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 #define MEM_CGROUP_RECLAIM_RETRIES 5
67 static struct mem_cgroup *root_mem_cgroup __read_mostly;
69 #ifdef CONFIG_MEMCG_SWAP
70 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
71 int do_swap_account __read_mostly;
73 /* for remember boot option*/
74 #ifdef CONFIG_MEMCG_SWAP_ENABLED
75 static int really_do_swap_account __initdata = 1;
77 static int really_do_swap_account __initdata = 0;
81 #define do_swap_account 0
86 * Statistics for memory cgroup.
88 enum mem_cgroup_stat_index {
90 * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
92 MEM_CGROUP_STAT_CACHE, /* # of pages charged as cache */
93 MEM_CGROUP_STAT_RSS, /* # of pages charged as anon rss */
94 MEM_CGROUP_STAT_FILE_MAPPED, /* # of pages charged as file rss */
95 MEM_CGROUP_STAT_SWAP, /* # of pages, swapped out */
96 MEM_CGROUP_STAT_NSTATS,
99 static const char * const mem_cgroup_stat_names[] = {
106 enum mem_cgroup_events_index {
107 MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */
108 MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */
109 MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */
110 MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */
111 MEM_CGROUP_EVENTS_NSTATS,
114 static const char * const mem_cgroup_events_names[] = {
122 * Per memcg event counter is incremented at every pagein/pageout. With THP,
123 * it will be incremated by the number of pages. This counter is used for
124 * for trigger some periodic events. This is straightforward and better
125 * than using jiffies etc. to handle periodic memcg event.
127 enum mem_cgroup_events_target {
128 MEM_CGROUP_TARGET_THRESH,
129 MEM_CGROUP_TARGET_SOFTLIMIT,
130 MEM_CGROUP_TARGET_NUMAINFO,
133 #define THRESHOLDS_EVENTS_TARGET 128
134 #define SOFTLIMIT_EVENTS_TARGET 1024
135 #define NUMAINFO_EVENTS_TARGET 1024
137 struct mem_cgroup_stat_cpu {
138 long count[MEM_CGROUP_STAT_NSTATS];
139 unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
140 unsigned long nr_page_events;
141 unsigned long targets[MEM_CGROUP_NTARGETS];
144 struct mem_cgroup_reclaim_iter {
145 /* css_id of the last scanned hierarchy member */
147 /* scan generation, increased every round-trip */
148 unsigned int generation;
152 * per-zone information in memory controller.
154 struct mem_cgroup_per_zone {
155 struct lruvec lruvec;
156 unsigned long lru_size[NR_LRU_LISTS];
158 struct mem_cgroup_reclaim_iter reclaim_iter[DEF_PRIORITY + 1];
160 struct rb_node tree_node; /* RB tree node */
161 unsigned long long usage_in_excess;/* Set to the value by which */
162 /* the soft limit is exceeded*/
164 struct mem_cgroup *memcg; /* Back pointer, we cannot */
165 /* use container_of */
168 struct mem_cgroup_per_node {
169 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
172 struct mem_cgroup_lru_info {
173 struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
177 * Cgroups above their limits are maintained in a RB-Tree, independent of
178 * their hierarchy representation
181 struct mem_cgroup_tree_per_zone {
182 struct rb_root rb_root;
186 struct mem_cgroup_tree_per_node {
187 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
190 struct mem_cgroup_tree {
191 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
194 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
196 struct mem_cgroup_threshold {
197 struct eventfd_ctx *eventfd;
202 struct mem_cgroup_threshold_ary {
203 /* An array index points to threshold just below or equal to usage. */
204 int current_threshold;
205 /* Size of entries[] */
207 /* Array of thresholds */
208 struct mem_cgroup_threshold entries[0];
211 struct mem_cgroup_thresholds {
212 /* Primary thresholds array */
213 struct mem_cgroup_threshold_ary *primary;
215 * Spare threshold array.
216 * This is needed to make mem_cgroup_unregister_event() "never fail".
217 * It must be able to store at least primary->size - 1 entries.
219 struct mem_cgroup_threshold_ary *spare;
223 struct mem_cgroup_eventfd_list {
224 struct list_head list;
225 struct eventfd_ctx *eventfd;
228 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
229 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
232 * The memory controller data structure. The memory controller controls both
233 * page cache and RSS per cgroup. We would eventually like to provide
234 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
235 * to help the administrator determine what knobs to tune.
237 * TODO: Add a water mark for the memory controller. Reclaim will begin when
238 * we hit the water mark. May be even add a low water mark, such that
239 * no reclaim occurs from a cgroup at it's low water mark, this is
240 * a feature that will be implemented much later in the future.
243 struct cgroup_subsys_state css;
245 * the counter to account for memory usage
247 struct res_counter res;
251 * the counter to account for mem+swap usage.
253 struct res_counter memsw;
256 * rcu_freeing is used only when freeing struct mem_cgroup,
257 * so put it into a union to avoid wasting more memory.
258 * It must be disjoint from the css field. It could be
259 * in a union with the res field, but res plays a much
260 * larger part in mem_cgroup life than memsw, and might
261 * be of interest, even at time of free, when debugging.
262 * So share rcu_head with the less interesting memsw.
264 struct rcu_head rcu_freeing;
266 * We also need some space for a worker in deferred freeing.
267 * By the time we call it, rcu_freeing is no longer in use.
269 struct work_struct work_freeing;
273 * the counter to account for kernel memory usage.
275 struct res_counter kmem;
277 * Per cgroup active and inactive list, similar to the
278 * per zone LRU lists.
280 struct mem_cgroup_lru_info info;
281 int last_scanned_node;
283 nodemask_t scan_nodes;
284 atomic_t numainfo_events;
285 atomic_t numainfo_updating;
288 * Should the accounting and control be hierarchical, per subtree?
291 unsigned long kmem_account_flags; /* See KMEM_ACCOUNTED_*, below */
299 /* OOM-Killer disable */
300 int oom_kill_disable;
302 /* set when res.limit == memsw.limit */
303 bool memsw_is_minimum;
305 /* protect arrays of thresholds */
306 struct mutex thresholds_lock;
308 /* thresholds for memory usage. RCU-protected */
309 struct mem_cgroup_thresholds thresholds;
311 /* thresholds for mem+swap usage. RCU-protected */
312 struct mem_cgroup_thresholds memsw_thresholds;
314 /* For oom notifier event fd */
315 struct list_head oom_notify;
318 * Should we move charges of a task when a task is moved into this
319 * mem_cgroup ? And what type of charges should we move ?
321 unsigned long move_charge_at_immigrate;
323 * set > 0 if pages under this cgroup are moving to other cgroup.
325 atomic_t moving_account;
326 /* taken only while moving_account > 0 */
327 spinlock_t move_lock;
331 struct mem_cgroup_stat_cpu __percpu *stat;
333 * used when a cpu is offlined or other synchronizations
334 * See mem_cgroup_read_stat().
336 struct mem_cgroup_stat_cpu nocpu_base;
337 spinlock_t pcp_counter_lock;
339 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
340 struct tcp_memcontrol tcp_mem;
342 #if defined(CONFIG_MEMCG_KMEM)
343 /* analogous to slab_common's slab_caches list. per-memcg */
344 struct list_head memcg_slab_caches;
345 /* Not a spinlock, we can take a lot of time walking the list */
346 struct mutex slab_caches_mutex;
347 /* Index in the kmem_cache->memcg_params->memcg_caches array */
352 /* internal only representation about the status of kmem accounting. */
354 KMEM_ACCOUNTED_ACTIVE = 0, /* accounted by this cgroup itself */
355 KMEM_ACCOUNTED_ACTIVATED, /* static key enabled. */
356 KMEM_ACCOUNTED_DEAD, /* dead memcg with pending kmem charges */
359 /* We account when limit is on, but only after call sites are patched */
360 #define KMEM_ACCOUNTED_MASK \
361 ((1 << KMEM_ACCOUNTED_ACTIVE) | (1 << KMEM_ACCOUNTED_ACTIVATED))
363 #ifdef CONFIG_MEMCG_KMEM
364 static inline void memcg_kmem_set_active(struct mem_cgroup *memcg)
366 set_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
369 static bool memcg_kmem_is_active(struct mem_cgroup *memcg)
371 return test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
374 static void memcg_kmem_set_activated(struct mem_cgroup *memcg)
376 set_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
379 static void memcg_kmem_clear_activated(struct mem_cgroup *memcg)
381 clear_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
384 static void memcg_kmem_mark_dead(struct mem_cgroup *memcg)
386 if (test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags))
387 set_bit(KMEM_ACCOUNTED_DEAD, &memcg->kmem_account_flags);
390 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup *memcg)
392 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD,
393 &memcg->kmem_account_flags);
397 /* Stuffs for move charges at task migration. */
399 * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
400 * left-shifted bitmap of these types.
403 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
404 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
408 /* "mc" and its members are protected by cgroup_mutex */
409 static struct move_charge_struct {
410 spinlock_t lock; /* for from, to */
411 struct mem_cgroup *from;
412 struct mem_cgroup *to;
413 unsigned long precharge;
414 unsigned long moved_charge;
415 unsigned long moved_swap;
416 struct task_struct *moving_task; /* a task moving charges */
417 wait_queue_head_t waitq; /* a waitq for other context */
419 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
420 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
423 static bool move_anon(void)
425 return test_bit(MOVE_CHARGE_TYPE_ANON,
426 &mc.to->move_charge_at_immigrate);
429 static bool move_file(void)
431 return test_bit(MOVE_CHARGE_TYPE_FILE,
432 &mc.to->move_charge_at_immigrate);
436 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
437 * limit reclaim to prevent infinite loops, if they ever occur.
439 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
440 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
443 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
444 MEM_CGROUP_CHARGE_TYPE_ANON,
445 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
446 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
450 /* for encoding cft->private value on file */
458 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
459 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
460 #define MEMFILE_ATTR(val) ((val) & 0xffff)
461 /* Used for OOM nofiier */
462 #define OOM_CONTROL (0)
465 * Reclaim flags for mem_cgroup_hierarchical_reclaim
467 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
468 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
469 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
470 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
472 static void mem_cgroup_get(struct mem_cgroup *memcg);
473 static void mem_cgroup_put(struct mem_cgroup *memcg);
476 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
478 return container_of(s, struct mem_cgroup, css);
481 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
483 return (memcg == root_mem_cgroup);
486 /* Writing them here to avoid exposing memcg's inner layout */
487 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
489 void sock_update_memcg(struct sock *sk)
491 if (mem_cgroup_sockets_enabled) {
492 struct mem_cgroup *memcg;
493 struct cg_proto *cg_proto;
495 BUG_ON(!sk->sk_prot->proto_cgroup);
497 /* Socket cloning can throw us here with sk_cgrp already
498 * filled. It won't however, necessarily happen from
499 * process context. So the test for root memcg given
500 * the current task's memcg won't help us in this case.
502 * Respecting the original socket's memcg is a better
503 * decision in this case.
506 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
507 mem_cgroup_get(sk->sk_cgrp->memcg);
512 memcg = mem_cgroup_from_task(current);
513 cg_proto = sk->sk_prot->proto_cgroup(memcg);
514 if (!mem_cgroup_is_root(memcg) && memcg_proto_active(cg_proto)) {
515 mem_cgroup_get(memcg);
516 sk->sk_cgrp = cg_proto;
521 EXPORT_SYMBOL(sock_update_memcg);
523 void sock_release_memcg(struct sock *sk)
525 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
526 struct mem_cgroup *memcg;
527 WARN_ON(!sk->sk_cgrp->memcg);
528 memcg = sk->sk_cgrp->memcg;
529 mem_cgroup_put(memcg);
533 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
535 if (!memcg || mem_cgroup_is_root(memcg))
538 return &memcg->tcp_mem.cg_proto;
540 EXPORT_SYMBOL(tcp_proto_cgroup);
542 static void disarm_sock_keys(struct mem_cgroup *memcg)
544 if (!memcg_proto_activated(&memcg->tcp_mem.cg_proto))
546 static_key_slow_dec(&memcg_socket_limit_enabled);
549 static void disarm_sock_keys(struct mem_cgroup *memcg)
554 #ifdef CONFIG_MEMCG_KMEM
556 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
557 * There are two main reasons for not using the css_id for this:
558 * 1) this works better in sparse environments, where we have a lot of memcgs,
559 * but only a few kmem-limited. Or also, if we have, for instance, 200
560 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
561 * 200 entry array for that.
563 * 2) In order not to violate the cgroup API, we would like to do all memory
564 * allocation in ->create(). At that point, we haven't yet allocated the
565 * css_id. Having a separate index prevents us from messing with the cgroup
568 * The current size of the caches array is stored in
569 * memcg_limited_groups_array_size. It will double each time we have to
572 static struct ida kmem_limited_groups;
573 int memcg_limited_groups_array_size;
576 * MIN_SIZE is different than 1, because we would like to avoid going through
577 * the alloc/free process all the time. In a small machine, 4 kmem-limited
578 * cgroups is a reasonable guess. In the future, it could be a parameter or
579 * tunable, but that is strictly not necessary.
581 * MAX_SIZE should be as large as the number of css_ids. Ideally, we could get
582 * this constant directly from cgroup, but it is understandable that this is
583 * better kept as an internal representation in cgroup.c. In any case, the
584 * css_id space is not getting any smaller, and we don't have to necessarily
585 * increase ours as well if it increases.
587 #define MEMCG_CACHES_MIN_SIZE 4
588 #define MEMCG_CACHES_MAX_SIZE 65535
591 * A lot of the calls to the cache allocation functions are expected to be
592 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
593 * conditional to this static branch, we'll have to allow modules that does
594 * kmem_cache_alloc and the such to see this symbol as well
596 struct static_key memcg_kmem_enabled_key;
597 EXPORT_SYMBOL(memcg_kmem_enabled_key);
599 static void disarm_kmem_keys(struct mem_cgroup *memcg)
601 if (memcg_kmem_is_active(memcg)) {
602 static_key_slow_dec(&memcg_kmem_enabled_key);
603 ida_simple_remove(&kmem_limited_groups, memcg->kmemcg_id);
606 * This check can't live in kmem destruction function,
607 * since the charges will outlive the cgroup
609 WARN_ON(res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0);
612 static void disarm_kmem_keys(struct mem_cgroup *memcg)
615 #endif /* CONFIG_MEMCG_KMEM */
617 static void disarm_static_keys(struct mem_cgroup *memcg)
619 disarm_sock_keys(memcg);
620 disarm_kmem_keys(memcg);
623 static void drain_all_stock_async(struct mem_cgroup *memcg);
625 static struct mem_cgroup_per_zone *
626 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
628 return &memcg->info.nodeinfo[nid]->zoneinfo[zid];
631 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
636 static struct mem_cgroup_per_zone *
637 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
639 int nid = page_to_nid(page);
640 int zid = page_zonenum(page);
642 return mem_cgroup_zoneinfo(memcg, nid, zid);
645 static struct mem_cgroup_tree_per_zone *
646 soft_limit_tree_node_zone(int nid, int zid)
648 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
651 static struct mem_cgroup_tree_per_zone *
652 soft_limit_tree_from_page(struct page *page)
654 int nid = page_to_nid(page);
655 int zid = page_zonenum(page);
657 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
661 __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
662 struct mem_cgroup_per_zone *mz,
663 struct mem_cgroup_tree_per_zone *mctz,
664 unsigned long long new_usage_in_excess)
666 struct rb_node **p = &mctz->rb_root.rb_node;
667 struct rb_node *parent = NULL;
668 struct mem_cgroup_per_zone *mz_node;
673 mz->usage_in_excess = new_usage_in_excess;
674 if (!mz->usage_in_excess)
678 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
680 if (mz->usage_in_excess < mz_node->usage_in_excess)
683 * We can't avoid mem cgroups that are over their soft
684 * limit by the same amount
686 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
689 rb_link_node(&mz->tree_node, parent, p);
690 rb_insert_color(&mz->tree_node, &mctz->rb_root);
695 __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
696 struct mem_cgroup_per_zone *mz,
697 struct mem_cgroup_tree_per_zone *mctz)
701 rb_erase(&mz->tree_node, &mctz->rb_root);
706 mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
707 struct mem_cgroup_per_zone *mz,
708 struct mem_cgroup_tree_per_zone *mctz)
710 spin_lock(&mctz->lock);
711 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
712 spin_unlock(&mctz->lock);
716 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
718 unsigned long long excess;
719 struct mem_cgroup_per_zone *mz;
720 struct mem_cgroup_tree_per_zone *mctz;
721 int nid = page_to_nid(page);
722 int zid = page_zonenum(page);
723 mctz = soft_limit_tree_from_page(page);
726 * Necessary to update all ancestors when hierarchy is used.
727 * because their event counter is not touched.
729 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
730 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
731 excess = res_counter_soft_limit_excess(&memcg->res);
733 * We have to update the tree if mz is on RB-tree or
734 * mem is over its softlimit.
736 if (excess || mz->on_tree) {
737 spin_lock(&mctz->lock);
738 /* if on-tree, remove it */
740 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
742 * Insert again. mz->usage_in_excess will be updated.
743 * If excess is 0, no tree ops.
745 __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
746 spin_unlock(&mctz->lock);
751 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
754 struct mem_cgroup_per_zone *mz;
755 struct mem_cgroup_tree_per_zone *mctz;
757 for_each_node(node) {
758 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
759 mz = mem_cgroup_zoneinfo(memcg, node, zone);
760 mctz = soft_limit_tree_node_zone(node, zone);
761 mem_cgroup_remove_exceeded(memcg, mz, mctz);
766 static struct mem_cgroup_per_zone *
767 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
769 struct rb_node *rightmost = NULL;
770 struct mem_cgroup_per_zone *mz;
774 rightmost = rb_last(&mctz->rb_root);
776 goto done; /* Nothing to reclaim from */
778 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
780 * Remove the node now but someone else can add it back,
781 * we will to add it back at the end of reclaim to its correct
782 * position in the tree.
784 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
785 if (!res_counter_soft_limit_excess(&mz->memcg->res) ||
786 !css_tryget(&mz->memcg->css))
792 static struct mem_cgroup_per_zone *
793 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
795 struct mem_cgroup_per_zone *mz;
797 spin_lock(&mctz->lock);
798 mz = __mem_cgroup_largest_soft_limit_node(mctz);
799 spin_unlock(&mctz->lock);
804 * Implementation Note: reading percpu statistics for memcg.
806 * Both of vmstat[] and percpu_counter has threshold and do periodic
807 * synchronization to implement "quick" read. There are trade-off between
808 * reading cost and precision of value. Then, we may have a chance to implement
809 * a periodic synchronizion of counter in memcg's counter.
811 * But this _read() function is used for user interface now. The user accounts
812 * memory usage by memory cgroup and he _always_ requires exact value because
813 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
814 * have to visit all online cpus and make sum. So, for now, unnecessary
815 * synchronization is not implemented. (just implemented for cpu hotplug)
817 * If there are kernel internal actions which can make use of some not-exact
818 * value, and reading all cpu value can be performance bottleneck in some
819 * common workload, threashold and synchonization as vmstat[] should be
822 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
823 enum mem_cgroup_stat_index idx)
829 for_each_online_cpu(cpu)
830 val += per_cpu(memcg->stat->count[idx], cpu);
831 #ifdef CONFIG_HOTPLUG_CPU
832 spin_lock(&memcg->pcp_counter_lock);
833 val += memcg->nocpu_base.count[idx];
834 spin_unlock(&memcg->pcp_counter_lock);
840 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
843 int val = (charge) ? 1 : -1;
844 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
847 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
848 enum mem_cgroup_events_index idx)
850 unsigned long val = 0;
853 for_each_online_cpu(cpu)
854 val += per_cpu(memcg->stat->events[idx], cpu);
855 #ifdef CONFIG_HOTPLUG_CPU
856 spin_lock(&memcg->pcp_counter_lock);
857 val += memcg->nocpu_base.events[idx];
858 spin_unlock(&memcg->pcp_counter_lock);
863 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
864 bool anon, int nr_pages)
869 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
870 * counted as CACHE even if it's on ANON LRU.
873 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
876 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
879 /* pagein of a big page is an event. So, ignore page size */
881 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
883 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
884 nr_pages = -nr_pages; /* for event */
887 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
893 mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
895 struct mem_cgroup_per_zone *mz;
897 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
898 return mz->lru_size[lru];
902 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
903 unsigned int lru_mask)
905 struct mem_cgroup_per_zone *mz;
907 unsigned long ret = 0;
909 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
912 if (BIT(lru) & lru_mask)
913 ret += mz->lru_size[lru];
919 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
920 int nid, unsigned int lru_mask)
925 for (zid = 0; zid < MAX_NR_ZONES; zid++)
926 total += mem_cgroup_zone_nr_lru_pages(memcg,
932 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
933 unsigned int lru_mask)
938 for_each_node_state(nid, N_HIGH_MEMORY)
939 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
943 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
944 enum mem_cgroup_events_target target)
946 unsigned long val, next;
948 val = __this_cpu_read(memcg->stat->nr_page_events);
949 next = __this_cpu_read(memcg->stat->targets[target]);
950 /* from time_after() in jiffies.h */
951 if ((long)next - (long)val < 0) {
953 case MEM_CGROUP_TARGET_THRESH:
954 next = val + THRESHOLDS_EVENTS_TARGET;
956 case MEM_CGROUP_TARGET_SOFTLIMIT:
957 next = val + SOFTLIMIT_EVENTS_TARGET;
959 case MEM_CGROUP_TARGET_NUMAINFO:
960 next = val + NUMAINFO_EVENTS_TARGET;
965 __this_cpu_write(memcg->stat->targets[target], next);
972 * Check events in order.
975 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
978 /* threshold event is triggered in finer grain than soft limit */
979 if (unlikely(mem_cgroup_event_ratelimit(memcg,
980 MEM_CGROUP_TARGET_THRESH))) {
982 bool do_numainfo __maybe_unused;
984 do_softlimit = mem_cgroup_event_ratelimit(memcg,
985 MEM_CGROUP_TARGET_SOFTLIMIT);
987 do_numainfo = mem_cgroup_event_ratelimit(memcg,
988 MEM_CGROUP_TARGET_NUMAINFO);
992 mem_cgroup_threshold(memcg);
993 if (unlikely(do_softlimit))
994 mem_cgroup_update_tree(memcg, page);
996 if (unlikely(do_numainfo))
997 atomic_inc(&memcg->numainfo_events);
1003 struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
1005 return mem_cgroup_from_css(
1006 cgroup_subsys_state(cont, mem_cgroup_subsys_id));
1009 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1012 * mm_update_next_owner() may clear mm->owner to NULL
1013 * if it races with swapoff, page migration, etc.
1014 * So this can be called with p == NULL.
1019 return mem_cgroup_from_css(task_subsys_state(p, mem_cgroup_subsys_id));
1022 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
1024 struct mem_cgroup *memcg = NULL;
1029 * Because we have no locks, mm->owner's may be being moved to other
1030 * cgroup. We use css_tryget() here even if this looks
1031 * pessimistic (rather than adding locks here).
1035 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1036 if (unlikely(!memcg))
1038 } while (!css_tryget(&memcg->css));
1044 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1045 * @root: hierarchy root
1046 * @prev: previously returned memcg, NULL on first invocation
1047 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1049 * Returns references to children of the hierarchy below @root, or
1050 * @root itself, or %NULL after a full round-trip.
1052 * Caller must pass the return value in @prev on subsequent
1053 * invocations for reference counting, or use mem_cgroup_iter_break()
1054 * to cancel a hierarchy walk before the round-trip is complete.
1056 * Reclaimers can specify a zone and a priority level in @reclaim to
1057 * divide up the memcgs in the hierarchy among all concurrent
1058 * reclaimers operating on the same zone and priority.
1060 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1061 struct mem_cgroup *prev,
1062 struct mem_cgroup_reclaim_cookie *reclaim)
1064 struct mem_cgroup *memcg = NULL;
1067 if (mem_cgroup_disabled())
1071 root = root_mem_cgroup;
1073 if (prev && !reclaim)
1074 id = css_id(&prev->css);
1076 if (prev && prev != root)
1077 css_put(&prev->css);
1079 if (!root->use_hierarchy && root != root_mem_cgroup) {
1086 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1087 struct cgroup_subsys_state *css;
1090 int nid = zone_to_nid(reclaim->zone);
1091 int zid = zone_idx(reclaim->zone);
1092 struct mem_cgroup_per_zone *mz;
1094 mz = mem_cgroup_zoneinfo(root, nid, zid);
1095 iter = &mz->reclaim_iter[reclaim->priority];
1096 if (prev && reclaim->generation != iter->generation)
1098 id = iter->position;
1102 css = css_get_next(&mem_cgroup_subsys, id + 1, &root->css, &id);
1104 if (css == &root->css || css_tryget(css))
1105 memcg = mem_cgroup_from_css(css);
1111 iter->position = id;
1114 else if (!prev && memcg)
1115 reclaim->generation = iter->generation;
1125 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1126 * @root: hierarchy root
1127 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1129 void mem_cgroup_iter_break(struct mem_cgroup *root,
1130 struct mem_cgroup *prev)
1133 root = root_mem_cgroup;
1134 if (prev && prev != root)
1135 css_put(&prev->css);
1139 * Iteration constructs for visiting all cgroups (under a tree). If
1140 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1141 * be used for reference counting.
1143 #define for_each_mem_cgroup_tree(iter, root) \
1144 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1146 iter = mem_cgroup_iter(root, iter, NULL))
1148 #define for_each_mem_cgroup(iter) \
1149 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1151 iter = mem_cgroup_iter(NULL, iter, NULL))
1153 void mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1155 struct mem_cgroup *memcg;
1161 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1162 if (unlikely(!memcg))
1167 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1170 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1178 EXPORT_SYMBOL(mem_cgroup_count_vm_event);
1181 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1182 * @zone: zone of the wanted lruvec
1183 * @memcg: memcg of the wanted lruvec
1185 * Returns the lru list vector holding pages for the given @zone and
1186 * @mem. This can be the global zone lruvec, if the memory controller
1189 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1190 struct mem_cgroup *memcg)
1192 struct mem_cgroup_per_zone *mz;
1194 if (mem_cgroup_disabled())
1195 return &zone->lruvec;
1197 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1202 * Following LRU functions are allowed to be used without PCG_LOCK.
1203 * Operations are called by routine of global LRU independently from memcg.
1204 * What we have to take care of here is validness of pc->mem_cgroup.
1206 * Changes to pc->mem_cgroup happens when
1209 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1210 * It is added to LRU before charge.
1211 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1212 * When moving account, the page is not on LRU. It's isolated.
1216 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1218 * @zone: zone of the page
1220 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1222 struct mem_cgroup_per_zone *mz;
1223 struct mem_cgroup *memcg;
1224 struct page_cgroup *pc;
1226 if (mem_cgroup_disabled())
1227 return &zone->lruvec;
1229 pc = lookup_page_cgroup(page);
1230 memcg = pc->mem_cgroup;
1233 * Surreptitiously switch any uncharged offlist page to root:
1234 * an uncharged page off lru does nothing to secure
1235 * its former mem_cgroup from sudden removal.
1237 * Our caller holds lru_lock, and PageCgroupUsed is updated
1238 * under page_cgroup lock: between them, they make all uses
1239 * of pc->mem_cgroup safe.
1241 if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1242 pc->mem_cgroup = memcg = root_mem_cgroup;
1244 mz = page_cgroup_zoneinfo(memcg, page);
1249 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1250 * @lruvec: mem_cgroup per zone lru vector
1251 * @lru: index of lru list the page is sitting on
1252 * @nr_pages: positive when adding or negative when removing
1254 * This function must be called when a page is added to or removed from an
1257 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1260 struct mem_cgroup_per_zone *mz;
1261 unsigned long *lru_size;
1263 if (mem_cgroup_disabled())
1266 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1267 lru_size = mz->lru_size + lru;
1268 *lru_size += nr_pages;
1269 VM_BUG_ON((long)(*lru_size) < 0);
1273 * Checks whether given mem is same or in the root_mem_cgroup's
1276 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1277 struct mem_cgroup *memcg)
1279 if (root_memcg == memcg)
1281 if (!root_memcg->use_hierarchy || !memcg)
1283 return css_is_ancestor(&memcg->css, &root_memcg->css);
1286 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1287 struct mem_cgroup *memcg)
1292 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1297 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *memcg)
1300 struct mem_cgroup *curr = NULL;
1301 struct task_struct *p;
1303 p = find_lock_task_mm(task);
1305 curr = try_get_mem_cgroup_from_mm(p->mm);
1309 * All threads may have already detached their mm's, but the oom
1310 * killer still needs to detect if they have already been oom
1311 * killed to prevent needlessly killing additional tasks.
1314 curr = mem_cgroup_from_task(task);
1316 css_get(&curr->css);
1322 * We should check use_hierarchy of "memcg" not "curr". Because checking
1323 * use_hierarchy of "curr" here make this function true if hierarchy is
1324 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1325 * hierarchy(even if use_hierarchy is disabled in "memcg").
1327 ret = mem_cgroup_same_or_subtree(memcg, curr);
1328 css_put(&curr->css);
1332 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1334 unsigned long inactive_ratio;
1335 unsigned long inactive;
1336 unsigned long active;
1339 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1340 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1342 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1344 inactive_ratio = int_sqrt(10 * gb);
1348 return inactive * inactive_ratio < active;
1351 int mem_cgroup_inactive_file_is_low(struct lruvec *lruvec)
1353 unsigned long active;
1354 unsigned long inactive;
1356 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_FILE);
1357 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_FILE);
1359 return (active > inactive);
1362 #define mem_cgroup_from_res_counter(counter, member) \
1363 container_of(counter, struct mem_cgroup, member)
1366 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1367 * @memcg: the memory cgroup
1369 * Returns the maximum amount of memory @mem can be charged with, in
1372 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1374 unsigned long long margin;
1376 margin = res_counter_margin(&memcg->res);
1377 if (do_swap_account)
1378 margin = min(margin, res_counter_margin(&memcg->memsw));
1379 return margin >> PAGE_SHIFT;
1382 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1384 struct cgroup *cgrp = memcg->css.cgroup;
1387 if (cgrp->parent == NULL)
1388 return vm_swappiness;
1390 return memcg->swappiness;
1394 * memcg->moving_account is used for checking possibility that some thread is
1395 * calling move_account(). When a thread on CPU-A starts moving pages under
1396 * a memcg, other threads should check memcg->moving_account under
1397 * rcu_read_lock(), like this:
1401 * memcg->moving_account+1 if (memcg->mocing_account)
1403 * synchronize_rcu() update something.
1408 /* for quick checking without looking up memcg */
1409 atomic_t memcg_moving __read_mostly;
1411 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1413 atomic_inc(&memcg_moving);
1414 atomic_inc(&memcg->moving_account);
1418 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1421 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1422 * We check NULL in callee rather than caller.
1425 atomic_dec(&memcg_moving);
1426 atomic_dec(&memcg->moving_account);
1431 * 2 routines for checking "mem" is under move_account() or not.
1433 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1434 * is used for avoiding races in accounting. If true,
1435 * pc->mem_cgroup may be overwritten.
1437 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1438 * under hierarchy of moving cgroups. This is for
1439 * waiting at hith-memory prressure caused by "move".
1442 static bool mem_cgroup_stolen(struct mem_cgroup *memcg)
1444 VM_BUG_ON(!rcu_read_lock_held());
1445 return atomic_read(&memcg->moving_account) > 0;
1448 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1450 struct mem_cgroup *from;
1451 struct mem_cgroup *to;
1454 * Unlike task_move routines, we access mc.to, mc.from not under
1455 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1457 spin_lock(&mc.lock);
1463 ret = mem_cgroup_same_or_subtree(memcg, from)
1464 || mem_cgroup_same_or_subtree(memcg, to);
1466 spin_unlock(&mc.lock);
1470 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1472 if (mc.moving_task && current != mc.moving_task) {
1473 if (mem_cgroup_under_move(memcg)) {
1475 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1476 /* moving charge context might have finished. */
1479 finish_wait(&mc.waitq, &wait);
1487 * Take this lock when
1488 * - a code tries to modify page's memcg while it's USED.
1489 * - a code tries to modify page state accounting in a memcg.
1490 * see mem_cgroup_stolen(), too.
1492 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1493 unsigned long *flags)
1495 spin_lock_irqsave(&memcg->move_lock, *flags);
1498 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1499 unsigned long *flags)
1501 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1505 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1506 * @memcg: The memory cgroup that went over limit
1507 * @p: Task that is going to be killed
1509 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1512 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1514 struct cgroup *task_cgrp;
1515 struct cgroup *mem_cgrp;
1517 * Need a buffer in BSS, can't rely on allocations. The code relies
1518 * on the assumption that OOM is serialized for memory controller.
1519 * If this assumption is broken, revisit this code.
1521 static char memcg_name[PATH_MAX];
1529 mem_cgrp = memcg->css.cgroup;
1530 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1532 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1535 * Unfortunately, we are unable to convert to a useful name
1536 * But we'll still print out the usage information
1543 printk(KERN_INFO "Task in %s killed", memcg_name);
1546 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1554 * Continues from above, so we don't need an KERN_ level
1556 printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1559 printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1560 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1561 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1562 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1563 printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1565 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1566 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1567 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1568 printk(KERN_INFO "kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1569 res_counter_read_u64(&memcg->kmem, RES_USAGE) >> 10,
1570 res_counter_read_u64(&memcg->kmem, RES_LIMIT) >> 10,
1571 res_counter_read_u64(&memcg->kmem, RES_FAILCNT));
1575 * This function returns the number of memcg under hierarchy tree. Returns
1576 * 1(self count) if no children.
1578 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1581 struct mem_cgroup *iter;
1583 for_each_mem_cgroup_tree(iter, memcg)
1589 * Return the memory (and swap, if configured) limit for a memcg.
1591 static u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1596 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1597 limit += total_swap_pages << PAGE_SHIFT;
1599 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1601 * If memsw is finite and limits the amount of swap space available
1602 * to this memcg, return that limit.
1604 return min(limit, memsw);
1607 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1610 struct mem_cgroup *iter;
1611 unsigned long chosen_points = 0;
1612 unsigned long totalpages;
1613 unsigned int points = 0;
1614 struct task_struct *chosen = NULL;
1617 * If current has a pending SIGKILL, then automatically select it. The
1618 * goal is to allow it to allocate so that it may quickly exit and free
1621 if (fatal_signal_pending(current)) {
1622 set_thread_flag(TIF_MEMDIE);
1626 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1627 totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
1628 for_each_mem_cgroup_tree(iter, memcg) {
1629 struct cgroup *cgroup = iter->css.cgroup;
1630 struct cgroup_iter it;
1631 struct task_struct *task;
1633 cgroup_iter_start(cgroup, &it);
1634 while ((task = cgroup_iter_next(cgroup, &it))) {
1635 switch (oom_scan_process_thread(task, totalpages, NULL,
1637 case OOM_SCAN_SELECT:
1639 put_task_struct(chosen);
1641 chosen_points = ULONG_MAX;
1642 get_task_struct(chosen);
1644 case OOM_SCAN_CONTINUE:
1646 case OOM_SCAN_ABORT:
1647 cgroup_iter_end(cgroup, &it);
1648 mem_cgroup_iter_break(memcg, iter);
1650 put_task_struct(chosen);
1655 points = oom_badness(task, memcg, NULL, totalpages);
1656 if (points > chosen_points) {
1658 put_task_struct(chosen);
1660 chosen_points = points;
1661 get_task_struct(chosen);
1664 cgroup_iter_end(cgroup, &it);
1669 points = chosen_points * 1000 / totalpages;
1670 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1671 NULL, "Memory cgroup out of memory");
1674 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1676 unsigned long flags)
1678 unsigned long total = 0;
1679 bool noswap = false;
1682 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1684 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1687 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1689 drain_all_stock_async(memcg);
1690 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1692 * Allow limit shrinkers, which are triggered directly
1693 * by userspace, to catch signals and stop reclaim
1694 * after minimal progress, regardless of the margin.
1696 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1698 if (mem_cgroup_margin(memcg))
1701 * If nothing was reclaimed after two attempts, there
1702 * may be no reclaimable pages in this hierarchy.
1711 * test_mem_cgroup_node_reclaimable
1712 * @memcg: the target memcg
1713 * @nid: the node ID to be checked.
1714 * @noswap : specify true here if the user wants flle only information.
1716 * This function returns whether the specified memcg contains any
1717 * reclaimable pages on a node. Returns true if there are any reclaimable
1718 * pages in the node.
1720 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1721 int nid, bool noswap)
1723 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1725 if (noswap || !total_swap_pages)
1727 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1732 #if MAX_NUMNODES > 1
1735 * Always updating the nodemask is not very good - even if we have an empty
1736 * list or the wrong list here, we can start from some node and traverse all
1737 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1740 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1744 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1745 * pagein/pageout changes since the last update.
1747 if (!atomic_read(&memcg->numainfo_events))
1749 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1752 /* make a nodemask where this memcg uses memory from */
1753 memcg->scan_nodes = node_states[N_HIGH_MEMORY];
1755 for_each_node_mask(nid, node_states[N_HIGH_MEMORY]) {
1757 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1758 node_clear(nid, memcg->scan_nodes);
1761 atomic_set(&memcg->numainfo_events, 0);
1762 atomic_set(&memcg->numainfo_updating, 0);
1766 * Selecting a node where we start reclaim from. Because what we need is just
1767 * reducing usage counter, start from anywhere is O,K. Considering
1768 * memory reclaim from current node, there are pros. and cons.
1770 * Freeing memory from current node means freeing memory from a node which
1771 * we'll use or we've used. So, it may make LRU bad. And if several threads
1772 * hit limits, it will see a contention on a node. But freeing from remote
1773 * node means more costs for memory reclaim because of memory latency.
1775 * Now, we use round-robin. Better algorithm is welcomed.
1777 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1781 mem_cgroup_may_update_nodemask(memcg);
1782 node = memcg->last_scanned_node;
1784 node = next_node(node, memcg->scan_nodes);
1785 if (node == MAX_NUMNODES)
1786 node = first_node(memcg->scan_nodes);
1788 * We call this when we hit limit, not when pages are added to LRU.
1789 * No LRU may hold pages because all pages are UNEVICTABLE or
1790 * memcg is too small and all pages are not on LRU. In that case,
1791 * we use curret node.
1793 if (unlikely(node == MAX_NUMNODES))
1794 node = numa_node_id();
1796 memcg->last_scanned_node = node;
1801 * Check all nodes whether it contains reclaimable pages or not.
1802 * For quick scan, we make use of scan_nodes. This will allow us to skip
1803 * unused nodes. But scan_nodes is lazily updated and may not cotain
1804 * enough new information. We need to do double check.
1806 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1811 * quick check...making use of scan_node.
1812 * We can skip unused nodes.
1814 if (!nodes_empty(memcg->scan_nodes)) {
1815 for (nid = first_node(memcg->scan_nodes);
1817 nid = next_node(nid, memcg->scan_nodes)) {
1819 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1824 * Check rest of nodes.
1826 for_each_node_state(nid, N_HIGH_MEMORY) {
1827 if (node_isset(nid, memcg->scan_nodes))
1829 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1836 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1841 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1843 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
1847 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1850 unsigned long *total_scanned)
1852 struct mem_cgroup *victim = NULL;
1855 unsigned long excess;
1856 unsigned long nr_scanned;
1857 struct mem_cgroup_reclaim_cookie reclaim = {
1862 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
1865 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1870 * If we have not been able to reclaim
1871 * anything, it might because there are
1872 * no reclaimable pages under this hierarchy
1877 * We want to do more targeted reclaim.
1878 * excess >> 2 is not to excessive so as to
1879 * reclaim too much, nor too less that we keep
1880 * coming back to reclaim from this cgroup
1882 if (total >= (excess >> 2) ||
1883 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1888 if (!mem_cgroup_reclaimable(victim, false))
1890 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1892 *total_scanned += nr_scanned;
1893 if (!res_counter_soft_limit_excess(&root_memcg->res))
1896 mem_cgroup_iter_break(root_memcg, victim);
1901 * Check OOM-Killer is already running under our hierarchy.
1902 * If someone is running, return false.
1903 * Has to be called with memcg_oom_lock
1905 static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg)
1907 struct mem_cgroup *iter, *failed = NULL;
1909 for_each_mem_cgroup_tree(iter, memcg) {
1910 if (iter->oom_lock) {
1912 * this subtree of our hierarchy is already locked
1913 * so we cannot give a lock.
1916 mem_cgroup_iter_break(memcg, iter);
1919 iter->oom_lock = true;
1926 * OK, we failed to lock the whole subtree so we have to clean up
1927 * what we set up to the failing subtree
1929 for_each_mem_cgroup_tree(iter, memcg) {
1930 if (iter == failed) {
1931 mem_cgroup_iter_break(memcg, iter);
1934 iter->oom_lock = false;
1940 * Has to be called with memcg_oom_lock
1942 static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1944 struct mem_cgroup *iter;
1946 for_each_mem_cgroup_tree(iter, memcg)
1947 iter->oom_lock = false;
1951 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1953 struct mem_cgroup *iter;
1955 for_each_mem_cgroup_tree(iter, memcg)
1956 atomic_inc(&iter->under_oom);
1959 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1961 struct mem_cgroup *iter;
1964 * When a new child is created while the hierarchy is under oom,
1965 * mem_cgroup_oom_lock() may not be called. We have to use
1966 * atomic_add_unless() here.
1968 for_each_mem_cgroup_tree(iter, memcg)
1969 atomic_add_unless(&iter->under_oom, -1, 0);
1972 static DEFINE_SPINLOCK(memcg_oom_lock);
1973 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1975 struct oom_wait_info {
1976 struct mem_cgroup *memcg;
1980 static int memcg_oom_wake_function(wait_queue_t *wait,
1981 unsigned mode, int sync, void *arg)
1983 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1984 struct mem_cgroup *oom_wait_memcg;
1985 struct oom_wait_info *oom_wait_info;
1987 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1988 oom_wait_memcg = oom_wait_info->memcg;
1991 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
1992 * Then we can use css_is_ancestor without taking care of RCU.
1994 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
1995 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
1997 return autoremove_wake_function(wait, mode, sync, arg);
2000 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
2002 /* for filtering, pass "memcg" as argument. */
2003 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
2006 static void memcg_oom_recover(struct mem_cgroup *memcg)
2008 if (memcg && atomic_read(&memcg->under_oom))
2009 memcg_wakeup_oom(memcg);
2013 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
2015 static bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask,
2018 struct oom_wait_info owait;
2019 bool locked, need_to_kill;
2021 owait.memcg = memcg;
2022 owait.wait.flags = 0;
2023 owait.wait.func = memcg_oom_wake_function;
2024 owait.wait.private = current;
2025 INIT_LIST_HEAD(&owait.wait.task_list);
2026 need_to_kill = true;
2027 mem_cgroup_mark_under_oom(memcg);
2029 /* At first, try to OOM lock hierarchy under memcg.*/
2030 spin_lock(&memcg_oom_lock);
2031 locked = mem_cgroup_oom_lock(memcg);
2033 * Even if signal_pending(), we can't quit charge() loop without
2034 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
2035 * under OOM is always welcomed, use TASK_KILLABLE here.
2037 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2038 if (!locked || memcg->oom_kill_disable)
2039 need_to_kill = false;
2041 mem_cgroup_oom_notify(memcg);
2042 spin_unlock(&memcg_oom_lock);
2045 finish_wait(&memcg_oom_waitq, &owait.wait);
2046 mem_cgroup_out_of_memory(memcg, mask, order);
2049 finish_wait(&memcg_oom_waitq, &owait.wait);
2051 spin_lock(&memcg_oom_lock);
2053 mem_cgroup_oom_unlock(memcg);
2054 memcg_wakeup_oom(memcg);
2055 spin_unlock(&memcg_oom_lock);
2057 mem_cgroup_unmark_under_oom(memcg);
2059 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
2061 /* Give chance to dying process */
2062 schedule_timeout_uninterruptible(1);
2067 * Currently used to update mapped file statistics, but the routine can be
2068 * generalized to update other statistics as well.
2070 * Notes: Race condition
2072 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2073 * it tends to be costly. But considering some conditions, we doesn't need
2074 * to do so _always_.
2076 * Considering "charge", lock_page_cgroup() is not required because all
2077 * file-stat operations happen after a page is attached to radix-tree. There
2078 * are no race with "charge".
2080 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2081 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2082 * if there are race with "uncharge". Statistics itself is properly handled
2085 * Considering "move", this is an only case we see a race. To make the race
2086 * small, we check mm->moving_account and detect there are possibility of race
2087 * If there is, we take a lock.
2090 void __mem_cgroup_begin_update_page_stat(struct page *page,
2091 bool *locked, unsigned long *flags)
2093 struct mem_cgroup *memcg;
2094 struct page_cgroup *pc;
2096 pc = lookup_page_cgroup(page);
2098 memcg = pc->mem_cgroup;
2099 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2102 * If this memory cgroup is not under account moving, we don't
2103 * need to take move_lock_mem_cgroup(). Because we already hold
2104 * rcu_read_lock(), any calls to move_account will be delayed until
2105 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2107 if (!mem_cgroup_stolen(memcg))
2110 move_lock_mem_cgroup(memcg, flags);
2111 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2112 move_unlock_mem_cgroup(memcg, flags);
2118 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
2120 struct page_cgroup *pc = lookup_page_cgroup(page);
2123 * It's guaranteed that pc->mem_cgroup never changes while
2124 * lock is held because a routine modifies pc->mem_cgroup
2125 * should take move_lock_mem_cgroup().
2127 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
2130 void mem_cgroup_update_page_stat(struct page *page,
2131 enum mem_cgroup_page_stat_item idx, int val)
2133 struct mem_cgroup *memcg;
2134 struct page_cgroup *pc = lookup_page_cgroup(page);
2135 unsigned long uninitialized_var(flags);
2137 if (mem_cgroup_disabled())
2140 memcg = pc->mem_cgroup;
2141 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2145 case MEMCG_NR_FILE_MAPPED:
2146 idx = MEM_CGROUP_STAT_FILE_MAPPED;
2152 this_cpu_add(memcg->stat->count[idx], val);
2156 * size of first charge trial. "32" comes from vmscan.c's magic value.
2157 * TODO: maybe necessary to use big numbers in big irons.
2159 #define CHARGE_BATCH 32U
2160 struct memcg_stock_pcp {
2161 struct mem_cgroup *cached; /* this never be root cgroup */
2162 unsigned int nr_pages;
2163 struct work_struct work;
2164 unsigned long flags;
2165 #define FLUSHING_CACHED_CHARGE 0
2167 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2168 static DEFINE_MUTEX(percpu_charge_mutex);
2171 * consume_stock: Try to consume stocked charge on this cpu.
2172 * @memcg: memcg to consume from.
2173 * @nr_pages: how many pages to charge.
2175 * The charges will only happen if @memcg matches the current cpu's memcg
2176 * stock, and at least @nr_pages are available in that stock. Failure to
2177 * service an allocation will refill the stock.
2179 * returns true if successful, false otherwise.
2181 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2183 struct memcg_stock_pcp *stock;
2186 if (nr_pages > CHARGE_BATCH)
2189 stock = &get_cpu_var(memcg_stock);
2190 if (memcg == stock->cached && stock->nr_pages >= nr_pages)
2191 stock->nr_pages -= nr_pages;
2192 else /* need to call res_counter_charge */
2194 put_cpu_var(memcg_stock);
2199 * Returns stocks cached in percpu to res_counter and reset cached information.
2201 static void drain_stock(struct memcg_stock_pcp *stock)
2203 struct mem_cgroup *old = stock->cached;
2205 if (stock->nr_pages) {
2206 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2208 res_counter_uncharge(&old->res, bytes);
2209 if (do_swap_account)
2210 res_counter_uncharge(&old->memsw, bytes);
2211 stock->nr_pages = 0;
2213 stock->cached = NULL;
2217 * This must be called under preempt disabled or must be called by
2218 * a thread which is pinned to local cpu.
2220 static void drain_local_stock(struct work_struct *dummy)
2222 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2224 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2228 * Cache charges(val) which is from res_counter, to local per_cpu area.
2229 * This will be consumed by consume_stock() function, later.
2231 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2233 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2235 if (stock->cached != memcg) { /* reset if necessary */
2237 stock->cached = memcg;
2239 stock->nr_pages += nr_pages;
2240 put_cpu_var(memcg_stock);
2244 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2245 * of the hierarchy under it. sync flag says whether we should block
2246 * until the work is done.
2248 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2252 /* Notify other cpus that system-wide "drain" is running */
2255 for_each_online_cpu(cpu) {
2256 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2257 struct mem_cgroup *memcg;
2259 memcg = stock->cached;
2260 if (!memcg || !stock->nr_pages)
2262 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2264 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2266 drain_local_stock(&stock->work);
2268 schedule_work_on(cpu, &stock->work);
2276 for_each_online_cpu(cpu) {
2277 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2278 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2279 flush_work(&stock->work);
2286 * Tries to drain stocked charges in other cpus. This function is asynchronous
2287 * and just put a work per cpu for draining localy on each cpu. Caller can
2288 * expects some charges will be back to res_counter later but cannot wait for
2291 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2294 * If someone calls draining, avoid adding more kworker runs.
2296 if (!mutex_trylock(&percpu_charge_mutex))
2298 drain_all_stock(root_memcg, false);
2299 mutex_unlock(&percpu_charge_mutex);
2302 /* This is a synchronous drain interface. */
2303 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2305 /* called when force_empty is called */
2306 mutex_lock(&percpu_charge_mutex);
2307 drain_all_stock(root_memcg, true);
2308 mutex_unlock(&percpu_charge_mutex);
2312 * This function drains percpu counter value from DEAD cpu and
2313 * move it to local cpu. Note that this function can be preempted.
2315 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2319 spin_lock(&memcg->pcp_counter_lock);
2320 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2321 long x = per_cpu(memcg->stat->count[i], cpu);
2323 per_cpu(memcg->stat->count[i], cpu) = 0;
2324 memcg->nocpu_base.count[i] += x;
2326 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2327 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2329 per_cpu(memcg->stat->events[i], cpu) = 0;
2330 memcg->nocpu_base.events[i] += x;
2332 spin_unlock(&memcg->pcp_counter_lock);
2335 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2336 unsigned long action,
2339 int cpu = (unsigned long)hcpu;
2340 struct memcg_stock_pcp *stock;
2341 struct mem_cgroup *iter;
2343 if (action == CPU_ONLINE)
2346 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2349 for_each_mem_cgroup(iter)
2350 mem_cgroup_drain_pcp_counter(iter, cpu);
2352 stock = &per_cpu(memcg_stock, cpu);
2358 /* See __mem_cgroup_try_charge() for details */
2360 CHARGE_OK, /* success */
2361 CHARGE_RETRY, /* need to retry but retry is not bad */
2362 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2363 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2364 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2367 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2368 unsigned int nr_pages, unsigned int min_pages,
2371 unsigned long csize = nr_pages * PAGE_SIZE;
2372 struct mem_cgroup *mem_over_limit;
2373 struct res_counter *fail_res;
2374 unsigned long flags = 0;
2377 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2380 if (!do_swap_account)
2382 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2386 res_counter_uncharge(&memcg->res, csize);
2387 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2388 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2390 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2392 * Never reclaim on behalf of optional batching, retry with a
2393 * single page instead.
2395 if (nr_pages > min_pages)
2396 return CHARGE_RETRY;
2398 if (!(gfp_mask & __GFP_WAIT))
2399 return CHARGE_WOULDBLOCK;
2401 if (gfp_mask & __GFP_NORETRY)
2402 return CHARGE_NOMEM;
2404 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2405 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2406 return CHARGE_RETRY;
2408 * Even though the limit is exceeded at this point, reclaim
2409 * may have been able to free some pages. Retry the charge
2410 * before killing the task.
2412 * Only for regular pages, though: huge pages are rather
2413 * unlikely to succeed so close to the limit, and we fall back
2414 * to regular pages anyway in case of failure.
2416 if (nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER) && ret)
2417 return CHARGE_RETRY;
2420 * At task move, charge accounts can be doubly counted. So, it's
2421 * better to wait until the end of task_move if something is going on.
2423 if (mem_cgroup_wait_acct_move(mem_over_limit))
2424 return CHARGE_RETRY;
2426 /* If we don't need to call oom-killer at el, return immediately */
2428 return CHARGE_NOMEM;
2430 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask, get_order(csize)))
2431 return CHARGE_OOM_DIE;
2433 return CHARGE_RETRY;
2437 * __mem_cgroup_try_charge() does
2438 * 1. detect memcg to be charged against from passed *mm and *ptr,
2439 * 2. update res_counter
2440 * 3. call memory reclaim if necessary.
2442 * In some special case, if the task is fatal, fatal_signal_pending() or
2443 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2444 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2445 * as possible without any hazards. 2: all pages should have a valid
2446 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2447 * pointer, that is treated as a charge to root_mem_cgroup.
2449 * So __mem_cgroup_try_charge() will return
2450 * 0 ... on success, filling *ptr with a valid memcg pointer.
2451 * -ENOMEM ... charge failure because of resource limits.
2452 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2454 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2455 * the oom-killer can be invoked.
2457 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2459 unsigned int nr_pages,
2460 struct mem_cgroup **ptr,
2463 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2464 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2465 struct mem_cgroup *memcg = NULL;
2469 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2470 * in system level. So, allow to go ahead dying process in addition to
2473 if (unlikely(test_thread_flag(TIF_MEMDIE)
2474 || fatal_signal_pending(current)))
2478 * We always charge the cgroup the mm_struct belongs to.
2479 * The mm_struct's mem_cgroup changes on task migration if the
2480 * thread group leader migrates. It's possible that mm is not
2481 * set, if so charge the root memcg (happens for pagecache usage).
2484 *ptr = root_mem_cgroup;
2486 if (*ptr) { /* css should be a valid one */
2488 if (mem_cgroup_is_root(memcg))
2490 if (consume_stock(memcg, nr_pages))
2492 css_get(&memcg->css);
2494 struct task_struct *p;
2497 p = rcu_dereference(mm->owner);
2499 * Because we don't have task_lock(), "p" can exit.
2500 * In that case, "memcg" can point to root or p can be NULL with
2501 * race with swapoff. Then, we have small risk of mis-accouning.
2502 * But such kind of mis-account by race always happens because
2503 * we don't have cgroup_mutex(). It's overkill and we allo that
2505 * (*) swapoff at el will charge against mm-struct not against
2506 * task-struct. So, mm->owner can be NULL.
2508 memcg = mem_cgroup_from_task(p);
2510 memcg = root_mem_cgroup;
2511 if (mem_cgroup_is_root(memcg)) {
2515 if (consume_stock(memcg, nr_pages)) {
2517 * It seems dagerous to access memcg without css_get().
2518 * But considering how consume_stok works, it's not
2519 * necessary. If consume_stock success, some charges
2520 * from this memcg are cached on this cpu. So, we
2521 * don't need to call css_get()/css_tryget() before
2522 * calling consume_stock().
2527 /* after here, we may be blocked. we need to get refcnt */
2528 if (!css_tryget(&memcg->css)) {
2538 /* If killed, bypass charge */
2539 if (fatal_signal_pending(current)) {
2540 css_put(&memcg->css);
2545 if (oom && !nr_oom_retries) {
2547 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2550 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, nr_pages,
2555 case CHARGE_RETRY: /* not in OOM situation but retry */
2557 css_put(&memcg->css);
2560 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2561 css_put(&memcg->css);
2563 case CHARGE_NOMEM: /* OOM routine works */
2565 css_put(&memcg->css);
2568 /* If oom, we never return -ENOMEM */
2571 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2572 css_put(&memcg->css);
2575 } while (ret != CHARGE_OK);
2577 if (batch > nr_pages)
2578 refill_stock(memcg, batch - nr_pages);
2579 css_put(&memcg->css);
2587 *ptr = root_mem_cgroup;
2592 * Somemtimes we have to undo a charge we got by try_charge().
2593 * This function is for that and do uncharge, put css's refcnt.
2594 * gotten by try_charge().
2596 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2597 unsigned int nr_pages)
2599 if (!mem_cgroup_is_root(memcg)) {
2600 unsigned long bytes = nr_pages * PAGE_SIZE;
2602 res_counter_uncharge(&memcg->res, bytes);
2603 if (do_swap_account)
2604 res_counter_uncharge(&memcg->memsw, bytes);
2609 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2610 * This is useful when moving usage to parent cgroup.
2612 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2613 unsigned int nr_pages)
2615 unsigned long bytes = nr_pages * PAGE_SIZE;
2617 if (mem_cgroup_is_root(memcg))
2620 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2621 if (do_swap_account)
2622 res_counter_uncharge_until(&memcg->memsw,
2623 memcg->memsw.parent, bytes);
2627 * A helper function to get mem_cgroup from ID. must be called under
2628 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2629 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2630 * called against removed memcg.)
2632 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2634 struct cgroup_subsys_state *css;
2636 /* ID 0 is unused ID */
2639 css = css_lookup(&mem_cgroup_subsys, id);
2642 return mem_cgroup_from_css(css);
2645 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2647 struct mem_cgroup *memcg = NULL;
2648 struct page_cgroup *pc;
2652 VM_BUG_ON(!PageLocked(page));
2654 pc = lookup_page_cgroup(page);
2655 lock_page_cgroup(pc);
2656 if (PageCgroupUsed(pc)) {
2657 memcg = pc->mem_cgroup;
2658 if (memcg && !css_tryget(&memcg->css))
2660 } else if (PageSwapCache(page)) {
2661 ent.val = page_private(page);
2662 id = lookup_swap_cgroup_id(ent);
2664 memcg = mem_cgroup_lookup(id);
2665 if (memcg && !css_tryget(&memcg->css))
2669 unlock_page_cgroup(pc);
2673 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2675 unsigned int nr_pages,
2676 enum charge_type ctype,
2679 struct page_cgroup *pc = lookup_page_cgroup(page);
2680 struct zone *uninitialized_var(zone);
2681 struct lruvec *lruvec;
2682 bool was_on_lru = false;
2685 lock_page_cgroup(pc);
2686 VM_BUG_ON(PageCgroupUsed(pc));
2688 * we don't need page_cgroup_lock about tail pages, becase they are not
2689 * accessed by any other context at this point.
2693 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2694 * may already be on some other mem_cgroup's LRU. Take care of it.
2697 zone = page_zone(page);
2698 spin_lock_irq(&zone->lru_lock);
2699 if (PageLRU(page)) {
2700 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2702 del_page_from_lru_list(page, lruvec, page_lru(page));
2707 pc->mem_cgroup = memcg;
2709 * We access a page_cgroup asynchronously without lock_page_cgroup().
2710 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2711 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2712 * before USED bit, we need memory barrier here.
2713 * See mem_cgroup_add_lru_list(), etc.
2716 SetPageCgroupUsed(pc);
2720 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2721 VM_BUG_ON(PageLRU(page));
2723 add_page_to_lru_list(page, lruvec, page_lru(page));
2725 spin_unlock_irq(&zone->lru_lock);
2728 if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
2733 mem_cgroup_charge_statistics(memcg, anon, nr_pages);
2734 unlock_page_cgroup(pc);
2737 * "charge_statistics" updated event counter. Then, check it.
2738 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2739 * if they exceeds softlimit.
2741 memcg_check_events(memcg, page);
2744 static DEFINE_MUTEX(set_limit_mutex);
2746 #ifdef CONFIG_MEMCG_KMEM
2747 static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg)
2749 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) &&
2750 (memcg->kmem_account_flags & KMEM_ACCOUNTED_MASK);
2754 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2755 * in the memcg_cache_params struct.
2757 static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
2759 struct kmem_cache *cachep;
2761 VM_BUG_ON(p->is_root_cache);
2762 cachep = p->root_cache;
2763 return cachep->memcg_params->memcg_caches[memcg_cache_id(p->memcg)];
2766 #ifdef CONFIG_SLABINFO
2767 static int mem_cgroup_slabinfo_read(struct cgroup *cont, struct cftype *cft,
2770 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
2771 struct memcg_cache_params *params;
2773 if (!memcg_can_account_kmem(memcg))
2776 print_slabinfo_header(m);
2778 mutex_lock(&memcg->slab_caches_mutex);
2779 list_for_each_entry(params, &memcg->memcg_slab_caches, list)
2780 cache_show(memcg_params_to_cache(params), m);
2781 mutex_unlock(&memcg->slab_caches_mutex);
2787 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size)
2789 struct res_counter *fail_res;
2790 struct mem_cgroup *_memcg;
2794 ret = res_counter_charge(&memcg->kmem, size, &fail_res);
2799 * Conditions under which we can wait for the oom_killer. Those are
2800 * the same conditions tested by the core page allocator
2802 may_oom = (gfp & __GFP_FS) && !(gfp & __GFP_NORETRY);
2805 ret = __mem_cgroup_try_charge(NULL, gfp, size >> PAGE_SHIFT,
2808 if (ret == -EINTR) {
2810 * __mem_cgroup_try_charge() chosed to bypass to root due to
2811 * OOM kill or fatal signal. Since our only options are to
2812 * either fail the allocation or charge it to this cgroup, do
2813 * it as a temporary condition. But we can't fail. From a
2814 * kmem/slab perspective, the cache has already been selected,
2815 * by mem_cgroup_kmem_get_cache(), so it is too late to change
2818 * This condition will only trigger if the task entered
2819 * memcg_charge_kmem in a sane state, but was OOM-killed during
2820 * __mem_cgroup_try_charge() above. Tasks that were already
2821 * dying when the allocation triggers should have been already
2822 * directed to the root cgroup in memcontrol.h
2824 res_counter_charge_nofail(&memcg->res, size, &fail_res);
2825 if (do_swap_account)
2826 res_counter_charge_nofail(&memcg->memsw, size,
2830 res_counter_uncharge(&memcg->kmem, size);
2835 static void memcg_uncharge_kmem(struct mem_cgroup *memcg, u64 size)
2837 res_counter_uncharge(&memcg->res, size);
2838 if (do_swap_account)
2839 res_counter_uncharge(&memcg->memsw, size);
2842 if (res_counter_uncharge(&memcg->kmem, size))
2845 if (memcg_kmem_test_and_clear_dead(memcg))
2846 mem_cgroup_put(memcg);
2849 void memcg_cache_list_add(struct mem_cgroup *memcg, struct kmem_cache *cachep)
2854 mutex_lock(&memcg->slab_caches_mutex);
2855 list_add(&cachep->memcg_params->list, &memcg->memcg_slab_caches);
2856 mutex_unlock(&memcg->slab_caches_mutex);
2860 * helper for acessing a memcg's index. It will be used as an index in the
2861 * child cache array in kmem_cache, and also to derive its name. This function
2862 * will return -1 when this is not a kmem-limited memcg.
2864 int memcg_cache_id(struct mem_cgroup *memcg)
2866 return memcg ? memcg->kmemcg_id : -1;
2870 * This ends up being protected by the set_limit mutex, during normal
2871 * operation, because that is its main call site.
2873 * But when we create a new cache, we can call this as well if its parent
2874 * is kmem-limited. That will have to hold set_limit_mutex as well.
2876 int memcg_update_cache_sizes(struct mem_cgroup *memcg)
2880 num = ida_simple_get(&kmem_limited_groups,
2881 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2885 * After this point, kmem_accounted (that we test atomically in
2886 * the beginning of this conditional), is no longer 0. This
2887 * guarantees only one process will set the following boolean
2888 * to true. We don't need test_and_set because we're protected
2889 * by the set_limit_mutex anyway.
2891 memcg_kmem_set_activated(memcg);
2893 ret = memcg_update_all_caches(num+1);
2895 ida_simple_remove(&kmem_limited_groups, num);
2896 memcg_kmem_clear_activated(memcg);
2900 memcg->kmemcg_id = num;
2901 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
2902 mutex_init(&memcg->slab_caches_mutex);
2906 static size_t memcg_caches_array_size(int num_groups)
2909 if (num_groups <= 0)
2912 size = 2 * num_groups;
2913 if (size < MEMCG_CACHES_MIN_SIZE)
2914 size = MEMCG_CACHES_MIN_SIZE;
2915 else if (size > MEMCG_CACHES_MAX_SIZE)
2916 size = MEMCG_CACHES_MAX_SIZE;
2922 * We should update the current array size iff all caches updates succeed. This
2923 * can only be done from the slab side. The slab mutex needs to be held when
2926 void memcg_update_array_size(int num)
2928 if (num > memcg_limited_groups_array_size)
2929 memcg_limited_groups_array_size = memcg_caches_array_size(num);
2932 int memcg_update_cache_size(struct kmem_cache *s, int num_groups)
2934 struct memcg_cache_params *cur_params = s->memcg_params;
2936 VM_BUG_ON(s->memcg_params && !s->memcg_params->is_root_cache);
2938 if (num_groups > memcg_limited_groups_array_size) {
2940 ssize_t size = memcg_caches_array_size(num_groups);
2942 size *= sizeof(void *);
2943 size += sizeof(struct memcg_cache_params);
2945 s->memcg_params = kzalloc(size, GFP_KERNEL);
2946 if (!s->memcg_params) {
2947 s->memcg_params = cur_params;
2951 s->memcg_params->is_root_cache = true;
2954 * There is the chance it will be bigger than
2955 * memcg_limited_groups_array_size, if we failed an allocation
2956 * in a cache, in which case all caches updated before it, will
2957 * have a bigger array.
2959 * But if that is the case, the data after
2960 * memcg_limited_groups_array_size is certainly unused
2962 for (i = 0; i < memcg_limited_groups_array_size; i++) {
2963 if (!cur_params->memcg_caches[i])
2965 s->memcg_params->memcg_caches[i] =
2966 cur_params->memcg_caches[i];
2970 * Ideally, we would wait until all caches succeed, and only
2971 * then free the old one. But this is not worth the extra
2972 * pointer per-cache we'd have to have for this.
2974 * It is not a big deal if some caches are left with a size
2975 * bigger than the others. And all updates will reset this
2983 int memcg_register_cache(struct mem_cgroup *memcg, struct kmem_cache *s,
2984 struct kmem_cache *root_cache)
2986 size_t size = sizeof(struct memcg_cache_params);
2988 if (!memcg_kmem_enabled())
2992 size += memcg_limited_groups_array_size * sizeof(void *);
2994 s->memcg_params = kzalloc(size, GFP_KERNEL);
2995 if (!s->memcg_params)
2999 s->memcg_params->memcg = memcg;
3000 s->memcg_params->root_cache = root_cache;
3005 void memcg_release_cache(struct kmem_cache *s)
3007 struct kmem_cache *root;
3008 struct mem_cgroup *memcg;
3012 * This happens, for instance, when a root cache goes away before we
3015 if (!s->memcg_params)
3018 if (s->memcg_params->is_root_cache)
3021 memcg = s->memcg_params->memcg;
3022 id = memcg_cache_id(memcg);
3024 root = s->memcg_params->root_cache;
3025 root->memcg_params->memcg_caches[id] = NULL;
3026 mem_cgroup_put(memcg);
3028 mutex_lock(&memcg->slab_caches_mutex);
3029 list_del(&s->memcg_params->list);
3030 mutex_unlock(&memcg->slab_caches_mutex);
3033 kfree(s->memcg_params);
3037 * During the creation a new cache, we need to disable our accounting mechanism
3038 * altogether. This is true even if we are not creating, but rather just
3039 * enqueing new caches to be created.
3041 * This is because that process will trigger allocations; some visible, like
3042 * explicit kmallocs to auxiliary data structures, name strings and internal
3043 * cache structures; some well concealed, like INIT_WORK() that can allocate
3044 * objects during debug.
3046 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3047 * to it. This may not be a bounded recursion: since the first cache creation
3048 * failed to complete (waiting on the allocation), we'll just try to create the
3049 * cache again, failing at the same point.
3051 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3052 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3053 * inside the following two functions.
3055 static inline void memcg_stop_kmem_account(void)
3060 current->memcg_kmem_skip_account++;
3063 static inline void memcg_resume_kmem_account(void)
3068 current->memcg_kmem_skip_account--;
3071 static void kmem_cache_destroy_work_func(struct work_struct *w)
3073 struct kmem_cache *cachep;
3074 struct memcg_cache_params *p;
3075 struct delayed_work *dw = to_delayed_work(w);
3077 p = container_of(dw, struct memcg_cache_params, destroy);
3079 cachep = memcg_params_to_cache(p);
3082 * If we get down to 0 after shrink, we could delete right away.
3083 * However, memcg_release_pages() already puts us back in the workqueue
3084 * in that case. If we proceed deleting, we'll get a dangling
3085 * reference, and removing the object from the workqueue in that case
3086 * is unnecessary complication. We are not a fast path.
3088 * Note that this case is fundamentally different from racing with
3089 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3090 * kmem_cache_shrink, not only we would be reinserting a dead cache
3091 * into the queue, but doing so from inside the worker racing to
3094 * So if we aren't down to zero, we'll just schedule a worker and try
3097 if (atomic_read(&cachep->memcg_params->nr_pages) != 0) {
3098 kmem_cache_shrink(cachep);
3099 if (atomic_read(&cachep->memcg_params->nr_pages) == 0)
3101 /* Once per minute should be good enough. */
3102 schedule_delayed_work(&cachep->memcg_params->destroy, 60 * HZ);
3104 kmem_cache_destroy(cachep);
3107 void mem_cgroup_destroy_cache(struct kmem_cache *cachep)
3109 if (!cachep->memcg_params->dead)
3113 * There are many ways in which we can get here.
3115 * We can get to a memory-pressure situation while the delayed work is
3116 * still pending to run. The vmscan shrinkers can then release all
3117 * cache memory and get us to destruction. If this is the case, we'll
3118 * be executed twice, which is a bug (the second time will execute over
3119 * bogus data). In this case, cancelling the work should be fine.
3121 * But we can also get here from the worker itself, if
3122 * kmem_cache_shrink is enough to shake all the remaining objects and
3123 * get the page count to 0. In this case, we'll deadlock if we try to
3124 * cancel the work (the worker runs with an internal lock held, which
3125 * is the same lock we would hold for cancel_delayed_work_sync().)
3127 * Since we can't possibly know who got us here, just refrain from
3128 * running if there is already work pending
3130 if (delayed_work_pending(&cachep->memcg_params->destroy))
3133 * We have to defer the actual destroying to a workqueue, because
3134 * we might currently be in a context that cannot sleep.
3136 schedule_delayed_work(&cachep->memcg_params->destroy, 0);
3139 static char *memcg_cache_name(struct mem_cgroup *memcg, struct kmem_cache *s)
3142 struct dentry *dentry;
3145 dentry = rcu_dereference(memcg->css.cgroup->dentry);
3148 BUG_ON(dentry == NULL);
3150 name = kasprintf(GFP_KERNEL, "%s(%d:%s)", s->name,
3151 memcg_cache_id(memcg), dentry->d_name.name);
3156 static struct kmem_cache *kmem_cache_dup(struct mem_cgroup *memcg,
3157 struct kmem_cache *s)
3160 struct kmem_cache *new;
3162 name = memcg_cache_name(memcg, s);
3166 new = kmem_cache_create_memcg(memcg, name, s->object_size, s->align,
3167 (s->flags & ~SLAB_PANIC), s->ctor, s);
3170 new->allocflags |= __GFP_KMEMCG;
3177 * This lock protects updaters, not readers. We want readers to be as fast as
3178 * they can, and they will either see NULL or a valid cache value. Our model
3179 * allow them to see NULL, in which case the root memcg will be selected.
3181 * We need this lock because multiple allocations to the same cache from a non
3182 * will span more than one worker. Only one of them can create the cache.
3184 static DEFINE_MUTEX(memcg_cache_mutex);
3185 static struct kmem_cache *memcg_create_kmem_cache(struct mem_cgroup *memcg,
3186 struct kmem_cache *cachep)
3188 struct kmem_cache *new_cachep;
3191 BUG_ON(!memcg_can_account_kmem(memcg));
3193 idx = memcg_cache_id(memcg);
3195 mutex_lock(&memcg_cache_mutex);
3196 new_cachep = cachep->memcg_params->memcg_caches[idx];
3200 /* Don't block progress to enqueue caches for internal infrastructure */
3201 memcg_stop_kmem_account();
3202 new_cachep = kmem_cache_dup(memcg, cachep);
3203 memcg_resume_kmem_account();
3205 if (new_cachep == NULL) {
3206 new_cachep = cachep;
3210 mem_cgroup_get(memcg);
3211 atomic_set(&new_cachep->memcg_params->nr_pages , 0);
3213 cachep->memcg_params->memcg_caches[idx] = new_cachep;
3215 * the readers won't lock, make sure everybody sees the updated value,
3216 * so they won't put stuff in the queue again for no reason
3220 mutex_unlock(&memcg_cache_mutex);
3224 void kmem_cache_destroy_memcg_children(struct kmem_cache *s)
3226 struct kmem_cache *c;
3229 if (!s->memcg_params)
3231 if (!s->memcg_params->is_root_cache)
3235 * If the cache is being destroyed, we trust that there is no one else
3236 * requesting objects from it. Even if there are, the sanity checks in
3237 * kmem_cache_destroy should caught this ill-case.
3239 * Still, we don't want anyone else freeing memcg_caches under our
3240 * noses, which can happen if a new memcg comes to life. As usual,
3241 * we'll take the set_limit_mutex to protect ourselves against this.
3243 mutex_lock(&set_limit_mutex);
3244 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3245 c = s->memcg_params->memcg_caches[i];
3250 * We will now manually delete the caches, so to avoid races
3251 * we need to cancel all pending destruction workers and
3252 * proceed with destruction ourselves.
3254 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3255 * and that could spawn the workers again: it is likely that
3256 * the cache still have active pages until this very moment.
3257 * This would lead us back to mem_cgroup_destroy_cache.
3259 * But that will not execute at all if the "dead" flag is not
3260 * set, so flip it down to guarantee we are in control.
3262 c->memcg_params->dead = false;
3263 cancel_delayed_work_sync(&c->memcg_params->destroy);
3264 kmem_cache_destroy(c);
3266 mutex_unlock(&set_limit_mutex);
3269 struct create_work {
3270 struct mem_cgroup *memcg;
3271 struct kmem_cache *cachep;
3272 struct work_struct work;
3275 static void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3277 struct kmem_cache *cachep;
3278 struct memcg_cache_params *params;
3280 if (!memcg_kmem_is_active(memcg))
3283 mutex_lock(&memcg->slab_caches_mutex);
3284 list_for_each_entry(params, &memcg->memcg_slab_caches, list) {
3285 cachep = memcg_params_to_cache(params);
3286 cachep->memcg_params->dead = true;
3287 INIT_DELAYED_WORK(&cachep->memcg_params->destroy,
3288 kmem_cache_destroy_work_func);
3289 schedule_delayed_work(&cachep->memcg_params->destroy, 0);
3291 mutex_unlock(&memcg->slab_caches_mutex);
3294 static void memcg_create_cache_work_func(struct work_struct *w)
3296 struct create_work *cw;
3298 cw = container_of(w, struct create_work, work);
3299 memcg_create_kmem_cache(cw->memcg, cw->cachep);
3300 /* Drop the reference gotten when we enqueued. */
3301 css_put(&cw->memcg->css);
3306 * Enqueue the creation of a per-memcg kmem_cache.
3307 * Called with rcu_read_lock.
3309 static void __memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3310 struct kmem_cache *cachep)
3312 struct create_work *cw;
3314 cw = kmalloc(sizeof(struct create_work), GFP_NOWAIT);
3318 /* The corresponding put will be done in the workqueue. */
3319 if (!css_tryget(&memcg->css)) {
3325 cw->cachep = cachep;
3327 INIT_WORK(&cw->work, memcg_create_cache_work_func);
3328 schedule_work(&cw->work);
3331 static void memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3332 struct kmem_cache *cachep)
3335 * We need to stop accounting when we kmalloc, because if the
3336 * corresponding kmalloc cache is not yet created, the first allocation
3337 * in __memcg_create_cache_enqueue will recurse.
3339 * However, it is better to enclose the whole function. Depending on
3340 * the debugging options enabled, INIT_WORK(), for instance, can
3341 * trigger an allocation. This too, will make us recurse. Because at
3342 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3343 * the safest choice is to do it like this, wrapping the whole function.
3345 memcg_stop_kmem_account();
3346 __memcg_create_cache_enqueue(memcg, cachep);
3347 memcg_resume_kmem_account();
3350 * Return the kmem_cache we're supposed to use for a slab allocation.
3351 * We try to use the current memcg's version of the cache.
3353 * If the cache does not exist yet, if we are the first user of it,
3354 * we either create it immediately, if possible, or create it asynchronously
3356 * In the latter case, we will let the current allocation go through with
3357 * the original cache.
3359 * Can't be called in interrupt context or from kernel threads.
3360 * This function needs to be called with rcu_read_lock() held.
3362 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
3365 struct mem_cgroup *memcg;
3368 VM_BUG_ON(!cachep->memcg_params);
3369 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
3371 if (!current->mm || current->memcg_kmem_skip_account)
3375 memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
3378 if (!memcg_can_account_kmem(memcg))
3381 idx = memcg_cache_id(memcg);
3384 * barrier to mare sure we're always seeing the up to date value. The
3385 * code updating memcg_caches will issue a write barrier to match this.
3387 read_barrier_depends();
3388 if (unlikely(cachep->memcg_params->memcg_caches[idx] == NULL)) {
3390 * If we are in a safe context (can wait, and not in interrupt
3391 * context), we could be be predictable and return right away.
3392 * This would guarantee that the allocation being performed
3393 * already belongs in the new cache.
3395 * However, there are some clashes that can arrive from locking.
3396 * For instance, because we acquire the slab_mutex while doing
3397 * kmem_cache_dup, this means no further allocation could happen
3398 * with the slab_mutex held.
3400 * Also, because cache creation issue get_online_cpus(), this
3401 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3402 * that ends up reversed during cpu hotplug. (cpuset allocates
3403 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3404 * better to defer everything.
3406 memcg_create_cache_enqueue(memcg, cachep);
3410 return cachep->memcg_params->memcg_caches[idx];
3412 EXPORT_SYMBOL(__memcg_kmem_get_cache);
3415 * We need to verify if the allocation against current->mm->owner's memcg is
3416 * possible for the given order. But the page is not allocated yet, so we'll
3417 * need a further commit step to do the final arrangements.
3419 * It is possible for the task to switch cgroups in this mean time, so at
3420 * commit time, we can't rely on task conversion any longer. We'll then use
3421 * the handle argument to return to the caller which cgroup we should commit
3422 * against. We could also return the memcg directly and avoid the pointer
3423 * passing, but a boolean return value gives better semantics considering
3424 * the compiled-out case as well.
3426 * Returning true means the allocation is possible.
3429 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3431 struct mem_cgroup *memcg;
3435 memcg = try_get_mem_cgroup_from_mm(current->mm);
3438 * very rare case described in mem_cgroup_from_task. Unfortunately there
3439 * isn't much we can do without complicating this too much, and it would
3440 * be gfp-dependent anyway. Just let it go
3442 if (unlikely(!memcg))
3445 if (!memcg_can_account_kmem(memcg)) {
3446 css_put(&memcg->css);
3450 ret = memcg_charge_kmem(memcg, gfp, PAGE_SIZE << order);
3454 css_put(&memcg->css);
3458 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3461 struct page_cgroup *pc;
3463 VM_BUG_ON(mem_cgroup_is_root(memcg));
3465 /* The page allocation failed. Revert */
3467 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3471 pc = lookup_page_cgroup(page);
3472 lock_page_cgroup(pc);
3473 pc->mem_cgroup = memcg;
3474 SetPageCgroupUsed(pc);
3475 unlock_page_cgroup(pc);
3478 void __memcg_kmem_uncharge_pages(struct page *page, int order)
3480 struct mem_cgroup *memcg = NULL;
3481 struct page_cgroup *pc;
3484 pc = lookup_page_cgroup(page);
3486 * Fast unlocked return. Theoretically might have changed, have to
3487 * check again after locking.
3489 if (!PageCgroupUsed(pc))
3492 lock_page_cgroup(pc);
3493 if (PageCgroupUsed(pc)) {
3494 memcg = pc->mem_cgroup;
3495 ClearPageCgroupUsed(pc);
3497 unlock_page_cgroup(pc);
3500 * We trust that only if there is a memcg associated with the page, it
3501 * is a valid allocation
3506 VM_BUG_ON(mem_cgroup_is_root(memcg));
3507 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3510 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3513 #endif /* CONFIG_MEMCG_KMEM */
3515 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3517 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3519 * Because tail pages are not marked as "used", set it. We're under
3520 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3521 * charge/uncharge will be never happen and move_account() is done under
3522 * compound_lock(), so we don't have to take care of races.
3524 void mem_cgroup_split_huge_fixup(struct page *head)
3526 struct page_cgroup *head_pc = lookup_page_cgroup(head);
3527 struct page_cgroup *pc;
3530 if (mem_cgroup_disabled())
3532 for (i = 1; i < HPAGE_PMD_NR; i++) {
3534 pc->mem_cgroup = head_pc->mem_cgroup;
3535 smp_wmb();/* see __commit_charge() */
3536 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
3539 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3542 * mem_cgroup_move_account - move account of the page
3544 * @nr_pages: number of regular pages (>1 for huge pages)
3545 * @pc: page_cgroup of the page.
3546 * @from: mem_cgroup which the page is moved from.
3547 * @to: mem_cgroup which the page is moved to. @from != @to.
3549 * The caller must confirm following.
3550 * - page is not on LRU (isolate_page() is useful.)
3551 * - compound_lock is held when nr_pages > 1
3553 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3556 static int mem_cgroup_move_account(struct page *page,
3557 unsigned int nr_pages,
3558 struct page_cgroup *pc,
3559 struct mem_cgroup *from,
3560 struct mem_cgroup *to)
3562 unsigned long flags;
3564 bool anon = PageAnon(page);
3566 VM_BUG_ON(from == to);
3567 VM_BUG_ON(PageLRU(page));
3569 * The page is isolated from LRU. So, collapse function
3570 * will not handle this page. But page splitting can happen.
3571 * Do this check under compound_page_lock(). The caller should
3575 if (nr_pages > 1 && !PageTransHuge(page))
3578 lock_page_cgroup(pc);
3581 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3584 move_lock_mem_cgroup(from, &flags);
3586 if (!anon && page_mapped(page)) {
3587 /* Update mapped_file data for mem_cgroup */
3589 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
3590 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
3593 mem_cgroup_charge_statistics(from, anon, -nr_pages);
3595 /* caller should have done css_get */
3596 pc->mem_cgroup = to;
3597 mem_cgroup_charge_statistics(to, anon, nr_pages);
3598 move_unlock_mem_cgroup(from, &flags);
3601 unlock_page_cgroup(pc);
3605 memcg_check_events(to, page);
3606 memcg_check_events(from, page);
3612 * mem_cgroup_move_parent - moves page to the parent group
3613 * @page: the page to move
3614 * @pc: page_cgroup of the page
3615 * @child: page's cgroup
3617 * move charges to its parent or the root cgroup if the group has no
3618 * parent (aka use_hierarchy==0).
3619 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3620 * mem_cgroup_move_account fails) the failure is always temporary and
3621 * it signals a race with a page removal/uncharge or migration. In the
3622 * first case the page is on the way out and it will vanish from the LRU
3623 * on the next attempt and the call should be retried later.
3624 * Isolation from the LRU fails only if page has been isolated from
3625 * the LRU since we looked at it and that usually means either global
3626 * reclaim or migration going on. The page will either get back to the
3628 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3629 * (!PageCgroupUsed) or moved to a different group. The page will
3630 * disappear in the next attempt.
3632 static int mem_cgroup_move_parent(struct page *page,
3633 struct page_cgroup *pc,
3634 struct mem_cgroup *child)
3636 struct mem_cgroup *parent;
3637 unsigned int nr_pages;
3638 unsigned long uninitialized_var(flags);
3641 VM_BUG_ON(mem_cgroup_is_root(child));
3644 if (!get_page_unless_zero(page))
3646 if (isolate_lru_page(page))
3649 nr_pages = hpage_nr_pages(page);
3651 parent = parent_mem_cgroup(child);
3653 * If no parent, move charges to root cgroup.
3656 parent = root_mem_cgroup;
3659 VM_BUG_ON(!PageTransHuge(page));
3660 flags = compound_lock_irqsave(page);
3663 ret = mem_cgroup_move_account(page, nr_pages,
3666 __mem_cgroup_cancel_local_charge(child, nr_pages);
3669 compound_unlock_irqrestore(page, flags);
3670 putback_lru_page(page);
3678 * Charge the memory controller for page usage.
3680 * 0 if the charge was successful
3681 * < 0 if the cgroup is over its limit
3683 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
3684 gfp_t gfp_mask, enum charge_type ctype)
3686 struct mem_cgroup *memcg = NULL;
3687 unsigned int nr_pages = 1;
3691 if (PageTransHuge(page)) {
3692 nr_pages <<= compound_order(page);
3693 VM_BUG_ON(!PageTransHuge(page));
3695 * Never OOM-kill a process for a huge page. The
3696 * fault handler will fall back to regular pages.
3701 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
3704 __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false);
3708 int mem_cgroup_newpage_charge(struct page *page,
3709 struct mm_struct *mm, gfp_t gfp_mask)
3711 if (mem_cgroup_disabled())
3713 VM_BUG_ON(page_mapped(page));
3714 VM_BUG_ON(page->mapping && !PageAnon(page));
3716 return mem_cgroup_charge_common(page, mm, gfp_mask,
3717 MEM_CGROUP_CHARGE_TYPE_ANON);
3721 * While swap-in, try_charge -> commit or cancel, the page is locked.
3722 * And when try_charge() successfully returns, one refcnt to memcg without
3723 * struct page_cgroup is acquired. This refcnt will be consumed by
3724 * "commit()" or removed by "cancel()"
3726 static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm,
3729 struct mem_cgroup **memcgp)
3731 struct mem_cgroup *memcg;
3732 struct page_cgroup *pc;
3735 pc = lookup_page_cgroup(page);
3737 * Every swap fault against a single page tries to charge the
3738 * page, bail as early as possible. shmem_unuse() encounters
3739 * already charged pages, too. The USED bit is protected by
3740 * the page lock, which serializes swap cache removal, which
3741 * in turn serializes uncharging.
3743 if (PageCgroupUsed(pc))
3745 if (!do_swap_account)
3747 memcg = try_get_mem_cgroup_from_page(page);
3751 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
3752 css_put(&memcg->css);
3757 ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
3763 int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page,
3764 gfp_t gfp_mask, struct mem_cgroup **memcgp)
3767 if (mem_cgroup_disabled())
3770 * A racing thread's fault, or swapoff, may have already
3771 * updated the pte, and even removed page from swap cache: in
3772 * those cases unuse_pte()'s pte_same() test will fail; but
3773 * there's also a KSM case which does need to charge the page.
3775 if (!PageSwapCache(page)) {
3778 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, memcgp, true);
3783 return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp);
3786 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
3788 if (mem_cgroup_disabled())
3792 __mem_cgroup_cancel_charge(memcg, 1);
3796 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
3797 enum charge_type ctype)
3799 if (mem_cgroup_disabled())
3804 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
3806 * Now swap is on-memory. This means this page may be
3807 * counted both as mem and swap....double count.
3808 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
3809 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
3810 * may call delete_from_swap_cache() before reach here.
3812 if (do_swap_account && PageSwapCache(page)) {
3813 swp_entry_t ent = {.val = page_private(page)};
3814 mem_cgroup_uncharge_swap(ent);
3818 void mem_cgroup_commit_charge_swapin(struct page *page,
3819 struct mem_cgroup *memcg)
3821 __mem_cgroup_commit_charge_swapin(page, memcg,
3822 MEM_CGROUP_CHARGE_TYPE_ANON);
3825 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
3828 struct mem_cgroup *memcg = NULL;
3829 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
3832 if (mem_cgroup_disabled())
3834 if (PageCompound(page))
3837 if (!PageSwapCache(page))
3838 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
3839 else { /* page is swapcache/shmem */
3840 ret = __mem_cgroup_try_charge_swapin(mm, page,
3843 __mem_cgroup_commit_charge_swapin(page, memcg, type);
3848 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
3849 unsigned int nr_pages,
3850 const enum charge_type ctype)
3852 struct memcg_batch_info *batch = NULL;
3853 bool uncharge_memsw = true;
3855 /* If swapout, usage of swap doesn't decrease */
3856 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
3857 uncharge_memsw = false;
3859 batch = ¤t->memcg_batch;
3861 * In usual, we do css_get() when we remember memcg pointer.
3862 * But in this case, we keep res->usage until end of a series of
3863 * uncharges. Then, it's ok to ignore memcg's refcnt.
3866 batch->memcg = memcg;
3868 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
3869 * In those cases, all pages freed continuously can be expected to be in
3870 * the same cgroup and we have chance to coalesce uncharges.
3871 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
3872 * because we want to do uncharge as soon as possible.
3875 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
3876 goto direct_uncharge;
3879 goto direct_uncharge;
3882 * In typical case, batch->memcg == mem. This means we can
3883 * merge a series of uncharges to an uncharge of res_counter.
3884 * If not, we uncharge res_counter ony by one.
3886 if (batch->memcg != memcg)
3887 goto direct_uncharge;
3888 /* remember freed charge and uncharge it later */
3891 batch->memsw_nr_pages++;
3894 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
3896 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
3897 if (unlikely(batch->memcg != memcg))
3898 memcg_oom_recover(memcg);
3902 * uncharge if !page_mapped(page)
3904 static struct mem_cgroup *
3905 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
3908 struct mem_cgroup *memcg = NULL;
3909 unsigned int nr_pages = 1;
3910 struct page_cgroup *pc;
3913 if (mem_cgroup_disabled())
3916 VM_BUG_ON(PageSwapCache(page));
3918 if (PageTransHuge(page)) {
3919 nr_pages <<= compound_order(page);
3920 VM_BUG_ON(!PageTransHuge(page));
3923 * Check if our page_cgroup is valid
3925 pc = lookup_page_cgroup(page);
3926 if (unlikely(!PageCgroupUsed(pc)))
3929 lock_page_cgroup(pc);
3931 memcg = pc->mem_cgroup;
3933 if (!PageCgroupUsed(pc))
3936 anon = PageAnon(page);
3939 case MEM_CGROUP_CHARGE_TYPE_ANON:
3941 * Generally PageAnon tells if it's the anon statistics to be
3942 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
3943 * used before page reached the stage of being marked PageAnon.
3947 case MEM_CGROUP_CHARGE_TYPE_DROP:
3948 /* See mem_cgroup_prepare_migration() */
3949 if (page_mapped(page))
3952 * Pages under migration may not be uncharged. But
3953 * end_migration() /must/ be the one uncharging the
3954 * unused post-migration page and so it has to call
3955 * here with the migration bit still set. See the
3956 * res_counter handling below.
3958 if (!end_migration && PageCgroupMigration(pc))
3961 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
3962 if (!PageAnon(page)) { /* Shared memory */
3963 if (page->mapping && !page_is_file_cache(page))
3965 } else if (page_mapped(page)) /* Anon */
3972 mem_cgroup_charge_statistics(memcg, anon, -nr_pages);
3974 ClearPageCgroupUsed(pc);
3976 * pc->mem_cgroup is not cleared here. It will be accessed when it's
3977 * freed from LRU. This is safe because uncharged page is expected not
3978 * to be reused (freed soon). Exception is SwapCache, it's handled by
3979 * special functions.
3982 unlock_page_cgroup(pc);
3984 * even after unlock, we have memcg->res.usage here and this memcg
3985 * will never be freed.
3987 memcg_check_events(memcg, page);
3988 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
3989 mem_cgroup_swap_statistics(memcg, true);
3990 mem_cgroup_get(memcg);
3993 * Migration does not charge the res_counter for the
3994 * replacement page, so leave it alone when phasing out the
3995 * page that is unused after the migration.
3997 if (!end_migration && !mem_cgroup_is_root(memcg))
3998 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
4003 unlock_page_cgroup(pc);
4007 void mem_cgroup_uncharge_page(struct page *page)
4010 if (page_mapped(page))
4012 VM_BUG_ON(page->mapping && !PageAnon(page));
4013 if (PageSwapCache(page))
4015 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
4018 void mem_cgroup_uncharge_cache_page(struct page *page)
4020 VM_BUG_ON(page_mapped(page));
4021 VM_BUG_ON(page->mapping);
4022 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
4026 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4027 * In that cases, pages are freed continuously and we can expect pages
4028 * are in the same memcg. All these calls itself limits the number of
4029 * pages freed at once, then uncharge_start/end() is called properly.
4030 * This may be called prural(2) times in a context,
4033 void mem_cgroup_uncharge_start(void)
4035 current->memcg_batch.do_batch++;
4036 /* We can do nest. */
4037 if (current->memcg_batch.do_batch == 1) {
4038 current->memcg_batch.memcg = NULL;
4039 current->memcg_batch.nr_pages = 0;
4040 current->memcg_batch.memsw_nr_pages = 0;
4044 void mem_cgroup_uncharge_end(void)
4046 struct memcg_batch_info *batch = ¤t->memcg_batch;
4048 if (!batch->do_batch)
4052 if (batch->do_batch) /* If stacked, do nothing. */
4058 * This "batch->memcg" is valid without any css_get/put etc...
4059 * bacause we hide charges behind us.
4061 if (batch->nr_pages)
4062 res_counter_uncharge(&batch->memcg->res,
4063 batch->nr_pages * PAGE_SIZE);
4064 if (batch->memsw_nr_pages)
4065 res_counter_uncharge(&batch->memcg->memsw,
4066 batch->memsw_nr_pages * PAGE_SIZE);
4067 memcg_oom_recover(batch->memcg);
4068 /* forget this pointer (for sanity check) */
4069 batch->memcg = NULL;
4074 * called after __delete_from_swap_cache() and drop "page" account.
4075 * memcg information is recorded to swap_cgroup of "ent"
4078 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
4080 struct mem_cgroup *memcg;
4081 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
4083 if (!swapout) /* this was a swap cache but the swap is unused ! */
4084 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
4086 memcg = __mem_cgroup_uncharge_common(page, ctype, false);
4089 * record memcg information, if swapout && memcg != NULL,
4090 * mem_cgroup_get() was called in uncharge().
4092 if (do_swap_account && swapout && memcg)
4093 swap_cgroup_record(ent, css_id(&memcg->css));
4097 #ifdef CONFIG_MEMCG_SWAP
4099 * called from swap_entry_free(). remove record in swap_cgroup and
4100 * uncharge "memsw" account.
4102 void mem_cgroup_uncharge_swap(swp_entry_t ent)
4104 struct mem_cgroup *memcg;
4107 if (!do_swap_account)
4110 id = swap_cgroup_record(ent, 0);
4112 memcg = mem_cgroup_lookup(id);
4115 * We uncharge this because swap is freed.
4116 * This memcg can be obsolete one. We avoid calling css_tryget
4118 if (!mem_cgroup_is_root(memcg))
4119 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
4120 mem_cgroup_swap_statistics(memcg, false);
4121 mem_cgroup_put(memcg);
4127 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4128 * @entry: swap entry to be moved
4129 * @from: mem_cgroup which the entry is moved from
4130 * @to: mem_cgroup which the entry is moved to
4132 * It succeeds only when the swap_cgroup's record for this entry is the same
4133 * as the mem_cgroup's id of @from.
4135 * Returns 0 on success, -EINVAL on failure.
4137 * The caller must have charged to @to, IOW, called res_counter_charge() about
4138 * both res and memsw, and called css_get().
4140 static int mem_cgroup_move_swap_account(swp_entry_t entry,
4141 struct mem_cgroup *from, struct mem_cgroup *to)
4143 unsigned short old_id, new_id;
4145 old_id = css_id(&from->css);
4146 new_id = css_id(&to->css);
4148 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
4149 mem_cgroup_swap_statistics(from, false);
4150 mem_cgroup_swap_statistics(to, true);
4152 * This function is only called from task migration context now.
4153 * It postpones res_counter and refcount handling till the end
4154 * of task migration(mem_cgroup_clear_mc()) for performance
4155 * improvement. But we cannot postpone mem_cgroup_get(to)
4156 * because if the process that has been moved to @to does
4157 * swap-in, the refcount of @to might be decreased to 0.
4165 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
4166 struct mem_cgroup *from, struct mem_cgroup *to)
4173 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4176 void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
4177 struct mem_cgroup **memcgp)
4179 struct mem_cgroup *memcg = NULL;
4180 struct page_cgroup *pc;
4181 enum charge_type ctype;
4185 VM_BUG_ON(PageTransHuge(page));
4186 if (mem_cgroup_disabled())
4189 pc = lookup_page_cgroup(page);
4190 lock_page_cgroup(pc);
4191 if (PageCgroupUsed(pc)) {
4192 memcg = pc->mem_cgroup;
4193 css_get(&memcg->css);
4195 * At migrating an anonymous page, its mapcount goes down
4196 * to 0 and uncharge() will be called. But, even if it's fully
4197 * unmapped, migration may fail and this page has to be
4198 * charged again. We set MIGRATION flag here and delay uncharge
4199 * until end_migration() is called
4201 * Corner Case Thinking
4203 * When the old page was mapped as Anon and it's unmap-and-freed
4204 * while migration was ongoing.
4205 * If unmap finds the old page, uncharge() of it will be delayed
4206 * until end_migration(). If unmap finds a new page, it's
4207 * uncharged when it make mapcount to be 1->0. If unmap code
4208 * finds swap_migration_entry, the new page will not be mapped
4209 * and end_migration() will find it(mapcount==0).
4212 * When the old page was mapped but migraion fails, the kernel
4213 * remaps it. A charge for it is kept by MIGRATION flag even
4214 * if mapcount goes down to 0. We can do remap successfully
4215 * without charging it again.
4218 * The "old" page is under lock_page() until the end of
4219 * migration, so, the old page itself will not be swapped-out.
4220 * If the new page is swapped out before end_migraton, our
4221 * hook to usual swap-out path will catch the event.
4224 SetPageCgroupMigration(pc);
4226 unlock_page_cgroup(pc);
4228 * If the page is not charged at this point,
4236 * We charge new page before it's used/mapped. So, even if unlock_page()
4237 * is called before end_migration, we can catch all events on this new
4238 * page. In the case new page is migrated but not remapped, new page's
4239 * mapcount will be finally 0 and we call uncharge in end_migration().
4242 ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
4244 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
4246 * The page is committed to the memcg, but it's not actually
4247 * charged to the res_counter since we plan on replacing the
4248 * old one and only one page is going to be left afterwards.
4250 __mem_cgroup_commit_charge(memcg, newpage, 1, ctype, false);
4253 /* remove redundant charge if migration failed*/
4254 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
4255 struct page *oldpage, struct page *newpage, bool migration_ok)
4257 struct page *used, *unused;
4258 struct page_cgroup *pc;
4264 if (!migration_ok) {
4271 anon = PageAnon(used);
4272 __mem_cgroup_uncharge_common(unused,
4273 anon ? MEM_CGROUP_CHARGE_TYPE_ANON
4274 : MEM_CGROUP_CHARGE_TYPE_CACHE,
4276 css_put(&memcg->css);
4278 * We disallowed uncharge of pages under migration because mapcount
4279 * of the page goes down to zero, temporarly.
4280 * Clear the flag and check the page should be charged.
4282 pc = lookup_page_cgroup(oldpage);
4283 lock_page_cgroup(pc);
4284 ClearPageCgroupMigration(pc);
4285 unlock_page_cgroup(pc);
4288 * If a page is a file cache, radix-tree replacement is very atomic
4289 * and we can skip this check. When it was an Anon page, its mapcount
4290 * goes down to 0. But because we added MIGRATION flage, it's not
4291 * uncharged yet. There are several case but page->mapcount check
4292 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4293 * check. (see prepare_charge() also)
4296 mem_cgroup_uncharge_page(used);
4300 * At replace page cache, newpage is not under any memcg but it's on
4301 * LRU. So, this function doesn't touch res_counter but handles LRU
4302 * in correct way. Both pages are locked so we cannot race with uncharge.
4304 void mem_cgroup_replace_page_cache(struct page *oldpage,
4305 struct page *newpage)
4307 struct mem_cgroup *memcg = NULL;
4308 struct page_cgroup *pc;
4309 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4311 if (mem_cgroup_disabled())
4314 pc = lookup_page_cgroup(oldpage);
4315 /* fix accounting on old pages */
4316 lock_page_cgroup(pc);
4317 if (PageCgroupUsed(pc)) {
4318 memcg = pc->mem_cgroup;
4319 mem_cgroup_charge_statistics(memcg, false, -1);
4320 ClearPageCgroupUsed(pc);
4322 unlock_page_cgroup(pc);
4325 * When called from shmem_replace_page(), in some cases the
4326 * oldpage has already been charged, and in some cases not.
4331 * Even if newpage->mapping was NULL before starting replacement,
4332 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4333 * LRU while we overwrite pc->mem_cgroup.
4335 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
4338 #ifdef CONFIG_DEBUG_VM
4339 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
4341 struct page_cgroup *pc;
4343 pc = lookup_page_cgroup(page);
4345 * Can be NULL while feeding pages into the page allocator for
4346 * the first time, i.e. during boot or memory hotplug;
4347 * or when mem_cgroup_disabled().
4349 if (likely(pc) && PageCgroupUsed(pc))
4354 bool mem_cgroup_bad_page_check(struct page *page)
4356 if (mem_cgroup_disabled())
4359 return lookup_page_cgroup_used(page) != NULL;
4362 void mem_cgroup_print_bad_page(struct page *page)
4364 struct page_cgroup *pc;
4366 pc = lookup_page_cgroup_used(page);
4368 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4369 pc, pc->flags, pc->mem_cgroup);
4374 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
4375 unsigned long long val)
4378 u64 memswlimit, memlimit;
4380 int children = mem_cgroup_count_children(memcg);
4381 u64 curusage, oldusage;
4385 * For keeping hierarchical_reclaim simple, how long we should retry
4386 * is depends on callers. We set our retry-count to be function
4387 * of # of children which we should visit in this loop.
4389 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
4391 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4394 while (retry_count) {
4395 if (signal_pending(current)) {
4400 * Rather than hide all in some function, I do this in
4401 * open coded manner. You see what this really does.
4402 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4404 mutex_lock(&set_limit_mutex);
4405 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4406 if (memswlimit < val) {
4408 mutex_unlock(&set_limit_mutex);
4412 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4416 ret = res_counter_set_limit(&memcg->res, val);
4418 if (memswlimit == val)
4419 memcg->memsw_is_minimum = true;
4421 memcg->memsw_is_minimum = false;
4423 mutex_unlock(&set_limit_mutex);
4428 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4429 MEM_CGROUP_RECLAIM_SHRINK);
4430 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4431 /* Usage is reduced ? */
4432 if (curusage >= oldusage)
4435 oldusage = curusage;
4437 if (!ret && enlarge)
4438 memcg_oom_recover(memcg);
4443 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
4444 unsigned long long val)
4447 u64 memlimit, memswlimit, oldusage, curusage;
4448 int children = mem_cgroup_count_children(memcg);
4452 /* see mem_cgroup_resize_res_limit */
4453 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
4454 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4455 while (retry_count) {
4456 if (signal_pending(current)) {
4461 * Rather than hide all in some function, I do this in
4462 * open coded manner. You see what this really does.
4463 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4465 mutex_lock(&set_limit_mutex);
4466 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4467 if (memlimit > val) {
4469 mutex_unlock(&set_limit_mutex);
4472 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4473 if (memswlimit < val)
4475 ret = res_counter_set_limit(&memcg->memsw, val);
4477 if (memlimit == val)
4478 memcg->memsw_is_minimum = true;
4480 memcg->memsw_is_minimum = false;
4482 mutex_unlock(&set_limit_mutex);
4487 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4488 MEM_CGROUP_RECLAIM_NOSWAP |
4489 MEM_CGROUP_RECLAIM_SHRINK);
4490 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4491 /* Usage is reduced ? */
4492 if (curusage >= oldusage)
4495 oldusage = curusage;
4497 if (!ret && enlarge)
4498 memcg_oom_recover(memcg);
4502 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
4504 unsigned long *total_scanned)
4506 unsigned long nr_reclaimed = 0;
4507 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
4508 unsigned long reclaimed;
4510 struct mem_cgroup_tree_per_zone *mctz;
4511 unsigned long long excess;
4512 unsigned long nr_scanned;
4517 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
4519 * This loop can run a while, specially if mem_cgroup's continuously
4520 * keep exceeding their soft limit and putting the system under
4527 mz = mem_cgroup_largest_soft_limit_node(mctz);
4532 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
4533 gfp_mask, &nr_scanned);
4534 nr_reclaimed += reclaimed;
4535 *total_scanned += nr_scanned;
4536 spin_lock(&mctz->lock);
4539 * If we failed to reclaim anything from this memory cgroup
4540 * it is time to move on to the next cgroup
4546 * Loop until we find yet another one.
4548 * By the time we get the soft_limit lock
4549 * again, someone might have aded the
4550 * group back on the RB tree. Iterate to
4551 * make sure we get a different mem.
4552 * mem_cgroup_largest_soft_limit_node returns
4553 * NULL if no other cgroup is present on
4557 __mem_cgroup_largest_soft_limit_node(mctz);
4559 css_put(&next_mz->memcg->css);
4560 else /* next_mz == NULL or other memcg */
4564 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
4565 excess = res_counter_soft_limit_excess(&mz->memcg->res);
4567 * One school of thought says that we should not add
4568 * back the node to the tree if reclaim returns 0.
4569 * But our reclaim could return 0, simply because due
4570 * to priority we are exposing a smaller subset of
4571 * memory to reclaim from. Consider this as a longer
4574 /* If excess == 0, no tree ops */
4575 __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess);
4576 spin_unlock(&mctz->lock);
4577 css_put(&mz->memcg->css);
4580 * Could not reclaim anything and there are no more
4581 * mem cgroups to try or we seem to be looping without
4582 * reclaiming anything.
4584 if (!nr_reclaimed &&
4586 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
4588 } while (!nr_reclaimed);
4590 css_put(&next_mz->memcg->css);
4591 return nr_reclaimed;
4595 * mem_cgroup_force_empty_list - clears LRU of a group
4596 * @memcg: group to clear
4599 * @lru: lru to to clear
4601 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4602 * reclaim the pages page themselves - pages are moved to the parent (or root)
4605 static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
4606 int node, int zid, enum lru_list lru)
4608 struct mem_cgroup_per_zone *mz;
4609 unsigned long flags;
4610 struct list_head *list;
4614 zone = &NODE_DATA(node)->node_zones[zid];
4615 mz = mem_cgroup_zoneinfo(memcg, node, zid);
4616 list = &mz->lruvec.lists[lru];
4620 struct page_cgroup *pc;
4623 spin_lock_irqsave(&zone->lru_lock, flags);
4624 if (list_empty(list)) {
4625 spin_unlock_irqrestore(&zone->lru_lock, flags);
4628 page = list_entry(list->prev, struct page, lru);
4630 list_move(&page->lru, list);
4632 spin_unlock_irqrestore(&zone->lru_lock, flags);
4635 spin_unlock_irqrestore(&zone->lru_lock, flags);
4637 pc = lookup_page_cgroup(page);
4639 if (mem_cgroup_move_parent(page, pc, memcg)) {
4640 /* found lock contention or "pc" is obsolete. */
4645 } while (!list_empty(list));
4649 * make mem_cgroup's charge to be 0 if there is no task by moving
4650 * all the charges and pages to the parent.
4651 * This enables deleting this mem_cgroup.
4653 * Caller is responsible for holding css reference on the memcg.
4655 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
4661 /* This is for making all *used* pages to be on LRU. */
4662 lru_add_drain_all();
4663 drain_all_stock_sync(memcg);
4664 mem_cgroup_start_move(memcg);
4665 for_each_node_state(node, N_HIGH_MEMORY) {
4666 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4669 mem_cgroup_force_empty_list(memcg,
4674 mem_cgroup_end_move(memcg);
4675 memcg_oom_recover(memcg);
4679 * Kernel memory may not necessarily be trackable to a specific
4680 * process. So they are not migrated, and therefore we can't
4681 * expect their value to drop to 0 here.
4682 * Having res filled up with kmem only is enough.
4684 * This is a safety check because mem_cgroup_force_empty_list
4685 * could have raced with mem_cgroup_replace_page_cache callers
4686 * so the lru seemed empty but the page could have been added
4687 * right after the check. RES_USAGE should be safe as we always
4688 * charge before adding to the LRU.
4690 usage = res_counter_read_u64(&memcg->res, RES_USAGE) -
4691 res_counter_read_u64(&memcg->kmem, RES_USAGE);
4692 } while (usage > 0);
4696 * Reclaims as many pages from the given memcg as possible and moves
4697 * the rest to the parent.
4699 * Caller is responsible for holding css reference for memcg.
4701 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
4703 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
4704 struct cgroup *cgrp = memcg->css.cgroup;
4706 /* returns EBUSY if there is a task or if we come here twice. */
4707 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
4710 /* we call try-to-free pages for make this cgroup empty */
4711 lru_add_drain_all();
4712 /* try to free all pages in this cgroup */
4713 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
4716 if (signal_pending(current))
4719 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
4723 /* maybe some writeback is necessary */
4724 congestion_wait(BLK_RW_ASYNC, HZ/10);
4729 mem_cgroup_reparent_charges(memcg);
4734 static int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
4736 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4739 if (mem_cgroup_is_root(memcg))
4741 css_get(&memcg->css);
4742 ret = mem_cgroup_force_empty(memcg);
4743 css_put(&memcg->css);
4749 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
4751 return mem_cgroup_from_cont(cont)->use_hierarchy;
4754 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
4758 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4759 struct cgroup *parent = cont->parent;
4760 struct mem_cgroup *parent_memcg = NULL;
4763 parent_memcg = mem_cgroup_from_cont(parent);
4767 if (memcg->use_hierarchy == val)
4771 * If parent's use_hierarchy is set, we can't make any modifications
4772 * in the child subtrees. If it is unset, then the change can
4773 * occur, provided the current cgroup has no children.
4775 * For the root cgroup, parent_mem is NULL, we allow value to be
4776 * set if there are no children.
4778 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
4779 (val == 1 || val == 0)) {
4780 if (list_empty(&cont->children))
4781 memcg->use_hierarchy = val;
4794 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
4795 enum mem_cgroup_stat_index idx)
4797 struct mem_cgroup *iter;
4800 /* Per-cpu values can be negative, use a signed accumulator */
4801 for_each_mem_cgroup_tree(iter, memcg)
4802 val += mem_cgroup_read_stat(iter, idx);
4804 if (val < 0) /* race ? */
4809 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
4813 if (!mem_cgroup_is_root(memcg)) {
4815 return res_counter_read_u64(&memcg->res, RES_USAGE);
4817 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
4820 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
4821 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
4824 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
4826 return val << PAGE_SHIFT;
4829 static ssize_t mem_cgroup_read(struct cgroup *cont, struct cftype *cft,
4830 struct file *file, char __user *buf,
4831 size_t nbytes, loff_t *ppos)
4833 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4839 type = MEMFILE_TYPE(cft->private);
4840 name = MEMFILE_ATTR(cft->private);
4842 if (!do_swap_account && type == _MEMSWAP)
4847 if (name == RES_USAGE)
4848 val = mem_cgroup_usage(memcg, false);
4850 val = res_counter_read_u64(&memcg->res, name);
4853 if (name == RES_USAGE)
4854 val = mem_cgroup_usage(memcg, true);
4856 val = res_counter_read_u64(&memcg->memsw, name);
4859 val = res_counter_read_u64(&memcg->kmem, name);
4865 len = scnprintf(str, sizeof(str), "%llu\n", (unsigned long long)val);
4866 return simple_read_from_buffer(buf, nbytes, ppos, str, len);
4869 static int memcg_update_kmem_limit(struct cgroup *cont, u64 val)
4872 #ifdef CONFIG_MEMCG_KMEM
4873 bool must_inc_static_branch = false;
4875 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4877 * For simplicity, we won't allow this to be disabled. It also can't
4878 * be changed if the cgroup has children already, or if tasks had
4881 * If tasks join before we set the limit, a person looking at
4882 * kmem.usage_in_bytes will have no way to determine when it took
4883 * place, which makes the value quite meaningless.
4885 * After it first became limited, changes in the value of the limit are
4886 * of course permitted.
4888 * Taking the cgroup_lock is really offensive, but it is so far the only
4889 * way to guarantee that no children will appear. There are plenty of
4890 * other offenders, and they should all go away. Fine grained locking
4891 * is probably the way to go here. When we are fully hierarchical, we
4892 * can also get rid of the use_hierarchy check.
4895 mutex_lock(&set_limit_mutex);
4896 if (!memcg->kmem_account_flags && val != RESOURCE_MAX) {
4897 if (cgroup_task_count(cont) || (memcg->use_hierarchy &&
4898 !list_empty(&cont->children))) {
4902 ret = res_counter_set_limit(&memcg->kmem, val);
4905 ret = memcg_update_cache_sizes(memcg);
4907 res_counter_set_limit(&memcg->kmem, RESOURCE_MAX);
4910 must_inc_static_branch = true;
4912 * kmem charges can outlive the cgroup. In the case of slab
4913 * pages, for instance, a page contain objects from various
4914 * processes, so it is unfeasible to migrate them away. We
4915 * need to reference count the memcg because of that.
4917 mem_cgroup_get(memcg);
4919 ret = res_counter_set_limit(&memcg->kmem, val);
4921 mutex_unlock(&set_limit_mutex);
4925 * We are by now familiar with the fact that we can't inc the static
4926 * branch inside cgroup_lock. See disarm functions for details. A
4927 * worker here is overkill, but also wrong: After the limit is set, we
4928 * must start accounting right away. Since this operation can't fail,
4929 * we can safely defer it to here - no rollback will be needed.
4931 * The boolean used to control this is also safe, because
4932 * KMEM_ACCOUNTED_ACTIVATED guarantees that only one process will be
4933 * able to set it to true;
4935 if (must_inc_static_branch) {
4936 static_key_slow_inc(&memcg_kmem_enabled_key);
4938 * setting the active bit after the inc will guarantee no one
4939 * starts accounting before all call sites are patched
4941 memcg_kmem_set_active(memcg);
4948 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
4951 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4955 memcg->kmem_account_flags = parent->kmem_account_flags;
4956 #ifdef CONFIG_MEMCG_KMEM
4958 * When that happen, we need to disable the static branch only on those
4959 * memcgs that enabled it. To achieve this, we would be forced to
4960 * complicate the code by keeping track of which memcgs were the ones
4961 * that actually enabled limits, and which ones got it from its
4964 * It is a lot simpler just to do static_key_slow_inc() on every child
4965 * that is accounted.
4967 if (!memcg_kmem_is_active(memcg))
4971 * destroy(), called if we fail, will issue static_key_slow_inc() and
4972 * mem_cgroup_put() if kmem is enabled. We have to either call them
4973 * unconditionally, or clear the KMEM_ACTIVE flag. I personally find
4974 * this more consistent, since it always leads to the same destroy path
4976 mem_cgroup_get(memcg);
4977 static_key_slow_inc(&memcg_kmem_enabled_key);
4979 mutex_lock(&set_limit_mutex);
4980 ret = memcg_update_cache_sizes(memcg);
4981 mutex_unlock(&set_limit_mutex);
4988 * The user of this function is...
4991 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
4994 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4997 unsigned long long val;
5000 type = MEMFILE_TYPE(cft->private);
5001 name = MEMFILE_ATTR(cft->private);
5003 if (!do_swap_account && type == _MEMSWAP)
5008 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
5012 /* This function does all necessary parse...reuse it */
5013 ret = res_counter_memparse_write_strategy(buffer, &val);
5017 ret = mem_cgroup_resize_limit(memcg, val);
5018 else if (type == _MEMSWAP)
5019 ret = mem_cgroup_resize_memsw_limit(memcg, val);
5020 else if (type == _KMEM)
5021 ret = memcg_update_kmem_limit(cont, val);
5025 case RES_SOFT_LIMIT:
5026 ret = res_counter_memparse_write_strategy(buffer, &val);
5030 * For memsw, soft limits are hard to implement in terms
5031 * of semantics, for now, we support soft limits for
5032 * control without swap
5035 ret = res_counter_set_soft_limit(&memcg->res, val);
5040 ret = -EINVAL; /* should be BUG() ? */
5046 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
5047 unsigned long long *mem_limit, unsigned long long *memsw_limit)
5049 struct cgroup *cgroup;
5050 unsigned long long min_limit, min_memsw_limit, tmp;
5052 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
5053 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5054 cgroup = memcg->css.cgroup;
5055 if (!memcg->use_hierarchy)
5058 while (cgroup->parent) {
5059 cgroup = cgroup->parent;
5060 memcg = mem_cgroup_from_cont(cgroup);
5061 if (!memcg->use_hierarchy)
5063 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
5064 min_limit = min(min_limit, tmp);
5065 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5066 min_memsw_limit = min(min_memsw_limit, tmp);
5069 *mem_limit = min_limit;
5070 *memsw_limit = min_memsw_limit;
5073 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
5075 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5079 type = MEMFILE_TYPE(event);
5080 name = MEMFILE_ATTR(event);
5082 if (!do_swap_account && type == _MEMSWAP)
5088 res_counter_reset_max(&memcg->res);
5089 else if (type == _MEMSWAP)
5090 res_counter_reset_max(&memcg->memsw);
5091 else if (type == _KMEM)
5092 res_counter_reset_max(&memcg->kmem);
5098 res_counter_reset_failcnt(&memcg->res);
5099 else if (type == _MEMSWAP)
5100 res_counter_reset_failcnt(&memcg->memsw);
5101 else if (type == _KMEM)
5102 res_counter_reset_failcnt(&memcg->kmem);
5111 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
5114 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
5118 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
5119 struct cftype *cft, u64 val)
5121 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5123 if (val >= (1 << NR_MOVE_TYPE))
5126 * We check this value several times in both in can_attach() and
5127 * attach(), so we need cgroup lock to prevent this value from being
5131 memcg->move_charge_at_immigrate = val;
5137 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
5138 struct cftype *cft, u64 val)
5145 static int memcg_numa_stat_show(struct cgroup *cont, struct cftype *cft,
5149 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
5150 unsigned long node_nr;
5151 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5153 total_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL);
5154 seq_printf(m, "total=%lu", total_nr);
5155 for_each_node_state(nid, N_HIGH_MEMORY) {
5156 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL);
5157 seq_printf(m, " N%d=%lu", nid, node_nr);
5161 file_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_FILE);
5162 seq_printf(m, "file=%lu", file_nr);
5163 for_each_node_state(nid, N_HIGH_MEMORY) {
5164 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5166 seq_printf(m, " N%d=%lu", nid, node_nr);
5170 anon_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_ANON);
5171 seq_printf(m, "anon=%lu", anon_nr);
5172 for_each_node_state(nid, N_HIGH_MEMORY) {
5173 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5175 seq_printf(m, " N%d=%lu", nid, node_nr);
5179 unevictable_nr = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
5180 seq_printf(m, "unevictable=%lu", unevictable_nr);
5181 for_each_node_state(nid, N_HIGH_MEMORY) {
5182 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5183 BIT(LRU_UNEVICTABLE));
5184 seq_printf(m, " N%d=%lu", nid, node_nr);
5189 #endif /* CONFIG_NUMA */
5191 static const char * const mem_cgroup_lru_names[] = {
5199 static inline void mem_cgroup_lru_names_not_uptodate(void)
5201 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
5204 static int memcg_stat_show(struct cgroup *cont, struct cftype *cft,
5207 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5208 struct mem_cgroup *mi;
5211 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5212 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5214 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
5215 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
5218 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
5219 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
5220 mem_cgroup_read_events(memcg, i));
5222 for (i = 0; i < NR_LRU_LISTS; i++)
5223 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
5224 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
5226 /* Hierarchical information */
5228 unsigned long long limit, memsw_limit;
5229 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
5230 seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
5231 if (do_swap_account)
5232 seq_printf(m, "hierarchical_memsw_limit %llu\n",
5236 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5239 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5241 for_each_mem_cgroup_tree(mi, memcg)
5242 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
5243 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
5246 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
5247 unsigned long long val = 0;
5249 for_each_mem_cgroup_tree(mi, memcg)
5250 val += mem_cgroup_read_events(mi, i);
5251 seq_printf(m, "total_%s %llu\n",
5252 mem_cgroup_events_names[i], val);
5255 for (i = 0; i < NR_LRU_LISTS; i++) {
5256 unsigned long long val = 0;
5258 for_each_mem_cgroup_tree(mi, memcg)
5259 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
5260 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
5263 #ifdef CONFIG_DEBUG_VM
5266 struct mem_cgroup_per_zone *mz;
5267 struct zone_reclaim_stat *rstat;
5268 unsigned long recent_rotated[2] = {0, 0};
5269 unsigned long recent_scanned[2] = {0, 0};
5271 for_each_online_node(nid)
5272 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
5273 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
5274 rstat = &mz->lruvec.reclaim_stat;
5276 recent_rotated[0] += rstat->recent_rotated[0];
5277 recent_rotated[1] += rstat->recent_rotated[1];
5278 recent_scanned[0] += rstat->recent_scanned[0];
5279 recent_scanned[1] += rstat->recent_scanned[1];
5281 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
5282 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
5283 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
5284 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
5291 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
5293 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5295 return mem_cgroup_swappiness(memcg);
5298 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
5301 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5302 struct mem_cgroup *parent;
5307 if (cgrp->parent == NULL)
5310 parent = mem_cgroup_from_cont(cgrp->parent);
5314 /* If under hierarchy, only empty-root can set this value */
5315 if ((parent->use_hierarchy) ||
5316 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
5321 memcg->swappiness = val;
5328 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
5330 struct mem_cgroup_threshold_ary *t;
5336 t = rcu_dereference(memcg->thresholds.primary);
5338 t = rcu_dereference(memcg->memsw_thresholds.primary);
5343 usage = mem_cgroup_usage(memcg, swap);
5346 * current_threshold points to threshold just below or equal to usage.
5347 * If it's not true, a threshold was crossed after last
5348 * call of __mem_cgroup_threshold().
5350 i = t->current_threshold;
5353 * Iterate backward over array of thresholds starting from
5354 * current_threshold and check if a threshold is crossed.
5355 * If none of thresholds below usage is crossed, we read
5356 * only one element of the array here.
5358 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
5359 eventfd_signal(t->entries[i].eventfd, 1);
5361 /* i = current_threshold + 1 */
5365 * Iterate forward over array of thresholds starting from
5366 * current_threshold+1 and check if a threshold is crossed.
5367 * If none of thresholds above usage is crossed, we read
5368 * only one element of the array here.
5370 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
5371 eventfd_signal(t->entries[i].eventfd, 1);
5373 /* Update current_threshold */
5374 t->current_threshold = i - 1;
5379 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
5382 __mem_cgroup_threshold(memcg, false);
5383 if (do_swap_account)
5384 __mem_cgroup_threshold(memcg, true);
5386 memcg = parent_mem_cgroup(memcg);
5390 static int compare_thresholds(const void *a, const void *b)
5392 const struct mem_cgroup_threshold *_a = a;
5393 const struct mem_cgroup_threshold *_b = b;
5395 return _a->threshold - _b->threshold;
5398 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
5400 struct mem_cgroup_eventfd_list *ev;
5402 list_for_each_entry(ev, &memcg->oom_notify, list)
5403 eventfd_signal(ev->eventfd, 1);
5407 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
5409 struct mem_cgroup *iter;
5411 for_each_mem_cgroup_tree(iter, memcg)
5412 mem_cgroup_oom_notify_cb(iter);
5415 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
5416 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5418 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5419 struct mem_cgroup_thresholds *thresholds;
5420 struct mem_cgroup_threshold_ary *new;
5421 enum res_type type = MEMFILE_TYPE(cft->private);
5422 u64 threshold, usage;
5425 ret = res_counter_memparse_write_strategy(args, &threshold);
5429 mutex_lock(&memcg->thresholds_lock);
5432 thresholds = &memcg->thresholds;
5433 else if (type == _MEMSWAP)
5434 thresholds = &memcg->memsw_thresholds;
5438 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5440 /* Check if a threshold crossed before adding a new one */
5441 if (thresholds->primary)
5442 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5444 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
5446 /* Allocate memory for new array of thresholds */
5447 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
5455 /* Copy thresholds (if any) to new array */
5456 if (thresholds->primary) {
5457 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
5458 sizeof(struct mem_cgroup_threshold));
5461 /* Add new threshold */
5462 new->entries[size - 1].eventfd = eventfd;
5463 new->entries[size - 1].threshold = threshold;
5465 /* Sort thresholds. Registering of new threshold isn't time-critical */
5466 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
5467 compare_thresholds, NULL);
5469 /* Find current threshold */
5470 new->current_threshold = -1;
5471 for (i = 0; i < size; i++) {
5472 if (new->entries[i].threshold <= usage) {
5474 * new->current_threshold will not be used until
5475 * rcu_assign_pointer(), so it's safe to increment
5478 ++new->current_threshold;
5483 /* Free old spare buffer and save old primary buffer as spare */
5484 kfree(thresholds->spare);
5485 thresholds->spare = thresholds->primary;
5487 rcu_assign_pointer(thresholds->primary, new);
5489 /* To be sure that nobody uses thresholds */
5493 mutex_unlock(&memcg->thresholds_lock);
5498 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
5499 struct cftype *cft, struct eventfd_ctx *eventfd)
5501 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5502 struct mem_cgroup_thresholds *thresholds;
5503 struct mem_cgroup_threshold_ary *new;
5504 enum res_type type = MEMFILE_TYPE(cft->private);
5508 mutex_lock(&memcg->thresholds_lock);
5510 thresholds = &memcg->thresholds;
5511 else if (type == _MEMSWAP)
5512 thresholds = &memcg->memsw_thresholds;
5516 if (!thresholds->primary)
5519 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5521 /* Check if a threshold crossed before removing */
5522 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5524 /* Calculate new number of threshold */
5526 for (i = 0; i < thresholds->primary->size; i++) {
5527 if (thresholds->primary->entries[i].eventfd != eventfd)
5531 new = thresholds->spare;
5533 /* Set thresholds array to NULL if we don't have thresholds */
5542 /* Copy thresholds and find current threshold */
5543 new->current_threshold = -1;
5544 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
5545 if (thresholds->primary->entries[i].eventfd == eventfd)
5548 new->entries[j] = thresholds->primary->entries[i];
5549 if (new->entries[j].threshold <= usage) {
5551 * new->current_threshold will not be used
5552 * until rcu_assign_pointer(), so it's safe to increment
5555 ++new->current_threshold;
5561 /* Swap primary and spare array */
5562 thresholds->spare = thresholds->primary;
5563 /* If all events are unregistered, free the spare array */
5565 kfree(thresholds->spare);
5566 thresholds->spare = NULL;
5569 rcu_assign_pointer(thresholds->primary, new);
5571 /* To be sure that nobody uses thresholds */
5574 mutex_unlock(&memcg->thresholds_lock);
5577 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
5578 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5580 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5581 struct mem_cgroup_eventfd_list *event;
5582 enum res_type type = MEMFILE_TYPE(cft->private);
5584 BUG_ON(type != _OOM_TYPE);
5585 event = kmalloc(sizeof(*event), GFP_KERNEL);
5589 spin_lock(&memcg_oom_lock);
5591 event->eventfd = eventfd;
5592 list_add(&event->list, &memcg->oom_notify);
5594 /* already in OOM ? */
5595 if (atomic_read(&memcg->under_oom))
5596 eventfd_signal(eventfd, 1);
5597 spin_unlock(&memcg_oom_lock);
5602 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
5603 struct cftype *cft, struct eventfd_ctx *eventfd)
5605 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5606 struct mem_cgroup_eventfd_list *ev, *tmp;
5607 enum res_type type = MEMFILE_TYPE(cft->private);
5609 BUG_ON(type != _OOM_TYPE);
5611 spin_lock(&memcg_oom_lock);
5613 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
5614 if (ev->eventfd == eventfd) {
5615 list_del(&ev->list);
5620 spin_unlock(&memcg_oom_lock);
5623 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
5624 struct cftype *cft, struct cgroup_map_cb *cb)
5626 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5628 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
5630 if (atomic_read(&memcg->under_oom))
5631 cb->fill(cb, "under_oom", 1);
5633 cb->fill(cb, "under_oom", 0);
5637 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
5638 struct cftype *cft, u64 val)
5640 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5641 struct mem_cgroup *parent;
5643 /* cannot set to root cgroup and only 0 and 1 are allowed */
5644 if (!cgrp->parent || !((val == 0) || (val == 1)))
5647 parent = mem_cgroup_from_cont(cgrp->parent);
5650 /* oom-kill-disable is a flag for subhierarchy. */
5651 if ((parent->use_hierarchy) ||
5652 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
5656 memcg->oom_kill_disable = val;
5658 memcg_oom_recover(memcg);
5663 #ifdef CONFIG_MEMCG_KMEM
5664 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5668 memcg->kmemcg_id = -1;
5669 ret = memcg_propagate_kmem(memcg);
5673 if (mem_cgroup_is_root(memcg))
5674 ida_init(&kmem_limited_groups);
5676 return mem_cgroup_sockets_init(memcg, ss);
5679 static void kmem_cgroup_destroy(struct mem_cgroup *memcg)
5681 mem_cgroup_sockets_destroy(memcg);
5683 memcg_kmem_mark_dead(memcg);
5685 if (res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0)
5689 * Charges already down to 0, undo mem_cgroup_get() done in the charge
5690 * path here, being careful not to race with memcg_uncharge_kmem: it is
5691 * possible that the charges went down to 0 between mark_dead and the
5692 * res_counter read, so in that case, we don't need the put
5694 if (memcg_kmem_test_and_clear_dead(memcg))
5695 mem_cgroup_put(memcg);
5698 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5703 static void kmem_cgroup_destroy(struct mem_cgroup *memcg)
5708 static struct cftype mem_cgroup_files[] = {
5710 .name = "usage_in_bytes",
5711 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
5712 .read = mem_cgroup_read,
5713 .register_event = mem_cgroup_usage_register_event,
5714 .unregister_event = mem_cgroup_usage_unregister_event,
5717 .name = "max_usage_in_bytes",
5718 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
5719 .trigger = mem_cgroup_reset,
5720 .read = mem_cgroup_read,
5723 .name = "limit_in_bytes",
5724 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
5725 .write_string = mem_cgroup_write,
5726 .read = mem_cgroup_read,
5729 .name = "soft_limit_in_bytes",
5730 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
5731 .write_string = mem_cgroup_write,
5732 .read = mem_cgroup_read,
5736 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
5737 .trigger = mem_cgroup_reset,
5738 .read = mem_cgroup_read,
5742 .read_seq_string = memcg_stat_show,
5745 .name = "force_empty",
5746 .trigger = mem_cgroup_force_empty_write,
5749 .name = "use_hierarchy",
5750 .write_u64 = mem_cgroup_hierarchy_write,
5751 .read_u64 = mem_cgroup_hierarchy_read,
5754 .name = "swappiness",
5755 .read_u64 = mem_cgroup_swappiness_read,
5756 .write_u64 = mem_cgroup_swappiness_write,
5759 .name = "move_charge_at_immigrate",
5760 .read_u64 = mem_cgroup_move_charge_read,
5761 .write_u64 = mem_cgroup_move_charge_write,
5764 .name = "oom_control",
5765 .read_map = mem_cgroup_oom_control_read,
5766 .write_u64 = mem_cgroup_oom_control_write,
5767 .register_event = mem_cgroup_oom_register_event,
5768 .unregister_event = mem_cgroup_oom_unregister_event,
5769 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
5773 .name = "numa_stat",
5774 .read_seq_string = memcg_numa_stat_show,
5777 #ifdef CONFIG_MEMCG_SWAP
5779 .name = "memsw.usage_in_bytes",
5780 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
5781 .read = mem_cgroup_read,
5782 .register_event = mem_cgroup_usage_register_event,
5783 .unregister_event = mem_cgroup_usage_unregister_event,
5786 .name = "memsw.max_usage_in_bytes",
5787 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
5788 .trigger = mem_cgroup_reset,
5789 .read = mem_cgroup_read,
5792 .name = "memsw.limit_in_bytes",
5793 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
5794 .write_string = mem_cgroup_write,
5795 .read = mem_cgroup_read,
5798 .name = "memsw.failcnt",
5799 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
5800 .trigger = mem_cgroup_reset,
5801 .read = mem_cgroup_read,
5804 #ifdef CONFIG_MEMCG_KMEM
5806 .name = "kmem.limit_in_bytes",
5807 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
5808 .write_string = mem_cgroup_write,
5809 .read = mem_cgroup_read,
5812 .name = "kmem.usage_in_bytes",
5813 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5814 .read = mem_cgroup_read,
5817 .name = "kmem.failcnt",
5818 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5819 .trigger = mem_cgroup_reset,
5820 .read = mem_cgroup_read,
5823 .name = "kmem.max_usage_in_bytes",
5824 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5825 .trigger = mem_cgroup_reset,
5826 .read = mem_cgroup_read,
5828 #ifdef CONFIG_SLABINFO
5830 .name = "kmem.slabinfo",
5831 .read_seq_string = mem_cgroup_slabinfo_read,
5835 { }, /* terminate */
5838 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
5840 struct mem_cgroup_per_node *pn;
5841 struct mem_cgroup_per_zone *mz;
5842 int zone, tmp = node;
5844 * This routine is called against possible nodes.
5845 * But it's BUG to call kmalloc() against offline node.
5847 * TODO: this routine can waste much memory for nodes which will
5848 * never be onlined. It's better to use memory hotplug callback
5851 if (!node_state(node, N_NORMAL_MEMORY))
5853 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5857 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5858 mz = &pn->zoneinfo[zone];
5859 lruvec_init(&mz->lruvec, &NODE_DATA(node)->node_zones[zone]);
5860 mz->usage_in_excess = 0;
5861 mz->on_tree = false;
5864 memcg->info.nodeinfo[node] = pn;
5868 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
5870 kfree(memcg->info.nodeinfo[node]);
5873 static struct mem_cgroup *mem_cgroup_alloc(void)
5875 struct mem_cgroup *memcg;
5876 int size = sizeof(struct mem_cgroup);
5878 /* Can be very big if MAX_NUMNODES is very big */
5879 if (size < PAGE_SIZE)
5880 memcg = kzalloc(size, GFP_KERNEL);
5882 memcg = vzalloc(size);
5887 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
5890 spin_lock_init(&memcg->pcp_counter_lock);
5894 if (size < PAGE_SIZE)
5902 * At destroying mem_cgroup, references from swap_cgroup can remain.
5903 * (scanning all at force_empty is too costly...)
5905 * Instead of clearing all references at force_empty, we remember
5906 * the number of reference from swap_cgroup and free mem_cgroup when
5907 * it goes down to 0.
5909 * Removal of cgroup itself succeeds regardless of refs from swap.
5912 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5915 int size = sizeof(struct mem_cgroup);
5917 mem_cgroup_remove_from_trees(memcg);
5918 free_css_id(&mem_cgroup_subsys, &memcg->css);
5921 free_mem_cgroup_per_zone_info(memcg, node);
5923 free_percpu(memcg->stat);
5926 * We need to make sure that (at least for now), the jump label
5927 * destruction code runs outside of the cgroup lock. This is because
5928 * get_online_cpus(), which is called from the static_branch update,
5929 * can't be called inside the cgroup_lock. cpusets are the ones
5930 * enforcing this dependency, so if they ever change, we might as well.
5932 * schedule_work() will guarantee this happens. Be careful if you need
5933 * to move this code around, and make sure it is outside
5936 disarm_static_keys(memcg);
5937 if (size < PAGE_SIZE)
5945 * Helpers for freeing a kmalloc()ed/vzalloc()ed mem_cgroup by RCU,
5946 * but in process context. The work_freeing structure is overlaid
5947 * on the rcu_freeing structure, which itself is overlaid on memsw.
5949 static void free_work(struct work_struct *work)
5951 struct mem_cgroup *memcg;
5953 memcg = container_of(work, struct mem_cgroup, work_freeing);
5954 __mem_cgroup_free(memcg);
5957 static void free_rcu(struct rcu_head *rcu_head)
5959 struct mem_cgroup *memcg;
5961 memcg = container_of(rcu_head, struct mem_cgroup, rcu_freeing);
5962 INIT_WORK(&memcg->work_freeing, free_work);
5963 schedule_work(&memcg->work_freeing);
5966 static void mem_cgroup_get(struct mem_cgroup *memcg)
5968 atomic_inc(&memcg->refcnt);
5971 static void __mem_cgroup_put(struct mem_cgroup *memcg, int count)
5973 if (atomic_sub_and_test(count, &memcg->refcnt)) {
5974 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5975 call_rcu(&memcg->rcu_freeing, free_rcu);
5977 mem_cgroup_put(parent);
5981 static void mem_cgroup_put(struct mem_cgroup *memcg)
5983 __mem_cgroup_put(memcg, 1);
5987 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
5989 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
5991 if (!memcg->res.parent)
5993 return mem_cgroup_from_res_counter(memcg->res.parent, res);
5995 EXPORT_SYMBOL(parent_mem_cgroup);
5997 #ifdef CONFIG_MEMCG_SWAP
5998 static void __init enable_swap_cgroup(void)
6000 if (!mem_cgroup_disabled() && really_do_swap_account)
6001 do_swap_account = 1;
6004 static void __init enable_swap_cgroup(void)
6009 static int mem_cgroup_soft_limit_tree_init(void)
6011 struct mem_cgroup_tree_per_node *rtpn;
6012 struct mem_cgroup_tree_per_zone *rtpz;
6013 int tmp, node, zone;
6015 for_each_node(node) {
6017 if (!node_state(node, N_NORMAL_MEMORY))
6019 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
6023 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6025 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6026 rtpz = &rtpn->rb_tree_per_zone[zone];
6027 rtpz->rb_root = RB_ROOT;
6028 spin_lock_init(&rtpz->lock);
6034 for_each_node(node) {
6035 if (!soft_limit_tree.rb_tree_per_node[node])
6037 kfree(soft_limit_tree.rb_tree_per_node[node]);
6038 soft_limit_tree.rb_tree_per_node[node] = NULL;
6044 static struct cgroup_subsys_state * __ref
6045 mem_cgroup_create(struct cgroup *cont)
6047 struct mem_cgroup *memcg, *parent;
6048 long error = -ENOMEM;
6051 memcg = mem_cgroup_alloc();
6053 return ERR_PTR(error);
6056 if (alloc_mem_cgroup_per_zone_info(memcg, node))
6060 if (cont->parent == NULL) {
6062 enable_swap_cgroup();
6064 if (mem_cgroup_soft_limit_tree_init())
6066 root_mem_cgroup = memcg;
6067 for_each_possible_cpu(cpu) {
6068 struct memcg_stock_pcp *stock =
6069 &per_cpu(memcg_stock, cpu);
6070 INIT_WORK(&stock->work, drain_local_stock);
6072 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
6074 parent = mem_cgroup_from_cont(cont->parent);
6075 memcg->use_hierarchy = parent->use_hierarchy;
6076 memcg->oom_kill_disable = parent->oom_kill_disable;
6079 if (parent && parent->use_hierarchy) {
6080 res_counter_init(&memcg->res, &parent->res);
6081 res_counter_init(&memcg->memsw, &parent->memsw);
6082 res_counter_init(&memcg->kmem, &parent->kmem);
6085 * We increment refcnt of the parent to ensure that we can
6086 * safely access it on res_counter_charge/uncharge.
6087 * This refcnt will be decremented when freeing this
6088 * mem_cgroup(see mem_cgroup_put).
6090 mem_cgroup_get(parent);
6092 res_counter_init(&memcg->res, NULL);
6093 res_counter_init(&memcg->memsw, NULL);
6094 res_counter_init(&memcg->kmem, NULL);
6096 * Deeper hierachy with use_hierarchy == false doesn't make
6097 * much sense so let cgroup subsystem know about this
6098 * unfortunate state in our controller.
6100 if (parent && parent != root_mem_cgroup)
6101 mem_cgroup_subsys.broken_hierarchy = true;
6103 memcg->last_scanned_node = MAX_NUMNODES;
6104 INIT_LIST_HEAD(&memcg->oom_notify);
6107 memcg->swappiness = mem_cgroup_swappiness(parent);
6108 atomic_set(&memcg->refcnt, 1);
6109 memcg->move_charge_at_immigrate = 0;
6110 mutex_init(&memcg->thresholds_lock);
6111 spin_lock_init(&memcg->move_lock);
6113 error = memcg_init_kmem(memcg, &mem_cgroup_subsys);
6116 * We call put now because our (and parent's) refcnts
6117 * are already in place. mem_cgroup_put() will internally
6118 * call __mem_cgroup_free, so return directly
6120 mem_cgroup_put(memcg);
6121 return ERR_PTR(error);
6125 __mem_cgroup_free(memcg);
6126 return ERR_PTR(error);
6129 static void mem_cgroup_pre_destroy(struct cgroup *cont)
6131 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
6133 mem_cgroup_reparent_charges(memcg);
6134 mem_cgroup_destroy_all_caches(memcg);
6137 static void mem_cgroup_destroy(struct cgroup *cont)
6139 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
6141 kmem_cgroup_destroy(memcg);
6143 mem_cgroup_put(memcg);
6147 /* Handlers for move charge at task migration. */
6148 #define PRECHARGE_COUNT_AT_ONCE 256
6149 static int mem_cgroup_do_precharge(unsigned long count)
6152 int batch_count = PRECHARGE_COUNT_AT_ONCE;
6153 struct mem_cgroup *memcg = mc.to;
6155 if (mem_cgroup_is_root(memcg)) {
6156 mc.precharge += count;
6157 /* we don't need css_get for root */
6160 /* try to charge at once */
6162 struct res_counter *dummy;
6164 * "memcg" cannot be under rmdir() because we've already checked
6165 * by cgroup_lock_live_cgroup() that it is not removed and we
6166 * are still under the same cgroup_mutex. So we can postpone
6169 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
6171 if (do_swap_account && res_counter_charge(&memcg->memsw,
6172 PAGE_SIZE * count, &dummy)) {
6173 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
6176 mc.precharge += count;
6180 /* fall back to one by one charge */
6182 if (signal_pending(current)) {
6186 if (!batch_count--) {
6187 batch_count = PRECHARGE_COUNT_AT_ONCE;
6190 ret = __mem_cgroup_try_charge(NULL,
6191 GFP_KERNEL, 1, &memcg, false);
6193 /* mem_cgroup_clear_mc() will do uncharge later */
6201 * get_mctgt_type - get target type of moving charge
6202 * @vma: the vma the pte to be checked belongs
6203 * @addr: the address corresponding to the pte to be checked
6204 * @ptent: the pte to be checked
6205 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6208 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6209 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6210 * move charge. if @target is not NULL, the page is stored in target->page
6211 * with extra refcnt got(Callers should handle it).
6212 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6213 * target for charge migration. if @target is not NULL, the entry is stored
6216 * Called with pte lock held.
6223 enum mc_target_type {
6229 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
6230 unsigned long addr, pte_t ptent)
6232 struct page *page = vm_normal_page(vma, addr, ptent);
6234 if (!page || !page_mapped(page))
6236 if (PageAnon(page)) {
6237 /* we don't move shared anon */
6240 } else if (!move_file())
6241 /* we ignore mapcount for file pages */
6243 if (!get_page_unless_zero(page))
6250 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6251 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6253 struct page *page = NULL;
6254 swp_entry_t ent = pte_to_swp_entry(ptent);
6256 if (!move_anon() || non_swap_entry(ent))
6259 * Because lookup_swap_cache() updates some statistics counter,
6260 * we call find_get_page() with swapper_space directly.
6262 page = find_get_page(&swapper_space, ent.val);
6263 if (do_swap_account)
6264 entry->val = ent.val;
6269 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6270 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6276 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
6277 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6279 struct page *page = NULL;
6280 struct address_space *mapping;
6283 if (!vma->vm_file) /* anonymous vma */
6288 mapping = vma->vm_file->f_mapping;
6289 if (pte_none(ptent))
6290 pgoff = linear_page_index(vma, addr);
6291 else /* pte_file(ptent) is true */
6292 pgoff = pte_to_pgoff(ptent);
6294 /* page is moved even if it's not RSS of this task(page-faulted). */
6295 page = find_get_page(mapping, pgoff);
6298 /* shmem/tmpfs may report page out on swap: account for that too. */
6299 if (radix_tree_exceptional_entry(page)) {
6300 swp_entry_t swap = radix_to_swp_entry(page);
6301 if (do_swap_account)
6303 page = find_get_page(&swapper_space, swap.val);
6309 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
6310 unsigned long addr, pte_t ptent, union mc_target *target)
6312 struct page *page = NULL;
6313 struct page_cgroup *pc;
6314 enum mc_target_type ret = MC_TARGET_NONE;
6315 swp_entry_t ent = { .val = 0 };
6317 if (pte_present(ptent))
6318 page = mc_handle_present_pte(vma, addr, ptent);
6319 else if (is_swap_pte(ptent))
6320 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
6321 else if (pte_none(ptent) || pte_file(ptent))
6322 page = mc_handle_file_pte(vma, addr, ptent, &ent);
6324 if (!page && !ent.val)
6327 pc = lookup_page_cgroup(page);
6329 * Do only loose check w/o page_cgroup lock.
6330 * mem_cgroup_move_account() checks the pc is valid or not under
6333 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6334 ret = MC_TARGET_PAGE;
6336 target->page = page;
6338 if (!ret || !target)
6341 /* There is a swap entry and a page doesn't exist or isn't charged */
6342 if (ent.val && !ret &&
6343 css_id(&mc.from->css) == lookup_swap_cgroup_id(ent)) {
6344 ret = MC_TARGET_SWAP;
6351 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6353 * We don't consider swapping or file mapped pages because THP does not
6354 * support them for now.
6355 * Caller should make sure that pmd_trans_huge(pmd) is true.
6357 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6358 unsigned long addr, pmd_t pmd, union mc_target *target)
6360 struct page *page = NULL;
6361 struct page_cgroup *pc;
6362 enum mc_target_type ret = MC_TARGET_NONE;
6364 page = pmd_page(pmd);
6365 VM_BUG_ON(!page || !PageHead(page));
6368 pc = lookup_page_cgroup(page);
6369 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6370 ret = MC_TARGET_PAGE;
6373 target->page = page;
6379 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6380 unsigned long addr, pmd_t pmd, union mc_target *target)
6382 return MC_TARGET_NONE;
6386 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
6387 unsigned long addr, unsigned long end,
6388 struct mm_walk *walk)
6390 struct vm_area_struct *vma = walk->private;
6394 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6395 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
6396 mc.precharge += HPAGE_PMD_NR;
6397 spin_unlock(&vma->vm_mm->page_table_lock);
6401 if (pmd_trans_unstable(pmd))
6403 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6404 for (; addr != end; pte++, addr += PAGE_SIZE)
6405 if (get_mctgt_type(vma, addr, *pte, NULL))
6406 mc.precharge++; /* increment precharge temporarily */
6407 pte_unmap_unlock(pte - 1, ptl);
6413 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
6415 unsigned long precharge;
6416 struct vm_area_struct *vma;
6418 down_read(&mm->mmap_sem);
6419 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6420 struct mm_walk mem_cgroup_count_precharge_walk = {
6421 .pmd_entry = mem_cgroup_count_precharge_pte_range,
6425 if (is_vm_hugetlb_page(vma))
6427 walk_page_range(vma->vm_start, vma->vm_end,
6428 &mem_cgroup_count_precharge_walk);
6430 up_read(&mm->mmap_sem);
6432 precharge = mc.precharge;
6438 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
6440 unsigned long precharge = mem_cgroup_count_precharge(mm);
6442 VM_BUG_ON(mc.moving_task);
6443 mc.moving_task = current;
6444 return mem_cgroup_do_precharge(precharge);
6447 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6448 static void __mem_cgroup_clear_mc(void)
6450 struct mem_cgroup *from = mc.from;
6451 struct mem_cgroup *to = mc.to;
6453 /* we must uncharge all the leftover precharges from mc.to */
6455 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
6459 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6460 * we must uncharge here.
6462 if (mc.moved_charge) {
6463 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
6464 mc.moved_charge = 0;
6466 /* we must fixup refcnts and charges */
6467 if (mc.moved_swap) {
6468 /* uncharge swap account from the old cgroup */
6469 if (!mem_cgroup_is_root(mc.from))
6470 res_counter_uncharge(&mc.from->memsw,
6471 PAGE_SIZE * mc.moved_swap);
6472 __mem_cgroup_put(mc.from, mc.moved_swap);
6474 if (!mem_cgroup_is_root(mc.to)) {
6476 * we charged both to->res and to->memsw, so we should
6479 res_counter_uncharge(&mc.to->res,
6480 PAGE_SIZE * mc.moved_swap);
6482 /* we've already done mem_cgroup_get(mc.to) */
6485 memcg_oom_recover(from);
6486 memcg_oom_recover(to);
6487 wake_up_all(&mc.waitq);
6490 static void mem_cgroup_clear_mc(void)
6492 struct mem_cgroup *from = mc.from;
6495 * we must clear moving_task before waking up waiters at the end of
6498 mc.moving_task = NULL;
6499 __mem_cgroup_clear_mc();
6500 spin_lock(&mc.lock);
6503 spin_unlock(&mc.lock);
6504 mem_cgroup_end_move(from);
6507 static int mem_cgroup_can_attach(struct cgroup *cgroup,
6508 struct cgroup_taskset *tset)
6510 struct task_struct *p = cgroup_taskset_first(tset);
6512 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup);
6514 if (memcg->move_charge_at_immigrate) {
6515 struct mm_struct *mm;
6516 struct mem_cgroup *from = mem_cgroup_from_task(p);
6518 VM_BUG_ON(from == memcg);
6520 mm = get_task_mm(p);
6523 /* We move charges only when we move a owner of the mm */
6524 if (mm->owner == p) {
6527 VM_BUG_ON(mc.precharge);
6528 VM_BUG_ON(mc.moved_charge);
6529 VM_BUG_ON(mc.moved_swap);
6530 mem_cgroup_start_move(from);
6531 spin_lock(&mc.lock);
6534 spin_unlock(&mc.lock);
6535 /* We set mc.moving_task later */
6537 ret = mem_cgroup_precharge_mc(mm);
6539 mem_cgroup_clear_mc();
6546 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
6547 struct cgroup_taskset *tset)
6549 mem_cgroup_clear_mc();
6552 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6553 unsigned long addr, unsigned long end,
6554 struct mm_walk *walk)
6557 struct vm_area_struct *vma = walk->private;
6560 enum mc_target_type target_type;
6561 union mc_target target;
6563 struct page_cgroup *pc;
6566 * We don't take compound_lock() here but no race with splitting thp
6568 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6569 * under splitting, which means there's no concurrent thp split,
6570 * - if another thread runs into split_huge_page() just after we
6571 * entered this if-block, the thread must wait for page table lock
6572 * to be unlocked in __split_huge_page_splitting(), where the main
6573 * part of thp split is not executed yet.
6575 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6576 if (mc.precharge < HPAGE_PMD_NR) {
6577 spin_unlock(&vma->vm_mm->page_table_lock);
6580 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6581 if (target_type == MC_TARGET_PAGE) {
6583 if (!isolate_lru_page(page)) {
6584 pc = lookup_page_cgroup(page);
6585 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
6586 pc, mc.from, mc.to)) {
6587 mc.precharge -= HPAGE_PMD_NR;
6588 mc.moved_charge += HPAGE_PMD_NR;
6590 putback_lru_page(page);
6594 spin_unlock(&vma->vm_mm->page_table_lock);
6598 if (pmd_trans_unstable(pmd))
6601 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6602 for (; addr != end; addr += PAGE_SIZE) {
6603 pte_t ptent = *(pte++);
6609 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6610 case MC_TARGET_PAGE:
6612 if (isolate_lru_page(page))
6614 pc = lookup_page_cgroup(page);
6615 if (!mem_cgroup_move_account(page, 1, pc,
6618 /* we uncharge from mc.from later. */
6621 putback_lru_page(page);
6622 put: /* get_mctgt_type() gets the page */
6625 case MC_TARGET_SWAP:
6627 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6629 /* we fixup refcnts and charges later. */
6637 pte_unmap_unlock(pte - 1, ptl);
6642 * We have consumed all precharges we got in can_attach().
6643 * We try charge one by one, but don't do any additional
6644 * charges to mc.to if we have failed in charge once in attach()
6647 ret = mem_cgroup_do_precharge(1);
6655 static void mem_cgroup_move_charge(struct mm_struct *mm)
6657 struct vm_area_struct *vma;
6659 lru_add_drain_all();
6661 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
6663 * Someone who are holding the mmap_sem might be waiting in
6664 * waitq. So we cancel all extra charges, wake up all waiters,
6665 * and retry. Because we cancel precharges, we might not be able
6666 * to move enough charges, but moving charge is a best-effort
6667 * feature anyway, so it wouldn't be a big problem.
6669 __mem_cgroup_clear_mc();
6673 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6675 struct mm_walk mem_cgroup_move_charge_walk = {
6676 .pmd_entry = mem_cgroup_move_charge_pte_range,
6680 if (is_vm_hugetlb_page(vma))
6682 ret = walk_page_range(vma->vm_start, vma->vm_end,
6683 &mem_cgroup_move_charge_walk);
6686 * means we have consumed all precharges and failed in
6687 * doing additional charge. Just abandon here.
6691 up_read(&mm->mmap_sem);
6694 static void mem_cgroup_move_task(struct cgroup *cont,
6695 struct cgroup_taskset *tset)
6697 struct task_struct *p = cgroup_taskset_first(tset);
6698 struct mm_struct *mm = get_task_mm(p);
6702 mem_cgroup_move_charge(mm);
6706 mem_cgroup_clear_mc();
6708 #else /* !CONFIG_MMU */
6709 static int mem_cgroup_can_attach(struct cgroup *cgroup,
6710 struct cgroup_taskset *tset)
6714 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
6715 struct cgroup_taskset *tset)
6718 static void mem_cgroup_move_task(struct cgroup *cont,
6719 struct cgroup_taskset *tset)
6724 struct cgroup_subsys mem_cgroup_subsys = {
6726 .subsys_id = mem_cgroup_subsys_id,
6727 .create = mem_cgroup_create,
6728 .pre_destroy = mem_cgroup_pre_destroy,
6729 .destroy = mem_cgroup_destroy,
6730 .can_attach = mem_cgroup_can_attach,
6731 .cancel_attach = mem_cgroup_cancel_attach,
6732 .attach = mem_cgroup_move_task,
6733 .base_cftypes = mem_cgroup_files,
6738 #ifdef CONFIG_MEMCG_SWAP
6739 static int __init enable_swap_account(char *s)
6741 /* consider enabled if no parameter or 1 is given */
6742 if (!strcmp(s, "1"))
6743 really_do_swap_account = 1;
6744 else if (!strcmp(s, "0"))
6745 really_do_swap_account = 0;
6748 __setup("swapaccount=", enable_swap_account);