1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * Kernel Memory Controller
14 * Copyright (C) 2012 Parallels Inc. and Google Inc.
15 * Authors: Glauber Costa and Suleiman Souhlal
17 * This program is free software; you can redistribute it and/or modify
18 * it under the terms of the GNU General Public License as published by
19 * the Free Software Foundation; either version 2 of the License, or
20 * (at your option) any later version.
22 * This program is distributed in the hope that it will be useful,
23 * but WITHOUT ANY WARRANTY; without even the implied warranty of
24 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
25 * GNU General Public License for more details.
28 #include <linux/res_counter.h>
29 #include <linux/memcontrol.h>
30 #include <linux/cgroup.h>
32 #include <linux/hugetlb.h>
33 #include <linux/pagemap.h>
34 #include <linux/smp.h>
35 #include <linux/page-flags.h>
36 #include <linux/backing-dev.h>
37 #include <linux/bit_spinlock.h>
38 #include <linux/rcupdate.h>
39 #include <linux/limits.h>
40 #include <linux/export.h>
41 #include <linux/mutex.h>
42 #include <linux/rbtree.h>
43 #include <linux/slab.h>
44 #include <linux/swap.h>
45 #include <linux/swapops.h>
46 #include <linux/spinlock.h>
47 #include <linux/eventfd.h>
48 #include <linux/sort.h>
50 #include <linux/seq_file.h>
51 #include <linux/vmalloc.h>
52 #include <linux/mm_inline.h>
53 #include <linux/page_cgroup.h>
54 #include <linux/cpu.h>
55 #include <linux/oom.h>
59 #include <net/tcp_memcontrol.h>
61 #include <asm/uaccess.h>
63 #include <trace/events/vmscan.h>
65 struct cgroup_subsys mem_cgroup_subsys __read_mostly;
66 EXPORT_SYMBOL(mem_cgroup_subsys);
68 #define MEM_CGROUP_RECLAIM_RETRIES 5
69 static struct mem_cgroup *root_mem_cgroup __read_mostly;
71 #ifdef CONFIG_MEMCG_SWAP
72 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
73 int do_swap_account __read_mostly;
75 /* for remember boot option*/
76 #ifdef CONFIG_MEMCG_SWAP_ENABLED
77 static int really_do_swap_account __initdata = 1;
79 static int really_do_swap_account __initdata = 0;
83 #define do_swap_account 0
88 * Statistics for memory cgroup.
90 enum mem_cgroup_stat_index {
92 * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
94 MEM_CGROUP_STAT_CACHE, /* # of pages charged as cache */
95 MEM_CGROUP_STAT_RSS, /* # of pages charged as anon rss */
96 MEM_CGROUP_STAT_FILE_MAPPED, /* # of pages charged as file rss */
97 MEM_CGROUP_STAT_SWAP, /* # of pages, swapped out */
98 MEM_CGROUP_STAT_NSTATS,
101 static const char * const mem_cgroup_stat_names[] = {
108 enum mem_cgroup_events_index {
109 MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */
110 MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */
111 MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */
112 MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */
113 MEM_CGROUP_EVENTS_NSTATS,
116 static const char * const mem_cgroup_events_names[] = {
124 * Per memcg event counter is incremented at every pagein/pageout. With THP,
125 * it will be incremated by the number of pages. This counter is used for
126 * for trigger some periodic events. This is straightforward and better
127 * than using jiffies etc. to handle periodic memcg event.
129 enum mem_cgroup_events_target {
130 MEM_CGROUP_TARGET_THRESH,
131 MEM_CGROUP_TARGET_SOFTLIMIT,
132 MEM_CGROUP_TARGET_NUMAINFO,
135 #define THRESHOLDS_EVENTS_TARGET 128
136 #define SOFTLIMIT_EVENTS_TARGET 1024
137 #define NUMAINFO_EVENTS_TARGET 1024
139 struct mem_cgroup_stat_cpu {
140 long count[MEM_CGROUP_STAT_NSTATS];
141 unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
142 unsigned long nr_page_events;
143 unsigned long targets[MEM_CGROUP_NTARGETS];
146 struct mem_cgroup_reclaim_iter {
147 /* css_id of the last scanned hierarchy member */
149 /* scan generation, increased every round-trip */
150 unsigned int generation;
154 * per-zone information in memory controller.
156 struct mem_cgroup_per_zone {
157 struct lruvec lruvec;
158 unsigned long lru_size[NR_LRU_LISTS];
160 struct mem_cgroup_reclaim_iter reclaim_iter[DEF_PRIORITY + 1];
162 struct rb_node tree_node; /* RB tree node */
163 unsigned long long usage_in_excess;/* Set to the value by which */
164 /* the soft limit is exceeded*/
166 struct mem_cgroup *memcg; /* Back pointer, we cannot */
167 /* use container_of */
170 struct mem_cgroup_per_node {
171 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
174 struct mem_cgroup_lru_info {
175 struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
179 * Cgroups above their limits are maintained in a RB-Tree, independent of
180 * their hierarchy representation
183 struct mem_cgroup_tree_per_zone {
184 struct rb_root rb_root;
188 struct mem_cgroup_tree_per_node {
189 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
192 struct mem_cgroup_tree {
193 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
196 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
198 struct mem_cgroup_threshold {
199 struct eventfd_ctx *eventfd;
204 struct mem_cgroup_threshold_ary {
205 /* An array index points to threshold just below or equal to usage. */
206 int current_threshold;
207 /* Size of entries[] */
209 /* Array of thresholds */
210 struct mem_cgroup_threshold entries[0];
213 struct mem_cgroup_thresholds {
214 /* Primary thresholds array */
215 struct mem_cgroup_threshold_ary *primary;
217 * Spare threshold array.
218 * This is needed to make mem_cgroup_unregister_event() "never fail".
219 * It must be able to store at least primary->size - 1 entries.
221 struct mem_cgroup_threshold_ary *spare;
225 struct mem_cgroup_eventfd_list {
226 struct list_head list;
227 struct eventfd_ctx *eventfd;
230 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
231 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
234 * The memory controller data structure. The memory controller controls both
235 * page cache and RSS per cgroup. We would eventually like to provide
236 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
237 * to help the administrator determine what knobs to tune.
239 * TODO: Add a water mark for the memory controller. Reclaim will begin when
240 * we hit the water mark. May be even add a low water mark, such that
241 * no reclaim occurs from a cgroup at it's low water mark, this is
242 * a feature that will be implemented much later in the future.
245 struct cgroup_subsys_state css;
247 * the counter to account for memory usage
249 struct res_counter res;
253 * the counter to account for mem+swap usage.
255 struct res_counter memsw;
258 * rcu_freeing is used only when freeing struct mem_cgroup,
259 * so put it into a union to avoid wasting more memory.
260 * It must be disjoint from the css field. It could be
261 * in a union with the res field, but res plays a much
262 * larger part in mem_cgroup life than memsw, and might
263 * be of interest, even at time of free, when debugging.
264 * So share rcu_head with the less interesting memsw.
266 struct rcu_head rcu_freeing;
268 * We also need some space for a worker in deferred freeing.
269 * By the time we call it, rcu_freeing is no longer in use.
271 struct work_struct work_freeing;
275 * the counter to account for kernel memory usage.
277 struct res_counter kmem;
279 * Per cgroup active and inactive list, similar to the
280 * per zone LRU lists.
282 struct mem_cgroup_lru_info info;
283 int last_scanned_node;
285 nodemask_t scan_nodes;
286 atomic_t numainfo_events;
287 atomic_t numainfo_updating;
290 * Should the accounting and control be hierarchical, per subtree?
293 unsigned long kmem_account_flags; /* See KMEM_ACCOUNTED_*, below */
301 /* OOM-Killer disable */
302 int oom_kill_disable;
304 /* set when res.limit == memsw.limit */
305 bool memsw_is_minimum;
307 /* protect arrays of thresholds */
308 struct mutex thresholds_lock;
310 /* thresholds for memory usage. RCU-protected */
311 struct mem_cgroup_thresholds thresholds;
313 /* thresholds for mem+swap usage. RCU-protected */
314 struct mem_cgroup_thresholds memsw_thresholds;
316 /* For oom notifier event fd */
317 struct list_head oom_notify;
320 * Should we move charges of a task when a task is moved into this
321 * mem_cgroup ? And what type of charges should we move ?
323 unsigned long move_charge_at_immigrate;
325 * set > 0 if pages under this cgroup are moving to other cgroup.
327 atomic_t moving_account;
328 /* taken only while moving_account > 0 */
329 spinlock_t move_lock;
333 struct mem_cgroup_stat_cpu __percpu *stat;
335 * used when a cpu is offlined or other synchronizations
336 * See mem_cgroup_read_stat().
338 struct mem_cgroup_stat_cpu nocpu_base;
339 spinlock_t pcp_counter_lock;
341 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
342 struct tcp_memcontrol tcp_mem;
344 #if defined(CONFIG_MEMCG_KMEM)
345 /* analogous to slab_common's slab_caches list. per-memcg */
346 struct list_head memcg_slab_caches;
347 /* Not a spinlock, we can take a lot of time walking the list */
348 struct mutex slab_caches_mutex;
349 /* Index in the kmem_cache->memcg_params->memcg_caches array */
354 /* internal only representation about the status of kmem accounting. */
356 KMEM_ACCOUNTED_ACTIVE = 0, /* accounted by this cgroup itself */
357 KMEM_ACCOUNTED_ACTIVATED, /* static key enabled. */
358 KMEM_ACCOUNTED_DEAD, /* dead memcg with pending kmem charges */
361 /* We account when limit is on, but only after call sites are patched */
362 #define KMEM_ACCOUNTED_MASK \
363 ((1 << KMEM_ACCOUNTED_ACTIVE) | (1 << KMEM_ACCOUNTED_ACTIVATED))
365 #ifdef CONFIG_MEMCG_KMEM
366 static inline void memcg_kmem_set_active(struct mem_cgroup *memcg)
368 set_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
371 static bool memcg_kmem_is_active(struct mem_cgroup *memcg)
373 return test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
376 static void memcg_kmem_set_activated(struct mem_cgroup *memcg)
378 set_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
381 static void memcg_kmem_clear_activated(struct mem_cgroup *memcg)
383 clear_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
386 static void memcg_kmem_mark_dead(struct mem_cgroup *memcg)
388 if (test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags))
389 set_bit(KMEM_ACCOUNTED_DEAD, &memcg->kmem_account_flags);
392 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup *memcg)
394 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD,
395 &memcg->kmem_account_flags);
399 /* Stuffs for move charges at task migration. */
401 * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
402 * left-shifted bitmap of these types.
405 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
406 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
410 /* "mc" and its members are protected by cgroup_mutex */
411 static struct move_charge_struct {
412 spinlock_t lock; /* for from, to */
413 struct mem_cgroup *from;
414 struct mem_cgroup *to;
415 unsigned long precharge;
416 unsigned long moved_charge;
417 unsigned long moved_swap;
418 struct task_struct *moving_task; /* a task moving charges */
419 wait_queue_head_t waitq; /* a waitq for other context */
421 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
422 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
425 static bool move_anon(void)
427 return test_bit(MOVE_CHARGE_TYPE_ANON,
428 &mc.to->move_charge_at_immigrate);
431 static bool move_file(void)
433 return test_bit(MOVE_CHARGE_TYPE_FILE,
434 &mc.to->move_charge_at_immigrate);
438 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
439 * limit reclaim to prevent infinite loops, if they ever occur.
441 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
442 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
445 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
446 MEM_CGROUP_CHARGE_TYPE_ANON,
447 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
448 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
452 /* for encoding cft->private value on file */
460 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
461 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
462 #define MEMFILE_ATTR(val) ((val) & 0xffff)
463 /* Used for OOM nofiier */
464 #define OOM_CONTROL (0)
467 * Reclaim flags for mem_cgroup_hierarchical_reclaim
469 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
470 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
471 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
472 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
474 static void mem_cgroup_get(struct mem_cgroup *memcg);
475 static void mem_cgroup_put(struct mem_cgroup *memcg);
478 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
480 return container_of(s, struct mem_cgroup, css);
483 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
485 return (memcg == root_mem_cgroup);
488 /* Writing them here to avoid exposing memcg's inner layout */
489 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
491 void sock_update_memcg(struct sock *sk)
493 if (mem_cgroup_sockets_enabled) {
494 struct mem_cgroup *memcg;
495 struct cg_proto *cg_proto;
497 BUG_ON(!sk->sk_prot->proto_cgroup);
499 /* Socket cloning can throw us here with sk_cgrp already
500 * filled. It won't however, necessarily happen from
501 * process context. So the test for root memcg given
502 * the current task's memcg won't help us in this case.
504 * Respecting the original socket's memcg is a better
505 * decision in this case.
508 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
509 mem_cgroup_get(sk->sk_cgrp->memcg);
514 memcg = mem_cgroup_from_task(current);
515 cg_proto = sk->sk_prot->proto_cgroup(memcg);
516 if (!mem_cgroup_is_root(memcg) && memcg_proto_active(cg_proto)) {
517 mem_cgroup_get(memcg);
518 sk->sk_cgrp = cg_proto;
523 EXPORT_SYMBOL(sock_update_memcg);
525 void sock_release_memcg(struct sock *sk)
527 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
528 struct mem_cgroup *memcg;
529 WARN_ON(!sk->sk_cgrp->memcg);
530 memcg = sk->sk_cgrp->memcg;
531 mem_cgroup_put(memcg);
535 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
537 if (!memcg || mem_cgroup_is_root(memcg))
540 return &memcg->tcp_mem.cg_proto;
542 EXPORT_SYMBOL(tcp_proto_cgroup);
544 static void disarm_sock_keys(struct mem_cgroup *memcg)
546 if (!memcg_proto_activated(&memcg->tcp_mem.cg_proto))
548 static_key_slow_dec(&memcg_socket_limit_enabled);
551 static void disarm_sock_keys(struct mem_cgroup *memcg)
556 #ifdef CONFIG_MEMCG_KMEM
558 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
559 * There are two main reasons for not using the css_id for this:
560 * 1) this works better in sparse environments, where we have a lot of memcgs,
561 * but only a few kmem-limited. Or also, if we have, for instance, 200
562 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
563 * 200 entry array for that.
565 * 2) In order not to violate the cgroup API, we would like to do all memory
566 * allocation in ->create(). At that point, we haven't yet allocated the
567 * css_id. Having a separate index prevents us from messing with the cgroup
570 * The current size of the caches array is stored in
571 * memcg_limited_groups_array_size. It will double each time we have to
574 static DEFINE_IDA(kmem_limited_groups);
575 static int memcg_limited_groups_array_size;
577 * MIN_SIZE is different than 1, because we would like to avoid going through
578 * the alloc/free process all the time. In a small machine, 4 kmem-limited
579 * cgroups is a reasonable guess. In the future, it could be a parameter or
580 * tunable, but that is strictly not necessary.
582 * MAX_SIZE should be as large as the number of css_ids. Ideally, we could get
583 * this constant directly from cgroup, but it is understandable that this is
584 * better kept as an internal representation in cgroup.c. In any case, the
585 * css_id space is not getting any smaller, and we don't have to necessarily
586 * increase ours as well if it increases.
588 #define MEMCG_CACHES_MIN_SIZE 4
589 #define MEMCG_CACHES_MAX_SIZE 65535
592 * A lot of the calls to the cache allocation functions are expected to be
593 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
594 * conditional to this static branch, we'll have to allow modules that does
595 * kmem_cache_alloc and the such to see this symbol as well
597 struct static_key memcg_kmem_enabled_key;
598 EXPORT_SYMBOL(memcg_kmem_enabled_key);
600 static void disarm_kmem_keys(struct mem_cgroup *memcg)
602 if (memcg_kmem_is_active(memcg)) {
603 static_key_slow_dec(&memcg_kmem_enabled_key);
604 ida_simple_remove(&kmem_limited_groups, memcg->kmemcg_id);
607 * This check can't live in kmem destruction function,
608 * since the charges will outlive the cgroup
610 WARN_ON(res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0);
613 static void disarm_kmem_keys(struct mem_cgroup *memcg)
616 #endif /* CONFIG_MEMCG_KMEM */
618 static void disarm_static_keys(struct mem_cgroup *memcg)
620 disarm_sock_keys(memcg);
621 disarm_kmem_keys(memcg);
624 static void drain_all_stock_async(struct mem_cgroup *memcg);
626 static struct mem_cgroup_per_zone *
627 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
629 return &memcg->info.nodeinfo[nid]->zoneinfo[zid];
632 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
637 static struct mem_cgroup_per_zone *
638 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
640 int nid = page_to_nid(page);
641 int zid = page_zonenum(page);
643 return mem_cgroup_zoneinfo(memcg, nid, zid);
646 static struct mem_cgroup_tree_per_zone *
647 soft_limit_tree_node_zone(int nid, int zid)
649 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
652 static struct mem_cgroup_tree_per_zone *
653 soft_limit_tree_from_page(struct page *page)
655 int nid = page_to_nid(page);
656 int zid = page_zonenum(page);
658 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
662 __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
663 struct mem_cgroup_per_zone *mz,
664 struct mem_cgroup_tree_per_zone *mctz,
665 unsigned long long new_usage_in_excess)
667 struct rb_node **p = &mctz->rb_root.rb_node;
668 struct rb_node *parent = NULL;
669 struct mem_cgroup_per_zone *mz_node;
674 mz->usage_in_excess = new_usage_in_excess;
675 if (!mz->usage_in_excess)
679 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
681 if (mz->usage_in_excess < mz_node->usage_in_excess)
684 * We can't avoid mem cgroups that are over their soft
685 * limit by the same amount
687 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
690 rb_link_node(&mz->tree_node, parent, p);
691 rb_insert_color(&mz->tree_node, &mctz->rb_root);
696 __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
697 struct mem_cgroup_per_zone *mz,
698 struct mem_cgroup_tree_per_zone *mctz)
702 rb_erase(&mz->tree_node, &mctz->rb_root);
707 mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
708 struct mem_cgroup_per_zone *mz,
709 struct mem_cgroup_tree_per_zone *mctz)
711 spin_lock(&mctz->lock);
712 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
713 spin_unlock(&mctz->lock);
717 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
719 unsigned long long excess;
720 struct mem_cgroup_per_zone *mz;
721 struct mem_cgroup_tree_per_zone *mctz;
722 int nid = page_to_nid(page);
723 int zid = page_zonenum(page);
724 mctz = soft_limit_tree_from_page(page);
727 * Necessary to update all ancestors when hierarchy is used.
728 * because their event counter is not touched.
730 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
731 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
732 excess = res_counter_soft_limit_excess(&memcg->res);
734 * We have to update the tree if mz is on RB-tree or
735 * mem is over its softlimit.
737 if (excess || mz->on_tree) {
738 spin_lock(&mctz->lock);
739 /* if on-tree, remove it */
741 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
743 * Insert again. mz->usage_in_excess will be updated.
744 * If excess is 0, no tree ops.
746 __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
747 spin_unlock(&mctz->lock);
752 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
755 struct mem_cgroup_per_zone *mz;
756 struct mem_cgroup_tree_per_zone *mctz;
758 for_each_node(node) {
759 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
760 mz = mem_cgroup_zoneinfo(memcg, node, zone);
761 mctz = soft_limit_tree_node_zone(node, zone);
762 mem_cgroup_remove_exceeded(memcg, mz, mctz);
767 static struct mem_cgroup_per_zone *
768 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
770 struct rb_node *rightmost = NULL;
771 struct mem_cgroup_per_zone *mz;
775 rightmost = rb_last(&mctz->rb_root);
777 goto done; /* Nothing to reclaim from */
779 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
781 * Remove the node now but someone else can add it back,
782 * we will to add it back at the end of reclaim to its correct
783 * position in the tree.
785 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
786 if (!res_counter_soft_limit_excess(&mz->memcg->res) ||
787 !css_tryget(&mz->memcg->css))
793 static struct mem_cgroup_per_zone *
794 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
796 struct mem_cgroup_per_zone *mz;
798 spin_lock(&mctz->lock);
799 mz = __mem_cgroup_largest_soft_limit_node(mctz);
800 spin_unlock(&mctz->lock);
805 * Implementation Note: reading percpu statistics for memcg.
807 * Both of vmstat[] and percpu_counter has threshold and do periodic
808 * synchronization to implement "quick" read. There are trade-off between
809 * reading cost and precision of value. Then, we may have a chance to implement
810 * a periodic synchronizion of counter in memcg's counter.
812 * But this _read() function is used for user interface now. The user accounts
813 * memory usage by memory cgroup and he _always_ requires exact value because
814 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
815 * have to visit all online cpus and make sum. So, for now, unnecessary
816 * synchronization is not implemented. (just implemented for cpu hotplug)
818 * If there are kernel internal actions which can make use of some not-exact
819 * value, and reading all cpu value can be performance bottleneck in some
820 * common workload, threashold and synchonization as vmstat[] should be
823 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
824 enum mem_cgroup_stat_index idx)
830 for_each_online_cpu(cpu)
831 val += per_cpu(memcg->stat->count[idx], cpu);
832 #ifdef CONFIG_HOTPLUG_CPU
833 spin_lock(&memcg->pcp_counter_lock);
834 val += memcg->nocpu_base.count[idx];
835 spin_unlock(&memcg->pcp_counter_lock);
841 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
844 int val = (charge) ? 1 : -1;
845 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
848 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
849 enum mem_cgroup_events_index idx)
851 unsigned long val = 0;
854 for_each_online_cpu(cpu)
855 val += per_cpu(memcg->stat->events[idx], cpu);
856 #ifdef CONFIG_HOTPLUG_CPU
857 spin_lock(&memcg->pcp_counter_lock);
858 val += memcg->nocpu_base.events[idx];
859 spin_unlock(&memcg->pcp_counter_lock);
864 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
865 bool anon, int nr_pages)
870 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
871 * counted as CACHE even if it's on ANON LRU.
874 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
877 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
880 /* pagein of a big page is an event. So, ignore page size */
882 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
884 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
885 nr_pages = -nr_pages; /* for event */
888 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
894 mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
896 struct mem_cgroup_per_zone *mz;
898 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
899 return mz->lru_size[lru];
903 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
904 unsigned int lru_mask)
906 struct mem_cgroup_per_zone *mz;
908 unsigned long ret = 0;
910 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
913 if (BIT(lru) & lru_mask)
914 ret += mz->lru_size[lru];
920 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
921 int nid, unsigned int lru_mask)
926 for (zid = 0; zid < MAX_NR_ZONES; zid++)
927 total += mem_cgroup_zone_nr_lru_pages(memcg,
933 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
934 unsigned int lru_mask)
939 for_each_node_state(nid, N_MEMORY)
940 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
944 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
945 enum mem_cgroup_events_target target)
947 unsigned long val, next;
949 val = __this_cpu_read(memcg->stat->nr_page_events);
950 next = __this_cpu_read(memcg->stat->targets[target]);
951 /* from time_after() in jiffies.h */
952 if ((long)next - (long)val < 0) {
954 case MEM_CGROUP_TARGET_THRESH:
955 next = val + THRESHOLDS_EVENTS_TARGET;
957 case MEM_CGROUP_TARGET_SOFTLIMIT:
958 next = val + SOFTLIMIT_EVENTS_TARGET;
960 case MEM_CGROUP_TARGET_NUMAINFO:
961 next = val + NUMAINFO_EVENTS_TARGET;
966 __this_cpu_write(memcg->stat->targets[target], next);
973 * Check events in order.
976 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
979 /* threshold event is triggered in finer grain than soft limit */
980 if (unlikely(mem_cgroup_event_ratelimit(memcg,
981 MEM_CGROUP_TARGET_THRESH))) {
983 bool do_numainfo __maybe_unused;
985 do_softlimit = mem_cgroup_event_ratelimit(memcg,
986 MEM_CGROUP_TARGET_SOFTLIMIT);
988 do_numainfo = mem_cgroup_event_ratelimit(memcg,
989 MEM_CGROUP_TARGET_NUMAINFO);
993 mem_cgroup_threshold(memcg);
994 if (unlikely(do_softlimit))
995 mem_cgroup_update_tree(memcg, page);
997 if (unlikely(do_numainfo))
998 atomic_inc(&memcg->numainfo_events);
1004 struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
1006 return mem_cgroup_from_css(
1007 cgroup_subsys_state(cont, mem_cgroup_subsys_id));
1010 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1013 * mm_update_next_owner() may clear mm->owner to NULL
1014 * if it races with swapoff, page migration, etc.
1015 * So this can be called with p == NULL.
1020 return mem_cgroup_from_css(task_subsys_state(p, mem_cgroup_subsys_id));
1023 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
1025 struct mem_cgroup *memcg = NULL;
1030 * Because we have no locks, mm->owner's may be being moved to other
1031 * cgroup. We use css_tryget() here even if this looks
1032 * pessimistic (rather than adding locks here).
1036 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1037 if (unlikely(!memcg))
1039 } while (!css_tryget(&memcg->css));
1045 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1046 * @root: hierarchy root
1047 * @prev: previously returned memcg, NULL on first invocation
1048 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1050 * Returns references to children of the hierarchy below @root, or
1051 * @root itself, or %NULL after a full round-trip.
1053 * Caller must pass the return value in @prev on subsequent
1054 * invocations for reference counting, or use mem_cgroup_iter_break()
1055 * to cancel a hierarchy walk before the round-trip is complete.
1057 * Reclaimers can specify a zone and a priority level in @reclaim to
1058 * divide up the memcgs in the hierarchy among all concurrent
1059 * reclaimers operating on the same zone and priority.
1061 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1062 struct mem_cgroup *prev,
1063 struct mem_cgroup_reclaim_cookie *reclaim)
1065 struct mem_cgroup *memcg = NULL;
1068 if (mem_cgroup_disabled())
1072 root = root_mem_cgroup;
1074 if (prev && !reclaim)
1075 id = css_id(&prev->css);
1077 if (prev && prev != root)
1078 css_put(&prev->css);
1080 if (!root->use_hierarchy && root != root_mem_cgroup) {
1087 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1088 struct cgroup_subsys_state *css;
1091 int nid = zone_to_nid(reclaim->zone);
1092 int zid = zone_idx(reclaim->zone);
1093 struct mem_cgroup_per_zone *mz;
1095 mz = mem_cgroup_zoneinfo(root, nid, zid);
1096 iter = &mz->reclaim_iter[reclaim->priority];
1097 if (prev && reclaim->generation != iter->generation)
1099 id = iter->position;
1103 css = css_get_next(&mem_cgroup_subsys, id + 1, &root->css, &id);
1105 if (css == &root->css || css_tryget(css))
1106 memcg = mem_cgroup_from_css(css);
1112 iter->position = id;
1115 else if (!prev && memcg)
1116 reclaim->generation = iter->generation;
1126 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1127 * @root: hierarchy root
1128 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1130 void mem_cgroup_iter_break(struct mem_cgroup *root,
1131 struct mem_cgroup *prev)
1134 root = root_mem_cgroup;
1135 if (prev && prev != root)
1136 css_put(&prev->css);
1140 * Iteration constructs for visiting all cgroups (under a tree). If
1141 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1142 * be used for reference counting.
1144 #define for_each_mem_cgroup_tree(iter, root) \
1145 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1147 iter = mem_cgroup_iter(root, iter, NULL))
1149 #define for_each_mem_cgroup(iter) \
1150 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1152 iter = mem_cgroup_iter(NULL, iter, NULL))
1154 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1156 struct mem_cgroup *memcg;
1159 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1160 if (unlikely(!memcg))
1165 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1168 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1176 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1179 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1180 * @zone: zone of the wanted lruvec
1181 * @memcg: memcg of the wanted lruvec
1183 * Returns the lru list vector holding pages for the given @zone and
1184 * @mem. This can be the global zone lruvec, if the memory controller
1187 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1188 struct mem_cgroup *memcg)
1190 struct mem_cgroup_per_zone *mz;
1191 struct lruvec *lruvec;
1193 if (mem_cgroup_disabled()) {
1194 lruvec = &zone->lruvec;
1198 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1199 lruvec = &mz->lruvec;
1202 * Since a node can be onlined after the mem_cgroup was created,
1203 * we have to be prepared to initialize lruvec->zone here;
1204 * and if offlined then reonlined, we need to reinitialize it.
1206 if (unlikely(lruvec->zone != zone))
1207 lruvec->zone = zone;
1212 * Following LRU functions are allowed to be used without PCG_LOCK.
1213 * Operations are called by routine of global LRU independently from memcg.
1214 * What we have to take care of here is validness of pc->mem_cgroup.
1216 * Changes to pc->mem_cgroup happens when
1219 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1220 * It is added to LRU before charge.
1221 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1222 * When moving account, the page is not on LRU. It's isolated.
1226 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1228 * @zone: zone of the page
1230 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1232 struct mem_cgroup_per_zone *mz;
1233 struct mem_cgroup *memcg;
1234 struct page_cgroup *pc;
1235 struct lruvec *lruvec;
1237 if (mem_cgroup_disabled()) {
1238 lruvec = &zone->lruvec;
1242 pc = lookup_page_cgroup(page);
1243 memcg = pc->mem_cgroup;
1246 * Surreptitiously switch any uncharged offlist page to root:
1247 * an uncharged page off lru does nothing to secure
1248 * its former mem_cgroup from sudden removal.
1250 * Our caller holds lru_lock, and PageCgroupUsed is updated
1251 * under page_cgroup lock: between them, they make all uses
1252 * of pc->mem_cgroup safe.
1254 if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1255 pc->mem_cgroup = memcg = root_mem_cgroup;
1257 mz = page_cgroup_zoneinfo(memcg, page);
1258 lruvec = &mz->lruvec;
1261 * Since a node can be onlined after the mem_cgroup was created,
1262 * we have to be prepared to initialize lruvec->zone here;
1263 * and if offlined then reonlined, we need to reinitialize it.
1265 if (unlikely(lruvec->zone != zone))
1266 lruvec->zone = zone;
1271 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1272 * @lruvec: mem_cgroup per zone lru vector
1273 * @lru: index of lru list the page is sitting on
1274 * @nr_pages: positive when adding or negative when removing
1276 * This function must be called when a page is added to or removed from an
1279 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1282 struct mem_cgroup_per_zone *mz;
1283 unsigned long *lru_size;
1285 if (mem_cgroup_disabled())
1288 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1289 lru_size = mz->lru_size + lru;
1290 *lru_size += nr_pages;
1291 VM_BUG_ON((long)(*lru_size) < 0);
1295 * Checks whether given mem is same or in the root_mem_cgroup's
1298 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1299 struct mem_cgroup *memcg)
1301 if (root_memcg == memcg)
1303 if (!root_memcg->use_hierarchy || !memcg)
1305 return css_is_ancestor(&memcg->css, &root_memcg->css);
1308 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1309 struct mem_cgroup *memcg)
1314 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1319 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *memcg)
1322 struct mem_cgroup *curr = NULL;
1323 struct task_struct *p;
1325 p = find_lock_task_mm(task);
1327 curr = try_get_mem_cgroup_from_mm(p->mm);
1331 * All threads may have already detached their mm's, but the oom
1332 * killer still needs to detect if they have already been oom
1333 * killed to prevent needlessly killing additional tasks.
1336 curr = mem_cgroup_from_task(task);
1338 css_get(&curr->css);
1344 * We should check use_hierarchy of "memcg" not "curr". Because checking
1345 * use_hierarchy of "curr" here make this function true if hierarchy is
1346 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1347 * hierarchy(even if use_hierarchy is disabled in "memcg").
1349 ret = mem_cgroup_same_or_subtree(memcg, curr);
1350 css_put(&curr->css);
1354 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1356 unsigned long inactive_ratio;
1357 unsigned long inactive;
1358 unsigned long active;
1361 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1362 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1364 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1366 inactive_ratio = int_sqrt(10 * gb);
1370 return inactive * inactive_ratio < active;
1373 int mem_cgroup_inactive_file_is_low(struct lruvec *lruvec)
1375 unsigned long active;
1376 unsigned long inactive;
1378 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_FILE);
1379 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_FILE);
1381 return (active > inactive);
1384 #define mem_cgroup_from_res_counter(counter, member) \
1385 container_of(counter, struct mem_cgroup, member)
1388 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1389 * @memcg: the memory cgroup
1391 * Returns the maximum amount of memory @mem can be charged with, in
1394 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1396 unsigned long long margin;
1398 margin = res_counter_margin(&memcg->res);
1399 if (do_swap_account)
1400 margin = min(margin, res_counter_margin(&memcg->memsw));
1401 return margin >> PAGE_SHIFT;
1404 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1406 struct cgroup *cgrp = memcg->css.cgroup;
1409 if (cgrp->parent == NULL)
1410 return vm_swappiness;
1412 return memcg->swappiness;
1416 * memcg->moving_account is used for checking possibility that some thread is
1417 * calling move_account(). When a thread on CPU-A starts moving pages under
1418 * a memcg, other threads should check memcg->moving_account under
1419 * rcu_read_lock(), like this:
1423 * memcg->moving_account+1 if (memcg->mocing_account)
1425 * synchronize_rcu() update something.
1430 /* for quick checking without looking up memcg */
1431 atomic_t memcg_moving __read_mostly;
1433 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1435 atomic_inc(&memcg_moving);
1436 atomic_inc(&memcg->moving_account);
1440 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1443 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1444 * We check NULL in callee rather than caller.
1447 atomic_dec(&memcg_moving);
1448 atomic_dec(&memcg->moving_account);
1453 * 2 routines for checking "mem" is under move_account() or not.
1455 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1456 * is used for avoiding races in accounting. If true,
1457 * pc->mem_cgroup may be overwritten.
1459 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1460 * under hierarchy of moving cgroups. This is for
1461 * waiting at hith-memory prressure caused by "move".
1464 static bool mem_cgroup_stolen(struct mem_cgroup *memcg)
1466 VM_BUG_ON(!rcu_read_lock_held());
1467 return atomic_read(&memcg->moving_account) > 0;
1470 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1472 struct mem_cgroup *from;
1473 struct mem_cgroup *to;
1476 * Unlike task_move routines, we access mc.to, mc.from not under
1477 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1479 spin_lock(&mc.lock);
1485 ret = mem_cgroup_same_or_subtree(memcg, from)
1486 || mem_cgroup_same_or_subtree(memcg, to);
1488 spin_unlock(&mc.lock);
1492 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1494 if (mc.moving_task && current != mc.moving_task) {
1495 if (mem_cgroup_under_move(memcg)) {
1497 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1498 /* moving charge context might have finished. */
1501 finish_wait(&mc.waitq, &wait);
1509 * Take this lock when
1510 * - a code tries to modify page's memcg while it's USED.
1511 * - a code tries to modify page state accounting in a memcg.
1512 * see mem_cgroup_stolen(), too.
1514 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1515 unsigned long *flags)
1517 spin_lock_irqsave(&memcg->move_lock, *flags);
1520 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1521 unsigned long *flags)
1523 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1527 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1528 * @memcg: The memory cgroup that went over limit
1529 * @p: Task that is going to be killed
1531 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1534 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1536 struct cgroup *task_cgrp;
1537 struct cgroup *mem_cgrp;
1539 * Need a buffer in BSS, can't rely on allocations. The code relies
1540 * on the assumption that OOM is serialized for memory controller.
1541 * If this assumption is broken, revisit this code.
1543 static char memcg_name[PATH_MAX];
1551 mem_cgrp = memcg->css.cgroup;
1552 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1554 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1557 * Unfortunately, we are unable to convert to a useful name
1558 * But we'll still print out the usage information
1565 printk(KERN_INFO "Task in %s killed", memcg_name);
1568 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1576 * Continues from above, so we don't need an KERN_ level
1578 printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1581 printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1582 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1583 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1584 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1585 printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1587 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1588 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1589 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1590 printk(KERN_INFO "kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1591 res_counter_read_u64(&memcg->kmem, RES_USAGE) >> 10,
1592 res_counter_read_u64(&memcg->kmem, RES_LIMIT) >> 10,
1593 res_counter_read_u64(&memcg->kmem, RES_FAILCNT));
1597 * This function returns the number of memcg under hierarchy tree. Returns
1598 * 1(self count) if no children.
1600 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1603 struct mem_cgroup *iter;
1605 for_each_mem_cgroup_tree(iter, memcg)
1611 * Return the memory (and swap, if configured) limit for a memcg.
1613 static u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1617 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1620 * Do not consider swap space if we cannot swap due to swappiness
1622 if (mem_cgroup_swappiness(memcg)) {
1625 limit += total_swap_pages << PAGE_SHIFT;
1626 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1629 * If memsw is finite and limits the amount of swap space
1630 * available to this memcg, return that limit.
1632 limit = min(limit, memsw);
1638 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1641 struct mem_cgroup *iter;
1642 unsigned long chosen_points = 0;
1643 unsigned long totalpages;
1644 unsigned int points = 0;
1645 struct task_struct *chosen = NULL;
1648 * If current has a pending SIGKILL, then automatically select it. The
1649 * goal is to allow it to allocate so that it may quickly exit and free
1652 if (fatal_signal_pending(current)) {
1653 set_thread_flag(TIF_MEMDIE);
1657 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1658 totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
1659 for_each_mem_cgroup_tree(iter, memcg) {
1660 struct cgroup *cgroup = iter->css.cgroup;
1661 struct cgroup_iter it;
1662 struct task_struct *task;
1664 cgroup_iter_start(cgroup, &it);
1665 while ((task = cgroup_iter_next(cgroup, &it))) {
1666 switch (oom_scan_process_thread(task, totalpages, NULL,
1668 case OOM_SCAN_SELECT:
1670 put_task_struct(chosen);
1672 chosen_points = ULONG_MAX;
1673 get_task_struct(chosen);
1675 case OOM_SCAN_CONTINUE:
1677 case OOM_SCAN_ABORT:
1678 cgroup_iter_end(cgroup, &it);
1679 mem_cgroup_iter_break(memcg, iter);
1681 put_task_struct(chosen);
1686 points = oom_badness(task, memcg, NULL, totalpages);
1687 if (points > chosen_points) {
1689 put_task_struct(chosen);
1691 chosen_points = points;
1692 get_task_struct(chosen);
1695 cgroup_iter_end(cgroup, &it);
1700 points = chosen_points * 1000 / totalpages;
1701 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1702 NULL, "Memory cgroup out of memory");
1705 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1707 unsigned long flags)
1709 unsigned long total = 0;
1710 bool noswap = false;
1713 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1715 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1718 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1720 drain_all_stock_async(memcg);
1721 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1723 * Allow limit shrinkers, which are triggered directly
1724 * by userspace, to catch signals and stop reclaim
1725 * after minimal progress, regardless of the margin.
1727 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1729 if (mem_cgroup_margin(memcg))
1732 * If nothing was reclaimed after two attempts, there
1733 * may be no reclaimable pages in this hierarchy.
1742 * test_mem_cgroup_node_reclaimable
1743 * @memcg: the target memcg
1744 * @nid: the node ID to be checked.
1745 * @noswap : specify true here if the user wants flle only information.
1747 * This function returns whether the specified memcg contains any
1748 * reclaimable pages on a node. Returns true if there are any reclaimable
1749 * pages in the node.
1751 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1752 int nid, bool noswap)
1754 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1756 if (noswap || !total_swap_pages)
1758 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1763 #if MAX_NUMNODES > 1
1766 * Always updating the nodemask is not very good - even if we have an empty
1767 * list or the wrong list here, we can start from some node and traverse all
1768 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1771 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1775 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1776 * pagein/pageout changes since the last update.
1778 if (!atomic_read(&memcg->numainfo_events))
1780 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1783 /* make a nodemask where this memcg uses memory from */
1784 memcg->scan_nodes = node_states[N_MEMORY];
1786 for_each_node_mask(nid, node_states[N_MEMORY]) {
1788 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1789 node_clear(nid, memcg->scan_nodes);
1792 atomic_set(&memcg->numainfo_events, 0);
1793 atomic_set(&memcg->numainfo_updating, 0);
1797 * Selecting a node where we start reclaim from. Because what we need is just
1798 * reducing usage counter, start from anywhere is O,K. Considering
1799 * memory reclaim from current node, there are pros. and cons.
1801 * Freeing memory from current node means freeing memory from a node which
1802 * we'll use or we've used. So, it may make LRU bad. And if several threads
1803 * hit limits, it will see a contention on a node. But freeing from remote
1804 * node means more costs for memory reclaim because of memory latency.
1806 * Now, we use round-robin. Better algorithm is welcomed.
1808 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1812 mem_cgroup_may_update_nodemask(memcg);
1813 node = memcg->last_scanned_node;
1815 node = next_node(node, memcg->scan_nodes);
1816 if (node == MAX_NUMNODES)
1817 node = first_node(memcg->scan_nodes);
1819 * We call this when we hit limit, not when pages are added to LRU.
1820 * No LRU may hold pages because all pages are UNEVICTABLE or
1821 * memcg is too small and all pages are not on LRU. In that case,
1822 * we use curret node.
1824 if (unlikely(node == MAX_NUMNODES))
1825 node = numa_node_id();
1827 memcg->last_scanned_node = node;
1832 * Check all nodes whether it contains reclaimable pages or not.
1833 * For quick scan, we make use of scan_nodes. This will allow us to skip
1834 * unused nodes. But scan_nodes is lazily updated and may not cotain
1835 * enough new information. We need to do double check.
1837 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1842 * quick check...making use of scan_node.
1843 * We can skip unused nodes.
1845 if (!nodes_empty(memcg->scan_nodes)) {
1846 for (nid = first_node(memcg->scan_nodes);
1848 nid = next_node(nid, memcg->scan_nodes)) {
1850 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1855 * Check rest of nodes.
1857 for_each_node_state(nid, N_MEMORY) {
1858 if (node_isset(nid, memcg->scan_nodes))
1860 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1867 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1872 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1874 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
1878 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1881 unsigned long *total_scanned)
1883 struct mem_cgroup *victim = NULL;
1886 unsigned long excess;
1887 unsigned long nr_scanned;
1888 struct mem_cgroup_reclaim_cookie reclaim = {
1893 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
1896 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1901 * If we have not been able to reclaim
1902 * anything, it might because there are
1903 * no reclaimable pages under this hierarchy
1908 * We want to do more targeted reclaim.
1909 * excess >> 2 is not to excessive so as to
1910 * reclaim too much, nor too less that we keep
1911 * coming back to reclaim from this cgroup
1913 if (total >= (excess >> 2) ||
1914 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1919 if (!mem_cgroup_reclaimable(victim, false))
1921 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1923 *total_scanned += nr_scanned;
1924 if (!res_counter_soft_limit_excess(&root_memcg->res))
1927 mem_cgroup_iter_break(root_memcg, victim);
1932 * Check OOM-Killer is already running under our hierarchy.
1933 * If someone is running, return false.
1934 * Has to be called with memcg_oom_lock
1936 static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg)
1938 struct mem_cgroup *iter, *failed = NULL;
1940 for_each_mem_cgroup_tree(iter, memcg) {
1941 if (iter->oom_lock) {
1943 * this subtree of our hierarchy is already locked
1944 * so we cannot give a lock.
1947 mem_cgroup_iter_break(memcg, iter);
1950 iter->oom_lock = true;
1957 * OK, we failed to lock the whole subtree so we have to clean up
1958 * what we set up to the failing subtree
1960 for_each_mem_cgroup_tree(iter, memcg) {
1961 if (iter == failed) {
1962 mem_cgroup_iter_break(memcg, iter);
1965 iter->oom_lock = false;
1971 * Has to be called with memcg_oom_lock
1973 static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1975 struct mem_cgroup *iter;
1977 for_each_mem_cgroup_tree(iter, memcg)
1978 iter->oom_lock = false;
1982 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1984 struct mem_cgroup *iter;
1986 for_each_mem_cgroup_tree(iter, memcg)
1987 atomic_inc(&iter->under_oom);
1990 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1992 struct mem_cgroup *iter;
1995 * When a new child is created while the hierarchy is under oom,
1996 * mem_cgroup_oom_lock() may not be called. We have to use
1997 * atomic_add_unless() here.
1999 for_each_mem_cgroup_tree(iter, memcg)
2000 atomic_add_unless(&iter->under_oom, -1, 0);
2003 static DEFINE_SPINLOCK(memcg_oom_lock);
2004 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
2006 struct oom_wait_info {
2007 struct mem_cgroup *memcg;
2011 static int memcg_oom_wake_function(wait_queue_t *wait,
2012 unsigned mode, int sync, void *arg)
2014 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
2015 struct mem_cgroup *oom_wait_memcg;
2016 struct oom_wait_info *oom_wait_info;
2018 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
2019 oom_wait_memcg = oom_wait_info->memcg;
2022 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2023 * Then we can use css_is_ancestor without taking care of RCU.
2025 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
2026 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
2028 return autoremove_wake_function(wait, mode, sync, arg);
2031 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
2033 /* for filtering, pass "memcg" as argument. */
2034 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
2037 static void memcg_oom_recover(struct mem_cgroup *memcg)
2039 if (memcg && atomic_read(&memcg->under_oom))
2040 memcg_wakeup_oom(memcg);
2044 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
2046 static bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask,
2049 struct oom_wait_info owait;
2050 bool locked, need_to_kill;
2052 owait.memcg = memcg;
2053 owait.wait.flags = 0;
2054 owait.wait.func = memcg_oom_wake_function;
2055 owait.wait.private = current;
2056 INIT_LIST_HEAD(&owait.wait.task_list);
2057 need_to_kill = true;
2058 mem_cgroup_mark_under_oom(memcg);
2060 /* At first, try to OOM lock hierarchy under memcg.*/
2061 spin_lock(&memcg_oom_lock);
2062 locked = mem_cgroup_oom_lock(memcg);
2064 * Even if signal_pending(), we can't quit charge() loop without
2065 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
2066 * under OOM is always welcomed, use TASK_KILLABLE here.
2068 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2069 if (!locked || memcg->oom_kill_disable)
2070 need_to_kill = false;
2072 mem_cgroup_oom_notify(memcg);
2073 spin_unlock(&memcg_oom_lock);
2076 finish_wait(&memcg_oom_waitq, &owait.wait);
2077 mem_cgroup_out_of_memory(memcg, mask, order);
2080 finish_wait(&memcg_oom_waitq, &owait.wait);
2082 spin_lock(&memcg_oom_lock);
2084 mem_cgroup_oom_unlock(memcg);
2085 memcg_wakeup_oom(memcg);
2086 spin_unlock(&memcg_oom_lock);
2088 mem_cgroup_unmark_under_oom(memcg);
2090 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
2092 /* Give chance to dying process */
2093 schedule_timeout_uninterruptible(1);
2098 * Currently used to update mapped file statistics, but the routine can be
2099 * generalized to update other statistics as well.
2101 * Notes: Race condition
2103 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2104 * it tends to be costly. But considering some conditions, we doesn't need
2105 * to do so _always_.
2107 * Considering "charge", lock_page_cgroup() is not required because all
2108 * file-stat operations happen after a page is attached to radix-tree. There
2109 * are no race with "charge".
2111 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2112 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2113 * if there are race with "uncharge". Statistics itself is properly handled
2116 * Considering "move", this is an only case we see a race. To make the race
2117 * small, we check mm->moving_account and detect there are possibility of race
2118 * If there is, we take a lock.
2121 void __mem_cgroup_begin_update_page_stat(struct page *page,
2122 bool *locked, unsigned long *flags)
2124 struct mem_cgroup *memcg;
2125 struct page_cgroup *pc;
2127 pc = lookup_page_cgroup(page);
2129 memcg = pc->mem_cgroup;
2130 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2133 * If this memory cgroup is not under account moving, we don't
2134 * need to take move_lock_mem_cgroup(). Because we already hold
2135 * rcu_read_lock(), any calls to move_account will be delayed until
2136 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2138 if (!mem_cgroup_stolen(memcg))
2141 move_lock_mem_cgroup(memcg, flags);
2142 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2143 move_unlock_mem_cgroup(memcg, flags);
2149 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
2151 struct page_cgroup *pc = lookup_page_cgroup(page);
2154 * It's guaranteed that pc->mem_cgroup never changes while
2155 * lock is held because a routine modifies pc->mem_cgroup
2156 * should take move_lock_mem_cgroup().
2158 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
2161 void mem_cgroup_update_page_stat(struct page *page,
2162 enum mem_cgroup_page_stat_item idx, int val)
2164 struct mem_cgroup *memcg;
2165 struct page_cgroup *pc = lookup_page_cgroup(page);
2166 unsigned long uninitialized_var(flags);
2168 if (mem_cgroup_disabled())
2171 memcg = pc->mem_cgroup;
2172 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2176 case MEMCG_NR_FILE_MAPPED:
2177 idx = MEM_CGROUP_STAT_FILE_MAPPED;
2183 this_cpu_add(memcg->stat->count[idx], val);
2187 * size of first charge trial. "32" comes from vmscan.c's magic value.
2188 * TODO: maybe necessary to use big numbers in big irons.
2190 #define CHARGE_BATCH 32U
2191 struct memcg_stock_pcp {
2192 struct mem_cgroup *cached; /* this never be root cgroup */
2193 unsigned int nr_pages;
2194 struct work_struct work;
2195 unsigned long flags;
2196 #define FLUSHING_CACHED_CHARGE 0
2198 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2199 static DEFINE_MUTEX(percpu_charge_mutex);
2202 * consume_stock: Try to consume stocked charge on this cpu.
2203 * @memcg: memcg to consume from.
2204 * @nr_pages: how many pages to charge.
2206 * The charges will only happen if @memcg matches the current cpu's memcg
2207 * stock, and at least @nr_pages are available in that stock. Failure to
2208 * service an allocation will refill the stock.
2210 * returns true if successful, false otherwise.
2212 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2214 struct memcg_stock_pcp *stock;
2217 if (nr_pages > CHARGE_BATCH)
2220 stock = &get_cpu_var(memcg_stock);
2221 if (memcg == stock->cached && stock->nr_pages >= nr_pages)
2222 stock->nr_pages -= nr_pages;
2223 else /* need to call res_counter_charge */
2225 put_cpu_var(memcg_stock);
2230 * Returns stocks cached in percpu to res_counter and reset cached information.
2232 static void drain_stock(struct memcg_stock_pcp *stock)
2234 struct mem_cgroup *old = stock->cached;
2236 if (stock->nr_pages) {
2237 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2239 res_counter_uncharge(&old->res, bytes);
2240 if (do_swap_account)
2241 res_counter_uncharge(&old->memsw, bytes);
2242 stock->nr_pages = 0;
2244 stock->cached = NULL;
2248 * This must be called under preempt disabled or must be called by
2249 * a thread which is pinned to local cpu.
2251 static void drain_local_stock(struct work_struct *dummy)
2253 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2255 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2259 * Cache charges(val) which is from res_counter, to local per_cpu area.
2260 * This will be consumed by consume_stock() function, later.
2262 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2264 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2266 if (stock->cached != memcg) { /* reset if necessary */
2268 stock->cached = memcg;
2270 stock->nr_pages += nr_pages;
2271 put_cpu_var(memcg_stock);
2275 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2276 * of the hierarchy under it. sync flag says whether we should block
2277 * until the work is done.
2279 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2283 /* Notify other cpus that system-wide "drain" is running */
2286 for_each_online_cpu(cpu) {
2287 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2288 struct mem_cgroup *memcg;
2290 memcg = stock->cached;
2291 if (!memcg || !stock->nr_pages)
2293 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2295 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2297 drain_local_stock(&stock->work);
2299 schedule_work_on(cpu, &stock->work);
2307 for_each_online_cpu(cpu) {
2308 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2309 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2310 flush_work(&stock->work);
2317 * Tries to drain stocked charges in other cpus. This function is asynchronous
2318 * and just put a work per cpu for draining localy on each cpu. Caller can
2319 * expects some charges will be back to res_counter later but cannot wait for
2322 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2325 * If someone calls draining, avoid adding more kworker runs.
2327 if (!mutex_trylock(&percpu_charge_mutex))
2329 drain_all_stock(root_memcg, false);
2330 mutex_unlock(&percpu_charge_mutex);
2333 /* This is a synchronous drain interface. */
2334 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2336 /* called when force_empty is called */
2337 mutex_lock(&percpu_charge_mutex);
2338 drain_all_stock(root_memcg, true);
2339 mutex_unlock(&percpu_charge_mutex);
2343 * This function drains percpu counter value from DEAD cpu and
2344 * move it to local cpu. Note that this function can be preempted.
2346 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2350 spin_lock(&memcg->pcp_counter_lock);
2351 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2352 long x = per_cpu(memcg->stat->count[i], cpu);
2354 per_cpu(memcg->stat->count[i], cpu) = 0;
2355 memcg->nocpu_base.count[i] += x;
2357 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2358 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2360 per_cpu(memcg->stat->events[i], cpu) = 0;
2361 memcg->nocpu_base.events[i] += x;
2363 spin_unlock(&memcg->pcp_counter_lock);
2366 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2367 unsigned long action,
2370 int cpu = (unsigned long)hcpu;
2371 struct memcg_stock_pcp *stock;
2372 struct mem_cgroup *iter;
2374 if (action == CPU_ONLINE)
2377 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2380 for_each_mem_cgroup(iter)
2381 mem_cgroup_drain_pcp_counter(iter, cpu);
2383 stock = &per_cpu(memcg_stock, cpu);
2389 /* See __mem_cgroup_try_charge() for details */
2391 CHARGE_OK, /* success */
2392 CHARGE_RETRY, /* need to retry but retry is not bad */
2393 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2394 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2395 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2398 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2399 unsigned int nr_pages, unsigned int min_pages,
2402 unsigned long csize = nr_pages * PAGE_SIZE;
2403 struct mem_cgroup *mem_over_limit;
2404 struct res_counter *fail_res;
2405 unsigned long flags = 0;
2408 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2411 if (!do_swap_account)
2413 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2417 res_counter_uncharge(&memcg->res, csize);
2418 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2419 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2421 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2423 * Never reclaim on behalf of optional batching, retry with a
2424 * single page instead.
2426 if (nr_pages > min_pages)
2427 return CHARGE_RETRY;
2429 if (!(gfp_mask & __GFP_WAIT))
2430 return CHARGE_WOULDBLOCK;
2432 if (gfp_mask & __GFP_NORETRY)
2433 return CHARGE_NOMEM;
2435 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2436 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2437 return CHARGE_RETRY;
2439 * Even though the limit is exceeded at this point, reclaim
2440 * may have been able to free some pages. Retry the charge
2441 * before killing the task.
2443 * Only for regular pages, though: huge pages are rather
2444 * unlikely to succeed so close to the limit, and we fall back
2445 * to regular pages anyway in case of failure.
2447 if (nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER) && ret)
2448 return CHARGE_RETRY;
2451 * At task move, charge accounts can be doubly counted. So, it's
2452 * better to wait until the end of task_move if something is going on.
2454 if (mem_cgroup_wait_acct_move(mem_over_limit))
2455 return CHARGE_RETRY;
2457 /* If we don't need to call oom-killer at el, return immediately */
2459 return CHARGE_NOMEM;
2461 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask, get_order(csize)))
2462 return CHARGE_OOM_DIE;
2464 return CHARGE_RETRY;
2468 * __mem_cgroup_try_charge() does
2469 * 1. detect memcg to be charged against from passed *mm and *ptr,
2470 * 2. update res_counter
2471 * 3. call memory reclaim if necessary.
2473 * In some special case, if the task is fatal, fatal_signal_pending() or
2474 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2475 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2476 * as possible without any hazards. 2: all pages should have a valid
2477 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2478 * pointer, that is treated as a charge to root_mem_cgroup.
2480 * So __mem_cgroup_try_charge() will return
2481 * 0 ... on success, filling *ptr with a valid memcg pointer.
2482 * -ENOMEM ... charge failure because of resource limits.
2483 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2485 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2486 * the oom-killer can be invoked.
2488 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2490 unsigned int nr_pages,
2491 struct mem_cgroup **ptr,
2494 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2495 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2496 struct mem_cgroup *memcg = NULL;
2500 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2501 * in system level. So, allow to go ahead dying process in addition to
2504 if (unlikely(test_thread_flag(TIF_MEMDIE)
2505 || fatal_signal_pending(current)))
2509 * We always charge the cgroup the mm_struct belongs to.
2510 * The mm_struct's mem_cgroup changes on task migration if the
2511 * thread group leader migrates. It's possible that mm is not
2512 * set, if so charge the root memcg (happens for pagecache usage).
2515 *ptr = root_mem_cgroup;
2517 if (*ptr) { /* css should be a valid one */
2519 if (mem_cgroup_is_root(memcg))
2521 if (consume_stock(memcg, nr_pages))
2523 css_get(&memcg->css);
2525 struct task_struct *p;
2528 p = rcu_dereference(mm->owner);
2530 * Because we don't have task_lock(), "p" can exit.
2531 * In that case, "memcg" can point to root or p can be NULL with
2532 * race with swapoff. Then, we have small risk of mis-accouning.
2533 * But such kind of mis-account by race always happens because
2534 * we don't have cgroup_mutex(). It's overkill and we allo that
2536 * (*) swapoff at el will charge against mm-struct not against
2537 * task-struct. So, mm->owner can be NULL.
2539 memcg = mem_cgroup_from_task(p);
2541 memcg = root_mem_cgroup;
2542 if (mem_cgroup_is_root(memcg)) {
2546 if (consume_stock(memcg, nr_pages)) {
2548 * It seems dagerous to access memcg without css_get().
2549 * But considering how consume_stok works, it's not
2550 * necessary. If consume_stock success, some charges
2551 * from this memcg are cached on this cpu. So, we
2552 * don't need to call css_get()/css_tryget() before
2553 * calling consume_stock().
2558 /* after here, we may be blocked. we need to get refcnt */
2559 if (!css_tryget(&memcg->css)) {
2569 /* If killed, bypass charge */
2570 if (fatal_signal_pending(current)) {
2571 css_put(&memcg->css);
2576 if (oom && !nr_oom_retries) {
2578 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2581 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, nr_pages,
2586 case CHARGE_RETRY: /* not in OOM situation but retry */
2588 css_put(&memcg->css);
2591 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2592 css_put(&memcg->css);
2594 case CHARGE_NOMEM: /* OOM routine works */
2596 css_put(&memcg->css);
2599 /* If oom, we never return -ENOMEM */
2602 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2603 css_put(&memcg->css);
2606 } while (ret != CHARGE_OK);
2608 if (batch > nr_pages)
2609 refill_stock(memcg, batch - nr_pages);
2610 css_put(&memcg->css);
2618 *ptr = root_mem_cgroup;
2623 * Somemtimes we have to undo a charge we got by try_charge().
2624 * This function is for that and do uncharge, put css's refcnt.
2625 * gotten by try_charge().
2627 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2628 unsigned int nr_pages)
2630 if (!mem_cgroup_is_root(memcg)) {
2631 unsigned long bytes = nr_pages * PAGE_SIZE;
2633 res_counter_uncharge(&memcg->res, bytes);
2634 if (do_swap_account)
2635 res_counter_uncharge(&memcg->memsw, bytes);
2640 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2641 * This is useful when moving usage to parent cgroup.
2643 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2644 unsigned int nr_pages)
2646 unsigned long bytes = nr_pages * PAGE_SIZE;
2648 if (mem_cgroup_is_root(memcg))
2651 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2652 if (do_swap_account)
2653 res_counter_uncharge_until(&memcg->memsw,
2654 memcg->memsw.parent, bytes);
2658 * A helper function to get mem_cgroup from ID. must be called under
2659 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2660 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2661 * called against removed memcg.)
2663 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2665 struct cgroup_subsys_state *css;
2667 /* ID 0 is unused ID */
2670 css = css_lookup(&mem_cgroup_subsys, id);
2673 return mem_cgroup_from_css(css);
2676 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2678 struct mem_cgroup *memcg = NULL;
2679 struct page_cgroup *pc;
2683 VM_BUG_ON(!PageLocked(page));
2685 pc = lookup_page_cgroup(page);
2686 lock_page_cgroup(pc);
2687 if (PageCgroupUsed(pc)) {
2688 memcg = pc->mem_cgroup;
2689 if (memcg && !css_tryget(&memcg->css))
2691 } else if (PageSwapCache(page)) {
2692 ent.val = page_private(page);
2693 id = lookup_swap_cgroup_id(ent);
2695 memcg = mem_cgroup_lookup(id);
2696 if (memcg && !css_tryget(&memcg->css))
2700 unlock_page_cgroup(pc);
2704 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2706 unsigned int nr_pages,
2707 enum charge_type ctype,
2710 struct page_cgroup *pc = lookup_page_cgroup(page);
2711 struct zone *uninitialized_var(zone);
2712 struct lruvec *lruvec;
2713 bool was_on_lru = false;
2716 lock_page_cgroup(pc);
2717 VM_BUG_ON(PageCgroupUsed(pc));
2719 * we don't need page_cgroup_lock about tail pages, becase they are not
2720 * accessed by any other context at this point.
2724 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2725 * may already be on some other mem_cgroup's LRU. Take care of it.
2728 zone = page_zone(page);
2729 spin_lock_irq(&zone->lru_lock);
2730 if (PageLRU(page)) {
2731 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2733 del_page_from_lru_list(page, lruvec, page_lru(page));
2738 pc->mem_cgroup = memcg;
2740 * We access a page_cgroup asynchronously without lock_page_cgroup().
2741 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2742 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2743 * before USED bit, we need memory barrier here.
2744 * See mem_cgroup_add_lru_list(), etc.
2747 SetPageCgroupUsed(pc);
2751 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2752 VM_BUG_ON(PageLRU(page));
2754 add_page_to_lru_list(page, lruvec, page_lru(page));
2756 spin_unlock_irq(&zone->lru_lock);
2759 if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
2764 mem_cgroup_charge_statistics(memcg, anon, nr_pages);
2765 unlock_page_cgroup(pc);
2768 * "charge_statistics" updated event counter. Then, check it.
2769 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2770 * if they exceeds softlimit.
2772 memcg_check_events(memcg, page);
2775 #ifdef CONFIG_MEMCG_KMEM
2776 static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg)
2778 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) &&
2779 (memcg->kmem_account_flags & KMEM_ACCOUNTED_MASK);
2782 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size)
2784 struct res_counter *fail_res;
2785 struct mem_cgroup *_memcg;
2789 ret = res_counter_charge(&memcg->kmem, size, &fail_res);
2794 * Conditions under which we can wait for the oom_killer. Those are
2795 * the same conditions tested by the core page allocator
2797 may_oom = (gfp & __GFP_FS) && !(gfp & __GFP_NORETRY);
2800 ret = __mem_cgroup_try_charge(NULL, gfp, size >> PAGE_SHIFT,
2803 if (ret == -EINTR) {
2805 * __mem_cgroup_try_charge() chosed to bypass to root due to
2806 * OOM kill or fatal signal. Since our only options are to
2807 * either fail the allocation or charge it to this cgroup, do
2808 * it as a temporary condition. But we can't fail. From a
2809 * kmem/slab perspective, the cache has already been selected,
2810 * by mem_cgroup_kmem_get_cache(), so it is too late to change
2813 * This condition will only trigger if the task entered
2814 * memcg_charge_kmem in a sane state, but was OOM-killed during
2815 * __mem_cgroup_try_charge() above. Tasks that were already
2816 * dying when the allocation triggers should have been already
2817 * directed to the root cgroup in memcontrol.h
2819 res_counter_charge_nofail(&memcg->res, size, &fail_res);
2820 if (do_swap_account)
2821 res_counter_charge_nofail(&memcg->memsw, size,
2825 res_counter_uncharge(&memcg->kmem, size);
2830 static void memcg_uncharge_kmem(struct mem_cgroup *memcg, u64 size)
2832 res_counter_uncharge(&memcg->res, size);
2833 if (do_swap_account)
2834 res_counter_uncharge(&memcg->memsw, size);
2837 if (res_counter_uncharge(&memcg->kmem, size))
2840 if (memcg_kmem_test_and_clear_dead(memcg))
2841 mem_cgroup_put(memcg);
2844 void memcg_cache_list_add(struct mem_cgroup *memcg, struct kmem_cache *cachep)
2849 mutex_lock(&memcg->slab_caches_mutex);
2850 list_add(&cachep->memcg_params->list, &memcg->memcg_slab_caches);
2851 mutex_unlock(&memcg->slab_caches_mutex);
2855 * helper for acessing a memcg's index. It will be used as an index in the
2856 * child cache array in kmem_cache, and also to derive its name. This function
2857 * will return -1 when this is not a kmem-limited memcg.
2859 int memcg_cache_id(struct mem_cgroup *memcg)
2861 return memcg ? memcg->kmemcg_id : -1;
2865 * This ends up being protected by the set_limit mutex, during normal
2866 * operation, because that is its main call site.
2868 * But when we create a new cache, we can call this as well if its parent
2869 * is kmem-limited. That will have to hold set_limit_mutex as well.
2871 int memcg_update_cache_sizes(struct mem_cgroup *memcg)
2875 num = ida_simple_get(&kmem_limited_groups,
2876 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2880 * After this point, kmem_accounted (that we test atomically in
2881 * the beginning of this conditional), is no longer 0. This
2882 * guarantees only one process will set the following boolean
2883 * to true. We don't need test_and_set because we're protected
2884 * by the set_limit_mutex anyway.
2886 memcg_kmem_set_activated(memcg);
2888 ret = memcg_update_all_caches(num+1);
2890 ida_simple_remove(&kmem_limited_groups, num);
2891 memcg_kmem_clear_activated(memcg);
2895 memcg->kmemcg_id = num;
2896 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
2897 mutex_init(&memcg->slab_caches_mutex);
2901 static size_t memcg_caches_array_size(int num_groups)
2904 if (num_groups <= 0)
2907 size = 2 * num_groups;
2908 if (size < MEMCG_CACHES_MIN_SIZE)
2909 size = MEMCG_CACHES_MIN_SIZE;
2910 else if (size > MEMCG_CACHES_MAX_SIZE)
2911 size = MEMCG_CACHES_MAX_SIZE;
2917 * We should update the current array size iff all caches updates succeed. This
2918 * can only be done from the slab side. The slab mutex needs to be held when
2921 void memcg_update_array_size(int num)
2923 if (num > memcg_limited_groups_array_size)
2924 memcg_limited_groups_array_size = memcg_caches_array_size(num);
2927 int memcg_update_cache_size(struct kmem_cache *s, int num_groups)
2929 struct memcg_cache_params *cur_params = s->memcg_params;
2931 VM_BUG_ON(s->memcg_params && !s->memcg_params->is_root_cache);
2933 if (num_groups > memcg_limited_groups_array_size) {
2935 ssize_t size = memcg_caches_array_size(num_groups);
2937 size *= sizeof(void *);
2938 size += sizeof(struct memcg_cache_params);
2940 s->memcg_params = kzalloc(size, GFP_KERNEL);
2941 if (!s->memcg_params) {
2942 s->memcg_params = cur_params;
2946 s->memcg_params->is_root_cache = true;
2949 * There is the chance it will be bigger than
2950 * memcg_limited_groups_array_size, if we failed an allocation
2951 * in a cache, in which case all caches updated before it, will
2952 * have a bigger array.
2954 * But if that is the case, the data after
2955 * memcg_limited_groups_array_size is certainly unused
2957 for (i = 0; i < memcg_limited_groups_array_size; i++) {
2958 if (!cur_params->memcg_caches[i])
2960 s->memcg_params->memcg_caches[i] =
2961 cur_params->memcg_caches[i];
2965 * Ideally, we would wait until all caches succeed, and only
2966 * then free the old one. But this is not worth the extra
2967 * pointer per-cache we'd have to have for this.
2969 * It is not a big deal if some caches are left with a size
2970 * bigger than the others. And all updates will reset this
2978 int memcg_register_cache(struct mem_cgroup *memcg, struct kmem_cache *s)
2980 size_t size = sizeof(struct memcg_cache_params);
2982 if (!memcg_kmem_enabled())
2986 size += memcg_limited_groups_array_size * sizeof(void *);
2988 s->memcg_params = kzalloc(size, GFP_KERNEL);
2989 if (!s->memcg_params)
2993 s->memcg_params->memcg = memcg;
2997 void memcg_release_cache(struct kmem_cache *s)
2999 struct kmem_cache *root;
3000 struct mem_cgroup *memcg;
3004 * This happens, for instance, when a root cache goes away before we
3007 if (!s->memcg_params)
3010 if (s->memcg_params->is_root_cache)
3013 memcg = s->memcg_params->memcg;
3014 id = memcg_cache_id(memcg);
3016 root = s->memcg_params->root_cache;
3017 root->memcg_params->memcg_caches[id] = NULL;
3018 mem_cgroup_put(memcg);
3020 mutex_lock(&memcg->slab_caches_mutex);
3021 list_del(&s->memcg_params->list);
3022 mutex_unlock(&memcg->slab_caches_mutex);
3025 kfree(s->memcg_params);
3028 static char *memcg_cache_name(struct mem_cgroup *memcg, struct kmem_cache *s)
3031 struct dentry *dentry;
3034 dentry = rcu_dereference(memcg->css.cgroup->dentry);
3037 BUG_ON(dentry == NULL);
3039 name = kasprintf(GFP_KERNEL, "%s(%d:%s)", s->name,
3040 memcg_cache_id(memcg), dentry->d_name.name);
3045 static struct kmem_cache *kmem_cache_dup(struct mem_cgroup *memcg,
3046 struct kmem_cache *s)
3049 struct kmem_cache *new;
3051 name = memcg_cache_name(memcg, s);
3055 new = kmem_cache_create_memcg(memcg, name, s->object_size, s->align,
3056 (s->flags & ~SLAB_PANIC), s->ctor);
3063 * This lock protects updaters, not readers. We want readers to be as fast as
3064 * they can, and they will either see NULL or a valid cache value. Our model
3065 * allow them to see NULL, in which case the root memcg will be selected.
3067 * We need this lock because multiple allocations to the same cache from a non
3068 * will span more than one worker. Only one of them can create the cache.
3070 static DEFINE_MUTEX(memcg_cache_mutex);
3071 static struct kmem_cache *memcg_create_kmem_cache(struct mem_cgroup *memcg,
3072 struct kmem_cache *cachep)
3074 struct kmem_cache *new_cachep;
3077 BUG_ON(!memcg_can_account_kmem(memcg));
3079 idx = memcg_cache_id(memcg);
3081 mutex_lock(&memcg_cache_mutex);
3082 new_cachep = cachep->memcg_params->memcg_caches[idx];
3086 new_cachep = kmem_cache_dup(memcg, cachep);
3088 if (new_cachep == NULL) {
3089 new_cachep = cachep;
3093 mem_cgroup_get(memcg);
3094 new_cachep->memcg_params->root_cache = cachep;
3096 cachep->memcg_params->memcg_caches[idx] = new_cachep;
3098 * the readers won't lock, make sure everybody sees the updated value,
3099 * so they won't put stuff in the queue again for no reason
3103 mutex_unlock(&memcg_cache_mutex);
3107 struct create_work {
3108 struct mem_cgroup *memcg;
3109 struct kmem_cache *cachep;
3110 struct work_struct work;
3113 static void memcg_create_cache_work_func(struct work_struct *w)
3115 struct create_work *cw;
3117 cw = container_of(w, struct create_work, work);
3118 memcg_create_kmem_cache(cw->memcg, cw->cachep);
3119 /* Drop the reference gotten when we enqueued. */
3120 css_put(&cw->memcg->css);
3125 * Enqueue the creation of a per-memcg kmem_cache.
3126 * Called with rcu_read_lock.
3128 static void memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3129 struct kmem_cache *cachep)
3131 struct create_work *cw;
3133 cw = kmalloc(sizeof(struct create_work), GFP_NOWAIT);
3137 /* The corresponding put will be done in the workqueue. */
3138 if (!css_tryget(&memcg->css)) {
3144 cw->cachep = cachep;
3146 INIT_WORK(&cw->work, memcg_create_cache_work_func);
3147 schedule_work(&cw->work);
3151 * Return the kmem_cache we're supposed to use for a slab allocation.
3152 * We try to use the current memcg's version of the cache.
3154 * If the cache does not exist yet, if we are the first user of it,
3155 * we either create it immediately, if possible, or create it asynchronously
3157 * In the latter case, we will let the current allocation go through with
3158 * the original cache.
3160 * Can't be called in interrupt context or from kernel threads.
3161 * This function needs to be called with rcu_read_lock() held.
3163 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
3166 struct mem_cgroup *memcg;
3169 VM_BUG_ON(!cachep->memcg_params);
3170 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
3173 memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
3176 if (!memcg_can_account_kmem(memcg))
3179 idx = memcg_cache_id(memcg);
3182 * barrier to mare sure we're always seeing the up to date value. The
3183 * code updating memcg_caches will issue a write barrier to match this.
3185 read_barrier_depends();
3186 if (unlikely(cachep->memcg_params->memcg_caches[idx] == NULL)) {
3188 * If we are in a safe context (can wait, and not in interrupt
3189 * context), we could be be predictable and return right away.
3190 * This would guarantee that the allocation being performed
3191 * already belongs in the new cache.
3193 * However, there are some clashes that can arrive from locking.
3194 * For instance, because we acquire the slab_mutex while doing
3195 * kmem_cache_dup, this means no further allocation could happen
3196 * with the slab_mutex held.
3198 * Also, because cache creation issue get_online_cpus(), this
3199 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3200 * that ends up reversed during cpu hotplug. (cpuset allocates
3201 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3202 * better to defer everything.
3204 memcg_create_cache_enqueue(memcg, cachep);
3208 return cachep->memcg_params->memcg_caches[idx];
3210 EXPORT_SYMBOL(__memcg_kmem_get_cache);
3213 * We need to verify if the allocation against current->mm->owner's memcg is
3214 * possible for the given order. But the page is not allocated yet, so we'll
3215 * need a further commit step to do the final arrangements.
3217 * It is possible for the task to switch cgroups in this mean time, so at
3218 * commit time, we can't rely on task conversion any longer. We'll then use
3219 * the handle argument to return to the caller which cgroup we should commit
3220 * against. We could also return the memcg directly and avoid the pointer
3221 * passing, but a boolean return value gives better semantics considering
3222 * the compiled-out case as well.
3224 * Returning true means the allocation is possible.
3227 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3229 struct mem_cgroup *memcg;
3233 memcg = try_get_mem_cgroup_from_mm(current->mm);
3236 * very rare case described in mem_cgroup_from_task. Unfortunately there
3237 * isn't much we can do without complicating this too much, and it would
3238 * be gfp-dependent anyway. Just let it go
3240 if (unlikely(!memcg))
3243 if (!memcg_can_account_kmem(memcg)) {
3244 css_put(&memcg->css);
3248 ret = memcg_charge_kmem(memcg, gfp, PAGE_SIZE << order);
3252 css_put(&memcg->css);
3256 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3259 struct page_cgroup *pc;
3261 VM_BUG_ON(mem_cgroup_is_root(memcg));
3263 /* The page allocation failed. Revert */
3265 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3269 pc = lookup_page_cgroup(page);
3270 lock_page_cgroup(pc);
3271 pc->mem_cgroup = memcg;
3272 SetPageCgroupUsed(pc);
3273 unlock_page_cgroup(pc);
3276 void __memcg_kmem_uncharge_pages(struct page *page, int order)
3278 struct mem_cgroup *memcg = NULL;
3279 struct page_cgroup *pc;
3282 pc = lookup_page_cgroup(page);
3284 * Fast unlocked return. Theoretically might have changed, have to
3285 * check again after locking.
3287 if (!PageCgroupUsed(pc))
3290 lock_page_cgroup(pc);
3291 if (PageCgroupUsed(pc)) {
3292 memcg = pc->mem_cgroup;
3293 ClearPageCgroupUsed(pc);
3295 unlock_page_cgroup(pc);
3298 * We trust that only if there is a memcg associated with the page, it
3299 * is a valid allocation
3304 VM_BUG_ON(mem_cgroup_is_root(memcg));
3305 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3307 #endif /* CONFIG_MEMCG_KMEM */
3309 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3311 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3313 * Because tail pages are not marked as "used", set it. We're under
3314 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3315 * charge/uncharge will be never happen and move_account() is done under
3316 * compound_lock(), so we don't have to take care of races.
3318 void mem_cgroup_split_huge_fixup(struct page *head)
3320 struct page_cgroup *head_pc = lookup_page_cgroup(head);
3321 struct page_cgroup *pc;
3324 if (mem_cgroup_disabled())
3326 for (i = 1; i < HPAGE_PMD_NR; i++) {
3328 pc->mem_cgroup = head_pc->mem_cgroup;
3329 smp_wmb();/* see __commit_charge() */
3330 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
3333 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3336 * mem_cgroup_move_account - move account of the page
3338 * @nr_pages: number of regular pages (>1 for huge pages)
3339 * @pc: page_cgroup of the page.
3340 * @from: mem_cgroup which the page is moved from.
3341 * @to: mem_cgroup which the page is moved to. @from != @to.
3343 * The caller must confirm following.
3344 * - page is not on LRU (isolate_page() is useful.)
3345 * - compound_lock is held when nr_pages > 1
3347 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3350 static int mem_cgroup_move_account(struct page *page,
3351 unsigned int nr_pages,
3352 struct page_cgroup *pc,
3353 struct mem_cgroup *from,
3354 struct mem_cgroup *to)
3356 unsigned long flags;
3358 bool anon = PageAnon(page);
3360 VM_BUG_ON(from == to);
3361 VM_BUG_ON(PageLRU(page));
3363 * The page is isolated from LRU. So, collapse function
3364 * will not handle this page. But page splitting can happen.
3365 * Do this check under compound_page_lock(). The caller should
3369 if (nr_pages > 1 && !PageTransHuge(page))
3372 lock_page_cgroup(pc);
3375 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3378 move_lock_mem_cgroup(from, &flags);
3380 if (!anon && page_mapped(page)) {
3381 /* Update mapped_file data for mem_cgroup */
3383 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
3384 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
3387 mem_cgroup_charge_statistics(from, anon, -nr_pages);
3389 /* caller should have done css_get */
3390 pc->mem_cgroup = to;
3391 mem_cgroup_charge_statistics(to, anon, nr_pages);
3392 move_unlock_mem_cgroup(from, &flags);
3395 unlock_page_cgroup(pc);
3399 memcg_check_events(to, page);
3400 memcg_check_events(from, page);
3406 * mem_cgroup_move_parent - moves page to the parent group
3407 * @page: the page to move
3408 * @pc: page_cgroup of the page
3409 * @child: page's cgroup
3411 * move charges to its parent or the root cgroup if the group has no
3412 * parent (aka use_hierarchy==0).
3413 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3414 * mem_cgroup_move_account fails) the failure is always temporary and
3415 * it signals a race with a page removal/uncharge or migration. In the
3416 * first case the page is on the way out and it will vanish from the LRU
3417 * on the next attempt and the call should be retried later.
3418 * Isolation from the LRU fails only if page has been isolated from
3419 * the LRU since we looked at it and that usually means either global
3420 * reclaim or migration going on. The page will either get back to the
3422 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3423 * (!PageCgroupUsed) or moved to a different group. The page will
3424 * disappear in the next attempt.
3426 static int mem_cgroup_move_parent(struct page *page,
3427 struct page_cgroup *pc,
3428 struct mem_cgroup *child)
3430 struct mem_cgroup *parent;
3431 unsigned int nr_pages;
3432 unsigned long uninitialized_var(flags);
3435 VM_BUG_ON(mem_cgroup_is_root(child));
3438 if (!get_page_unless_zero(page))
3440 if (isolate_lru_page(page))
3443 nr_pages = hpage_nr_pages(page);
3445 parent = parent_mem_cgroup(child);
3447 * If no parent, move charges to root cgroup.
3450 parent = root_mem_cgroup;
3453 VM_BUG_ON(!PageTransHuge(page));
3454 flags = compound_lock_irqsave(page);
3457 ret = mem_cgroup_move_account(page, nr_pages,
3460 __mem_cgroup_cancel_local_charge(child, nr_pages);
3463 compound_unlock_irqrestore(page, flags);
3464 putback_lru_page(page);
3472 * Charge the memory controller for page usage.
3474 * 0 if the charge was successful
3475 * < 0 if the cgroup is over its limit
3477 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
3478 gfp_t gfp_mask, enum charge_type ctype)
3480 struct mem_cgroup *memcg = NULL;
3481 unsigned int nr_pages = 1;
3485 if (PageTransHuge(page)) {
3486 nr_pages <<= compound_order(page);
3487 VM_BUG_ON(!PageTransHuge(page));
3489 * Never OOM-kill a process for a huge page. The
3490 * fault handler will fall back to regular pages.
3495 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
3498 __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false);
3502 int mem_cgroup_newpage_charge(struct page *page,
3503 struct mm_struct *mm, gfp_t gfp_mask)
3505 if (mem_cgroup_disabled())
3507 VM_BUG_ON(page_mapped(page));
3508 VM_BUG_ON(page->mapping && !PageAnon(page));
3510 return mem_cgroup_charge_common(page, mm, gfp_mask,
3511 MEM_CGROUP_CHARGE_TYPE_ANON);
3515 * While swap-in, try_charge -> commit or cancel, the page is locked.
3516 * And when try_charge() successfully returns, one refcnt to memcg without
3517 * struct page_cgroup is acquired. This refcnt will be consumed by
3518 * "commit()" or removed by "cancel()"
3520 static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm,
3523 struct mem_cgroup **memcgp)
3525 struct mem_cgroup *memcg;
3526 struct page_cgroup *pc;
3529 pc = lookup_page_cgroup(page);
3531 * Every swap fault against a single page tries to charge the
3532 * page, bail as early as possible. shmem_unuse() encounters
3533 * already charged pages, too. The USED bit is protected by
3534 * the page lock, which serializes swap cache removal, which
3535 * in turn serializes uncharging.
3537 if (PageCgroupUsed(pc))
3539 if (!do_swap_account)
3541 memcg = try_get_mem_cgroup_from_page(page);
3545 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
3546 css_put(&memcg->css);
3551 ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
3557 int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page,
3558 gfp_t gfp_mask, struct mem_cgroup **memcgp)
3561 if (mem_cgroup_disabled())
3564 * A racing thread's fault, or swapoff, may have already
3565 * updated the pte, and even removed page from swap cache: in
3566 * those cases unuse_pte()'s pte_same() test will fail; but
3567 * there's also a KSM case which does need to charge the page.
3569 if (!PageSwapCache(page)) {
3572 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, memcgp, true);
3577 return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp);
3580 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
3582 if (mem_cgroup_disabled())
3586 __mem_cgroup_cancel_charge(memcg, 1);
3590 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
3591 enum charge_type ctype)
3593 if (mem_cgroup_disabled())
3598 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
3600 * Now swap is on-memory. This means this page may be
3601 * counted both as mem and swap....double count.
3602 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
3603 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
3604 * may call delete_from_swap_cache() before reach here.
3606 if (do_swap_account && PageSwapCache(page)) {
3607 swp_entry_t ent = {.val = page_private(page)};
3608 mem_cgroup_uncharge_swap(ent);
3612 void mem_cgroup_commit_charge_swapin(struct page *page,
3613 struct mem_cgroup *memcg)
3615 __mem_cgroup_commit_charge_swapin(page, memcg,
3616 MEM_CGROUP_CHARGE_TYPE_ANON);
3619 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
3622 struct mem_cgroup *memcg = NULL;
3623 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
3626 if (mem_cgroup_disabled())
3628 if (PageCompound(page))
3631 if (!PageSwapCache(page))
3632 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
3633 else { /* page is swapcache/shmem */
3634 ret = __mem_cgroup_try_charge_swapin(mm, page,
3637 __mem_cgroup_commit_charge_swapin(page, memcg, type);
3642 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
3643 unsigned int nr_pages,
3644 const enum charge_type ctype)
3646 struct memcg_batch_info *batch = NULL;
3647 bool uncharge_memsw = true;
3649 /* If swapout, usage of swap doesn't decrease */
3650 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
3651 uncharge_memsw = false;
3653 batch = ¤t->memcg_batch;
3655 * In usual, we do css_get() when we remember memcg pointer.
3656 * But in this case, we keep res->usage until end of a series of
3657 * uncharges. Then, it's ok to ignore memcg's refcnt.
3660 batch->memcg = memcg;
3662 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
3663 * In those cases, all pages freed continuously can be expected to be in
3664 * the same cgroup and we have chance to coalesce uncharges.
3665 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
3666 * because we want to do uncharge as soon as possible.
3669 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
3670 goto direct_uncharge;
3673 goto direct_uncharge;
3676 * In typical case, batch->memcg == mem. This means we can
3677 * merge a series of uncharges to an uncharge of res_counter.
3678 * If not, we uncharge res_counter ony by one.
3680 if (batch->memcg != memcg)
3681 goto direct_uncharge;
3682 /* remember freed charge and uncharge it later */
3685 batch->memsw_nr_pages++;
3688 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
3690 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
3691 if (unlikely(batch->memcg != memcg))
3692 memcg_oom_recover(memcg);
3696 * uncharge if !page_mapped(page)
3698 static struct mem_cgroup *
3699 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
3702 struct mem_cgroup *memcg = NULL;
3703 unsigned int nr_pages = 1;
3704 struct page_cgroup *pc;
3707 if (mem_cgroup_disabled())
3710 VM_BUG_ON(PageSwapCache(page));
3712 if (PageTransHuge(page)) {
3713 nr_pages <<= compound_order(page);
3714 VM_BUG_ON(!PageTransHuge(page));
3717 * Check if our page_cgroup is valid
3719 pc = lookup_page_cgroup(page);
3720 if (unlikely(!PageCgroupUsed(pc)))
3723 lock_page_cgroup(pc);
3725 memcg = pc->mem_cgroup;
3727 if (!PageCgroupUsed(pc))
3730 anon = PageAnon(page);
3733 case MEM_CGROUP_CHARGE_TYPE_ANON:
3735 * Generally PageAnon tells if it's the anon statistics to be
3736 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
3737 * used before page reached the stage of being marked PageAnon.
3741 case MEM_CGROUP_CHARGE_TYPE_DROP:
3742 /* See mem_cgroup_prepare_migration() */
3743 if (page_mapped(page))
3746 * Pages under migration may not be uncharged. But
3747 * end_migration() /must/ be the one uncharging the
3748 * unused post-migration page and so it has to call
3749 * here with the migration bit still set. See the
3750 * res_counter handling below.
3752 if (!end_migration && PageCgroupMigration(pc))
3755 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
3756 if (!PageAnon(page)) { /* Shared memory */
3757 if (page->mapping && !page_is_file_cache(page))
3759 } else if (page_mapped(page)) /* Anon */
3766 mem_cgroup_charge_statistics(memcg, anon, -nr_pages);
3768 ClearPageCgroupUsed(pc);
3770 * pc->mem_cgroup is not cleared here. It will be accessed when it's
3771 * freed from LRU. This is safe because uncharged page is expected not
3772 * to be reused (freed soon). Exception is SwapCache, it's handled by
3773 * special functions.
3776 unlock_page_cgroup(pc);
3778 * even after unlock, we have memcg->res.usage here and this memcg
3779 * will never be freed.
3781 memcg_check_events(memcg, page);
3782 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
3783 mem_cgroup_swap_statistics(memcg, true);
3784 mem_cgroup_get(memcg);
3787 * Migration does not charge the res_counter for the
3788 * replacement page, so leave it alone when phasing out the
3789 * page that is unused after the migration.
3791 if (!end_migration && !mem_cgroup_is_root(memcg))
3792 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
3797 unlock_page_cgroup(pc);
3801 void mem_cgroup_uncharge_page(struct page *page)
3804 if (page_mapped(page))
3806 VM_BUG_ON(page->mapping && !PageAnon(page));
3807 if (PageSwapCache(page))
3809 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
3812 void mem_cgroup_uncharge_cache_page(struct page *page)
3814 VM_BUG_ON(page_mapped(page));
3815 VM_BUG_ON(page->mapping);
3816 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
3820 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
3821 * In that cases, pages are freed continuously and we can expect pages
3822 * are in the same memcg. All these calls itself limits the number of
3823 * pages freed at once, then uncharge_start/end() is called properly.
3824 * This may be called prural(2) times in a context,
3827 void mem_cgroup_uncharge_start(void)
3829 current->memcg_batch.do_batch++;
3830 /* We can do nest. */
3831 if (current->memcg_batch.do_batch == 1) {
3832 current->memcg_batch.memcg = NULL;
3833 current->memcg_batch.nr_pages = 0;
3834 current->memcg_batch.memsw_nr_pages = 0;
3838 void mem_cgroup_uncharge_end(void)
3840 struct memcg_batch_info *batch = ¤t->memcg_batch;
3842 if (!batch->do_batch)
3846 if (batch->do_batch) /* If stacked, do nothing. */
3852 * This "batch->memcg" is valid without any css_get/put etc...
3853 * bacause we hide charges behind us.
3855 if (batch->nr_pages)
3856 res_counter_uncharge(&batch->memcg->res,
3857 batch->nr_pages * PAGE_SIZE);
3858 if (batch->memsw_nr_pages)
3859 res_counter_uncharge(&batch->memcg->memsw,
3860 batch->memsw_nr_pages * PAGE_SIZE);
3861 memcg_oom_recover(batch->memcg);
3862 /* forget this pointer (for sanity check) */
3863 batch->memcg = NULL;
3868 * called after __delete_from_swap_cache() and drop "page" account.
3869 * memcg information is recorded to swap_cgroup of "ent"
3872 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
3874 struct mem_cgroup *memcg;
3875 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
3877 if (!swapout) /* this was a swap cache but the swap is unused ! */
3878 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
3880 memcg = __mem_cgroup_uncharge_common(page, ctype, false);
3883 * record memcg information, if swapout && memcg != NULL,
3884 * mem_cgroup_get() was called in uncharge().
3886 if (do_swap_account && swapout && memcg)
3887 swap_cgroup_record(ent, css_id(&memcg->css));
3891 #ifdef CONFIG_MEMCG_SWAP
3893 * called from swap_entry_free(). remove record in swap_cgroup and
3894 * uncharge "memsw" account.
3896 void mem_cgroup_uncharge_swap(swp_entry_t ent)
3898 struct mem_cgroup *memcg;
3901 if (!do_swap_account)
3904 id = swap_cgroup_record(ent, 0);
3906 memcg = mem_cgroup_lookup(id);
3909 * We uncharge this because swap is freed.
3910 * This memcg can be obsolete one. We avoid calling css_tryget
3912 if (!mem_cgroup_is_root(memcg))
3913 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
3914 mem_cgroup_swap_statistics(memcg, false);
3915 mem_cgroup_put(memcg);
3921 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3922 * @entry: swap entry to be moved
3923 * @from: mem_cgroup which the entry is moved from
3924 * @to: mem_cgroup which the entry is moved to
3926 * It succeeds only when the swap_cgroup's record for this entry is the same
3927 * as the mem_cgroup's id of @from.
3929 * Returns 0 on success, -EINVAL on failure.
3931 * The caller must have charged to @to, IOW, called res_counter_charge() about
3932 * both res and memsw, and called css_get().
3934 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3935 struct mem_cgroup *from, struct mem_cgroup *to)
3937 unsigned short old_id, new_id;
3939 old_id = css_id(&from->css);
3940 new_id = css_id(&to->css);
3942 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3943 mem_cgroup_swap_statistics(from, false);
3944 mem_cgroup_swap_statistics(to, true);
3946 * This function is only called from task migration context now.
3947 * It postpones res_counter and refcount handling till the end
3948 * of task migration(mem_cgroup_clear_mc()) for performance
3949 * improvement. But we cannot postpone mem_cgroup_get(to)
3950 * because if the process that has been moved to @to does
3951 * swap-in, the refcount of @to might be decreased to 0.
3959 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3960 struct mem_cgroup *from, struct mem_cgroup *to)
3967 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
3970 void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
3971 struct mem_cgroup **memcgp)
3973 struct mem_cgroup *memcg = NULL;
3974 unsigned int nr_pages = 1;
3975 struct page_cgroup *pc;
3976 enum charge_type ctype;
3980 if (mem_cgroup_disabled())
3983 if (PageTransHuge(page))
3984 nr_pages <<= compound_order(page);
3986 pc = lookup_page_cgroup(page);
3987 lock_page_cgroup(pc);
3988 if (PageCgroupUsed(pc)) {
3989 memcg = pc->mem_cgroup;
3990 css_get(&memcg->css);
3992 * At migrating an anonymous page, its mapcount goes down
3993 * to 0 and uncharge() will be called. But, even if it's fully
3994 * unmapped, migration may fail and this page has to be
3995 * charged again. We set MIGRATION flag here and delay uncharge
3996 * until end_migration() is called
3998 * Corner Case Thinking
4000 * When the old page was mapped as Anon and it's unmap-and-freed
4001 * while migration was ongoing.
4002 * If unmap finds the old page, uncharge() of it will be delayed
4003 * until end_migration(). If unmap finds a new page, it's
4004 * uncharged when it make mapcount to be 1->0. If unmap code
4005 * finds swap_migration_entry, the new page will not be mapped
4006 * and end_migration() will find it(mapcount==0).
4009 * When the old page was mapped but migraion fails, the kernel
4010 * remaps it. A charge for it is kept by MIGRATION flag even
4011 * if mapcount goes down to 0. We can do remap successfully
4012 * without charging it again.
4015 * The "old" page is under lock_page() until the end of
4016 * migration, so, the old page itself will not be swapped-out.
4017 * If the new page is swapped out before end_migraton, our
4018 * hook to usual swap-out path will catch the event.
4021 SetPageCgroupMigration(pc);
4023 unlock_page_cgroup(pc);
4025 * If the page is not charged at this point,
4033 * We charge new page before it's used/mapped. So, even if unlock_page()
4034 * is called before end_migration, we can catch all events on this new
4035 * page. In the case new page is migrated but not remapped, new page's
4036 * mapcount will be finally 0 and we call uncharge in end_migration().
4039 ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
4041 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
4043 * The page is committed to the memcg, but it's not actually
4044 * charged to the res_counter since we plan on replacing the
4045 * old one and only one page is going to be left afterwards.
4047 __mem_cgroup_commit_charge(memcg, newpage, nr_pages, ctype, false);
4050 /* remove redundant charge if migration failed*/
4051 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
4052 struct page *oldpage, struct page *newpage, bool migration_ok)
4054 struct page *used, *unused;
4055 struct page_cgroup *pc;
4061 if (!migration_ok) {
4068 anon = PageAnon(used);
4069 __mem_cgroup_uncharge_common(unused,
4070 anon ? MEM_CGROUP_CHARGE_TYPE_ANON
4071 : MEM_CGROUP_CHARGE_TYPE_CACHE,
4073 css_put(&memcg->css);
4075 * We disallowed uncharge of pages under migration because mapcount
4076 * of the page goes down to zero, temporarly.
4077 * Clear the flag and check the page should be charged.
4079 pc = lookup_page_cgroup(oldpage);
4080 lock_page_cgroup(pc);
4081 ClearPageCgroupMigration(pc);
4082 unlock_page_cgroup(pc);
4085 * If a page is a file cache, radix-tree replacement is very atomic
4086 * and we can skip this check. When it was an Anon page, its mapcount
4087 * goes down to 0. But because we added MIGRATION flage, it's not
4088 * uncharged yet. There are several case but page->mapcount check
4089 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4090 * check. (see prepare_charge() also)
4093 mem_cgroup_uncharge_page(used);
4097 * At replace page cache, newpage is not under any memcg but it's on
4098 * LRU. So, this function doesn't touch res_counter but handles LRU
4099 * in correct way. Both pages are locked so we cannot race with uncharge.
4101 void mem_cgroup_replace_page_cache(struct page *oldpage,
4102 struct page *newpage)
4104 struct mem_cgroup *memcg = NULL;
4105 struct page_cgroup *pc;
4106 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4108 if (mem_cgroup_disabled())
4111 pc = lookup_page_cgroup(oldpage);
4112 /* fix accounting on old pages */
4113 lock_page_cgroup(pc);
4114 if (PageCgroupUsed(pc)) {
4115 memcg = pc->mem_cgroup;
4116 mem_cgroup_charge_statistics(memcg, false, -1);
4117 ClearPageCgroupUsed(pc);
4119 unlock_page_cgroup(pc);
4122 * When called from shmem_replace_page(), in some cases the
4123 * oldpage has already been charged, and in some cases not.
4128 * Even if newpage->mapping was NULL before starting replacement,
4129 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4130 * LRU while we overwrite pc->mem_cgroup.
4132 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
4135 #ifdef CONFIG_DEBUG_VM
4136 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
4138 struct page_cgroup *pc;
4140 pc = lookup_page_cgroup(page);
4142 * Can be NULL while feeding pages into the page allocator for
4143 * the first time, i.e. during boot or memory hotplug;
4144 * or when mem_cgroup_disabled().
4146 if (likely(pc) && PageCgroupUsed(pc))
4151 bool mem_cgroup_bad_page_check(struct page *page)
4153 if (mem_cgroup_disabled())
4156 return lookup_page_cgroup_used(page) != NULL;
4159 void mem_cgroup_print_bad_page(struct page *page)
4161 struct page_cgroup *pc;
4163 pc = lookup_page_cgroup_used(page);
4165 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4166 pc, pc->flags, pc->mem_cgroup);
4171 static DEFINE_MUTEX(set_limit_mutex);
4173 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
4174 unsigned long long val)
4177 u64 memswlimit, memlimit;
4179 int children = mem_cgroup_count_children(memcg);
4180 u64 curusage, oldusage;
4184 * For keeping hierarchical_reclaim simple, how long we should retry
4185 * is depends on callers. We set our retry-count to be function
4186 * of # of children which we should visit in this loop.
4188 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
4190 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4193 while (retry_count) {
4194 if (signal_pending(current)) {
4199 * Rather than hide all in some function, I do this in
4200 * open coded manner. You see what this really does.
4201 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4203 mutex_lock(&set_limit_mutex);
4204 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4205 if (memswlimit < val) {
4207 mutex_unlock(&set_limit_mutex);
4211 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4215 ret = res_counter_set_limit(&memcg->res, val);
4217 if (memswlimit == val)
4218 memcg->memsw_is_minimum = true;
4220 memcg->memsw_is_minimum = false;
4222 mutex_unlock(&set_limit_mutex);
4227 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4228 MEM_CGROUP_RECLAIM_SHRINK);
4229 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4230 /* Usage is reduced ? */
4231 if (curusage >= oldusage)
4234 oldusage = curusage;
4236 if (!ret && enlarge)
4237 memcg_oom_recover(memcg);
4242 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
4243 unsigned long long val)
4246 u64 memlimit, memswlimit, oldusage, curusage;
4247 int children = mem_cgroup_count_children(memcg);
4251 /* see mem_cgroup_resize_res_limit */
4252 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
4253 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4254 while (retry_count) {
4255 if (signal_pending(current)) {
4260 * Rather than hide all in some function, I do this in
4261 * open coded manner. You see what this really does.
4262 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4264 mutex_lock(&set_limit_mutex);
4265 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4266 if (memlimit > val) {
4268 mutex_unlock(&set_limit_mutex);
4271 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4272 if (memswlimit < val)
4274 ret = res_counter_set_limit(&memcg->memsw, val);
4276 if (memlimit == val)
4277 memcg->memsw_is_minimum = true;
4279 memcg->memsw_is_minimum = false;
4281 mutex_unlock(&set_limit_mutex);
4286 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4287 MEM_CGROUP_RECLAIM_NOSWAP |
4288 MEM_CGROUP_RECLAIM_SHRINK);
4289 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4290 /* Usage is reduced ? */
4291 if (curusage >= oldusage)
4294 oldusage = curusage;
4296 if (!ret && enlarge)
4297 memcg_oom_recover(memcg);
4301 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
4303 unsigned long *total_scanned)
4305 unsigned long nr_reclaimed = 0;
4306 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
4307 unsigned long reclaimed;
4309 struct mem_cgroup_tree_per_zone *mctz;
4310 unsigned long long excess;
4311 unsigned long nr_scanned;
4316 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
4318 * This loop can run a while, specially if mem_cgroup's continuously
4319 * keep exceeding their soft limit and putting the system under
4326 mz = mem_cgroup_largest_soft_limit_node(mctz);
4331 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
4332 gfp_mask, &nr_scanned);
4333 nr_reclaimed += reclaimed;
4334 *total_scanned += nr_scanned;
4335 spin_lock(&mctz->lock);
4338 * If we failed to reclaim anything from this memory cgroup
4339 * it is time to move on to the next cgroup
4345 * Loop until we find yet another one.
4347 * By the time we get the soft_limit lock
4348 * again, someone might have aded the
4349 * group back on the RB tree. Iterate to
4350 * make sure we get a different mem.
4351 * mem_cgroup_largest_soft_limit_node returns
4352 * NULL if no other cgroup is present on
4356 __mem_cgroup_largest_soft_limit_node(mctz);
4358 css_put(&next_mz->memcg->css);
4359 else /* next_mz == NULL or other memcg */
4363 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
4364 excess = res_counter_soft_limit_excess(&mz->memcg->res);
4366 * One school of thought says that we should not add
4367 * back the node to the tree if reclaim returns 0.
4368 * But our reclaim could return 0, simply because due
4369 * to priority we are exposing a smaller subset of
4370 * memory to reclaim from. Consider this as a longer
4373 /* If excess == 0, no tree ops */
4374 __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess);
4375 spin_unlock(&mctz->lock);
4376 css_put(&mz->memcg->css);
4379 * Could not reclaim anything and there are no more
4380 * mem cgroups to try or we seem to be looping without
4381 * reclaiming anything.
4383 if (!nr_reclaimed &&
4385 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
4387 } while (!nr_reclaimed);
4389 css_put(&next_mz->memcg->css);
4390 return nr_reclaimed;
4394 * mem_cgroup_force_empty_list - clears LRU of a group
4395 * @memcg: group to clear
4398 * @lru: lru to to clear
4400 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4401 * reclaim the pages page themselves - pages are moved to the parent (or root)
4404 static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
4405 int node, int zid, enum lru_list lru)
4407 struct lruvec *lruvec;
4408 unsigned long flags;
4409 struct list_head *list;
4413 zone = &NODE_DATA(node)->node_zones[zid];
4414 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
4415 list = &lruvec->lists[lru];
4419 struct page_cgroup *pc;
4422 spin_lock_irqsave(&zone->lru_lock, flags);
4423 if (list_empty(list)) {
4424 spin_unlock_irqrestore(&zone->lru_lock, flags);
4427 page = list_entry(list->prev, struct page, lru);
4429 list_move(&page->lru, list);
4431 spin_unlock_irqrestore(&zone->lru_lock, flags);
4434 spin_unlock_irqrestore(&zone->lru_lock, flags);
4436 pc = lookup_page_cgroup(page);
4438 if (mem_cgroup_move_parent(page, pc, memcg)) {
4439 /* found lock contention or "pc" is obsolete. */
4444 } while (!list_empty(list));
4448 * make mem_cgroup's charge to be 0 if there is no task by moving
4449 * all the charges and pages to the parent.
4450 * This enables deleting this mem_cgroup.
4452 * Caller is responsible for holding css reference on the memcg.
4454 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
4460 /* This is for making all *used* pages to be on LRU. */
4461 lru_add_drain_all();
4462 drain_all_stock_sync(memcg);
4463 mem_cgroup_start_move(memcg);
4464 for_each_node_state(node, N_MEMORY) {
4465 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4468 mem_cgroup_force_empty_list(memcg,
4473 mem_cgroup_end_move(memcg);
4474 memcg_oom_recover(memcg);
4478 * Kernel memory may not necessarily be trackable to a specific
4479 * process. So they are not migrated, and therefore we can't
4480 * expect their value to drop to 0 here.
4481 * Having res filled up with kmem only is enough.
4483 * This is a safety check because mem_cgroup_force_empty_list
4484 * could have raced with mem_cgroup_replace_page_cache callers
4485 * so the lru seemed empty but the page could have been added
4486 * right after the check. RES_USAGE should be safe as we always
4487 * charge before adding to the LRU.
4489 usage = res_counter_read_u64(&memcg->res, RES_USAGE) -
4490 res_counter_read_u64(&memcg->kmem, RES_USAGE);
4491 } while (usage > 0);
4495 * Reclaims as many pages from the given memcg as possible and moves
4496 * the rest to the parent.
4498 * Caller is responsible for holding css reference for memcg.
4500 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
4502 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
4503 struct cgroup *cgrp = memcg->css.cgroup;
4505 /* returns EBUSY if there is a task or if we come here twice. */
4506 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
4509 /* we call try-to-free pages for make this cgroup empty */
4510 lru_add_drain_all();
4511 /* try to free all pages in this cgroup */
4512 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
4515 if (signal_pending(current))
4518 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
4522 /* maybe some writeback is necessary */
4523 congestion_wait(BLK_RW_ASYNC, HZ/10);
4528 mem_cgroup_reparent_charges(memcg);
4533 static int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
4535 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4538 if (mem_cgroup_is_root(memcg))
4540 css_get(&memcg->css);
4541 ret = mem_cgroup_force_empty(memcg);
4542 css_put(&memcg->css);
4548 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
4550 return mem_cgroup_from_cont(cont)->use_hierarchy;
4553 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
4557 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4558 struct cgroup *parent = cont->parent;
4559 struct mem_cgroup *parent_memcg = NULL;
4562 parent_memcg = mem_cgroup_from_cont(parent);
4566 if (memcg->use_hierarchy == val)
4570 * If parent's use_hierarchy is set, we can't make any modifications
4571 * in the child subtrees. If it is unset, then the change can
4572 * occur, provided the current cgroup has no children.
4574 * For the root cgroup, parent_mem is NULL, we allow value to be
4575 * set if there are no children.
4577 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
4578 (val == 1 || val == 0)) {
4579 if (list_empty(&cont->children))
4580 memcg->use_hierarchy = val;
4593 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
4594 enum mem_cgroup_stat_index idx)
4596 struct mem_cgroup *iter;
4599 /* Per-cpu values can be negative, use a signed accumulator */
4600 for_each_mem_cgroup_tree(iter, memcg)
4601 val += mem_cgroup_read_stat(iter, idx);
4603 if (val < 0) /* race ? */
4608 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
4612 if (!mem_cgroup_is_root(memcg)) {
4614 return res_counter_read_u64(&memcg->res, RES_USAGE);
4616 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
4619 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
4620 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
4623 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
4625 return val << PAGE_SHIFT;
4628 static ssize_t mem_cgroup_read(struct cgroup *cont, struct cftype *cft,
4629 struct file *file, char __user *buf,
4630 size_t nbytes, loff_t *ppos)
4632 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4638 type = MEMFILE_TYPE(cft->private);
4639 name = MEMFILE_ATTR(cft->private);
4641 if (!do_swap_account && type == _MEMSWAP)
4646 if (name == RES_USAGE)
4647 val = mem_cgroup_usage(memcg, false);
4649 val = res_counter_read_u64(&memcg->res, name);
4652 if (name == RES_USAGE)
4653 val = mem_cgroup_usage(memcg, true);
4655 val = res_counter_read_u64(&memcg->memsw, name);
4658 val = res_counter_read_u64(&memcg->kmem, name);
4664 len = scnprintf(str, sizeof(str), "%llu\n", (unsigned long long)val);
4665 return simple_read_from_buffer(buf, nbytes, ppos, str, len);
4668 static int memcg_update_kmem_limit(struct cgroup *cont, u64 val)
4671 #ifdef CONFIG_MEMCG_KMEM
4672 bool must_inc_static_branch = false;
4674 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4676 * For simplicity, we won't allow this to be disabled. It also can't
4677 * be changed if the cgroup has children already, or if tasks had
4680 * If tasks join before we set the limit, a person looking at
4681 * kmem.usage_in_bytes will have no way to determine when it took
4682 * place, which makes the value quite meaningless.
4684 * After it first became limited, changes in the value of the limit are
4685 * of course permitted.
4687 * Taking the cgroup_lock is really offensive, but it is so far the only
4688 * way to guarantee that no children will appear. There are plenty of
4689 * other offenders, and they should all go away. Fine grained locking
4690 * is probably the way to go here. When we are fully hierarchical, we
4691 * can also get rid of the use_hierarchy check.
4694 mutex_lock(&set_limit_mutex);
4695 if (!memcg->kmem_account_flags && val != RESOURCE_MAX) {
4696 if (cgroup_task_count(cont) || (memcg->use_hierarchy &&
4697 !list_empty(&cont->children))) {
4701 ret = res_counter_set_limit(&memcg->kmem, val);
4704 ret = memcg_update_cache_sizes(memcg);
4706 res_counter_set_limit(&memcg->kmem, RESOURCE_MAX);
4709 must_inc_static_branch = true;
4711 * kmem charges can outlive the cgroup. In the case of slab
4712 * pages, for instance, a page contain objects from various
4713 * processes, so it is unfeasible to migrate them away. We
4714 * need to reference count the memcg because of that.
4716 mem_cgroup_get(memcg);
4718 ret = res_counter_set_limit(&memcg->kmem, val);
4720 mutex_unlock(&set_limit_mutex);
4724 * We are by now familiar with the fact that we can't inc the static
4725 * branch inside cgroup_lock. See disarm functions for details. A
4726 * worker here is overkill, but also wrong: After the limit is set, we
4727 * must start accounting right away. Since this operation can't fail,
4728 * we can safely defer it to here - no rollback will be needed.
4730 * The boolean used to control this is also safe, because
4731 * KMEM_ACCOUNTED_ACTIVATED guarantees that only one process will be
4732 * able to set it to true;
4734 if (must_inc_static_branch) {
4735 static_key_slow_inc(&memcg_kmem_enabled_key);
4737 * setting the active bit after the inc will guarantee no one
4738 * starts accounting before all call sites are patched
4740 memcg_kmem_set_active(memcg);
4747 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
4750 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4754 memcg->kmem_account_flags = parent->kmem_account_flags;
4755 #ifdef CONFIG_MEMCG_KMEM
4757 * When that happen, we need to disable the static branch only on those
4758 * memcgs that enabled it. To achieve this, we would be forced to
4759 * complicate the code by keeping track of which memcgs were the ones
4760 * that actually enabled limits, and which ones got it from its
4763 * It is a lot simpler just to do static_key_slow_inc() on every child
4764 * that is accounted.
4766 if (!memcg_kmem_is_active(memcg))
4770 * destroy(), called if we fail, will issue static_key_slow_inc() and
4771 * mem_cgroup_put() if kmem is enabled. We have to either call them
4772 * unconditionally, or clear the KMEM_ACTIVE flag. I personally find
4773 * this more consistent, since it always leads to the same destroy path
4775 mem_cgroup_get(memcg);
4776 static_key_slow_inc(&memcg_kmem_enabled_key);
4778 mutex_lock(&set_limit_mutex);
4779 ret = memcg_update_cache_sizes(memcg);
4780 mutex_unlock(&set_limit_mutex);
4787 * The user of this function is...
4790 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
4793 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4796 unsigned long long val;
4799 type = MEMFILE_TYPE(cft->private);
4800 name = MEMFILE_ATTR(cft->private);
4802 if (!do_swap_account && type == _MEMSWAP)
4807 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
4811 /* This function does all necessary parse...reuse it */
4812 ret = res_counter_memparse_write_strategy(buffer, &val);
4816 ret = mem_cgroup_resize_limit(memcg, val);
4817 else if (type == _MEMSWAP)
4818 ret = mem_cgroup_resize_memsw_limit(memcg, val);
4819 else if (type == _KMEM)
4820 ret = memcg_update_kmem_limit(cont, val);
4824 case RES_SOFT_LIMIT:
4825 ret = res_counter_memparse_write_strategy(buffer, &val);
4829 * For memsw, soft limits are hard to implement in terms
4830 * of semantics, for now, we support soft limits for
4831 * control without swap
4834 ret = res_counter_set_soft_limit(&memcg->res, val);
4839 ret = -EINVAL; /* should be BUG() ? */
4845 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
4846 unsigned long long *mem_limit, unsigned long long *memsw_limit)
4848 struct cgroup *cgroup;
4849 unsigned long long min_limit, min_memsw_limit, tmp;
4851 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4852 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4853 cgroup = memcg->css.cgroup;
4854 if (!memcg->use_hierarchy)
4857 while (cgroup->parent) {
4858 cgroup = cgroup->parent;
4859 memcg = mem_cgroup_from_cont(cgroup);
4860 if (!memcg->use_hierarchy)
4862 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
4863 min_limit = min(min_limit, tmp);
4864 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4865 min_memsw_limit = min(min_memsw_limit, tmp);
4868 *mem_limit = min_limit;
4869 *memsw_limit = min_memsw_limit;
4872 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
4874 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4878 type = MEMFILE_TYPE(event);
4879 name = MEMFILE_ATTR(event);
4881 if (!do_swap_account && type == _MEMSWAP)
4887 res_counter_reset_max(&memcg->res);
4888 else if (type == _MEMSWAP)
4889 res_counter_reset_max(&memcg->memsw);
4890 else if (type == _KMEM)
4891 res_counter_reset_max(&memcg->kmem);
4897 res_counter_reset_failcnt(&memcg->res);
4898 else if (type == _MEMSWAP)
4899 res_counter_reset_failcnt(&memcg->memsw);
4900 else if (type == _KMEM)
4901 res_counter_reset_failcnt(&memcg->kmem);
4910 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
4913 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
4917 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4918 struct cftype *cft, u64 val)
4920 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4922 if (val >= (1 << NR_MOVE_TYPE))
4925 * We check this value several times in both in can_attach() and
4926 * attach(), so we need cgroup lock to prevent this value from being
4930 memcg->move_charge_at_immigrate = val;
4936 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4937 struct cftype *cft, u64 val)
4944 static int memcg_numa_stat_show(struct cgroup *cont, struct cftype *cft,
4948 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
4949 unsigned long node_nr;
4950 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4952 total_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL);
4953 seq_printf(m, "total=%lu", total_nr);
4954 for_each_node_state(nid, N_MEMORY) {
4955 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL);
4956 seq_printf(m, " N%d=%lu", nid, node_nr);
4960 file_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_FILE);
4961 seq_printf(m, "file=%lu", file_nr);
4962 for_each_node_state(nid, N_MEMORY) {
4963 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4965 seq_printf(m, " N%d=%lu", nid, node_nr);
4969 anon_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_ANON);
4970 seq_printf(m, "anon=%lu", anon_nr);
4971 for_each_node_state(nid, N_MEMORY) {
4972 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4974 seq_printf(m, " N%d=%lu", nid, node_nr);
4978 unevictable_nr = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
4979 seq_printf(m, "unevictable=%lu", unevictable_nr);
4980 for_each_node_state(nid, N_MEMORY) {
4981 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4982 BIT(LRU_UNEVICTABLE));
4983 seq_printf(m, " N%d=%lu", nid, node_nr);
4988 #endif /* CONFIG_NUMA */
4990 static const char * const mem_cgroup_lru_names[] = {
4998 static inline void mem_cgroup_lru_names_not_uptodate(void)
5000 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
5003 static int memcg_stat_show(struct cgroup *cont, struct cftype *cft,
5006 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5007 struct mem_cgroup *mi;
5010 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5011 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5013 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
5014 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
5017 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
5018 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
5019 mem_cgroup_read_events(memcg, i));
5021 for (i = 0; i < NR_LRU_LISTS; i++)
5022 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
5023 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
5025 /* Hierarchical information */
5027 unsigned long long limit, memsw_limit;
5028 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
5029 seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
5030 if (do_swap_account)
5031 seq_printf(m, "hierarchical_memsw_limit %llu\n",
5035 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5038 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5040 for_each_mem_cgroup_tree(mi, memcg)
5041 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
5042 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
5045 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
5046 unsigned long long val = 0;
5048 for_each_mem_cgroup_tree(mi, memcg)
5049 val += mem_cgroup_read_events(mi, i);
5050 seq_printf(m, "total_%s %llu\n",
5051 mem_cgroup_events_names[i], val);
5054 for (i = 0; i < NR_LRU_LISTS; i++) {
5055 unsigned long long val = 0;
5057 for_each_mem_cgroup_tree(mi, memcg)
5058 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
5059 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
5062 #ifdef CONFIG_DEBUG_VM
5065 struct mem_cgroup_per_zone *mz;
5066 struct zone_reclaim_stat *rstat;
5067 unsigned long recent_rotated[2] = {0, 0};
5068 unsigned long recent_scanned[2] = {0, 0};
5070 for_each_online_node(nid)
5071 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
5072 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
5073 rstat = &mz->lruvec.reclaim_stat;
5075 recent_rotated[0] += rstat->recent_rotated[0];
5076 recent_rotated[1] += rstat->recent_rotated[1];
5077 recent_scanned[0] += rstat->recent_scanned[0];
5078 recent_scanned[1] += rstat->recent_scanned[1];
5080 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
5081 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
5082 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
5083 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
5090 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
5092 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5094 return mem_cgroup_swappiness(memcg);
5097 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
5100 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5101 struct mem_cgroup *parent;
5106 if (cgrp->parent == NULL)
5109 parent = mem_cgroup_from_cont(cgrp->parent);
5113 /* If under hierarchy, only empty-root can set this value */
5114 if ((parent->use_hierarchy) ||
5115 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
5120 memcg->swappiness = val;
5127 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
5129 struct mem_cgroup_threshold_ary *t;
5135 t = rcu_dereference(memcg->thresholds.primary);
5137 t = rcu_dereference(memcg->memsw_thresholds.primary);
5142 usage = mem_cgroup_usage(memcg, swap);
5145 * current_threshold points to threshold just below or equal to usage.
5146 * If it's not true, a threshold was crossed after last
5147 * call of __mem_cgroup_threshold().
5149 i = t->current_threshold;
5152 * Iterate backward over array of thresholds starting from
5153 * current_threshold and check if a threshold is crossed.
5154 * If none of thresholds below usage is crossed, we read
5155 * only one element of the array here.
5157 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
5158 eventfd_signal(t->entries[i].eventfd, 1);
5160 /* i = current_threshold + 1 */
5164 * Iterate forward over array of thresholds starting from
5165 * current_threshold+1 and check if a threshold is crossed.
5166 * If none of thresholds above usage is crossed, we read
5167 * only one element of the array here.
5169 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
5170 eventfd_signal(t->entries[i].eventfd, 1);
5172 /* Update current_threshold */
5173 t->current_threshold = i - 1;
5178 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
5181 __mem_cgroup_threshold(memcg, false);
5182 if (do_swap_account)
5183 __mem_cgroup_threshold(memcg, true);
5185 memcg = parent_mem_cgroup(memcg);
5189 static int compare_thresholds(const void *a, const void *b)
5191 const struct mem_cgroup_threshold *_a = a;
5192 const struct mem_cgroup_threshold *_b = b;
5194 return _a->threshold - _b->threshold;
5197 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
5199 struct mem_cgroup_eventfd_list *ev;
5201 list_for_each_entry(ev, &memcg->oom_notify, list)
5202 eventfd_signal(ev->eventfd, 1);
5206 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
5208 struct mem_cgroup *iter;
5210 for_each_mem_cgroup_tree(iter, memcg)
5211 mem_cgroup_oom_notify_cb(iter);
5214 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
5215 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5217 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5218 struct mem_cgroup_thresholds *thresholds;
5219 struct mem_cgroup_threshold_ary *new;
5220 enum res_type type = MEMFILE_TYPE(cft->private);
5221 u64 threshold, usage;
5224 ret = res_counter_memparse_write_strategy(args, &threshold);
5228 mutex_lock(&memcg->thresholds_lock);
5231 thresholds = &memcg->thresholds;
5232 else if (type == _MEMSWAP)
5233 thresholds = &memcg->memsw_thresholds;
5237 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5239 /* Check if a threshold crossed before adding a new one */
5240 if (thresholds->primary)
5241 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5243 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
5245 /* Allocate memory for new array of thresholds */
5246 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
5254 /* Copy thresholds (if any) to new array */
5255 if (thresholds->primary) {
5256 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
5257 sizeof(struct mem_cgroup_threshold));
5260 /* Add new threshold */
5261 new->entries[size - 1].eventfd = eventfd;
5262 new->entries[size - 1].threshold = threshold;
5264 /* Sort thresholds. Registering of new threshold isn't time-critical */
5265 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
5266 compare_thresholds, NULL);
5268 /* Find current threshold */
5269 new->current_threshold = -1;
5270 for (i = 0; i < size; i++) {
5271 if (new->entries[i].threshold <= usage) {
5273 * new->current_threshold will not be used until
5274 * rcu_assign_pointer(), so it's safe to increment
5277 ++new->current_threshold;
5282 /* Free old spare buffer and save old primary buffer as spare */
5283 kfree(thresholds->spare);
5284 thresholds->spare = thresholds->primary;
5286 rcu_assign_pointer(thresholds->primary, new);
5288 /* To be sure that nobody uses thresholds */
5292 mutex_unlock(&memcg->thresholds_lock);
5297 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
5298 struct cftype *cft, struct eventfd_ctx *eventfd)
5300 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5301 struct mem_cgroup_thresholds *thresholds;
5302 struct mem_cgroup_threshold_ary *new;
5303 enum res_type type = MEMFILE_TYPE(cft->private);
5307 mutex_lock(&memcg->thresholds_lock);
5309 thresholds = &memcg->thresholds;
5310 else if (type == _MEMSWAP)
5311 thresholds = &memcg->memsw_thresholds;
5315 if (!thresholds->primary)
5318 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5320 /* Check if a threshold crossed before removing */
5321 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5323 /* Calculate new number of threshold */
5325 for (i = 0; i < thresholds->primary->size; i++) {
5326 if (thresholds->primary->entries[i].eventfd != eventfd)
5330 new = thresholds->spare;
5332 /* Set thresholds array to NULL if we don't have thresholds */
5341 /* Copy thresholds and find current threshold */
5342 new->current_threshold = -1;
5343 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
5344 if (thresholds->primary->entries[i].eventfd == eventfd)
5347 new->entries[j] = thresholds->primary->entries[i];
5348 if (new->entries[j].threshold <= usage) {
5350 * new->current_threshold will not be used
5351 * until rcu_assign_pointer(), so it's safe to increment
5354 ++new->current_threshold;
5360 /* Swap primary and spare array */
5361 thresholds->spare = thresholds->primary;
5362 /* If all events are unregistered, free the spare array */
5364 kfree(thresholds->spare);
5365 thresholds->spare = NULL;
5368 rcu_assign_pointer(thresholds->primary, new);
5370 /* To be sure that nobody uses thresholds */
5373 mutex_unlock(&memcg->thresholds_lock);
5376 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
5377 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5379 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5380 struct mem_cgroup_eventfd_list *event;
5381 enum res_type type = MEMFILE_TYPE(cft->private);
5383 BUG_ON(type != _OOM_TYPE);
5384 event = kmalloc(sizeof(*event), GFP_KERNEL);
5388 spin_lock(&memcg_oom_lock);
5390 event->eventfd = eventfd;
5391 list_add(&event->list, &memcg->oom_notify);
5393 /* already in OOM ? */
5394 if (atomic_read(&memcg->under_oom))
5395 eventfd_signal(eventfd, 1);
5396 spin_unlock(&memcg_oom_lock);
5401 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
5402 struct cftype *cft, struct eventfd_ctx *eventfd)
5404 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5405 struct mem_cgroup_eventfd_list *ev, *tmp;
5406 enum res_type type = MEMFILE_TYPE(cft->private);
5408 BUG_ON(type != _OOM_TYPE);
5410 spin_lock(&memcg_oom_lock);
5412 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
5413 if (ev->eventfd == eventfd) {
5414 list_del(&ev->list);
5419 spin_unlock(&memcg_oom_lock);
5422 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
5423 struct cftype *cft, struct cgroup_map_cb *cb)
5425 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5427 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
5429 if (atomic_read(&memcg->under_oom))
5430 cb->fill(cb, "under_oom", 1);
5432 cb->fill(cb, "under_oom", 0);
5436 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
5437 struct cftype *cft, u64 val)
5439 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5440 struct mem_cgroup *parent;
5442 /* cannot set to root cgroup and only 0 and 1 are allowed */
5443 if (!cgrp->parent || !((val == 0) || (val == 1)))
5446 parent = mem_cgroup_from_cont(cgrp->parent);
5449 /* oom-kill-disable is a flag for subhierarchy. */
5450 if ((parent->use_hierarchy) ||
5451 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
5455 memcg->oom_kill_disable = val;
5457 memcg_oom_recover(memcg);
5462 #ifdef CONFIG_MEMCG_KMEM
5463 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5467 memcg->kmemcg_id = -1;
5468 ret = memcg_propagate_kmem(memcg);
5472 return mem_cgroup_sockets_init(memcg, ss);
5475 static void kmem_cgroup_destroy(struct mem_cgroup *memcg)
5477 mem_cgroup_sockets_destroy(memcg);
5479 memcg_kmem_mark_dead(memcg);
5481 if (res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0)
5485 * Charges already down to 0, undo mem_cgroup_get() done in the charge
5486 * path here, being careful not to race with memcg_uncharge_kmem: it is
5487 * possible that the charges went down to 0 between mark_dead and the
5488 * res_counter read, so in that case, we don't need the put
5490 if (memcg_kmem_test_and_clear_dead(memcg))
5491 mem_cgroup_put(memcg);
5494 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5499 static void kmem_cgroup_destroy(struct mem_cgroup *memcg)
5504 static struct cftype mem_cgroup_files[] = {
5506 .name = "usage_in_bytes",
5507 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
5508 .read = mem_cgroup_read,
5509 .register_event = mem_cgroup_usage_register_event,
5510 .unregister_event = mem_cgroup_usage_unregister_event,
5513 .name = "max_usage_in_bytes",
5514 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
5515 .trigger = mem_cgroup_reset,
5516 .read = mem_cgroup_read,
5519 .name = "limit_in_bytes",
5520 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
5521 .write_string = mem_cgroup_write,
5522 .read = mem_cgroup_read,
5525 .name = "soft_limit_in_bytes",
5526 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
5527 .write_string = mem_cgroup_write,
5528 .read = mem_cgroup_read,
5532 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
5533 .trigger = mem_cgroup_reset,
5534 .read = mem_cgroup_read,
5538 .read_seq_string = memcg_stat_show,
5541 .name = "force_empty",
5542 .trigger = mem_cgroup_force_empty_write,
5545 .name = "use_hierarchy",
5546 .write_u64 = mem_cgroup_hierarchy_write,
5547 .read_u64 = mem_cgroup_hierarchy_read,
5550 .name = "swappiness",
5551 .read_u64 = mem_cgroup_swappiness_read,
5552 .write_u64 = mem_cgroup_swappiness_write,
5555 .name = "move_charge_at_immigrate",
5556 .read_u64 = mem_cgroup_move_charge_read,
5557 .write_u64 = mem_cgroup_move_charge_write,
5560 .name = "oom_control",
5561 .read_map = mem_cgroup_oom_control_read,
5562 .write_u64 = mem_cgroup_oom_control_write,
5563 .register_event = mem_cgroup_oom_register_event,
5564 .unregister_event = mem_cgroup_oom_unregister_event,
5565 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
5569 .name = "numa_stat",
5570 .read_seq_string = memcg_numa_stat_show,
5573 #ifdef CONFIG_MEMCG_SWAP
5575 .name = "memsw.usage_in_bytes",
5576 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
5577 .read = mem_cgroup_read,
5578 .register_event = mem_cgroup_usage_register_event,
5579 .unregister_event = mem_cgroup_usage_unregister_event,
5582 .name = "memsw.max_usage_in_bytes",
5583 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
5584 .trigger = mem_cgroup_reset,
5585 .read = mem_cgroup_read,
5588 .name = "memsw.limit_in_bytes",
5589 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
5590 .write_string = mem_cgroup_write,
5591 .read = mem_cgroup_read,
5594 .name = "memsw.failcnt",
5595 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
5596 .trigger = mem_cgroup_reset,
5597 .read = mem_cgroup_read,
5600 #ifdef CONFIG_MEMCG_KMEM
5602 .name = "kmem.limit_in_bytes",
5603 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
5604 .write_string = mem_cgroup_write,
5605 .read = mem_cgroup_read,
5608 .name = "kmem.usage_in_bytes",
5609 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5610 .read = mem_cgroup_read,
5613 .name = "kmem.failcnt",
5614 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5615 .trigger = mem_cgroup_reset,
5616 .read = mem_cgroup_read,
5619 .name = "kmem.max_usage_in_bytes",
5620 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5621 .trigger = mem_cgroup_reset,
5622 .read = mem_cgroup_read,
5625 { }, /* terminate */
5628 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
5630 struct mem_cgroup_per_node *pn;
5631 struct mem_cgroup_per_zone *mz;
5632 int zone, tmp = node;
5634 * This routine is called against possible nodes.
5635 * But it's BUG to call kmalloc() against offline node.
5637 * TODO: this routine can waste much memory for nodes which will
5638 * never be onlined. It's better to use memory hotplug callback
5641 if (!node_state(node, N_NORMAL_MEMORY))
5643 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5647 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5648 mz = &pn->zoneinfo[zone];
5649 lruvec_init(&mz->lruvec);
5650 mz->usage_in_excess = 0;
5651 mz->on_tree = false;
5654 memcg->info.nodeinfo[node] = pn;
5658 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
5660 kfree(memcg->info.nodeinfo[node]);
5663 static struct mem_cgroup *mem_cgroup_alloc(void)
5665 struct mem_cgroup *memcg;
5666 int size = sizeof(struct mem_cgroup);
5668 /* Can be very big if MAX_NUMNODES is very big */
5669 if (size < PAGE_SIZE)
5670 memcg = kzalloc(size, GFP_KERNEL);
5672 memcg = vzalloc(size);
5677 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
5680 spin_lock_init(&memcg->pcp_counter_lock);
5684 if (size < PAGE_SIZE)
5692 * At destroying mem_cgroup, references from swap_cgroup can remain.
5693 * (scanning all at force_empty is too costly...)
5695 * Instead of clearing all references at force_empty, we remember
5696 * the number of reference from swap_cgroup and free mem_cgroup when
5697 * it goes down to 0.
5699 * Removal of cgroup itself succeeds regardless of refs from swap.
5702 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5705 int size = sizeof(struct mem_cgroup);
5707 mem_cgroup_remove_from_trees(memcg);
5708 free_css_id(&mem_cgroup_subsys, &memcg->css);
5711 free_mem_cgroup_per_zone_info(memcg, node);
5713 free_percpu(memcg->stat);
5716 * We need to make sure that (at least for now), the jump label
5717 * destruction code runs outside of the cgroup lock. This is because
5718 * get_online_cpus(), which is called from the static_branch update,
5719 * can't be called inside the cgroup_lock. cpusets are the ones
5720 * enforcing this dependency, so if they ever change, we might as well.
5722 * schedule_work() will guarantee this happens. Be careful if you need
5723 * to move this code around, and make sure it is outside
5726 disarm_static_keys(memcg);
5727 if (size < PAGE_SIZE)
5735 * Helpers for freeing a kmalloc()ed/vzalloc()ed mem_cgroup by RCU,
5736 * but in process context. The work_freeing structure is overlaid
5737 * on the rcu_freeing structure, which itself is overlaid on memsw.
5739 static void free_work(struct work_struct *work)
5741 struct mem_cgroup *memcg;
5743 memcg = container_of(work, struct mem_cgroup, work_freeing);
5744 __mem_cgroup_free(memcg);
5747 static void free_rcu(struct rcu_head *rcu_head)
5749 struct mem_cgroup *memcg;
5751 memcg = container_of(rcu_head, struct mem_cgroup, rcu_freeing);
5752 INIT_WORK(&memcg->work_freeing, free_work);
5753 schedule_work(&memcg->work_freeing);
5756 static void mem_cgroup_get(struct mem_cgroup *memcg)
5758 atomic_inc(&memcg->refcnt);
5761 static void __mem_cgroup_put(struct mem_cgroup *memcg, int count)
5763 if (atomic_sub_and_test(count, &memcg->refcnt)) {
5764 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5765 call_rcu(&memcg->rcu_freeing, free_rcu);
5767 mem_cgroup_put(parent);
5771 static void mem_cgroup_put(struct mem_cgroup *memcg)
5773 __mem_cgroup_put(memcg, 1);
5777 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
5779 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
5781 if (!memcg->res.parent)
5783 return mem_cgroup_from_res_counter(memcg->res.parent, res);
5785 EXPORT_SYMBOL(parent_mem_cgroup);
5787 #ifdef CONFIG_MEMCG_SWAP
5788 static void __init enable_swap_cgroup(void)
5790 if (!mem_cgroup_disabled() && really_do_swap_account)
5791 do_swap_account = 1;
5794 static void __init enable_swap_cgroup(void)
5799 static int mem_cgroup_soft_limit_tree_init(void)
5801 struct mem_cgroup_tree_per_node *rtpn;
5802 struct mem_cgroup_tree_per_zone *rtpz;
5803 int tmp, node, zone;
5805 for_each_node(node) {
5807 if (!node_state(node, N_NORMAL_MEMORY))
5809 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
5813 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5815 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5816 rtpz = &rtpn->rb_tree_per_zone[zone];
5817 rtpz->rb_root = RB_ROOT;
5818 spin_lock_init(&rtpz->lock);
5824 for_each_node(node) {
5825 if (!soft_limit_tree.rb_tree_per_node[node])
5827 kfree(soft_limit_tree.rb_tree_per_node[node]);
5828 soft_limit_tree.rb_tree_per_node[node] = NULL;
5834 static struct cgroup_subsys_state * __ref
5835 mem_cgroup_css_alloc(struct cgroup *cont)
5837 struct mem_cgroup *memcg, *parent;
5838 long error = -ENOMEM;
5841 memcg = mem_cgroup_alloc();
5843 return ERR_PTR(error);
5846 if (alloc_mem_cgroup_per_zone_info(memcg, node))
5850 if (cont->parent == NULL) {
5852 enable_swap_cgroup();
5854 if (mem_cgroup_soft_limit_tree_init())
5856 root_mem_cgroup = memcg;
5857 for_each_possible_cpu(cpu) {
5858 struct memcg_stock_pcp *stock =
5859 &per_cpu(memcg_stock, cpu);
5860 INIT_WORK(&stock->work, drain_local_stock);
5862 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5864 parent = mem_cgroup_from_cont(cont->parent);
5865 memcg->use_hierarchy = parent->use_hierarchy;
5866 memcg->oom_kill_disable = parent->oom_kill_disable;
5869 if (parent && parent->use_hierarchy) {
5870 res_counter_init(&memcg->res, &parent->res);
5871 res_counter_init(&memcg->memsw, &parent->memsw);
5872 res_counter_init(&memcg->kmem, &parent->kmem);
5875 * We increment refcnt of the parent to ensure that we can
5876 * safely access it on res_counter_charge/uncharge.
5877 * This refcnt will be decremented when freeing this
5878 * mem_cgroup(see mem_cgroup_put).
5880 mem_cgroup_get(parent);
5882 res_counter_init(&memcg->res, NULL);
5883 res_counter_init(&memcg->memsw, NULL);
5884 res_counter_init(&memcg->kmem, NULL);
5886 * Deeper hierachy with use_hierarchy == false doesn't make
5887 * much sense so let cgroup subsystem know about this
5888 * unfortunate state in our controller.
5890 if (parent && parent != root_mem_cgroup)
5891 mem_cgroup_subsys.broken_hierarchy = true;
5893 memcg->last_scanned_node = MAX_NUMNODES;
5894 INIT_LIST_HEAD(&memcg->oom_notify);
5897 memcg->swappiness = mem_cgroup_swappiness(parent);
5898 atomic_set(&memcg->refcnt, 1);
5899 memcg->move_charge_at_immigrate = 0;
5900 mutex_init(&memcg->thresholds_lock);
5901 spin_lock_init(&memcg->move_lock);
5903 error = memcg_init_kmem(memcg, &mem_cgroup_subsys);
5906 * We call put now because our (and parent's) refcnts
5907 * are already in place. mem_cgroup_put() will internally
5908 * call __mem_cgroup_free, so return directly
5910 mem_cgroup_put(memcg);
5911 return ERR_PTR(error);
5915 __mem_cgroup_free(memcg);
5916 return ERR_PTR(error);
5919 static void mem_cgroup_css_offline(struct cgroup *cont)
5921 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5923 mem_cgroup_reparent_charges(memcg);
5926 static void mem_cgroup_css_free(struct cgroup *cont)
5928 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5930 kmem_cgroup_destroy(memcg);
5932 mem_cgroup_put(memcg);
5936 /* Handlers for move charge at task migration. */
5937 #define PRECHARGE_COUNT_AT_ONCE 256
5938 static int mem_cgroup_do_precharge(unsigned long count)
5941 int batch_count = PRECHARGE_COUNT_AT_ONCE;
5942 struct mem_cgroup *memcg = mc.to;
5944 if (mem_cgroup_is_root(memcg)) {
5945 mc.precharge += count;
5946 /* we don't need css_get for root */
5949 /* try to charge at once */
5951 struct res_counter *dummy;
5953 * "memcg" cannot be under rmdir() because we've already checked
5954 * by cgroup_lock_live_cgroup() that it is not removed and we
5955 * are still under the same cgroup_mutex. So we can postpone
5958 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
5960 if (do_swap_account && res_counter_charge(&memcg->memsw,
5961 PAGE_SIZE * count, &dummy)) {
5962 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
5965 mc.precharge += count;
5969 /* fall back to one by one charge */
5971 if (signal_pending(current)) {
5975 if (!batch_count--) {
5976 batch_count = PRECHARGE_COUNT_AT_ONCE;
5979 ret = __mem_cgroup_try_charge(NULL,
5980 GFP_KERNEL, 1, &memcg, false);
5982 /* mem_cgroup_clear_mc() will do uncharge later */
5990 * get_mctgt_type - get target type of moving charge
5991 * @vma: the vma the pte to be checked belongs
5992 * @addr: the address corresponding to the pte to be checked
5993 * @ptent: the pte to be checked
5994 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5997 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5998 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5999 * move charge. if @target is not NULL, the page is stored in target->page
6000 * with extra refcnt got(Callers should handle it).
6001 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6002 * target for charge migration. if @target is not NULL, the entry is stored
6005 * Called with pte lock held.
6012 enum mc_target_type {
6018 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
6019 unsigned long addr, pte_t ptent)
6021 struct page *page = vm_normal_page(vma, addr, ptent);
6023 if (!page || !page_mapped(page))
6025 if (PageAnon(page)) {
6026 /* we don't move shared anon */
6029 } else if (!move_file())
6030 /* we ignore mapcount for file pages */
6032 if (!get_page_unless_zero(page))
6039 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6040 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6042 struct page *page = NULL;
6043 swp_entry_t ent = pte_to_swp_entry(ptent);
6045 if (!move_anon() || non_swap_entry(ent))
6048 * Because lookup_swap_cache() updates some statistics counter,
6049 * we call find_get_page() with swapper_space directly.
6051 page = find_get_page(&swapper_space, ent.val);
6052 if (do_swap_account)
6053 entry->val = ent.val;
6058 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6059 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6065 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
6066 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6068 struct page *page = NULL;
6069 struct address_space *mapping;
6072 if (!vma->vm_file) /* anonymous vma */
6077 mapping = vma->vm_file->f_mapping;
6078 if (pte_none(ptent))
6079 pgoff = linear_page_index(vma, addr);
6080 else /* pte_file(ptent) is true */
6081 pgoff = pte_to_pgoff(ptent);
6083 /* page is moved even if it's not RSS of this task(page-faulted). */
6084 page = find_get_page(mapping, pgoff);
6087 /* shmem/tmpfs may report page out on swap: account for that too. */
6088 if (radix_tree_exceptional_entry(page)) {
6089 swp_entry_t swap = radix_to_swp_entry(page);
6090 if (do_swap_account)
6092 page = find_get_page(&swapper_space, swap.val);
6098 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
6099 unsigned long addr, pte_t ptent, union mc_target *target)
6101 struct page *page = NULL;
6102 struct page_cgroup *pc;
6103 enum mc_target_type ret = MC_TARGET_NONE;
6104 swp_entry_t ent = { .val = 0 };
6106 if (pte_present(ptent))
6107 page = mc_handle_present_pte(vma, addr, ptent);
6108 else if (is_swap_pte(ptent))
6109 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
6110 else if (pte_none(ptent) || pte_file(ptent))
6111 page = mc_handle_file_pte(vma, addr, ptent, &ent);
6113 if (!page && !ent.val)
6116 pc = lookup_page_cgroup(page);
6118 * Do only loose check w/o page_cgroup lock.
6119 * mem_cgroup_move_account() checks the pc is valid or not under
6122 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6123 ret = MC_TARGET_PAGE;
6125 target->page = page;
6127 if (!ret || !target)
6130 /* There is a swap entry and a page doesn't exist or isn't charged */
6131 if (ent.val && !ret &&
6132 css_id(&mc.from->css) == lookup_swap_cgroup_id(ent)) {
6133 ret = MC_TARGET_SWAP;
6140 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6142 * We don't consider swapping or file mapped pages because THP does not
6143 * support them for now.
6144 * Caller should make sure that pmd_trans_huge(pmd) is true.
6146 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6147 unsigned long addr, pmd_t pmd, union mc_target *target)
6149 struct page *page = NULL;
6150 struct page_cgroup *pc;
6151 enum mc_target_type ret = MC_TARGET_NONE;
6153 page = pmd_page(pmd);
6154 VM_BUG_ON(!page || !PageHead(page));
6157 pc = lookup_page_cgroup(page);
6158 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6159 ret = MC_TARGET_PAGE;
6162 target->page = page;
6168 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6169 unsigned long addr, pmd_t pmd, union mc_target *target)
6171 return MC_TARGET_NONE;
6175 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
6176 unsigned long addr, unsigned long end,
6177 struct mm_walk *walk)
6179 struct vm_area_struct *vma = walk->private;
6183 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6184 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
6185 mc.precharge += HPAGE_PMD_NR;
6186 spin_unlock(&vma->vm_mm->page_table_lock);
6190 if (pmd_trans_unstable(pmd))
6192 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6193 for (; addr != end; pte++, addr += PAGE_SIZE)
6194 if (get_mctgt_type(vma, addr, *pte, NULL))
6195 mc.precharge++; /* increment precharge temporarily */
6196 pte_unmap_unlock(pte - 1, ptl);
6202 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
6204 unsigned long precharge;
6205 struct vm_area_struct *vma;
6207 down_read(&mm->mmap_sem);
6208 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6209 struct mm_walk mem_cgroup_count_precharge_walk = {
6210 .pmd_entry = mem_cgroup_count_precharge_pte_range,
6214 if (is_vm_hugetlb_page(vma))
6216 walk_page_range(vma->vm_start, vma->vm_end,
6217 &mem_cgroup_count_precharge_walk);
6219 up_read(&mm->mmap_sem);
6221 precharge = mc.precharge;
6227 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
6229 unsigned long precharge = mem_cgroup_count_precharge(mm);
6231 VM_BUG_ON(mc.moving_task);
6232 mc.moving_task = current;
6233 return mem_cgroup_do_precharge(precharge);
6236 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6237 static void __mem_cgroup_clear_mc(void)
6239 struct mem_cgroup *from = mc.from;
6240 struct mem_cgroup *to = mc.to;
6242 /* we must uncharge all the leftover precharges from mc.to */
6244 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
6248 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6249 * we must uncharge here.
6251 if (mc.moved_charge) {
6252 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
6253 mc.moved_charge = 0;
6255 /* we must fixup refcnts and charges */
6256 if (mc.moved_swap) {
6257 /* uncharge swap account from the old cgroup */
6258 if (!mem_cgroup_is_root(mc.from))
6259 res_counter_uncharge(&mc.from->memsw,
6260 PAGE_SIZE * mc.moved_swap);
6261 __mem_cgroup_put(mc.from, mc.moved_swap);
6263 if (!mem_cgroup_is_root(mc.to)) {
6265 * we charged both to->res and to->memsw, so we should
6268 res_counter_uncharge(&mc.to->res,
6269 PAGE_SIZE * mc.moved_swap);
6271 /* we've already done mem_cgroup_get(mc.to) */
6274 memcg_oom_recover(from);
6275 memcg_oom_recover(to);
6276 wake_up_all(&mc.waitq);
6279 static void mem_cgroup_clear_mc(void)
6281 struct mem_cgroup *from = mc.from;
6284 * we must clear moving_task before waking up waiters at the end of
6287 mc.moving_task = NULL;
6288 __mem_cgroup_clear_mc();
6289 spin_lock(&mc.lock);
6292 spin_unlock(&mc.lock);
6293 mem_cgroup_end_move(from);
6296 static int mem_cgroup_can_attach(struct cgroup *cgroup,
6297 struct cgroup_taskset *tset)
6299 struct task_struct *p = cgroup_taskset_first(tset);
6301 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup);
6303 if (memcg->move_charge_at_immigrate) {
6304 struct mm_struct *mm;
6305 struct mem_cgroup *from = mem_cgroup_from_task(p);
6307 VM_BUG_ON(from == memcg);
6309 mm = get_task_mm(p);
6312 /* We move charges only when we move a owner of the mm */
6313 if (mm->owner == p) {
6316 VM_BUG_ON(mc.precharge);
6317 VM_BUG_ON(mc.moved_charge);
6318 VM_BUG_ON(mc.moved_swap);
6319 mem_cgroup_start_move(from);
6320 spin_lock(&mc.lock);
6323 spin_unlock(&mc.lock);
6324 /* We set mc.moving_task later */
6326 ret = mem_cgroup_precharge_mc(mm);
6328 mem_cgroup_clear_mc();
6335 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
6336 struct cgroup_taskset *tset)
6338 mem_cgroup_clear_mc();
6341 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6342 unsigned long addr, unsigned long end,
6343 struct mm_walk *walk)
6346 struct vm_area_struct *vma = walk->private;
6349 enum mc_target_type target_type;
6350 union mc_target target;
6352 struct page_cgroup *pc;
6355 * We don't take compound_lock() here but no race with splitting thp
6357 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6358 * under splitting, which means there's no concurrent thp split,
6359 * - if another thread runs into split_huge_page() just after we
6360 * entered this if-block, the thread must wait for page table lock
6361 * to be unlocked in __split_huge_page_splitting(), where the main
6362 * part of thp split is not executed yet.
6364 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6365 if (mc.precharge < HPAGE_PMD_NR) {
6366 spin_unlock(&vma->vm_mm->page_table_lock);
6369 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6370 if (target_type == MC_TARGET_PAGE) {
6372 if (!isolate_lru_page(page)) {
6373 pc = lookup_page_cgroup(page);
6374 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
6375 pc, mc.from, mc.to)) {
6376 mc.precharge -= HPAGE_PMD_NR;
6377 mc.moved_charge += HPAGE_PMD_NR;
6379 putback_lru_page(page);
6383 spin_unlock(&vma->vm_mm->page_table_lock);
6387 if (pmd_trans_unstable(pmd))
6390 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6391 for (; addr != end; addr += PAGE_SIZE) {
6392 pte_t ptent = *(pte++);
6398 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6399 case MC_TARGET_PAGE:
6401 if (isolate_lru_page(page))
6403 pc = lookup_page_cgroup(page);
6404 if (!mem_cgroup_move_account(page, 1, pc,
6407 /* we uncharge from mc.from later. */
6410 putback_lru_page(page);
6411 put: /* get_mctgt_type() gets the page */
6414 case MC_TARGET_SWAP:
6416 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6418 /* we fixup refcnts and charges later. */
6426 pte_unmap_unlock(pte - 1, ptl);
6431 * We have consumed all precharges we got in can_attach().
6432 * We try charge one by one, but don't do any additional
6433 * charges to mc.to if we have failed in charge once in attach()
6436 ret = mem_cgroup_do_precharge(1);
6444 static void mem_cgroup_move_charge(struct mm_struct *mm)
6446 struct vm_area_struct *vma;
6448 lru_add_drain_all();
6450 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
6452 * Someone who are holding the mmap_sem might be waiting in
6453 * waitq. So we cancel all extra charges, wake up all waiters,
6454 * and retry. Because we cancel precharges, we might not be able
6455 * to move enough charges, but moving charge is a best-effort
6456 * feature anyway, so it wouldn't be a big problem.
6458 __mem_cgroup_clear_mc();
6462 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6464 struct mm_walk mem_cgroup_move_charge_walk = {
6465 .pmd_entry = mem_cgroup_move_charge_pte_range,
6469 if (is_vm_hugetlb_page(vma))
6471 ret = walk_page_range(vma->vm_start, vma->vm_end,
6472 &mem_cgroup_move_charge_walk);
6475 * means we have consumed all precharges and failed in
6476 * doing additional charge. Just abandon here.
6480 up_read(&mm->mmap_sem);
6483 static void mem_cgroup_move_task(struct cgroup *cont,
6484 struct cgroup_taskset *tset)
6486 struct task_struct *p = cgroup_taskset_first(tset);
6487 struct mm_struct *mm = get_task_mm(p);
6491 mem_cgroup_move_charge(mm);
6495 mem_cgroup_clear_mc();
6497 #else /* !CONFIG_MMU */
6498 static int mem_cgroup_can_attach(struct cgroup *cgroup,
6499 struct cgroup_taskset *tset)
6503 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
6504 struct cgroup_taskset *tset)
6507 static void mem_cgroup_move_task(struct cgroup *cont,
6508 struct cgroup_taskset *tset)
6513 struct cgroup_subsys mem_cgroup_subsys = {
6515 .subsys_id = mem_cgroup_subsys_id,
6516 .css_alloc = mem_cgroup_css_alloc,
6517 .css_offline = mem_cgroup_css_offline,
6518 .css_free = mem_cgroup_css_free,
6519 .can_attach = mem_cgroup_can_attach,
6520 .cancel_attach = mem_cgroup_cancel_attach,
6521 .attach = mem_cgroup_move_task,
6522 .base_cftypes = mem_cgroup_files,
6527 #ifdef CONFIG_MEMCG_SWAP
6528 static int __init enable_swap_account(char *s)
6530 /* consider enabled if no parameter or 1 is given */
6531 if (!strcmp(s, "1"))
6532 really_do_swap_account = 1;
6533 else if (!strcmp(s, "0"))
6534 really_do_swap_account = 0;
6537 __setup("swapaccount=", enable_swap_account);