcg to be charged against from passed *mm and *ptr,
2254 * 2. update res_counter
2255 * 3. call memory reclaim if necessary.
2257 * In some special case, if the task is fatal, fatal_signal_pending() or
2258 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2259 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2260 * as possible without any hazards. 2: all pages should have a valid
2261 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2262 * pointer, that is treated as a charge to root_mem_cgroup.
2264 * So __mem_cgroup_try_charge() will return
2265 * 0 ... on success, filling *ptr with a valid memcg pointer.
2266 * -ENOMEM ... charge failure because of resource limits.
2267 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2269 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2270 * the oom-killer can be invoked.
2272 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2274 unsigned int nr_pages,
2275 struct mem_cgroup **ptr,
2278 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2279 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2280 struct mem_cgroup *memcg = NULL;
2284 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2285 * in system level. So, allow to go ahead dying process in addition to
2288 if (unlikely(test_thread_flag(TIF_MEMDIE)
2289 || fatal_signal_pending(current)))
2293 * We always charge the cgroup the mm_struct belongs to.
2294 * The mm_struct's mem_cgroup changes on task migration if the
2295 * thread group leader migrates. It's possible that mm is not
2296 * set, if so charge the init_mm (happens for pagecache usage).
2299 *ptr = root_mem_cgroup;
2301 if (*ptr) { /* css should be a valid one */
2303 VM_BUG_ON(css_is_removed(&memcg->css));
2304 if (mem_cgroup_is_root(memcg))
2306 if (nr_pages == 1 && consume_stock(memcg))
2308 css_get(&memcg->css);
2310 struct task_struct *p;
2313 p = rcu_dereference(mm->owner);
2315 * Because we don't have task_lock(), "p" can exit.
2316 * In that case, "memcg" can point to root or p can be NULL with
2317 * race with swapoff. Then, we have small risk of mis-accouning.
2318 * But such kind of mis-account by race always happens because
2319 * we don't have cgroup_mutex(). It's overkill and we allo that
2321 * (*) swapoff at el will charge against mm-struct not against
2322 * task-struct. So, mm->owner can be NULL.
2324 memcg = mem_cgroup_from_task(p);
2326 memcg = root_mem_cgroup;
2327 if (mem_cgroup_is_root(memcg)) {
2331 if (nr_pages == 1 && consume_stock(memcg)) {
2333 * It seems dagerous to access memcg without css_get().
2334 * But considering how consume_stok works, it's not
2335 * necessary. If consume_stock success, some charges
2336 * from this memcg are cached on this cpu. So, we
2337 * don't need to call css_get()/css_tryget() before
2338 * calling consume_stock().
2343 /* after here, we may be blocked. we need to get refcnt */
2344 if (!css_tryget(&memcg->css)) {
2354 /* If killed, bypass charge */
2355 if (fatal_signal_pending(current)) {
2356 css_put(&memcg->css);
2361 if (oom && !nr_oom_retries) {
2363 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2366 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, oom_check);
2370 case CHARGE_RETRY: /* not in OOM situation but retry */
2372 css_put(&memcg->css);
2375 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2376 css_put(&memcg->css);
2378 case CHARGE_NOMEM: /* OOM routine works */
2380 css_put(&memcg->css);
2383 /* If oom, we never return -ENOMEM */
2386 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2387 css_put(&memcg->css);
2390 } while (ret != CHARGE_OK);
2392 if (batch > nr_pages)
2393 refill_stock(memcg, batch - nr_pages);
2394 css_put(&memcg->css);
2402 *ptr = root_mem_cgroup;
2407 * Somemtimes we have to undo a charge we got by try_charge().
2408 * This function is for that and do uncharge, put css's refcnt.
2409 * gotten by try_charge().
2411 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2412 unsigned int nr_pages)
2414 if (!mem_cgroup_is_root(memcg)) {
2415 unsigned long bytes = nr_pages * PAGE_SIZE;
2417 res_counter_uncharge(&memcg->res, bytes);
2418 if (do_swap_account)
2419 res_counter_uncharge(&memcg->memsw, bytes);
2424 * A helper function to get mem_cgroup from ID. must be called under
2425 * rcu_read_lock(). The caller must check css_is_removed() or some if
2426 * it's concern. (dropping refcnt from swap can be called against removed
2429 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2431 struct cgroup_subsys_state *css;
2433 /* ID 0 is unused ID */
2436 css = css_lookup(&mem_cgroup_subsys, id);
2439 return container_of(css, struct mem_cgroup, css);
2442 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2444 struct mem_cgroup *memcg = NULL;
2445 struct page_cgroup *pc;
2449 VM_BUG_ON(!PageLocked(page));
2451 pc = lookup_page_cgroup(page);
2452 lock_page_cgroup(pc);
2453 if (PageCgroupUsed(pc)) {
2454 memcg = pc->mem_cgroup;
2455 if (memcg && !css_tryget(&memcg->css))
2457 } else if (PageSwapCache(page)) {
2458 ent.val = page_private(page);
2459 id = lookup_swap_cgroup_id(ent);
2461 memcg = mem_cgroup_lookup(id);
2462 if (memcg && !css_tryget(&memcg->css))
2466 unlock_page_cgroup(pc);
2470 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2472 unsigned int nr_pages,
2473 enum charge_type ctype,
2476 struct page_cgroup *pc = lookup_page_cgroup(page);
2477 struct zone *uninitialized_var(zone);
2478 bool was_on_lru = false;
2481 lock_page_cgroup(pc);
2482 if (unlikely(PageCgroupUsed(pc))) {
2483 unlock_page_cgroup(pc);
2484 __mem_cgroup_cancel_charge(memcg, nr_pages);
2488 * we don't need page_cgroup_lock about tail pages, becase they are not
2489 * accessed by any other context at this point.
2493 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2494 * may already be on some other mem_cgroup's LRU. Take care of it.
2497 zone = page_zone(page);
2498 spin_lock_irq(&zone->lru_lock);
2499 if (PageLRU(page)) {
2501 del_page_from_lru_list(zone, page, page_lru(page));
2506 pc->mem_cgroup = memcg;
2508 * We access a page_cgroup asynchronously without lock_page_cgroup().
2509 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2510 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2511 * before USED bit, we need memory barrier here.
2512 * See mem_cgroup_add_lru_list(), etc.
2515 SetPageCgroupUsed(pc);
2519 VM_BUG_ON(PageLRU(page));
2521 add_page_to_lru_list(zone, page, page_lru(page));
2523 spin_unlock_irq(&zone->lru_lock);
2526 if (ctype == MEM_CGROUP_CHARGE_TYPE_MAPPED)
2531 mem_cgroup_charge_statistics(memcg, anon, nr_pages);
2532 unlock_page_cgroup(pc);
2535 * "charge_statistics" updated event counter. Then, check it.
2536 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2537 * if they exceeds softlimit.
2539 memcg_check_events(memcg, page);
2542 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2544 #define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MIGRATION))
2546 * Because tail pages are not marked as "used", set it. We're under
2547 * zone->lru_lock, 'splitting on pmd' and compound_lock.
2548 * charge/uncharge will be never happen and move_account() is done under
2549 * compound_lock(), so we don't have to take care of races.
2551 void mem_cgroup_split_huge_fixup(struct page *head)
2553 struct page_cgroup *head_pc = lookup_page_cgroup(head);
2554 struct page_cgroup *pc;
2557 if (mem_cgroup_disabled())
2559 for (i = 1; i < HPAGE_PMD_NR; i++) {
2561 pc->mem_cgroup = head_pc->mem_cgroup;
2562 smp_wmb();/* see __commit_charge() */
2563 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2566 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2569 * mem_cgroup_move_account - move account of the page
2571 * @nr_pages: number of regular pages (>1 for huge pages)
2572 * @pc: page_cgroup of the page.
2573 * @from: mem_cgroup which the page is moved from.
2574 * @to: mem_cgroup which the page is moved to. @from != @to.
2575 * @uncharge: whether we should call uncharge and css_put against @from.
2577 * The caller must confirm following.
2578 * - page is not on LRU (isolate_page() is useful.)
2579 * - compound_lock is held when nr_pages > 1
2581 * This function doesn't do "charge" nor css_get to new cgroup. It should be
2582 * done by a caller(__mem_cgroup_try_charge would be useful). If @uncharge is
2583 * true, this function does "uncharge" from old cgroup, but it doesn't if
2584 * @uncharge is false, so a caller should do "uncharge".
2586 static int mem_cgroup_move_account(struct page *page,
2587 unsigned int nr_pages,
2588 struct page_cgroup *pc,
2589 struct mem_cgroup *from,
2590 struct mem_cgroup *to,
2593 unsigned long flags;
2595 bool anon = PageAnon(page);
2597 VM_BUG_ON(from == to);
2598 VM_BUG_ON(PageLRU(page));
2600 * The page is isolated from LRU. So, collapse function
2601 * will not handle this page. But page splitting can happen.
2602 * Do this check under compound_page_lock(). The caller should
2606 if (nr_pages > 1 && !PageTransHuge(page))
2609 lock_page_cgroup(pc);
2612 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
2615 move_lock_mem_cgroup(from, &flags);
2617 if (!anon && page_mapped(page)) {
2618 /* Update mapped_file data for mem_cgroup */
2620 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2621 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2624 mem_cgroup_charge_statistics(from, anon, -nr_pages);
2626 /* This is not "cancel", but cancel_charge does all we need. */
2627 __mem_cgroup_cancel_charge(from, nr_pages);
2629 /* caller should have done css_get */
2630 pc->mem_cgroup = to;
2631 mem_cgroup_charge_statistics(to, anon, nr_pages);
2633 * We charges against "to" which may not have any tasks. Then, "to"
2634 * can be under rmdir(). But in current implementation, caller of
2635 * this function is just force_empty() and move charge, so it's
2636 * guaranteed that "to" is never removed. So, we don't check rmdir
2639 move_unlock_mem_cgroup(from, &flags);
2642 unlock_page_cgroup(pc);
2646 memcg_check_events(to, page);
2647 memcg_check_events(from, page);
2653 * move charges to its parent.
2656 static int mem_cgroup_move_parent(struct page *page,
2657 struct page_cgroup *pc,
2658 struct mem_cgroup *child,
2661 struct cgroup *cg = child->css.cgroup;
2662 struct cgroup *pcg = cg->parent;
2663 struct mem_cgroup *parent;
2664 unsigned int nr_pages;
2665 unsigned long uninitialized_var(flags);
2673 if (!get_page_unless_zero(page))
2675 if (isolate_lru_page(page))
2678 nr_pages = hpage_nr_pages(page);
2680 parent = mem_cgroup_from_cont(pcg);
2681 ret = __mem_cgroup_try_charge(NULL, gfp_mask, nr_pages, &parent, false);
2686 flags = compound_lock_irqsave(page);
2688 ret = mem_cgroup_move_account(page, nr_pages, pc, child, parent, true);
2690 __mem_cgroup_cancel_charge(parent, nr_pages);
2693 compound_unlock_irqrestore(page, flags);
2695 putback_lru_page(page);
2703 * Charge the memory controller for page usage.
2705 * 0 if the charge was successful
2706 * < 0 if the cgroup is over its limit
2708 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2709 gfp_t gfp_mask, enum charge_type ctype)
2711 struct mem_cgroup *memcg = NULL;
2712 unsigned int nr_pages = 1;
2716 if (PageTransHuge(page)) {
2717 nr_pages <<= compound_order(page);
2718 VM_BUG_ON(!PageTransHuge(page));
2720 * Never OOM-kill a process for a huge page. The
2721 * fault handler will fall back to regular pages.
2726 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
2729 __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false);
2733 int mem_cgroup_newpage_charge(struct page *page,
2734 struct mm_struct *mm, gfp_t gfp_mask)
2736 if (mem_cgroup_disabled())
2738 VM_BUG_ON(page_mapped(page));
2739 VM_BUG_ON(page->mapping && !PageAnon(page));
2741 return mem_cgroup_charge_common(page, mm, gfp_mask,
2742 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2746 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2747 enum charge_type ctype);
2749 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2752 struct mem_cgroup *memcg = NULL;
2753 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
2756 if (mem_cgroup_disabled())
2758 if (PageCompound(page))
2763 if (!page_is_file_cache(page))
2764 type = MEM_CGROUP_CHARGE_TYPE_SHMEM;
2766 if (!PageSwapCache(page))
2767 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
2768 else { /* page is swapcache/shmem */
2769 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &memcg);
2771 __mem_cgroup_commit_charge_swapin(page, memcg, type);
2777 * While swap-in, try_charge -> commit or cancel, the page is locked.
2778 * And when try_charge() successfully returns, one refcnt to memcg without
2779 * struct page_cgroup is acquired. This refcnt will be consumed by
2780 * "commit()" or removed by "cancel()"
2782 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2784 gfp_t mask, struct mem_cgroup **memcgp)
2786 struct mem_cgroup *memcg;
2791 if (mem_cgroup_disabled())
2794 if (!do_swap_account)
2797 * A racing thread's fault, or swapoff, may have already updated
2798 * the pte, and even removed page from swap cache: in those cases
2799 * do_swap_page()'s pte_same() test will fail; but there's also a
2800 * KSM case which does need to charge the page.
2802 if (!PageSwapCache(page))
2804 memcg = try_get_mem_cgroup_from_page(page);
2808 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
2809 css_put(&memcg->css);
2816 ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
2823 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
2824 enum charge_type ctype)
2826 if (mem_cgroup_disabled())
2830 cgroup_exclude_rmdir(&memcg->css);
2832 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
2834 * Now swap is on-memory. This means this page may be
2835 * counted both as mem and swap....double count.
2836 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2837 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2838 * may call delete_from_swap_cache() before reach here.
2840 if (do_swap_account && PageSwapCache(page)) {
2841 swp_entry_t ent = {.val = page_private(page)};
2842 mem_cgroup_uncharge_swap(ent);
2845 * At swapin, we may charge account against cgroup which has no tasks.
2846 * So, rmdir()->pre_destroy() can be called while we do this charge.
2847 * In that case, we need to call pre_destroy() again. check it here.
2849 cgroup_release_and_wakeup_rmdir(&memcg->css);
2852 void mem_cgroup_commit_charge_swapin(struct page *page,
2853 struct mem_cgroup *memcg)
2855 __mem_cgroup_commit_charge_swapin(page, memcg,
2856 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2859 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
2861 if (mem_cgroup_disabled())
2865 __mem_cgroup_cancel_charge(memcg, 1);
2868 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
2869 unsigned int nr_pages,
2870 const enum charge_type ctype)
2872 struct memcg_batch_info *batch = NULL;
2873 bool uncharge_memsw = true;
2875 /* If swapout, usage of swap doesn't decrease */
2876 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2877 uncharge_memsw = false;
2879 batch = ¤t->memcg_batch;
2881 * In usual, we do css_get() when we remember memcg pointer.
2882 * But in this case, we keep res->usage until end of a series of
2883 * uncharges. Then, it's ok to ignore memcg's refcnt.
2886 batch->memcg = memcg;
2888 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2889 * In those cases, all pages freed continuously can be expected to be in
2890 * the same cgroup and we have chance to coalesce uncharges.
2891 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2892 * because we want to do uncharge as soon as possible.
2895 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
2896 goto direct_uncharge;
2899 goto direct_uncharge;
2902 * In typical case, batch->memcg == mem. This means we can
2903 * merge a series of uncharges to an uncharge of res_counter.
2904 * If not, we uncharge res_counter ony by one.
2906 if (batch->memcg != memcg)
2907 goto direct_uncharge;
2908 /* remember freed charge and uncharge it later */
2911 batch->memsw_nr_pages++;
2914 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
2916 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
2917 if (unlikely(batch->memcg != memcg))
2918 memcg_oom_recover(memcg);
2922 * uncharge if !page_mapped(page)
2924 static struct mem_cgroup *
2925 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
2927 struct mem_cgroup *memcg = NULL;
2928 unsigned int nr_pages = 1;
2929 struct page_cgroup *pc;
2932 if (mem_cgroup_disabled())
2935 if (PageSwapCache(page))
2938 if (PageTransHuge(page)) {
2939 nr_pages <<= compound_order(page);
2940 VM_BUG_ON(!PageTransHuge(page));
2943 * Check if our page_cgroup is valid
2945 pc = lookup_page_cgroup(page);
2946 if (unlikely(!PageCgroupUsed(pc)))
2949 lock_page_cgroup(pc);
2951 memcg = pc->mem_cgroup;
2953 if (!PageCgroupUsed(pc))
2956 anon = PageAnon(page);
2959 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2961 * Generally PageAnon tells if it's the anon statistics to be
2962 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
2963 * used before page reached the stage of being marked PageAnon.
2967 case MEM_CGROUP_CHARGE_TYPE_DROP:
2968 /* See mem_cgroup_prepare_migration() */
2969 if (page_mapped(page) || PageCgroupMigration(pc))
2972 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
2973 if (!PageAnon(page)) { /* Shared memory */
2974 if (page->mapping && !page_is_file_cache(page))
2976 } else if (page_mapped(page)) /* Anon */
2983 mem_cgroup_charge_statistics(memcg, anon, -nr_pages);
2985 ClearPageCgroupUsed(pc);
2987 * pc->mem_cgroup is not cleared here. It will be accessed when it's
2988 * freed from LRU. This is safe because uncharged page is expected not
2989 * to be reused (freed soon). Exception is SwapCache, it's handled by
2990 * special functions.
2993 unlock_page_cgroup(pc);
2995 * even after unlock, we have memcg->res.usage here and this memcg
2996 * will never be freed.
2998 memcg_check_events(memcg, page);
2999 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
3000 mem_cgroup_swap_statistics(memcg, true);
3001 mem_cgroup_get(memcg);
3003 if (!mem_cgroup_is_root(memcg))
3004 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
3009 unlock_page_cgroup(pc);
3013 void mem_cgroup_uncharge_page(struct page *page)
3016 if (page_mapped(page))
3018 VM_BUG_ON(page->mapping && !PageAnon(page));
3019 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
3022 void mem_cgroup_uncharge_cache_page(struct page *page)
3024 VM_BUG_ON(page_mapped(page));
3025 VM_BUG_ON(page->mapping);
3026 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
3030 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
3031 * In that cases, pages are freed continuously and we can expect pages
3032 * are in the same memcg. All these calls itself limits the number of
3033 * pages freed at once, then uncharge_start/end() is called properly.
3034 * This may be called prural(2) times in a context,
3037 void mem_cgroup_uncharge_start(void)
3039 current->memcg_batch.do_batch++;
3040 /* We can do nest. */
3041 if (current->memcg_batch.do_batch == 1) {
3042 current->memcg_batch.memcg = NULL;
3043 current->memcg_batch.nr_pages = 0;
3044 current->memcg_batch.memsw_nr_pages = 0;
3048 void mem_cgroup_uncharge_end(void)
3050 struct memcg_batch_info *batch = ¤t->memcg_batch;
3052 if (!batch->do_batch)
3056 if (batch->do_batch) /* If stacked, do nothing. */
3062 * This "batch->memcg" is valid without any css_get/put etc...
3063 * bacause we hide charges behind us.
3065 if (batch->nr_pages)
3066 res_counter_uncharge(&batch->memcg->res,
3067 batch->nr_pages * PAGE_SIZE);
3068 if (batch->memsw_nr_pages)
3069 res_counter_uncharge(&batch->memcg->memsw,
3070 batch->memsw_nr_pages * PAGE_SIZE);
3071 memcg_oom_recover(batch->memcg);
3072 /* forget this pointer (for sanity check) */
3073 batch->memcg = NULL;
3078 * called after __delete_from_swap_cache() and drop "page" account.
3079 * memcg information is recorded to swap_cgroup of "ent"
3082 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
3084 struct mem_cgroup *memcg;
3085 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
3087 if (!swapout) /* this was a swap cache but the swap is unused ! */
3088 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
3090 memcg = __mem_cgroup_uncharge_common(page, ctype);
3093 * record memcg information, if swapout && memcg != NULL,
3094 * mem_cgroup_get() was called in uncharge().
3096 if (do_swap_account && swapout && memcg)
3097 swap_cgroup_record(ent, css_id(&memcg->css));
3101 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3103 * called from swap_entry_free(). remove record in swap_cgroup and
3104 * uncharge "memsw" account.
3106 void mem_cgroup_uncharge_swap(swp_entry_t ent)
3108 struct mem_cgroup *memcg;
3111 if (!do_swap_account)
3114 id = swap_cgroup_record(ent, 0);
3116 memcg = mem_cgroup_lookup(id);
3119 * We uncharge this because swap is freed.
3120 * This memcg can be obsolete one. We avoid calling css_tryget
3122 if (!mem_cgroup_is_root(memcg))
3123 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
3124 mem_cgroup_swap_statistics(memcg, false);
3125 mem_cgroup_put(memcg);
3131 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3132 * @entry: swap entry to be moved
3133 * @from: mem_cgroup which the entry is moved from
3134 * @to: mem_cgroup which the entry is moved to
3136 * It succeeds only when the swap_cgroup's record for this entry is the same
3137 * as the mem_cgroup's id of @from.
3139 * Returns 0 on success, -EINVAL on failure.
3141 * The caller must have charged to @to, IOW, called res_counter_charge() about
3142 * both res and memsw, and called css_get().
3144 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3145 struct mem_cgroup *from, struct mem_cgroup *to)
3147 unsigned short old_id, new_id;
3149 old_id = css_id(&from->css);
3150 new_id = css_id(&to->css);
3152 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3153 mem_cgroup_swap_statistics(from, false);
3154 mem_cgroup_swap_statistics(to, true);
3156 * This function is only called from task migration context now.
3157 * It postpones res_counter and refcount handling till the end
3158 * of task migration(mem_cgroup_clear_mc()) for performance
3159 * improvement. But we cannot postpone mem_cgroup_get(to)
3160 * because if the process that has been moved to @to does
3161 * swap-in, the refcount of @to might be decreased to 0.
3169 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3170 struct mem_cgroup *from, struct mem_cgroup *to)
3177 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
3180 int mem_cgroup_prepare_migration(struct page *page,
3181 struct page *newpage, struct mem_cgroup **memcgp, gfp_t gfp_mask)
3183 struct mem_cgroup *memcg = NULL;
3184 struct page_cgroup *pc;
3185 enum charge_type ctype;
3190 VM_BUG_ON(PageTransHuge(page));
3191 if (mem_cgroup_disabled())
3194 pc = lookup_page_cgroup(page);
3195 lock_page_cgroup(pc);
3196 if (PageCgroupUsed(pc)) {
3197 memcg = pc->mem_cgroup;
3198 css_get(&memcg->css);
3200 * At migrating an anonymous page, its mapcount goes down
3201 * to 0 and uncharge() will be called. But, even if it's fully
3202 * unmapped, migration may fail and this page has to be
3203 * charged again. We set MIGRATION flag here and delay uncharge
3204 * until end_migration() is called
3206 * Corner Case Thinking
3208 * When the old page was mapped as Anon and it's unmap-and-freed
3209 * while migration was ongoing.
3210 * If unmap finds the old page, uncharge() of it will be delayed
3211 * until end_migration(). If unmap finds a new page, it's
3212 * uncharged when it make mapcount to be 1->0. If unmap code
3213 * finds swap_migration_entry, the new page will not be mapped
3214 * and end_migration() will find it(mapcount==0).
3217 * When the old page was mapped but migraion fails, the kernel
3218 * remaps it. A charge for it is kept by MIGRATION flag even
3219 * if mapcount goes down to 0. We can do remap successfully
3220 * without charging it again.
3223 * The "old" page is under lock_page() until the end of
3224 * migration, so, the old page itself will not be swapped-out.
3225 * If the new page is swapped out before end_migraton, our
3226 * hook to usual swap-out path will catch the event.
3229 SetPageCgroupMigration(pc);
3231 unlock_page_cgroup(pc);
3233 * If the page is not charged at this point,
3240 ret = __mem_cgroup_try_charge(NULL, gfp_mask, 1, memcgp, false);
3241 css_put(&memcg->css);/* drop extra refcnt */
3243 if (PageAnon(page)) {
3244 lock_page_cgroup(pc);
3245 ClearPageCgroupMigration(pc);
3246 unlock_page_cgroup(pc);
3248 * The old page may be fully unmapped while we kept it.
3250 mem_cgroup_uncharge_page(page);
3252 /* we'll need to revisit this error code (we have -EINTR) */
3256 * We charge new page before it's used/mapped. So, even if unlock_page()
3257 * is called before end_migration, we can catch all events on this new
3258 * page. In the case new page is migrated but not remapped, new page's
3259 * mapcount will be finally 0 and we call uncharge in end_migration().
3262 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
3263 else if (page_is_file_cache(page))
3264 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
3266 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3267 __mem_cgroup_commit_charge(memcg, newpage, 1, ctype, false);
3271 /* remove redundant charge if migration failed*/
3272 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
3273 struct page *oldpage, struct page *newpage, bool migration_ok)
3275 struct page *used, *unused;
3276 struct page_cgroup *pc;
3281 /* blocks rmdir() */
3282 cgroup_exclude_rmdir(&memcg->css);
3283 if (!migration_ok) {
3291 * We disallowed uncharge of pages under migration because mapcount
3292 * of the page goes down to zero, temporarly.
3293 * Clear the flag and check the page should be charged.
3295 pc = lookup_page_cgroup(oldpage);
3296 lock_page_cgroup(pc);
3297 ClearPageCgroupMigration(pc);
3298 unlock_page_cgroup(pc);
3299 anon = PageAnon(used);
3300 __mem_cgroup_uncharge_common(unused,
3301 anon ? MEM_CGROUP_CHARGE_TYPE_MAPPED
3302 : MEM_CGROUP_CHARGE_TYPE_CACHE);
3305 * If a page is a file cache, radix-tree replacement is very atomic
3306 * and we can skip this check. When it was an Anon page, its mapcount
3307 * goes down to 0. But because we added MIGRATION flage, it's not
3308 * uncharged yet. There are several case but page->mapcount check
3309 * and USED bit check in mem_cgroup_uncharge_page() will do enough
3310 * check. (see prepare_charge() also)
3313 mem_cgroup_uncharge_page(used);
3315 * At migration, we may charge account against cgroup which has no
3317 * So, rmdir()->pre_destroy() can be called while we do this charge.
3318 * In that case, we need to call pre_destroy() again. check it here.
3320 cgroup_release_and_wakeup_rmdir(&memcg->css);
3324 * At replace page cache, newpage is not under any memcg but it's on
3325 * LRU. So, this function doesn't touch res_counter but handles LRU
3326 * in correct way. Both pages are locked so we cannot race with uncharge.
3328 void mem_cgroup_replace_page_cache(struct page *oldpage,
3329 struct page *newpage)
3331 struct mem_cgroup *memcg = NULL;
3332 struct page_cgroup *pc;
3333 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
3335 if (mem_cgroup_disabled())
3338 pc = lookup_page_cgroup(oldpage);
3339 /* fix accounting on old pages */
3340 lock_page_cgroup(pc);
3341 if (PageCgroupUsed(pc)) {
3342 memcg = pc->mem_cgroup;
3343 mem_cgroup_charge_statistics(memcg, false, -1);
3344 ClearPageCgroupUsed(pc);
3346 unlock_page_cgroup(pc);
3349 * When called from shmem_replace_page(), in some cases the
3350 * oldpage has already been charged, and in some cases not.
3355 if (PageSwapBacked(oldpage))
3356 type = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3359 * Even if newpage->mapping was NULL before starting replacement,
3360 * the newpage may be on LRU(or pagevec for LRU) already. We lock
3361 * LRU while we overwrite pc->mem_cgroup.
3363 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
3366 #ifdef CONFIG_DEBUG_VM
3367 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3369 struct page_cgroup *pc;
3371 pc = lookup_page_cgroup(page);
3373 * Can be NULL while feeding pages into the page allocator for
3374 * the first time, i.e. during boot or memory hotplug;
3375 * or when mem_cgroup_disabled().
3377 if (likely(pc) && PageCgroupUsed(pc))
3382 bool mem_cgroup_bad_page_check(struct page *page)
3384 if (mem_cgroup_disabled())
3387 return lookup_page_cgroup_used(page) != NULL;
3390 void mem_cgroup_print_bad_page(struct page *page)
3392 struct page_cgroup *pc;
3394 pc = lookup_page_cgroup_used(page);
3396 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
3397 pc, pc->flags, pc->mem_cgroup);
3402 static DEFINE_MUTEX(set_limit_mutex);
3404 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3405 unsigned long long val)
3408 u64 memswlimit, memlimit;
3410 int children = mem_cgroup_count_children(memcg);
3411 u64 curusage, oldusage;
3415 * For keeping hierarchical_reclaim simple, how long we should retry
3416 * is depends on callers. We set our retry-count to be function
3417 * of # of children which we should visit in this loop.
3419 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3421 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3424 while (retry_count) {
3425 if (signal_pending(current)) {
3430 * Rather than hide all in some function, I do this in
3431 * open coded manner. You see what this really does.
3432 * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3434 mutex_lock(&set_limit_mutex);
3435 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3436 if (memswlimit < val) {
3438 mutex_unlock(&set_limit_mutex);
3442 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3446 ret = res_counter_set_limit(&memcg->res, val);
3448 if (memswlimit == val)
3449 memcg->memsw_is_minimum = true;
3451 memcg->memsw_is_minimum = false;
3453 mutex_unlock(&set_limit_mutex);
3458 mem_cgroup_reclaim(memcg, GFP_KERNEL,
3459 MEM_CGROUP_RECLAIM_SHRINK);
3460 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3461 /* Usage is reduced ? */
3462 if (curusage >= oldusage)
3465 oldusage = curusage;
3467 if (!ret && enlarge)
3468 memcg_oom_recover(memcg);
3473 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3474 unsigned long long val)
3477 u64 memlimit, memswlimit, oldusage, curusage;
3478 int children = mem_cgroup_count_children(memcg);
3482 /* see mem_cgroup_resize_res_limit */
3483 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3484 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3485 while (retry_count) {
3486 if (signal_pending(current)) {
3491 * Rather than hide all in some function, I do this in
3492 * open coded manner. You see what this really does.
3493 * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3495 mutex_lock(&set_limit_mutex);
3496 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3497 if (memlimit > val) {
3499 mutex_unlock(&set_limit_mutex);
3502 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3503 if (memswlimit < val)
3505 ret = res_counter_set_limit(&memcg->memsw, val);
3507 if (memlimit == val)
3508 memcg->memsw_is_minimum = true;
3510 memcg->memsw_is_minimum = false;
3512 mutex_unlock(&set_limit_mutex);
3517 mem_cgroup_reclaim(memcg, GFP_KERNEL,
3518 MEM_CGROUP_RECLAIM_NOSWAP |
3519 MEM_CGROUP_RECLAIM_SHRINK);
3520 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3521 /* Usage is reduced ? */
3522 if (curusage >= oldusage)
3525 oldusage = curusage;
3527 if (!ret && enlarge)
3528 memcg_oom_recover(memcg);
3532 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3534 unsigned long *total_scanned)
3536 unsigned long nr_reclaimed = 0;
3537 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3538 unsigned long reclaimed;
3540 struct mem_cgroup_tree_per_zone *mctz;
3541 unsigned long long excess;
3542 unsigned long nr_scanned;
3547 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3549 * This loop can run a while, specially if mem_cgroup's continuously
3550 * keep exceeding their soft limit and putting the system under
3557 mz = mem_cgroup_largest_soft_limit_node(mctz);
3562 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
3563 gfp_mask, &nr_scanned);
3564 nr_reclaimed += reclaimed;
3565 *total_scanned += nr_scanned;
3566 spin_lock(&mctz->lock);
3569 * If we failed to reclaim anything from this memory cgroup
3570 * it is time to move on to the next cgroup
3576 * Loop until we find yet another one.
3578 * By the time we get the soft_limit lock
3579 * again, someone might have aded the
3580 * group back on the RB tree. Iterate to
3581 * make sure we get a different mem.
3582 * mem_cgroup_largest_soft_limit_node returns
3583 * NULL if no other cgroup is present on
3587 __mem_cgroup_largest_soft_limit_node(mctz);
3589 css_put(&next_mz->memcg->css);
3590 else /* next_mz == NULL or other memcg */
3594 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
3595 excess = res_counter_soft_limit_excess(&mz->memcg->res);
3597 * One school of thought says that we should not add
3598 * back the node to the tree if reclaim returns 0.
3599 * But our reclaim could return 0, simply because due
3600 * to priority we are exposing a smaller subset of
3601 * memory to reclaim from. Consider this as a longer
3604 /* If excess == 0, no tree ops */
3605 __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess);
3606 spin_unlock(&mctz->lock);
3607 css_put(&mz->memcg->css);
3610 * Could not reclaim anything and there are no more
3611 * mem cgroups to try or we seem to be looping without
3612 * reclaiming anything.
3614 if (!nr_reclaimed &&
3616 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3618 } while (!nr_reclaimed);
3620 css_put(&next_mz->memcg->css);
3621 return nr_reclaimed;
3625 * This routine traverse page_cgroup in given list and drop them all.
3626 * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3628 static int mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
3629 int node, int zid, enum lru_list lru)
3631 struct mem_cgroup_per_zone *mz;
3632 unsigned long flags, loop;
3633 struct list_head *list;
3638 zone = &NODE_DATA(node)->node_zones[zid];
3639 mz = mem_cgroup_zoneinfo(memcg, node, zid);
3640 list = &mz->lruvec.lists[lru];
3642 loop = mz->lru_size[lru];
3643 /* give some margin against EBUSY etc...*/
3647 struct page_cgroup *pc;
3651 spin_lock_irqsave(&zone->lru_lock, flags);
3652 if (list_empty(list)) {
3653 spin_unlock_irqrestore(&zone->lru_lock, flags);
3656 page = list_entry(list->prev, struct page, lru);
3658 list_move(&page->lru, list);
3660 spin_unlock_irqrestore(&zone->lru_lock, flags);
3663 spin_unlock_irqrestore(&zone->lru_lock, flags);
3665 pc = lookup_page_cgroup(page);
3667 ret = mem_cgroup_move_parent(page, pc, memcg, GFP_KERNEL);
3668 if (ret == -ENOMEM || ret == -EINTR)
3671 if (ret == -EBUSY || ret == -EINVAL) {
3672 /* found lock contention or "pc" is obsolete. */
3679 if (!ret && !list_empty(list))
3685 * make mem_cgroup's charge to be 0 if there is no task.
3686 * This enables deleting this mem_cgroup.
3688 static int mem_cgroup_force_empty(struct mem_cgroup *memcg, bool free_all)
3691 int node, zid, shrink;
3692 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3693 struct cgroup *cgrp = memcg->css.cgroup;
3695 css_get(&memcg->css);
3698 /* should free all ? */
3704 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3707 if (signal_pending(current))
3709 /* This is for making all *used* pages to be on LRU. */
3710 lru_add_drain_all();
3711 drain_all_stock_sync(memcg);
3713 mem_cgroup_start_move(memcg);
3714 for_each_node_state(node, N_HIGH_MEMORY) {
3715 for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3718 ret = mem_cgroup_force_empty_list(memcg,
3727 mem_cgroup_end_move(memcg);
3728 memcg_oom_recover(memcg);
3729 /* it seems parent cgroup doesn't have enough mem */
3733 /* "ret" should also be checked to ensure all lists are empty. */
3734 } while (res_counter_read_u64(&memcg->res, RES_USAGE) > 0 || ret);
3736 css_put(&memcg->css);
3740 /* returns EBUSY if there is a task or if we come here twice. */
3741 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3745 /* we call try-to-free pages for make this cgroup empty */
3746 lru_add_drain_all();
3747 /* try to free all pages in this cgroup */
3749 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
3752 if (signal_pending(current)) {
3756 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
3760 /* maybe some writeback is necessary */
3761 congestion_wait(BLK_RW_ASYNC, HZ/10);
3766 /* try move_account...there may be some *locked* pages. */
3770 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3772 return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3776 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3778 return mem_cgroup_from_cont(cont)->use_hierarchy;
3781 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3785 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3786 struct cgroup *parent = cont->parent;
3787 struct mem_cgroup *parent_memcg = NULL;
3790 parent_memcg = mem_cgroup_from_cont(parent);
3794 * If parent's use_hierarchy is set, we can't make any modifications
3795 * in the child subtrees. If it is unset, then the change can
3796 * occur, provided the current cgroup has no children.
3798 * For the root cgroup, parent_mem is NULL, we allow value to be
3799 * set if there are no children.
3801 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3802 (val == 1 || val == 0)) {
3803 if (list_empty(&cont->children))
3804 memcg->use_hierarchy = val;
3815 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
3816 enum mem_cgroup_stat_index idx)
3818 struct mem_cgroup *iter;
3821 /* Per-cpu values can be negative, use a signed accumulator */
3822 for_each_mem_cgroup_tree(iter, memcg)
3823 val += mem_cgroup_read_stat(iter, idx);
3825 if (val < 0) /* race ? */
3830 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3834 if (!mem_cgroup_is_root(memcg)) {
3836 return res_counter_read_u64(&memcg->res, RES_USAGE);
3838 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
3841 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
3842 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
3845 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
3847 return val << PAGE_SHIFT;
3850 static ssize_t mem_cgroup_read(struct cgroup *cont, struct cftype *cft,
3851 struct file *file, char __user *buf,
3852 size_t nbytes, loff_t *ppos)
3854 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3857 int type, name, len;
3859 type = MEMFILE_TYPE(cft->private);
3860 name = MEMFILE_ATTR(cft->private);
3862 if (!do_swap_account && type == _MEMSWAP)
3867 if (name == RES_USAGE)
3868 val = mem_cgroup_usage(memcg, false);
3870 val = res_counter_read_u64(&memcg->res, name);
3873 if (name == RES_USAGE)
3874 val = mem_cgroup_usage(memcg, true);
3876 val = res_counter_read_u64(&memcg->memsw, name);
3882 len = scnprintf(str, sizeof(str), "%llu\n", (unsigned long long)val);
3883 return simple_read_from_buffer(buf, nbytes, ppos, str, len);
3886 * The user of this function is...
3889 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
3892 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3894 unsigned long long val;
3897 type = MEMFILE_TYPE(cft->private);
3898 name = MEMFILE_ATTR(cft->private);
3900 if (!do_swap_account && type == _MEMSWAP)
3905 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3909 /* This function does all necessary parse...reuse it */
3910 ret = res_counter_memparse_write_strategy(buffer, &val);
3914 ret = mem_cgroup_resize_limit(memcg, val);
3916 ret = mem_cgroup_resize_memsw_limit(memcg, val);
3918 case RES_SOFT_LIMIT:
3919 ret = res_counter_memparse_write_strategy(buffer, &val);
3923 * For memsw, soft limits are hard to implement in terms
3924 * of semantics, for now, we support soft limits for
3925 * control without swap
3928 ret = res_counter_set_soft_limit(&memcg->res, val);
3933 ret = -EINVAL; /* should be BUG() ? */
3939 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
3940 unsigned long long *mem_limit, unsigned long long *memsw_limit)
3942 struct cgroup *cgroup;
3943 unsigned long long min_limit, min_memsw_limit, tmp;
3945 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3946 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3947 cgroup = memcg->css.cgroup;
3948 if (!memcg->use_hierarchy)
3951 while (cgroup->parent) {
3952 cgroup = cgroup->parent;
3953 memcg = mem_cgroup_from_cont(cgroup);
3954 if (!memcg->use_hierarchy)
3956 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
3957 min_limit = min(min_limit, tmp);
3958 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3959 min_memsw_limit = min(min_memsw_limit, tmp);
3962 *mem_limit = min_limit;
3963 *memsw_limit = min_memsw_limit;
3966 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
3968 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3971 type = MEMFILE_TYPE(event);
3972 name = MEMFILE_ATTR(event);
3974 if (!do_swap_account && type == _MEMSWAP)
3980 res_counter_reset_max(&memcg->res);
3982 res_counter_reset_max(&memcg->memsw);
3986 res_counter_reset_failcnt(&memcg->res);
3988 res_counter_reset_failcnt(&memcg->memsw);
3995 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
3998 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
4002 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4003 struct cftype *cft, u64 val)
4005 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4007 if (val >= (1 << NR_MOVE_TYPE))
4010 * We check this value several times in both in can_attach() and
4011 * attach(), so we need cgroup lock to prevent this value from being
4015 memcg->move_charge_at_immigrate = val;
4021 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4022 struct cftype *cft, u64 val)
4029 /* For read statistics */
4047 struct mcs_total_stat {
4048 s64 stat[NR_MCS_STAT];
4054 } memcg_stat_strings[NR_MCS_STAT] = {
4055 {"cache", "total_cache"},
4056 {"rss", "total_rss"},
4057 {"mapped_file", "total_mapped_file"},
4058 {"pgpgin", "total_pgpgin"},
4059 {"pgpgout", "total_pgpgout"},
4060 {"swap", "total_swap"},
4061 {"pgfault", "total_pgfault"},
4062 {"pgmajfault", "total_pgmajfault"},
4063 {"inactive_anon", "total_inactive_anon"},
4064 {"active_anon", "total_active_anon"},
4065 {"inactive_file", "total_inactive_file"},
4066 {"active_file", "total_active_file"},
4067 {"unevictable", "total_unevictable"}
4072 mem_cgroup_get_local_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4077 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_CACHE);
4078 s->stat[MCS_CACHE] += val * PAGE_SIZE;
4079 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_RSS);
4080 s->stat[MCS_RSS] += val * PAGE_SIZE;
4081 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_FILE_MAPPED);
4082 s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
4083 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGIN);
4084 s->stat[MCS_PGPGIN] += val;
4085 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGOUT);
4086 s->stat[MCS_PGPGOUT] += val;
4087 if (do_swap_account) {
4088 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
4089 s->stat[MCS_SWAP] += val * PAGE_SIZE;
4091 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGFAULT);
4092 s->stat[MCS_PGFAULT] += val;
4093 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGMAJFAULT);
4094 s->stat[MCS_PGMAJFAULT] += val;
4097 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_ANON));
4098 s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
4099 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_ANON));
4100 s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
4101 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_FILE));
4102 s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
4103 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_FILE));
4104 s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
4105 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
4106 s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
4110 mem_cgroup_get_total_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4112 struct mem_cgroup *iter;
4114 for_each_mem_cgroup_tree(iter, memcg)
4115 mem_cgroup_get_local_stat(iter, s);
4119 static int mem_control_numa_stat_show(struct seq_file *m, void *arg)
4122 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
4123 unsigned long node_nr;
4124 struct cgroup *cont = m->private;
4125 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4127 total_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL);
4128 seq_printf(m, "total=%lu", total_nr);
4129 for_each_node_state(nid, N_HIGH_MEMORY) {
4130 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL);
4131 seq_printf(m, " N%d=%lu", nid, node_nr);
4135 file_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_FILE);
4136 seq_printf(m, "file=%lu", file_nr);
4137 for_each_node_state(nid, N_HIGH_MEMORY) {
4138 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4140 seq_printf(m, " N%d=%lu", nid, node_nr);
4144 anon_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_ANON);
4145 seq_printf(m, "anon=%lu", anon_nr);
4146 for_each_node_state(nid, N_HIGH_MEMORY) {
4147 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4149 seq_printf(m, " N%d=%lu", nid, node_nr);
4153 unevictable_nr = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
4154 seq_printf(m, "unevictable=%lu", unevictable_nr);
4155 for_each_node_state(nid, N_HIGH_MEMORY) {
4156 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4157 BIT(LRU_UNEVICTABLE));
4158 seq_printf(m, " N%d=%lu", nid, node_nr);
4163 #endif /* CONFIG_NUMA */
4165 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
4166 struct cgroup_map_cb *cb)
4168 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4169 struct mcs_total_stat mystat;
4172 memset(&mystat, 0, sizeof(mystat));
4173 mem_cgroup_get_local_stat(memcg, &mystat);
4176 for (i = 0; i < NR_MCS_STAT; i++) {
4177 if (i == MCS_SWAP && !do_swap_account)
4179 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
4182 /* Hierarchical information */
4184 unsigned long long limit, memsw_limit;
4185 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
4186 cb->fill(cb, "hierarchical_memory_limit", limit);
4187 if (do_swap_account)
4188 cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
4191 memset(&mystat, 0, sizeof(mystat));
4192 mem_cgroup_get_total_stat(memcg, &mystat);
4193 for (i = 0; i < NR_MCS_STAT; i++) {
4194 if (i == MCS_SWAP && !do_swap_account)
4196 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
4199 #ifdef CONFIG_DEBUG_VM
4202 struct mem_cgroup_per_zone *mz;
4203 struct zone_reclaim_stat *rstat;
4204 unsigned long recent_rotated[2] = {0, 0};
4205 unsigned long recent_scanned[2] = {0, 0};
4207 for_each_online_node(nid)
4208 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4209 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
4210 rstat = &mz->lruvec.reclaim_stat;
4212 recent_rotated[0] += rstat->recent_rotated[0];
4213 recent_rotated[1] += rstat->recent_rotated[1];
4214 recent_scanned[0] += rstat->recent_scanned[0];
4215 recent_scanned[1] += rstat->recent_scanned[1];
4217 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
4218 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
4219 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
4220 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
4227 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
4229 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4231 return mem_cgroup_swappiness(memcg);
4234 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
4237 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4238 struct mem_cgroup *parent;
4243 if (cgrp->parent == NULL)
4246 parent = mem_cgroup_from_cont(cgrp->parent);
4250 /* If under hierarchy, only empty-root can set this value */
4251 if ((parent->use_hierarchy) ||
4252 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4257 memcg->swappiness = val;
4264 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4266 struct mem_cgroup_threshold_ary *t;
4272 t = rcu_dereference(memcg->thresholds.primary);
4274 t = rcu_dereference(memcg->memsw_thresholds.primary);
4279 usage = mem_cgroup_usage(memcg, swap);
4282 * current_threshold points to threshold just below usage.
4283 * If it's not true, a threshold was crossed after last
4284 * call of __mem_cgroup_threshold().
4286 i = t->current_threshold;
4289 * Iterate backward over array of thresholds starting from
4290 * current_threshold and check if a threshold is crossed.
4291 * If none of thresholds below usage is crossed, we read
4292 * only one element of the array here.
4294 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4295 eventfd_signal(t->entries[i].eventfd, 1);
4297 /* i = current_threshold + 1 */
4301 * Iterate forward over array of thresholds starting from
4302 * current_threshold+1 and check if a threshold is crossed.
4303 * If none of thresholds above usage is crossed, we read
4304 * only one element of the array here.
4306 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4307 eventfd_signal(t->entries[i].eventfd, 1);
4309 /* Update current_threshold */
4310 t->current_threshold = i - 1;
4315 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4318 __mem_cgroup_threshold(memcg, false);
4319 if (do_swap_account)
4320 __mem_cgroup_threshold(memcg, true);
4322 memcg = parent_mem_cgroup(memcg);
4326 static int compare_thresholds(const void *a, const void *b)
4328 const struct mem_cgroup_threshold *_a = a;
4329 const struct mem_cgroup_threshold *_b = b;
4331 return _a->threshold - _b->threshold;
4334 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4336 struct mem_cgroup_eventfd_list *ev;
4338 list_for_each_entry(ev, &memcg->oom_notify, list)
4339 eventfd_signal(ev->eventfd, 1);
4343 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4345 struct mem_cgroup *iter;
4347 for_each_mem_cgroup_tree(iter, memcg)
4348 mem_cgroup_oom_notify_cb(iter);
4351 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
4352 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4354 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4355 struct mem_cgroup_thresholds *thresholds;
4356 struct mem_cgroup_threshold_ary *new;
4357 int type = MEMFILE_TYPE(cft->private);
4358 u64 threshold, usage;
4361 ret = res_counter_memparse_write_strategy(args, &threshold);
4365 mutex_lock(&memcg->thresholds_lock);
4368 thresholds = &memcg->thresholds;
4369 else if (type == _MEMSWAP)
4370 thresholds = &memcg->memsw_thresholds;
4374 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4376 /* Check if a threshold crossed before adding a new one */
4377 if (thresholds->primary)
4378 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4380 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4382 /* Allocate memory for new array of thresholds */
4383 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4391 /* Copy thresholds (if any) to new array */
4392 if (thresholds->primary) {
4393 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4394 sizeof(struct mem_cgroup_threshold));
4397 /* Add new threshold */
4398 new->entries[size - 1].eventfd = eventfd;
4399 new->entries[size - 1].threshold = threshold;
4401 /* Sort thresholds. Registering of new threshold isn't time-critical */
4402 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4403 compare_thresholds, NULL);
4405 /* Find current threshold */
4406 new->current_threshold = -1;
4407 for (i = 0; i < size; i++) {
4408 if (new->entries[i].threshold < usage) {
4410 * new->current_threshold will not be used until
4411 * rcu_assign_pointer(), so it's safe to increment
4414 ++new->current_threshold;
4418 /* Free old spare buffer and save old primary buffer as spare */
4419 kfree(thresholds->spare);
4420 thresholds->spare = thresholds->primary;
4422 rcu_assign_pointer(thresholds->primary, new);
4424 /* To be sure that nobody uses thresholds */
4428 mutex_unlock(&memcg->thresholds_lock);
4433 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
4434 struct cftype *cft, struct eventfd_ctx *eventfd)
4436 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4437 struct mem_cgroup_thresholds *thresholds;
4438 struct mem_cgroup_threshold_ary *new;
4439 int type = MEMFILE_TYPE(cft->private);
4443 mutex_lock(&memcg->thresholds_lock);
4445 thresholds = &memcg->thresholds;
4446 else if (type == _MEMSWAP)
4447 thresholds = &memcg->memsw_thresholds;
4451 if (!thresholds->primary)
4454 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4456 /* Check if a threshold crossed before removing */
4457 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4459 /* Calculate new number of threshold */
4461 for (i = 0; i < thresholds->primary->size; i++) {
4462 if (thresholds->primary->entries[i].eventfd != eventfd)
4466 new = thresholds->spare;
4468 /* Set thresholds array to NULL if we don't have thresholds */
4477 /* Copy thresholds and find current threshold */
4478 new->current_threshold = -1;
4479 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4480 if (thresholds->primary->entries[i].eventfd == eventfd)
4483 new->entries[j] = thresholds->primary->entries[i];
4484 if (new->entries[j].threshold < usage) {
4486 * new->current_threshold will not be used
4487 * until rcu_assign_pointer(), so it's safe to increment
4490 ++new->current_threshold;
4496 /* Swap primary and spare array */
4497 thresholds->spare = thresholds->primary;
4498 /* If all events are unregistered, free the spare array */
4500 kfree(thresholds->spare);
4501 thresholds->spare = NULL;
4504 rcu_assign_pointer(thresholds->primary, new);
4506 /* To be sure that nobody uses thresholds */
4509 mutex_unlock(&memcg->thresholds_lock);
4512 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4513 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4515 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4516 struct mem_cgroup_eventfd_list *event;
4517 int type = MEMFILE_TYPE(cft->private);
4519 BUG_ON(type != _OOM_TYPE);
4520 event = kmalloc(sizeof(*event), GFP_KERNEL);
4524 spin_lock(&memcg_oom_lock);
4526 event->eventfd = eventfd;
4527 list_add(&event->list, &memcg->oom_notify);
4529 /* already in OOM ? */
4530 if (atomic_read(&memcg->under_oom))
4531 eventfd_signal(eventfd, 1);
4532 spin_unlock(&memcg_oom_lock);
4537 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4538 struct cftype *cft, struct eventfd_ctx *eventfd)
4540 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4541 struct mem_cgroup_eventfd_list *ev, *tmp;
4542 int type = MEMFILE_TYPE(cft->private);
4544 BUG_ON(type != _OOM_TYPE);
4546 spin_lock(&memcg_oom_lock);
4548 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4549 if (ev->eventfd == eventfd) {
4550 list_del(&ev->list);
4555 spin_unlock(&memcg_oom_lock);
4558 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4559 struct cftype *cft, struct cgroup_map_cb *cb)
4561 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4563 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
4565 if (atomic_read(&memcg->under_oom))
4566 cb->fill(cb, "under_oom", 1);
4568 cb->fill(cb, "under_oom", 0);
4572 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4573 struct cftype *cft, u64 val)
4575 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4576 struct mem_cgroup *parent;
4578 /* cannot set to root cgroup and only 0 and 1 are allowed */
4579 if (!cgrp->parent || !((val == 0) || (val == 1)))
4582 parent = mem_cgroup_from_cont(cgrp->parent);
4585 /* oom-kill-disable is a flag for subhierarchy. */
4586 if ((parent->use_hierarchy) ||
4587 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4591 memcg->oom_kill_disable = val;
4593 memcg_oom_recover(memcg);
4599 static const struct file_operations mem_control_numa_stat_file_operations = {
4601 .llseek = seq_lseek,
4602 .release = single_release,
4605 static int mem_control_numa_stat_open(struct inode *unused, struct file *file)
4607 struct cgroup *cont = file->f_dentry->d_parent->d_fsdata;
4609 file->f_op = &mem_control_numa_stat_file_operations;
4610 return single_open(file, mem_control_numa_stat_show, cont);
4612 #endif /* CONFIG_NUMA */
4614 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_KMEM
4615 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
4617 return mem_cgroup_sockets_init(memcg, ss);
4620 static void kmem_cgroup_destroy(struct mem_cgroup *memcg)
4622 mem_cgroup_sockets_destroy(memcg);
4625 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
4630 static void kmem_cgroup_destroy(struct mem_cgroup *memcg)
4635 static struct cftype mem_cgroup_files[] = {
4637 .name = "usage_in_bytes",
4638 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4639 .read = mem_cgroup_read,
4640 .register_event = mem_cgroup_usage_register_event,
4641 .unregister_event = mem_cgroup_usage_unregister_event,
4644 .name = "max_usage_in_bytes",
4645 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4646 .trigger = mem_cgroup_reset,
4647 .read = mem_cgroup_read,
4650 .name = "limit_in_bytes",
4651 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4652 .write_string = mem_cgroup_write,
4653 .read = mem_cgroup_read,
4656 .name = "soft_limit_in_bytes",
4657 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4658 .write_string = mem_cgroup_write,
4659 .read = mem_cgroup_read,
4663 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4664 .trigger = mem_cgroup_reset,
4665 .read = mem_cgroup_read,
4669 .read_map = mem_control_stat_show,
4672 .name = "force_empty",
4673 .trigger = mem_cgroup_force_empty_write,
4676 .name = "use_hierarchy",
4677 .write_u64 = mem_cgroup_hierarchy_write,
4678 .read_u64 = mem_cgroup_hierarchy_read,
4681 .name = "swappiness",
4682 .read_u64 = mem_cgroup_swappiness_read,
4683 .write_u64 = mem_cgroup_swappiness_write,
4686 .name = "move_charge_at_immigrate",
4687 .read_u64 = mem_cgroup_move_charge_read,
4688 .write_u64 = mem_cgroup_move_charge_write,
4691 .name = "oom_control",
4692 .read_map = mem_cgroup_oom_control_read,
4693 .write_u64 = mem_cgroup_oom_control_write,
4694 .register_event = mem_cgroup_oom_register_event,
4695 .unregister_event = mem_cgroup_oom_unregister_event,
4696 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4700 .name = "numa_stat",
4701 .open = mem_control_numa_stat_open,
4705 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4707 .name = "memsw.usage_in_bytes",
4708 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4709 .read = mem_cgroup_read,
4710 .register_event = mem_cgroup_usage_register_event,
4711 .unregister_event = mem_cgroup_usage_unregister_event,
4714 .name = "memsw.max_usage_in_bytes",
4715 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4716 .trigger = mem_cgroup_reset,
4717 .read = mem_cgroup_read,
4720 .name = "memsw.limit_in_bytes",
4721 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4722 .write_string = mem_cgroup_write,
4723 .read = mem_cgroup_read,
4726 .name = "memsw.failcnt",
4727 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4728 .trigger = mem_cgroup_reset,
4729 .read = mem_cgroup_read,
4732 { }, /* terminate */
4735 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4737 struct mem_cgroup_per_node *pn;
4738 struct mem_cgroup_per_zone *mz;
4740 int zone, tmp = node;
4742 * This routine is called against possible nodes.
4743 * But it's BUG to call kmalloc() against offline node.
4745 * TODO: this routine can waste much memory for nodes which will
4746 * never be onlined. It's better to use memory hotplug callback
4749 if (!node_state(node, N_NORMAL_MEMORY))
4751 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4755 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4756 mz = &pn->zoneinfo[zone];
4758 INIT_LIST_HEAD(&mz->lruvec.lists[lru]);
4759 mz->usage_in_excess = 0;
4760 mz->on_tree = false;
4763 memcg->info.nodeinfo[node] = pn;
4767 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4769 kfree(memcg->info.nodeinfo[node]);
4772 static struct mem_cgroup *mem_cgroup_alloc(void)
4774 struct mem_cgroup *memcg;
4775 int size = sizeof(struct mem_cgroup);
4777 /* Can be very big if MAX_NUMNODES is very big */
4778 if (size < PAGE_SIZE)
4779 memcg = kzalloc(size, GFP_KERNEL);
4781 memcg = vzalloc(size);
4786 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4789 spin_lock_init(&memcg->pcp_counter_lock);
4793 if (size < PAGE_SIZE)
4801 * Helpers for freeing a vzalloc()ed mem_cgroup by RCU,
4802 * but in process context. The work_freeing structure is overlaid
4803 * on the rcu_freeing structure, which itself is overlaid on memsw.
4805 static void vfree_work(struct work_struct *work)
4807 struct mem_cgroup *memcg;
4809 memcg = container_of(work, struct mem_cgroup, work_freeing);
4812 static void vfree_rcu(struct rcu_head *rcu_head)
4814 struct mem_cgroup *memcg;
4816 memcg = container_of(rcu_head, struct mem_cgroup, rcu_freeing);
4817 INIT_WORK(&memcg->work_freeing, vfree_work);
4818 schedule_work(&memcg->work_freeing);
4822 * At destroying mem_cgroup, references from swap_cgroup can remain.
4823 * (scanning all at force_empty is too costly...)
4825 * Instead of clearing all references at force_empty, we remember
4826 * the number of reference from swap_cgroup and free mem_cgroup when
4827 * it goes down to 0.
4829 * Removal of cgroup itself succeeds regardless of refs from swap.
4832 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4836 mem_cgroup_remove_from_trees(memcg);
4837 free_css_id(&mem_cgroup_subsys, &memcg->css);
4840 free_mem_cgroup_per_zone_info(memcg, node);
4842 free_percpu(memcg->stat);
4843 if (sizeof(struct mem_cgroup) < PAGE_SIZE)
4844 kfree_rcu(memcg, rcu_freeing);
4846 call_rcu(&memcg->rcu_freeing, vfree_rcu);
4849 static void mem_cgroup_get(struct mem_cgroup *memcg)
4851 atomic_inc(&memcg->refcnt);
4854 static void __mem_cgroup_put(struct mem_cgroup *memcg, int count)
4856 if (atomic_sub_and_test(count, &memcg->refcnt)) {
4857 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4858 __mem_cgroup_free(memcg);
4860 mem_cgroup_put(parent);
4864 static void mem_cgroup_put(struct mem_cgroup *memcg)
4866 __mem_cgroup_put(memcg, 1);
4870 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4872 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4874 if (!memcg->res.parent)
4876 return mem_cgroup_from_res_counter(memcg->res.parent, res);
4878 EXPORT_SYMBOL(parent_mem_cgroup);
4880 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4881 static void __init enable_swap_cgroup(void)
4883 if (!mem_cgroup_disabled() && really_do_swap_account)
4884 do_swap_account = 1;
4887 static void __init enable_swap_cgroup(void)
4892 static int mem_cgroup_soft_limit_tree_init(void)
4894 struct mem_cgroup_tree_per_node *rtpn;
4895 struct mem_cgroup_tree_per_zone *rtpz;
4896 int tmp, node, zone;
4898 for_each_node(node) {
4900 if (!node_state(node, N_NORMAL_MEMORY))
4902 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
4906 soft_limit_tree.rb_tree_per_node[node] = rtpn;
4908 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4909 rtpz = &rtpn->rb_tree_per_zone[zone];
4910 rtpz->rb_root = RB_ROOT;
4911 spin_lock_init(&rtpz->lock);
4917 for_each_node(node) {
4918 if (!soft_limit_tree.rb_tree_per_node[node])
4920 kfree(soft_limit_tree.rb_tree_per_node[node]);
4921 soft_limit_tree.rb_tree_per_node[node] = NULL;
4927 static struct cgroup_subsys_state * __ref
4928 mem_cgroup_create(struct cgroup *cont)
4930 struct mem_cgroup *memcg, *parent;
4931 long error = -ENOMEM;
4934 memcg = mem_cgroup_alloc();
4936 return ERR_PTR(error);
4939 if (alloc_mem_cgroup_per_zone_info(memcg, node))
4943 if (cont->parent == NULL) {
4945 enable_swap_cgroup();
4947 if (mem_cgroup_soft_limit_tree_init())
4949 root_mem_cgroup = memcg;
4950 for_each_possible_cpu(cpu) {
4951 struct memcg_stock_pcp *stock =
4952 &per_cpu(memcg_stock, cpu);
4953 INIT_WORK(&stock->work, drain_local_stock);
4955 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
4957 parent = mem_cgroup_from_cont(cont->parent);
4958 memcg->use_hierarchy = parent->use_hierarchy;
4959 memcg->oom_kill_disable = parent->oom_kill_disable;
4962 if (parent && parent->use_hierarchy) {
4963 res_counter_init(&memcg->res, &parent->res);
4964 res_counter_init(&memcg->memsw, &parent->memsw);
4966 * We increment refcnt of the parent to ensure that we can
4967 * safely access it on res_counter_charge/uncharge.
4968 * This refcnt will be decremented when freeing this
4969 * mem_cgroup(see mem_cgroup_put).
4971 mem_cgroup_get(parent);
4973 res_counter_init(&memcg->res, NULL);
4974 res_counter_init(&memcg->memsw, NULL);
4976 memcg->last_scanned_node = MAX_NUMNODES;
4977 INIT_LIST_HEAD(&memcg->oom_notify);
4980 memcg->swappiness = mem_cgroup_swappiness(parent);
4981 atomic_set(&memcg->refcnt, 1);
4982 memcg->move_charge_at_immigrate = 0;
4983 mutex_init(&memcg->thresholds_lock);
4984 spin_lock_init(&memcg->move_lock);
4986 error = memcg_init_kmem(memcg, &mem_cgroup_subsys);
4989 * We call put now because our (and parent's) refcnts
4990 * are already in place. mem_cgroup_put() will internally
4991 * call __mem_cgroup_free, so return directly
4993 mem_cgroup_put(memcg);
4994 return ERR_PTR(error);
4998 __mem_cgroup_free(memcg);
4999 return ERR_PTR(error);
5002 static int mem_cgroup_pre_destroy(struct cgroup *cont)
5004 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5006 return mem_cgroup_force_empty(memcg, false);
5009 static void mem_cgroup_destroy(struct cgroup *cont)
5011 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5013 kmem_cgroup_destroy(memcg);
5015 mem_cgroup_put(memcg);
5019 /* Handlers for move charge at task migration. */
5020 #define PRECHARGE_COUNT_AT_ONCE 256
5021 static int mem_cgroup_do_precharge(unsigned long count)
5024 int batch_count = PRECHARGE_COUNT_AT_ONCE;
5025 struct mem_cgroup *memcg = mc.to;
5027 if (mem_cgroup_is_root(memcg)) {
5028 mc.precharge += count;
5029 /* we don't need css_get for root */
5032 /* try to charge at once */
5034 struct res_counter *dummy;
5036 * "memcg" cannot be under rmdir() because we've already checked
5037 * by cgroup_lock_live_cgroup() that it is not removed and we
5038 * are still under the same cgroup_mutex. So we can postpone
5041 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
5043 if (do_swap_account && res_counter_charge(&memcg->memsw,
5044 PAGE_SIZE * count, &dummy)) {
5045 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
5048 mc.precharge += count;
5052 /* fall back to one by one charge */
5054 if (signal_pending(current)) {
5058 if (!batch_count--) {
5059 batch_count = PRECHARGE_COUNT_AT_ONCE;
5062 ret = __mem_cgroup_try_charge(NULL,
5063 GFP_KERNEL, 1, &memcg, false);
5065 /* mem_cgroup_clear_mc() will do uncharge later */
5073 * get_mctgt_type - get target type of moving charge
5074 * @vma: the vma the pte to be checked belongs
5075 * @addr: the address corresponding to the pte to be checked
5076 * @ptent: the pte to be checked
5077 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5080 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5081 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5082 * move charge. if @target is not NULL, the page is stored in target->page
5083 * with extra refcnt got(Callers should handle it).
5084 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5085 * target for charge migration. if @target is not NULL, the entry is stored
5088 * Called with pte lock held.
5095 enum mc_target_type {
5101 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5102 unsigned long addr, pte_t ptent)
5104 struct page *page = vm_normal_page(vma, addr, ptent);
5106 if (!page || !page_mapped(page))
5108 if (PageAnon(page)) {
5109 /* we don't move shared anon */
5112 } else if (!move_file())
5113 /* we ignore mapcount for file pages */
5115 if (!get_page_unless_zero(page))
5122 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5123 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5125 struct page *page = NULL;
5126 swp_entry_t ent = pte_to_swp_entry(ptent);
5128 if (!move_anon() || non_swap_entry(ent))
5131 * Because lookup_swap_cache() updates some statistics counter,
5132 * we call find_get_page() with swapper_space directly.
5134 page = find_get_page(&swapper_space, ent.val);
5135 if (do_swap_account)
5136 entry->val = ent.val;
5141 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5142 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5148 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5149 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5151 struct page *page = NULL;
5152 struct inode *inode;
5153 struct address_space *mapping;
5156 if (!vma->vm_file) /* anonymous vma */
5161 inode = vma->vm_file->f_path.dentry->d_inode;
5162 mapping = vma->vm_file->f_mapping;
5163 if (pte_none(ptent))
5164 pgoff = linear_page_index(vma, addr);
5165 else /* pte_file(ptent) is true */
5166 pgoff = pte_to_pgoff(ptent);
5168 /* page is moved even if it's not RSS of this task(page-faulted). */
5169 page = find_get_page(mapping, pgoff);
5172 /* shmem/tmpfs may report page out on swap: account for that too. */
5173 if (radix_tree_exceptional_entry(page)) {
5174 swp_entry_t swap = radix_to_swp_entry(page);
5175 if (do_swap_account)
5177 page = find_get_page(&swapper_space, swap.val);
5183 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5184 unsigned long addr, pte_t ptent, union mc_target *target)
5186 struct page *page = NULL;
5187 struct page_cgroup *pc;
5188 enum mc_target_type ret = MC_TARGET_NONE;
5189 swp_entry_t ent = { .val = 0 };
5191 if (pte_present(ptent))
5192 page = mc_handle_present_pte(vma, addr, ptent);
5193 else if (is_swap_pte(ptent))
5194 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5195 else if (pte_none(ptent) || pte_file(ptent))
5196 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5198 if (!page && !ent.val)
5201 pc = lookup_page_cgroup(page);
5203 * Do only loose check w/o page_cgroup lock.
5204 * mem_cgroup_move_account() checks the pc is valid or not under
5207 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5208 ret = MC_TARGET_PAGE;
5210 target->page = page;
5212 if (!ret || !target)
5215 /* There is a swap entry and a page doesn't exist or isn't charged */
5216 if (ent.val && !ret &&
5217 css_id(&mc.from->css) == lookup_swap_cgroup_id(ent)) {
5218 ret = MC_TARGET_SWAP;
5225 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5227 * We don't consider swapping or file mapped pages because THP does not
5228 * support them for now.
5229 * Caller should make sure that pmd_trans_huge(pmd) is true.
5231 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5232 unsigned long addr, pmd_t pmd, union mc_target *target)
5234 struct page *page = NULL;
5235 struct page_cgroup *pc;
5236 enum mc_target_type ret = MC_TARGET_NONE;
5238 page = pmd_page(pmd);
5239 VM_BUG_ON(!page || !PageHead(page));
5242 pc = lookup_page_cgroup(page);
5243 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5244 ret = MC_TARGET_PAGE;
5247 target->page = page;
5253 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5254 unsigned long addr, pmd_t pmd, union mc_target *target)
5256 return MC_TARGET_NONE;
5260 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5261 unsigned long addr, unsigned long end,
5262 struct mm_walk *walk)
5264 struct vm_area_struct *vma = walk->private;
5268 if (pmd_trans_huge_lock(pmd, vma) == 1) {
5269 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5270 mc.precharge += HPAGE_PMD_NR;
5271 spin_unlock(&vma->vm_mm->page_table_lock);
5275 if (pmd_trans_unstable(pmd))
5277 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5278 for (; addr != end; pte++, addr += PAGE_SIZE)
5279 if (get_mctgt_type(vma, addr, *pte, NULL))
5280 mc.precharge++; /* increment precharge temporarily */
5281 pte_unmap_unlock(pte - 1, ptl);
5287 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5289 unsigned long precharge;
5290 struct vm_area_struct *vma;
5292 down_read(&mm->mmap_sem);
5293 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5294 struct mm_walk mem_cgroup_count_precharge_walk = {
5295 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5299 if (is_vm_hugetlb_page(vma))
5301 walk_page_range(vma->vm_start, vma->vm_end,
5302 &mem_cgroup_count_precharge_walk);
5304 up_read(&mm->mmap_sem);
5306 precharge = mc.precharge;
5312 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5314 unsigned long precharge = mem_cgroup_count_precharge(mm);
5316 VM_BUG_ON(mc.moving_task);
5317 mc.moving_task = current;
5318 return mem_cgroup_do_precharge(precharge);
5321 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5322 static void __mem_cgroup_clear_mc(void)
5324 struct mem_cgroup *from = mc.from;
5325 struct mem_cgroup *to = mc.to;
5327 /* we must uncharge all the leftover precharges from mc.to */
5329 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
5333 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5334 * we must uncharge here.
5336 if (mc.moved_charge) {
5337 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
5338 mc.moved_charge = 0;
5340 /* we must fixup refcnts and charges */
5341 if (mc.moved_swap) {
5342 /* uncharge swap account from the old cgroup */
5343 if (!mem_cgroup_is_root(mc.from))
5344 res_counter_uncharge(&mc.from->memsw,
5345 PAGE_SIZE * mc.moved_swap);
5346 __mem_cgroup_put(mc.from, mc.moved_swap);
5348 if (!mem_cgroup_is_root(mc.to)) {
5350 * we charged both to->res and to->memsw, so we should
5353 res_counter_uncharge(&mc.to->res,
5354 PAGE_SIZE * mc.moved_swap);
5356 /* we've already done mem_cgroup_get(mc.to) */
5359 memcg_oom_recover(from);
5360 memcg_oom_recover(to);
5361 wake_up_all(&mc.waitq);
5364 static void mem_cgroup_clear_mc(void)
5366 struct mem_cgroup *from = mc.from;
5369 * we must clear moving_task before waking up waiters at the end of
5372 mc.moving_task = NULL;
5373 __mem_cgroup_clear_mc();
5374 spin_lock(&mc.lock);
5377 spin_unlock(&mc.lock);
5378 mem_cgroup_end_move(from);
5381 static int mem_cgroup_can_attach(struct cgroup *cgroup,
5382 struct cgroup_taskset *tset)
5384 struct task_struct *p = cgroup_taskset_first(tset);
5386 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup);
5388 if (memcg->move_charge_at_immigrate) {
5389 struct mm_struct *mm;
5390 struct mem_cgroup *from = mem_cgroup_from_task(p);
5392 VM_BUG_ON(from == memcg);
5394 mm = get_task_mm(p);
5397 /* We move charges only when we move a owner of the mm */
5398 if (mm->owner == p) {
5401 VM_BUG_ON(mc.precharge);
5402 VM_BUG_ON(mc.moved_charge);
5403 VM_BUG_ON(mc.moved_swap);
5404 mem_cgroup_start_move(from);
5405 spin_lock(&mc.lock);
5408 spin_unlock(&mc.lock);
5409 /* We set mc.moving_task later */
5411 ret = mem_cgroup_precharge_mc(mm);
5413 mem_cgroup_clear_mc();
5420 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
5421 struct cgroup_taskset *tset)
5423 mem_cgroup_clear_mc();
5426 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5427 unsigned long addr, unsigned long end,
5428 struct mm_walk *walk)
5431 struct vm_area_struct *vma = walk->private;
5434 enum mc_target_type target_type;
5435 union mc_target target;
5437 struct page_cgroup *pc;
5440 * We don't take compound_lock() here but no race with splitting thp
5442 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
5443 * under splitting, which means there's no concurrent thp split,
5444 * - if another thread runs into split_huge_page() just after we
5445 * entered this if-block, the thread must wait for page table lock
5446 * to be unlocked in __split_huge_page_splitting(), where the main
5447 * part of thp split is not executed yet.
5449 if (pmd_trans_huge_lock(pmd, vma) == 1) {
5450 if (mc.precharge < HPAGE_PMD_NR) {
5451 spin_unlock(&vma->vm_mm->page_table_lock);
5454 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5455 if (target_type == MC_TARGET_PAGE) {
5457 if (!isolate_lru_page(page)) {
5458 pc = lookup_page_cgroup(page);
5459 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
5462 mc.precharge -= HPAGE_PMD_NR;
5463 mc.moved_charge += HPAGE_PMD_NR;
5465 putback_lru_page(page);
5469 spin_unlock(&vma->vm_mm->page_table_lock);
5473 if (pmd_trans_unstable(pmd))
5476 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5477 for (; addr != end; addr += PAGE_SIZE) {
5478 pte_t ptent = *(pte++);
5484 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5485 case MC_TARGET_PAGE:
5487 if (isolate_lru_page(page))
5489 pc = lookup_page_cgroup(page);
5490 if (!mem_cgroup_move_account(page, 1, pc,
5491 mc.from, mc.to, false)) {
5493 /* we uncharge from mc.from later. */
5496 putback_lru_page(page);
5497 put: /* get_mctgt_type() gets the page */
5500 case MC_TARGET_SWAP:
5502 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5504 /* we fixup refcnts and charges later. */
5512 pte_unmap_unlock(pte - 1, ptl);
5517 * We have consumed all precharges we got in can_attach().
5518 * We try charge one by one, but don't do any additional
5519 * charges to mc.to if we have failed in charge once in attach()
5522 ret = mem_cgroup_do_precharge(1);
5530 static void mem_cgroup_move_charge(struct mm_struct *mm)
5532 struct vm_area_struct *vma;
5534 lru_add_drain_all();
5536 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5538 * Someone who are holding the mmap_sem might be waiting in
5539 * waitq. So we cancel all extra charges, wake up all waiters,
5540 * and retry. Because we cancel precharges, we might not be able
5541 * to move enough charges, but moving charge is a best-effort
5542 * feature anyway, so it wouldn't be a big problem.
5544 __mem_cgroup_clear_mc();
5548 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5550 struct mm_walk mem_cgroup_move_charge_walk = {
5551 .pmd_entry = mem_cgroup_move_charge_pte_range,
5555 if (is_vm_hugetlb_page(vma))
5557 ret = walk_page_range(vma->vm_start, vma->vm_end,
5558 &mem_cgroup_move_charge_walk);
5561 * means we have consumed all precharges and failed in
5562 * doing additional charge. Just abandon here.
5566 up_read(&mm->mmap_sem);
5569 static void mem_cgroup_move_task(struct cgroup *cont,
5570 struct cgroup_taskset *tset)
5572 struct task_struct *p = cgroup_taskset_first(tset);
5573 struct mm_struct *mm = get_task_mm(p);
5577 mem_cgroup_move_charge(mm);
5581 mem_cgroup_clear_mc();
5583 #else /* !CONFIG_MMU */
5584 static int mem_cgroup_can_attach(struct cgroup *cgroup,
5585 struct cgroup_taskset *tset)
5589 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
5590 struct cgroup_taskset *tset)
5593 static void mem_cgroup_move_task(struct cgroup *cont,
5594 struct cgroup_taskset *tset)
5599 struct cgroup_subsys mem_cgroup_subsys = {
5601 .subsys_id = mem_cgroup_subsys_id,
5602 .create = mem_cgroup_create,
5603 .pre_destroy = mem_cgroup_pre_destroy,
5604 .destroy = mem_cgroup_destroy,
5605 .can_attach = mem_cgroup_can_attach,
5606 .cancel_attach = mem_cgroup_cancel_attach,
5607 .attach = mem_cgroup_move_task,
5608 .base_cftypes = mem_cgroup_files,
5611 .__DEPRECATED_clear_css_refs = true,
5614 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5615 static int __init enable_swap_account(char *s)
5617 /* consider enabled if no parameter or 1 is given */
5618 if (!strcmp(s, "1"))
5619 really_do_swap_account = 1;
5620 else if (!strcmp(s, "0"))
5621 really_do_swap_account = 0;
5624 __setup("swapaccount=", enable_swap_account);