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[] = {
123 static const char * const mem_cgroup_lru_names[] = {
132 * Per memcg event counter is incremented at every pagein/pageout. With THP,
133 * it will be incremated by the number of pages. This counter is used for
134 * for trigger some periodic events. This is straightforward and better
135 * than using jiffies etc. to handle periodic memcg event.
137 enum mem_cgroup_events_target {
138 MEM_CGROUP_TARGET_THRESH,
139 MEM_CGROUP_TARGET_SOFTLIMIT,
140 MEM_CGROUP_TARGET_NUMAINFO,
143 #define THRESHOLDS_EVENTS_TARGET 128
144 #define SOFTLIMIT_EVENTS_TARGET 1024
145 #define NUMAINFO_EVENTS_TARGET 1024
147 struct mem_cgroup_stat_cpu {
148 long count[MEM_CGROUP_STAT_NSTATS];
149 unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
150 unsigned long nr_page_events;
151 unsigned long targets[MEM_CGROUP_NTARGETS];
154 struct mem_cgroup_reclaim_iter {
156 * last scanned hierarchy member. Valid only if last_dead_count
157 * matches memcg->dead_count of the hierarchy root group.
159 struct mem_cgroup *last_visited;
160 unsigned long last_dead_count;
162 /* scan generation, increased every round-trip */
163 unsigned int generation;
167 * per-zone information in memory controller.
169 struct mem_cgroup_per_zone {
170 struct lruvec lruvec;
171 unsigned long lru_size[NR_LRU_LISTS];
173 struct mem_cgroup_reclaim_iter reclaim_iter[DEF_PRIORITY + 1];
175 struct rb_node tree_node; /* RB tree node */
176 unsigned long long usage_in_excess;/* Set to the value by which */
177 /* the soft limit is exceeded*/
179 struct mem_cgroup *memcg; /* Back pointer, we cannot */
180 /* use container_of */
183 struct mem_cgroup_per_node {
184 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
187 struct mem_cgroup_lru_info {
188 struct mem_cgroup_per_node *nodeinfo[0];
192 * Cgroups above their limits are maintained in a RB-Tree, independent of
193 * their hierarchy representation
196 struct mem_cgroup_tree_per_zone {
197 struct rb_root rb_root;
201 struct mem_cgroup_tree_per_node {
202 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
205 struct mem_cgroup_tree {
206 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
209 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
211 struct mem_cgroup_threshold {
212 struct eventfd_ctx *eventfd;
217 struct mem_cgroup_threshold_ary {
218 /* An array index points to threshold just below or equal to usage. */
219 int current_threshold;
220 /* Size of entries[] */
222 /* Array of thresholds */
223 struct mem_cgroup_threshold entries[0];
226 struct mem_cgroup_thresholds {
227 /* Primary thresholds array */
228 struct mem_cgroup_threshold_ary *primary;
230 * Spare threshold array.
231 * This is needed to make mem_cgroup_unregister_event() "never fail".
232 * It must be able to store at least primary->size - 1 entries.
234 struct mem_cgroup_threshold_ary *spare;
238 struct mem_cgroup_eventfd_list {
239 struct list_head list;
240 struct eventfd_ctx *eventfd;
243 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
244 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
247 * The memory controller data structure. The memory controller controls both
248 * page cache and RSS per cgroup. We would eventually like to provide
249 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
250 * to help the administrator determine what knobs to tune.
252 * TODO: Add a water mark for the memory controller. Reclaim will begin when
253 * we hit the water mark. May be even add a low water mark, such that
254 * no reclaim occurs from a cgroup at it's low water mark, this is
255 * a feature that will be implemented much later in the future.
258 struct cgroup_subsys_state css;
260 * the counter to account for memory usage
262 struct res_counter res;
266 * the counter to account for mem+swap usage.
268 struct res_counter memsw;
271 * rcu_freeing is used only when freeing struct mem_cgroup,
272 * so put it into a union to avoid wasting more memory.
273 * It must be disjoint from the css field. It could be
274 * in a union with the res field, but res plays a much
275 * larger part in mem_cgroup life than memsw, and might
276 * be of interest, even at time of free, when debugging.
277 * So share rcu_head with the less interesting memsw.
279 struct rcu_head rcu_freeing;
281 * We also need some space for a worker in deferred freeing.
282 * By the time we call it, rcu_freeing is no longer in use.
284 struct work_struct work_freeing;
288 * the counter to account for kernel memory usage.
290 struct res_counter kmem;
292 * Should the accounting and control be hierarchical, per subtree?
295 unsigned long kmem_account_flags; /* See KMEM_ACCOUNTED_*, below */
303 /* OOM-Killer disable */
304 int oom_kill_disable;
306 /* set when res.limit == memsw.limit */
307 bool memsw_is_minimum;
309 /* protect arrays of thresholds */
310 struct mutex thresholds_lock;
312 /* thresholds for memory usage. RCU-protected */
313 struct mem_cgroup_thresholds thresholds;
315 /* thresholds for mem+swap usage. RCU-protected */
316 struct mem_cgroup_thresholds memsw_thresholds;
318 /* For oom notifier event fd */
319 struct list_head oom_notify;
322 * Should we move charges of a task when a task is moved into this
323 * mem_cgroup ? And what type of charges should we move ?
325 unsigned long move_charge_at_immigrate;
327 * set > 0 if pages under this cgroup are moving to other cgroup.
329 atomic_t moving_account;
330 /* taken only while moving_account > 0 */
331 spinlock_t move_lock;
335 struct mem_cgroup_stat_cpu __percpu *stat;
337 * used when a cpu is offlined or other synchronizations
338 * See mem_cgroup_read_stat().
340 struct mem_cgroup_stat_cpu nocpu_base;
341 spinlock_t pcp_counter_lock;
344 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
345 struct tcp_memcontrol tcp_mem;
347 #if defined(CONFIG_MEMCG_KMEM)
348 /* analogous to slab_common's slab_caches list. per-memcg */
349 struct list_head memcg_slab_caches;
350 /* Not a spinlock, we can take a lot of time walking the list */
351 struct mutex slab_caches_mutex;
352 /* Index in the kmem_cache->memcg_params->memcg_caches array */
356 int last_scanned_node;
358 nodemask_t scan_nodes;
359 atomic_t numainfo_events;
360 atomic_t numainfo_updating;
363 * Per cgroup active and inactive list, similar to the
364 * per zone LRU lists.
366 * WARNING: This has to be the last element of the struct. Don't
367 * add new fields after this point.
369 struct mem_cgroup_lru_info info;
372 static size_t memcg_size(void)
374 return sizeof(struct mem_cgroup) +
375 nr_node_ids * sizeof(struct mem_cgroup_per_node);
378 /* internal only representation about the status of kmem accounting. */
380 KMEM_ACCOUNTED_ACTIVE = 0, /* accounted by this cgroup itself */
381 KMEM_ACCOUNTED_ACTIVATED, /* static key enabled. */
382 KMEM_ACCOUNTED_DEAD, /* dead memcg with pending kmem charges */
385 /* We account when limit is on, but only after call sites are patched */
386 #define KMEM_ACCOUNTED_MASK \
387 ((1 << KMEM_ACCOUNTED_ACTIVE) | (1 << KMEM_ACCOUNTED_ACTIVATED))
389 #ifdef CONFIG_MEMCG_KMEM
390 static inline void memcg_kmem_set_active(struct mem_cgroup *memcg)
392 set_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
395 static bool memcg_kmem_is_active(struct mem_cgroup *memcg)
397 return test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
400 static void memcg_kmem_set_activated(struct mem_cgroup *memcg)
402 set_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
405 static void memcg_kmem_clear_activated(struct mem_cgroup *memcg)
407 clear_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
410 static void memcg_kmem_mark_dead(struct mem_cgroup *memcg)
412 if (test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags))
413 set_bit(KMEM_ACCOUNTED_DEAD, &memcg->kmem_account_flags);
416 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup *memcg)
418 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD,
419 &memcg->kmem_account_flags);
423 /* Stuffs for move charges at task migration. */
425 * Types of charges to be moved. "move_charge_at_immitgrate" and
426 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
429 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
430 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
434 /* "mc" and its members are protected by cgroup_mutex */
435 static struct move_charge_struct {
436 spinlock_t lock; /* for from, to */
437 struct mem_cgroup *from;
438 struct mem_cgroup *to;
439 unsigned long immigrate_flags;
440 unsigned long precharge;
441 unsigned long moved_charge;
442 unsigned long moved_swap;
443 struct task_struct *moving_task; /* a task moving charges */
444 wait_queue_head_t waitq; /* a waitq for other context */
446 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
447 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
450 static bool move_anon(void)
452 return test_bit(MOVE_CHARGE_TYPE_ANON, &mc.immigrate_flags);
455 static bool move_file(void)
457 return test_bit(MOVE_CHARGE_TYPE_FILE, &mc.immigrate_flags);
461 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
462 * limit reclaim to prevent infinite loops, if they ever occur.
464 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
465 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
468 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
469 MEM_CGROUP_CHARGE_TYPE_ANON,
470 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
471 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
475 /* for encoding cft->private value on file */
483 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
484 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
485 #define MEMFILE_ATTR(val) ((val) & 0xffff)
486 /* Used for OOM nofiier */
487 #define OOM_CONTROL (0)
490 * Reclaim flags for mem_cgroup_hierarchical_reclaim
492 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
493 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
494 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
495 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
498 * The memcg_create_mutex will be held whenever a new cgroup is created.
499 * As a consequence, any change that needs to protect against new child cgroups
500 * appearing has to hold it as well.
502 static DEFINE_MUTEX(memcg_create_mutex);
504 static void mem_cgroup_get(struct mem_cgroup *memcg);
505 static void mem_cgroup_put(struct mem_cgroup *memcg);
508 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
510 return container_of(s, struct mem_cgroup, css);
513 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
515 return (memcg == root_mem_cgroup);
518 /* Writing them here to avoid exposing memcg's inner layout */
519 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
521 void sock_update_memcg(struct sock *sk)
523 if (mem_cgroup_sockets_enabled) {
524 struct mem_cgroup *memcg;
525 struct cg_proto *cg_proto;
527 BUG_ON(!sk->sk_prot->proto_cgroup);
529 /* Socket cloning can throw us here with sk_cgrp already
530 * filled. It won't however, necessarily happen from
531 * process context. So the test for root memcg given
532 * the current task's memcg won't help us in this case.
534 * Respecting the original socket's memcg is a better
535 * decision in this case.
538 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
539 mem_cgroup_get(sk->sk_cgrp->memcg);
544 memcg = mem_cgroup_from_task(current);
545 cg_proto = sk->sk_prot->proto_cgroup(memcg);
546 if (!mem_cgroup_is_root(memcg) && memcg_proto_active(cg_proto)) {
547 mem_cgroup_get(memcg);
548 sk->sk_cgrp = cg_proto;
553 EXPORT_SYMBOL(sock_update_memcg);
555 void sock_release_memcg(struct sock *sk)
557 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
558 struct mem_cgroup *memcg;
559 WARN_ON(!sk->sk_cgrp->memcg);
560 memcg = sk->sk_cgrp->memcg;
561 mem_cgroup_put(memcg);
565 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
567 if (!memcg || mem_cgroup_is_root(memcg))
570 return &memcg->tcp_mem.cg_proto;
572 EXPORT_SYMBOL(tcp_proto_cgroup);
574 static void disarm_sock_keys(struct mem_cgroup *memcg)
576 if (!memcg_proto_activated(&memcg->tcp_mem.cg_proto))
578 static_key_slow_dec(&memcg_socket_limit_enabled);
581 static void disarm_sock_keys(struct mem_cgroup *memcg)
586 #ifdef CONFIG_MEMCG_KMEM
588 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
589 * There are two main reasons for not using the css_id for this:
590 * 1) this works better in sparse environments, where we have a lot of memcgs,
591 * but only a few kmem-limited. Or also, if we have, for instance, 200
592 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
593 * 200 entry array for that.
595 * 2) In order not to violate the cgroup API, we would like to do all memory
596 * allocation in ->create(). At that point, we haven't yet allocated the
597 * css_id. Having a separate index prevents us from messing with the cgroup
600 * The current size of the caches array is stored in
601 * memcg_limited_groups_array_size. It will double each time we have to
604 static DEFINE_IDA(kmem_limited_groups);
605 int memcg_limited_groups_array_size;
608 * MIN_SIZE is different than 1, because we would like to avoid going through
609 * the alloc/free process all the time. In a small machine, 4 kmem-limited
610 * cgroups is a reasonable guess. In the future, it could be a parameter or
611 * tunable, but that is strictly not necessary.
613 * MAX_SIZE should be as large as the number of css_ids. Ideally, we could get
614 * this constant directly from cgroup, but it is understandable that this is
615 * better kept as an internal representation in cgroup.c. In any case, the
616 * css_id space is not getting any smaller, and we don't have to necessarily
617 * increase ours as well if it increases.
619 #define MEMCG_CACHES_MIN_SIZE 4
620 #define MEMCG_CACHES_MAX_SIZE 65535
623 * A lot of the calls to the cache allocation functions are expected to be
624 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
625 * conditional to this static branch, we'll have to allow modules that does
626 * kmem_cache_alloc and the such to see this symbol as well
628 struct static_key memcg_kmem_enabled_key;
629 EXPORT_SYMBOL(memcg_kmem_enabled_key);
631 static void disarm_kmem_keys(struct mem_cgroup *memcg)
633 if (memcg_kmem_is_active(memcg)) {
634 static_key_slow_dec(&memcg_kmem_enabled_key);
635 ida_simple_remove(&kmem_limited_groups, memcg->kmemcg_id);
638 * This check can't live in kmem destruction function,
639 * since the charges will outlive the cgroup
641 WARN_ON(res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0);
644 static void disarm_kmem_keys(struct mem_cgroup *memcg)
647 #endif /* CONFIG_MEMCG_KMEM */
649 static void disarm_static_keys(struct mem_cgroup *memcg)
651 disarm_sock_keys(memcg);
652 disarm_kmem_keys(memcg);
655 static void drain_all_stock_async(struct mem_cgroup *memcg);
657 static struct mem_cgroup_per_zone *
658 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
660 VM_BUG_ON((unsigned)nid >= nr_node_ids);
661 return &memcg->info.nodeinfo[nid]->zoneinfo[zid];
664 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
669 static struct mem_cgroup_per_zone *
670 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
672 int nid = page_to_nid(page);
673 int zid = page_zonenum(page);
675 return mem_cgroup_zoneinfo(memcg, nid, zid);
678 static struct mem_cgroup_tree_per_zone *
679 soft_limit_tree_node_zone(int nid, int zid)
681 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
684 static struct mem_cgroup_tree_per_zone *
685 soft_limit_tree_from_page(struct page *page)
687 int nid = page_to_nid(page);
688 int zid = page_zonenum(page);
690 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
694 __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
695 struct mem_cgroup_per_zone *mz,
696 struct mem_cgroup_tree_per_zone *mctz,
697 unsigned long long new_usage_in_excess)
699 struct rb_node **p = &mctz->rb_root.rb_node;
700 struct rb_node *parent = NULL;
701 struct mem_cgroup_per_zone *mz_node;
706 mz->usage_in_excess = new_usage_in_excess;
707 if (!mz->usage_in_excess)
711 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
713 if (mz->usage_in_excess < mz_node->usage_in_excess)
716 * We can't avoid mem cgroups that are over their soft
717 * limit by the same amount
719 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
722 rb_link_node(&mz->tree_node, parent, p);
723 rb_insert_color(&mz->tree_node, &mctz->rb_root);
728 __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
729 struct mem_cgroup_per_zone *mz,
730 struct mem_cgroup_tree_per_zone *mctz)
734 rb_erase(&mz->tree_node, &mctz->rb_root);
739 mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
740 struct mem_cgroup_per_zone *mz,
741 struct mem_cgroup_tree_per_zone *mctz)
743 spin_lock(&mctz->lock);
744 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
745 spin_unlock(&mctz->lock);
749 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
751 unsigned long long excess;
752 struct mem_cgroup_per_zone *mz;
753 struct mem_cgroup_tree_per_zone *mctz;
754 int nid = page_to_nid(page);
755 int zid = page_zonenum(page);
756 mctz = soft_limit_tree_from_page(page);
759 * Necessary to update all ancestors when hierarchy is used.
760 * because their event counter is not touched.
762 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
763 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
764 excess = res_counter_soft_limit_excess(&memcg->res);
766 * We have to update the tree if mz is on RB-tree or
767 * mem is over its softlimit.
769 if (excess || mz->on_tree) {
770 spin_lock(&mctz->lock);
771 /* if on-tree, remove it */
773 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
775 * Insert again. mz->usage_in_excess will be updated.
776 * If excess is 0, no tree ops.
778 __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
779 spin_unlock(&mctz->lock);
784 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
787 struct mem_cgroup_per_zone *mz;
788 struct mem_cgroup_tree_per_zone *mctz;
790 for_each_node(node) {
791 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
792 mz = mem_cgroup_zoneinfo(memcg, node, zone);
793 mctz = soft_limit_tree_node_zone(node, zone);
794 mem_cgroup_remove_exceeded(memcg, mz, mctz);
799 static struct mem_cgroup_per_zone *
800 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
802 struct rb_node *rightmost = NULL;
803 struct mem_cgroup_per_zone *mz;
807 rightmost = rb_last(&mctz->rb_root);
809 goto done; /* Nothing to reclaim from */
811 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
813 * Remove the node now but someone else can add it back,
814 * we will to add it back at the end of reclaim to its correct
815 * position in the tree.
817 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
818 if (!res_counter_soft_limit_excess(&mz->memcg->res) ||
819 !css_tryget(&mz->memcg->css))
825 static struct mem_cgroup_per_zone *
826 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
828 struct mem_cgroup_per_zone *mz;
830 spin_lock(&mctz->lock);
831 mz = __mem_cgroup_largest_soft_limit_node(mctz);
832 spin_unlock(&mctz->lock);
837 * Implementation Note: reading percpu statistics for memcg.
839 * Both of vmstat[] and percpu_counter has threshold and do periodic
840 * synchronization to implement "quick" read. There are trade-off between
841 * reading cost and precision of value. Then, we may have a chance to implement
842 * a periodic synchronizion of counter in memcg's counter.
844 * But this _read() function is used for user interface now. The user accounts
845 * memory usage by memory cgroup and he _always_ requires exact value because
846 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
847 * have to visit all online cpus and make sum. So, for now, unnecessary
848 * synchronization is not implemented. (just implemented for cpu hotplug)
850 * If there are kernel internal actions which can make use of some not-exact
851 * value, and reading all cpu value can be performance bottleneck in some
852 * common workload, threashold and synchonization as vmstat[] should be
855 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
856 enum mem_cgroup_stat_index idx)
862 for_each_online_cpu(cpu)
863 val += per_cpu(memcg->stat->count[idx], cpu);
864 #ifdef CONFIG_HOTPLUG_CPU
865 spin_lock(&memcg->pcp_counter_lock);
866 val += memcg->nocpu_base.count[idx];
867 spin_unlock(&memcg->pcp_counter_lock);
873 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
876 int val = (charge) ? 1 : -1;
877 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
880 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
881 enum mem_cgroup_events_index idx)
883 unsigned long val = 0;
886 for_each_online_cpu(cpu)
887 val += per_cpu(memcg->stat->events[idx], cpu);
888 #ifdef CONFIG_HOTPLUG_CPU
889 spin_lock(&memcg->pcp_counter_lock);
890 val += memcg->nocpu_base.events[idx];
891 spin_unlock(&memcg->pcp_counter_lock);
896 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
897 bool anon, int nr_pages)
902 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
903 * counted as CACHE even if it's on ANON LRU.
906 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
909 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
912 /* pagein of a big page is an event. So, ignore page size */
914 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
916 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
917 nr_pages = -nr_pages; /* for event */
920 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
926 mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
928 struct mem_cgroup_per_zone *mz;
930 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
931 return mz->lru_size[lru];
935 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
936 unsigned int lru_mask)
938 struct mem_cgroup_per_zone *mz;
940 unsigned long ret = 0;
942 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
945 if (BIT(lru) & lru_mask)
946 ret += mz->lru_size[lru];
952 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
953 int nid, unsigned int lru_mask)
958 for (zid = 0; zid < MAX_NR_ZONES; zid++)
959 total += mem_cgroup_zone_nr_lru_pages(memcg,
965 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
966 unsigned int lru_mask)
971 for_each_node_state(nid, N_MEMORY)
972 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
976 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
977 enum mem_cgroup_events_target target)
979 unsigned long val, next;
981 val = __this_cpu_read(memcg->stat->nr_page_events);
982 next = __this_cpu_read(memcg->stat->targets[target]);
983 /* from time_after() in jiffies.h */
984 if ((long)next - (long)val < 0) {
986 case MEM_CGROUP_TARGET_THRESH:
987 next = val + THRESHOLDS_EVENTS_TARGET;
989 case MEM_CGROUP_TARGET_SOFTLIMIT:
990 next = val + SOFTLIMIT_EVENTS_TARGET;
992 case MEM_CGROUP_TARGET_NUMAINFO:
993 next = val + NUMAINFO_EVENTS_TARGET;
998 __this_cpu_write(memcg->stat->targets[target], next);
1005 * Check events in order.
1008 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
1011 /* threshold event is triggered in finer grain than soft limit */
1012 if (unlikely(mem_cgroup_event_ratelimit(memcg,
1013 MEM_CGROUP_TARGET_THRESH))) {
1015 bool do_numainfo __maybe_unused;
1017 do_softlimit = mem_cgroup_event_ratelimit(memcg,
1018 MEM_CGROUP_TARGET_SOFTLIMIT);
1019 #if MAX_NUMNODES > 1
1020 do_numainfo = mem_cgroup_event_ratelimit(memcg,
1021 MEM_CGROUP_TARGET_NUMAINFO);
1025 mem_cgroup_threshold(memcg);
1026 if (unlikely(do_softlimit))
1027 mem_cgroup_update_tree(memcg, page);
1028 #if MAX_NUMNODES > 1
1029 if (unlikely(do_numainfo))
1030 atomic_inc(&memcg->numainfo_events);
1036 struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
1038 return mem_cgroup_from_css(
1039 cgroup_subsys_state(cont, mem_cgroup_subsys_id));
1042 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1045 * mm_update_next_owner() may clear mm->owner to NULL
1046 * if it races with swapoff, page migration, etc.
1047 * So this can be called with p == NULL.
1052 return mem_cgroup_from_css(task_subsys_state(p, mem_cgroup_subsys_id));
1055 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
1057 struct mem_cgroup *memcg = NULL;
1062 * Because we have no locks, mm->owner's may be being moved to other
1063 * cgroup. We use css_tryget() here even if this looks
1064 * pessimistic (rather than adding locks here).
1068 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1069 if (unlikely(!memcg))
1071 } while (!css_tryget(&memcg->css));
1077 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1078 * @root: hierarchy root
1079 * @prev: previously returned memcg, NULL on first invocation
1080 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1082 * Returns references to children of the hierarchy below @root, or
1083 * @root itself, or %NULL after a full round-trip.
1085 * Caller must pass the return value in @prev on subsequent
1086 * invocations for reference counting, or use mem_cgroup_iter_break()
1087 * to cancel a hierarchy walk before the round-trip is complete.
1089 * Reclaimers can specify a zone and a priority level in @reclaim to
1090 * divide up the memcgs in the hierarchy among all concurrent
1091 * reclaimers operating on the same zone and priority.
1093 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1094 struct mem_cgroup *prev,
1095 struct mem_cgroup_reclaim_cookie *reclaim)
1097 struct mem_cgroup *memcg = NULL;
1098 struct mem_cgroup *last_visited = NULL;
1099 unsigned long uninitialized_var(dead_count);
1101 if (mem_cgroup_disabled())
1105 root = root_mem_cgroup;
1107 if (prev && !reclaim)
1108 last_visited = prev;
1110 if (!root->use_hierarchy && root != root_mem_cgroup) {
1118 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1119 struct cgroup_subsys_state *css = NULL;
1122 int nid = zone_to_nid(reclaim->zone);
1123 int zid = zone_idx(reclaim->zone);
1124 struct mem_cgroup_per_zone *mz;
1126 mz = mem_cgroup_zoneinfo(root, nid, zid);
1127 iter = &mz->reclaim_iter[reclaim->priority];
1128 last_visited = iter->last_visited;
1129 if (prev && reclaim->generation != iter->generation) {
1130 iter->last_visited = NULL;
1135 * If the dead_count mismatches, a destruction
1136 * has happened or is happening concurrently.
1137 * If the dead_count matches, a destruction
1138 * might still happen concurrently, but since
1139 * we checked under RCU, that destruction
1140 * won't free the object until we release the
1141 * RCU reader lock. Thus, the dead_count
1142 * check verifies the pointer is still valid,
1143 * css_tryget() verifies the cgroup pointed to
1146 dead_count = atomic_read(&root->dead_count);
1148 last_visited = iter->last_visited;
1150 if ((dead_count != iter->last_dead_count) ||
1151 !css_tryget(&last_visited->css)) {
1152 last_visited = NULL;
1158 * Root is not visited by cgroup iterators so it needs an
1161 if (!last_visited) {
1164 struct cgroup *prev_cgroup, *next_cgroup;
1166 prev_cgroup = (last_visited == root) ? NULL
1167 : last_visited->css.cgroup;
1168 next_cgroup = cgroup_next_descendant_pre(prev_cgroup,
1171 css = cgroup_subsys_state(next_cgroup,
1172 mem_cgroup_subsys_id);
1176 * Even if we found a group we have to make sure it is alive.
1177 * css && !memcg means that the groups should be skipped and
1178 * we should continue the tree walk.
1179 * last_visited css is safe to use because it is protected by
1180 * css_get and the tree walk is rcu safe.
1182 if (css == &root->css || (css && css_tryget(css)))
1183 memcg = mem_cgroup_from_css(css);
1186 struct mem_cgroup *curr = memcg;
1189 css_put(&last_visited->css);
1192 curr = mem_cgroup_from_css(css);
1194 iter->last_visited = curr;
1196 iter->last_dead_count = dead_count;
1200 else if (!prev && memcg)
1201 reclaim->generation = iter->generation;
1202 } else if (css && !memcg) {
1203 last_visited = mem_cgroup_from_css(css);
1212 if (prev && prev != root)
1213 css_put(&prev->css);
1219 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1220 * @root: hierarchy root
1221 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1223 void mem_cgroup_iter_break(struct mem_cgroup *root,
1224 struct mem_cgroup *prev)
1227 root = root_mem_cgroup;
1228 if (prev && prev != root)
1229 css_put(&prev->css);
1233 * Iteration constructs for visiting all cgroups (under a tree). If
1234 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1235 * be used for reference counting.
1237 #define for_each_mem_cgroup_tree(iter, root) \
1238 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1240 iter = mem_cgroup_iter(root, iter, NULL))
1242 #define for_each_mem_cgroup(iter) \
1243 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1245 iter = mem_cgroup_iter(NULL, iter, NULL))
1247 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1249 struct mem_cgroup *memcg;
1252 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1253 if (unlikely(!memcg))
1258 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1261 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1269 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1272 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1273 * @zone: zone of the wanted lruvec
1274 * @memcg: memcg of the wanted lruvec
1276 * Returns the lru list vector holding pages for the given @zone and
1277 * @mem. This can be the global zone lruvec, if the memory controller
1280 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1281 struct mem_cgroup *memcg)
1283 struct mem_cgroup_per_zone *mz;
1284 struct lruvec *lruvec;
1286 if (mem_cgroup_disabled()) {
1287 lruvec = &zone->lruvec;
1291 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1292 lruvec = &mz->lruvec;
1295 * Since a node can be onlined after the mem_cgroup was created,
1296 * we have to be prepared to initialize lruvec->zone here;
1297 * and if offlined then reonlined, we need to reinitialize it.
1299 if (unlikely(lruvec->zone != zone))
1300 lruvec->zone = zone;
1305 * Following LRU functions are allowed to be used without PCG_LOCK.
1306 * Operations are called by routine of global LRU independently from memcg.
1307 * What we have to take care of here is validness of pc->mem_cgroup.
1309 * Changes to pc->mem_cgroup happens when
1312 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1313 * It is added to LRU before charge.
1314 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1315 * When moving account, the page is not on LRU. It's isolated.
1319 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1321 * @zone: zone of the page
1323 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1325 struct mem_cgroup_per_zone *mz;
1326 struct mem_cgroup *memcg;
1327 struct page_cgroup *pc;
1328 struct lruvec *lruvec;
1330 if (mem_cgroup_disabled()) {
1331 lruvec = &zone->lruvec;
1335 pc = lookup_page_cgroup(page);
1336 memcg = pc->mem_cgroup;
1339 * Surreptitiously switch any uncharged offlist page to root:
1340 * an uncharged page off lru does nothing to secure
1341 * its former mem_cgroup from sudden removal.
1343 * Our caller holds lru_lock, and PageCgroupUsed is updated
1344 * under page_cgroup lock: between them, they make all uses
1345 * of pc->mem_cgroup safe.
1347 if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1348 pc->mem_cgroup = memcg = root_mem_cgroup;
1350 mz = page_cgroup_zoneinfo(memcg, page);
1351 lruvec = &mz->lruvec;
1354 * Since a node can be onlined after the mem_cgroup was created,
1355 * we have to be prepared to initialize lruvec->zone here;
1356 * and if offlined then reonlined, we need to reinitialize it.
1358 if (unlikely(lruvec->zone != zone))
1359 lruvec->zone = zone;
1364 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1365 * @lruvec: mem_cgroup per zone lru vector
1366 * @lru: index of lru list the page is sitting on
1367 * @nr_pages: positive when adding or negative when removing
1369 * This function must be called when a page is added to or removed from an
1372 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1375 struct mem_cgroup_per_zone *mz;
1376 unsigned long *lru_size;
1378 if (mem_cgroup_disabled())
1381 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1382 lru_size = mz->lru_size + lru;
1383 *lru_size += nr_pages;
1384 VM_BUG_ON((long)(*lru_size) < 0);
1388 * Checks whether given mem is same or in the root_mem_cgroup's
1391 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1392 struct mem_cgroup *memcg)
1394 if (root_memcg == memcg)
1396 if (!root_memcg->use_hierarchy || !memcg)
1398 return css_is_ancestor(&memcg->css, &root_memcg->css);
1401 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1402 struct mem_cgroup *memcg)
1407 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1412 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *memcg)
1415 struct mem_cgroup *curr = NULL;
1416 struct task_struct *p;
1418 p = find_lock_task_mm(task);
1420 curr = try_get_mem_cgroup_from_mm(p->mm);
1424 * All threads may have already detached their mm's, but the oom
1425 * killer still needs to detect if they have already been oom
1426 * killed to prevent needlessly killing additional tasks.
1429 curr = mem_cgroup_from_task(task);
1431 css_get(&curr->css);
1437 * We should check use_hierarchy of "memcg" not "curr". Because checking
1438 * use_hierarchy of "curr" here make this function true if hierarchy is
1439 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1440 * hierarchy(even if use_hierarchy is disabled in "memcg").
1442 ret = mem_cgroup_same_or_subtree(memcg, curr);
1443 css_put(&curr->css);
1447 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1449 unsigned long inactive_ratio;
1450 unsigned long inactive;
1451 unsigned long active;
1454 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1455 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1457 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1459 inactive_ratio = int_sqrt(10 * gb);
1463 return inactive * inactive_ratio < active;
1466 #define mem_cgroup_from_res_counter(counter, member) \
1467 container_of(counter, struct mem_cgroup, member)
1470 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1471 * @memcg: the memory cgroup
1473 * Returns the maximum amount of memory @mem can be charged with, in
1476 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1478 unsigned long long margin;
1480 margin = res_counter_margin(&memcg->res);
1481 if (do_swap_account)
1482 margin = min(margin, res_counter_margin(&memcg->memsw));
1483 return margin >> PAGE_SHIFT;
1486 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1488 struct cgroup *cgrp = memcg->css.cgroup;
1491 if (cgrp->parent == NULL)
1492 return vm_swappiness;
1494 return memcg->swappiness;
1498 * memcg->moving_account is used for checking possibility that some thread is
1499 * calling move_account(). When a thread on CPU-A starts moving pages under
1500 * a memcg, other threads should check memcg->moving_account under
1501 * rcu_read_lock(), like this:
1505 * memcg->moving_account+1 if (memcg->mocing_account)
1507 * synchronize_rcu() update something.
1512 /* for quick checking without looking up memcg */
1513 atomic_t memcg_moving __read_mostly;
1515 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1517 atomic_inc(&memcg_moving);
1518 atomic_inc(&memcg->moving_account);
1522 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1525 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1526 * We check NULL in callee rather than caller.
1529 atomic_dec(&memcg_moving);
1530 atomic_dec(&memcg->moving_account);
1535 * 2 routines for checking "mem" is under move_account() or not.
1537 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1538 * is used for avoiding races in accounting. If true,
1539 * pc->mem_cgroup may be overwritten.
1541 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1542 * under hierarchy of moving cgroups. This is for
1543 * waiting at hith-memory prressure caused by "move".
1546 static bool mem_cgroup_stolen(struct mem_cgroup *memcg)
1548 VM_BUG_ON(!rcu_read_lock_held());
1549 return atomic_read(&memcg->moving_account) > 0;
1552 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1554 struct mem_cgroup *from;
1555 struct mem_cgroup *to;
1558 * Unlike task_move routines, we access mc.to, mc.from not under
1559 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1561 spin_lock(&mc.lock);
1567 ret = mem_cgroup_same_or_subtree(memcg, from)
1568 || mem_cgroup_same_or_subtree(memcg, to);
1570 spin_unlock(&mc.lock);
1574 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1576 if (mc.moving_task && current != mc.moving_task) {
1577 if (mem_cgroup_under_move(memcg)) {
1579 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1580 /* moving charge context might have finished. */
1583 finish_wait(&mc.waitq, &wait);
1591 * Take this lock when
1592 * - a code tries to modify page's memcg while it's USED.
1593 * - a code tries to modify page state accounting in a memcg.
1594 * see mem_cgroup_stolen(), too.
1596 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1597 unsigned long *flags)
1599 spin_lock_irqsave(&memcg->move_lock, *flags);
1602 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1603 unsigned long *flags)
1605 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1608 #define K(x) ((x) << (PAGE_SHIFT-10))
1610 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1611 * @memcg: The memory cgroup that went over limit
1612 * @p: Task that is going to be killed
1614 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1617 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1619 struct cgroup *task_cgrp;
1620 struct cgroup *mem_cgrp;
1622 * Need a buffer in BSS, can't rely on allocations. The code relies
1623 * on the assumption that OOM is serialized for memory controller.
1624 * If this assumption is broken, revisit this code.
1626 static char memcg_name[PATH_MAX];
1628 struct mem_cgroup *iter;
1636 mem_cgrp = memcg->css.cgroup;
1637 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1639 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1642 * Unfortunately, we are unable to convert to a useful name
1643 * But we'll still print out the usage information
1650 pr_info("Task in %s killed", memcg_name);
1653 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1661 * Continues from above, so we don't need an KERN_ level
1663 pr_cont(" as a result of limit of %s\n", memcg_name);
1666 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1667 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1668 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1669 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1670 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1671 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1672 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1673 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1674 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1675 res_counter_read_u64(&memcg->kmem, RES_USAGE) >> 10,
1676 res_counter_read_u64(&memcg->kmem, RES_LIMIT) >> 10,
1677 res_counter_read_u64(&memcg->kmem, RES_FAILCNT));
1679 for_each_mem_cgroup_tree(iter, memcg) {
1680 pr_info("Memory cgroup stats");
1683 ret = cgroup_path(iter->css.cgroup, memcg_name, PATH_MAX);
1685 pr_cont(" for %s", memcg_name);
1689 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1690 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1692 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1693 K(mem_cgroup_read_stat(iter, i)));
1696 for (i = 0; i < NR_LRU_LISTS; i++)
1697 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1698 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1705 * This function returns the number of memcg under hierarchy tree. Returns
1706 * 1(self count) if no children.
1708 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1711 struct mem_cgroup *iter;
1713 for_each_mem_cgroup_tree(iter, memcg)
1719 * Return the memory (and swap, if configured) limit for a memcg.
1721 static u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1725 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1728 * Do not consider swap space if we cannot swap due to swappiness
1730 if (mem_cgroup_swappiness(memcg)) {
1733 limit += total_swap_pages << PAGE_SHIFT;
1734 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1737 * If memsw is finite and limits the amount of swap space
1738 * available to this memcg, return that limit.
1740 limit = min(limit, memsw);
1746 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1749 struct mem_cgroup *iter;
1750 unsigned long chosen_points = 0;
1751 unsigned long totalpages;
1752 unsigned int points = 0;
1753 struct task_struct *chosen = NULL;
1756 * If current has a pending SIGKILL, then automatically select it. The
1757 * goal is to allow it to allocate so that it may quickly exit and free
1760 if (fatal_signal_pending(current)) {
1761 set_thread_flag(TIF_MEMDIE);
1765 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1766 totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
1767 for_each_mem_cgroup_tree(iter, memcg) {
1768 struct cgroup *cgroup = iter->css.cgroup;
1769 struct cgroup_iter it;
1770 struct task_struct *task;
1772 cgroup_iter_start(cgroup, &it);
1773 while ((task = cgroup_iter_next(cgroup, &it))) {
1774 switch (oom_scan_process_thread(task, totalpages, NULL,
1776 case OOM_SCAN_SELECT:
1778 put_task_struct(chosen);
1780 chosen_points = ULONG_MAX;
1781 get_task_struct(chosen);
1783 case OOM_SCAN_CONTINUE:
1785 case OOM_SCAN_ABORT:
1786 cgroup_iter_end(cgroup, &it);
1787 mem_cgroup_iter_break(memcg, iter);
1789 put_task_struct(chosen);
1794 points = oom_badness(task, memcg, NULL, totalpages);
1795 if (points > chosen_points) {
1797 put_task_struct(chosen);
1799 chosen_points = points;
1800 get_task_struct(chosen);
1803 cgroup_iter_end(cgroup, &it);
1808 points = chosen_points * 1000 / totalpages;
1809 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1810 NULL, "Memory cgroup out of memory");
1813 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1815 unsigned long flags)
1817 unsigned long total = 0;
1818 bool noswap = false;
1821 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1823 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1826 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1828 drain_all_stock_async(memcg);
1829 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1831 * Allow limit shrinkers, which are triggered directly
1832 * by userspace, to catch signals and stop reclaim
1833 * after minimal progress, regardless of the margin.
1835 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1837 if (mem_cgroup_margin(memcg))
1840 * If nothing was reclaimed after two attempts, there
1841 * may be no reclaimable pages in this hierarchy.
1850 * test_mem_cgroup_node_reclaimable
1851 * @memcg: the target memcg
1852 * @nid: the node ID to be checked.
1853 * @noswap : specify true here if the user wants flle only information.
1855 * This function returns whether the specified memcg contains any
1856 * reclaimable pages on a node. Returns true if there are any reclaimable
1857 * pages in the node.
1859 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1860 int nid, bool noswap)
1862 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1864 if (noswap || !total_swap_pages)
1866 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1871 #if MAX_NUMNODES > 1
1874 * Always updating the nodemask is not very good - even if we have an empty
1875 * list or the wrong list here, we can start from some node and traverse all
1876 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1879 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1883 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1884 * pagein/pageout changes since the last update.
1886 if (!atomic_read(&memcg->numainfo_events))
1888 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1891 /* make a nodemask where this memcg uses memory from */
1892 memcg->scan_nodes = node_states[N_MEMORY];
1894 for_each_node_mask(nid, node_states[N_MEMORY]) {
1896 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1897 node_clear(nid, memcg->scan_nodes);
1900 atomic_set(&memcg->numainfo_events, 0);
1901 atomic_set(&memcg->numainfo_updating, 0);
1905 * Selecting a node where we start reclaim from. Because what we need is just
1906 * reducing usage counter, start from anywhere is O,K. Considering
1907 * memory reclaim from current node, there are pros. and cons.
1909 * Freeing memory from current node means freeing memory from a node which
1910 * we'll use or we've used. So, it may make LRU bad. And if several threads
1911 * hit limits, it will see a contention on a node. But freeing from remote
1912 * node means more costs for memory reclaim because of memory latency.
1914 * Now, we use round-robin. Better algorithm is welcomed.
1916 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1920 mem_cgroup_may_update_nodemask(memcg);
1921 node = memcg->last_scanned_node;
1923 node = next_node(node, memcg->scan_nodes);
1924 if (node == MAX_NUMNODES)
1925 node = first_node(memcg->scan_nodes);
1927 * We call this when we hit limit, not when pages are added to LRU.
1928 * No LRU may hold pages because all pages are UNEVICTABLE or
1929 * memcg is too small and all pages are not on LRU. In that case,
1930 * we use curret node.
1932 if (unlikely(node == MAX_NUMNODES))
1933 node = numa_node_id();
1935 memcg->last_scanned_node = node;
1940 * Check all nodes whether it contains reclaimable pages or not.
1941 * For quick scan, we make use of scan_nodes. This will allow us to skip
1942 * unused nodes. But scan_nodes is lazily updated and may not cotain
1943 * enough new information. We need to do double check.
1945 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1950 * quick check...making use of scan_node.
1951 * We can skip unused nodes.
1953 if (!nodes_empty(memcg->scan_nodes)) {
1954 for (nid = first_node(memcg->scan_nodes);
1956 nid = next_node(nid, memcg->scan_nodes)) {
1958 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1963 * Check rest of nodes.
1965 for_each_node_state(nid, N_MEMORY) {
1966 if (node_isset(nid, memcg->scan_nodes))
1968 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1975 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1980 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1982 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
1986 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1989 unsigned long *total_scanned)
1991 struct mem_cgroup *victim = NULL;
1994 unsigned long excess;
1995 unsigned long nr_scanned;
1996 struct mem_cgroup_reclaim_cookie reclaim = {
2001 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
2004 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
2009 * If we have not been able to reclaim
2010 * anything, it might because there are
2011 * no reclaimable pages under this hierarchy
2016 * We want to do more targeted reclaim.
2017 * excess >> 2 is not to excessive so as to
2018 * reclaim too much, nor too less that we keep
2019 * coming back to reclaim from this cgroup
2021 if (total >= (excess >> 2) ||
2022 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
2027 if (!mem_cgroup_reclaimable(victim, false))
2029 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
2031 *total_scanned += nr_scanned;
2032 if (!res_counter_soft_limit_excess(&root_memcg->res))
2035 mem_cgroup_iter_break(root_memcg, victim);
2040 * Check OOM-Killer is already running under our hierarchy.
2041 * If someone is running, return false.
2042 * Has to be called with memcg_oom_lock
2044 static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg)
2046 struct mem_cgroup *iter, *failed = NULL;
2048 for_each_mem_cgroup_tree(iter, memcg) {
2049 if (iter->oom_lock) {
2051 * this subtree of our hierarchy is already locked
2052 * so we cannot give a lock.
2055 mem_cgroup_iter_break(memcg, iter);
2058 iter->oom_lock = true;
2065 * OK, we failed to lock the whole subtree so we have to clean up
2066 * what we set up to the failing subtree
2068 for_each_mem_cgroup_tree(iter, memcg) {
2069 if (iter == failed) {
2070 mem_cgroup_iter_break(memcg, iter);
2073 iter->oom_lock = false;
2079 * Has to be called with memcg_oom_lock
2081 static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
2083 struct mem_cgroup *iter;
2085 for_each_mem_cgroup_tree(iter, memcg)
2086 iter->oom_lock = false;
2090 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
2092 struct mem_cgroup *iter;
2094 for_each_mem_cgroup_tree(iter, memcg)
2095 atomic_inc(&iter->under_oom);
2098 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
2100 struct mem_cgroup *iter;
2103 * When a new child is created while the hierarchy is under oom,
2104 * mem_cgroup_oom_lock() may not be called. We have to use
2105 * atomic_add_unless() here.
2107 for_each_mem_cgroup_tree(iter, memcg)
2108 atomic_add_unless(&iter->under_oom, -1, 0);
2111 static DEFINE_SPINLOCK(memcg_oom_lock);
2112 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
2114 struct oom_wait_info {
2115 struct mem_cgroup *memcg;
2119 static int memcg_oom_wake_function(wait_queue_t *wait,
2120 unsigned mode, int sync, void *arg)
2122 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
2123 struct mem_cgroup *oom_wait_memcg;
2124 struct oom_wait_info *oom_wait_info;
2126 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
2127 oom_wait_memcg = oom_wait_info->memcg;
2130 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2131 * Then we can use css_is_ancestor without taking care of RCU.
2133 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
2134 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
2136 return autoremove_wake_function(wait, mode, sync, arg);
2139 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
2141 /* for filtering, pass "memcg" as argument. */
2142 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
2145 static void memcg_oom_recover(struct mem_cgroup *memcg)
2147 if (memcg && atomic_read(&memcg->under_oom))
2148 memcg_wakeup_oom(memcg);
2152 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
2154 static bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask,
2157 struct oom_wait_info owait;
2158 bool locked, need_to_kill;
2160 owait.memcg = memcg;
2161 owait.wait.flags = 0;
2162 owait.wait.func = memcg_oom_wake_function;
2163 owait.wait.private = current;
2164 INIT_LIST_HEAD(&owait.wait.task_list);
2165 need_to_kill = true;
2166 mem_cgroup_mark_under_oom(memcg);
2168 /* At first, try to OOM lock hierarchy under memcg.*/
2169 spin_lock(&memcg_oom_lock);
2170 locked = mem_cgroup_oom_lock(memcg);
2172 * Even if signal_pending(), we can't quit charge() loop without
2173 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
2174 * under OOM is always welcomed, use TASK_KILLABLE here.
2176 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2177 if (!locked || memcg->oom_kill_disable)
2178 need_to_kill = false;
2180 mem_cgroup_oom_notify(memcg);
2181 spin_unlock(&memcg_oom_lock);
2184 finish_wait(&memcg_oom_waitq, &owait.wait);
2185 mem_cgroup_out_of_memory(memcg, mask, order);
2188 finish_wait(&memcg_oom_waitq, &owait.wait);
2190 spin_lock(&memcg_oom_lock);
2192 mem_cgroup_oom_unlock(memcg);
2193 memcg_wakeup_oom(memcg);
2194 spin_unlock(&memcg_oom_lock);
2196 mem_cgroup_unmark_under_oom(memcg);
2198 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
2200 /* Give chance to dying process */
2201 schedule_timeout_uninterruptible(1);
2206 * Currently used to update mapped file statistics, but the routine can be
2207 * generalized to update other statistics as well.
2209 * Notes: Race condition
2211 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2212 * it tends to be costly. But considering some conditions, we doesn't need
2213 * to do so _always_.
2215 * Considering "charge", lock_page_cgroup() is not required because all
2216 * file-stat operations happen after a page is attached to radix-tree. There
2217 * are no race with "charge".
2219 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2220 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2221 * if there are race with "uncharge". Statistics itself is properly handled
2224 * Considering "move", this is an only case we see a race. To make the race
2225 * small, we check mm->moving_account and detect there are possibility of race
2226 * If there is, we take a lock.
2229 void __mem_cgroup_begin_update_page_stat(struct page *page,
2230 bool *locked, unsigned long *flags)
2232 struct mem_cgroup *memcg;
2233 struct page_cgroup *pc;
2235 pc = lookup_page_cgroup(page);
2237 memcg = pc->mem_cgroup;
2238 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2241 * If this memory cgroup is not under account moving, we don't
2242 * need to take move_lock_mem_cgroup(). Because we already hold
2243 * rcu_read_lock(), any calls to move_account will be delayed until
2244 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2246 if (!mem_cgroup_stolen(memcg))
2249 move_lock_mem_cgroup(memcg, flags);
2250 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2251 move_unlock_mem_cgroup(memcg, flags);
2257 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
2259 struct page_cgroup *pc = lookup_page_cgroup(page);
2262 * It's guaranteed that pc->mem_cgroup never changes while
2263 * lock is held because a routine modifies pc->mem_cgroup
2264 * should take move_lock_mem_cgroup().
2266 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
2269 void mem_cgroup_update_page_stat(struct page *page,
2270 enum mem_cgroup_page_stat_item idx, int val)
2272 struct mem_cgroup *memcg;
2273 struct page_cgroup *pc = lookup_page_cgroup(page);
2274 unsigned long uninitialized_var(flags);
2276 if (mem_cgroup_disabled())
2279 memcg = pc->mem_cgroup;
2280 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2284 case MEMCG_NR_FILE_MAPPED:
2285 idx = MEM_CGROUP_STAT_FILE_MAPPED;
2291 this_cpu_add(memcg->stat->count[idx], val);
2295 * size of first charge trial. "32" comes from vmscan.c's magic value.
2296 * TODO: maybe necessary to use big numbers in big irons.
2298 #define CHARGE_BATCH 32U
2299 struct memcg_stock_pcp {
2300 struct mem_cgroup *cached; /* this never be root cgroup */
2301 unsigned int nr_pages;
2302 struct work_struct work;
2303 unsigned long flags;
2304 #define FLUSHING_CACHED_CHARGE 0
2306 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2307 static DEFINE_MUTEX(percpu_charge_mutex);
2310 * consume_stock: Try to consume stocked charge on this cpu.
2311 * @memcg: memcg to consume from.
2312 * @nr_pages: how many pages to charge.
2314 * The charges will only happen if @memcg matches the current cpu's memcg
2315 * stock, and at least @nr_pages are available in that stock. Failure to
2316 * service an allocation will refill the stock.
2318 * returns true if successful, false otherwise.
2320 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2322 struct memcg_stock_pcp *stock;
2325 if (nr_pages > CHARGE_BATCH)
2328 stock = &get_cpu_var(memcg_stock);
2329 if (memcg == stock->cached && stock->nr_pages >= nr_pages)
2330 stock->nr_pages -= nr_pages;
2331 else /* need to call res_counter_charge */
2333 put_cpu_var(memcg_stock);
2338 * Returns stocks cached in percpu to res_counter and reset cached information.
2340 static void drain_stock(struct memcg_stock_pcp *stock)
2342 struct mem_cgroup *old = stock->cached;
2344 if (stock->nr_pages) {
2345 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2347 res_counter_uncharge(&old->res, bytes);
2348 if (do_swap_account)
2349 res_counter_uncharge(&old->memsw, bytes);
2350 stock->nr_pages = 0;
2352 stock->cached = NULL;
2356 * This must be called under preempt disabled or must be called by
2357 * a thread which is pinned to local cpu.
2359 static void drain_local_stock(struct work_struct *dummy)
2361 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2363 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2366 static void __init memcg_stock_init(void)
2370 for_each_possible_cpu(cpu) {
2371 struct memcg_stock_pcp *stock =
2372 &per_cpu(memcg_stock, cpu);
2373 INIT_WORK(&stock->work, drain_local_stock);
2378 * Cache charges(val) which is from res_counter, to local per_cpu area.
2379 * This will be consumed by consume_stock() function, later.
2381 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2383 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2385 if (stock->cached != memcg) { /* reset if necessary */
2387 stock->cached = memcg;
2389 stock->nr_pages += nr_pages;
2390 put_cpu_var(memcg_stock);
2394 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2395 * of the hierarchy under it. sync flag says whether we should block
2396 * until the work is done.
2398 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2402 /* Notify other cpus that system-wide "drain" is running */
2405 for_each_online_cpu(cpu) {
2406 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2407 struct mem_cgroup *memcg;
2409 memcg = stock->cached;
2410 if (!memcg || !stock->nr_pages)
2412 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2414 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2416 drain_local_stock(&stock->work);
2418 schedule_work_on(cpu, &stock->work);
2426 for_each_online_cpu(cpu) {
2427 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2428 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2429 flush_work(&stock->work);
2436 * Tries to drain stocked charges in other cpus. This function is asynchronous
2437 * and just put a work per cpu for draining localy on each cpu. Caller can
2438 * expects some charges will be back to res_counter later but cannot wait for
2441 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2444 * If someone calls draining, avoid adding more kworker runs.
2446 if (!mutex_trylock(&percpu_charge_mutex))
2448 drain_all_stock(root_memcg, false);
2449 mutex_unlock(&percpu_charge_mutex);
2452 /* This is a synchronous drain interface. */
2453 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2455 /* called when force_empty is called */
2456 mutex_lock(&percpu_charge_mutex);
2457 drain_all_stock(root_memcg, true);
2458 mutex_unlock(&percpu_charge_mutex);
2462 * This function drains percpu counter value from DEAD cpu and
2463 * move it to local cpu. Note that this function can be preempted.
2465 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2469 spin_lock(&memcg->pcp_counter_lock);
2470 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2471 long x = per_cpu(memcg->stat->count[i], cpu);
2473 per_cpu(memcg->stat->count[i], cpu) = 0;
2474 memcg->nocpu_base.count[i] += x;
2476 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2477 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2479 per_cpu(memcg->stat->events[i], cpu) = 0;
2480 memcg->nocpu_base.events[i] += x;
2482 spin_unlock(&memcg->pcp_counter_lock);
2485 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2486 unsigned long action,
2489 int cpu = (unsigned long)hcpu;
2490 struct memcg_stock_pcp *stock;
2491 struct mem_cgroup *iter;
2493 if (action == CPU_ONLINE)
2496 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2499 for_each_mem_cgroup(iter)
2500 mem_cgroup_drain_pcp_counter(iter, cpu);
2502 stock = &per_cpu(memcg_stock, cpu);
2508 /* See __mem_cgroup_try_charge() for details */
2510 CHARGE_OK, /* success */
2511 CHARGE_RETRY, /* need to retry but retry is not bad */
2512 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2513 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2514 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2517 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2518 unsigned int nr_pages, unsigned int min_pages,
2521 unsigned long csize = nr_pages * PAGE_SIZE;
2522 struct mem_cgroup *mem_over_limit;
2523 struct res_counter *fail_res;
2524 unsigned long flags = 0;
2527 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2530 if (!do_swap_account)
2532 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2536 res_counter_uncharge(&memcg->res, csize);
2537 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2538 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2540 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2542 * Never reclaim on behalf of optional batching, retry with a
2543 * single page instead.
2545 if (nr_pages > min_pages)
2546 return CHARGE_RETRY;
2548 if (!(gfp_mask & __GFP_WAIT))
2549 return CHARGE_WOULDBLOCK;
2551 if (gfp_mask & __GFP_NORETRY)
2552 return CHARGE_NOMEM;
2554 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2555 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2556 return CHARGE_RETRY;
2558 * Even though the limit is exceeded at this point, reclaim
2559 * may have been able to free some pages. Retry the charge
2560 * before killing the task.
2562 * Only for regular pages, though: huge pages are rather
2563 * unlikely to succeed so close to the limit, and we fall back
2564 * to regular pages anyway in case of failure.
2566 if (nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER) && ret)
2567 return CHARGE_RETRY;
2570 * At task move, charge accounts can be doubly counted. So, it's
2571 * better to wait until the end of task_move if something is going on.
2573 if (mem_cgroup_wait_acct_move(mem_over_limit))
2574 return CHARGE_RETRY;
2576 /* If we don't need to call oom-killer at el, return immediately */
2578 return CHARGE_NOMEM;
2580 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask, get_order(csize)))
2581 return CHARGE_OOM_DIE;
2583 return CHARGE_RETRY;
2587 * __mem_cgroup_try_charge() does
2588 * 1. detect memcg to be charged against from passed *mm and *ptr,
2589 * 2. update res_counter
2590 * 3. call memory reclaim if necessary.
2592 * In some special case, if the task is fatal, fatal_signal_pending() or
2593 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2594 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2595 * as possible without any hazards. 2: all pages should have a valid
2596 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2597 * pointer, that is treated as a charge to root_mem_cgroup.
2599 * So __mem_cgroup_try_charge() will return
2600 * 0 ... on success, filling *ptr with a valid memcg pointer.
2601 * -ENOMEM ... charge failure because of resource limits.
2602 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2604 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2605 * the oom-killer can be invoked.
2607 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2609 unsigned int nr_pages,
2610 struct mem_cgroup **ptr,
2613 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2614 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2615 struct mem_cgroup *memcg = NULL;
2619 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2620 * in system level. So, allow to go ahead dying process in addition to
2623 if (unlikely(test_thread_flag(TIF_MEMDIE)
2624 || fatal_signal_pending(current)))
2628 * We always charge the cgroup the mm_struct belongs to.
2629 * The mm_struct's mem_cgroup changes on task migration if the
2630 * thread group leader migrates. It's possible that mm is not
2631 * set, if so charge the root memcg (happens for pagecache usage).
2634 *ptr = root_mem_cgroup;
2636 if (*ptr) { /* css should be a valid one */
2638 if (mem_cgroup_is_root(memcg))
2640 if (consume_stock(memcg, nr_pages))
2642 css_get(&memcg->css);
2644 struct task_struct *p;
2647 p = rcu_dereference(mm->owner);
2649 * Because we don't have task_lock(), "p" can exit.
2650 * In that case, "memcg" can point to root or p can be NULL with
2651 * race with swapoff. Then, we have small risk of mis-accouning.
2652 * But such kind of mis-account by race always happens because
2653 * we don't have cgroup_mutex(). It's overkill and we allo that
2655 * (*) swapoff at el will charge against mm-struct not against
2656 * task-struct. So, mm->owner can be NULL.
2658 memcg = mem_cgroup_from_task(p);
2660 memcg = root_mem_cgroup;
2661 if (mem_cgroup_is_root(memcg)) {
2665 if (consume_stock(memcg, nr_pages)) {
2667 * It seems dagerous to access memcg without css_get().
2668 * But considering how consume_stok works, it's not
2669 * necessary. If consume_stock success, some charges
2670 * from this memcg are cached on this cpu. So, we
2671 * don't need to call css_get()/css_tryget() before
2672 * calling consume_stock().
2677 /* after here, we may be blocked. we need to get refcnt */
2678 if (!css_tryget(&memcg->css)) {
2688 /* If killed, bypass charge */
2689 if (fatal_signal_pending(current)) {
2690 css_put(&memcg->css);
2695 if (oom && !nr_oom_retries) {
2697 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2700 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, nr_pages,
2705 case CHARGE_RETRY: /* not in OOM situation but retry */
2707 css_put(&memcg->css);
2710 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2711 css_put(&memcg->css);
2713 case CHARGE_NOMEM: /* OOM routine works */
2715 css_put(&memcg->css);
2718 /* If oom, we never return -ENOMEM */
2721 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2722 css_put(&memcg->css);
2725 } while (ret != CHARGE_OK);
2727 if (batch > nr_pages)
2728 refill_stock(memcg, batch - nr_pages);
2729 css_put(&memcg->css);
2737 *ptr = root_mem_cgroup;
2742 * Somemtimes we have to undo a charge we got by try_charge().
2743 * This function is for that and do uncharge, put css's refcnt.
2744 * gotten by try_charge().
2746 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2747 unsigned int nr_pages)
2749 if (!mem_cgroup_is_root(memcg)) {
2750 unsigned long bytes = nr_pages * PAGE_SIZE;
2752 res_counter_uncharge(&memcg->res, bytes);
2753 if (do_swap_account)
2754 res_counter_uncharge(&memcg->memsw, bytes);
2759 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2760 * This is useful when moving usage to parent cgroup.
2762 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2763 unsigned int nr_pages)
2765 unsigned long bytes = nr_pages * PAGE_SIZE;
2767 if (mem_cgroup_is_root(memcg))
2770 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2771 if (do_swap_account)
2772 res_counter_uncharge_until(&memcg->memsw,
2773 memcg->memsw.parent, bytes);
2777 * A helper function to get mem_cgroup from ID. must be called under
2778 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2779 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2780 * called against removed memcg.)
2782 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2784 struct cgroup_subsys_state *css;
2786 /* ID 0 is unused ID */
2789 css = css_lookup(&mem_cgroup_subsys, id);
2792 return mem_cgroup_from_css(css);
2795 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2797 struct mem_cgroup *memcg = NULL;
2798 struct page_cgroup *pc;
2802 VM_BUG_ON(!PageLocked(page));
2804 pc = lookup_page_cgroup(page);
2805 lock_page_cgroup(pc);
2806 if (PageCgroupUsed(pc)) {
2807 memcg = pc->mem_cgroup;
2808 if (memcg && !css_tryget(&memcg->css))
2810 } else if (PageSwapCache(page)) {
2811 ent.val = page_private(page);
2812 id = lookup_swap_cgroup_id(ent);
2814 memcg = mem_cgroup_lookup(id);
2815 if (memcg && !css_tryget(&memcg->css))
2819 unlock_page_cgroup(pc);
2823 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2825 unsigned int nr_pages,
2826 enum charge_type ctype,
2829 struct page_cgroup *pc = lookup_page_cgroup(page);
2830 struct zone *uninitialized_var(zone);
2831 struct lruvec *lruvec;
2832 bool was_on_lru = false;
2835 lock_page_cgroup(pc);
2836 VM_BUG_ON(PageCgroupUsed(pc));
2838 * we don't need page_cgroup_lock about tail pages, becase they are not
2839 * accessed by any other context at this point.
2843 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2844 * may already be on some other mem_cgroup's LRU. Take care of it.
2847 zone = page_zone(page);
2848 spin_lock_irq(&zone->lru_lock);
2849 if (PageLRU(page)) {
2850 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2852 del_page_from_lru_list(page, lruvec, page_lru(page));
2857 pc->mem_cgroup = memcg;
2859 * We access a page_cgroup asynchronously without lock_page_cgroup().
2860 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2861 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2862 * before USED bit, we need memory barrier here.
2863 * See mem_cgroup_add_lru_list(), etc.
2866 SetPageCgroupUsed(pc);
2870 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2871 VM_BUG_ON(PageLRU(page));
2873 add_page_to_lru_list(page, lruvec, page_lru(page));
2875 spin_unlock_irq(&zone->lru_lock);
2878 if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
2883 mem_cgroup_charge_statistics(memcg, anon, nr_pages);
2884 unlock_page_cgroup(pc);
2887 * "charge_statistics" updated event counter. Then, check it.
2888 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2889 * if they exceeds softlimit.
2891 memcg_check_events(memcg, page);
2894 static DEFINE_MUTEX(set_limit_mutex);
2896 #ifdef CONFIG_MEMCG_KMEM
2897 static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg)
2899 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) &&
2900 (memcg->kmem_account_flags & KMEM_ACCOUNTED_MASK);
2904 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2905 * in the memcg_cache_params struct.
2907 static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
2909 struct kmem_cache *cachep;
2911 VM_BUG_ON(p->is_root_cache);
2912 cachep = p->root_cache;
2913 return cachep->memcg_params->memcg_caches[memcg_cache_id(p->memcg)];
2916 #ifdef CONFIG_SLABINFO
2917 static int mem_cgroup_slabinfo_read(struct cgroup *cont, struct cftype *cft,
2920 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
2921 struct memcg_cache_params *params;
2923 if (!memcg_can_account_kmem(memcg))
2926 print_slabinfo_header(m);
2928 mutex_lock(&memcg->slab_caches_mutex);
2929 list_for_each_entry(params, &memcg->memcg_slab_caches, list)
2930 cache_show(memcg_params_to_cache(params), m);
2931 mutex_unlock(&memcg->slab_caches_mutex);
2937 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size)
2939 struct res_counter *fail_res;
2940 struct mem_cgroup *_memcg;
2944 ret = res_counter_charge(&memcg->kmem, size, &fail_res);
2949 * Conditions under which we can wait for the oom_killer. Those are
2950 * the same conditions tested by the core page allocator
2952 may_oom = (gfp & __GFP_FS) && !(gfp & __GFP_NORETRY);
2955 ret = __mem_cgroup_try_charge(NULL, gfp, size >> PAGE_SHIFT,
2958 if (ret == -EINTR) {
2960 * __mem_cgroup_try_charge() chosed to bypass to root due to
2961 * OOM kill or fatal signal. Since our only options are to
2962 * either fail the allocation or charge it to this cgroup, do
2963 * it as a temporary condition. But we can't fail. From a
2964 * kmem/slab perspective, the cache has already been selected,
2965 * by mem_cgroup_kmem_get_cache(), so it is too late to change
2968 * This condition will only trigger if the task entered
2969 * memcg_charge_kmem in a sane state, but was OOM-killed during
2970 * __mem_cgroup_try_charge() above. Tasks that were already
2971 * dying when the allocation triggers should have been already
2972 * directed to the root cgroup in memcontrol.h
2974 res_counter_charge_nofail(&memcg->res, size, &fail_res);
2975 if (do_swap_account)
2976 res_counter_charge_nofail(&memcg->memsw, size,
2980 res_counter_uncharge(&memcg->kmem, size);
2985 static void memcg_uncharge_kmem(struct mem_cgroup *memcg, u64 size)
2987 res_counter_uncharge(&memcg->res, size);
2988 if (do_swap_account)
2989 res_counter_uncharge(&memcg->memsw, size);
2992 if (res_counter_uncharge(&memcg->kmem, size))
2995 if (memcg_kmem_test_and_clear_dead(memcg))
2996 mem_cgroup_put(memcg);
2999 void memcg_cache_list_add(struct mem_cgroup *memcg, struct kmem_cache *cachep)
3004 mutex_lock(&memcg->slab_caches_mutex);
3005 list_add(&cachep->memcg_params->list, &memcg->memcg_slab_caches);
3006 mutex_unlock(&memcg->slab_caches_mutex);
3010 * helper for acessing a memcg's index. It will be used as an index in the
3011 * child cache array in kmem_cache, and also to derive its name. This function
3012 * will return -1 when this is not a kmem-limited memcg.
3014 int memcg_cache_id(struct mem_cgroup *memcg)
3016 return memcg ? memcg->kmemcg_id : -1;
3020 * This ends up being protected by the set_limit mutex, during normal
3021 * operation, because that is its main call site.
3023 * But when we create a new cache, we can call this as well if its parent
3024 * is kmem-limited. That will have to hold set_limit_mutex as well.
3026 int memcg_update_cache_sizes(struct mem_cgroup *memcg)
3030 num = ida_simple_get(&kmem_limited_groups,
3031 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
3035 * After this point, kmem_accounted (that we test atomically in
3036 * the beginning of this conditional), is no longer 0. This
3037 * guarantees only one process will set the following boolean
3038 * to true. We don't need test_and_set because we're protected
3039 * by the set_limit_mutex anyway.
3041 memcg_kmem_set_activated(memcg);
3043 ret = memcg_update_all_caches(num+1);
3045 ida_simple_remove(&kmem_limited_groups, num);
3046 memcg_kmem_clear_activated(memcg);
3050 memcg->kmemcg_id = num;
3051 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
3052 mutex_init(&memcg->slab_caches_mutex);
3056 static size_t memcg_caches_array_size(int num_groups)
3059 if (num_groups <= 0)
3062 size = 2 * num_groups;
3063 if (size < MEMCG_CACHES_MIN_SIZE)
3064 size = MEMCG_CACHES_MIN_SIZE;
3065 else if (size > MEMCG_CACHES_MAX_SIZE)
3066 size = MEMCG_CACHES_MAX_SIZE;
3072 * We should update the current array size iff all caches updates succeed. This
3073 * can only be done from the slab side. The slab mutex needs to be held when
3076 void memcg_update_array_size(int num)
3078 if (num > memcg_limited_groups_array_size)
3079 memcg_limited_groups_array_size = memcg_caches_array_size(num);
3082 static void kmem_cache_destroy_work_func(struct work_struct *w);
3084 int memcg_update_cache_size(struct kmem_cache *s, int num_groups)
3086 struct memcg_cache_params *cur_params = s->memcg_params;
3088 VM_BUG_ON(s->memcg_params && !s->memcg_params->is_root_cache);
3090 if (num_groups > memcg_limited_groups_array_size) {
3092 ssize_t size = memcg_caches_array_size(num_groups);
3094 size *= sizeof(void *);
3095 size += sizeof(struct memcg_cache_params);
3097 s->memcg_params = kzalloc(size, GFP_KERNEL);
3098 if (!s->memcg_params) {
3099 s->memcg_params = cur_params;
3103 INIT_WORK(&s->memcg_params->destroy,
3104 kmem_cache_destroy_work_func);
3105 s->memcg_params->is_root_cache = true;
3108 * There is the chance it will be bigger than
3109 * memcg_limited_groups_array_size, if we failed an allocation
3110 * in a cache, in which case all caches updated before it, will
3111 * have a bigger array.
3113 * But if that is the case, the data after
3114 * memcg_limited_groups_array_size is certainly unused
3116 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3117 if (!cur_params->memcg_caches[i])
3119 s->memcg_params->memcg_caches[i] =
3120 cur_params->memcg_caches[i];
3124 * Ideally, we would wait until all caches succeed, and only
3125 * then free the old one. But this is not worth the extra
3126 * pointer per-cache we'd have to have for this.
3128 * It is not a big deal if some caches are left with a size
3129 * bigger than the others. And all updates will reset this
3137 int memcg_register_cache(struct mem_cgroup *memcg, struct kmem_cache *s,
3138 struct kmem_cache *root_cache)
3140 size_t size = sizeof(struct memcg_cache_params);
3142 if (!memcg_kmem_enabled())
3146 size += memcg_limited_groups_array_size * sizeof(void *);
3148 s->memcg_params = kzalloc(size, GFP_KERNEL);
3149 if (!s->memcg_params)
3152 INIT_WORK(&s->memcg_params->destroy,
3153 kmem_cache_destroy_work_func);
3155 s->memcg_params->memcg = memcg;
3156 s->memcg_params->root_cache = root_cache;
3158 s->memcg_params->is_root_cache = true;
3163 void memcg_release_cache(struct kmem_cache *s)
3165 struct kmem_cache *root;
3166 struct mem_cgroup *memcg;
3170 * This happens, for instance, when a root cache goes away before we
3173 if (!s->memcg_params)
3176 if (s->memcg_params->is_root_cache)
3179 memcg = s->memcg_params->memcg;
3180 id = memcg_cache_id(memcg);
3182 root = s->memcg_params->root_cache;
3183 root->memcg_params->memcg_caches[id] = NULL;
3184 mem_cgroup_put(memcg);
3186 mutex_lock(&memcg->slab_caches_mutex);
3187 list_del(&s->memcg_params->list);
3188 mutex_unlock(&memcg->slab_caches_mutex);
3191 kfree(s->memcg_params);
3195 * During the creation a new cache, we need to disable our accounting mechanism
3196 * altogether. This is true even if we are not creating, but rather just
3197 * enqueing new caches to be created.
3199 * This is because that process will trigger allocations; some visible, like
3200 * explicit kmallocs to auxiliary data structures, name strings and internal
3201 * cache structures; some well concealed, like INIT_WORK() that can allocate
3202 * objects during debug.
3204 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3205 * to it. This may not be a bounded recursion: since the first cache creation
3206 * failed to complete (waiting on the allocation), we'll just try to create the
3207 * cache again, failing at the same point.
3209 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3210 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3211 * inside the following two functions.
3213 static inline void memcg_stop_kmem_account(void)
3215 VM_BUG_ON(!current->mm);
3216 current->memcg_kmem_skip_account++;
3219 static inline void memcg_resume_kmem_account(void)
3221 VM_BUG_ON(!current->mm);
3222 current->memcg_kmem_skip_account--;
3225 static void kmem_cache_destroy_work_func(struct work_struct *w)
3227 struct kmem_cache *cachep;
3228 struct memcg_cache_params *p;
3230 p = container_of(w, struct memcg_cache_params, destroy);
3232 cachep = memcg_params_to_cache(p);
3235 * If we get down to 0 after shrink, we could delete right away.
3236 * However, memcg_release_pages() already puts us back in the workqueue
3237 * in that case. If we proceed deleting, we'll get a dangling
3238 * reference, and removing the object from the workqueue in that case
3239 * is unnecessary complication. We are not a fast path.
3241 * Note that this case is fundamentally different from racing with
3242 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3243 * kmem_cache_shrink, not only we would be reinserting a dead cache
3244 * into the queue, but doing so from inside the worker racing to
3247 * So if we aren't down to zero, we'll just schedule a worker and try
3250 if (atomic_read(&cachep->memcg_params->nr_pages) != 0) {
3251 kmem_cache_shrink(cachep);
3252 if (atomic_read(&cachep->memcg_params->nr_pages) == 0)
3255 kmem_cache_destroy(cachep);
3258 void mem_cgroup_destroy_cache(struct kmem_cache *cachep)
3260 if (!cachep->memcg_params->dead)
3264 * There are many ways in which we can get here.
3266 * We can get to a memory-pressure situation while the delayed work is
3267 * still pending to run. The vmscan shrinkers can then release all
3268 * cache memory and get us to destruction. If this is the case, we'll
3269 * be executed twice, which is a bug (the second time will execute over
3270 * bogus data). In this case, cancelling the work should be fine.
3272 * But we can also get here from the worker itself, if
3273 * kmem_cache_shrink is enough to shake all the remaining objects and
3274 * get the page count to 0. In this case, we'll deadlock if we try to
3275 * cancel the work (the worker runs with an internal lock held, which
3276 * is the same lock we would hold for cancel_work_sync().)
3278 * Since we can't possibly know who got us here, just refrain from
3279 * running if there is already work pending
3281 if (work_pending(&cachep->memcg_params->destroy))
3284 * We have to defer the actual destroying to a workqueue, because
3285 * we might currently be in a context that cannot sleep.
3287 schedule_work(&cachep->memcg_params->destroy);
3290 static char *memcg_cache_name(struct mem_cgroup *memcg, struct kmem_cache *s)
3293 struct dentry *dentry;
3296 dentry = rcu_dereference(memcg->css.cgroup->dentry);
3299 BUG_ON(dentry == NULL);
3301 name = kasprintf(GFP_KERNEL, "%s(%d:%s)", s->name,
3302 memcg_cache_id(memcg), dentry->d_name.name);
3307 static struct kmem_cache *kmem_cache_dup(struct mem_cgroup *memcg,
3308 struct kmem_cache *s)
3311 struct kmem_cache *new;
3313 name = memcg_cache_name(memcg, s);
3317 new = kmem_cache_create_memcg(memcg, name, s->object_size, s->align,
3318 (s->flags & ~SLAB_PANIC), s->ctor, s);
3321 new->allocflags |= __GFP_KMEMCG;
3328 * This lock protects updaters, not readers. We want readers to be as fast as
3329 * they can, and they will either see NULL or a valid cache value. Our model
3330 * allow them to see NULL, in which case the root memcg will be selected.
3332 * We need this lock because multiple allocations to the same cache from a non
3333 * will span more than one worker. Only one of them can create the cache.
3335 static DEFINE_MUTEX(memcg_cache_mutex);
3336 static struct kmem_cache *memcg_create_kmem_cache(struct mem_cgroup *memcg,
3337 struct kmem_cache *cachep)
3339 struct kmem_cache *new_cachep;
3342 BUG_ON(!memcg_can_account_kmem(memcg));
3344 idx = memcg_cache_id(memcg);
3346 mutex_lock(&memcg_cache_mutex);
3347 new_cachep = cachep->memcg_params->memcg_caches[idx];
3351 new_cachep = kmem_cache_dup(memcg, cachep);
3352 if (new_cachep == NULL) {
3353 new_cachep = cachep;
3357 mem_cgroup_get(memcg);
3358 atomic_set(&new_cachep->memcg_params->nr_pages , 0);
3360 cachep->memcg_params->memcg_caches[idx] = new_cachep;
3362 * the readers won't lock, make sure everybody sees the updated value,
3363 * so they won't put stuff in the queue again for no reason
3367 mutex_unlock(&memcg_cache_mutex);
3371 void kmem_cache_destroy_memcg_children(struct kmem_cache *s)
3373 struct kmem_cache *c;
3376 if (!s->memcg_params)
3378 if (!s->memcg_params->is_root_cache)
3382 * If the cache is being destroyed, we trust that there is no one else
3383 * requesting objects from it. Even if there are, the sanity checks in
3384 * kmem_cache_destroy should caught this ill-case.
3386 * Still, we don't want anyone else freeing memcg_caches under our
3387 * noses, which can happen if a new memcg comes to life. As usual,
3388 * we'll take the set_limit_mutex to protect ourselves against this.
3390 mutex_lock(&set_limit_mutex);
3391 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3392 c = s->memcg_params->memcg_caches[i];
3397 * We will now manually delete the caches, so to avoid races
3398 * we need to cancel all pending destruction workers and
3399 * proceed with destruction ourselves.
3401 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3402 * and that could spawn the workers again: it is likely that
3403 * the cache still have active pages until this very moment.
3404 * This would lead us back to mem_cgroup_destroy_cache.
3406 * But that will not execute at all if the "dead" flag is not
3407 * set, so flip it down to guarantee we are in control.
3409 c->memcg_params->dead = false;
3410 cancel_work_sync(&c->memcg_params->destroy);
3411 kmem_cache_destroy(c);
3413 mutex_unlock(&set_limit_mutex);
3416 struct create_work {
3417 struct mem_cgroup *memcg;
3418 struct kmem_cache *cachep;
3419 struct work_struct work;
3422 static void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3424 struct kmem_cache *cachep;
3425 struct memcg_cache_params *params;
3427 if (!memcg_kmem_is_active(memcg))
3430 mutex_lock(&memcg->slab_caches_mutex);
3431 list_for_each_entry(params, &memcg->memcg_slab_caches, list) {
3432 cachep = memcg_params_to_cache(params);
3433 cachep->memcg_params->dead = true;
3434 schedule_work(&cachep->memcg_params->destroy);
3436 mutex_unlock(&memcg->slab_caches_mutex);
3439 static void memcg_create_cache_work_func(struct work_struct *w)
3441 struct create_work *cw;
3443 cw = container_of(w, struct create_work, work);
3444 memcg_create_kmem_cache(cw->memcg, cw->cachep);
3445 /* Drop the reference gotten when we enqueued. */
3446 css_put(&cw->memcg->css);
3451 * Enqueue the creation of a per-memcg kmem_cache.
3452 * Called with rcu_read_lock.
3454 static void __memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3455 struct kmem_cache *cachep)
3457 struct create_work *cw;
3459 cw = kmalloc(sizeof(struct create_work), GFP_NOWAIT);
3463 /* The corresponding put will be done in the workqueue. */
3464 if (!css_tryget(&memcg->css)) {
3470 cw->cachep = cachep;
3472 INIT_WORK(&cw->work, memcg_create_cache_work_func);
3473 schedule_work(&cw->work);
3476 static void memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3477 struct kmem_cache *cachep)
3480 * We need to stop accounting when we kmalloc, because if the
3481 * corresponding kmalloc cache is not yet created, the first allocation
3482 * in __memcg_create_cache_enqueue will recurse.
3484 * However, it is better to enclose the whole function. Depending on
3485 * the debugging options enabled, INIT_WORK(), for instance, can
3486 * trigger an allocation. This too, will make us recurse. Because at
3487 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3488 * the safest choice is to do it like this, wrapping the whole function.
3490 memcg_stop_kmem_account();
3491 __memcg_create_cache_enqueue(memcg, cachep);
3492 memcg_resume_kmem_account();
3495 * Return the kmem_cache we're supposed to use for a slab allocation.
3496 * We try to use the current memcg's version of the cache.
3498 * If the cache does not exist yet, if we are the first user of it,
3499 * we either create it immediately, if possible, or create it asynchronously
3501 * In the latter case, we will let the current allocation go through with
3502 * the original cache.
3504 * Can't be called in interrupt context or from kernel threads.
3505 * This function needs to be called with rcu_read_lock() held.
3507 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
3510 struct mem_cgroup *memcg;
3513 VM_BUG_ON(!cachep->memcg_params);
3514 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
3516 if (!current->mm || current->memcg_kmem_skip_account)
3520 memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
3523 if (!memcg_can_account_kmem(memcg))
3526 idx = memcg_cache_id(memcg);
3529 * barrier to mare sure we're always seeing the up to date value. The
3530 * code updating memcg_caches will issue a write barrier to match this.
3532 read_barrier_depends();
3533 if (unlikely(cachep->memcg_params->memcg_caches[idx] == NULL)) {
3535 * If we are in a safe context (can wait, and not in interrupt
3536 * context), we could be be predictable and return right away.
3537 * This would guarantee that the allocation being performed
3538 * already belongs in the new cache.
3540 * However, there are some clashes that can arrive from locking.
3541 * For instance, because we acquire the slab_mutex while doing
3542 * kmem_cache_dup, this means no further allocation could happen
3543 * with the slab_mutex held.
3545 * Also, because cache creation issue get_online_cpus(), this
3546 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3547 * that ends up reversed during cpu hotplug. (cpuset allocates
3548 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3549 * better to defer everything.
3551 memcg_create_cache_enqueue(memcg, cachep);
3555 return cachep->memcg_params->memcg_caches[idx];
3557 EXPORT_SYMBOL(__memcg_kmem_get_cache);
3560 * We need to verify if the allocation against current->mm->owner's memcg is
3561 * possible for the given order. But the page is not allocated yet, so we'll
3562 * need a further commit step to do the final arrangements.
3564 * It is possible for the task to switch cgroups in this mean time, so at
3565 * commit time, we can't rely on task conversion any longer. We'll then use
3566 * the handle argument to return to the caller which cgroup we should commit
3567 * against. We could also return the memcg directly and avoid the pointer
3568 * passing, but a boolean return value gives better semantics considering
3569 * the compiled-out case as well.
3571 * Returning true means the allocation is possible.
3574 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3576 struct mem_cgroup *memcg;
3580 memcg = try_get_mem_cgroup_from_mm(current->mm);
3583 * very rare case described in mem_cgroup_from_task. Unfortunately there
3584 * isn't much we can do without complicating this too much, and it would
3585 * be gfp-dependent anyway. Just let it go
3587 if (unlikely(!memcg))
3590 if (!memcg_can_account_kmem(memcg)) {
3591 css_put(&memcg->css);
3595 ret = memcg_charge_kmem(memcg, gfp, PAGE_SIZE << order);
3599 css_put(&memcg->css);
3603 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3606 struct page_cgroup *pc;
3608 VM_BUG_ON(mem_cgroup_is_root(memcg));
3610 /* The page allocation failed. Revert */
3612 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3616 pc = lookup_page_cgroup(page);
3617 lock_page_cgroup(pc);
3618 pc->mem_cgroup = memcg;
3619 SetPageCgroupUsed(pc);
3620 unlock_page_cgroup(pc);
3623 void __memcg_kmem_uncharge_pages(struct page *page, int order)
3625 struct mem_cgroup *memcg = NULL;
3626 struct page_cgroup *pc;
3629 pc = lookup_page_cgroup(page);
3631 * Fast unlocked return. Theoretically might have changed, have to
3632 * check again after locking.
3634 if (!PageCgroupUsed(pc))
3637 lock_page_cgroup(pc);
3638 if (PageCgroupUsed(pc)) {
3639 memcg = pc->mem_cgroup;
3640 ClearPageCgroupUsed(pc);
3642 unlock_page_cgroup(pc);
3645 * We trust that only if there is a memcg associated with the page, it
3646 * is a valid allocation
3651 VM_BUG_ON(mem_cgroup_is_root(memcg));
3652 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3655 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3658 #endif /* CONFIG_MEMCG_KMEM */
3660 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3662 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3664 * Because tail pages are not marked as "used", set it. We're under
3665 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3666 * charge/uncharge will be never happen and move_account() is done under
3667 * compound_lock(), so we don't have to take care of races.
3669 void mem_cgroup_split_huge_fixup(struct page *head)
3671 struct page_cgroup *head_pc = lookup_page_cgroup(head);
3672 struct page_cgroup *pc;
3675 if (mem_cgroup_disabled())
3677 for (i = 1; i < HPAGE_PMD_NR; i++) {
3679 pc->mem_cgroup = head_pc->mem_cgroup;
3680 smp_wmb();/* see __commit_charge() */
3681 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
3684 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3687 * mem_cgroup_move_account - move account of the page
3689 * @nr_pages: number of regular pages (>1 for huge pages)
3690 * @pc: page_cgroup of the page.
3691 * @from: mem_cgroup which the page is moved from.
3692 * @to: mem_cgroup which the page is moved to. @from != @to.
3694 * The caller must confirm following.
3695 * - page is not on LRU (isolate_page() is useful.)
3696 * - compound_lock is held when nr_pages > 1
3698 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3701 static int mem_cgroup_move_account(struct page *page,
3702 unsigned int nr_pages,
3703 struct page_cgroup *pc,
3704 struct mem_cgroup *from,
3705 struct mem_cgroup *to)
3707 unsigned long flags;
3709 bool anon = PageAnon(page);
3711 VM_BUG_ON(from == to);
3712 VM_BUG_ON(PageLRU(page));
3714 * The page is isolated from LRU. So, collapse function
3715 * will not handle this page. But page splitting can happen.
3716 * Do this check under compound_page_lock(). The caller should
3720 if (nr_pages > 1 && !PageTransHuge(page))
3723 lock_page_cgroup(pc);
3726 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3729 move_lock_mem_cgroup(from, &flags);
3731 if (!anon && page_mapped(page)) {
3732 /* Update mapped_file data for mem_cgroup */
3734 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
3735 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
3738 mem_cgroup_charge_statistics(from, anon, -nr_pages);
3740 /* caller should have done css_get */
3741 pc->mem_cgroup = to;
3742 mem_cgroup_charge_statistics(to, anon, nr_pages);
3743 move_unlock_mem_cgroup(from, &flags);
3746 unlock_page_cgroup(pc);
3750 memcg_check_events(to, page);
3751 memcg_check_events(from, page);
3757 * mem_cgroup_move_parent - moves page to the parent group
3758 * @page: the page to move
3759 * @pc: page_cgroup of the page
3760 * @child: page's cgroup
3762 * move charges to its parent or the root cgroup if the group has no
3763 * parent (aka use_hierarchy==0).
3764 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3765 * mem_cgroup_move_account fails) the failure is always temporary and
3766 * it signals a race with a page removal/uncharge or migration. In the
3767 * first case the page is on the way out and it will vanish from the LRU
3768 * on the next attempt and the call should be retried later.
3769 * Isolation from the LRU fails only if page has been isolated from
3770 * the LRU since we looked at it and that usually means either global
3771 * reclaim or migration going on. The page will either get back to the
3773 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3774 * (!PageCgroupUsed) or moved to a different group. The page will
3775 * disappear in the next attempt.
3777 static int mem_cgroup_move_parent(struct page *page,
3778 struct page_cgroup *pc,
3779 struct mem_cgroup *child)
3781 struct mem_cgroup *parent;
3782 unsigned int nr_pages;
3783 unsigned long uninitialized_var(flags);
3786 VM_BUG_ON(mem_cgroup_is_root(child));
3789 if (!get_page_unless_zero(page))
3791 if (isolate_lru_page(page))
3794 nr_pages = hpage_nr_pages(page);
3796 parent = parent_mem_cgroup(child);
3798 * If no parent, move charges to root cgroup.
3801 parent = root_mem_cgroup;
3804 VM_BUG_ON(!PageTransHuge(page));
3805 flags = compound_lock_irqsave(page);
3808 ret = mem_cgroup_move_account(page, nr_pages,
3811 __mem_cgroup_cancel_local_charge(child, nr_pages);
3814 compound_unlock_irqrestore(page, flags);
3815 putback_lru_page(page);
3823 * Charge the memory controller for page usage.
3825 * 0 if the charge was successful
3826 * < 0 if the cgroup is over its limit
3828 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
3829 gfp_t gfp_mask, enum charge_type ctype)
3831 struct mem_cgroup *memcg = NULL;
3832 unsigned int nr_pages = 1;
3836 if (PageTransHuge(page)) {
3837 nr_pages <<= compound_order(page);
3838 VM_BUG_ON(!PageTransHuge(page));
3840 * Never OOM-kill a process for a huge page. The
3841 * fault handler will fall back to regular pages.
3846 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
3849 __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false);
3853 int mem_cgroup_newpage_charge(struct page *page,
3854 struct mm_struct *mm, gfp_t gfp_mask)
3856 if (mem_cgroup_disabled())
3858 VM_BUG_ON(page_mapped(page));
3859 VM_BUG_ON(page->mapping && !PageAnon(page));
3861 return mem_cgroup_charge_common(page, mm, gfp_mask,
3862 MEM_CGROUP_CHARGE_TYPE_ANON);
3866 * While swap-in, try_charge -> commit or cancel, the page is locked.
3867 * And when try_charge() successfully returns, one refcnt to memcg without
3868 * struct page_cgroup is acquired. This refcnt will be consumed by
3869 * "commit()" or removed by "cancel()"
3871 static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm,
3874 struct mem_cgroup **memcgp)
3876 struct mem_cgroup *memcg;
3877 struct page_cgroup *pc;
3880 pc = lookup_page_cgroup(page);
3882 * Every swap fault against a single page tries to charge the
3883 * page, bail as early as possible. shmem_unuse() encounters
3884 * already charged pages, too. The USED bit is protected by
3885 * the page lock, which serializes swap cache removal, which
3886 * in turn serializes uncharging.
3888 if (PageCgroupUsed(pc))
3890 if (!do_swap_account)
3892 memcg = try_get_mem_cgroup_from_page(page);
3896 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
3897 css_put(&memcg->css);
3902 ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
3908 int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page,
3909 gfp_t gfp_mask, struct mem_cgroup **memcgp)
3912 if (mem_cgroup_disabled())
3915 * A racing thread's fault, or swapoff, may have already
3916 * updated the pte, and even removed page from swap cache: in
3917 * those cases unuse_pte()'s pte_same() test will fail; but
3918 * there's also a KSM case which does need to charge the page.
3920 if (!PageSwapCache(page)) {
3923 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, memcgp, true);
3928 return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp);
3931 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
3933 if (mem_cgroup_disabled())
3937 __mem_cgroup_cancel_charge(memcg, 1);
3941 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
3942 enum charge_type ctype)
3944 if (mem_cgroup_disabled())
3949 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
3951 * Now swap is on-memory. This means this page may be
3952 * counted both as mem and swap....double count.
3953 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
3954 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
3955 * may call delete_from_swap_cache() before reach here.
3957 if (do_swap_account && PageSwapCache(page)) {
3958 swp_entry_t ent = {.val = page_private(page)};
3959 mem_cgroup_uncharge_swap(ent);
3963 void mem_cgroup_commit_charge_swapin(struct page *page,
3964 struct mem_cgroup *memcg)
3966 __mem_cgroup_commit_charge_swapin(page, memcg,
3967 MEM_CGROUP_CHARGE_TYPE_ANON);
3970 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
3973 struct mem_cgroup *memcg = NULL;
3974 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
3977 if (mem_cgroup_disabled())
3979 if (PageCompound(page))
3982 if (!PageSwapCache(page))
3983 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
3984 else { /* page is swapcache/shmem */
3985 ret = __mem_cgroup_try_charge_swapin(mm, page,
3988 __mem_cgroup_commit_charge_swapin(page, memcg, type);
3993 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
3994 unsigned int nr_pages,
3995 const enum charge_type ctype)
3997 struct memcg_batch_info *batch = NULL;
3998 bool uncharge_memsw = true;
4000 /* If swapout, usage of swap doesn't decrease */
4001 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
4002 uncharge_memsw = false;
4004 batch = ¤t->memcg_batch;
4006 * In usual, we do css_get() when we remember memcg pointer.
4007 * But in this case, we keep res->usage until end of a series of
4008 * uncharges. Then, it's ok to ignore memcg's refcnt.
4011 batch->memcg = memcg;
4013 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
4014 * In those cases, all pages freed continuously can be expected to be in
4015 * the same cgroup and we have chance to coalesce uncharges.
4016 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
4017 * because we want to do uncharge as soon as possible.
4020 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
4021 goto direct_uncharge;
4024 goto direct_uncharge;
4027 * In typical case, batch->memcg == mem. This means we can
4028 * merge a series of uncharges to an uncharge of res_counter.
4029 * If not, we uncharge res_counter ony by one.
4031 if (batch->memcg != memcg)
4032 goto direct_uncharge;
4033 /* remember freed charge and uncharge it later */
4036 batch->memsw_nr_pages++;
4039 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
4041 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
4042 if (unlikely(batch->memcg != memcg))
4043 memcg_oom_recover(memcg);
4047 * uncharge if !page_mapped(page)
4049 static struct mem_cgroup *
4050 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
4053 struct mem_cgroup *memcg = NULL;
4054 unsigned int nr_pages = 1;
4055 struct page_cgroup *pc;
4058 if (mem_cgroup_disabled())
4061 VM_BUG_ON(PageSwapCache(page));
4063 if (PageTransHuge(page)) {
4064 nr_pages <<= compound_order(page);
4065 VM_BUG_ON(!PageTransHuge(page));
4068 * Check if our page_cgroup is valid
4070 pc = lookup_page_cgroup(page);
4071 if (unlikely(!PageCgroupUsed(pc)))
4074 lock_page_cgroup(pc);
4076 memcg = pc->mem_cgroup;
4078 if (!PageCgroupUsed(pc))
4081 anon = PageAnon(page);
4084 case MEM_CGROUP_CHARGE_TYPE_ANON:
4086 * Generally PageAnon tells if it's the anon statistics to be
4087 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
4088 * used before page reached the stage of being marked PageAnon.
4092 case MEM_CGROUP_CHARGE_TYPE_DROP:
4093 /* See mem_cgroup_prepare_migration() */
4094 if (page_mapped(page))
4097 * Pages under migration may not be uncharged. But
4098 * end_migration() /must/ be the one uncharging the
4099 * unused post-migration page and so it has to call
4100 * here with the migration bit still set. See the
4101 * res_counter handling below.
4103 if (!end_migration && PageCgroupMigration(pc))
4106 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
4107 if (!PageAnon(page)) { /* Shared memory */
4108 if (page->mapping && !page_is_file_cache(page))
4110 } else if (page_mapped(page)) /* Anon */
4117 mem_cgroup_charge_statistics(memcg, anon, -nr_pages);
4119 ClearPageCgroupUsed(pc);
4121 * pc->mem_cgroup is not cleared here. It will be accessed when it's
4122 * freed from LRU. This is safe because uncharged page is expected not
4123 * to be reused (freed soon). Exception is SwapCache, it's handled by
4124 * special functions.
4127 unlock_page_cgroup(pc);
4129 * even after unlock, we have memcg->res.usage here and this memcg
4130 * will never be freed.
4132 memcg_check_events(memcg, page);
4133 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
4134 mem_cgroup_swap_statistics(memcg, true);
4135 mem_cgroup_get(memcg);
4138 * Migration does not charge the res_counter for the
4139 * replacement page, so leave it alone when phasing out the
4140 * page that is unused after the migration.
4142 if (!end_migration && !mem_cgroup_is_root(memcg))
4143 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
4148 unlock_page_cgroup(pc);
4152 void mem_cgroup_uncharge_page(struct page *page)
4155 if (page_mapped(page))
4157 VM_BUG_ON(page->mapping && !PageAnon(page));
4158 if (PageSwapCache(page))
4160 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
4163 void mem_cgroup_uncharge_cache_page(struct page *page)
4165 VM_BUG_ON(page_mapped(page));
4166 VM_BUG_ON(page->mapping);
4167 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
4171 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4172 * In that cases, pages are freed continuously and we can expect pages
4173 * are in the same memcg. All these calls itself limits the number of
4174 * pages freed at once, then uncharge_start/end() is called properly.
4175 * This may be called prural(2) times in a context,
4178 void mem_cgroup_uncharge_start(void)
4180 current->memcg_batch.do_batch++;
4181 /* We can do nest. */
4182 if (current->memcg_batch.do_batch == 1) {
4183 current->memcg_batch.memcg = NULL;
4184 current->memcg_batch.nr_pages = 0;
4185 current->memcg_batch.memsw_nr_pages = 0;
4189 void mem_cgroup_uncharge_end(void)
4191 struct memcg_batch_info *batch = ¤t->memcg_batch;
4193 if (!batch->do_batch)
4197 if (batch->do_batch) /* If stacked, do nothing. */
4203 * This "batch->memcg" is valid without any css_get/put etc...
4204 * bacause we hide charges behind us.
4206 if (batch->nr_pages)
4207 res_counter_uncharge(&batch->memcg->res,
4208 batch->nr_pages * PAGE_SIZE);
4209 if (batch->memsw_nr_pages)
4210 res_counter_uncharge(&batch->memcg->memsw,
4211 batch->memsw_nr_pages * PAGE_SIZE);
4212 memcg_oom_recover(batch->memcg);
4213 /* forget this pointer (for sanity check) */
4214 batch->memcg = NULL;
4219 * called after __delete_from_swap_cache() and drop "page" account.
4220 * memcg information is recorded to swap_cgroup of "ent"
4223 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
4225 struct mem_cgroup *memcg;
4226 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
4228 if (!swapout) /* this was a swap cache but the swap is unused ! */
4229 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
4231 memcg = __mem_cgroup_uncharge_common(page, ctype, false);
4234 * record memcg information, if swapout && memcg != NULL,
4235 * mem_cgroup_get() was called in uncharge().
4237 if (do_swap_account && swapout && memcg)
4238 swap_cgroup_record(ent, css_id(&memcg->css));
4242 #ifdef CONFIG_MEMCG_SWAP
4244 * called from swap_entry_free(). remove record in swap_cgroup and
4245 * uncharge "memsw" account.
4247 void mem_cgroup_uncharge_swap(swp_entry_t ent)
4249 struct mem_cgroup *memcg;
4252 if (!do_swap_account)
4255 id = swap_cgroup_record(ent, 0);
4257 memcg = mem_cgroup_lookup(id);
4260 * We uncharge this because swap is freed.
4261 * This memcg can be obsolete one. We avoid calling css_tryget
4263 if (!mem_cgroup_is_root(memcg))
4264 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
4265 mem_cgroup_swap_statistics(memcg, false);
4266 mem_cgroup_put(memcg);
4272 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4273 * @entry: swap entry to be moved
4274 * @from: mem_cgroup which the entry is moved from
4275 * @to: mem_cgroup which the entry is moved to
4277 * It succeeds only when the swap_cgroup's record for this entry is the same
4278 * as the mem_cgroup's id of @from.
4280 * Returns 0 on success, -EINVAL on failure.
4282 * The caller must have charged to @to, IOW, called res_counter_charge() about
4283 * both res and memsw, and called css_get().
4285 static int mem_cgroup_move_swap_account(swp_entry_t entry,
4286 struct mem_cgroup *from, struct mem_cgroup *to)
4288 unsigned short old_id, new_id;
4290 old_id = css_id(&from->css);
4291 new_id = css_id(&to->css);
4293 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
4294 mem_cgroup_swap_statistics(from, false);
4295 mem_cgroup_swap_statistics(to, true);
4297 * This function is only called from task migration context now.
4298 * It postpones res_counter and refcount handling till the end
4299 * of task migration(mem_cgroup_clear_mc()) for performance
4300 * improvement. But we cannot postpone mem_cgroup_get(to)
4301 * because if the process that has been moved to @to does
4302 * swap-in, the refcount of @to might be decreased to 0.
4310 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
4311 struct mem_cgroup *from, struct mem_cgroup *to)
4318 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4321 void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
4322 struct mem_cgroup **memcgp)
4324 struct mem_cgroup *memcg = NULL;
4325 unsigned int nr_pages = 1;
4326 struct page_cgroup *pc;
4327 enum charge_type ctype;
4331 if (mem_cgroup_disabled())
4334 if (PageTransHuge(page))
4335 nr_pages <<= compound_order(page);
4337 pc = lookup_page_cgroup(page);
4338 lock_page_cgroup(pc);
4339 if (PageCgroupUsed(pc)) {
4340 memcg = pc->mem_cgroup;
4341 css_get(&memcg->css);
4343 * At migrating an anonymous page, its mapcount goes down
4344 * to 0 and uncharge() will be called. But, even if it's fully
4345 * unmapped, migration may fail and this page has to be
4346 * charged again. We set MIGRATION flag here and delay uncharge
4347 * until end_migration() is called
4349 * Corner Case Thinking
4351 * When the old page was mapped as Anon and it's unmap-and-freed
4352 * while migration was ongoing.
4353 * If unmap finds the old page, uncharge() of it will be delayed
4354 * until end_migration(). If unmap finds a new page, it's
4355 * uncharged when it make mapcount to be 1->0. If unmap code
4356 * finds swap_migration_entry, the new page will not be mapped
4357 * and end_migration() will find it(mapcount==0).
4360 * When the old page was mapped but migraion fails, the kernel
4361 * remaps it. A charge for it is kept by MIGRATION flag even
4362 * if mapcount goes down to 0. We can do remap successfully
4363 * without charging it again.
4366 * The "old" page is under lock_page() until the end of
4367 * migration, so, the old page itself will not be swapped-out.
4368 * If the new page is swapped out before end_migraton, our
4369 * hook to usual swap-out path will catch the event.
4372 SetPageCgroupMigration(pc);
4374 unlock_page_cgroup(pc);
4376 * If the page is not charged at this point,
4384 * We charge new page before it's used/mapped. So, even if unlock_page()
4385 * is called before end_migration, we can catch all events on this new
4386 * page. In the case new page is migrated but not remapped, new page's
4387 * mapcount will be finally 0 and we call uncharge in end_migration().
4390 ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
4392 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
4394 * The page is committed to the memcg, but it's not actually
4395 * charged to the res_counter since we plan on replacing the
4396 * old one and only one page is going to be left afterwards.
4398 __mem_cgroup_commit_charge(memcg, newpage, nr_pages, ctype, false);
4401 /* remove redundant charge if migration failed*/
4402 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
4403 struct page *oldpage, struct page *newpage, bool migration_ok)
4405 struct page *used, *unused;
4406 struct page_cgroup *pc;
4412 if (!migration_ok) {
4419 anon = PageAnon(used);
4420 __mem_cgroup_uncharge_common(unused,
4421 anon ? MEM_CGROUP_CHARGE_TYPE_ANON
4422 : MEM_CGROUP_CHARGE_TYPE_CACHE,
4424 css_put(&memcg->css);
4426 * We disallowed uncharge of pages under migration because mapcount
4427 * of the page goes down to zero, temporarly.
4428 * Clear the flag and check the page should be charged.
4430 pc = lookup_page_cgroup(oldpage);
4431 lock_page_cgroup(pc);
4432 ClearPageCgroupMigration(pc);
4433 unlock_page_cgroup(pc);
4436 * If a page is a file cache, radix-tree replacement is very atomic
4437 * and we can skip this check. When it was an Anon page, its mapcount
4438 * goes down to 0. But because we added MIGRATION flage, it's not
4439 * uncharged yet. There are several case but page->mapcount check
4440 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4441 * check. (see prepare_charge() also)
4444 mem_cgroup_uncharge_page(used);
4448 * At replace page cache, newpage is not under any memcg but it's on
4449 * LRU. So, this function doesn't touch res_counter but handles LRU
4450 * in correct way. Both pages are locked so we cannot race with uncharge.
4452 void mem_cgroup_replace_page_cache(struct page *oldpage,
4453 struct page *newpage)
4455 struct mem_cgroup *memcg = NULL;
4456 struct page_cgroup *pc;
4457 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4459 if (mem_cgroup_disabled())
4462 pc = lookup_page_cgroup(oldpage);
4463 /* fix accounting on old pages */
4464 lock_page_cgroup(pc);
4465 if (PageCgroupUsed(pc)) {
4466 memcg = pc->mem_cgroup;
4467 mem_cgroup_charge_statistics(memcg, false, -1);
4468 ClearPageCgroupUsed(pc);
4470 unlock_page_cgroup(pc);
4473 * When called from shmem_replace_page(), in some cases the
4474 * oldpage has already been charged, and in some cases not.
4479 * Even if newpage->mapping was NULL before starting replacement,
4480 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4481 * LRU while we overwrite pc->mem_cgroup.
4483 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
4486 #ifdef CONFIG_DEBUG_VM
4487 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
4489 struct page_cgroup *pc;
4491 pc = lookup_page_cgroup(page);
4493 * Can be NULL while feeding pages into the page allocator for
4494 * the first time, i.e. during boot or memory hotplug;
4495 * or when mem_cgroup_disabled().
4497 if (likely(pc) && PageCgroupUsed(pc))
4502 bool mem_cgroup_bad_page_check(struct page *page)
4504 if (mem_cgroup_disabled())
4507 return lookup_page_cgroup_used(page) != NULL;
4510 void mem_cgroup_print_bad_page(struct page *page)
4512 struct page_cgroup *pc;
4514 pc = lookup_page_cgroup_used(page);
4516 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4517 pc, pc->flags, pc->mem_cgroup);
4522 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
4523 unsigned long long val)
4526 u64 memswlimit, memlimit;
4528 int children = mem_cgroup_count_children(memcg);
4529 u64 curusage, oldusage;
4533 * For keeping hierarchical_reclaim simple, how long we should retry
4534 * is depends on callers. We set our retry-count to be function
4535 * of # of children which we should visit in this loop.
4537 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
4539 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4542 while (retry_count) {
4543 if (signal_pending(current)) {
4548 * Rather than hide all in some function, I do this in
4549 * open coded manner. You see what this really does.
4550 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4552 mutex_lock(&set_limit_mutex);
4553 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4554 if (memswlimit < val) {
4556 mutex_unlock(&set_limit_mutex);
4560 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4564 ret = res_counter_set_limit(&memcg->res, val);
4566 if (memswlimit == val)
4567 memcg->memsw_is_minimum = true;
4569 memcg->memsw_is_minimum = false;
4571 mutex_unlock(&set_limit_mutex);
4576 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4577 MEM_CGROUP_RECLAIM_SHRINK);
4578 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4579 /* Usage is reduced ? */
4580 if (curusage >= oldusage)
4583 oldusage = curusage;
4585 if (!ret && enlarge)
4586 memcg_oom_recover(memcg);
4591 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
4592 unsigned long long val)
4595 u64 memlimit, memswlimit, oldusage, curusage;
4596 int children = mem_cgroup_count_children(memcg);
4600 /* see mem_cgroup_resize_res_limit */
4601 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
4602 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4603 while (retry_count) {
4604 if (signal_pending(current)) {
4609 * Rather than hide all in some function, I do this in
4610 * open coded manner. You see what this really does.
4611 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4613 mutex_lock(&set_limit_mutex);
4614 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4615 if (memlimit > val) {
4617 mutex_unlock(&set_limit_mutex);
4620 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4621 if (memswlimit < val)
4623 ret = res_counter_set_limit(&memcg->memsw, val);
4625 if (memlimit == val)
4626 memcg->memsw_is_minimum = true;
4628 memcg->memsw_is_minimum = false;
4630 mutex_unlock(&set_limit_mutex);
4635 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4636 MEM_CGROUP_RECLAIM_NOSWAP |
4637 MEM_CGROUP_RECLAIM_SHRINK);
4638 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4639 /* Usage is reduced ? */
4640 if (curusage >= oldusage)
4643 oldusage = curusage;
4645 if (!ret && enlarge)
4646 memcg_oom_recover(memcg);
4650 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
4652 unsigned long *total_scanned)
4654 unsigned long nr_reclaimed = 0;
4655 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
4656 unsigned long reclaimed;
4658 struct mem_cgroup_tree_per_zone *mctz;
4659 unsigned long long excess;
4660 unsigned long nr_scanned;
4665 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
4667 * This loop can run a while, specially if mem_cgroup's continuously
4668 * keep exceeding their soft limit and putting the system under
4675 mz = mem_cgroup_largest_soft_limit_node(mctz);
4680 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
4681 gfp_mask, &nr_scanned);
4682 nr_reclaimed += reclaimed;
4683 *total_scanned += nr_scanned;
4684 spin_lock(&mctz->lock);
4687 * If we failed to reclaim anything from this memory cgroup
4688 * it is time to move on to the next cgroup
4694 * Loop until we find yet another one.
4696 * By the time we get the soft_limit lock
4697 * again, someone might have aded the
4698 * group back on the RB tree. Iterate to
4699 * make sure we get a different mem.
4700 * mem_cgroup_largest_soft_limit_node returns
4701 * NULL if no other cgroup is present on
4705 __mem_cgroup_largest_soft_limit_node(mctz);
4707 css_put(&next_mz->memcg->css);
4708 else /* next_mz == NULL or other memcg */
4712 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
4713 excess = res_counter_soft_limit_excess(&mz->memcg->res);
4715 * One school of thought says that we should not add
4716 * back the node to the tree if reclaim returns 0.
4717 * But our reclaim could return 0, simply because due
4718 * to priority we are exposing a smaller subset of
4719 * memory to reclaim from. Consider this as a longer
4722 /* If excess == 0, no tree ops */
4723 __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess);
4724 spin_unlock(&mctz->lock);
4725 css_put(&mz->memcg->css);
4728 * Could not reclaim anything and there are no more
4729 * mem cgroups to try or we seem to be looping without
4730 * reclaiming anything.
4732 if (!nr_reclaimed &&
4734 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
4736 } while (!nr_reclaimed);
4738 css_put(&next_mz->memcg->css);
4739 return nr_reclaimed;
4743 * mem_cgroup_force_empty_list - clears LRU of a group
4744 * @memcg: group to clear
4747 * @lru: lru to to clear
4749 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4750 * reclaim the pages page themselves - pages are moved to the parent (or root)
4753 static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
4754 int node, int zid, enum lru_list lru)
4756 struct lruvec *lruvec;
4757 unsigned long flags;
4758 struct list_head *list;
4762 zone = &NODE_DATA(node)->node_zones[zid];
4763 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
4764 list = &lruvec->lists[lru];
4768 struct page_cgroup *pc;
4771 spin_lock_irqsave(&zone->lru_lock, flags);
4772 if (list_empty(list)) {
4773 spin_unlock_irqrestore(&zone->lru_lock, flags);
4776 page = list_entry(list->prev, struct page, lru);
4778 list_move(&page->lru, list);
4780 spin_unlock_irqrestore(&zone->lru_lock, flags);
4783 spin_unlock_irqrestore(&zone->lru_lock, flags);
4785 pc = lookup_page_cgroup(page);
4787 if (mem_cgroup_move_parent(page, pc, memcg)) {
4788 /* found lock contention or "pc" is obsolete. */
4793 } while (!list_empty(list));
4797 * make mem_cgroup's charge to be 0 if there is no task by moving
4798 * all the charges and pages to the parent.
4799 * This enables deleting this mem_cgroup.
4801 * Caller is responsible for holding css reference on the memcg.
4803 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
4809 /* This is for making all *used* pages to be on LRU. */
4810 lru_add_drain_all();
4811 drain_all_stock_sync(memcg);
4812 mem_cgroup_start_move(memcg);
4813 for_each_node_state(node, N_MEMORY) {
4814 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4817 mem_cgroup_force_empty_list(memcg,
4822 mem_cgroup_end_move(memcg);
4823 memcg_oom_recover(memcg);
4827 * Kernel memory may not necessarily be trackable to a specific
4828 * process. So they are not migrated, and therefore we can't
4829 * expect their value to drop to 0 here.
4830 * Having res filled up with kmem only is enough.
4832 * This is a safety check because mem_cgroup_force_empty_list
4833 * could have raced with mem_cgroup_replace_page_cache callers
4834 * so the lru seemed empty but the page could have been added
4835 * right after the check. RES_USAGE should be safe as we always
4836 * charge before adding to the LRU.
4838 usage = res_counter_read_u64(&memcg->res, RES_USAGE) -
4839 res_counter_read_u64(&memcg->kmem, RES_USAGE);
4840 } while (usage > 0);
4844 * This mainly exists for tests during the setting of set of use_hierarchy.
4845 * Since this is the very setting we are changing, the current hierarchy value
4848 static inline bool __memcg_has_children(struct mem_cgroup *memcg)
4852 /* bounce at first found */
4853 cgroup_for_each_child(pos, memcg->css.cgroup)
4859 * Must be called with memcg_create_mutex held, unless the cgroup is guaranteed
4860 * to be already dead (as in mem_cgroup_force_empty, for instance). This is
4861 * from mem_cgroup_count_children(), in the sense that we don't really care how
4862 * many children we have; we only need to know if we have any. It also counts
4863 * any memcg without hierarchy as infertile.
4865 static inline bool memcg_has_children(struct mem_cgroup *memcg)
4867 return memcg->use_hierarchy && __memcg_has_children(memcg);
4871 * Reclaims as many pages from the given memcg as possible and moves
4872 * the rest to the parent.
4874 * Caller is responsible for holding css reference for memcg.
4876 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
4878 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
4879 struct cgroup *cgrp = memcg->css.cgroup;
4881 /* returns EBUSY if there is a task or if we come here twice. */
4882 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
4885 /* we call try-to-free pages for make this cgroup empty */
4886 lru_add_drain_all();
4887 /* try to free all pages in this cgroup */
4888 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
4891 if (signal_pending(current))
4894 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
4898 /* maybe some writeback is necessary */
4899 congestion_wait(BLK_RW_ASYNC, HZ/10);
4904 mem_cgroup_reparent_charges(memcg);
4909 static int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
4911 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4914 if (mem_cgroup_is_root(memcg))
4916 css_get(&memcg->css);
4917 ret = mem_cgroup_force_empty(memcg);
4918 css_put(&memcg->css);
4924 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
4926 return mem_cgroup_from_cont(cont)->use_hierarchy;
4929 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
4933 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4934 struct cgroup *parent = cont->parent;
4935 struct mem_cgroup *parent_memcg = NULL;
4938 parent_memcg = mem_cgroup_from_cont(parent);
4940 mutex_lock(&memcg_create_mutex);
4942 if (memcg->use_hierarchy == val)
4946 * If parent's use_hierarchy is set, we can't make any modifications
4947 * in the child subtrees. If it is unset, then the change can
4948 * occur, provided the current cgroup has no children.
4950 * For the root cgroup, parent_mem is NULL, we allow value to be
4951 * set if there are no children.
4953 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
4954 (val == 1 || val == 0)) {
4955 if (!__memcg_has_children(memcg))
4956 memcg->use_hierarchy = val;
4963 mutex_unlock(&memcg_create_mutex);
4969 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
4970 enum mem_cgroup_stat_index idx)
4972 struct mem_cgroup *iter;
4975 /* Per-cpu values can be negative, use a signed accumulator */
4976 for_each_mem_cgroup_tree(iter, memcg)
4977 val += mem_cgroup_read_stat(iter, idx);
4979 if (val < 0) /* race ? */
4984 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
4988 if (!mem_cgroup_is_root(memcg)) {
4990 return res_counter_read_u64(&memcg->res, RES_USAGE);
4992 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
4995 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
4996 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
4999 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
5001 return val << PAGE_SHIFT;
5004 static ssize_t mem_cgroup_read(struct cgroup *cont, struct cftype *cft,
5005 struct file *file, char __user *buf,
5006 size_t nbytes, loff_t *ppos)
5008 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5014 type = MEMFILE_TYPE(cft->private);
5015 name = MEMFILE_ATTR(cft->private);
5017 if (!do_swap_account && type == _MEMSWAP)
5022 if (name == RES_USAGE)
5023 val = mem_cgroup_usage(memcg, false);
5025 val = res_counter_read_u64(&memcg->res, name);
5028 if (name == RES_USAGE)
5029 val = mem_cgroup_usage(memcg, true);
5031 val = res_counter_read_u64(&memcg->memsw, name);
5034 val = res_counter_read_u64(&memcg->kmem, name);
5040 len = scnprintf(str, sizeof(str), "%llu\n", (unsigned long long)val);
5041 return simple_read_from_buffer(buf, nbytes, ppos, str, len);
5044 static int memcg_update_kmem_limit(struct cgroup *cont, u64 val)
5047 #ifdef CONFIG_MEMCG_KMEM
5048 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5050 * For simplicity, we won't allow this to be disabled. It also can't
5051 * be changed if the cgroup has children already, or if tasks had
5054 * If tasks join before we set the limit, a person looking at
5055 * kmem.usage_in_bytes will have no way to determine when it took
5056 * place, which makes the value quite meaningless.
5058 * After it first became limited, changes in the value of the limit are
5059 * of course permitted.
5061 mutex_lock(&memcg_create_mutex);
5062 mutex_lock(&set_limit_mutex);
5063 if (!memcg->kmem_account_flags && val != RESOURCE_MAX) {
5064 if (cgroup_task_count(cont) || memcg_has_children(memcg)) {
5068 ret = res_counter_set_limit(&memcg->kmem, val);
5071 ret = memcg_update_cache_sizes(memcg);
5073 res_counter_set_limit(&memcg->kmem, RESOURCE_MAX);
5076 static_key_slow_inc(&memcg_kmem_enabled_key);
5078 * setting the active bit after the inc will guarantee no one
5079 * starts accounting before all call sites are patched
5081 memcg_kmem_set_active(memcg);
5084 * kmem charges can outlive the cgroup. In the case of slab
5085 * pages, for instance, a page contain objects from various
5086 * processes, so it is unfeasible to migrate them away. We
5087 * need to reference count the memcg because of that.
5089 mem_cgroup_get(memcg);
5091 ret = res_counter_set_limit(&memcg->kmem, val);
5093 mutex_unlock(&set_limit_mutex);
5094 mutex_unlock(&memcg_create_mutex);
5099 #ifdef CONFIG_MEMCG_KMEM
5100 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
5103 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5107 memcg->kmem_account_flags = parent->kmem_account_flags;
5109 * When that happen, we need to disable the static branch only on those
5110 * memcgs that enabled it. To achieve this, we would be forced to
5111 * complicate the code by keeping track of which memcgs were the ones
5112 * that actually enabled limits, and which ones got it from its
5115 * It is a lot simpler just to do static_key_slow_inc() on every child
5116 * that is accounted.
5118 if (!memcg_kmem_is_active(memcg))
5122 * destroy(), called if we fail, will issue static_key_slow_inc() and
5123 * mem_cgroup_put() if kmem is enabled. We have to either call them
5124 * unconditionally, or clear the KMEM_ACTIVE flag. I personally find
5125 * this more consistent, since it always leads to the same destroy path
5127 mem_cgroup_get(memcg);
5128 static_key_slow_inc(&memcg_kmem_enabled_key);
5130 mutex_lock(&set_limit_mutex);
5131 ret = memcg_update_cache_sizes(memcg);
5132 mutex_unlock(&set_limit_mutex);
5136 #endif /* CONFIG_MEMCG_KMEM */
5139 * The user of this function is...
5142 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
5145 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5148 unsigned long long val;
5151 type = MEMFILE_TYPE(cft->private);
5152 name = MEMFILE_ATTR(cft->private);
5154 if (!do_swap_account && type == _MEMSWAP)
5159 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
5163 /* This function does all necessary parse...reuse it */
5164 ret = res_counter_memparse_write_strategy(buffer, &val);
5168 ret = mem_cgroup_resize_limit(memcg, val);
5169 else if (type == _MEMSWAP)
5170 ret = mem_cgroup_resize_memsw_limit(memcg, val);
5171 else if (type == _KMEM)
5172 ret = memcg_update_kmem_limit(cont, val);
5176 case RES_SOFT_LIMIT:
5177 ret = res_counter_memparse_write_strategy(buffer, &val);
5181 * For memsw, soft limits are hard to implement in terms
5182 * of semantics, for now, we support soft limits for
5183 * control without swap
5186 ret = res_counter_set_soft_limit(&memcg->res, val);
5191 ret = -EINVAL; /* should be BUG() ? */
5197 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
5198 unsigned long long *mem_limit, unsigned long long *memsw_limit)
5200 struct cgroup *cgroup;
5201 unsigned long long min_limit, min_memsw_limit, tmp;
5203 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
5204 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5205 cgroup = memcg->css.cgroup;
5206 if (!memcg->use_hierarchy)
5209 while (cgroup->parent) {
5210 cgroup = cgroup->parent;
5211 memcg = mem_cgroup_from_cont(cgroup);
5212 if (!memcg->use_hierarchy)
5214 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
5215 min_limit = min(min_limit, tmp);
5216 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5217 min_memsw_limit = min(min_memsw_limit, tmp);
5220 *mem_limit = min_limit;
5221 *memsw_limit = min_memsw_limit;
5224 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
5226 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5230 type = MEMFILE_TYPE(event);
5231 name = MEMFILE_ATTR(event);
5233 if (!do_swap_account && type == _MEMSWAP)
5239 res_counter_reset_max(&memcg->res);
5240 else if (type == _MEMSWAP)
5241 res_counter_reset_max(&memcg->memsw);
5242 else if (type == _KMEM)
5243 res_counter_reset_max(&memcg->kmem);
5249 res_counter_reset_failcnt(&memcg->res);
5250 else if (type == _MEMSWAP)
5251 res_counter_reset_failcnt(&memcg->memsw);
5252 else if (type == _KMEM)
5253 res_counter_reset_failcnt(&memcg->kmem);
5262 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
5265 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
5269 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
5270 struct cftype *cft, u64 val)
5272 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5274 if (val >= (1 << NR_MOVE_TYPE))
5278 * No kind of locking is needed in here, because ->can_attach() will
5279 * check this value once in the beginning of the process, and then carry
5280 * on with stale data. This means that changes to this value will only
5281 * affect task migrations starting after the change.
5283 memcg->move_charge_at_immigrate = val;
5287 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
5288 struct cftype *cft, u64 val)
5295 static int memcg_numa_stat_show(struct cgroup *cont, struct cftype *cft,
5299 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
5300 unsigned long node_nr;
5301 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5303 total_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL);
5304 seq_printf(m, "total=%lu", total_nr);
5305 for_each_node_state(nid, N_MEMORY) {
5306 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL);
5307 seq_printf(m, " N%d=%lu", nid, node_nr);
5311 file_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_FILE);
5312 seq_printf(m, "file=%lu", file_nr);
5313 for_each_node_state(nid, N_MEMORY) {
5314 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5316 seq_printf(m, " N%d=%lu", nid, node_nr);
5320 anon_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_ANON);
5321 seq_printf(m, "anon=%lu", anon_nr);
5322 for_each_node_state(nid, N_MEMORY) {
5323 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5325 seq_printf(m, " N%d=%lu", nid, node_nr);
5329 unevictable_nr = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
5330 seq_printf(m, "unevictable=%lu", unevictable_nr);
5331 for_each_node_state(nid, N_MEMORY) {
5332 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5333 BIT(LRU_UNEVICTABLE));
5334 seq_printf(m, " N%d=%lu", nid, node_nr);
5339 #endif /* CONFIG_NUMA */
5341 static inline void mem_cgroup_lru_names_not_uptodate(void)
5343 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
5346 static int memcg_stat_show(struct cgroup *cont, struct cftype *cft,
5349 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5350 struct mem_cgroup *mi;
5353 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5354 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5356 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
5357 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
5360 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
5361 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
5362 mem_cgroup_read_events(memcg, i));
5364 for (i = 0; i < NR_LRU_LISTS; i++)
5365 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
5366 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
5368 /* Hierarchical information */
5370 unsigned long long limit, memsw_limit;
5371 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
5372 seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
5373 if (do_swap_account)
5374 seq_printf(m, "hierarchical_memsw_limit %llu\n",
5378 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5381 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5383 for_each_mem_cgroup_tree(mi, memcg)
5384 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
5385 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
5388 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
5389 unsigned long long val = 0;
5391 for_each_mem_cgroup_tree(mi, memcg)
5392 val += mem_cgroup_read_events(mi, i);
5393 seq_printf(m, "total_%s %llu\n",
5394 mem_cgroup_events_names[i], val);
5397 for (i = 0; i < NR_LRU_LISTS; i++) {
5398 unsigned long long val = 0;
5400 for_each_mem_cgroup_tree(mi, memcg)
5401 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
5402 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
5405 #ifdef CONFIG_DEBUG_VM
5408 struct mem_cgroup_per_zone *mz;
5409 struct zone_reclaim_stat *rstat;
5410 unsigned long recent_rotated[2] = {0, 0};
5411 unsigned long recent_scanned[2] = {0, 0};
5413 for_each_online_node(nid)
5414 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
5415 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
5416 rstat = &mz->lruvec.reclaim_stat;
5418 recent_rotated[0] += rstat->recent_rotated[0];
5419 recent_rotated[1] += rstat->recent_rotated[1];
5420 recent_scanned[0] += rstat->recent_scanned[0];
5421 recent_scanned[1] += rstat->recent_scanned[1];
5423 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
5424 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
5425 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
5426 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
5433 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
5435 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5437 return mem_cgroup_swappiness(memcg);
5440 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
5443 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5444 struct mem_cgroup *parent;
5449 if (cgrp->parent == NULL)
5452 parent = mem_cgroup_from_cont(cgrp->parent);
5454 mutex_lock(&memcg_create_mutex);
5456 /* If under hierarchy, only empty-root can set this value */
5457 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5458 mutex_unlock(&memcg_create_mutex);
5462 memcg->swappiness = val;
5464 mutex_unlock(&memcg_create_mutex);
5469 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
5471 struct mem_cgroup_threshold_ary *t;
5477 t = rcu_dereference(memcg->thresholds.primary);
5479 t = rcu_dereference(memcg->memsw_thresholds.primary);
5484 usage = mem_cgroup_usage(memcg, swap);
5487 * current_threshold points to threshold just below or equal to usage.
5488 * If it's not true, a threshold was crossed after last
5489 * call of __mem_cgroup_threshold().
5491 i = t->current_threshold;
5494 * Iterate backward over array of thresholds starting from
5495 * current_threshold and check if a threshold is crossed.
5496 * If none of thresholds below usage is crossed, we read
5497 * only one element of the array here.
5499 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
5500 eventfd_signal(t->entries[i].eventfd, 1);
5502 /* i = current_threshold + 1 */
5506 * Iterate forward over array of thresholds starting from
5507 * current_threshold+1 and check if a threshold is crossed.
5508 * If none of thresholds above usage is crossed, we read
5509 * only one element of the array here.
5511 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
5512 eventfd_signal(t->entries[i].eventfd, 1);
5514 /* Update current_threshold */
5515 t->current_threshold = i - 1;
5520 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
5523 __mem_cgroup_threshold(memcg, false);
5524 if (do_swap_account)
5525 __mem_cgroup_threshold(memcg, true);
5527 memcg = parent_mem_cgroup(memcg);
5531 static int compare_thresholds(const void *a, const void *b)
5533 const struct mem_cgroup_threshold *_a = a;
5534 const struct mem_cgroup_threshold *_b = b;
5536 return _a->threshold - _b->threshold;
5539 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
5541 struct mem_cgroup_eventfd_list *ev;
5543 list_for_each_entry(ev, &memcg->oom_notify, list)
5544 eventfd_signal(ev->eventfd, 1);
5548 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
5550 struct mem_cgroup *iter;
5552 for_each_mem_cgroup_tree(iter, memcg)
5553 mem_cgroup_oom_notify_cb(iter);
5556 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
5557 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5559 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5560 struct mem_cgroup_thresholds *thresholds;
5561 struct mem_cgroup_threshold_ary *new;
5562 enum res_type type = MEMFILE_TYPE(cft->private);
5563 u64 threshold, usage;
5566 ret = res_counter_memparse_write_strategy(args, &threshold);
5570 mutex_lock(&memcg->thresholds_lock);
5573 thresholds = &memcg->thresholds;
5574 else if (type == _MEMSWAP)
5575 thresholds = &memcg->memsw_thresholds;
5579 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5581 /* Check if a threshold crossed before adding a new one */
5582 if (thresholds->primary)
5583 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5585 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
5587 /* Allocate memory for new array of thresholds */
5588 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
5596 /* Copy thresholds (if any) to new array */
5597 if (thresholds->primary) {
5598 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
5599 sizeof(struct mem_cgroup_threshold));
5602 /* Add new threshold */
5603 new->entries[size - 1].eventfd = eventfd;
5604 new->entries[size - 1].threshold = threshold;
5606 /* Sort thresholds. Registering of new threshold isn't time-critical */
5607 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
5608 compare_thresholds, NULL);
5610 /* Find current threshold */
5611 new->current_threshold = -1;
5612 for (i = 0; i < size; i++) {
5613 if (new->entries[i].threshold <= usage) {
5615 * new->current_threshold will not be used until
5616 * rcu_assign_pointer(), so it's safe to increment
5619 ++new->current_threshold;
5624 /* Free old spare buffer and save old primary buffer as spare */
5625 kfree(thresholds->spare);
5626 thresholds->spare = thresholds->primary;
5628 rcu_assign_pointer(thresholds->primary, new);
5630 /* To be sure that nobody uses thresholds */
5634 mutex_unlock(&memcg->thresholds_lock);
5639 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
5640 struct cftype *cft, struct eventfd_ctx *eventfd)
5642 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5643 struct mem_cgroup_thresholds *thresholds;
5644 struct mem_cgroup_threshold_ary *new;
5645 enum res_type type = MEMFILE_TYPE(cft->private);
5649 mutex_lock(&memcg->thresholds_lock);
5651 thresholds = &memcg->thresholds;
5652 else if (type == _MEMSWAP)
5653 thresholds = &memcg->memsw_thresholds;
5657 if (!thresholds->primary)
5660 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5662 /* Check if a threshold crossed before removing */
5663 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5665 /* Calculate new number of threshold */
5667 for (i = 0; i < thresholds->primary->size; i++) {
5668 if (thresholds->primary->entries[i].eventfd != eventfd)
5672 new = thresholds->spare;
5674 /* Set thresholds array to NULL if we don't have thresholds */
5683 /* Copy thresholds and find current threshold */
5684 new->current_threshold = -1;
5685 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
5686 if (thresholds->primary->entries[i].eventfd == eventfd)
5689 new->entries[j] = thresholds->primary->entries[i];
5690 if (new->entries[j].threshold <= usage) {
5692 * new->current_threshold will not be used
5693 * until rcu_assign_pointer(), so it's safe to increment
5696 ++new->current_threshold;
5702 /* Swap primary and spare array */
5703 thresholds->spare = thresholds->primary;
5704 /* If all events are unregistered, free the spare array */
5706 kfree(thresholds->spare);
5707 thresholds->spare = NULL;
5710 rcu_assign_pointer(thresholds->primary, new);
5712 /* To be sure that nobody uses thresholds */
5715 mutex_unlock(&memcg->thresholds_lock);
5718 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
5719 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5721 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5722 struct mem_cgroup_eventfd_list *event;
5723 enum res_type type = MEMFILE_TYPE(cft->private);
5725 BUG_ON(type != _OOM_TYPE);
5726 event = kmalloc(sizeof(*event), GFP_KERNEL);
5730 spin_lock(&memcg_oom_lock);
5732 event->eventfd = eventfd;
5733 list_add(&event->list, &memcg->oom_notify);
5735 /* already in OOM ? */
5736 if (atomic_read(&memcg->under_oom))
5737 eventfd_signal(eventfd, 1);
5738 spin_unlock(&memcg_oom_lock);
5743 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
5744 struct cftype *cft, struct eventfd_ctx *eventfd)
5746 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5747 struct mem_cgroup_eventfd_list *ev, *tmp;
5748 enum res_type type = MEMFILE_TYPE(cft->private);
5750 BUG_ON(type != _OOM_TYPE);
5752 spin_lock(&memcg_oom_lock);
5754 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
5755 if (ev->eventfd == eventfd) {
5756 list_del(&ev->list);
5761 spin_unlock(&memcg_oom_lock);
5764 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
5765 struct cftype *cft, struct cgroup_map_cb *cb)
5767 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5769 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
5771 if (atomic_read(&memcg->under_oom))
5772 cb->fill(cb, "under_oom", 1);
5774 cb->fill(cb, "under_oom", 0);
5778 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
5779 struct cftype *cft, u64 val)
5781 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5782 struct mem_cgroup *parent;
5784 /* cannot set to root cgroup and only 0 and 1 are allowed */
5785 if (!cgrp->parent || !((val == 0) || (val == 1)))
5788 parent = mem_cgroup_from_cont(cgrp->parent);
5790 mutex_lock(&memcg_create_mutex);
5791 /* oom-kill-disable is a flag for subhierarchy. */
5792 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5793 mutex_unlock(&memcg_create_mutex);
5796 memcg->oom_kill_disable = val;
5798 memcg_oom_recover(memcg);
5799 mutex_unlock(&memcg_create_mutex);
5803 #ifdef CONFIG_MEMCG_KMEM
5804 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5808 memcg->kmemcg_id = -1;
5809 ret = memcg_propagate_kmem(memcg);
5813 return mem_cgroup_sockets_init(memcg, ss);
5816 static void kmem_cgroup_destroy(struct mem_cgroup *memcg)
5818 mem_cgroup_sockets_destroy(memcg);
5820 memcg_kmem_mark_dead(memcg);
5822 if (res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0)
5826 * Charges already down to 0, undo mem_cgroup_get() done in the charge
5827 * path here, being careful not to race with memcg_uncharge_kmem: it is
5828 * possible that the charges went down to 0 between mark_dead and the
5829 * res_counter read, so in that case, we don't need the put
5831 if (memcg_kmem_test_and_clear_dead(memcg))
5832 mem_cgroup_put(memcg);
5835 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5840 static void kmem_cgroup_destroy(struct mem_cgroup *memcg)
5845 static struct cftype mem_cgroup_files[] = {
5847 .name = "usage_in_bytes",
5848 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
5849 .read = mem_cgroup_read,
5850 .register_event = mem_cgroup_usage_register_event,
5851 .unregister_event = mem_cgroup_usage_unregister_event,
5854 .name = "max_usage_in_bytes",
5855 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
5856 .trigger = mem_cgroup_reset,
5857 .read = mem_cgroup_read,
5860 .name = "limit_in_bytes",
5861 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
5862 .write_string = mem_cgroup_write,
5863 .read = mem_cgroup_read,
5866 .name = "soft_limit_in_bytes",
5867 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
5868 .write_string = mem_cgroup_write,
5869 .read = mem_cgroup_read,
5873 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
5874 .trigger = mem_cgroup_reset,
5875 .read = mem_cgroup_read,
5879 .read_seq_string = memcg_stat_show,
5882 .name = "force_empty",
5883 .trigger = mem_cgroup_force_empty_write,
5886 .name = "use_hierarchy",
5887 .write_u64 = mem_cgroup_hierarchy_write,
5888 .read_u64 = mem_cgroup_hierarchy_read,
5891 .name = "swappiness",
5892 .read_u64 = mem_cgroup_swappiness_read,
5893 .write_u64 = mem_cgroup_swappiness_write,
5896 .name = "move_charge_at_immigrate",
5897 .read_u64 = mem_cgroup_move_charge_read,
5898 .write_u64 = mem_cgroup_move_charge_write,
5901 .name = "oom_control",
5902 .read_map = mem_cgroup_oom_control_read,
5903 .write_u64 = mem_cgroup_oom_control_write,
5904 .register_event = mem_cgroup_oom_register_event,
5905 .unregister_event = mem_cgroup_oom_unregister_event,
5906 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
5910 .name = "numa_stat",
5911 .read_seq_string = memcg_numa_stat_show,
5914 #ifdef CONFIG_MEMCG_KMEM
5916 .name = "kmem.limit_in_bytes",
5917 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
5918 .write_string = mem_cgroup_write,
5919 .read = mem_cgroup_read,
5922 .name = "kmem.usage_in_bytes",
5923 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5924 .read = mem_cgroup_read,
5927 .name = "kmem.failcnt",
5928 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5929 .trigger = mem_cgroup_reset,
5930 .read = mem_cgroup_read,
5933 .name = "kmem.max_usage_in_bytes",
5934 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5935 .trigger = mem_cgroup_reset,
5936 .read = mem_cgroup_read,
5938 #ifdef CONFIG_SLABINFO
5940 .name = "kmem.slabinfo",
5941 .read_seq_string = mem_cgroup_slabinfo_read,
5945 { }, /* terminate */
5948 #ifdef CONFIG_MEMCG_SWAP
5949 static struct cftype memsw_cgroup_files[] = {
5951 .name = "memsw.usage_in_bytes",
5952 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
5953 .read = mem_cgroup_read,
5954 .register_event = mem_cgroup_usage_register_event,
5955 .unregister_event = mem_cgroup_usage_unregister_event,
5958 .name = "memsw.max_usage_in_bytes",
5959 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
5960 .trigger = mem_cgroup_reset,
5961 .read = mem_cgroup_read,
5964 .name = "memsw.limit_in_bytes",
5965 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
5966 .write_string = mem_cgroup_write,
5967 .read = mem_cgroup_read,
5970 .name = "memsw.failcnt",
5971 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
5972 .trigger = mem_cgroup_reset,
5973 .read = mem_cgroup_read,
5975 { }, /* terminate */
5978 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
5980 struct mem_cgroup_per_node *pn;
5981 struct mem_cgroup_per_zone *mz;
5982 int zone, tmp = node;
5984 * This routine is called against possible nodes.
5985 * But it's BUG to call kmalloc() against offline node.
5987 * TODO: this routine can waste much memory for nodes which will
5988 * never be onlined. It's better to use memory hotplug callback
5991 if (!node_state(node, N_NORMAL_MEMORY))
5993 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5997 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5998 mz = &pn->zoneinfo[zone];
5999 lruvec_init(&mz->lruvec);
6000 mz->usage_in_excess = 0;
6001 mz->on_tree = false;
6004 memcg->info.nodeinfo[node] = pn;
6008 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6010 kfree(memcg->info.nodeinfo[node]);
6013 static struct mem_cgroup *mem_cgroup_alloc(void)
6015 struct mem_cgroup *memcg;
6016 size_t size = memcg_size();
6018 /* Can be very big if nr_node_ids is very big */
6019 if (size < PAGE_SIZE)
6020 memcg = kzalloc(size, GFP_KERNEL);
6022 memcg = vzalloc(size);
6027 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
6030 spin_lock_init(&memcg->pcp_counter_lock);
6034 if (size < PAGE_SIZE)
6042 * At destroying mem_cgroup, references from swap_cgroup can remain.
6043 * (scanning all at force_empty is too costly...)
6045 * Instead of clearing all references at force_empty, we remember
6046 * the number of reference from swap_cgroup and free mem_cgroup when
6047 * it goes down to 0.
6049 * Removal of cgroup itself succeeds regardless of refs from swap.
6052 static void __mem_cgroup_free(struct mem_cgroup *memcg)
6055 size_t size = memcg_size();
6057 mem_cgroup_remove_from_trees(memcg);
6058 free_css_id(&mem_cgroup_subsys, &memcg->css);
6061 free_mem_cgroup_per_zone_info(memcg, node);
6063 free_percpu(memcg->stat);
6066 * We need to make sure that (at least for now), the jump label
6067 * destruction code runs outside of the cgroup lock. This is because
6068 * get_online_cpus(), which is called from the static_branch update,
6069 * can't be called inside the cgroup_lock. cpusets are the ones
6070 * enforcing this dependency, so if they ever change, we might as well.
6072 * schedule_work() will guarantee this happens. Be careful if you need
6073 * to move this code around, and make sure it is outside
6076 disarm_static_keys(memcg);
6077 if (size < PAGE_SIZE)
6085 * Helpers for freeing a kmalloc()ed/vzalloc()ed mem_cgroup by RCU,
6086 * but in process context. The work_freeing structure is overlaid
6087 * on the rcu_freeing structure, which itself is overlaid on memsw.
6089 static void free_work(struct work_struct *work)
6091 struct mem_cgroup *memcg;
6093 memcg = container_of(work, struct mem_cgroup, work_freeing);
6094 __mem_cgroup_free(memcg);
6097 static void free_rcu(struct rcu_head *rcu_head)
6099 struct mem_cgroup *memcg;
6101 memcg = container_of(rcu_head, struct mem_cgroup, rcu_freeing);
6102 INIT_WORK(&memcg->work_freeing, free_work);
6103 schedule_work(&memcg->work_freeing);
6106 static void mem_cgroup_get(struct mem_cgroup *memcg)
6108 atomic_inc(&memcg->refcnt);
6111 static void __mem_cgroup_put(struct mem_cgroup *memcg, int count)
6113 if (atomic_sub_and_test(count, &memcg->refcnt)) {
6114 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
6115 call_rcu(&memcg->rcu_freeing, free_rcu);
6117 mem_cgroup_put(parent);
6121 static void mem_cgroup_put(struct mem_cgroup *memcg)
6123 __mem_cgroup_put(memcg, 1);
6127 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6129 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
6131 if (!memcg->res.parent)
6133 return mem_cgroup_from_res_counter(memcg->res.parent, res);
6135 EXPORT_SYMBOL(parent_mem_cgroup);
6137 static void __init mem_cgroup_soft_limit_tree_init(void)
6139 struct mem_cgroup_tree_per_node *rtpn;
6140 struct mem_cgroup_tree_per_zone *rtpz;
6141 int tmp, node, zone;
6143 for_each_node(node) {
6145 if (!node_state(node, N_NORMAL_MEMORY))
6147 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
6150 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6152 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6153 rtpz = &rtpn->rb_tree_per_zone[zone];
6154 rtpz->rb_root = RB_ROOT;
6155 spin_lock_init(&rtpz->lock);
6160 static struct cgroup_subsys_state * __ref
6161 mem_cgroup_css_alloc(struct cgroup *cont)
6163 struct mem_cgroup *memcg;
6164 long error = -ENOMEM;
6167 memcg = mem_cgroup_alloc();
6169 return ERR_PTR(error);
6172 if (alloc_mem_cgroup_per_zone_info(memcg, node))
6176 if (cont->parent == NULL) {
6177 root_mem_cgroup = memcg;
6178 res_counter_init(&memcg->res, NULL);
6179 res_counter_init(&memcg->memsw, NULL);
6180 res_counter_init(&memcg->kmem, NULL);
6183 memcg->last_scanned_node = MAX_NUMNODES;
6184 INIT_LIST_HEAD(&memcg->oom_notify);
6185 atomic_set(&memcg->refcnt, 1);
6186 memcg->move_charge_at_immigrate = 0;
6187 mutex_init(&memcg->thresholds_lock);
6188 spin_lock_init(&memcg->move_lock);
6193 __mem_cgroup_free(memcg);
6194 return ERR_PTR(error);
6198 mem_cgroup_css_online(struct cgroup *cont)
6200 struct mem_cgroup *memcg, *parent;
6206 mutex_lock(&memcg_create_mutex);
6207 memcg = mem_cgroup_from_cont(cont);
6208 parent = mem_cgroup_from_cont(cont->parent);
6210 memcg->use_hierarchy = parent->use_hierarchy;
6211 memcg->oom_kill_disable = parent->oom_kill_disable;
6212 memcg->swappiness = mem_cgroup_swappiness(parent);
6214 if (parent->use_hierarchy) {
6215 res_counter_init(&memcg->res, &parent->res);
6216 res_counter_init(&memcg->memsw, &parent->memsw);
6217 res_counter_init(&memcg->kmem, &parent->kmem);
6220 * We increment refcnt of the parent to ensure that we can
6221 * safely access it on res_counter_charge/uncharge.
6222 * This refcnt will be decremented when freeing this
6223 * mem_cgroup(see mem_cgroup_put).
6225 mem_cgroup_get(parent);
6227 res_counter_init(&memcg->res, NULL);
6228 res_counter_init(&memcg->memsw, NULL);
6229 res_counter_init(&memcg->kmem, NULL);
6231 * Deeper hierachy with use_hierarchy == false doesn't make
6232 * much sense so let cgroup subsystem know about this
6233 * unfortunate state in our controller.
6235 if (parent != root_mem_cgroup)
6236 mem_cgroup_subsys.broken_hierarchy = true;
6239 error = memcg_init_kmem(memcg, &mem_cgroup_subsys);
6240 mutex_unlock(&memcg_create_mutex);
6243 * We call put now because our (and parent's) refcnts
6244 * are already in place. mem_cgroup_put() will internally
6245 * call __mem_cgroup_free, so return directly
6247 mem_cgroup_put(memcg);
6248 if (parent->use_hierarchy)
6249 mem_cgroup_put(parent);
6255 * Announce all parents that a group from their hierarchy is gone.
6257 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup *memcg)
6259 struct mem_cgroup *parent = memcg;
6261 while ((parent = parent_mem_cgroup(parent)))
6262 atomic_inc(&parent->dead_count);
6265 * if the root memcg is not hierarchical we have to check it
6268 if (!root_mem_cgroup->use_hierarchy)
6269 atomic_inc(&root_mem_cgroup->dead_count);
6272 static void mem_cgroup_css_offline(struct cgroup *cont)
6274 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
6276 mem_cgroup_invalidate_reclaim_iterators(memcg);
6277 mem_cgroup_reparent_charges(memcg);
6278 mem_cgroup_destroy_all_caches(memcg);
6281 static void mem_cgroup_css_free(struct cgroup *cont)
6283 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
6285 kmem_cgroup_destroy(memcg);
6287 mem_cgroup_put(memcg);
6291 /* Handlers for move charge at task migration. */
6292 #define PRECHARGE_COUNT_AT_ONCE 256
6293 static int mem_cgroup_do_precharge(unsigned long count)
6296 int batch_count = PRECHARGE_COUNT_AT_ONCE;
6297 struct mem_cgroup *memcg = mc.to;
6299 if (mem_cgroup_is_root(memcg)) {
6300 mc.precharge += count;
6301 /* we don't need css_get for root */
6304 /* try to charge at once */
6306 struct res_counter *dummy;
6308 * "memcg" cannot be under rmdir() because we've already checked
6309 * by cgroup_lock_live_cgroup() that it is not removed and we
6310 * are still under the same cgroup_mutex. So we can postpone
6313 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
6315 if (do_swap_account && res_counter_charge(&memcg->memsw,
6316 PAGE_SIZE * count, &dummy)) {
6317 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
6320 mc.precharge += count;
6324 /* fall back to one by one charge */
6326 if (signal_pending(current)) {
6330 if (!batch_count--) {
6331 batch_count = PRECHARGE_COUNT_AT_ONCE;
6334 ret = __mem_cgroup_try_charge(NULL,
6335 GFP_KERNEL, 1, &memcg, false);
6337 /* mem_cgroup_clear_mc() will do uncharge later */
6345 * get_mctgt_type - get target type of moving charge
6346 * @vma: the vma the pte to be checked belongs
6347 * @addr: the address corresponding to the pte to be checked
6348 * @ptent: the pte to be checked
6349 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6352 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6353 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6354 * move charge. if @target is not NULL, the page is stored in target->page
6355 * with extra refcnt got(Callers should handle it).
6356 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6357 * target for charge migration. if @target is not NULL, the entry is stored
6360 * Called with pte lock held.
6367 enum mc_target_type {
6373 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
6374 unsigned long addr, pte_t ptent)
6376 struct page *page = vm_normal_page(vma, addr, ptent);
6378 if (!page || !page_mapped(page))
6380 if (PageAnon(page)) {
6381 /* we don't move shared anon */
6384 } else if (!move_file())
6385 /* we ignore mapcount for file pages */
6387 if (!get_page_unless_zero(page))
6394 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6395 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6397 struct page *page = NULL;
6398 swp_entry_t ent = pte_to_swp_entry(ptent);
6400 if (!move_anon() || non_swap_entry(ent))
6403 * Because lookup_swap_cache() updates some statistics counter,
6404 * we call find_get_page() with swapper_space directly.
6406 page = find_get_page(swap_address_space(ent), ent.val);
6407 if (do_swap_account)
6408 entry->val = ent.val;
6413 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6414 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6420 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
6421 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6423 struct page *page = NULL;
6424 struct address_space *mapping;
6427 if (!vma->vm_file) /* anonymous vma */
6432 mapping = vma->vm_file->f_mapping;
6433 if (pte_none(ptent))
6434 pgoff = linear_page_index(vma, addr);
6435 else /* pte_file(ptent) is true */
6436 pgoff = pte_to_pgoff(ptent);
6438 /* page is moved even if it's not RSS of this task(page-faulted). */
6439 page = find_get_page(mapping, pgoff);
6442 /* shmem/tmpfs may report page out on swap: account for that too. */
6443 if (radix_tree_exceptional_entry(page)) {
6444 swp_entry_t swap = radix_to_swp_entry(page);
6445 if (do_swap_account)
6447 page = find_get_page(swap_address_space(swap), swap.val);
6453 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
6454 unsigned long addr, pte_t ptent, union mc_target *target)
6456 struct page *page = NULL;
6457 struct page_cgroup *pc;
6458 enum mc_target_type ret = MC_TARGET_NONE;
6459 swp_entry_t ent = { .val = 0 };
6461 if (pte_present(ptent))
6462 page = mc_handle_present_pte(vma, addr, ptent);
6463 else if (is_swap_pte(ptent))
6464 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
6465 else if (pte_none(ptent) || pte_file(ptent))
6466 page = mc_handle_file_pte(vma, addr, ptent, &ent);
6468 if (!page && !ent.val)
6471 pc = lookup_page_cgroup(page);
6473 * Do only loose check w/o page_cgroup lock.
6474 * mem_cgroup_move_account() checks the pc is valid or not under
6477 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6478 ret = MC_TARGET_PAGE;
6480 target->page = page;
6482 if (!ret || !target)
6485 /* There is a swap entry and a page doesn't exist or isn't charged */
6486 if (ent.val && !ret &&
6487 css_id(&mc.from->css) == lookup_swap_cgroup_id(ent)) {
6488 ret = MC_TARGET_SWAP;
6495 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6497 * We don't consider swapping or file mapped pages because THP does not
6498 * support them for now.
6499 * Caller should make sure that pmd_trans_huge(pmd) is true.
6501 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6502 unsigned long addr, pmd_t pmd, union mc_target *target)
6504 struct page *page = NULL;
6505 struct page_cgroup *pc;
6506 enum mc_target_type ret = MC_TARGET_NONE;
6508 page = pmd_page(pmd);
6509 VM_BUG_ON(!page || !PageHead(page));
6512 pc = lookup_page_cgroup(page);
6513 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6514 ret = MC_TARGET_PAGE;
6517 target->page = page;
6523 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6524 unsigned long addr, pmd_t pmd, union mc_target *target)
6526 return MC_TARGET_NONE;
6530 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
6531 unsigned long addr, unsigned long end,
6532 struct mm_walk *walk)
6534 struct vm_area_struct *vma = walk->private;
6538 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6539 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
6540 mc.precharge += HPAGE_PMD_NR;
6541 spin_unlock(&vma->vm_mm->page_table_lock);
6545 if (pmd_trans_unstable(pmd))
6547 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6548 for (; addr != end; pte++, addr += PAGE_SIZE)
6549 if (get_mctgt_type(vma, addr, *pte, NULL))
6550 mc.precharge++; /* increment precharge temporarily */
6551 pte_unmap_unlock(pte - 1, ptl);
6557 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
6559 unsigned long precharge;
6560 struct vm_area_struct *vma;
6562 down_read(&mm->mmap_sem);
6563 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6564 struct mm_walk mem_cgroup_count_precharge_walk = {
6565 .pmd_entry = mem_cgroup_count_precharge_pte_range,
6569 if (is_vm_hugetlb_page(vma))
6571 walk_page_range(vma->vm_start, vma->vm_end,
6572 &mem_cgroup_count_precharge_walk);
6574 up_read(&mm->mmap_sem);
6576 precharge = mc.precharge;
6582 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
6584 unsigned long precharge = mem_cgroup_count_precharge(mm);
6586 VM_BUG_ON(mc.moving_task);
6587 mc.moving_task = current;
6588 return mem_cgroup_do_precharge(precharge);
6591 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6592 static void __mem_cgroup_clear_mc(void)
6594 struct mem_cgroup *from = mc.from;
6595 struct mem_cgroup *to = mc.to;
6597 /* we must uncharge all the leftover precharges from mc.to */
6599 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
6603 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6604 * we must uncharge here.
6606 if (mc.moved_charge) {
6607 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
6608 mc.moved_charge = 0;
6610 /* we must fixup refcnts and charges */
6611 if (mc.moved_swap) {
6612 /* uncharge swap account from the old cgroup */
6613 if (!mem_cgroup_is_root(mc.from))
6614 res_counter_uncharge(&mc.from->memsw,
6615 PAGE_SIZE * mc.moved_swap);
6616 __mem_cgroup_put(mc.from, mc.moved_swap);
6618 if (!mem_cgroup_is_root(mc.to)) {
6620 * we charged both to->res and to->memsw, so we should
6623 res_counter_uncharge(&mc.to->res,
6624 PAGE_SIZE * mc.moved_swap);
6626 /* we've already done mem_cgroup_get(mc.to) */
6629 memcg_oom_recover(from);
6630 memcg_oom_recover(to);
6631 wake_up_all(&mc.waitq);
6634 static void mem_cgroup_clear_mc(void)
6636 struct mem_cgroup *from = mc.from;
6639 * we must clear moving_task before waking up waiters at the end of
6642 mc.moving_task = NULL;
6643 __mem_cgroup_clear_mc();
6644 spin_lock(&mc.lock);
6647 spin_unlock(&mc.lock);
6648 mem_cgroup_end_move(from);
6651 static int mem_cgroup_can_attach(struct cgroup *cgroup,
6652 struct cgroup_taskset *tset)
6654 struct task_struct *p = cgroup_taskset_first(tset);
6656 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup);
6657 unsigned long move_charge_at_immigrate;
6660 * We are now commited to this value whatever it is. Changes in this
6661 * tunable will only affect upcoming migrations, not the current one.
6662 * So we need to save it, and keep it going.
6664 move_charge_at_immigrate = memcg->move_charge_at_immigrate;
6665 if (move_charge_at_immigrate) {
6666 struct mm_struct *mm;
6667 struct mem_cgroup *from = mem_cgroup_from_task(p);
6669 VM_BUG_ON(from == memcg);
6671 mm = get_task_mm(p);
6674 /* We move charges only when we move a owner of the mm */
6675 if (mm->owner == p) {
6678 VM_BUG_ON(mc.precharge);
6679 VM_BUG_ON(mc.moved_charge);
6680 VM_BUG_ON(mc.moved_swap);
6681 mem_cgroup_start_move(from);
6682 spin_lock(&mc.lock);
6685 mc.immigrate_flags = move_charge_at_immigrate;
6686 spin_unlock(&mc.lock);
6687 /* We set mc.moving_task later */
6689 ret = mem_cgroup_precharge_mc(mm);
6691 mem_cgroup_clear_mc();
6698 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
6699 struct cgroup_taskset *tset)
6701 mem_cgroup_clear_mc();
6704 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6705 unsigned long addr, unsigned long end,
6706 struct mm_walk *walk)
6709 struct vm_area_struct *vma = walk->private;
6712 enum mc_target_type target_type;
6713 union mc_target target;
6715 struct page_cgroup *pc;
6718 * We don't take compound_lock() here but no race with splitting thp
6720 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6721 * under splitting, which means there's no concurrent thp split,
6722 * - if another thread runs into split_huge_page() just after we
6723 * entered this if-block, the thread must wait for page table lock
6724 * to be unlocked in __split_huge_page_splitting(), where the main
6725 * part of thp split is not executed yet.
6727 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6728 if (mc.precharge < HPAGE_PMD_NR) {
6729 spin_unlock(&vma->vm_mm->page_table_lock);
6732 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6733 if (target_type == MC_TARGET_PAGE) {
6735 if (!isolate_lru_page(page)) {
6736 pc = lookup_page_cgroup(page);
6737 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
6738 pc, mc.from, mc.to)) {
6739 mc.precharge -= HPAGE_PMD_NR;
6740 mc.moved_charge += HPAGE_PMD_NR;
6742 putback_lru_page(page);
6746 spin_unlock(&vma->vm_mm->page_table_lock);
6750 if (pmd_trans_unstable(pmd))
6753 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6754 for (; addr != end; addr += PAGE_SIZE) {
6755 pte_t ptent = *(pte++);
6761 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6762 case MC_TARGET_PAGE:
6764 if (isolate_lru_page(page))
6766 pc = lookup_page_cgroup(page);
6767 if (!mem_cgroup_move_account(page, 1, pc,
6770 /* we uncharge from mc.from later. */
6773 putback_lru_page(page);
6774 put: /* get_mctgt_type() gets the page */
6777 case MC_TARGET_SWAP:
6779 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6781 /* we fixup refcnts and charges later. */
6789 pte_unmap_unlock(pte - 1, ptl);
6794 * We have consumed all precharges we got in can_attach().
6795 * We try charge one by one, but don't do any additional
6796 * charges to mc.to if we have failed in charge once in attach()
6799 ret = mem_cgroup_do_precharge(1);
6807 static void mem_cgroup_move_charge(struct mm_struct *mm)
6809 struct vm_area_struct *vma;
6811 lru_add_drain_all();
6813 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
6815 * Someone who are holding the mmap_sem might be waiting in
6816 * waitq. So we cancel all extra charges, wake up all waiters,
6817 * and retry. Because we cancel precharges, we might not be able
6818 * to move enough charges, but moving charge is a best-effort
6819 * feature anyway, so it wouldn't be a big problem.
6821 __mem_cgroup_clear_mc();
6825 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6827 struct mm_walk mem_cgroup_move_charge_walk = {
6828 .pmd_entry = mem_cgroup_move_charge_pte_range,
6832 if (is_vm_hugetlb_page(vma))
6834 ret = walk_page_range(vma->vm_start, vma->vm_end,
6835 &mem_cgroup_move_charge_walk);
6838 * means we have consumed all precharges and failed in
6839 * doing additional charge. Just abandon here.
6843 up_read(&mm->mmap_sem);
6846 static void mem_cgroup_move_task(struct cgroup *cont,
6847 struct cgroup_taskset *tset)
6849 struct task_struct *p = cgroup_taskset_first(tset);
6850 struct mm_struct *mm = get_task_mm(p);
6854 mem_cgroup_move_charge(mm);
6858 mem_cgroup_clear_mc();
6860 #else /* !CONFIG_MMU */
6861 static int mem_cgroup_can_attach(struct cgroup *cgroup,
6862 struct cgroup_taskset *tset)
6866 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
6867 struct cgroup_taskset *tset)
6870 static void mem_cgroup_move_task(struct cgroup *cont,
6871 struct cgroup_taskset *tset)
6876 struct cgroup_subsys mem_cgroup_subsys = {
6878 .subsys_id = mem_cgroup_subsys_id,
6879 .css_alloc = mem_cgroup_css_alloc,
6880 .css_online = mem_cgroup_css_online,
6881 .css_offline = mem_cgroup_css_offline,
6882 .css_free = mem_cgroup_css_free,
6883 .can_attach = mem_cgroup_can_attach,
6884 .cancel_attach = mem_cgroup_cancel_attach,
6885 .attach = mem_cgroup_move_task,
6886 .base_cftypes = mem_cgroup_files,
6891 #ifdef CONFIG_MEMCG_SWAP
6892 static int __init enable_swap_account(char *s)
6894 /* consider enabled if no parameter or 1 is given */
6895 if (!strcmp(s, "1"))
6896 really_do_swap_account = 1;
6897 else if (!strcmp(s, "0"))
6898 really_do_swap_account = 0;
6901 __setup("swapaccount=", enable_swap_account);
6903 static void __init memsw_file_init(void)
6905 WARN_ON(cgroup_add_cftypes(&mem_cgroup_subsys, memsw_cgroup_files));
6908 static void __init enable_swap_cgroup(void)
6910 if (!mem_cgroup_disabled() && really_do_swap_account) {
6911 do_swap_account = 1;
6917 static void __init enable_swap_cgroup(void)
6923 * subsys_initcall() for memory controller.
6925 * Some parts like hotcpu_notifier() have to be initialized from this context
6926 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
6927 * everything that doesn't depend on a specific mem_cgroup structure should
6928 * be initialized from here.
6930 static int __init mem_cgroup_init(void)
6932 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
6933 enable_swap_cgroup();
6934 mem_cgroup_soft_limit_tree_init();
6938 subsys_initcall(mem_cgroup_init);