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 DEFINE_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;
1193 struct lruvec *lruvec;
1195 if (mem_cgroup_disabled()) {
1196 lruvec = &zone->lruvec;
1200 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1201 lruvec = &mz->lruvec;
1204 * Since a node can be onlined after the mem_cgroup was created,
1205 * we have to be prepared to initialize lruvec->zone here;
1206 * and if offlined then reonlined, we need to reinitialize it.
1208 if (unlikely(lruvec->zone != zone))
1209 lruvec->zone = zone;
1214 * Following LRU functions are allowed to be used without PCG_LOCK.
1215 * Operations are called by routine of global LRU independently from memcg.
1216 * What we have to take care of here is validness of pc->mem_cgroup.
1218 * Changes to pc->mem_cgroup happens when
1221 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1222 * It is added to LRU before charge.
1223 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1224 * When moving account, the page is not on LRU. It's isolated.
1228 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1230 * @zone: zone of the page
1232 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1234 struct mem_cgroup_per_zone *mz;
1235 struct mem_cgroup *memcg;
1236 struct page_cgroup *pc;
1237 struct lruvec *lruvec;
1239 if (mem_cgroup_disabled()) {
1240 lruvec = &zone->lruvec;
1244 pc = lookup_page_cgroup(page);
1245 memcg = pc->mem_cgroup;
1248 * Surreptitiously switch any uncharged offlist page to root:
1249 * an uncharged page off lru does nothing to secure
1250 * its former mem_cgroup from sudden removal.
1252 * Our caller holds lru_lock, and PageCgroupUsed is updated
1253 * under page_cgroup lock: between them, they make all uses
1254 * of pc->mem_cgroup safe.
1256 if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1257 pc->mem_cgroup = memcg = root_mem_cgroup;
1259 mz = page_cgroup_zoneinfo(memcg, page);
1260 lruvec = &mz->lruvec;
1263 * Since a node can be onlined after the mem_cgroup was created,
1264 * we have to be prepared to initialize lruvec->zone here;
1265 * and if offlined then reonlined, we need to reinitialize it.
1267 if (unlikely(lruvec->zone != zone))
1268 lruvec->zone = zone;
1273 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1274 * @lruvec: mem_cgroup per zone lru vector
1275 * @lru: index of lru list the page is sitting on
1276 * @nr_pages: positive when adding or negative when removing
1278 * This function must be called when a page is added to or removed from an
1281 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1284 struct mem_cgroup_per_zone *mz;
1285 unsigned long *lru_size;
1287 if (mem_cgroup_disabled())
1290 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1291 lru_size = mz->lru_size + lru;
1292 *lru_size += nr_pages;
1293 VM_BUG_ON((long)(*lru_size) < 0);
1297 * Checks whether given mem is same or in the root_mem_cgroup's
1300 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1301 struct mem_cgroup *memcg)
1303 if (root_memcg == memcg)
1305 if (!root_memcg->use_hierarchy || !memcg)
1307 return css_is_ancestor(&memcg->css, &root_memcg->css);
1310 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1311 struct mem_cgroup *memcg)
1316 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1321 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *memcg)
1324 struct mem_cgroup *curr = NULL;
1325 struct task_struct *p;
1327 p = find_lock_task_mm(task);
1329 curr = try_get_mem_cgroup_from_mm(p->mm);
1333 * All threads may have already detached their mm's, but the oom
1334 * killer still needs to detect if they have already been oom
1335 * killed to prevent needlessly killing additional tasks.
1338 curr = mem_cgroup_from_task(task);
1340 css_get(&curr->css);
1346 * We should check use_hierarchy of "memcg" not "curr". Because checking
1347 * use_hierarchy of "curr" here make this function true if hierarchy is
1348 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1349 * hierarchy(even if use_hierarchy is disabled in "memcg").
1351 ret = mem_cgroup_same_or_subtree(memcg, curr);
1352 css_put(&curr->css);
1356 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1358 unsigned long inactive_ratio;
1359 unsigned long inactive;
1360 unsigned long active;
1363 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1364 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1366 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1368 inactive_ratio = int_sqrt(10 * gb);
1372 return inactive * inactive_ratio < active;
1375 int mem_cgroup_inactive_file_is_low(struct lruvec *lruvec)
1377 unsigned long active;
1378 unsigned long inactive;
1380 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_FILE);
1381 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_FILE);
1383 return (active > inactive);
1386 #define mem_cgroup_from_res_counter(counter, member) \
1387 container_of(counter, struct mem_cgroup, member)
1390 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1391 * @memcg: the memory cgroup
1393 * Returns the maximum amount of memory @mem can be charged with, in
1396 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1398 unsigned long long margin;
1400 margin = res_counter_margin(&memcg->res);
1401 if (do_swap_account)
1402 margin = min(margin, res_counter_margin(&memcg->memsw));
1403 return margin >> PAGE_SHIFT;
1406 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1408 struct cgroup *cgrp = memcg->css.cgroup;
1411 if (cgrp->parent == NULL)
1412 return vm_swappiness;
1414 return memcg->swappiness;
1418 * memcg->moving_account is used for checking possibility that some thread is
1419 * calling move_account(). When a thread on CPU-A starts moving pages under
1420 * a memcg, other threads should check memcg->moving_account under
1421 * rcu_read_lock(), like this:
1425 * memcg->moving_account+1 if (memcg->mocing_account)
1427 * synchronize_rcu() update something.
1432 /* for quick checking without looking up memcg */
1433 atomic_t memcg_moving __read_mostly;
1435 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1437 atomic_inc(&memcg_moving);
1438 atomic_inc(&memcg->moving_account);
1442 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1445 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1446 * We check NULL in callee rather than caller.
1449 atomic_dec(&memcg_moving);
1450 atomic_dec(&memcg->moving_account);
1455 * 2 routines for checking "mem" is under move_account() or not.
1457 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1458 * is used for avoiding races in accounting. If true,
1459 * pc->mem_cgroup may be overwritten.
1461 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1462 * under hierarchy of moving cgroups. This is for
1463 * waiting at hith-memory prressure caused by "move".
1466 static bool mem_cgroup_stolen(struct mem_cgroup *memcg)
1468 VM_BUG_ON(!rcu_read_lock_held());
1469 return atomic_read(&memcg->moving_account) > 0;
1472 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1474 struct mem_cgroup *from;
1475 struct mem_cgroup *to;
1478 * Unlike task_move routines, we access mc.to, mc.from not under
1479 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1481 spin_lock(&mc.lock);
1487 ret = mem_cgroup_same_or_subtree(memcg, from)
1488 || mem_cgroup_same_or_subtree(memcg, to);
1490 spin_unlock(&mc.lock);
1494 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1496 if (mc.moving_task && current != mc.moving_task) {
1497 if (mem_cgroup_under_move(memcg)) {
1499 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1500 /* moving charge context might have finished. */
1503 finish_wait(&mc.waitq, &wait);
1511 * Take this lock when
1512 * - a code tries to modify page's memcg while it's USED.
1513 * - a code tries to modify page state accounting in a memcg.
1514 * see mem_cgroup_stolen(), too.
1516 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1517 unsigned long *flags)
1519 spin_lock_irqsave(&memcg->move_lock, *flags);
1522 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1523 unsigned long *flags)
1525 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1529 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1530 * @memcg: The memory cgroup that went over limit
1531 * @p: Task that is going to be killed
1533 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1536 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1538 struct cgroup *task_cgrp;
1539 struct cgroup *mem_cgrp;
1541 * Need a buffer in BSS, can't rely on allocations. The code relies
1542 * on the assumption that OOM is serialized for memory controller.
1543 * If this assumption is broken, revisit this code.
1545 static char memcg_name[PATH_MAX];
1553 mem_cgrp = memcg->css.cgroup;
1554 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1556 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1559 * Unfortunately, we are unable to convert to a useful name
1560 * But we'll still print out the usage information
1567 printk(KERN_INFO "Task in %s killed", memcg_name);
1570 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1578 * Continues from above, so we don't need an KERN_ level
1580 printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1583 printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1584 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1585 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1586 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1587 printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1589 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1590 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1591 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1592 printk(KERN_INFO "kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1593 res_counter_read_u64(&memcg->kmem, RES_USAGE) >> 10,
1594 res_counter_read_u64(&memcg->kmem, RES_LIMIT) >> 10,
1595 res_counter_read_u64(&memcg->kmem, RES_FAILCNT));
1599 * This function returns the number of memcg under hierarchy tree. Returns
1600 * 1(self count) if no children.
1602 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1605 struct mem_cgroup *iter;
1607 for_each_mem_cgroup_tree(iter, memcg)
1613 * Return the memory (and swap, if configured) limit for a memcg.
1615 static u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1619 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1622 * Do not consider swap space if we cannot swap due to swappiness
1624 if (mem_cgroup_swappiness(memcg)) {
1627 limit += total_swap_pages << PAGE_SHIFT;
1628 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1631 * If memsw is finite and limits the amount of swap space
1632 * available to this memcg, return that limit.
1634 limit = min(limit, memsw);
1640 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1643 struct mem_cgroup *iter;
1644 unsigned long chosen_points = 0;
1645 unsigned long totalpages;
1646 unsigned int points = 0;
1647 struct task_struct *chosen = NULL;
1650 * If current has a pending SIGKILL, then automatically select it. The
1651 * goal is to allow it to allocate so that it may quickly exit and free
1654 if (fatal_signal_pending(current)) {
1655 set_thread_flag(TIF_MEMDIE);
1659 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1660 totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
1661 for_each_mem_cgroup_tree(iter, memcg) {
1662 struct cgroup *cgroup = iter->css.cgroup;
1663 struct cgroup_iter it;
1664 struct task_struct *task;
1666 cgroup_iter_start(cgroup, &it);
1667 while ((task = cgroup_iter_next(cgroup, &it))) {
1668 switch (oom_scan_process_thread(task, totalpages, NULL,
1670 case OOM_SCAN_SELECT:
1672 put_task_struct(chosen);
1674 chosen_points = ULONG_MAX;
1675 get_task_struct(chosen);
1677 case OOM_SCAN_CONTINUE:
1679 case OOM_SCAN_ABORT:
1680 cgroup_iter_end(cgroup, &it);
1681 mem_cgroup_iter_break(memcg, iter);
1683 put_task_struct(chosen);
1688 points = oom_badness(task, memcg, NULL, totalpages);
1689 if (points > chosen_points) {
1691 put_task_struct(chosen);
1693 chosen_points = points;
1694 get_task_struct(chosen);
1697 cgroup_iter_end(cgroup, &it);
1702 points = chosen_points * 1000 / totalpages;
1703 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1704 NULL, "Memory cgroup out of memory");
1707 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1709 unsigned long flags)
1711 unsigned long total = 0;
1712 bool noswap = false;
1715 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1717 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1720 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1722 drain_all_stock_async(memcg);
1723 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1725 * Allow limit shrinkers, which are triggered directly
1726 * by userspace, to catch signals and stop reclaim
1727 * after minimal progress, regardless of the margin.
1729 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1731 if (mem_cgroup_margin(memcg))
1734 * If nothing was reclaimed after two attempts, there
1735 * may be no reclaimable pages in this hierarchy.
1744 * test_mem_cgroup_node_reclaimable
1745 * @memcg: the target memcg
1746 * @nid: the node ID to be checked.
1747 * @noswap : specify true here if the user wants flle only information.
1749 * This function returns whether the specified memcg contains any
1750 * reclaimable pages on a node. Returns true if there are any reclaimable
1751 * pages in the node.
1753 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1754 int nid, bool noswap)
1756 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1758 if (noswap || !total_swap_pages)
1760 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1765 #if MAX_NUMNODES > 1
1768 * Always updating the nodemask is not very good - even if we have an empty
1769 * list or the wrong list here, we can start from some node and traverse all
1770 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1773 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1777 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1778 * pagein/pageout changes since the last update.
1780 if (!atomic_read(&memcg->numainfo_events))
1782 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1785 /* make a nodemask where this memcg uses memory from */
1786 memcg->scan_nodes = node_states[N_HIGH_MEMORY];
1788 for_each_node_mask(nid, node_states[N_HIGH_MEMORY]) {
1790 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1791 node_clear(nid, memcg->scan_nodes);
1794 atomic_set(&memcg->numainfo_events, 0);
1795 atomic_set(&memcg->numainfo_updating, 0);
1799 * Selecting a node where we start reclaim from. Because what we need is just
1800 * reducing usage counter, start from anywhere is O,K. Considering
1801 * memory reclaim from current node, there are pros. and cons.
1803 * Freeing memory from current node means freeing memory from a node which
1804 * we'll use or we've used. So, it may make LRU bad. And if several threads
1805 * hit limits, it will see a contention on a node. But freeing from remote
1806 * node means more costs for memory reclaim because of memory latency.
1808 * Now, we use round-robin. Better algorithm is welcomed.
1810 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1814 mem_cgroup_may_update_nodemask(memcg);
1815 node = memcg->last_scanned_node;
1817 node = next_node(node, memcg->scan_nodes);
1818 if (node == MAX_NUMNODES)
1819 node = first_node(memcg->scan_nodes);
1821 * We call this when we hit limit, not when pages are added to LRU.
1822 * No LRU may hold pages because all pages are UNEVICTABLE or
1823 * memcg is too small and all pages are not on LRU. In that case,
1824 * we use curret node.
1826 if (unlikely(node == MAX_NUMNODES))
1827 node = numa_node_id();
1829 memcg->last_scanned_node = node;
1834 * Check all nodes whether it contains reclaimable pages or not.
1835 * For quick scan, we make use of scan_nodes. This will allow us to skip
1836 * unused nodes. But scan_nodes is lazily updated and may not cotain
1837 * enough new information. We need to do double check.
1839 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1844 * quick check...making use of scan_node.
1845 * We can skip unused nodes.
1847 if (!nodes_empty(memcg->scan_nodes)) {
1848 for (nid = first_node(memcg->scan_nodes);
1850 nid = next_node(nid, memcg->scan_nodes)) {
1852 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1857 * Check rest of nodes.
1859 for_each_node_state(nid, N_HIGH_MEMORY) {
1860 if (node_isset(nid, memcg->scan_nodes))
1862 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1869 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1874 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1876 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
1880 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1883 unsigned long *total_scanned)
1885 struct mem_cgroup *victim = NULL;
1888 unsigned long excess;
1889 unsigned long nr_scanned;
1890 struct mem_cgroup_reclaim_cookie reclaim = {
1895 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
1898 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1903 * If we have not been able to reclaim
1904 * anything, it might because there are
1905 * no reclaimable pages under this hierarchy
1910 * We want to do more targeted reclaim.
1911 * excess >> 2 is not to excessive so as to
1912 * reclaim too much, nor too less that we keep
1913 * coming back to reclaim from this cgroup
1915 if (total >= (excess >> 2) ||
1916 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1921 if (!mem_cgroup_reclaimable(victim, false))
1923 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1925 *total_scanned += nr_scanned;
1926 if (!res_counter_soft_limit_excess(&root_memcg->res))
1929 mem_cgroup_iter_break(root_memcg, victim);
1934 * Check OOM-Killer is already running under our hierarchy.
1935 * If someone is running, return false.
1936 * Has to be called with memcg_oom_lock
1938 static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg)
1940 struct mem_cgroup *iter, *failed = NULL;
1942 for_each_mem_cgroup_tree(iter, memcg) {
1943 if (iter->oom_lock) {
1945 * this subtree of our hierarchy is already locked
1946 * so we cannot give a lock.
1949 mem_cgroup_iter_break(memcg, iter);
1952 iter->oom_lock = true;
1959 * OK, we failed to lock the whole subtree so we have to clean up
1960 * what we set up to the failing subtree
1962 for_each_mem_cgroup_tree(iter, memcg) {
1963 if (iter == failed) {
1964 mem_cgroup_iter_break(memcg, iter);
1967 iter->oom_lock = false;
1973 * Has to be called with memcg_oom_lock
1975 static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1977 struct mem_cgroup *iter;
1979 for_each_mem_cgroup_tree(iter, memcg)
1980 iter->oom_lock = false;
1984 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1986 struct mem_cgroup *iter;
1988 for_each_mem_cgroup_tree(iter, memcg)
1989 atomic_inc(&iter->under_oom);
1992 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1994 struct mem_cgroup *iter;
1997 * When a new child is created while the hierarchy is under oom,
1998 * mem_cgroup_oom_lock() may not be called. We have to use
1999 * atomic_add_unless() here.
2001 for_each_mem_cgroup_tree(iter, memcg)
2002 atomic_add_unless(&iter->under_oom, -1, 0);
2005 static DEFINE_SPINLOCK(memcg_oom_lock);
2006 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
2008 struct oom_wait_info {
2009 struct mem_cgroup *memcg;
2013 static int memcg_oom_wake_function(wait_queue_t *wait,
2014 unsigned mode, int sync, void *arg)
2016 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
2017 struct mem_cgroup *oom_wait_memcg;
2018 struct oom_wait_info *oom_wait_info;
2020 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
2021 oom_wait_memcg = oom_wait_info->memcg;
2024 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2025 * Then we can use css_is_ancestor without taking care of RCU.
2027 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
2028 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
2030 return autoremove_wake_function(wait, mode, sync, arg);
2033 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
2035 /* for filtering, pass "memcg" as argument. */
2036 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
2039 static void memcg_oom_recover(struct mem_cgroup *memcg)
2041 if (memcg && atomic_read(&memcg->under_oom))
2042 memcg_wakeup_oom(memcg);
2046 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
2048 static bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask,
2051 struct oom_wait_info owait;
2052 bool locked, need_to_kill;
2054 owait.memcg = memcg;
2055 owait.wait.flags = 0;
2056 owait.wait.func = memcg_oom_wake_function;
2057 owait.wait.private = current;
2058 INIT_LIST_HEAD(&owait.wait.task_list);
2059 need_to_kill = true;
2060 mem_cgroup_mark_under_oom(memcg);
2062 /* At first, try to OOM lock hierarchy under memcg.*/
2063 spin_lock(&memcg_oom_lock);
2064 locked = mem_cgroup_oom_lock(memcg);
2066 * Even if signal_pending(), we can't quit charge() loop without
2067 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
2068 * under OOM is always welcomed, use TASK_KILLABLE here.
2070 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2071 if (!locked || memcg->oom_kill_disable)
2072 need_to_kill = false;
2074 mem_cgroup_oom_notify(memcg);
2075 spin_unlock(&memcg_oom_lock);
2078 finish_wait(&memcg_oom_waitq, &owait.wait);
2079 mem_cgroup_out_of_memory(memcg, mask, order);
2082 finish_wait(&memcg_oom_waitq, &owait.wait);
2084 spin_lock(&memcg_oom_lock);
2086 mem_cgroup_oom_unlock(memcg);
2087 memcg_wakeup_oom(memcg);
2088 spin_unlock(&memcg_oom_lock);
2090 mem_cgroup_unmark_under_oom(memcg);
2092 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
2094 /* Give chance to dying process */
2095 schedule_timeout_uninterruptible(1);
2100 * Currently used to update mapped file statistics, but the routine can be
2101 * generalized to update other statistics as well.
2103 * Notes: Race condition
2105 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2106 * it tends to be costly. But considering some conditions, we doesn't need
2107 * to do so _always_.
2109 * Considering "charge", lock_page_cgroup() is not required because all
2110 * file-stat operations happen after a page is attached to radix-tree. There
2111 * are no race with "charge".
2113 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2114 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2115 * if there are race with "uncharge". Statistics itself is properly handled
2118 * Considering "move", this is an only case we see a race. To make the race
2119 * small, we check mm->moving_account and detect there are possibility of race
2120 * If there is, we take a lock.
2123 void __mem_cgroup_begin_update_page_stat(struct page *page,
2124 bool *locked, unsigned long *flags)
2126 struct mem_cgroup *memcg;
2127 struct page_cgroup *pc;
2129 pc = lookup_page_cgroup(page);
2131 memcg = pc->mem_cgroup;
2132 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2135 * If this memory cgroup is not under account moving, we don't
2136 * need to take move_lock_mem_cgroup(). Because we already hold
2137 * rcu_read_lock(), any calls to move_account will be delayed until
2138 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2140 if (!mem_cgroup_stolen(memcg))
2143 move_lock_mem_cgroup(memcg, flags);
2144 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2145 move_unlock_mem_cgroup(memcg, flags);
2151 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
2153 struct page_cgroup *pc = lookup_page_cgroup(page);
2156 * It's guaranteed that pc->mem_cgroup never changes while
2157 * lock is held because a routine modifies pc->mem_cgroup
2158 * should take move_lock_mem_cgroup().
2160 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
2163 void mem_cgroup_update_page_stat(struct page *page,
2164 enum mem_cgroup_page_stat_item idx, int val)
2166 struct mem_cgroup *memcg;
2167 struct page_cgroup *pc = lookup_page_cgroup(page);
2168 unsigned long uninitialized_var(flags);
2170 if (mem_cgroup_disabled())
2173 memcg = pc->mem_cgroup;
2174 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2178 case MEMCG_NR_FILE_MAPPED:
2179 idx = MEM_CGROUP_STAT_FILE_MAPPED;
2185 this_cpu_add(memcg->stat->count[idx], val);
2189 * size of first charge trial. "32" comes from vmscan.c's magic value.
2190 * TODO: maybe necessary to use big numbers in big irons.
2192 #define CHARGE_BATCH 32U
2193 struct memcg_stock_pcp {
2194 struct mem_cgroup *cached; /* this never be root cgroup */
2195 unsigned int nr_pages;
2196 struct work_struct work;
2197 unsigned long flags;
2198 #define FLUSHING_CACHED_CHARGE 0
2200 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2201 static DEFINE_MUTEX(percpu_charge_mutex);
2204 * consume_stock: Try to consume stocked charge on this cpu.
2205 * @memcg: memcg to consume from.
2206 * @nr_pages: how many pages to charge.
2208 * The charges will only happen if @memcg matches the current cpu's memcg
2209 * stock, and at least @nr_pages are available in that stock. Failure to
2210 * service an allocation will refill the stock.
2212 * returns true if successful, false otherwise.
2214 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2216 struct memcg_stock_pcp *stock;
2219 if (nr_pages > CHARGE_BATCH)
2222 stock = &get_cpu_var(memcg_stock);
2223 if (memcg == stock->cached && stock->nr_pages >= nr_pages)
2224 stock->nr_pages -= nr_pages;
2225 else /* need to call res_counter_charge */
2227 put_cpu_var(memcg_stock);
2232 * Returns stocks cached in percpu to res_counter and reset cached information.
2234 static void drain_stock(struct memcg_stock_pcp *stock)
2236 struct mem_cgroup *old = stock->cached;
2238 if (stock->nr_pages) {
2239 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2241 res_counter_uncharge(&old->res, bytes);
2242 if (do_swap_account)
2243 res_counter_uncharge(&old->memsw, bytes);
2244 stock->nr_pages = 0;
2246 stock->cached = NULL;
2250 * This must be called under preempt disabled or must be called by
2251 * a thread which is pinned to local cpu.
2253 static void drain_local_stock(struct work_struct *dummy)
2255 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2257 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2261 * Cache charges(val) which is from res_counter, to local per_cpu area.
2262 * This will be consumed by consume_stock() function, later.
2264 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2266 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2268 if (stock->cached != memcg) { /* reset if necessary */
2270 stock->cached = memcg;
2272 stock->nr_pages += nr_pages;
2273 put_cpu_var(memcg_stock);
2277 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2278 * of the hierarchy under it. sync flag says whether we should block
2279 * until the work is done.
2281 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2285 /* Notify other cpus that system-wide "drain" is running */
2288 for_each_online_cpu(cpu) {
2289 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2290 struct mem_cgroup *memcg;
2292 memcg = stock->cached;
2293 if (!memcg || !stock->nr_pages)
2295 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2297 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2299 drain_local_stock(&stock->work);
2301 schedule_work_on(cpu, &stock->work);
2309 for_each_online_cpu(cpu) {
2310 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2311 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2312 flush_work(&stock->work);
2319 * Tries to drain stocked charges in other cpus. This function is asynchronous
2320 * and just put a work per cpu for draining localy on each cpu. Caller can
2321 * expects some charges will be back to res_counter later but cannot wait for
2324 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2327 * If someone calls draining, avoid adding more kworker runs.
2329 if (!mutex_trylock(&percpu_charge_mutex))
2331 drain_all_stock(root_memcg, false);
2332 mutex_unlock(&percpu_charge_mutex);
2335 /* This is a synchronous drain interface. */
2336 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2338 /* called when force_empty is called */
2339 mutex_lock(&percpu_charge_mutex);
2340 drain_all_stock(root_memcg, true);
2341 mutex_unlock(&percpu_charge_mutex);
2345 * This function drains percpu counter value from DEAD cpu and
2346 * move it to local cpu. Note that this function can be preempted.
2348 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2352 spin_lock(&memcg->pcp_counter_lock);
2353 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2354 long x = per_cpu(memcg->stat->count[i], cpu);
2356 per_cpu(memcg->stat->count[i], cpu) = 0;
2357 memcg->nocpu_base.count[i] += x;
2359 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2360 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2362 per_cpu(memcg->stat->events[i], cpu) = 0;
2363 memcg->nocpu_base.events[i] += x;
2365 spin_unlock(&memcg->pcp_counter_lock);
2368 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2369 unsigned long action,
2372 int cpu = (unsigned long)hcpu;
2373 struct memcg_stock_pcp *stock;
2374 struct mem_cgroup *iter;
2376 if (action == CPU_ONLINE)
2379 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2382 for_each_mem_cgroup(iter)
2383 mem_cgroup_drain_pcp_counter(iter, cpu);
2385 stock = &per_cpu(memcg_stock, cpu);
2391 /* See __mem_cgroup_try_charge() for details */
2393 CHARGE_OK, /* success */
2394 CHARGE_RETRY, /* need to retry but retry is not bad */
2395 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2396 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2397 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2400 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2401 unsigned int nr_pages, unsigned int min_pages,
2404 unsigned long csize = nr_pages * PAGE_SIZE;
2405 struct mem_cgroup *mem_over_limit;
2406 struct res_counter *fail_res;
2407 unsigned long flags = 0;
2410 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2413 if (!do_swap_account)
2415 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2419 res_counter_uncharge(&memcg->res, csize);
2420 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2421 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2423 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2425 * Never reclaim on behalf of optional batching, retry with a
2426 * single page instead.
2428 if (nr_pages > min_pages)
2429 return CHARGE_RETRY;
2431 if (!(gfp_mask & __GFP_WAIT))
2432 return CHARGE_WOULDBLOCK;
2434 if (gfp_mask & __GFP_NORETRY)
2435 return CHARGE_NOMEM;
2437 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2438 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2439 return CHARGE_RETRY;
2441 * Even though the limit is exceeded at this point, reclaim
2442 * may have been able to free some pages. Retry the charge
2443 * before killing the task.
2445 * Only for regular pages, though: huge pages are rather
2446 * unlikely to succeed so close to the limit, and we fall back
2447 * to regular pages anyway in case of failure.
2449 if (nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER) && ret)
2450 return CHARGE_RETRY;
2453 * At task move, charge accounts can be doubly counted. So, it's
2454 * better to wait until the end of task_move if something is going on.
2456 if (mem_cgroup_wait_acct_move(mem_over_limit))
2457 return CHARGE_RETRY;
2459 /* If we don't need to call oom-killer at el, return immediately */
2461 return CHARGE_NOMEM;
2463 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask, get_order(csize)))
2464 return CHARGE_OOM_DIE;
2466 return CHARGE_RETRY;
2470 * __mem_cgroup_try_charge() does
2471 * 1. detect memcg to be charged against from passed *mm and *ptr,
2472 * 2. update res_counter
2473 * 3. call memory reclaim if necessary.
2475 * In some special case, if the task is fatal, fatal_signal_pending() or
2476 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2477 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2478 * as possible without any hazards. 2: all pages should have a valid
2479 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2480 * pointer, that is treated as a charge to root_mem_cgroup.
2482 * So __mem_cgroup_try_charge() will return
2483 * 0 ... on success, filling *ptr with a valid memcg pointer.
2484 * -ENOMEM ... charge failure because of resource limits.
2485 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2487 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2488 * the oom-killer can be invoked.
2490 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2492 unsigned int nr_pages,
2493 struct mem_cgroup **ptr,
2496 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2497 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2498 struct mem_cgroup *memcg = NULL;
2502 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2503 * in system level. So, allow to go ahead dying process in addition to
2506 if (unlikely(test_thread_flag(TIF_MEMDIE)
2507 || fatal_signal_pending(current)))
2511 * We always charge the cgroup the mm_struct belongs to.
2512 * The mm_struct's mem_cgroup changes on task migration if the
2513 * thread group leader migrates. It's possible that mm is not
2514 * set, if so charge the root memcg (happens for pagecache usage).
2517 *ptr = root_mem_cgroup;
2519 if (*ptr) { /* css should be a valid one */
2521 if (mem_cgroup_is_root(memcg))
2523 if (consume_stock(memcg, nr_pages))
2525 css_get(&memcg->css);
2527 struct task_struct *p;
2530 p = rcu_dereference(mm->owner);
2532 * Because we don't have task_lock(), "p" can exit.
2533 * In that case, "memcg" can point to root or p can be NULL with
2534 * race with swapoff. Then, we have small risk of mis-accouning.
2535 * But such kind of mis-account by race always happens because
2536 * we don't have cgroup_mutex(). It's overkill and we allo that
2538 * (*) swapoff at el will charge against mm-struct not against
2539 * task-struct. So, mm->owner can be NULL.
2541 memcg = mem_cgroup_from_task(p);
2543 memcg = root_mem_cgroup;
2544 if (mem_cgroup_is_root(memcg)) {
2548 if (consume_stock(memcg, nr_pages)) {
2550 * It seems dagerous to access memcg without css_get().
2551 * But considering how consume_stok works, it's not
2552 * necessary. If consume_stock success, some charges
2553 * from this memcg are cached on this cpu. So, we
2554 * don't need to call css_get()/css_tryget() before
2555 * calling consume_stock().
2560 /* after here, we may be blocked. we need to get refcnt */
2561 if (!css_tryget(&memcg->css)) {
2571 /* If killed, bypass charge */
2572 if (fatal_signal_pending(current)) {
2573 css_put(&memcg->css);
2578 if (oom && !nr_oom_retries) {
2580 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2583 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, nr_pages,
2588 case CHARGE_RETRY: /* not in OOM situation but retry */
2590 css_put(&memcg->css);
2593 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2594 css_put(&memcg->css);
2596 case CHARGE_NOMEM: /* OOM routine works */
2598 css_put(&memcg->css);
2601 /* If oom, we never return -ENOMEM */
2604 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2605 css_put(&memcg->css);
2608 } while (ret != CHARGE_OK);
2610 if (batch > nr_pages)
2611 refill_stock(memcg, batch - nr_pages);
2612 css_put(&memcg->css);
2620 *ptr = root_mem_cgroup;
2625 * Somemtimes we have to undo a charge we got by try_charge().
2626 * This function is for that and do uncharge, put css's refcnt.
2627 * gotten by try_charge().
2629 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2630 unsigned int nr_pages)
2632 if (!mem_cgroup_is_root(memcg)) {
2633 unsigned long bytes = nr_pages * PAGE_SIZE;
2635 res_counter_uncharge(&memcg->res, bytes);
2636 if (do_swap_account)
2637 res_counter_uncharge(&memcg->memsw, bytes);
2642 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2643 * This is useful when moving usage to parent cgroup.
2645 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2646 unsigned int nr_pages)
2648 unsigned long bytes = nr_pages * PAGE_SIZE;
2650 if (mem_cgroup_is_root(memcg))
2653 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2654 if (do_swap_account)
2655 res_counter_uncharge_until(&memcg->memsw,
2656 memcg->memsw.parent, bytes);
2660 * A helper function to get mem_cgroup from ID. must be called under
2661 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2662 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2663 * called against removed memcg.)
2665 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2667 struct cgroup_subsys_state *css;
2669 /* ID 0 is unused ID */
2672 css = css_lookup(&mem_cgroup_subsys, id);
2675 return mem_cgroup_from_css(css);
2678 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2680 struct mem_cgroup *memcg = NULL;
2681 struct page_cgroup *pc;
2685 VM_BUG_ON(!PageLocked(page));
2687 pc = lookup_page_cgroup(page);
2688 lock_page_cgroup(pc);
2689 if (PageCgroupUsed(pc)) {
2690 memcg = pc->mem_cgroup;
2691 if (memcg && !css_tryget(&memcg->css))
2693 } else if (PageSwapCache(page)) {
2694 ent.val = page_private(page);
2695 id = lookup_swap_cgroup_id(ent);
2697 memcg = mem_cgroup_lookup(id);
2698 if (memcg && !css_tryget(&memcg->css))
2702 unlock_page_cgroup(pc);
2706 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2708 unsigned int nr_pages,
2709 enum charge_type ctype,
2712 struct page_cgroup *pc = lookup_page_cgroup(page);
2713 struct zone *uninitialized_var(zone);
2714 struct lruvec *lruvec;
2715 bool was_on_lru = false;
2718 lock_page_cgroup(pc);
2719 VM_BUG_ON(PageCgroupUsed(pc));
2721 * we don't need page_cgroup_lock about tail pages, becase they are not
2722 * accessed by any other context at this point.
2726 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2727 * may already be on some other mem_cgroup's LRU. Take care of it.
2730 zone = page_zone(page);
2731 spin_lock_irq(&zone->lru_lock);
2732 if (PageLRU(page)) {
2733 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2735 del_page_from_lru_list(page, lruvec, page_lru(page));
2740 pc->mem_cgroup = memcg;
2742 * We access a page_cgroup asynchronously without lock_page_cgroup().
2743 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2744 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2745 * before USED bit, we need memory barrier here.
2746 * See mem_cgroup_add_lru_list(), etc.
2749 SetPageCgroupUsed(pc);
2753 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2754 VM_BUG_ON(PageLRU(page));
2756 add_page_to_lru_list(page, lruvec, page_lru(page));
2758 spin_unlock_irq(&zone->lru_lock);
2761 if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
2766 mem_cgroup_charge_statistics(memcg, anon, nr_pages);
2767 unlock_page_cgroup(pc);
2770 * "charge_statistics" updated event counter. Then, check it.
2771 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2772 * if they exceeds softlimit.
2774 memcg_check_events(memcg, page);
2777 static DEFINE_MUTEX(set_limit_mutex);
2779 #ifdef CONFIG_MEMCG_KMEM
2780 static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg)
2782 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) &&
2783 (memcg->kmem_account_flags & KMEM_ACCOUNTED_MASK);
2787 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2788 * in the memcg_cache_params struct.
2790 static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
2792 struct kmem_cache *cachep;
2794 VM_BUG_ON(p->is_root_cache);
2795 cachep = p->root_cache;
2796 return cachep->memcg_params->memcg_caches[memcg_cache_id(p->memcg)];
2799 #ifdef CONFIG_SLABINFO
2800 static int mem_cgroup_slabinfo_read(struct cgroup *cont, struct cftype *cft,
2803 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
2804 struct memcg_cache_params *params;
2806 if (!memcg_can_account_kmem(memcg))
2809 print_slabinfo_header(m);
2811 mutex_lock(&memcg->slab_caches_mutex);
2812 list_for_each_entry(params, &memcg->memcg_slab_caches, list)
2813 cache_show(memcg_params_to_cache(params), m);
2814 mutex_unlock(&memcg->slab_caches_mutex);
2820 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size)
2822 struct res_counter *fail_res;
2823 struct mem_cgroup *_memcg;
2827 ret = res_counter_charge(&memcg->kmem, size, &fail_res);
2832 * Conditions under which we can wait for the oom_killer. Those are
2833 * the same conditions tested by the core page allocator
2835 may_oom = (gfp & __GFP_FS) && !(gfp & __GFP_NORETRY);
2838 ret = __mem_cgroup_try_charge(NULL, gfp, size >> PAGE_SHIFT,
2841 if (ret == -EINTR) {
2843 * __mem_cgroup_try_charge() chosed to bypass to root due to
2844 * OOM kill or fatal signal. Since our only options are to
2845 * either fail the allocation or charge it to this cgroup, do
2846 * it as a temporary condition. But we can't fail. From a
2847 * kmem/slab perspective, the cache has already been selected,
2848 * by mem_cgroup_kmem_get_cache(), so it is too late to change
2851 * This condition will only trigger if the task entered
2852 * memcg_charge_kmem in a sane state, but was OOM-killed during
2853 * __mem_cgroup_try_charge() above. Tasks that were already
2854 * dying when the allocation triggers should have been already
2855 * directed to the root cgroup in memcontrol.h
2857 res_counter_charge_nofail(&memcg->res, size, &fail_res);
2858 if (do_swap_account)
2859 res_counter_charge_nofail(&memcg->memsw, size,
2863 res_counter_uncharge(&memcg->kmem, size);
2868 static void memcg_uncharge_kmem(struct mem_cgroup *memcg, u64 size)
2870 res_counter_uncharge(&memcg->res, size);
2871 if (do_swap_account)
2872 res_counter_uncharge(&memcg->memsw, size);
2875 if (res_counter_uncharge(&memcg->kmem, size))
2878 if (memcg_kmem_test_and_clear_dead(memcg))
2879 mem_cgroup_put(memcg);
2882 void memcg_cache_list_add(struct mem_cgroup *memcg, struct kmem_cache *cachep)
2887 mutex_lock(&memcg->slab_caches_mutex);
2888 list_add(&cachep->memcg_params->list, &memcg->memcg_slab_caches);
2889 mutex_unlock(&memcg->slab_caches_mutex);
2893 * helper for acessing a memcg's index. It will be used as an index in the
2894 * child cache array in kmem_cache, and also to derive its name. This function
2895 * will return -1 when this is not a kmem-limited memcg.
2897 int memcg_cache_id(struct mem_cgroup *memcg)
2899 return memcg ? memcg->kmemcg_id : -1;
2903 * This ends up being protected by the set_limit mutex, during normal
2904 * operation, because that is its main call site.
2906 * But when we create a new cache, we can call this as well if its parent
2907 * is kmem-limited. That will have to hold set_limit_mutex as well.
2909 int memcg_update_cache_sizes(struct mem_cgroup *memcg)
2913 num = ida_simple_get(&kmem_limited_groups,
2914 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2918 * After this point, kmem_accounted (that we test atomically in
2919 * the beginning of this conditional), is no longer 0. This
2920 * guarantees only one process will set the following boolean
2921 * to true. We don't need test_and_set because we're protected
2922 * by the set_limit_mutex anyway.
2924 memcg_kmem_set_activated(memcg);
2926 ret = memcg_update_all_caches(num+1);
2928 ida_simple_remove(&kmem_limited_groups, num);
2929 memcg_kmem_clear_activated(memcg);
2933 memcg->kmemcg_id = num;
2934 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
2935 mutex_init(&memcg->slab_caches_mutex);
2939 static size_t memcg_caches_array_size(int num_groups)
2942 if (num_groups <= 0)
2945 size = 2 * num_groups;
2946 if (size < MEMCG_CACHES_MIN_SIZE)
2947 size = MEMCG_CACHES_MIN_SIZE;
2948 else if (size > MEMCG_CACHES_MAX_SIZE)
2949 size = MEMCG_CACHES_MAX_SIZE;
2955 * We should update the current array size iff all caches updates succeed. This
2956 * can only be done from the slab side. The slab mutex needs to be held when
2959 void memcg_update_array_size(int num)
2961 if (num > memcg_limited_groups_array_size)
2962 memcg_limited_groups_array_size = memcg_caches_array_size(num);
2965 int memcg_update_cache_size(struct kmem_cache *s, int num_groups)
2967 struct memcg_cache_params *cur_params = s->memcg_params;
2969 VM_BUG_ON(s->memcg_params && !s->memcg_params->is_root_cache);
2971 if (num_groups > memcg_limited_groups_array_size) {
2973 ssize_t size = memcg_caches_array_size(num_groups);
2975 size *= sizeof(void *);
2976 size += sizeof(struct memcg_cache_params);
2978 s->memcg_params = kzalloc(size, GFP_KERNEL);
2979 if (!s->memcg_params) {
2980 s->memcg_params = cur_params;
2984 s->memcg_params->is_root_cache = true;
2987 * There is the chance it will be bigger than
2988 * memcg_limited_groups_array_size, if we failed an allocation
2989 * in a cache, in which case all caches updated before it, will
2990 * have a bigger array.
2992 * But if that is the case, the data after
2993 * memcg_limited_groups_array_size is certainly unused
2995 for (i = 0; i < memcg_limited_groups_array_size; i++) {
2996 if (!cur_params->memcg_caches[i])
2998 s->memcg_params->memcg_caches[i] =
2999 cur_params->memcg_caches[i];
3003 * Ideally, we would wait until all caches succeed, and only
3004 * then free the old one. But this is not worth the extra
3005 * pointer per-cache we'd have to have for this.
3007 * It is not a big deal if some caches are left with a size
3008 * bigger than the others. And all updates will reset this
3016 int memcg_register_cache(struct mem_cgroup *memcg, struct kmem_cache *s,
3017 struct kmem_cache *root_cache)
3019 size_t size = sizeof(struct memcg_cache_params);
3021 if (!memcg_kmem_enabled())
3025 size += memcg_limited_groups_array_size * sizeof(void *);
3027 s->memcg_params = kzalloc(size, GFP_KERNEL);
3028 if (!s->memcg_params)
3032 s->memcg_params->memcg = memcg;
3033 s->memcg_params->root_cache = root_cache;
3038 void memcg_release_cache(struct kmem_cache *s)
3040 struct kmem_cache *root;
3041 struct mem_cgroup *memcg;
3045 * This happens, for instance, when a root cache goes away before we
3048 if (!s->memcg_params)
3051 if (s->memcg_params->is_root_cache)
3054 memcg = s->memcg_params->memcg;
3055 id = memcg_cache_id(memcg);
3057 root = s->memcg_params->root_cache;
3058 root->memcg_params->memcg_caches[id] = NULL;
3059 mem_cgroup_put(memcg);
3061 mutex_lock(&memcg->slab_caches_mutex);
3062 list_del(&s->memcg_params->list);
3063 mutex_unlock(&memcg->slab_caches_mutex);
3066 kfree(s->memcg_params);
3070 * During the creation a new cache, we need to disable our accounting mechanism
3071 * altogether. This is true even if we are not creating, but rather just
3072 * enqueing new caches to be created.
3074 * This is because that process will trigger allocations; some visible, like
3075 * explicit kmallocs to auxiliary data structures, name strings and internal
3076 * cache structures; some well concealed, like INIT_WORK() that can allocate
3077 * objects during debug.
3079 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3080 * to it. This may not be a bounded recursion: since the first cache creation
3081 * failed to complete (waiting on the allocation), we'll just try to create the
3082 * cache again, failing at the same point.
3084 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3085 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3086 * inside the following two functions.
3088 static inline void memcg_stop_kmem_account(void)
3090 VM_BUG_ON(!current->mm);
3091 current->memcg_kmem_skip_account++;
3094 static inline void memcg_resume_kmem_account(void)
3096 VM_BUG_ON(!current->mm);
3097 current->memcg_kmem_skip_account--;
3100 static void kmem_cache_destroy_work_func(struct work_struct *w)
3102 struct kmem_cache *cachep;
3103 struct memcg_cache_params *p;
3105 p = container_of(w, struct memcg_cache_params, destroy);
3107 cachep = memcg_params_to_cache(p);
3110 * If we get down to 0 after shrink, we could delete right away.
3111 * However, memcg_release_pages() already puts us back in the workqueue
3112 * in that case. If we proceed deleting, we'll get a dangling
3113 * reference, and removing the object from the workqueue in that case
3114 * is unnecessary complication. We are not a fast path.
3116 * Note that this case is fundamentally different from racing with
3117 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3118 * kmem_cache_shrink, not only we would be reinserting a dead cache
3119 * into the queue, but doing so from inside the worker racing to
3122 * So if we aren't down to zero, we'll just schedule a worker and try
3125 if (atomic_read(&cachep->memcg_params->nr_pages) != 0) {
3126 kmem_cache_shrink(cachep);
3127 if (atomic_read(&cachep->memcg_params->nr_pages) == 0)
3130 kmem_cache_destroy(cachep);
3133 void mem_cgroup_destroy_cache(struct kmem_cache *cachep)
3135 if (!cachep->memcg_params->dead)
3139 * There are many ways in which we can get here.
3141 * We can get to a memory-pressure situation while the delayed work is
3142 * still pending to run. The vmscan shrinkers can then release all
3143 * cache memory and get us to destruction. If this is the case, we'll
3144 * be executed twice, which is a bug (the second time will execute over
3145 * bogus data). In this case, cancelling the work should be fine.
3147 * But we can also get here from the worker itself, if
3148 * kmem_cache_shrink is enough to shake all the remaining objects and
3149 * get the page count to 0. In this case, we'll deadlock if we try to
3150 * cancel the work (the worker runs with an internal lock held, which
3151 * is the same lock we would hold for cancel_work_sync().)
3153 * Since we can't possibly know who got us here, just refrain from
3154 * running if there is already work pending
3156 if (work_pending(&cachep->memcg_params->destroy))
3159 * We have to defer the actual destroying to a workqueue, because
3160 * we might currently be in a context that cannot sleep.
3162 schedule_work(&cachep->memcg_params->destroy);
3165 static char *memcg_cache_name(struct mem_cgroup *memcg, struct kmem_cache *s)
3168 struct dentry *dentry;
3171 dentry = rcu_dereference(memcg->css.cgroup->dentry);
3174 BUG_ON(dentry == NULL);
3176 name = kasprintf(GFP_KERNEL, "%s(%d:%s)", s->name,
3177 memcg_cache_id(memcg), dentry->d_name.name);
3182 static struct kmem_cache *kmem_cache_dup(struct mem_cgroup *memcg,
3183 struct kmem_cache *s)
3186 struct kmem_cache *new;
3188 name = memcg_cache_name(memcg, s);
3192 new = kmem_cache_create_memcg(memcg, name, s->object_size, s->align,
3193 (s->flags & ~SLAB_PANIC), s->ctor, s);
3196 new->allocflags |= __GFP_KMEMCG;
3203 * This lock protects updaters, not readers. We want readers to be as fast as
3204 * they can, and they will either see NULL or a valid cache value. Our model
3205 * allow them to see NULL, in which case the root memcg will be selected.
3207 * We need this lock because multiple allocations to the same cache from a non
3208 * will span more than one worker. Only one of them can create the cache.
3210 static DEFINE_MUTEX(memcg_cache_mutex);
3211 static struct kmem_cache *memcg_create_kmem_cache(struct mem_cgroup *memcg,
3212 struct kmem_cache *cachep)
3214 struct kmem_cache *new_cachep;
3217 BUG_ON(!memcg_can_account_kmem(memcg));
3219 idx = memcg_cache_id(memcg);
3221 mutex_lock(&memcg_cache_mutex);
3222 new_cachep = cachep->memcg_params->memcg_caches[idx];
3226 new_cachep = kmem_cache_dup(memcg, cachep);
3227 if (new_cachep == NULL) {
3228 new_cachep = cachep;
3232 mem_cgroup_get(memcg);
3233 atomic_set(&new_cachep->memcg_params->nr_pages , 0);
3235 cachep->memcg_params->memcg_caches[idx] = new_cachep;
3237 * the readers won't lock, make sure everybody sees the updated value,
3238 * so they won't put stuff in the queue again for no reason
3242 mutex_unlock(&memcg_cache_mutex);
3246 void kmem_cache_destroy_memcg_children(struct kmem_cache *s)
3248 struct kmem_cache *c;
3251 if (!s->memcg_params)
3253 if (!s->memcg_params->is_root_cache)
3257 * If the cache is being destroyed, we trust that there is no one else
3258 * requesting objects from it. Even if there are, the sanity checks in
3259 * kmem_cache_destroy should caught this ill-case.
3261 * Still, we don't want anyone else freeing memcg_caches under our
3262 * noses, which can happen if a new memcg comes to life. As usual,
3263 * we'll take the set_limit_mutex to protect ourselves against this.
3265 mutex_lock(&set_limit_mutex);
3266 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3267 c = s->memcg_params->memcg_caches[i];
3272 * We will now manually delete the caches, so to avoid races
3273 * we need to cancel all pending destruction workers and
3274 * proceed with destruction ourselves.
3276 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3277 * and that could spawn the workers again: it is likely that
3278 * the cache still have active pages until this very moment.
3279 * This would lead us back to mem_cgroup_destroy_cache.
3281 * But that will not execute at all if the "dead" flag is not
3282 * set, so flip it down to guarantee we are in control.
3284 c->memcg_params->dead = false;
3285 cancel_work_sync(&c->memcg_params->destroy);
3286 kmem_cache_destroy(c);
3288 mutex_unlock(&set_limit_mutex);
3291 struct create_work {
3292 struct mem_cgroup *memcg;
3293 struct kmem_cache *cachep;
3294 struct work_struct work;
3297 static void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3299 struct kmem_cache *cachep;
3300 struct memcg_cache_params *params;
3302 if (!memcg_kmem_is_active(memcg))
3305 mutex_lock(&memcg->slab_caches_mutex);
3306 list_for_each_entry(params, &memcg->memcg_slab_caches, list) {
3307 cachep = memcg_params_to_cache(params);
3308 cachep->memcg_params->dead = true;
3309 INIT_WORK(&cachep->memcg_params->destroy,
3310 kmem_cache_destroy_work_func);
3311 schedule_work(&cachep->memcg_params->destroy);
3313 mutex_unlock(&memcg->slab_caches_mutex);
3316 static void memcg_create_cache_work_func(struct work_struct *w)
3318 struct create_work *cw;
3320 cw = container_of(w, struct create_work, work);
3321 memcg_create_kmem_cache(cw->memcg, cw->cachep);
3322 /* Drop the reference gotten when we enqueued. */
3323 css_put(&cw->memcg->css);
3328 * Enqueue the creation of a per-memcg kmem_cache.
3329 * Called with rcu_read_lock.
3331 static void __memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3332 struct kmem_cache *cachep)
3334 struct create_work *cw;
3336 cw = kmalloc(sizeof(struct create_work), GFP_NOWAIT);
3340 /* The corresponding put will be done in the workqueue. */
3341 if (!css_tryget(&memcg->css)) {
3347 cw->cachep = cachep;
3349 INIT_WORK(&cw->work, memcg_create_cache_work_func);
3350 schedule_work(&cw->work);
3353 static void memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3354 struct kmem_cache *cachep)
3357 * We need to stop accounting when we kmalloc, because if the
3358 * corresponding kmalloc cache is not yet created, the first allocation
3359 * in __memcg_create_cache_enqueue will recurse.
3361 * However, it is better to enclose the whole function. Depending on
3362 * the debugging options enabled, INIT_WORK(), for instance, can
3363 * trigger an allocation. This too, will make us recurse. Because at
3364 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3365 * the safest choice is to do it like this, wrapping the whole function.
3367 memcg_stop_kmem_account();
3368 __memcg_create_cache_enqueue(memcg, cachep);
3369 memcg_resume_kmem_account();
3372 * Return the kmem_cache we're supposed to use for a slab allocation.
3373 * We try to use the current memcg's version of the cache.
3375 * If the cache does not exist yet, if we are the first user of it,
3376 * we either create it immediately, if possible, or create it asynchronously
3378 * In the latter case, we will let the current allocation go through with
3379 * the original cache.
3381 * Can't be called in interrupt context or from kernel threads.
3382 * This function needs to be called with rcu_read_lock() held.
3384 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
3387 struct mem_cgroup *memcg;
3390 VM_BUG_ON(!cachep->memcg_params);
3391 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
3393 if (!current->mm || current->memcg_kmem_skip_account)
3397 memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
3400 if (!memcg_can_account_kmem(memcg))
3403 idx = memcg_cache_id(memcg);
3406 * barrier to mare sure we're always seeing the up to date value. The
3407 * code updating memcg_caches will issue a write barrier to match this.
3409 read_barrier_depends();
3410 if (unlikely(cachep->memcg_params->memcg_caches[idx] == NULL)) {
3412 * If we are in a safe context (can wait, and not in interrupt
3413 * context), we could be be predictable and return right away.
3414 * This would guarantee that the allocation being performed
3415 * already belongs in the new cache.
3417 * However, there are some clashes that can arrive from locking.
3418 * For instance, because we acquire the slab_mutex while doing
3419 * kmem_cache_dup, this means no further allocation could happen
3420 * with the slab_mutex held.
3422 * Also, because cache creation issue get_online_cpus(), this
3423 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3424 * that ends up reversed during cpu hotplug. (cpuset allocates
3425 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3426 * better to defer everything.
3428 memcg_create_cache_enqueue(memcg, cachep);
3432 return cachep->memcg_params->memcg_caches[idx];
3434 EXPORT_SYMBOL(__memcg_kmem_get_cache);
3437 * We need to verify if the allocation against current->mm->owner's memcg is
3438 * possible for the given order. But the page is not allocated yet, so we'll
3439 * need a further commit step to do the final arrangements.
3441 * It is possible for the task to switch cgroups in this mean time, so at
3442 * commit time, we can't rely on task conversion any longer. We'll then use
3443 * the handle argument to return to the caller which cgroup we should commit
3444 * against. We could also return the memcg directly and avoid the pointer
3445 * passing, but a boolean return value gives better semantics considering
3446 * the compiled-out case as well.
3448 * Returning true means the allocation is possible.
3451 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3453 struct mem_cgroup *memcg;
3457 memcg = try_get_mem_cgroup_from_mm(current->mm);
3460 * very rare case described in mem_cgroup_from_task. Unfortunately there
3461 * isn't much we can do without complicating this too much, and it would
3462 * be gfp-dependent anyway. Just let it go
3464 if (unlikely(!memcg))
3467 if (!memcg_can_account_kmem(memcg)) {
3468 css_put(&memcg->css);
3472 ret = memcg_charge_kmem(memcg, gfp, PAGE_SIZE << order);
3476 css_put(&memcg->css);
3480 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3483 struct page_cgroup *pc;
3485 VM_BUG_ON(mem_cgroup_is_root(memcg));
3487 /* The page allocation failed. Revert */
3489 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3493 pc = lookup_page_cgroup(page);
3494 lock_page_cgroup(pc);
3495 pc->mem_cgroup = memcg;
3496 SetPageCgroupUsed(pc);
3497 unlock_page_cgroup(pc);
3500 void __memcg_kmem_uncharge_pages(struct page *page, int order)
3502 struct mem_cgroup *memcg = NULL;
3503 struct page_cgroup *pc;
3506 pc = lookup_page_cgroup(page);
3508 * Fast unlocked return. Theoretically might have changed, have to
3509 * check again after locking.
3511 if (!PageCgroupUsed(pc))
3514 lock_page_cgroup(pc);
3515 if (PageCgroupUsed(pc)) {
3516 memcg = pc->mem_cgroup;
3517 ClearPageCgroupUsed(pc);
3519 unlock_page_cgroup(pc);
3522 * We trust that only if there is a memcg associated with the page, it
3523 * is a valid allocation
3528 VM_BUG_ON(mem_cgroup_is_root(memcg));
3529 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3532 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3535 #endif /* CONFIG_MEMCG_KMEM */
3537 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3539 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3541 * Because tail pages are not marked as "used", set it. We're under
3542 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3543 * charge/uncharge will be never happen and move_account() is done under
3544 * compound_lock(), so we don't have to take care of races.
3546 void mem_cgroup_split_huge_fixup(struct page *head)
3548 struct page_cgroup *head_pc = lookup_page_cgroup(head);
3549 struct page_cgroup *pc;
3552 if (mem_cgroup_disabled())
3554 for (i = 1; i < HPAGE_PMD_NR; i++) {
3556 pc->mem_cgroup = head_pc->mem_cgroup;
3557 smp_wmb();/* see __commit_charge() */
3558 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
3561 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3564 * mem_cgroup_move_account - move account of the page
3566 * @nr_pages: number of regular pages (>1 for huge pages)
3567 * @pc: page_cgroup of the page.
3568 * @from: mem_cgroup which the page is moved from.
3569 * @to: mem_cgroup which the page is moved to. @from != @to.
3571 * The caller must confirm following.
3572 * - page is not on LRU (isolate_page() is useful.)
3573 * - compound_lock is held when nr_pages > 1
3575 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3578 static int mem_cgroup_move_account(struct page *page,
3579 unsigned int nr_pages,
3580 struct page_cgroup *pc,
3581 struct mem_cgroup *from,
3582 struct mem_cgroup *to)
3584 unsigned long flags;
3586 bool anon = PageAnon(page);
3588 VM_BUG_ON(from == to);
3589 VM_BUG_ON(PageLRU(page));
3591 * The page is isolated from LRU. So, collapse function
3592 * will not handle this page. But page splitting can happen.
3593 * Do this check under compound_page_lock(). The caller should
3597 if (nr_pages > 1 && !PageTransHuge(page))
3600 lock_page_cgroup(pc);
3603 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3606 move_lock_mem_cgroup(from, &flags);
3608 if (!anon && page_mapped(page)) {
3609 /* Update mapped_file data for mem_cgroup */
3611 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
3612 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
3615 mem_cgroup_charge_statistics(from, anon, -nr_pages);
3617 /* caller should have done css_get */
3618 pc->mem_cgroup = to;
3619 mem_cgroup_charge_statistics(to, anon, nr_pages);
3620 move_unlock_mem_cgroup(from, &flags);
3623 unlock_page_cgroup(pc);
3627 memcg_check_events(to, page);
3628 memcg_check_events(from, page);
3634 * mem_cgroup_move_parent - moves page to the parent group
3635 * @page: the page to move
3636 * @pc: page_cgroup of the page
3637 * @child: page's cgroup
3639 * move charges to its parent or the root cgroup if the group has no
3640 * parent (aka use_hierarchy==0).
3641 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3642 * mem_cgroup_move_account fails) the failure is always temporary and
3643 * it signals a race with a page removal/uncharge or migration. In the
3644 * first case the page is on the way out and it will vanish from the LRU
3645 * on the next attempt and the call should be retried later.
3646 * Isolation from the LRU fails only if page has been isolated from
3647 * the LRU since we looked at it and that usually means either global
3648 * reclaim or migration going on. The page will either get back to the
3650 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3651 * (!PageCgroupUsed) or moved to a different group. The page will
3652 * disappear in the next attempt.
3654 static int mem_cgroup_move_parent(struct page *page,
3655 struct page_cgroup *pc,
3656 struct mem_cgroup *child)
3658 struct mem_cgroup *parent;
3659 unsigned int nr_pages;
3660 unsigned long uninitialized_var(flags);
3663 VM_BUG_ON(mem_cgroup_is_root(child));
3666 if (!get_page_unless_zero(page))
3668 if (isolate_lru_page(page))
3671 nr_pages = hpage_nr_pages(page);
3673 parent = parent_mem_cgroup(child);
3675 * If no parent, move charges to root cgroup.
3678 parent = root_mem_cgroup;
3681 VM_BUG_ON(!PageTransHuge(page));
3682 flags = compound_lock_irqsave(page);
3685 ret = mem_cgroup_move_account(page, nr_pages,
3688 __mem_cgroup_cancel_local_charge(child, nr_pages);
3691 compound_unlock_irqrestore(page, flags);
3692 putback_lru_page(page);
3700 * Charge the memory controller for page usage.
3702 * 0 if the charge was successful
3703 * < 0 if the cgroup is over its limit
3705 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
3706 gfp_t gfp_mask, enum charge_type ctype)
3708 struct mem_cgroup *memcg = NULL;
3709 unsigned int nr_pages = 1;
3713 if (PageTransHuge(page)) {
3714 nr_pages <<= compound_order(page);
3715 VM_BUG_ON(!PageTransHuge(page));
3717 * Never OOM-kill a process for a huge page. The
3718 * fault handler will fall back to regular pages.
3723 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
3726 __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false);
3730 int mem_cgroup_newpage_charge(struct page *page,
3731 struct mm_struct *mm, gfp_t gfp_mask)
3733 if (mem_cgroup_disabled())
3735 VM_BUG_ON(page_mapped(page));
3736 VM_BUG_ON(page->mapping && !PageAnon(page));
3738 return mem_cgroup_charge_common(page, mm, gfp_mask,
3739 MEM_CGROUP_CHARGE_TYPE_ANON);
3743 * While swap-in, try_charge -> commit or cancel, the page is locked.
3744 * And when try_charge() successfully returns, one refcnt to memcg without
3745 * struct page_cgroup is acquired. This refcnt will be consumed by
3746 * "commit()" or removed by "cancel()"
3748 static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm,
3751 struct mem_cgroup **memcgp)
3753 struct mem_cgroup *memcg;
3754 struct page_cgroup *pc;
3757 pc = lookup_page_cgroup(page);
3759 * Every swap fault against a single page tries to charge the
3760 * page, bail as early as possible. shmem_unuse() encounters
3761 * already charged pages, too. The USED bit is protected by
3762 * the page lock, which serializes swap cache removal, which
3763 * in turn serializes uncharging.
3765 if (PageCgroupUsed(pc))
3767 if (!do_swap_account)
3769 memcg = try_get_mem_cgroup_from_page(page);
3773 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
3774 css_put(&memcg->css);
3779 ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
3785 int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page,
3786 gfp_t gfp_mask, struct mem_cgroup **memcgp)
3789 if (mem_cgroup_disabled())
3792 * A racing thread's fault, or swapoff, may have already
3793 * updated the pte, and even removed page from swap cache: in
3794 * those cases unuse_pte()'s pte_same() test will fail; but
3795 * there's also a KSM case which does need to charge the page.
3797 if (!PageSwapCache(page)) {
3800 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, memcgp, true);
3805 return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp);
3808 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
3810 if (mem_cgroup_disabled())
3814 __mem_cgroup_cancel_charge(memcg, 1);
3818 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
3819 enum charge_type ctype)
3821 if (mem_cgroup_disabled())
3826 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
3828 * Now swap is on-memory. This means this page may be
3829 * counted both as mem and swap....double count.
3830 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
3831 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
3832 * may call delete_from_swap_cache() before reach here.
3834 if (do_swap_account && PageSwapCache(page)) {
3835 swp_entry_t ent = {.val = page_private(page)};
3836 mem_cgroup_uncharge_swap(ent);
3840 void mem_cgroup_commit_charge_swapin(struct page *page,
3841 struct mem_cgroup *memcg)
3843 __mem_cgroup_commit_charge_swapin(page, memcg,
3844 MEM_CGROUP_CHARGE_TYPE_ANON);
3847 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
3850 struct mem_cgroup *memcg = NULL;
3851 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
3854 if (mem_cgroup_disabled())
3856 if (PageCompound(page))
3859 if (!PageSwapCache(page))
3860 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
3861 else { /* page is swapcache/shmem */
3862 ret = __mem_cgroup_try_charge_swapin(mm, page,
3865 __mem_cgroup_commit_charge_swapin(page, memcg, type);
3870 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
3871 unsigned int nr_pages,
3872 const enum charge_type ctype)
3874 struct memcg_batch_info *batch = NULL;
3875 bool uncharge_memsw = true;
3877 /* If swapout, usage of swap doesn't decrease */
3878 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
3879 uncharge_memsw = false;
3881 batch = ¤t->memcg_batch;
3883 * In usual, we do css_get() when we remember memcg pointer.
3884 * But in this case, we keep res->usage until end of a series of
3885 * uncharges. Then, it's ok to ignore memcg's refcnt.
3888 batch->memcg = memcg;
3890 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
3891 * In those cases, all pages freed continuously can be expected to be in
3892 * the same cgroup and we have chance to coalesce uncharges.
3893 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
3894 * because we want to do uncharge as soon as possible.
3897 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
3898 goto direct_uncharge;
3901 goto direct_uncharge;
3904 * In typical case, batch->memcg == mem. This means we can
3905 * merge a series of uncharges to an uncharge of res_counter.
3906 * If not, we uncharge res_counter ony by one.
3908 if (batch->memcg != memcg)
3909 goto direct_uncharge;
3910 /* remember freed charge and uncharge it later */
3913 batch->memsw_nr_pages++;
3916 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
3918 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
3919 if (unlikely(batch->memcg != memcg))
3920 memcg_oom_recover(memcg);
3924 * uncharge if !page_mapped(page)
3926 static struct mem_cgroup *
3927 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
3930 struct mem_cgroup *memcg = NULL;
3931 unsigned int nr_pages = 1;
3932 struct page_cgroup *pc;
3935 if (mem_cgroup_disabled())
3938 VM_BUG_ON(PageSwapCache(page));
3940 if (PageTransHuge(page)) {
3941 nr_pages <<= compound_order(page);
3942 VM_BUG_ON(!PageTransHuge(page));
3945 * Check if our page_cgroup is valid
3947 pc = lookup_page_cgroup(page);
3948 if (unlikely(!PageCgroupUsed(pc)))
3951 lock_page_cgroup(pc);
3953 memcg = pc->mem_cgroup;
3955 if (!PageCgroupUsed(pc))
3958 anon = PageAnon(page);
3961 case MEM_CGROUP_CHARGE_TYPE_ANON:
3963 * Generally PageAnon tells if it's the anon statistics to be
3964 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
3965 * used before page reached the stage of being marked PageAnon.
3969 case MEM_CGROUP_CHARGE_TYPE_DROP:
3970 /* See mem_cgroup_prepare_migration() */
3971 if (page_mapped(page))
3974 * Pages under migration may not be uncharged. But
3975 * end_migration() /must/ be the one uncharging the
3976 * unused post-migration page and so it has to call
3977 * here with the migration bit still set. See the
3978 * res_counter handling below.
3980 if (!end_migration && PageCgroupMigration(pc))
3983 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
3984 if (!PageAnon(page)) { /* Shared memory */
3985 if (page->mapping && !page_is_file_cache(page))
3987 } else if (page_mapped(page)) /* Anon */
3994 mem_cgroup_charge_statistics(memcg, anon, -nr_pages);
3996 ClearPageCgroupUsed(pc);
3998 * pc->mem_cgroup is not cleared here. It will be accessed when it's
3999 * freed from LRU. This is safe because uncharged page is expected not
4000 * to be reused (freed soon). Exception is SwapCache, it's handled by
4001 * special functions.
4004 unlock_page_cgroup(pc);
4006 * even after unlock, we have memcg->res.usage here and this memcg
4007 * will never be freed.
4009 memcg_check_events(memcg, page);
4010 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
4011 mem_cgroup_swap_statistics(memcg, true);
4012 mem_cgroup_get(memcg);
4015 * Migration does not charge the res_counter for the
4016 * replacement page, so leave it alone when phasing out the
4017 * page that is unused after the migration.
4019 if (!end_migration && !mem_cgroup_is_root(memcg))
4020 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
4025 unlock_page_cgroup(pc);
4029 void mem_cgroup_uncharge_page(struct page *page)
4032 if (page_mapped(page))
4034 VM_BUG_ON(page->mapping && !PageAnon(page));
4035 if (PageSwapCache(page))
4037 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
4040 void mem_cgroup_uncharge_cache_page(struct page *page)
4042 VM_BUG_ON(page_mapped(page));
4043 VM_BUG_ON(page->mapping);
4044 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
4048 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4049 * In that cases, pages are freed continuously and we can expect pages
4050 * are in the same memcg. All these calls itself limits the number of
4051 * pages freed at once, then uncharge_start/end() is called properly.
4052 * This may be called prural(2) times in a context,
4055 void mem_cgroup_uncharge_start(void)
4057 current->memcg_batch.do_batch++;
4058 /* We can do nest. */
4059 if (current->memcg_batch.do_batch == 1) {
4060 current->memcg_batch.memcg = NULL;
4061 current->memcg_batch.nr_pages = 0;
4062 current->memcg_batch.memsw_nr_pages = 0;
4066 void mem_cgroup_uncharge_end(void)
4068 struct memcg_batch_info *batch = ¤t->memcg_batch;
4070 if (!batch->do_batch)
4074 if (batch->do_batch) /* If stacked, do nothing. */
4080 * This "batch->memcg" is valid without any css_get/put etc...
4081 * bacause we hide charges behind us.
4083 if (batch->nr_pages)
4084 res_counter_uncharge(&batch->memcg->res,
4085 batch->nr_pages * PAGE_SIZE);
4086 if (batch->memsw_nr_pages)
4087 res_counter_uncharge(&batch->memcg->memsw,
4088 batch->memsw_nr_pages * PAGE_SIZE);
4089 memcg_oom_recover(batch->memcg);
4090 /* forget this pointer (for sanity check) */
4091 batch->memcg = NULL;
4096 * called after __delete_from_swap_cache() and drop "page" account.
4097 * memcg information is recorded to swap_cgroup of "ent"
4100 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
4102 struct mem_cgroup *memcg;
4103 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
4105 if (!swapout) /* this was a swap cache but the swap is unused ! */
4106 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
4108 memcg = __mem_cgroup_uncharge_common(page, ctype, false);
4111 * record memcg information, if swapout && memcg != NULL,
4112 * mem_cgroup_get() was called in uncharge().
4114 if (do_swap_account && swapout && memcg)
4115 swap_cgroup_record(ent, css_id(&memcg->css));
4119 #ifdef CONFIG_MEMCG_SWAP
4121 * called from swap_entry_free(). remove record in swap_cgroup and
4122 * uncharge "memsw" account.
4124 void mem_cgroup_uncharge_swap(swp_entry_t ent)
4126 struct mem_cgroup *memcg;
4129 if (!do_swap_account)
4132 id = swap_cgroup_record(ent, 0);
4134 memcg = mem_cgroup_lookup(id);
4137 * We uncharge this because swap is freed.
4138 * This memcg can be obsolete one. We avoid calling css_tryget
4140 if (!mem_cgroup_is_root(memcg))
4141 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
4142 mem_cgroup_swap_statistics(memcg, false);
4143 mem_cgroup_put(memcg);
4149 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4150 * @entry: swap entry to be moved
4151 * @from: mem_cgroup which the entry is moved from
4152 * @to: mem_cgroup which the entry is moved to
4154 * It succeeds only when the swap_cgroup's record for this entry is the same
4155 * as the mem_cgroup's id of @from.
4157 * Returns 0 on success, -EINVAL on failure.
4159 * The caller must have charged to @to, IOW, called res_counter_charge() about
4160 * both res and memsw, and called css_get().
4162 static int mem_cgroup_move_swap_account(swp_entry_t entry,
4163 struct mem_cgroup *from, struct mem_cgroup *to)
4165 unsigned short old_id, new_id;
4167 old_id = css_id(&from->css);
4168 new_id = css_id(&to->css);
4170 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
4171 mem_cgroup_swap_statistics(from, false);
4172 mem_cgroup_swap_statistics(to, true);
4174 * This function is only called from task migration context now.
4175 * It postpones res_counter and refcount handling till the end
4176 * of task migration(mem_cgroup_clear_mc()) for performance
4177 * improvement. But we cannot postpone mem_cgroup_get(to)
4178 * because if the process that has been moved to @to does
4179 * swap-in, the refcount of @to might be decreased to 0.
4187 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
4188 struct mem_cgroup *from, struct mem_cgroup *to)
4195 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4198 void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
4199 struct mem_cgroup **memcgp)
4201 struct mem_cgroup *memcg = NULL;
4202 struct page_cgroup *pc;
4203 enum charge_type ctype;
4207 VM_BUG_ON(PageTransHuge(page));
4208 if (mem_cgroup_disabled())
4211 pc = lookup_page_cgroup(page);
4212 lock_page_cgroup(pc);
4213 if (PageCgroupUsed(pc)) {
4214 memcg = pc->mem_cgroup;
4215 css_get(&memcg->css);
4217 * At migrating an anonymous page, its mapcount goes down
4218 * to 0 and uncharge() will be called. But, even if it's fully
4219 * unmapped, migration may fail and this page has to be
4220 * charged again. We set MIGRATION flag here and delay uncharge
4221 * until end_migration() is called
4223 * Corner Case Thinking
4225 * When the old page was mapped as Anon and it's unmap-and-freed
4226 * while migration was ongoing.
4227 * If unmap finds the old page, uncharge() of it will be delayed
4228 * until end_migration(). If unmap finds a new page, it's
4229 * uncharged when it make mapcount to be 1->0. If unmap code
4230 * finds swap_migration_entry, the new page will not be mapped
4231 * and end_migration() will find it(mapcount==0).
4234 * When the old page was mapped but migraion fails, the kernel
4235 * remaps it. A charge for it is kept by MIGRATION flag even
4236 * if mapcount goes down to 0. We can do remap successfully
4237 * without charging it again.
4240 * The "old" page is under lock_page() until the end of
4241 * migration, so, the old page itself will not be swapped-out.
4242 * If the new page is swapped out before end_migraton, our
4243 * hook to usual swap-out path will catch the event.
4246 SetPageCgroupMigration(pc);
4248 unlock_page_cgroup(pc);
4250 * If the page is not charged at this point,
4258 * We charge new page before it's used/mapped. So, even if unlock_page()
4259 * is called before end_migration, we can catch all events on this new
4260 * page. In the case new page is migrated but not remapped, new page's
4261 * mapcount will be finally 0 and we call uncharge in end_migration().
4264 ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
4266 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
4268 * The page is committed to the memcg, but it's not actually
4269 * charged to the res_counter since we plan on replacing the
4270 * old one and only one page is going to be left afterwards.
4272 __mem_cgroup_commit_charge(memcg, newpage, 1, ctype, false);
4275 /* remove redundant charge if migration failed*/
4276 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
4277 struct page *oldpage, struct page *newpage, bool migration_ok)
4279 struct page *used, *unused;
4280 struct page_cgroup *pc;
4286 if (!migration_ok) {
4293 anon = PageAnon(used);
4294 __mem_cgroup_uncharge_common(unused,
4295 anon ? MEM_CGROUP_CHARGE_TYPE_ANON
4296 : MEM_CGROUP_CHARGE_TYPE_CACHE,
4298 css_put(&memcg->css);
4300 * We disallowed uncharge of pages under migration because mapcount
4301 * of the page goes down to zero, temporarly.
4302 * Clear the flag and check the page should be charged.
4304 pc = lookup_page_cgroup(oldpage);
4305 lock_page_cgroup(pc);
4306 ClearPageCgroupMigration(pc);
4307 unlock_page_cgroup(pc);
4310 * If a page is a file cache, radix-tree replacement is very atomic
4311 * and we can skip this check. When it was an Anon page, its mapcount
4312 * goes down to 0. But because we added MIGRATION flage, it's not
4313 * uncharged yet. There are several case but page->mapcount check
4314 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4315 * check. (see prepare_charge() also)
4318 mem_cgroup_uncharge_page(used);
4322 * At replace page cache, newpage is not under any memcg but it's on
4323 * LRU. So, this function doesn't touch res_counter but handles LRU
4324 * in correct way. Both pages are locked so we cannot race with uncharge.
4326 void mem_cgroup_replace_page_cache(struct page *oldpage,
4327 struct page *newpage)
4329 struct mem_cgroup *memcg = NULL;
4330 struct page_cgroup *pc;
4331 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4333 if (mem_cgroup_disabled())
4336 pc = lookup_page_cgroup(oldpage);
4337 /* fix accounting on old pages */
4338 lock_page_cgroup(pc);
4339 if (PageCgroupUsed(pc)) {
4340 memcg = pc->mem_cgroup;
4341 mem_cgroup_charge_statistics(memcg, false, -1);
4342 ClearPageCgroupUsed(pc);
4344 unlock_page_cgroup(pc);
4347 * When called from shmem_replace_page(), in some cases the
4348 * oldpage has already been charged, and in some cases not.
4353 * Even if newpage->mapping was NULL before starting replacement,
4354 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4355 * LRU while we overwrite pc->mem_cgroup.
4357 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
4360 #ifdef CONFIG_DEBUG_VM
4361 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
4363 struct page_cgroup *pc;
4365 pc = lookup_page_cgroup(page);
4367 * Can be NULL while feeding pages into the page allocator for
4368 * the first time, i.e. during boot or memory hotplug;
4369 * or when mem_cgroup_disabled().
4371 if (likely(pc) && PageCgroupUsed(pc))
4376 bool mem_cgroup_bad_page_check(struct page *page)
4378 if (mem_cgroup_disabled())
4381 return lookup_page_cgroup_used(page) != NULL;
4384 void mem_cgroup_print_bad_page(struct page *page)
4386 struct page_cgroup *pc;
4388 pc = lookup_page_cgroup_used(page);
4390 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4391 pc, pc->flags, pc->mem_cgroup);
4396 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
4397 unsigned long long val)
4400 u64 memswlimit, memlimit;
4402 int children = mem_cgroup_count_children(memcg);
4403 u64 curusage, oldusage;
4407 * For keeping hierarchical_reclaim simple, how long we should retry
4408 * is depends on callers. We set our retry-count to be function
4409 * of # of children which we should visit in this loop.
4411 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
4413 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4416 while (retry_count) {
4417 if (signal_pending(current)) {
4422 * Rather than hide all in some function, I do this in
4423 * open coded manner. You see what this really does.
4424 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4426 mutex_lock(&set_limit_mutex);
4427 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4428 if (memswlimit < val) {
4430 mutex_unlock(&set_limit_mutex);
4434 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4438 ret = res_counter_set_limit(&memcg->res, val);
4440 if (memswlimit == val)
4441 memcg->memsw_is_minimum = true;
4443 memcg->memsw_is_minimum = false;
4445 mutex_unlock(&set_limit_mutex);
4450 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4451 MEM_CGROUP_RECLAIM_SHRINK);
4452 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4453 /* Usage is reduced ? */
4454 if (curusage >= oldusage)
4457 oldusage = curusage;
4459 if (!ret && enlarge)
4460 memcg_oom_recover(memcg);
4465 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
4466 unsigned long long val)
4469 u64 memlimit, memswlimit, oldusage, curusage;
4470 int children = mem_cgroup_count_children(memcg);
4474 /* see mem_cgroup_resize_res_limit */
4475 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
4476 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4477 while (retry_count) {
4478 if (signal_pending(current)) {
4483 * Rather than hide all in some function, I do this in
4484 * open coded manner. You see what this really does.
4485 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4487 mutex_lock(&set_limit_mutex);
4488 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4489 if (memlimit > val) {
4491 mutex_unlock(&set_limit_mutex);
4494 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4495 if (memswlimit < val)
4497 ret = res_counter_set_limit(&memcg->memsw, val);
4499 if (memlimit == val)
4500 memcg->memsw_is_minimum = true;
4502 memcg->memsw_is_minimum = false;
4504 mutex_unlock(&set_limit_mutex);
4509 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4510 MEM_CGROUP_RECLAIM_NOSWAP |
4511 MEM_CGROUP_RECLAIM_SHRINK);
4512 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4513 /* Usage is reduced ? */
4514 if (curusage >= oldusage)
4517 oldusage = curusage;
4519 if (!ret && enlarge)
4520 memcg_oom_recover(memcg);
4524 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
4526 unsigned long *total_scanned)
4528 unsigned long nr_reclaimed = 0;
4529 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
4530 unsigned long reclaimed;
4532 struct mem_cgroup_tree_per_zone *mctz;
4533 unsigned long long excess;
4534 unsigned long nr_scanned;
4539 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
4541 * This loop can run a while, specially if mem_cgroup's continuously
4542 * keep exceeding their soft limit and putting the system under
4549 mz = mem_cgroup_largest_soft_limit_node(mctz);
4554 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
4555 gfp_mask, &nr_scanned);
4556 nr_reclaimed += reclaimed;
4557 *total_scanned += nr_scanned;
4558 spin_lock(&mctz->lock);
4561 * If we failed to reclaim anything from this memory cgroup
4562 * it is time to move on to the next cgroup
4568 * Loop until we find yet another one.
4570 * By the time we get the soft_limit lock
4571 * again, someone might have aded the
4572 * group back on the RB tree. Iterate to
4573 * make sure we get a different mem.
4574 * mem_cgroup_largest_soft_limit_node returns
4575 * NULL if no other cgroup is present on
4579 __mem_cgroup_largest_soft_limit_node(mctz);
4581 css_put(&next_mz->memcg->css);
4582 else /* next_mz == NULL or other memcg */
4586 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
4587 excess = res_counter_soft_limit_excess(&mz->memcg->res);
4589 * One school of thought says that we should not add
4590 * back the node to the tree if reclaim returns 0.
4591 * But our reclaim could return 0, simply because due
4592 * to priority we are exposing a smaller subset of
4593 * memory to reclaim from. Consider this as a longer
4596 /* If excess == 0, no tree ops */
4597 __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess);
4598 spin_unlock(&mctz->lock);
4599 css_put(&mz->memcg->css);
4602 * Could not reclaim anything and there are no more
4603 * mem cgroups to try or we seem to be looping without
4604 * reclaiming anything.
4606 if (!nr_reclaimed &&
4608 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
4610 } while (!nr_reclaimed);
4612 css_put(&next_mz->memcg->css);
4613 return nr_reclaimed;
4617 * mem_cgroup_force_empty_list - clears LRU of a group
4618 * @memcg: group to clear
4621 * @lru: lru to to clear
4623 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4624 * reclaim the pages page themselves - pages are moved to the parent (or root)
4627 static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
4628 int node, int zid, enum lru_list lru)
4630 struct lruvec *lruvec;
4631 unsigned long flags;
4632 struct list_head *list;
4636 zone = &NODE_DATA(node)->node_zones[zid];
4637 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
4638 list = &lruvec->lists[lru];
4642 struct page_cgroup *pc;
4645 spin_lock_irqsave(&zone->lru_lock, flags);
4646 if (list_empty(list)) {
4647 spin_unlock_irqrestore(&zone->lru_lock, flags);
4650 page = list_entry(list->prev, struct page, lru);
4652 list_move(&page->lru, list);
4654 spin_unlock_irqrestore(&zone->lru_lock, flags);
4657 spin_unlock_irqrestore(&zone->lru_lock, flags);
4659 pc = lookup_page_cgroup(page);
4661 if (mem_cgroup_move_parent(page, pc, memcg)) {
4662 /* found lock contention or "pc" is obsolete. */
4667 } while (!list_empty(list));
4671 * make mem_cgroup's charge to be 0 if there is no task by moving
4672 * all the charges and pages to the parent.
4673 * This enables deleting this mem_cgroup.
4675 * Caller is responsible for holding css reference on the memcg.
4677 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
4683 /* This is for making all *used* pages to be on LRU. */
4684 lru_add_drain_all();
4685 drain_all_stock_sync(memcg);
4686 mem_cgroup_start_move(memcg);
4687 for_each_node_state(node, N_HIGH_MEMORY) {
4688 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4691 mem_cgroup_force_empty_list(memcg,
4696 mem_cgroup_end_move(memcg);
4697 memcg_oom_recover(memcg);
4701 * Kernel memory may not necessarily be trackable to a specific
4702 * process. So they are not migrated, and therefore we can't
4703 * expect their value to drop to 0 here.
4704 * Having res filled up with kmem only is enough.
4706 * This is a safety check because mem_cgroup_force_empty_list
4707 * could have raced with mem_cgroup_replace_page_cache callers
4708 * so the lru seemed empty but the page could have been added
4709 * right after the check. RES_USAGE should be safe as we always
4710 * charge before adding to the LRU.
4712 usage = res_counter_read_u64(&memcg->res, RES_USAGE) -
4713 res_counter_read_u64(&memcg->kmem, RES_USAGE);
4714 } while (usage > 0);
4718 * Reclaims as many pages from the given memcg as possible and moves
4719 * the rest to the parent.
4721 * Caller is responsible for holding css reference for memcg.
4723 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
4725 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
4726 struct cgroup *cgrp = memcg->css.cgroup;
4728 /* returns EBUSY if there is a task or if we come here twice. */
4729 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
4732 /* we call try-to-free pages for make this cgroup empty */
4733 lru_add_drain_all();
4734 /* try to free all pages in this cgroup */
4735 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
4738 if (signal_pending(current))
4741 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
4745 /* maybe some writeback is necessary */
4746 congestion_wait(BLK_RW_ASYNC, HZ/10);
4751 mem_cgroup_reparent_charges(memcg);
4756 static int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
4758 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4761 if (mem_cgroup_is_root(memcg))
4763 css_get(&memcg->css);
4764 ret = mem_cgroup_force_empty(memcg);
4765 css_put(&memcg->css);
4771 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
4773 return mem_cgroup_from_cont(cont)->use_hierarchy;
4776 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
4780 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4781 struct cgroup *parent = cont->parent;
4782 struct mem_cgroup *parent_memcg = NULL;
4785 parent_memcg = mem_cgroup_from_cont(parent);
4789 if (memcg->use_hierarchy == val)
4793 * If parent's use_hierarchy is set, we can't make any modifications
4794 * in the child subtrees. If it is unset, then the change can
4795 * occur, provided the current cgroup has no children.
4797 * For the root cgroup, parent_mem is NULL, we allow value to be
4798 * set if there are no children.
4800 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
4801 (val == 1 || val == 0)) {
4802 if (list_empty(&cont->children))
4803 memcg->use_hierarchy = val;
4816 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
4817 enum mem_cgroup_stat_index idx)
4819 struct mem_cgroup *iter;
4822 /* Per-cpu values can be negative, use a signed accumulator */
4823 for_each_mem_cgroup_tree(iter, memcg)
4824 val += mem_cgroup_read_stat(iter, idx);
4826 if (val < 0) /* race ? */
4831 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
4835 if (!mem_cgroup_is_root(memcg)) {
4837 return res_counter_read_u64(&memcg->res, RES_USAGE);
4839 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
4842 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
4843 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
4846 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
4848 return val << PAGE_SHIFT;
4851 static ssize_t mem_cgroup_read(struct cgroup *cont, struct cftype *cft,
4852 struct file *file, char __user *buf,
4853 size_t nbytes, loff_t *ppos)
4855 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4861 type = MEMFILE_TYPE(cft->private);
4862 name = MEMFILE_ATTR(cft->private);
4864 if (!do_swap_account && type == _MEMSWAP)
4869 if (name == RES_USAGE)
4870 val = mem_cgroup_usage(memcg, false);
4872 val = res_counter_read_u64(&memcg->res, name);
4875 if (name == RES_USAGE)
4876 val = mem_cgroup_usage(memcg, true);
4878 val = res_counter_read_u64(&memcg->memsw, name);
4881 val = res_counter_read_u64(&memcg->kmem, name);
4887 len = scnprintf(str, sizeof(str), "%llu\n", (unsigned long long)val);
4888 return simple_read_from_buffer(buf, nbytes, ppos, str, len);
4891 static int memcg_update_kmem_limit(struct cgroup *cont, u64 val)
4894 #ifdef CONFIG_MEMCG_KMEM
4895 bool must_inc_static_branch = false;
4897 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4899 * For simplicity, we won't allow this to be disabled. It also can't
4900 * be changed if the cgroup has children already, or if tasks had
4903 * If tasks join before we set the limit, a person looking at
4904 * kmem.usage_in_bytes will have no way to determine when it took
4905 * place, which makes the value quite meaningless.
4907 * After it first became limited, changes in the value of the limit are
4908 * of course permitted.
4910 * Taking the cgroup_lock is really offensive, but it is so far the only
4911 * way to guarantee that no children will appear. There are plenty of
4912 * other offenders, and they should all go away. Fine grained locking
4913 * is probably the way to go here. When we are fully hierarchical, we
4914 * can also get rid of the use_hierarchy check.
4917 mutex_lock(&set_limit_mutex);
4918 if (!memcg->kmem_account_flags && val != RESOURCE_MAX) {
4919 if (cgroup_task_count(cont) || (memcg->use_hierarchy &&
4920 !list_empty(&cont->children))) {
4924 ret = res_counter_set_limit(&memcg->kmem, val);
4927 ret = memcg_update_cache_sizes(memcg);
4929 res_counter_set_limit(&memcg->kmem, RESOURCE_MAX);
4932 must_inc_static_branch = true;
4934 * kmem charges can outlive the cgroup. In the case of slab
4935 * pages, for instance, a page contain objects from various
4936 * processes, so it is unfeasible to migrate them away. We
4937 * need to reference count the memcg because of that.
4939 mem_cgroup_get(memcg);
4941 ret = res_counter_set_limit(&memcg->kmem, val);
4943 mutex_unlock(&set_limit_mutex);
4947 * We are by now familiar with the fact that we can't inc the static
4948 * branch inside cgroup_lock. See disarm functions for details. A
4949 * worker here is overkill, but also wrong: After the limit is set, we
4950 * must start accounting right away. Since this operation can't fail,
4951 * we can safely defer it to here - no rollback will be needed.
4953 * The boolean used to control this is also safe, because
4954 * KMEM_ACCOUNTED_ACTIVATED guarantees that only one process will be
4955 * able to set it to true;
4957 if (must_inc_static_branch) {
4958 static_key_slow_inc(&memcg_kmem_enabled_key);
4960 * setting the active bit after the inc will guarantee no one
4961 * starts accounting before all call sites are patched
4963 memcg_kmem_set_active(memcg);
4970 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
4973 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4977 memcg->kmem_account_flags = parent->kmem_account_flags;
4978 #ifdef CONFIG_MEMCG_KMEM
4980 * When that happen, we need to disable the static branch only on those
4981 * memcgs that enabled it. To achieve this, we would be forced to
4982 * complicate the code by keeping track of which memcgs were the ones
4983 * that actually enabled limits, and which ones got it from its
4986 * It is a lot simpler just to do static_key_slow_inc() on every child
4987 * that is accounted.
4989 if (!memcg_kmem_is_active(memcg))
4993 * destroy(), called if we fail, will issue static_key_slow_inc() and
4994 * mem_cgroup_put() if kmem is enabled. We have to either call them
4995 * unconditionally, or clear the KMEM_ACTIVE flag. I personally find
4996 * this more consistent, since it always leads to the same destroy path
4998 mem_cgroup_get(memcg);
4999 static_key_slow_inc(&memcg_kmem_enabled_key);
5001 mutex_lock(&set_limit_mutex);
5002 ret = memcg_update_cache_sizes(memcg);
5003 mutex_unlock(&set_limit_mutex);
5010 * The user of this function is...
5013 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
5016 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5019 unsigned long long val;
5022 type = MEMFILE_TYPE(cft->private);
5023 name = MEMFILE_ATTR(cft->private);
5025 if (!do_swap_account && type == _MEMSWAP)
5030 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
5034 /* This function does all necessary parse...reuse it */
5035 ret = res_counter_memparse_write_strategy(buffer, &val);
5039 ret = mem_cgroup_resize_limit(memcg, val);
5040 else if (type == _MEMSWAP)
5041 ret = mem_cgroup_resize_memsw_limit(memcg, val);
5042 else if (type == _KMEM)
5043 ret = memcg_update_kmem_limit(cont, val);
5047 case RES_SOFT_LIMIT:
5048 ret = res_counter_memparse_write_strategy(buffer, &val);
5052 * For memsw, soft limits are hard to implement in terms
5053 * of semantics, for now, we support soft limits for
5054 * control without swap
5057 ret = res_counter_set_soft_limit(&memcg->res, val);
5062 ret = -EINVAL; /* should be BUG() ? */
5068 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
5069 unsigned long long *mem_limit, unsigned long long *memsw_limit)
5071 struct cgroup *cgroup;
5072 unsigned long long min_limit, min_memsw_limit, tmp;
5074 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
5075 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5076 cgroup = memcg->css.cgroup;
5077 if (!memcg->use_hierarchy)
5080 while (cgroup->parent) {
5081 cgroup = cgroup->parent;
5082 memcg = mem_cgroup_from_cont(cgroup);
5083 if (!memcg->use_hierarchy)
5085 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
5086 min_limit = min(min_limit, tmp);
5087 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5088 min_memsw_limit = min(min_memsw_limit, tmp);
5091 *mem_limit = min_limit;
5092 *memsw_limit = min_memsw_limit;
5095 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
5097 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5101 type = MEMFILE_TYPE(event);
5102 name = MEMFILE_ATTR(event);
5104 if (!do_swap_account && type == _MEMSWAP)
5110 res_counter_reset_max(&memcg->res);
5111 else if (type == _MEMSWAP)
5112 res_counter_reset_max(&memcg->memsw);
5113 else if (type == _KMEM)
5114 res_counter_reset_max(&memcg->kmem);
5120 res_counter_reset_failcnt(&memcg->res);
5121 else if (type == _MEMSWAP)
5122 res_counter_reset_failcnt(&memcg->memsw);
5123 else if (type == _KMEM)
5124 res_counter_reset_failcnt(&memcg->kmem);
5133 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
5136 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
5140 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
5141 struct cftype *cft, u64 val)
5143 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5145 if (val >= (1 << NR_MOVE_TYPE))
5148 * We check this value several times in both in can_attach() and
5149 * attach(), so we need cgroup lock to prevent this value from being
5153 memcg->move_charge_at_immigrate = val;
5159 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
5160 struct cftype *cft, u64 val)
5167 static int memcg_numa_stat_show(struct cgroup *cont, struct cftype *cft,
5171 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
5172 unsigned long node_nr;
5173 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5175 total_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL);
5176 seq_printf(m, "total=%lu", total_nr);
5177 for_each_node_state(nid, N_HIGH_MEMORY) {
5178 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL);
5179 seq_printf(m, " N%d=%lu", nid, node_nr);
5183 file_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_FILE);
5184 seq_printf(m, "file=%lu", file_nr);
5185 for_each_node_state(nid, N_HIGH_MEMORY) {
5186 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5188 seq_printf(m, " N%d=%lu", nid, node_nr);
5192 anon_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_ANON);
5193 seq_printf(m, "anon=%lu", anon_nr);
5194 for_each_node_state(nid, N_HIGH_MEMORY) {
5195 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5197 seq_printf(m, " N%d=%lu", nid, node_nr);
5201 unevictable_nr = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
5202 seq_printf(m, "unevictable=%lu", unevictable_nr);
5203 for_each_node_state(nid, N_HIGH_MEMORY) {
5204 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5205 BIT(LRU_UNEVICTABLE));
5206 seq_printf(m, " N%d=%lu", nid, node_nr);
5211 #endif /* CONFIG_NUMA */
5213 static const char * const mem_cgroup_lru_names[] = {
5221 static inline void mem_cgroup_lru_names_not_uptodate(void)
5223 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
5226 static int memcg_stat_show(struct cgroup *cont, struct cftype *cft,
5229 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5230 struct mem_cgroup *mi;
5233 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5234 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5236 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
5237 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
5240 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
5241 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
5242 mem_cgroup_read_events(memcg, i));
5244 for (i = 0; i < NR_LRU_LISTS; i++)
5245 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
5246 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
5248 /* Hierarchical information */
5250 unsigned long long limit, memsw_limit;
5251 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
5252 seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
5253 if (do_swap_account)
5254 seq_printf(m, "hierarchical_memsw_limit %llu\n",
5258 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5261 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5263 for_each_mem_cgroup_tree(mi, memcg)
5264 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
5265 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
5268 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
5269 unsigned long long val = 0;
5271 for_each_mem_cgroup_tree(mi, memcg)
5272 val += mem_cgroup_read_events(mi, i);
5273 seq_printf(m, "total_%s %llu\n",
5274 mem_cgroup_events_names[i], val);
5277 for (i = 0; i < NR_LRU_LISTS; i++) {
5278 unsigned long long val = 0;
5280 for_each_mem_cgroup_tree(mi, memcg)
5281 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
5282 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
5285 #ifdef CONFIG_DEBUG_VM
5288 struct mem_cgroup_per_zone *mz;
5289 struct zone_reclaim_stat *rstat;
5290 unsigned long recent_rotated[2] = {0, 0};
5291 unsigned long recent_scanned[2] = {0, 0};
5293 for_each_online_node(nid)
5294 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
5295 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
5296 rstat = &mz->lruvec.reclaim_stat;
5298 recent_rotated[0] += rstat->recent_rotated[0];
5299 recent_rotated[1] += rstat->recent_rotated[1];
5300 recent_scanned[0] += rstat->recent_scanned[0];
5301 recent_scanned[1] += rstat->recent_scanned[1];
5303 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
5304 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
5305 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
5306 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
5313 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
5315 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5317 return mem_cgroup_swappiness(memcg);
5320 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
5323 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5324 struct mem_cgroup *parent;
5329 if (cgrp->parent == NULL)
5332 parent = mem_cgroup_from_cont(cgrp->parent);
5336 /* If under hierarchy, only empty-root can set this value */
5337 if ((parent->use_hierarchy) ||
5338 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
5343 memcg->swappiness = val;
5350 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
5352 struct mem_cgroup_threshold_ary *t;
5358 t = rcu_dereference(memcg->thresholds.primary);
5360 t = rcu_dereference(memcg->memsw_thresholds.primary);
5365 usage = mem_cgroup_usage(memcg, swap);
5368 * current_threshold points to threshold just below or equal to usage.
5369 * If it's not true, a threshold was crossed after last
5370 * call of __mem_cgroup_threshold().
5372 i = t->current_threshold;
5375 * Iterate backward over array of thresholds starting from
5376 * current_threshold and check if a threshold is crossed.
5377 * If none of thresholds below usage is crossed, we read
5378 * only one element of the array here.
5380 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
5381 eventfd_signal(t->entries[i].eventfd, 1);
5383 /* i = current_threshold + 1 */
5387 * Iterate forward over array of thresholds starting from
5388 * current_threshold+1 and check if a threshold is crossed.
5389 * If none of thresholds above usage is crossed, we read
5390 * only one element of the array here.
5392 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
5393 eventfd_signal(t->entries[i].eventfd, 1);
5395 /* Update current_threshold */
5396 t->current_threshold = i - 1;
5401 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
5404 __mem_cgroup_threshold(memcg, false);
5405 if (do_swap_account)
5406 __mem_cgroup_threshold(memcg, true);
5408 memcg = parent_mem_cgroup(memcg);
5412 static int compare_thresholds(const void *a, const void *b)
5414 const struct mem_cgroup_threshold *_a = a;
5415 const struct mem_cgroup_threshold *_b = b;
5417 return _a->threshold - _b->threshold;
5420 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
5422 struct mem_cgroup_eventfd_list *ev;
5424 list_for_each_entry(ev, &memcg->oom_notify, list)
5425 eventfd_signal(ev->eventfd, 1);
5429 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
5431 struct mem_cgroup *iter;
5433 for_each_mem_cgroup_tree(iter, memcg)
5434 mem_cgroup_oom_notify_cb(iter);
5437 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
5438 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5440 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5441 struct mem_cgroup_thresholds *thresholds;
5442 struct mem_cgroup_threshold_ary *new;
5443 enum res_type type = MEMFILE_TYPE(cft->private);
5444 u64 threshold, usage;
5447 ret = res_counter_memparse_write_strategy(args, &threshold);
5451 mutex_lock(&memcg->thresholds_lock);
5454 thresholds = &memcg->thresholds;
5455 else if (type == _MEMSWAP)
5456 thresholds = &memcg->memsw_thresholds;
5460 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5462 /* Check if a threshold crossed before adding a new one */
5463 if (thresholds->primary)
5464 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5466 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
5468 /* Allocate memory for new array of thresholds */
5469 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
5477 /* Copy thresholds (if any) to new array */
5478 if (thresholds->primary) {
5479 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
5480 sizeof(struct mem_cgroup_threshold));
5483 /* Add new threshold */
5484 new->entries[size - 1].eventfd = eventfd;
5485 new->entries[size - 1].threshold = threshold;
5487 /* Sort thresholds. Registering of new threshold isn't time-critical */
5488 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
5489 compare_thresholds, NULL);
5491 /* Find current threshold */
5492 new->current_threshold = -1;
5493 for (i = 0; i < size; i++) {
5494 if (new->entries[i].threshold <= usage) {
5496 * new->current_threshold will not be used until
5497 * rcu_assign_pointer(), so it's safe to increment
5500 ++new->current_threshold;
5505 /* Free old spare buffer and save old primary buffer as spare */
5506 kfree(thresholds->spare);
5507 thresholds->spare = thresholds->primary;
5509 rcu_assign_pointer(thresholds->primary, new);
5511 /* To be sure that nobody uses thresholds */
5515 mutex_unlock(&memcg->thresholds_lock);
5520 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
5521 struct cftype *cft, struct eventfd_ctx *eventfd)
5523 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5524 struct mem_cgroup_thresholds *thresholds;
5525 struct mem_cgroup_threshold_ary *new;
5526 enum res_type type = MEMFILE_TYPE(cft->private);
5530 mutex_lock(&memcg->thresholds_lock);
5532 thresholds = &memcg->thresholds;
5533 else if (type == _MEMSWAP)
5534 thresholds = &memcg->memsw_thresholds;
5538 if (!thresholds->primary)
5541 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5543 /* Check if a threshold crossed before removing */
5544 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5546 /* Calculate new number of threshold */
5548 for (i = 0; i < thresholds->primary->size; i++) {
5549 if (thresholds->primary->entries[i].eventfd != eventfd)
5553 new = thresholds->spare;
5555 /* Set thresholds array to NULL if we don't have thresholds */
5564 /* Copy thresholds and find current threshold */
5565 new->current_threshold = -1;
5566 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
5567 if (thresholds->primary->entries[i].eventfd == eventfd)
5570 new->entries[j] = thresholds->primary->entries[i];
5571 if (new->entries[j].threshold <= usage) {
5573 * new->current_threshold will not be used
5574 * until rcu_assign_pointer(), so it's safe to increment
5577 ++new->current_threshold;
5583 /* Swap primary and spare array */
5584 thresholds->spare = thresholds->primary;
5585 /* If all events are unregistered, free the spare array */
5587 kfree(thresholds->spare);
5588 thresholds->spare = NULL;
5591 rcu_assign_pointer(thresholds->primary, new);
5593 /* To be sure that nobody uses thresholds */
5596 mutex_unlock(&memcg->thresholds_lock);
5599 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
5600 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5602 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5603 struct mem_cgroup_eventfd_list *event;
5604 enum res_type type = MEMFILE_TYPE(cft->private);
5606 BUG_ON(type != _OOM_TYPE);
5607 event = kmalloc(sizeof(*event), GFP_KERNEL);
5611 spin_lock(&memcg_oom_lock);
5613 event->eventfd = eventfd;
5614 list_add(&event->list, &memcg->oom_notify);
5616 /* already in OOM ? */
5617 if (atomic_read(&memcg->under_oom))
5618 eventfd_signal(eventfd, 1);
5619 spin_unlock(&memcg_oom_lock);
5624 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
5625 struct cftype *cft, struct eventfd_ctx *eventfd)
5627 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5628 struct mem_cgroup_eventfd_list *ev, *tmp;
5629 enum res_type type = MEMFILE_TYPE(cft->private);
5631 BUG_ON(type != _OOM_TYPE);
5633 spin_lock(&memcg_oom_lock);
5635 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
5636 if (ev->eventfd == eventfd) {
5637 list_del(&ev->list);
5642 spin_unlock(&memcg_oom_lock);
5645 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
5646 struct cftype *cft, struct cgroup_map_cb *cb)
5648 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5650 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
5652 if (atomic_read(&memcg->under_oom))
5653 cb->fill(cb, "under_oom", 1);
5655 cb->fill(cb, "under_oom", 0);
5659 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
5660 struct cftype *cft, u64 val)
5662 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5663 struct mem_cgroup *parent;
5665 /* cannot set to root cgroup and only 0 and 1 are allowed */
5666 if (!cgrp->parent || !((val == 0) || (val == 1)))
5669 parent = mem_cgroup_from_cont(cgrp->parent);
5672 /* oom-kill-disable is a flag for subhierarchy. */
5673 if ((parent->use_hierarchy) ||
5674 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
5678 memcg->oom_kill_disable = val;
5680 memcg_oom_recover(memcg);
5685 #ifdef CONFIG_MEMCG_KMEM
5686 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5690 memcg->kmemcg_id = -1;
5691 ret = memcg_propagate_kmem(memcg);
5695 return mem_cgroup_sockets_init(memcg, ss);
5698 static void kmem_cgroup_destroy(struct mem_cgroup *memcg)
5700 mem_cgroup_sockets_destroy(memcg);
5702 memcg_kmem_mark_dead(memcg);
5704 if (res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0)
5708 * Charges already down to 0, undo mem_cgroup_get() done in the charge
5709 * path here, being careful not to race with memcg_uncharge_kmem: it is
5710 * possible that the charges went down to 0 between mark_dead and the
5711 * res_counter read, so in that case, we don't need the put
5713 if (memcg_kmem_test_and_clear_dead(memcg))
5714 mem_cgroup_put(memcg);
5717 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5722 static void kmem_cgroup_destroy(struct mem_cgroup *memcg)
5727 static struct cftype mem_cgroup_files[] = {
5729 .name = "usage_in_bytes",
5730 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
5731 .read = mem_cgroup_read,
5732 .register_event = mem_cgroup_usage_register_event,
5733 .unregister_event = mem_cgroup_usage_unregister_event,
5736 .name = "max_usage_in_bytes",
5737 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
5738 .trigger = mem_cgroup_reset,
5739 .read = mem_cgroup_read,
5742 .name = "limit_in_bytes",
5743 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
5744 .write_string = mem_cgroup_write,
5745 .read = mem_cgroup_read,
5748 .name = "soft_limit_in_bytes",
5749 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
5750 .write_string = mem_cgroup_write,
5751 .read = mem_cgroup_read,
5755 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
5756 .trigger = mem_cgroup_reset,
5757 .read = mem_cgroup_read,
5761 .read_seq_string = memcg_stat_show,
5764 .name = "force_empty",
5765 .trigger = mem_cgroup_force_empty_write,
5768 .name = "use_hierarchy",
5769 .write_u64 = mem_cgroup_hierarchy_write,
5770 .read_u64 = mem_cgroup_hierarchy_read,
5773 .name = "swappiness",
5774 .read_u64 = mem_cgroup_swappiness_read,
5775 .write_u64 = mem_cgroup_swappiness_write,
5778 .name = "move_charge_at_immigrate",
5779 .read_u64 = mem_cgroup_move_charge_read,
5780 .write_u64 = mem_cgroup_move_charge_write,
5783 .name = "oom_control",
5784 .read_map = mem_cgroup_oom_control_read,
5785 .write_u64 = mem_cgroup_oom_control_write,
5786 .register_event = mem_cgroup_oom_register_event,
5787 .unregister_event = mem_cgroup_oom_unregister_event,
5788 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
5792 .name = "numa_stat",
5793 .read_seq_string = memcg_numa_stat_show,
5796 #ifdef CONFIG_MEMCG_SWAP
5798 .name = "memsw.usage_in_bytes",
5799 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
5800 .read = mem_cgroup_read,
5801 .register_event = mem_cgroup_usage_register_event,
5802 .unregister_event = mem_cgroup_usage_unregister_event,
5805 .name = "memsw.max_usage_in_bytes",
5806 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
5807 .trigger = mem_cgroup_reset,
5808 .read = mem_cgroup_read,
5811 .name = "memsw.limit_in_bytes",
5812 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
5813 .write_string = mem_cgroup_write,
5814 .read = mem_cgroup_read,
5817 .name = "memsw.failcnt",
5818 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
5819 .trigger = mem_cgroup_reset,
5820 .read = mem_cgroup_read,
5823 #ifdef CONFIG_MEMCG_KMEM
5825 .name = "kmem.limit_in_bytes",
5826 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
5827 .write_string = mem_cgroup_write,
5828 .read = mem_cgroup_read,
5831 .name = "kmem.usage_in_bytes",
5832 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5833 .read = mem_cgroup_read,
5836 .name = "kmem.failcnt",
5837 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5838 .trigger = mem_cgroup_reset,
5839 .read = mem_cgroup_read,
5842 .name = "kmem.max_usage_in_bytes",
5843 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5844 .trigger = mem_cgroup_reset,
5845 .read = mem_cgroup_read,
5847 #ifdef CONFIG_SLABINFO
5849 .name = "kmem.slabinfo",
5850 .read_seq_string = mem_cgroup_slabinfo_read,
5854 { }, /* terminate */
5857 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
5859 struct mem_cgroup_per_node *pn;
5860 struct mem_cgroup_per_zone *mz;
5861 int zone, tmp = node;
5863 * This routine is called against possible nodes.
5864 * But it's BUG to call kmalloc() against offline node.
5866 * TODO: this routine can waste much memory for nodes which will
5867 * never be onlined. It's better to use memory hotplug callback
5870 if (!node_state(node, N_NORMAL_MEMORY))
5872 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5876 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5877 mz = &pn->zoneinfo[zone];
5878 lruvec_init(&mz->lruvec);
5879 mz->usage_in_excess = 0;
5880 mz->on_tree = false;
5883 memcg->info.nodeinfo[node] = pn;
5887 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
5889 kfree(memcg->info.nodeinfo[node]);
5892 static struct mem_cgroup *mem_cgroup_alloc(void)
5894 struct mem_cgroup *memcg;
5895 int size = sizeof(struct mem_cgroup);
5897 /* Can be very big if MAX_NUMNODES is very big */
5898 if (size < PAGE_SIZE)
5899 memcg = kzalloc(size, GFP_KERNEL);
5901 memcg = vzalloc(size);
5906 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
5909 spin_lock_init(&memcg->pcp_counter_lock);
5913 if (size < PAGE_SIZE)
5921 * At destroying mem_cgroup, references from swap_cgroup can remain.
5922 * (scanning all at force_empty is too costly...)
5924 * Instead of clearing all references at force_empty, we remember
5925 * the number of reference from swap_cgroup and free mem_cgroup when
5926 * it goes down to 0.
5928 * Removal of cgroup itself succeeds regardless of refs from swap.
5931 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5934 int size = sizeof(struct mem_cgroup);
5936 mem_cgroup_remove_from_trees(memcg);
5937 free_css_id(&mem_cgroup_subsys, &memcg->css);
5940 free_mem_cgroup_per_zone_info(memcg, node);
5942 free_percpu(memcg->stat);
5945 * We need to make sure that (at least for now), the jump label
5946 * destruction code runs outside of the cgroup lock. This is because
5947 * get_online_cpus(), which is called from the static_branch update,
5948 * can't be called inside the cgroup_lock. cpusets are the ones
5949 * enforcing this dependency, so if they ever change, we might as well.
5951 * schedule_work() will guarantee this happens. Be careful if you need
5952 * to move this code around, and make sure it is outside
5955 disarm_static_keys(memcg);
5956 if (size < PAGE_SIZE)
5964 * Helpers for freeing a kmalloc()ed/vzalloc()ed mem_cgroup by RCU,
5965 * but in process context. The work_freeing structure is overlaid
5966 * on the rcu_freeing structure, which itself is overlaid on memsw.
5968 static void free_work(struct work_struct *work)
5970 struct mem_cgroup *memcg;
5972 memcg = container_of(work, struct mem_cgroup, work_freeing);
5973 __mem_cgroup_free(memcg);
5976 static void free_rcu(struct rcu_head *rcu_head)
5978 struct mem_cgroup *memcg;
5980 memcg = container_of(rcu_head, struct mem_cgroup, rcu_freeing);
5981 INIT_WORK(&memcg->work_freeing, free_work);
5982 schedule_work(&memcg->work_freeing);
5985 static void mem_cgroup_get(struct mem_cgroup *memcg)
5987 atomic_inc(&memcg->refcnt);
5990 static void __mem_cgroup_put(struct mem_cgroup *memcg, int count)
5992 if (atomic_sub_and_test(count, &memcg->refcnt)) {
5993 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5994 call_rcu(&memcg->rcu_freeing, free_rcu);
5996 mem_cgroup_put(parent);
6000 static void mem_cgroup_put(struct mem_cgroup *memcg)
6002 __mem_cgroup_put(memcg, 1);
6006 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6008 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
6010 if (!memcg->res.parent)
6012 return mem_cgroup_from_res_counter(memcg->res.parent, res);
6014 EXPORT_SYMBOL(parent_mem_cgroup);
6016 #ifdef CONFIG_MEMCG_SWAP
6017 static void __init enable_swap_cgroup(void)
6019 if (!mem_cgroup_disabled() && really_do_swap_account)
6020 do_swap_account = 1;
6023 static void __init enable_swap_cgroup(void)
6028 static int mem_cgroup_soft_limit_tree_init(void)
6030 struct mem_cgroup_tree_per_node *rtpn;
6031 struct mem_cgroup_tree_per_zone *rtpz;
6032 int tmp, node, zone;
6034 for_each_node(node) {
6036 if (!node_state(node, N_NORMAL_MEMORY))
6038 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
6042 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6044 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6045 rtpz = &rtpn->rb_tree_per_zone[zone];
6046 rtpz->rb_root = RB_ROOT;
6047 spin_lock_init(&rtpz->lock);
6053 for_each_node(node) {
6054 if (!soft_limit_tree.rb_tree_per_node[node])
6056 kfree(soft_limit_tree.rb_tree_per_node[node]);
6057 soft_limit_tree.rb_tree_per_node[node] = NULL;
6063 static struct cgroup_subsys_state * __ref
6064 mem_cgroup_create(struct cgroup *cont)
6066 struct mem_cgroup *memcg, *parent;
6067 long error = -ENOMEM;
6070 memcg = mem_cgroup_alloc();
6072 return ERR_PTR(error);
6075 if (alloc_mem_cgroup_per_zone_info(memcg, node))
6079 if (cont->parent == NULL) {
6081 enable_swap_cgroup();
6083 if (mem_cgroup_soft_limit_tree_init())
6085 root_mem_cgroup = memcg;
6086 for_each_possible_cpu(cpu) {
6087 struct memcg_stock_pcp *stock =
6088 &per_cpu(memcg_stock, cpu);
6089 INIT_WORK(&stock->work, drain_local_stock);
6091 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
6093 parent = mem_cgroup_from_cont(cont->parent);
6094 memcg->use_hierarchy = parent->use_hierarchy;
6095 memcg->oom_kill_disable = parent->oom_kill_disable;
6098 if (parent && parent->use_hierarchy) {
6099 res_counter_init(&memcg->res, &parent->res);
6100 res_counter_init(&memcg->memsw, &parent->memsw);
6101 res_counter_init(&memcg->kmem, &parent->kmem);
6104 * We increment refcnt of the parent to ensure that we can
6105 * safely access it on res_counter_charge/uncharge.
6106 * This refcnt will be decremented when freeing this
6107 * mem_cgroup(see mem_cgroup_put).
6109 mem_cgroup_get(parent);
6111 res_counter_init(&memcg->res, NULL);
6112 res_counter_init(&memcg->memsw, NULL);
6113 res_counter_init(&memcg->kmem, NULL);
6115 * Deeper hierachy with use_hierarchy == false doesn't make
6116 * much sense so let cgroup subsystem know about this
6117 * unfortunate state in our controller.
6119 if (parent && parent != root_mem_cgroup)
6120 mem_cgroup_subsys.broken_hierarchy = true;
6122 memcg->last_scanned_node = MAX_NUMNODES;
6123 INIT_LIST_HEAD(&memcg->oom_notify);
6126 memcg->swappiness = mem_cgroup_swappiness(parent);
6127 atomic_set(&memcg->refcnt, 1);
6128 memcg->move_charge_at_immigrate = 0;
6129 mutex_init(&memcg->thresholds_lock);
6130 spin_lock_init(&memcg->move_lock);
6132 error = memcg_init_kmem(memcg, &mem_cgroup_subsys);
6135 * We call put now because our (and parent's) refcnts
6136 * are already in place. mem_cgroup_put() will internally
6137 * call __mem_cgroup_free, so return directly
6139 mem_cgroup_put(memcg);
6140 return ERR_PTR(error);
6144 __mem_cgroup_free(memcg);
6145 return ERR_PTR(error);
6148 static void mem_cgroup_pre_destroy(struct cgroup *cont)
6150 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
6152 mem_cgroup_reparent_charges(memcg);
6153 mem_cgroup_destroy_all_caches(memcg);
6156 static void mem_cgroup_destroy(struct cgroup *cont)
6158 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
6160 kmem_cgroup_destroy(memcg);
6162 mem_cgroup_put(memcg);
6166 /* Handlers for move charge at task migration. */
6167 #define PRECHARGE_COUNT_AT_ONCE 256
6168 static int mem_cgroup_do_precharge(unsigned long count)
6171 int batch_count = PRECHARGE_COUNT_AT_ONCE;
6172 struct mem_cgroup *memcg = mc.to;
6174 if (mem_cgroup_is_root(memcg)) {
6175 mc.precharge += count;
6176 /* we don't need css_get for root */
6179 /* try to charge at once */
6181 struct res_counter *dummy;
6183 * "memcg" cannot be under rmdir() because we've already checked
6184 * by cgroup_lock_live_cgroup() that it is not removed and we
6185 * are still under the same cgroup_mutex. So we can postpone
6188 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
6190 if (do_swap_account && res_counter_charge(&memcg->memsw,
6191 PAGE_SIZE * count, &dummy)) {
6192 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
6195 mc.precharge += count;
6199 /* fall back to one by one charge */
6201 if (signal_pending(current)) {
6205 if (!batch_count--) {
6206 batch_count = PRECHARGE_COUNT_AT_ONCE;
6209 ret = __mem_cgroup_try_charge(NULL,
6210 GFP_KERNEL, 1, &memcg, false);
6212 /* mem_cgroup_clear_mc() will do uncharge later */
6220 * get_mctgt_type - get target type of moving charge
6221 * @vma: the vma the pte to be checked belongs
6222 * @addr: the address corresponding to the pte to be checked
6223 * @ptent: the pte to be checked
6224 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6227 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6228 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6229 * move charge. if @target is not NULL, the page is stored in target->page
6230 * with extra refcnt got(Callers should handle it).
6231 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6232 * target for charge migration. if @target is not NULL, the entry is stored
6235 * Called with pte lock held.
6242 enum mc_target_type {
6248 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
6249 unsigned long addr, pte_t ptent)
6251 struct page *page = vm_normal_page(vma, addr, ptent);
6253 if (!page || !page_mapped(page))
6255 if (PageAnon(page)) {
6256 /* we don't move shared anon */
6259 } else if (!move_file())
6260 /* we ignore mapcount for file pages */
6262 if (!get_page_unless_zero(page))
6269 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6270 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6272 struct page *page = NULL;
6273 swp_entry_t ent = pte_to_swp_entry(ptent);
6275 if (!move_anon() || non_swap_entry(ent))
6278 * Because lookup_swap_cache() updates some statistics counter,
6279 * we call find_get_page() with swapper_space directly.
6281 page = find_get_page(&swapper_space, ent.val);
6282 if (do_swap_account)
6283 entry->val = ent.val;
6288 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6289 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6295 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
6296 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6298 struct page *page = NULL;
6299 struct address_space *mapping;
6302 if (!vma->vm_file) /* anonymous vma */
6307 mapping = vma->vm_file->f_mapping;
6308 if (pte_none(ptent))
6309 pgoff = linear_page_index(vma, addr);
6310 else /* pte_file(ptent) is true */
6311 pgoff = pte_to_pgoff(ptent);
6313 /* page is moved even if it's not RSS of this task(page-faulted). */
6314 page = find_get_page(mapping, pgoff);
6317 /* shmem/tmpfs may report page out on swap: account for that too. */
6318 if (radix_tree_exceptional_entry(page)) {
6319 swp_entry_t swap = radix_to_swp_entry(page);
6320 if (do_swap_account)
6322 page = find_get_page(&swapper_space, swap.val);
6328 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
6329 unsigned long addr, pte_t ptent, union mc_target *target)
6331 struct page *page = NULL;
6332 struct page_cgroup *pc;
6333 enum mc_target_type ret = MC_TARGET_NONE;
6334 swp_entry_t ent = { .val = 0 };
6336 if (pte_present(ptent))
6337 page = mc_handle_present_pte(vma, addr, ptent);
6338 else if (is_swap_pte(ptent))
6339 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
6340 else if (pte_none(ptent) || pte_file(ptent))
6341 page = mc_handle_file_pte(vma, addr, ptent, &ent);
6343 if (!page && !ent.val)
6346 pc = lookup_page_cgroup(page);
6348 * Do only loose check w/o page_cgroup lock.
6349 * mem_cgroup_move_account() checks the pc is valid or not under
6352 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6353 ret = MC_TARGET_PAGE;
6355 target->page = page;
6357 if (!ret || !target)
6360 /* There is a swap entry and a page doesn't exist or isn't charged */
6361 if (ent.val && !ret &&
6362 css_id(&mc.from->css) == lookup_swap_cgroup_id(ent)) {
6363 ret = MC_TARGET_SWAP;
6370 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6372 * We don't consider swapping or file mapped pages because THP does not
6373 * support them for now.
6374 * Caller should make sure that pmd_trans_huge(pmd) is true.
6376 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6377 unsigned long addr, pmd_t pmd, union mc_target *target)
6379 struct page *page = NULL;
6380 struct page_cgroup *pc;
6381 enum mc_target_type ret = MC_TARGET_NONE;
6383 page = pmd_page(pmd);
6384 VM_BUG_ON(!page || !PageHead(page));
6387 pc = lookup_page_cgroup(page);
6388 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6389 ret = MC_TARGET_PAGE;
6392 target->page = page;
6398 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6399 unsigned long addr, pmd_t pmd, union mc_target *target)
6401 return MC_TARGET_NONE;
6405 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
6406 unsigned long addr, unsigned long end,
6407 struct mm_walk *walk)
6409 struct vm_area_struct *vma = walk->private;
6413 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6414 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
6415 mc.precharge += HPAGE_PMD_NR;
6416 spin_unlock(&vma->vm_mm->page_table_lock);
6420 if (pmd_trans_unstable(pmd))
6422 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6423 for (; addr != end; pte++, addr += PAGE_SIZE)
6424 if (get_mctgt_type(vma, addr, *pte, NULL))
6425 mc.precharge++; /* increment precharge temporarily */
6426 pte_unmap_unlock(pte - 1, ptl);
6432 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
6434 unsigned long precharge;
6435 struct vm_area_struct *vma;
6437 down_read(&mm->mmap_sem);
6438 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6439 struct mm_walk mem_cgroup_count_precharge_walk = {
6440 .pmd_entry = mem_cgroup_count_precharge_pte_range,
6444 if (is_vm_hugetlb_page(vma))
6446 walk_page_range(vma->vm_start, vma->vm_end,
6447 &mem_cgroup_count_precharge_walk);
6449 up_read(&mm->mmap_sem);
6451 precharge = mc.precharge;
6457 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
6459 unsigned long precharge = mem_cgroup_count_precharge(mm);
6461 VM_BUG_ON(mc.moving_task);
6462 mc.moving_task = current;
6463 return mem_cgroup_do_precharge(precharge);
6466 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6467 static void __mem_cgroup_clear_mc(void)
6469 struct mem_cgroup *from = mc.from;
6470 struct mem_cgroup *to = mc.to;
6472 /* we must uncharge all the leftover precharges from mc.to */
6474 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
6478 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6479 * we must uncharge here.
6481 if (mc.moved_charge) {
6482 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
6483 mc.moved_charge = 0;
6485 /* we must fixup refcnts and charges */
6486 if (mc.moved_swap) {
6487 /* uncharge swap account from the old cgroup */
6488 if (!mem_cgroup_is_root(mc.from))
6489 res_counter_uncharge(&mc.from->memsw,
6490 PAGE_SIZE * mc.moved_swap);
6491 __mem_cgroup_put(mc.from, mc.moved_swap);
6493 if (!mem_cgroup_is_root(mc.to)) {
6495 * we charged both to->res and to->memsw, so we should
6498 res_counter_uncharge(&mc.to->res,
6499 PAGE_SIZE * mc.moved_swap);
6501 /* we've already done mem_cgroup_get(mc.to) */
6504 memcg_oom_recover(from);
6505 memcg_oom_recover(to);
6506 wake_up_all(&mc.waitq);
6509 static void mem_cgroup_clear_mc(void)
6511 struct mem_cgroup *from = mc.from;
6514 * we must clear moving_task before waking up waiters at the end of
6517 mc.moving_task = NULL;
6518 __mem_cgroup_clear_mc();
6519 spin_lock(&mc.lock);
6522 spin_unlock(&mc.lock);
6523 mem_cgroup_end_move(from);
6526 static int mem_cgroup_can_attach(struct cgroup *cgroup,
6527 struct cgroup_taskset *tset)
6529 struct task_struct *p = cgroup_taskset_first(tset);
6531 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup);
6533 if (memcg->move_charge_at_immigrate) {
6534 struct mm_struct *mm;
6535 struct mem_cgroup *from = mem_cgroup_from_task(p);
6537 VM_BUG_ON(from == memcg);
6539 mm = get_task_mm(p);
6542 /* We move charges only when we move a owner of the mm */
6543 if (mm->owner == p) {
6546 VM_BUG_ON(mc.precharge);
6547 VM_BUG_ON(mc.moved_charge);
6548 VM_BUG_ON(mc.moved_swap);
6549 mem_cgroup_start_move(from);
6550 spin_lock(&mc.lock);
6553 spin_unlock(&mc.lock);
6554 /* We set mc.moving_task later */
6556 ret = mem_cgroup_precharge_mc(mm);
6558 mem_cgroup_clear_mc();
6565 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
6566 struct cgroup_taskset *tset)
6568 mem_cgroup_clear_mc();
6571 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6572 unsigned long addr, unsigned long end,
6573 struct mm_walk *walk)
6576 struct vm_area_struct *vma = walk->private;
6579 enum mc_target_type target_type;
6580 union mc_target target;
6582 struct page_cgroup *pc;
6585 * We don't take compound_lock() here but no race with splitting thp
6587 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6588 * under splitting, which means there's no concurrent thp split,
6589 * - if another thread runs into split_huge_page() just after we
6590 * entered this if-block, the thread must wait for page table lock
6591 * to be unlocked in __split_huge_page_splitting(), where the main
6592 * part of thp split is not executed yet.
6594 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6595 if (mc.precharge < HPAGE_PMD_NR) {
6596 spin_unlock(&vma->vm_mm->page_table_lock);
6599 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6600 if (target_type == MC_TARGET_PAGE) {
6602 if (!isolate_lru_page(page)) {
6603 pc = lookup_page_cgroup(page);
6604 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
6605 pc, mc.from, mc.to)) {
6606 mc.precharge -= HPAGE_PMD_NR;
6607 mc.moved_charge += HPAGE_PMD_NR;
6609 putback_lru_page(page);
6613 spin_unlock(&vma->vm_mm->page_table_lock);
6617 if (pmd_trans_unstable(pmd))
6620 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6621 for (; addr != end; addr += PAGE_SIZE) {
6622 pte_t ptent = *(pte++);
6628 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6629 case MC_TARGET_PAGE:
6631 if (isolate_lru_page(page))
6633 pc = lookup_page_cgroup(page);
6634 if (!mem_cgroup_move_account(page, 1, pc,
6637 /* we uncharge from mc.from later. */
6640 putback_lru_page(page);
6641 put: /* get_mctgt_type() gets the page */
6644 case MC_TARGET_SWAP:
6646 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6648 /* we fixup refcnts and charges later. */
6656 pte_unmap_unlock(pte - 1, ptl);
6661 * We have consumed all precharges we got in can_attach().
6662 * We try charge one by one, but don't do any additional
6663 * charges to mc.to if we have failed in charge once in attach()
6666 ret = mem_cgroup_do_precharge(1);
6674 static void mem_cgroup_move_charge(struct mm_struct *mm)
6676 struct vm_area_struct *vma;
6678 lru_add_drain_all();
6680 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
6682 * Someone who are holding the mmap_sem might be waiting in
6683 * waitq. So we cancel all extra charges, wake up all waiters,
6684 * and retry. Because we cancel precharges, we might not be able
6685 * to move enough charges, but moving charge is a best-effort
6686 * feature anyway, so it wouldn't be a big problem.
6688 __mem_cgroup_clear_mc();
6692 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6694 struct mm_walk mem_cgroup_move_charge_walk = {
6695 .pmd_entry = mem_cgroup_move_charge_pte_range,
6699 if (is_vm_hugetlb_page(vma))
6701 ret = walk_page_range(vma->vm_start, vma->vm_end,
6702 &mem_cgroup_move_charge_walk);
6705 * means we have consumed all precharges and failed in
6706 * doing additional charge. Just abandon here.
6710 up_read(&mm->mmap_sem);
6713 static void mem_cgroup_move_task(struct cgroup *cont,
6714 struct cgroup_taskset *tset)
6716 struct task_struct *p = cgroup_taskset_first(tset);
6717 struct mm_struct *mm = get_task_mm(p);
6721 mem_cgroup_move_charge(mm);
6725 mem_cgroup_clear_mc();
6727 #else /* !CONFIG_MMU */
6728 static int mem_cgroup_can_attach(struct cgroup *cgroup,
6729 struct cgroup_taskset *tset)
6733 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
6734 struct cgroup_taskset *tset)
6737 static void mem_cgroup_move_task(struct cgroup *cont,
6738 struct cgroup_taskset *tset)
6743 struct cgroup_subsys mem_cgroup_subsys = {
6745 .subsys_id = mem_cgroup_subsys_id,
6746 .create = mem_cgroup_create,
6747 .pre_destroy = mem_cgroup_pre_destroy,
6748 .destroy = mem_cgroup_destroy,
6749 .can_attach = mem_cgroup_can_attach,
6750 .cancel_attach = mem_cgroup_cancel_attach,
6751 .attach = mem_cgroup_move_task,
6752 .base_cftypes = mem_cgroup_files,
6757 #ifdef CONFIG_MEMCG_SWAP
6758 static int __init enable_swap_account(char *s)
6760 /* consider enabled if no parameter or 1 is given */
6761 if (!strcmp(s, "1"))
6762 really_do_swap_account = 1;
6763 else if (!strcmp(s, "0"))
6764 really_do_swap_account = 0;
6767 __setup("swapaccount=", enable_swap_account);