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/poll.h>
49 #include <linux/sort.h>
51 #include <linux/seq_file.h>
52 #include <linux/vmpressure.h>
53 #include <linux/mm_inline.h>
54 #include <linux/page_cgroup.h>
55 #include <linux/cpu.h>
56 #include <linux/oom.h>
57 #include <linux/lockdep.h>
58 #include <linux/file.h>
62 #include <net/tcp_memcontrol.h>
65 #include <asm/uaccess.h>
67 #include <trace/events/vmscan.h>
69 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
70 EXPORT_SYMBOL(memory_cgrp_subsys);
72 #define MEM_CGROUP_RECLAIM_RETRIES 5
73 static struct mem_cgroup *root_mem_cgroup __read_mostly;
75 #ifdef CONFIG_MEMCG_SWAP
76 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
77 int do_swap_account __read_mostly;
79 /* for remember boot option*/
80 #ifdef CONFIG_MEMCG_SWAP_ENABLED
81 static int really_do_swap_account __initdata = 1;
83 static int really_do_swap_account __initdata;
87 #define do_swap_account 0
91 static const char * const mem_cgroup_stat_names[] = {
100 enum mem_cgroup_events_index {
101 MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */
102 MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */
103 MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */
104 MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */
105 MEM_CGROUP_EVENTS_NSTATS,
108 static const char * const mem_cgroup_events_names[] = {
115 static const char * const mem_cgroup_lru_names[] = {
124 * Per memcg event counter is incremented at every pagein/pageout. With THP,
125 * it will be incremated by the number of pages. This counter is used for
126 * for trigger some periodic events. This is straightforward and better
127 * than using jiffies etc. to handle periodic memcg event.
129 enum mem_cgroup_events_target {
130 MEM_CGROUP_TARGET_THRESH,
131 MEM_CGROUP_TARGET_SOFTLIMIT,
132 MEM_CGROUP_TARGET_NUMAINFO,
135 #define THRESHOLDS_EVENTS_TARGET 128
136 #define SOFTLIMIT_EVENTS_TARGET 1024
137 #define NUMAINFO_EVENTS_TARGET 1024
139 struct mem_cgroup_stat_cpu {
140 long count[MEM_CGROUP_STAT_NSTATS];
141 unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
142 unsigned long nr_page_events;
143 unsigned long targets[MEM_CGROUP_NTARGETS];
146 struct mem_cgroup_reclaim_iter {
148 * last scanned hierarchy member. Valid only if last_dead_count
149 * matches memcg->dead_count of the hierarchy root group.
151 struct mem_cgroup *last_visited;
154 /* scan generation, increased every round-trip */
155 unsigned int generation;
159 * per-zone information in memory controller.
161 struct mem_cgroup_per_zone {
162 struct lruvec lruvec;
163 unsigned long lru_size[NR_LRU_LISTS];
165 struct mem_cgroup_reclaim_iter reclaim_iter[DEF_PRIORITY + 1];
167 struct rb_node tree_node; /* RB tree node */
168 unsigned long long usage_in_excess;/* Set to the value by which */
169 /* the soft limit is exceeded*/
171 struct mem_cgroup *memcg; /* Back pointer, we cannot */
172 /* use container_of */
175 struct mem_cgroup_per_node {
176 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
180 * Cgroups above their limits are maintained in a RB-Tree, independent of
181 * their hierarchy representation
184 struct mem_cgroup_tree_per_zone {
185 struct rb_root rb_root;
189 struct mem_cgroup_tree_per_node {
190 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
193 struct mem_cgroup_tree {
194 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
197 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
199 struct mem_cgroup_threshold {
200 struct eventfd_ctx *eventfd;
205 struct mem_cgroup_threshold_ary {
206 /* An array index points to threshold just below or equal to usage. */
207 int current_threshold;
208 /* Size of entries[] */
210 /* Array of thresholds */
211 struct mem_cgroup_threshold entries[0];
214 struct mem_cgroup_thresholds {
215 /* Primary thresholds array */
216 struct mem_cgroup_threshold_ary *primary;
218 * Spare threshold array.
219 * This is needed to make mem_cgroup_unregister_event() "never fail".
220 * It must be able to store at least primary->size - 1 entries.
222 struct mem_cgroup_threshold_ary *spare;
226 struct mem_cgroup_eventfd_list {
227 struct list_head list;
228 struct eventfd_ctx *eventfd;
232 * cgroup_event represents events which userspace want to receive.
234 struct mem_cgroup_event {
236 * memcg which the event belongs to.
238 struct mem_cgroup *memcg;
240 * eventfd to signal userspace about the event.
242 struct eventfd_ctx *eventfd;
244 * Each of these stored in a list by the cgroup.
246 struct list_head list;
248 * register_event() callback will be used to add new userspace
249 * waiter for changes related to this event. Use eventfd_signal()
250 * on eventfd to send notification to userspace.
252 int (*register_event)(struct mem_cgroup *memcg,
253 struct eventfd_ctx *eventfd, const char *args);
255 * unregister_event() callback will be called when userspace closes
256 * the eventfd or on cgroup removing. This callback must be set,
257 * if you want provide notification functionality.
259 void (*unregister_event)(struct mem_cgroup *memcg,
260 struct eventfd_ctx *eventfd);
262 * All fields below needed to unregister event when
263 * userspace closes eventfd.
266 wait_queue_head_t *wqh;
268 struct work_struct remove;
271 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
272 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
275 * The memory controller data structure. The memory controller controls both
276 * page cache and RSS per cgroup. We would eventually like to provide
277 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
278 * to help the administrator determine what knobs to tune.
280 * TODO: Add a water mark for the memory controller. Reclaim will begin when
281 * we hit the water mark. May be even add a low water mark, such that
282 * no reclaim occurs from a cgroup at it's low water mark, this is
283 * a feature that will be implemented much later in the future.
286 struct cgroup_subsys_state css;
288 * the counter to account for memory usage
290 struct res_counter res;
292 /* vmpressure notifications */
293 struct vmpressure vmpressure;
296 * the counter to account for mem+swap usage.
298 struct res_counter memsw;
301 * the counter to account for kernel memory usage.
303 struct res_counter kmem;
305 * Should the accounting and control be hierarchical, per subtree?
308 unsigned long kmem_account_flags; /* See KMEM_ACCOUNTED_*, below */
312 atomic_t oom_wakeups;
315 /* OOM-Killer disable */
316 int oom_kill_disable;
318 /* set when res.limit == memsw.limit */
319 bool memsw_is_minimum;
321 /* protect arrays of thresholds */
322 struct mutex thresholds_lock;
324 /* thresholds for memory usage. RCU-protected */
325 struct mem_cgroup_thresholds thresholds;
327 /* thresholds for mem+swap usage. RCU-protected */
328 struct mem_cgroup_thresholds memsw_thresholds;
330 /* For oom notifier event fd */
331 struct list_head oom_notify;
334 * Should we move charges of a task when a task is moved into this
335 * mem_cgroup ? And what type of charges should we move ?
337 unsigned long move_charge_at_immigrate;
339 * set > 0 if pages under this cgroup are moving to other cgroup.
341 atomic_t moving_account;
342 /* taken only while moving_account > 0 */
343 spinlock_t move_lock;
347 struct mem_cgroup_stat_cpu __percpu *stat;
349 * used when a cpu is offlined or other synchronizations
350 * See mem_cgroup_read_stat().
352 struct mem_cgroup_stat_cpu nocpu_base;
353 spinlock_t pcp_counter_lock;
356 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
357 struct cg_proto tcp_mem;
359 #if defined(CONFIG_MEMCG_KMEM)
360 /* analogous to slab_common's slab_caches list, but per-memcg;
361 * protected by memcg_slab_mutex */
362 struct list_head memcg_slab_caches;
363 /* Index in the kmem_cache->memcg_params->memcg_caches array */
367 int last_scanned_node;
369 nodemask_t scan_nodes;
370 atomic_t numainfo_events;
371 atomic_t numainfo_updating;
374 /* List of events which userspace want to receive */
375 struct list_head event_list;
376 spinlock_t event_list_lock;
378 struct mem_cgroup_per_node *nodeinfo[0];
379 /* WARNING: nodeinfo must be the last member here */
382 /* internal only representation about the status of kmem accounting. */
384 KMEM_ACCOUNTED_ACTIVE, /* accounted by this cgroup itself */
385 KMEM_ACCOUNTED_DEAD, /* dead memcg with pending kmem charges */
388 #ifdef CONFIG_MEMCG_KMEM
389 static inline void memcg_kmem_set_active(struct mem_cgroup *memcg)
391 set_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
394 static bool memcg_kmem_is_active(struct mem_cgroup *memcg)
396 return test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
399 static void memcg_kmem_mark_dead(struct mem_cgroup *memcg)
402 * Our caller must use css_get() first, because memcg_uncharge_kmem()
403 * will call css_put() if it sees the memcg is dead.
406 if (test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags))
407 set_bit(KMEM_ACCOUNTED_DEAD, &memcg->kmem_account_flags);
410 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup *memcg)
412 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD,
413 &memcg->kmem_account_flags);
417 /* Stuffs for move charges at task migration. */
419 * Types of charges to be moved. "move_charge_at_immitgrate" and
420 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
423 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
424 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
428 /* "mc" and its members are protected by cgroup_mutex */
429 static struct move_charge_struct {
430 spinlock_t lock; /* for from, to */
431 struct mem_cgroup *from;
432 struct mem_cgroup *to;
433 unsigned long immigrate_flags;
434 unsigned long precharge;
435 unsigned long moved_charge;
436 unsigned long moved_swap;
437 struct task_struct *moving_task; /* a task moving charges */
438 wait_queue_head_t waitq; /* a waitq for other context */
440 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
441 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
444 static bool move_anon(void)
446 return test_bit(MOVE_CHARGE_TYPE_ANON, &mc.immigrate_flags);
449 static bool move_file(void)
451 return test_bit(MOVE_CHARGE_TYPE_FILE, &mc.immigrate_flags);
455 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
456 * limit reclaim to prevent infinite loops, if they ever occur.
458 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
459 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
462 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
463 MEM_CGROUP_CHARGE_TYPE_ANON,
464 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
465 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
469 /* for encoding cft->private value on file */
477 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
478 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
479 #define MEMFILE_ATTR(val) ((val) & 0xffff)
480 /* Used for OOM nofiier */
481 #define OOM_CONTROL (0)
484 * Reclaim flags for mem_cgroup_hierarchical_reclaim
486 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
487 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
488 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
489 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
492 * The memcg_create_mutex will be held whenever a new cgroup is created.
493 * As a consequence, any change that needs to protect against new child cgroups
494 * appearing has to hold it as well.
496 static DEFINE_MUTEX(memcg_create_mutex);
498 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
500 return s ? container_of(s, struct mem_cgroup, css) : NULL;
503 /* Some nice accessors for the vmpressure. */
504 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
507 memcg = root_mem_cgroup;
508 return &memcg->vmpressure;
511 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
513 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
516 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
518 return (memcg == root_mem_cgroup);
522 * We restrict the id in the range of [1, 65535], so it can fit into
525 #define MEM_CGROUP_ID_MAX USHRT_MAX
527 static inline unsigned short mem_cgroup_id(struct mem_cgroup *memcg)
530 * The ID of the root cgroup is 0, but memcg treat 0 as an
531 * invalid ID, so we return (cgroup_id + 1).
533 return memcg->css.cgroup->id + 1;
536 static inline struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
538 struct cgroup_subsys_state *css;
540 css = css_from_id(id - 1, &memory_cgrp_subsys);
541 return mem_cgroup_from_css(css);
544 /* Writing them here to avoid exposing memcg's inner layout */
545 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
547 void sock_update_memcg(struct sock *sk)
549 if (mem_cgroup_sockets_enabled) {
550 struct mem_cgroup *memcg;
551 struct cg_proto *cg_proto;
553 BUG_ON(!sk->sk_prot->proto_cgroup);
555 /* Socket cloning can throw us here with sk_cgrp already
556 * filled. It won't however, necessarily happen from
557 * process context. So the test for root memcg given
558 * the current task's memcg won't help us in this case.
560 * Respecting the original socket's memcg is a better
561 * decision in this case.
564 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
565 css_get(&sk->sk_cgrp->memcg->css);
570 memcg = mem_cgroup_from_task(current);
571 cg_proto = sk->sk_prot->proto_cgroup(memcg);
572 if (!mem_cgroup_is_root(memcg) &&
573 memcg_proto_active(cg_proto) && css_tryget(&memcg->css)) {
574 sk->sk_cgrp = cg_proto;
579 EXPORT_SYMBOL(sock_update_memcg);
581 void sock_release_memcg(struct sock *sk)
583 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
584 struct mem_cgroup *memcg;
585 WARN_ON(!sk->sk_cgrp->memcg);
586 memcg = sk->sk_cgrp->memcg;
587 css_put(&sk->sk_cgrp->memcg->css);
591 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
593 if (!memcg || mem_cgroup_is_root(memcg))
596 return &memcg->tcp_mem;
598 EXPORT_SYMBOL(tcp_proto_cgroup);
600 static void disarm_sock_keys(struct mem_cgroup *memcg)
602 if (!memcg_proto_activated(&memcg->tcp_mem))
604 static_key_slow_dec(&memcg_socket_limit_enabled);
607 static void disarm_sock_keys(struct mem_cgroup *memcg)
612 #ifdef CONFIG_MEMCG_KMEM
614 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
615 * The main reason for not using cgroup id for this:
616 * this works better in sparse environments, where we have a lot of memcgs,
617 * but only a few kmem-limited. Or also, if we have, for instance, 200
618 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
619 * 200 entry array for that.
621 * The current size of the caches array is stored in
622 * memcg_limited_groups_array_size. It will double each time we have to
625 static DEFINE_IDA(kmem_limited_groups);
626 int memcg_limited_groups_array_size;
629 * MIN_SIZE is different than 1, because we would like to avoid going through
630 * the alloc/free process all the time. In a small machine, 4 kmem-limited
631 * cgroups is a reasonable guess. In the future, it could be a parameter or
632 * tunable, but that is strictly not necessary.
634 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
635 * this constant directly from cgroup, but it is understandable that this is
636 * better kept as an internal representation in cgroup.c. In any case, the
637 * cgrp_id space is not getting any smaller, and we don't have to necessarily
638 * increase ours as well if it increases.
640 #define MEMCG_CACHES_MIN_SIZE 4
641 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
644 * A lot of the calls to the cache allocation functions are expected to be
645 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
646 * conditional to this static branch, we'll have to allow modules that does
647 * kmem_cache_alloc and the such to see this symbol as well
649 struct static_key memcg_kmem_enabled_key;
650 EXPORT_SYMBOL(memcg_kmem_enabled_key);
652 static void disarm_kmem_keys(struct mem_cgroup *memcg)
654 if (memcg_kmem_is_active(memcg)) {
655 static_key_slow_dec(&memcg_kmem_enabled_key);
656 ida_simple_remove(&kmem_limited_groups, memcg->kmemcg_id);
659 * This check can't live in kmem destruction function,
660 * since the charges will outlive the cgroup
662 WARN_ON(res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0);
665 static void disarm_kmem_keys(struct mem_cgroup *memcg)
668 #endif /* CONFIG_MEMCG_KMEM */
670 static void disarm_static_keys(struct mem_cgroup *memcg)
672 disarm_sock_keys(memcg);
673 disarm_kmem_keys(memcg);
676 static void drain_all_stock_async(struct mem_cgroup *memcg);
678 static struct mem_cgroup_per_zone *
679 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
681 VM_BUG_ON((unsigned)nid >= nr_node_ids);
682 return &memcg->nodeinfo[nid]->zoneinfo[zid];
685 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
690 static struct mem_cgroup_per_zone *
691 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
693 int nid = page_to_nid(page);
694 int zid = page_zonenum(page);
696 return mem_cgroup_zoneinfo(memcg, nid, zid);
699 static struct mem_cgroup_tree_per_zone *
700 soft_limit_tree_node_zone(int nid, int zid)
702 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
705 static struct mem_cgroup_tree_per_zone *
706 soft_limit_tree_from_page(struct page *page)
708 int nid = page_to_nid(page);
709 int zid = page_zonenum(page);
711 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
715 __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
716 struct mem_cgroup_per_zone *mz,
717 struct mem_cgroup_tree_per_zone *mctz,
718 unsigned long long new_usage_in_excess)
720 struct rb_node **p = &mctz->rb_root.rb_node;
721 struct rb_node *parent = NULL;
722 struct mem_cgroup_per_zone *mz_node;
727 mz->usage_in_excess = new_usage_in_excess;
728 if (!mz->usage_in_excess)
732 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
734 if (mz->usage_in_excess < mz_node->usage_in_excess)
737 * We can't avoid mem cgroups that are over their soft
738 * limit by the same amount
740 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
743 rb_link_node(&mz->tree_node, parent, p);
744 rb_insert_color(&mz->tree_node, &mctz->rb_root);
749 __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
750 struct mem_cgroup_per_zone *mz,
751 struct mem_cgroup_tree_per_zone *mctz)
755 rb_erase(&mz->tree_node, &mctz->rb_root);
760 mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
761 struct mem_cgroup_per_zone *mz,
762 struct mem_cgroup_tree_per_zone *mctz)
764 spin_lock(&mctz->lock);
765 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
766 spin_unlock(&mctz->lock);
770 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
772 unsigned long long excess;
773 struct mem_cgroup_per_zone *mz;
774 struct mem_cgroup_tree_per_zone *mctz;
775 int nid = page_to_nid(page);
776 int zid = page_zonenum(page);
777 mctz = soft_limit_tree_from_page(page);
780 * Necessary to update all ancestors when hierarchy is used.
781 * because their event counter is not touched.
783 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
784 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
785 excess = res_counter_soft_limit_excess(&memcg->res);
787 * We have to update the tree if mz is on RB-tree or
788 * mem is over its softlimit.
790 if (excess || mz->on_tree) {
791 spin_lock(&mctz->lock);
792 /* if on-tree, remove it */
794 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
796 * Insert again. mz->usage_in_excess will be updated.
797 * If excess is 0, no tree ops.
799 __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
800 spin_unlock(&mctz->lock);
805 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
808 struct mem_cgroup_per_zone *mz;
809 struct mem_cgroup_tree_per_zone *mctz;
811 for_each_node(node) {
812 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
813 mz = mem_cgroup_zoneinfo(memcg, node, zone);
814 mctz = soft_limit_tree_node_zone(node, zone);
815 mem_cgroup_remove_exceeded(memcg, mz, mctz);
820 static struct mem_cgroup_per_zone *
821 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
823 struct rb_node *rightmost = NULL;
824 struct mem_cgroup_per_zone *mz;
828 rightmost = rb_last(&mctz->rb_root);
830 goto done; /* Nothing to reclaim from */
832 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
834 * Remove the node now but someone else can add it back,
835 * we will to add it back at the end of reclaim to its correct
836 * position in the tree.
838 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
839 if (!res_counter_soft_limit_excess(&mz->memcg->res) ||
840 !css_tryget(&mz->memcg->css))
846 static struct mem_cgroup_per_zone *
847 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
849 struct mem_cgroup_per_zone *mz;
851 spin_lock(&mctz->lock);
852 mz = __mem_cgroup_largest_soft_limit_node(mctz);
853 spin_unlock(&mctz->lock);
858 * Implementation Note: reading percpu statistics for memcg.
860 * Both of vmstat[] and percpu_counter has threshold and do periodic
861 * synchronization to implement "quick" read. There are trade-off between
862 * reading cost and precision of value. Then, we may have a chance to implement
863 * a periodic synchronizion of counter in memcg's counter.
865 * But this _read() function is used for user interface now. The user accounts
866 * memory usage by memory cgroup and he _always_ requires exact value because
867 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
868 * have to visit all online cpus and make sum. So, for now, unnecessary
869 * synchronization is not implemented. (just implemented for cpu hotplug)
871 * If there are kernel internal actions which can make use of some not-exact
872 * value, and reading all cpu value can be performance bottleneck in some
873 * common workload, threashold and synchonization as vmstat[] should be
876 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
877 enum mem_cgroup_stat_index idx)
883 for_each_online_cpu(cpu)
884 val += per_cpu(memcg->stat->count[idx], cpu);
885 #ifdef CONFIG_HOTPLUG_CPU
886 spin_lock(&memcg->pcp_counter_lock);
887 val += memcg->nocpu_base.count[idx];
888 spin_unlock(&memcg->pcp_counter_lock);
894 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
897 int val = (charge) ? 1 : -1;
898 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
901 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
902 enum mem_cgroup_events_index idx)
904 unsigned long val = 0;
908 for_each_online_cpu(cpu)
909 val += per_cpu(memcg->stat->events[idx], cpu);
910 #ifdef CONFIG_HOTPLUG_CPU
911 spin_lock(&memcg->pcp_counter_lock);
912 val += memcg->nocpu_base.events[idx];
913 spin_unlock(&memcg->pcp_counter_lock);
919 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
921 bool anon, int nr_pages)
924 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
925 * counted as CACHE even if it's on ANON LRU.
928 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
931 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
934 if (PageTransHuge(page))
935 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
938 /* pagein of a big page is an event. So, ignore page size */
940 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
942 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
943 nr_pages = -nr_pages; /* for event */
946 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
950 mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
952 struct mem_cgroup_per_zone *mz;
954 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
955 return mz->lru_size[lru];
959 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
960 unsigned int lru_mask)
962 struct mem_cgroup_per_zone *mz;
964 unsigned long ret = 0;
966 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
969 if (BIT(lru) & lru_mask)
970 ret += mz->lru_size[lru];
976 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
977 int nid, unsigned int lru_mask)
982 for (zid = 0; zid < MAX_NR_ZONES; zid++)
983 total += mem_cgroup_zone_nr_lru_pages(memcg,
989 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
990 unsigned int lru_mask)
995 for_each_node_state(nid, N_MEMORY)
996 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
1000 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
1001 enum mem_cgroup_events_target target)
1003 unsigned long val, next;
1005 val = __this_cpu_read(memcg->stat->nr_page_events);
1006 next = __this_cpu_read(memcg->stat->targets[target]);
1007 /* from time_after() in jiffies.h */
1008 if ((long)next - (long)val < 0) {
1010 case MEM_CGROUP_TARGET_THRESH:
1011 next = val + THRESHOLDS_EVENTS_TARGET;
1013 case MEM_CGROUP_TARGET_SOFTLIMIT:
1014 next = val + SOFTLIMIT_EVENTS_TARGET;
1016 case MEM_CGROUP_TARGET_NUMAINFO:
1017 next = val + NUMAINFO_EVENTS_TARGET;
1022 __this_cpu_write(memcg->stat->targets[target], next);
1029 * Check events in order.
1032 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
1035 /* threshold event is triggered in finer grain than soft limit */
1036 if (unlikely(mem_cgroup_event_ratelimit(memcg,
1037 MEM_CGROUP_TARGET_THRESH))) {
1039 bool do_numainfo __maybe_unused;
1041 do_softlimit = mem_cgroup_event_ratelimit(memcg,
1042 MEM_CGROUP_TARGET_SOFTLIMIT);
1043 #if MAX_NUMNODES > 1
1044 do_numainfo = mem_cgroup_event_ratelimit(memcg,
1045 MEM_CGROUP_TARGET_NUMAINFO);
1049 mem_cgroup_threshold(memcg);
1050 if (unlikely(do_softlimit))
1051 mem_cgroup_update_tree(memcg, page);
1052 #if MAX_NUMNODES > 1
1053 if (unlikely(do_numainfo))
1054 atomic_inc(&memcg->numainfo_events);
1060 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1063 * mm_update_next_owner() may clear mm->owner to NULL
1064 * if it races with swapoff, page migration, etc.
1065 * So this can be called with p == NULL.
1070 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
1073 static struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
1075 struct mem_cgroup *memcg = NULL;
1079 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1080 if (unlikely(!memcg))
1081 memcg = root_mem_cgroup;
1082 } while (!css_tryget(&memcg->css));
1088 * Returns a next (in a pre-order walk) alive memcg (with elevated css
1089 * ref. count) or NULL if the whole root's subtree has been visited.
1091 * helper function to be used by mem_cgroup_iter
1093 static struct mem_cgroup *__mem_cgroup_iter_next(struct mem_cgroup *root,
1094 struct mem_cgroup *last_visited)
1096 struct cgroup_subsys_state *prev_css, *next_css;
1098 prev_css = last_visited ? &last_visited->css : NULL;
1100 next_css = css_next_descendant_pre(prev_css, &root->css);
1103 * Even if we found a group we have to make sure it is
1104 * alive. css && !memcg means that the groups should be
1105 * skipped and we should continue the tree walk.
1106 * last_visited css is safe to use because it is
1107 * protected by css_get and the tree walk is rcu safe.
1109 * We do not take a reference on the root of the tree walk
1110 * because we might race with the root removal when it would
1111 * be the only node in the iterated hierarchy and mem_cgroup_iter
1112 * would end up in an endless loop because it expects that at
1113 * least one valid node will be returned. Root cannot disappear
1114 * because caller of the iterator should hold it already so
1115 * skipping css reference should be safe.
1118 if ((next_css == &root->css) ||
1119 ((next_css->flags & CSS_ONLINE) && css_tryget(next_css)))
1120 return mem_cgroup_from_css(next_css);
1122 prev_css = next_css;
1129 static void mem_cgroup_iter_invalidate(struct mem_cgroup *root)
1132 * When a group in the hierarchy below root is destroyed, the
1133 * hierarchy iterator can no longer be trusted since it might
1134 * have pointed to the destroyed group. Invalidate it.
1136 atomic_inc(&root->dead_count);
1139 static struct mem_cgroup *
1140 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter *iter,
1141 struct mem_cgroup *root,
1144 struct mem_cgroup *position = NULL;
1146 * A cgroup destruction happens in two stages: offlining and
1147 * release. They are separated by a RCU grace period.
1149 * If the iterator is valid, we may still race with an
1150 * offlining. The RCU lock ensures the object won't be
1151 * released, tryget will fail if we lost the race.
1153 *sequence = atomic_read(&root->dead_count);
1154 if (iter->last_dead_count == *sequence) {
1156 position = iter->last_visited;
1159 * We cannot take a reference to root because we might race
1160 * with root removal and returning NULL would end up in
1161 * an endless loop on the iterator user level when root
1162 * would be returned all the time.
1164 if (position && position != root &&
1165 !css_tryget(&position->css))
1171 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter *iter,
1172 struct mem_cgroup *last_visited,
1173 struct mem_cgroup *new_position,
1174 struct mem_cgroup *root,
1177 /* root reference counting symmetric to mem_cgroup_iter_load */
1178 if (last_visited && last_visited != root)
1179 css_put(&last_visited->css);
1181 * We store the sequence count from the time @last_visited was
1182 * loaded successfully instead of rereading it here so that we
1183 * don't lose destruction events in between. We could have
1184 * raced with the destruction of @new_position after all.
1186 iter->last_visited = new_position;
1188 iter->last_dead_count = sequence;
1192 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1193 * @root: hierarchy root
1194 * @prev: previously returned memcg, NULL on first invocation
1195 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1197 * Returns references to children of the hierarchy below @root, or
1198 * @root itself, or %NULL after a full round-trip.
1200 * Caller must pass the return value in @prev on subsequent
1201 * invocations for reference counting, or use mem_cgroup_iter_break()
1202 * to cancel a hierarchy walk before the round-trip is complete.
1204 * Reclaimers can specify a zone and a priority level in @reclaim to
1205 * divide up the memcgs in the hierarchy among all concurrent
1206 * reclaimers operating on the same zone and priority.
1208 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1209 struct mem_cgroup *prev,
1210 struct mem_cgroup_reclaim_cookie *reclaim)
1212 struct mem_cgroup *memcg = NULL;
1213 struct mem_cgroup *last_visited = NULL;
1215 if (mem_cgroup_disabled())
1219 root = root_mem_cgroup;
1221 if (prev && !reclaim)
1222 last_visited = prev;
1224 if (!root->use_hierarchy && root != root_mem_cgroup) {
1232 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1233 int uninitialized_var(seq);
1236 int nid = zone_to_nid(reclaim->zone);
1237 int zid = zone_idx(reclaim->zone);
1238 struct mem_cgroup_per_zone *mz;
1240 mz = mem_cgroup_zoneinfo(root, nid, zid);
1241 iter = &mz->reclaim_iter[reclaim->priority];
1242 if (prev && reclaim->generation != iter->generation) {
1243 iter->last_visited = NULL;
1247 last_visited = mem_cgroup_iter_load(iter, root, &seq);
1250 memcg = __mem_cgroup_iter_next(root, last_visited);
1253 mem_cgroup_iter_update(iter, last_visited, memcg, root,
1258 else if (!prev && memcg)
1259 reclaim->generation = iter->generation;
1268 if (prev && prev != root)
1269 css_put(&prev->css);
1275 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1276 * @root: hierarchy root
1277 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1279 void mem_cgroup_iter_break(struct mem_cgroup *root,
1280 struct mem_cgroup *prev)
1283 root = root_mem_cgroup;
1284 if (prev && prev != root)
1285 css_put(&prev->css);
1289 * Iteration constructs for visiting all cgroups (under a tree). If
1290 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1291 * be used for reference counting.
1293 #define for_each_mem_cgroup_tree(iter, root) \
1294 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1296 iter = mem_cgroup_iter(root, iter, NULL))
1298 #define for_each_mem_cgroup(iter) \
1299 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1301 iter = mem_cgroup_iter(NULL, iter, NULL))
1303 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1305 struct mem_cgroup *memcg;
1308 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1309 if (unlikely(!memcg))
1314 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1317 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1325 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1328 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1329 * @zone: zone of the wanted lruvec
1330 * @memcg: memcg of the wanted lruvec
1332 * Returns the lru list vector holding pages for the given @zone and
1333 * @mem. This can be the global zone lruvec, if the memory controller
1336 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1337 struct mem_cgroup *memcg)
1339 struct mem_cgroup_per_zone *mz;
1340 struct lruvec *lruvec;
1342 if (mem_cgroup_disabled()) {
1343 lruvec = &zone->lruvec;
1347 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1348 lruvec = &mz->lruvec;
1351 * Since a node can be onlined after the mem_cgroup was created,
1352 * we have to be prepared to initialize lruvec->zone here;
1353 * and if offlined then reonlined, we need to reinitialize it.
1355 if (unlikely(lruvec->zone != zone))
1356 lruvec->zone = zone;
1361 * Following LRU functions are allowed to be used without PCG_LOCK.
1362 * Operations are called by routine of global LRU independently from memcg.
1363 * What we have to take care of here is validness of pc->mem_cgroup.
1365 * Changes to pc->mem_cgroup happens when
1368 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1369 * It is added to LRU before charge.
1370 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1371 * When moving account, the page is not on LRU. It's isolated.
1375 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1377 * @zone: zone of the page
1379 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1381 struct mem_cgroup_per_zone *mz;
1382 struct mem_cgroup *memcg;
1383 struct page_cgroup *pc;
1384 struct lruvec *lruvec;
1386 if (mem_cgroup_disabled()) {
1387 lruvec = &zone->lruvec;
1391 pc = lookup_page_cgroup(page);
1392 memcg = pc->mem_cgroup;
1395 * Surreptitiously switch any uncharged offlist page to root:
1396 * an uncharged page off lru does nothing to secure
1397 * its former mem_cgroup from sudden removal.
1399 * Our caller holds lru_lock, and PageCgroupUsed is updated
1400 * under page_cgroup lock: between them, they make all uses
1401 * of pc->mem_cgroup safe.
1403 if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1404 pc->mem_cgroup = memcg = root_mem_cgroup;
1406 mz = page_cgroup_zoneinfo(memcg, page);
1407 lruvec = &mz->lruvec;
1410 * Since a node can be onlined after the mem_cgroup was created,
1411 * we have to be prepared to initialize lruvec->zone here;
1412 * and if offlined then reonlined, we need to reinitialize it.
1414 if (unlikely(lruvec->zone != zone))
1415 lruvec->zone = zone;
1420 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1421 * @lruvec: mem_cgroup per zone lru vector
1422 * @lru: index of lru list the page is sitting on
1423 * @nr_pages: positive when adding or negative when removing
1425 * This function must be called when a page is added to or removed from an
1428 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1431 struct mem_cgroup_per_zone *mz;
1432 unsigned long *lru_size;
1434 if (mem_cgroup_disabled())
1437 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1438 lru_size = mz->lru_size + lru;
1439 *lru_size += nr_pages;
1440 VM_BUG_ON((long)(*lru_size) < 0);
1444 * Checks whether given mem is same or in the root_mem_cgroup's
1447 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1448 struct mem_cgroup *memcg)
1450 if (root_memcg == memcg)
1452 if (!root_memcg->use_hierarchy || !memcg)
1454 return cgroup_is_descendant(memcg->css.cgroup, root_memcg->css.cgroup);
1457 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1458 struct mem_cgroup *memcg)
1463 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1468 bool task_in_mem_cgroup(struct task_struct *task,
1469 const struct mem_cgroup *memcg)
1471 struct mem_cgroup *curr = NULL;
1472 struct task_struct *p;
1475 p = find_lock_task_mm(task);
1477 curr = get_mem_cgroup_from_mm(p->mm);
1481 * All threads may have already detached their mm's, but the oom
1482 * killer still needs to detect if they have already been oom
1483 * killed to prevent needlessly killing additional tasks.
1486 curr = mem_cgroup_from_task(task);
1488 css_get(&curr->css);
1492 * We should check use_hierarchy of "memcg" not "curr". Because checking
1493 * use_hierarchy of "curr" here make this function true if hierarchy is
1494 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1495 * hierarchy(even if use_hierarchy is disabled in "memcg").
1497 ret = mem_cgroup_same_or_subtree(memcg, curr);
1498 css_put(&curr->css);
1502 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1504 unsigned long inactive_ratio;
1505 unsigned long inactive;
1506 unsigned long active;
1509 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1510 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1512 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1514 inactive_ratio = int_sqrt(10 * gb);
1518 return inactive * inactive_ratio < active;
1521 #define mem_cgroup_from_res_counter(counter, member) \
1522 container_of(counter, struct mem_cgroup, member)
1525 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1526 * @memcg: the memory cgroup
1528 * Returns the maximum amount of memory @mem can be charged with, in
1531 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1533 unsigned long long margin;
1535 margin = res_counter_margin(&memcg->res);
1536 if (do_swap_account)
1537 margin = min(margin, res_counter_margin(&memcg->memsw));
1538 return margin >> PAGE_SHIFT;
1541 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1544 if (!css_parent(&memcg->css))
1545 return vm_swappiness;
1547 return memcg->swappiness;
1551 * memcg->moving_account is used for checking possibility that some thread is
1552 * calling move_account(). When a thread on CPU-A starts moving pages under
1553 * a memcg, other threads should check memcg->moving_account under
1554 * rcu_read_lock(), like this:
1558 * memcg->moving_account+1 if (memcg->mocing_account)
1560 * synchronize_rcu() update something.
1565 /* for quick checking without looking up memcg */
1566 atomic_t memcg_moving __read_mostly;
1568 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1570 atomic_inc(&memcg_moving);
1571 atomic_inc(&memcg->moving_account);
1575 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1578 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1579 * We check NULL in callee rather than caller.
1582 atomic_dec(&memcg_moving);
1583 atomic_dec(&memcg->moving_account);
1588 * A routine for checking "mem" is under move_account() or not.
1590 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1591 * moving cgroups. This is for waiting at high-memory pressure
1594 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1596 struct mem_cgroup *from;
1597 struct mem_cgroup *to;
1600 * Unlike task_move routines, we access mc.to, mc.from not under
1601 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1603 spin_lock(&mc.lock);
1609 ret = mem_cgroup_same_or_subtree(memcg, from)
1610 || mem_cgroup_same_or_subtree(memcg, to);
1612 spin_unlock(&mc.lock);
1616 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1618 if (mc.moving_task && current != mc.moving_task) {
1619 if (mem_cgroup_under_move(memcg)) {
1621 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1622 /* moving charge context might have finished. */
1625 finish_wait(&mc.waitq, &wait);
1633 * Take this lock when
1634 * - a code tries to modify page's memcg while it's USED.
1635 * - a code tries to modify page state accounting in a memcg.
1637 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1638 unsigned long *flags)
1640 spin_lock_irqsave(&memcg->move_lock, *flags);
1643 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1644 unsigned long *flags)
1646 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1649 #define K(x) ((x) << (PAGE_SHIFT-10))
1651 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1652 * @memcg: The memory cgroup that went over limit
1653 * @p: Task that is going to be killed
1655 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1658 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1660 /* oom_info_lock ensures that parallel ooms do not interleave */
1661 static DEFINE_MUTEX(oom_info_lock);
1662 struct mem_cgroup *iter;
1668 mutex_lock(&oom_info_lock);
1671 pr_info("Task in ");
1672 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1673 pr_info(" killed as a result of limit of ");
1674 pr_cont_cgroup_path(memcg->css.cgroup);
1679 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1680 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1681 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1682 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1683 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1684 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1685 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1686 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1687 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1688 res_counter_read_u64(&memcg->kmem, RES_USAGE) >> 10,
1689 res_counter_read_u64(&memcg->kmem, RES_LIMIT) >> 10,
1690 res_counter_read_u64(&memcg->kmem, RES_FAILCNT));
1692 for_each_mem_cgroup_tree(iter, memcg) {
1693 pr_info("Memory cgroup stats for ");
1694 pr_cont_cgroup_path(iter->css.cgroup);
1697 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1698 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1700 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1701 K(mem_cgroup_read_stat(iter, i)));
1704 for (i = 0; i < NR_LRU_LISTS; i++)
1705 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1706 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1710 mutex_unlock(&oom_info_lock);
1714 * This function returns the number of memcg under hierarchy tree. Returns
1715 * 1(self count) if no children.
1717 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1720 struct mem_cgroup *iter;
1722 for_each_mem_cgroup_tree(iter, memcg)
1728 * Return the memory (and swap, if configured) limit for a memcg.
1730 static u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1734 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1737 * Do not consider swap space if we cannot swap due to swappiness
1739 if (mem_cgroup_swappiness(memcg)) {
1742 limit += total_swap_pages << PAGE_SHIFT;
1743 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1746 * If memsw is finite and limits the amount of swap space
1747 * available to this memcg, return that limit.
1749 limit = min(limit, memsw);
1755 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1758 struct mem_cgroup *iter;
1759 unsigned long chosen_points = 0;
1760 unsigned long totalpages;
1761 unsigned int points = 0;
1762 struct task_struct *chosen = NULL;
1765 * If current has a pending SIGKILL or is exiting, then automatically
1766 * select it. The goal is to allow it to allocate so that it may
1767 * quickly exit and free its memory.
1769 if (fatal_signal_pending(current) || current->flags & PF_EXITING) {
1770 set_thread_flag(TIF_MEMDIE);
1774 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1775 totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
1776 for_each_mem_cgroup_tree(iter, memcg) {
1777 struct css_task_iter it;
1778 struct task_struct *task;
1780 css_task_iter_start(&iter->css, &it);
1781 while ((task = css_task_iter_next(&it))) {
1782 switch (oom_scan_process_thread(task, totalpages, NULL,
1784 case OOM_SCAN_SELECT:
1786 put_task_struct(chosen);
1788 chosen_points = ULONG_MAX;
1789 get_task_struct(chosen);
1791 case OOM_SCAN_CONTINUE:
1793 case OOM_SCAN_ABORT:
1794 css_task_iter_end(&it);
1795 mem_cgroup_iter_break(memcg, iter);
1797 put_task_struct(chosen);
1802 points = oom_badness(task, memcg, NULL, totalpages);
1803 if (!points || points < chosen_points)
1805 /* Prefer thread group leaders for display purposes */
1806 if (points == chosen_points &&
1807 thread_group_leader(chosen))
1811 put_task_struct(chosen);
1813 chosen_points = points;
1814 get_task_struct(chosen);
1816 css_task_iter_end(&it);
1821 points = chosen_points * 1000 / totalpages;
1822 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1823 NULL, "Memory cgroup out of memory");
1826 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1828 unsigned long flags)
1830 unsigned long total = 0;
1831 bool noswap = false;
1834 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1836 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1839 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1841 drain_all_stock_async(memcg);
1842 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1844 * Allow limit shrinkers, which are triggered directly
1845 * by userspace, to catch signals and stop reclaim
1846 * after minimal progress, regardless of the margin.
1848 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1850 if (mem_cgroup_margin(memcg))
1853 * If nothing was reclaimed after two attempts, there
1854 * may be no reclaimable pages in this hierarchy.
1863 * test_mem_cgroup_node_reclaimable
1864 * @memcg: the target memcg
1865 * @nid: the node ID to be checked.
1866 * @noswap : specify true here if the user wants flle only information.
1868 * This function returns whether the specified memcg contains any
1869 * reclaimable pages on a node. Returns true if there are any reclaimable
1870 * pages in the node.
1872 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1873 int nid, bool noswap)
1875 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1877 if (noswap || !total_swap_pages)
1879 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1884 #if MAX_NUMNODES > 1
1887 * Always updating the nodemask is not very good - even if we have an empty
1888 * list or the wrong list here, we can start from some node and traverse all
1889 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1892 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1896 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1897 * pagein/pageout changes since the last update.
1899 if (!atomic_read(&memcg->numainfo_events))
1901 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1904 /* make a nodemask where this memcg uses memory from */
1905 memcg->scan_nodes = node_states[N_MEMORY];
1907 for_each_node_mask(nid, node_states[N_MEMORY]) {
1909 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1910 node_clear(nid, memcg->scan_nodes);
1913 atomic_set(&memcg->numainfo_events, 0);
1914 atomic_set(&memcg->numainfo_updating, 0);
1918 * Selecting a node where we start reclaim from. Because what we need is just
1919 * reducing usage counter, start from anywhere is O,K. Considering
1920 * memory reclaim from current node, there are pros. and cons.
1922 * Freeing memory from current node means freeing memory from a node which
1923 * we'll use or we've used. So, it may make LRU bad. And if several threads
1924 * hit limits, it will see a contention on a node. But freeing from remote
1925 * node means more costs for memory reclaim because of memory latency.
1927 * Now, we use round-robin. Better algorithm is welcomed.
1929 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1933 mem_cgroup_may_update_nodemask(memcg);
1934 node = memcg->last_scanned_node;
1936 node = next_node(node, memcg->scan_nodes);
1937 if (node == MAX_NUMNODES)
1938 node = first_node(memcg->scan_nodes);
1940 * We call this when we hit limit, not when pages are added to LRU.
1941 * No LRU may hold pages because all pages are UNEVICTABLE or
1942 * memcg is too small and all pages are not on LRU. In that case,
1943 * we use curret node.
1945 if (unlikely(node == MAX_NUMNODES))
1946 node = numa_node_id();
1948 memcg->last_scanned_node = node;
1953 * Check all nodes whether it contains reclaimable pages or not.
1954 * For quick scan, we make use of scan_nodes. This will allow us to skip
1955 * unused nodes. But scan_nodes is lazily updated and may not cotain
1956 * enough new information. We need to do double check.
1958 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1963 * quick check...making use of scan_node.
1964 * We can skip unused nodes.
1966 if (!nodes_empty(memcg->scan_nodes)) {
1967 for (nid = first_node(memcg->scan_nodes);
1969 nid = next_node(nid, memcg->scan_nodes)) {
1971 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1976 * Check rest of nodes.
1978 for_each_node_state(nid, N_MEMORY) {
1979 if (node_isset(nid, memcg->scan_nodes))
1981 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1988 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1993 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1995 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
1999 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
2002 unsigned long *total_scanned)
2004 struct mem_cgroup *victim = NULL;
2007 unsigned long excess;
2008 unsigned long nr_scanned;
2009 struct mem_cgroup_reclaim_cookie reclaim = {
2014 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
2017 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
2022 * If we have not been able to reclaim
2023 * anything, it might because there are
2024 * no reclaimable pages under this hierarchy
2029 * We want to do more targeted reclaim.
2030 * excess >> 2 is not to excessive so as to
2031 * reclaim too much, nor too less that we keep
2032 * coming back to reclaim from this cgroup
2034 if (total >= (excess >> 2) ||
2035 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
2040 if (!mem_cgroup_reclaimable(victim, false))
2042 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
2044 *total_scanned += nr_scanned;
2045 if (!res_counter_soft_limit_excess(&root_memcg->res))
2048 mem_cgroup_iter_break(root_memcg, victim);
2052 #ifdef CONFIG_LOCKDEP
2053 static struct lockdep_map memcg_oom_lock_dep_map = {
2054 .name = "memcg_oom_lock",
2058 static DEFINE_SPINLOCK(memcg_oom_lock);
2061 * Check OOM-Killer is already running under our hierarchy.
2062 * If someone is running, return false.
2064 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
2066 struct mem_cgroup *iter, *failed = NULL;
2068 spin_lock(&memcg_oom_lock);
2070 for_each_mem_cgroup_tree(iter, memcg) {
2071 if (iter->oom_lock) {
2073 * this subtree of our hierarchy is already locked
2074 * so we cannot give a lock.
2077 mem_cgroup_iter_break(memcg, iter);
2080 iter->oom_lock = true;
2085 * OK, we failed to lock the whole subtree so we have
2086 * to clean up what we set up to the failing subtree
2088 for_each_mem_cgroup_tree(iter, memcg) {
2089 if (iter == failed) {
2090 mem_cgroup_iter_break(memcg, iter);
2093 iter->oom_lock = false;
2096 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
2098 spin_unlock(&memcg_oom_lock);
2103 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
2105 struct mem_cgroup *iter;
2107 spin_lock(&memcg_oom_lock);
2108 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
2109 for_each_mem_cgroup_tree(iter, memcg)
2110 iter->oom_lock = false;
2111 spin_unlock(&memcg_oom_lock);
2114 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
2116 struct mem_cgroup *iter;
2118 for_each_mem_cgroup_tree(iter, memcg)
2119 atomic_inc(&iter->under_oom);
2122 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
2124 struct mem_cgroup *iter;
2127 * When a new child is created while the hierarchy is under oom,
2128 * mem_cgroup_oom_lock() may not be called. We have to use
2129 * atomic_add_unless() here.
2131 for_each_mem_cgroup_tree(iter, memcg)
2132 atomic_add_unless(&iter->under_oom, -1, 0);
2135 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
2137 struct oom_wait_info {
2138 struct mem_cgroup *memcg;
2142 static int memcg_oom_wake_function(wait_queue_t *wait,
2143 unsigned mode, int sync, void *arg)
2145 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
2146 struct mem_cgroup *oom_wait_memcg;
2147 struct oom_wait_info *oom_wait_info;
2149 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
2150 oom_wait_memcg = oom_wait_info->memcg;
2153 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2154 * Then we can use css_is_ancestor without taking care of RCU.
2156 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
2157 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
2159 return autoremove_wake_function(wait, mode, sync, arg);
2162 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
2164 atomic_inc(&memcg->oom_wakeups);
2165 /* for filtering, pass "memcg" as argument. */
2166 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
2169 static void memcg_oom_recover(struct mem_cgroup *memcg)
2171 if (memcg && atomic_read(&memcg->under_oom))
2172 memcg_wakeup_oom(memcg);
2175 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
2177 if (!current->memcg_oom.may_oom)
2180 * We are in the middle of the charge context here, so we
2181 * don't want to block when potentially sitting on a callstack
2182 * that holds all kinds of filesystem and mm locks.
2184 * Also, the caller may handle a failed allocation gracefully
2185 * (like optional page cache readahead) and so an OOM killer
2186 * invocation might not even be necessary.
2188 * That's why we don't do anything here except remember the
2189 * OOM context and then deal with it at the end of the page
2190 * fault when the stack is unwound, the locks are released,
2191 * and when we know whether the fault was overall successful.
2193 css_get(&memcg->css);
2194 current->memcg_oom.memcg = memcg;
2195 current->memcg_oom.gfp_mask = mask;
2196 current->memcg_oom.order = order;
2200 * mem_cgroup_oom_synchronize - complete memcg OOM handling
2201 * @handle: actually kill/wait or just clean up the OOM state
2203 * This has to be called at the end of a page fault if the memcg OOM
2204 * handler was enabled.
2206 * Memcg supports userspace OOM handling where failed allocations must
2207 * sleep on a waitqueue until the userspace task resolves the
2208 * situation. Sleeping directly in the charge context with all kinds
2209 * of locks held is not a good idea, instead we remember an OOM state
2210 * in the task and mem_cgroup_oom_synchronize() has to be called at
2211 * the end of the page fault to complete the OOM handling.
2213 * Returns %true if an ongoing memcg OOM situation was detected and
2214 * completed, %false otherwise.
2216 bool mem_cgroup_oom_synchronize(bool handle)
2218 struct mem_cgroup *memcg = current->memcg_oom.memcg;
2219 struct oom_wait_info owait;
2222 /* OOM is global, do not handle */
2229 owait.memcg = memcg;
2230 owait.wait.flags = 0;
2231 owait.wait.func = memcg_oom_wake_function;
2232 owait.wait.private = current;
2233 INIT_LIST_HEAD(&owait.wait.task_list);
2235 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2236 mem_cgroup_mark_under_oom(memcg);
2238 locked = mem_cgroup_oom_trylock(memcg);
2241 mem_cgroup_oom_notify(memcg);
2243 if (locked && !memcg->oom_kill_disable) {
2244 mem_cgroup_unmark_under_oom(memcg);
2245 finish_wait(&memcg_oom_waitq, &owait.wait);
2246 mem_cgroup_out_of_memory(memcg, current->memcg_oom.gfp_mask,
2247 current->memcg_oom.order);
2250 mem_cgroup_unmark_under_oom(memcg);
2251 finish_wait(&memcg_oom_waitq, &owait.wait);
2255 mem_cgroup_oom_unlock(memcg);
2257 * There is no guarantee that an OOM-lock contender
2258 * sees the wakeups triggered by the OOM kill
2259 * uncharges. Wake any sleepers explicitely.
2261 memcg_oom_recover(memcg);
2264 current->memcg_oom.memcg = NULL;
2265 css_put(&memcg->css);
2270 * Used to update mapped file or writeback or other statistics.
2272 * Notes: Race condition
2274 * We usually use lock_page_cgroup() for accessing page_cgroup member but
2275 * it tends to be costly. But considering some conditions, we doesn't need
2276 * to do so _always_.
2278 * Considering "charge", lock_page_cgroup() is not required because all
2279 * file-stat operations happen after a page is attached to radix-tree. There
2280 * are no race with "charge".
2282 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2283 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2284 * if there are race with "uncharge". Statistics itself is properly handled
2287 * Considering "move", this is an only case we see a race. To make the race
2288 * small, we check memcg->moving_account and detect there are possibility
2289 * of race or not. If there is, we take a lock.
2292 void __mem_cgroup_begin_update_page_stat(struct page *page,
2293 bool *locked, unsigned long *flags)
2295 struct mem_cgroup *memcg;
2296 struct page_cgroup *pc;
2298 pc = lookup_page_cgroup(page);
2300 memcg = pc->mem_cgroup;
2301 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2304 * If this memory cgroup is not under account moving, we don't
2305 * need to take move_lock_mem_cgroup(). Because we already hold
2306 * rcu_read_lock(), any calls to move_account will be delayed until
2307 * rcu_read_unlock().
2309 VM_BUG_ON(!rcu_read_lock_held());
2310 if (atomic_read(&memcg->moving_account) <= 0)
2313 move_lock_mem_cgroup(memcg, flags);
2314 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2315 move_unlock_mem_cgroup(memcg, flags);
2321 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
2323 struct page_cgroup *pc = lookup_page_cgroup(page);
2326 * It's guaranteed that pc->mem_cgroup never changes while
2327 * lock is held because a routine modifies pc->mem_cgroup
2328 * should take move_lock_mem_cgroup().
2330 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
2333 void mem_cgroup_update_page_stat(struct page *page,
2334 enum mem_cgroup_stat_index idx, int val)
2336 struct mem_cgroup *memcg;
2337 struct page_cgroup *pc = lookup_page_cgroup(page);
2338 unsigned long uninitialized_var(flags);
2340 if (mem_cgroup_disabled())
2343 VM_BUG_ON(!rcu_read_lock_held());
2344 memcg = pc->mem_cgroup;
2345 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2348 this_cpu_add(memcg->stat->count[idx], val);
2352 * size of first charge trial. "32" comes from vmscan.c's magic value.
2353 * TODO: maybe necessary to use big numbers in big irons.
2355 #define CHARGE_BATCH 32U
2356 struct memcg_stock_pcp {
2357 struct mem_cgroup *cached; /* this never be root cgroup */
2358 unsigned int nr_pages;
2359 struct work_struct work;
2360 unsigned long flags;
2361 #define FLUSHING_CACHED_CHARGE 0
2363 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2364 static DEFINE_MUTEX(percpu_charge_mutex);
2367 * consume_stock: Try to consume stocked charge on this cpu.
2368 * @memcg: memcg to consume from.
2369 * @nr_pages: how many pages to charge.
2371 * The charges will only happen if @memcg matches the current cpu's memcg
2372 * stock, and at least @nr_pages are available in that stock. Failure to
2373 * service an allocation will refill the stock.
2375 * returns true if successful, false otherwise.
2377 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2379 struct memcg_stock_pcp *stock;
2382 if (nr_pages > CHARGE_BATCH)
2385 stock = &get_cpu_var(memcg_stock);
2386 if (memcg == stock->cached && stock->nr_pages >= nr_pages)
2387 stock->nr_pages -= nr_pages;
2388 else /* need to call res_counter_charge */
2390 put_cpu_var(memcg_stock);
2395 * Returns stocks cached in percpu to res_counter and reset cached information.
2397 static void drain_stock(struct memcg_stock_pcp *stock)
2399 struct mem_cgroup *old = stock->cached;
2401 if (stock->nr_pages) {
2402 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2404 res_counter_uncharge(&old->res, bytes);
2405 if (do_swap_account)
2406 res_counter_uncharge(&old->memsw, bytes);
2407 stock->nr_pages = 0;
2409 stock->cached = NULL;
2413 * This must be called under preempt disabled or must be called by
2414 * a thread which is pinned to local cpu.
2416 static void drain_local_stock(struct work_struct *dummy)
2418 struct memcg_stock_pcp *stock = this_cpu_ptr(&memcg_stock);
2420 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2423 static void __init memcg_stock_init(void)
2427 for_each_possible_cpu(cpu) {
2428 struct memcg_stock_pcp *stock =
2429 &per_cpu(memcg_stock, cpu);
2430 INIT_WORK(&stock->work, drain_local_stock);
2435 * Cache charges(val) which is from res_counter, to local per_cpu area.
2436 * This will be consumed by consume_stock() function, later.
2438 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2440 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2442 if (stock->cached != memcg) { /* reset if necessary */
2444 stock->cached = memcg;
2446 stock->nr_pages += nr_pages;
2447 put_cpu_var(memcg_stock);
2451 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2452 * of the hierarchy under it. sync flag says whether we should block
2453 * until the work is done.
2455 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2459 /* Notify other cpus that system-wide "drain" is running */
2462 for_each_online_cpu(cpu) {
2463 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2464 struct mem_cgroup *memcg;
2466 memcg = stock->cached;
2467 if (!memcg || !stock->nr_pages)
2469 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2471 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2473 drain_local_stock(&stock->work);
2475 schedule_work_on(cpu, &stock->work);
2483 for_each_online_cpu(cpu) {
2484 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2485 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2486 flush_work(&stock->work);
2493 * Tries to drain stocked charges in other cpus. This function is asynchronous
2494 * and just put a work per cpu for draining localy on each cpu. Caller can
2495 * expects some charges will be back to res_counter later but cannot wait for
2498 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2501 * If someone calls draining, avoid adding more kworker runs.
2503 if (!mutex_trylock(&percpu_charge_mutex))
2505 drain_all_stock(root_memcg, false);
2506 mutex_unlock(&percpu_charge_mutex);
2509 /* This is a synchronous drain interface. */
2510 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2512 /* called when force_empty is called */
2513 mutex_lock(&percpu_charge_mutex);
2514 drain_all_stock(root_memcg, true);
2515 mutex_unlock(&percpu_charge_mutex);
2519 * This function drains percpu counter value from DEAD cpu and
2520 * move it to local cpu. Note that this function can be preempted.
2522 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2526 spin_lock(&memcg->pcp_counter_lock);
2527 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2528 long x = per_cpu(memcg->stat->count[i], cpu);
2530 per_cpu(memcg->stat->count[i], cpu) = 0;
2531 memcg->nocpu_base.count[i] += x;
2533 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2534 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2536 per_cpu(memcg->stat->events[i], cpu) = 0;
2537 memcg->nocpu_base.events[i] += x;
2539 spin_unlock(&memcg->pcp_counter_lock);
2542 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
2543 unsigned long action,
2546 int cpu = (unsigned long)hcpu;
2547 struct memcg_stock_pcp *stock;
2548 struct mem_cgroup *iter;
2550 if (action == CPU_ONLINE)
2553 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2556 for_each_mem_cgroup(iter)
2557 mem_cgroup_drain_pcp_counter(iter, cpu);
2559 stock = &per_cpu(memcg_stock, cpu);
2565 /* See mem_cgroup_try_charge() for details */
2567 CHARGE_OK, /* success */
2568 CHARGE_RETRY, /* need to retry but retry is not bad */
2569 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2570 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2573 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2574 unsigned int nr_pages, unsigned int min_pages,
2577 unsigned long csize = nr_pages * PAGE_SIZE;
2578 struct mem_cgroup *mem_over_limit;
2579 struct res_counter *fail_res;
2580 unsigned long flags = 0;
2583 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2586 if (!do_swap_account)
2588 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2592 res_counter_uncharge(&memcg->res, csize);
2593 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2594 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2596 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2598 * Never reclaim on behalf of optional batching, retry with a
2599 * single page instead.
2601 if (nr_pages > min_pages)
2602 return CHARGE_RETRY;
2604 if (!(gfp_mask & __GFP_WAIT))
2605 return CHARGE_WOULDBLOCK;
2607 if (gfp_mask & __GFP_NORETRY)
2608 return CHARGE_NOMEM;
2610 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2611 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2612 return CHARGE_RETRY;
2614 * Even though the limit is exceeded at this point, reclaim
2615 * may have been able to free some pages. Retry the charge
2616 * before killing the task.
2618 * Only for regular pages, though: huge pages are rather
2619 * unlikely to succeed so close to the limit, and we fall back
2620 * to regular pages anyway in case of failure.
2622 if (nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER) && ret)
2623 return CHARGE_RETRY;
2626 * At task move, charge accounts can be doubly counted. So, it's
2627 * better to wait until the end of task_move if something is going on.
2629 if (mem_cgroup_wait_acct_move(mem_over_limit))
2630 return CHARGE_RETRY;
2633 mem_cgroup_oom(mem_over_limit, gfp_mask, get_order(csize));
2635 return CHARGE_NOMEM;
2639 * mem_cgroup_try_charge - try charging a memcg
2640 * @memcg: memcg to charge
2641 * @nr_pages: number of pages to charge
2642 * @oom: trigger OOM if reclaim fails
2644 * Returns 0 if @memcg was charged successfully, -EINTR if the charge
2645 * was bypassed to root_mem_cgroup, and -ENOMEM if the charge failed.
2647 static int mem_cgroup_try_charge(struct mem_cgroup *memcg,
2649 unsigned int nr_pages,
2652 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2653 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2656 if (mem_cgroup_is_root(memcg))
2659 * Unlike in global OOM situations, memcg is not in a physical
2660 * memory shortage. Allow dying and OOM-killed tasks to
2661 * bypass the last charges so that they can exit quickly and
2662 * free their memory.
2664 if (unlikely(test_thread_flag(TIF_MEMDIE) ||
2665 fatal_signal_pending(current) ||
2666 current->flags & PF_EXITING))
2669 if (unlikely(task_in_memcg_oom(current)))
2672 if (gfp_mask & __GFP_NOFAIL)
2675 if (consume_stock(memcg, nr_pages))
2679 bool invoke_oom = oom && !nr_oom_retries;
2681 /* If killed, bypass charge */
2682 if (fatal_signal_pending(current))
2685 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch,
2686 nr_pages, invoke_oom);
2690 case CHARGE_RETRY: /* not in OOM situation but retry */
2693 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2695 case CHARGE_NOMEM: /* OOM routine works */
2696 if (!oom || invoke_oom)
2701 } while (ret != CHARGE_OK);
2703 if (batch > nr_pages)
2704 refill_stock(memcg, batch - nr_pages);
2708 if (!(gfp_mask & __GFP_NOFAIL))
2715 * mem_cgroup_try_charge_mm - try charging a mm
2716 * @mm: mm_struct to charge
2717 * @nr_pages: number of pages to charge
2718 * @oom: trigger OOM if reclaim fails
2720 * Returns the charged mem_cgroup associated with the given mm_struct or
2721 * NULL the charge failed.
2723 static struct mem_cgroup *mem_cgroup_try_charge_mm(struct mm_struct *mm,
2725 unsigned int nr_pages,
2729 struct mem_cgroup *memcg;
2732 memcg = get_mem_cgroup_from_mm(mm);
2733 ret = mem_cgroup_try_charge(memcg, gfp_mask, nr_pages, oom);
2734 css_put(&memcg->css);
2736 memcg = root_mem_cgroup;
2744 * Somemtimes we have to undo a charge we got by try_charge().
2745 * This function is for that and do uncharge, put css's refcnt.
2746 * gotten by try_charge().
2748 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2749 unsigned int nr_pages)
2751 if (!mem_cgroup_is_root(memcg)) {
2752 unsigned long bytes = nr_pages * PAGE_SIZE;
2754 res_counter_uncharge(&memcg->res, bytes);
2755 if (do_swap_account)
2756 res_counter_uncharge(&memcg->memsw, bytes);
2761 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2762 * This is useful when moving usage to parent cgroup.
2764 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2765 unsigned int nr_pages)
2767 unsigned long bytes = nr_pages * PAGE_SIZE;
2769 if (mem_cgroup_is_root(memcg))
2772 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2773 if (do_swap_account)
2774 res_counter_uncharge_until(&memcg->memsw,
2775 memcg->memsw.parent, bytes);
2779 * A helper function to get mem_cgroup from ID. must be called under
2780 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2781 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2782 * called against removed memcg.)
2784 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2786 /* ID 0 is unused ID */
2789 return mem_cgroup_from_id(id);
2792 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2794 struct mem_cgroup *memcg = NULL;
2795 struct page_cgroup *pc;
2799 VM_BUG_ON_PAGE(!PageLocked(page), page);
2801 pc = lookup_page_cgroup(page);
2802 lock_page_cgroup(pc);
2803 if (PageCgroupUsed(pc)) {
2804 memcg = pc->mem_cgroup;
2805 if (memcg && !css_tryget(&memcg->css))
2807 } else if (PageSwapCache(page)) {
2808 ent.val = page_private(page);
2809 id = lookup_swap_cgroup_id(ent);
2811 memcg = mem_cgroup_lookup(id);
2812 if (memcg && !css_tryget(&memcg->css))
2816 unlock_page_cgroup(pc);
2820 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2822 unsigned int nr_pages,
2823 enum charge_type ctype,
2826 struct page_cgroup *pc = lookup_page_cgroup(page);
2827 struct zone *uninitialized_var(zone);
2828 struct lruvec *lruvec;
2829 bool was_on_lru = false;
2832 lock_page_cgroup(pc);
2833 VM_BUG_ON_PAGE(PageCgroupUsed(pc), page);
2835 * we don't need page_cgroup_lock about tail pages, becase they are not
2836 * accessed by any other context at this point.
2840 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2841 * may already be on some other mem_cgroup's LRU. Take care of it.
2844 zone = page_zone(page);
2845 spin_lock_irq(&zone->lru_lock);
2846 if (PageLRU(page)) {
2847 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2849 del_page_from_lru_list(page, lruvec, page_lru(page));
2854 pc->mem_cgroup = memcg;
2856 * We access a page_cgroup asynchronously without lock_page_cgroup().
2857 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2858 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2859 * before USED bit, we need memory barrier here.
2860 * See mem_cgroup_add_lru_list(), etc.
2863 SetPageCgroupUsed(pc);
2867 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2868 VM_BUG_ON_PAGE(PageLRU(page), page);
2870 add_page_to_lru_list(page, lruvec, page_lru(page));
2872 spin_unlock_irq(&zone->lru_lock);
2875 if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
2880 mem_cgroup_charge_statistics(memcg, page, anon, nr_pages);
2881 unlock_page_cgroup(pc);
2884 * "charge_statistics" updated event counter. Then, check it.
2885 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2886 * if they exceeds softlimit.
2888 memcg_check_events(memcg, page);
2891 static DEFINE_MUTEX(set_limit_mutex);
2893 #ifdef CONFIG_MEMCG_KMEM
2895 * The memcg_slab_mutex is held whenever a per memcg kmem cache is created or
2896 * destroyed. It protects memcg_caches arrays and memcg_slab_caches lists.
2898 static DEFINE_MUTEX(memcg_slab_mutex);
2900 static DEFINE_MUTEX(activate_kmem_mutex);
2902 static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg)
2904 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) &&
2905 memcg_kmem_is_active(memcg);
2909 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2910 * in the memcg_cache_params struct.
2912 static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
2914 struct kmem_cache *cachep;
2916 VM_BUG_ON(p->is_root_cache);
2917 cachep = p->root_cache;
2918 return cache_from_memcg_idx(cachep, memcg_cache_id(p->memcg));
2921 #ifdef CONFIG_SLABINFO
2922 static int mem_cgroup_slabinfo_read(struct seq_file *m, void *v)
2924 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
2925 struct memcg_cache_params *params;
2927 if (!memcg_can_account_kmem(memcg))
2930 print_slabinfo_header(m);
2932 mutex_lock(&memcg_slab_mutex);
2933 list_for_each_entry(params, &memcg->memcg_slab_caches, list)
2934 cache_show(memcg_params_to_cache(params), m);
2935 mutex_unlock(&memcg_slab_mutex);
2941 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size)
2943 struct res_counter *fail_res;
2946 ret = res_counter_charge(&memcg->kmem, size, &fail_res);
2950 ret = mem_cgroup_try_charge(memcg, gfp, size >> PAGE_SHIFT,
2951 oom_gfp_allowed(gfp));
2952 if (ret == -EINTR) {
2954 * mem_cgroup_try_charge() chosed to bypass to root due to
2955 * OOM kill or fatal signal. Since our only options are to
2956 * either fail the allocation or charge it to this cgroup, do
2957 * it as a temporary condition. But we can't fail. From a
2958 * kmem/slab perspective, the cache has already been selected,
2959 * by mem_cgroup_kmem_get_cache(), so it is too late to change
2962 * This condition will only trigger if the task entered
2963 * memcg_charge_kmem in a sane state, but was OOM-killed during
2964 * mem_cgroup_try_charge() above. Tasks that were already
2965 * dying when the allocation triggers should have been already
2966 * directed to the root cgroup in memcontrol.h
2968 res_counter_charge_nofail(&memcg->res, size, &fail_res);
2969 if (do_swap_account)
2970 res_counter_charge_nofail(&memcg->memsw, size,
2974 res_counter_uncharge(&memcg->kmem, size);
2979 static void memcg_uncharge_kmem(struct mem_cgroup *memcg, u64 size)
2981 res_counter_uncharge(&memcg->res, size);
2982 if (do_swap_account)
2983 res_counter_uncharge(&memcg->memsw, size);
2986 if (res_counter_uncharge(&memcg->kmem, size))
2990 * Releases a reference taken in kmem_cgroup_css_offline in case
2991 * this last uncharge is racing with the offlining code or it is
2992 * outliving the memcg existence.
2994 * The memory barrier imposed by test&clear is paired with the
2995 * explicit one in memcg_kmem_mark_dead().
2997 if (memcg_kmem_test_and_clear_dead(memcg))
2998 css_put(&memcg->css);
3002 * helper for acessing a memcg's index. It will be used as an index in the
3003 * child cache array in kmem_cache, and also to derive its name. This function
3004 * will return -1 when this is not a kmem-limited memcg.
3006 int memcg_cache_id(struct mem_cgroup *memcg)
3008 return memcg ? memcg->kmemcg_id : -1;
3011 static size_t memcg_caches_array_size(int num_groups)
3014 if (num_groups <= 0)
3017 size = 2 * num_groups;
3018 if (size < MEMCG_CACHES_MIN_SIZE)
3019 size = MEMCG_CACHES_MIN_SIZE;
3020 else if (size > MEMCG_CACHES_MAX_SIZE)
3021 size = MEMCG_CACHES_MAX_SIZE;
3027 * We should update the current array size iff all caches updates succeed. This
3028 * can only be done from the slab side. The slab mutex needs to be held when
3031 void memcg_update_array_size(int num)
3033 if (num > memcg_limited_groups_array_size)
3034 memcg_limited_groups_array_size = memcg_caches_array_size(num);
3037 int memcg_update_cache_size(struct kmem_cache *s, int num_groups)
3039 struct memcg_cache_params *cur_params = s->memcg_params;
3041 VM_BUG_ON(!is_root_cache(s));
3043 if (num_groups > memcg_limited_groups_array_size) {
3045 struct memcg_cache_params *new_params;
3046 ssize_t size = memcg_caches_array_size(num_groups);
3048 size *= sizeof(void *);
3049 size += offsetof(struct memcg_cache_params, memcg_caches);
3051 new_params = kzalloc(size, GFP_KERNEL);
3055 new_params->is_root_cache = true;
3058 * There is the chance it will be bigger than
3059 * memcg_limited_groups_array_size, if we failed an allocation
3060 * in a cache, in which case all caches updated before it, will
3061 * have a bigger array.
3063 * But if that is the case, the data after
3064 * memcg_limited_groups_array_size is certainly unused
3066 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3067 if (!cur_params->memcg_caches[i])
3069 new_params->memcg_caches[i] =
3070 cur_params->memcg_caches[i];
3074 * Ideally, we would wait until all caches succeed, and only
3075 * then free the old one. But this is not worth the extra
3076 * pointer per-cache we'd have to have for this.
3078 * It is not a big deal if some caches are left with a size
3079 * bigger than the others. And all updates will reset this
3082 rcu_assign_pointer(s->memcg_params, new_params);
3084 kfree_rcu(cur_params, rcu_head);
3089 int memcg_alloc_cache_params(struct mem_cgroup *memcg, struct kmem_cache *s,
3090 struct kmem_cache *root_cache)
3094 if (!memcg_kmem_enabled())
3098 size = offsetof(struct memcg_cache_params, memcg_caches);
3099 size += memcg_limited_groups_array_size * sizeof(void *);
3101 size = sizeof(struct memcg_cache_params);
3103 s->memcg_params = kzalloc(size, GFP_KERNEL);
3104 if (!s->memcg_params)
3108 s->memcg_params->memcg = memcg;
3109 s->memcg_params->root_cache = root_cache;
3110 css_get(&memcg->css);
3112 s->memcg_params->is_root_cache = true;
3117 void memcg_free_cache_params(struct kmem_cache *s)
3119 if (!s->memcg_params)
3121 if (!s->memcg_params->is_root_cache)
3122 css_put(&s->memcg_params->memcg->css);
3123 kfree(s->memcg_params);
3126 static void memcg_register_cache(struct mem_cgroup *memcg,
3127 struct kmem_cache *root_cache)
3129 static char memcg_name_buf[NAME_MAX + 1]; /* protected by
3131 struct kmem_cache *cachep;
3134 lockdep_assert_held(&memcg_slab_mutex);
3136 id = memcg_cache_id(memcg);
3139 * Since per-memcg caches are created asynchronously on first
3140 * allocation (see memcg_kmem_get_cache()), several threads can try to
3141 * create the same cache, but only one of them may succeed.
3143 if (cache_from_memcg_idx(root_cache, id))
3146 cgroup_name(memcg->css.cgroup, memcg_name_buf, NAME_MAX + 1);
3147 cachep = memcg_create_kmem_cache(memcg, root_cache, memcg_name_buf);
3149 * If we could not create a memcg cache, do not complain, because
3150 * that's not critical at all as we can always proceed with the root
3156 list_add(&cachep->memcg_params->list, &memcg->memcg_slab_caches);
3159 * Since readers won't lock (see cache_from_memcg_idx()), we need a
3160 * barrier here to ensure nobody will see the kmem_cache partially
3165 BUG_ON(root_cache->memcg_params->memcg_caches[id]);
3166 root_cache->memcg_params->memcg_caches[id] = cachep;
3169 static void memcg_unregister_cache(struct kmem_cache *cachep)
3171 struct kmem_cache *root_cache;
3172 struct mem_cgroup *memcg;
3175 lockdep_assert_held(&memcg_slab_mutex);
3177 BUG_ON(is_root_cache(cachep));
3179 root_cache = cachep->memcg_params->root_cache;
3180 memcg = cachep->memcg_params->memcg;
3181 id = memcg_cache_id(memcg);
3183 BUG_ON(root_cache->memcg_params->memcg_caches[id] != cachep);
3184 root_cache->memcg_params->memcg_caches[id] = NULL;
3186 list_del(&cachep->memcg_params->list);
3188 kmem_cache_destroy(cachep);
3192 * During the creation a new cache, we need to disable our accounting mechanism
3193 * altogether. This is true even if we are not creating, but rather just
3194 * enqueing new caches to be created.
3196 * This is because that process will trigger allocations; some visible, like
3197 * explicit kmallocs to auxiliary data structures, name strings and internal
3198 * cache structures; some well concealed, like INIT_WORK() that can allocate
3199 * objects during debug.
3201 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3202 * to it. This may not be a bounded recursion: since the first cache creation
3203 * failed to complete (waiting on the allocation), we'll just try to create the
3204 * cache again, failing at the same point.
3206 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3207 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3208 * inside the following two functions.
3210 static inline void memcg_stop_kmem_account(void)
3212 VM_BUG_ON(!current->mm);
3213 current->memcg_kmem_skip_account++;
3216 static inline void memcg_resume_kmem_account(void)
3218 VM_BUG_ON(!current->mm);
3219 current->memcg_kmem_skip_account--;
3222 int __memcg_cleanup_cache_params(struct kmem_cache *s)
3224 struct kmem_cache *c;
3227 mutex_lock(&memcg_slab_mutex);
3228 for_each_memcg_cache_index(i) {
3229 c = cache_from_memcg_idx(s, i);
3233 memcg_unregister_cache(c);
3235 if (cache_from_memcg_idx(s, i))
3238 mutex_unlock(&memcg_slab_mutex);
3242 static void memcg_unregister_all_caches(struct mem_cgroup *memcg)
3244 struct kmem_cache *cachep;
3245 struct memcg_cache_params *params, *tmp;
3247 if (!memcg_kmem_is_active(memcg))
3250 mutex_lock(&memcg_slab_mutex);
3251 list_for_each_entry_safe(params, tmp, &memcg->memcg_slab_caches, list) {
3252 cachep = memcg_params_to_cache(params);
3253 kmem_cache_shrink(cachep);
3254 if (atomic_read(&cachep->memcg_params->nr_pages) == 0)
3255 memcg_unregister_cache(cachep);
3257 mutex_unlock(&memcg_slab_mutex);
3260 struct memcg_register_cache_work {
3261 struct mem_cgroup *memcg;
3262 struct kmem_cache *cachep;
3263 struct work_struct work;
3266 static void memcg_register_cache_func(struct work_struct *w)
3268 struct memcg_register_cache_work *cw =
3269 container_of(w, struct memcg_register_cache_work, work);
3270 struct mem_cgroup *memcg = cw->memcg;
3271 struct kmem_cache *cachep = cw->cachep;
3273 mutex_lock(&memcg_slab_mutex);
3274 memcg_register_cache(memcg, cachep);
3275 mutex_unlock(&memcg_slab_mutex);
3277 css_put(&memcg->css);
3282 * Enqueue the creation of a per-memcg kmem_cache.
3284 static void __memcg_schedule_register_cache(struct mem_cgroup *memcg,
3285 struct kmem_cache *cachep)
3287 struct memcg_register_cache_work *cw;
3289 cw = kmalloc(sizeof(*cw), GFP_NOWAIT);
3291 css_put(&memcg->css);
3296 cw->cachep = cachep;
3298 INIT_WORK(&cw->work, memcg_register_cache_func);
3299 schedule_work(&cw->work);
3302 static void memcg_schedule_register_cache(struct mem_cgroup *memcg,
3303 struct kmem_cache *cachep)
3306 * We need to stop accounting when we kmalloc, because if the
3307 * corresponding kmalloc cache is not yet created, the first allocation
3308 * in __memcg_schedule_register_cache will recurse.
3310 * However, it is better to enclose the whole function. Depending on
3311 * the debugging options enabled, INIT_WORK(), for instance, can
3312 * trigger an allocation. This too, will make us recurse. Because at
3313 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3314 * the safest choice is to do it like this, wrapping the whole function.
3316 memcg_stop_kmem_account();
3317 __memcg_schedule_register_cache(memcg, cachep);
3318 memcg_resume_kmem_account();
3321 int __memcg_charge_slab(struct kmem_cache *cachep, gfp_t gfp, int order)
3325 res = memcg_charge_kmem(cachep->memcg_params->memcg, gfp,
3326 PAGE_SIZE << order);
3328 atomic_add(1 << order, &cachep->memcg_params->nr_pages);
3332 void __memcg_uncharge_slab(struct kmem_cache *cachep, int order)
3334 memcg_uncharge_kmem(cachep->memcg_params->memcg, PAGE_SIZE << order);
3335 atomic_sub(1 << order, &cachep->memcg_params->nr_pages);
3339 * Return the kmem_cache we're supposed to use for a slab allocation.
3340 * We try to use the current memcg's version of the cache.
3342 * If the cache does not exist yet, if we are the first user of it,
3343 * we either create it immediately, if possible, or create it asynchronously
3345 * In the latter case, we will let the current allocation go through with
3346 * the original cache.
3348 * Can't be called in interrupt context or from kernel threads.
3349 * This function needs to be called with rcu_read_lock() held.
3351 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
3354 struct mem_cgroup *memcg;
3355 struct kmem_cache *memcg_cachep;
3357 VM_BUG_ON(!cachep->memcg_params);
3358 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
3360 if (!current->mm || current->memcg_kmem_skip_account)
3364 memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
3366 if (!memcg_can_account_kmem(memcg))
3369 memcg_cachep = cache_from_memcg_idx(cachep, memcg_cache_id(memcg));
3370 if (likely(memcg_cachep)) {
3371 cachep = memcg_cachep;
3375 /* The corresponding put will be done in the workqueue. */
3376 if (!css_tryget(&memcg->css))
3381 * If we are in a safe context (can wait, and not in interrupt
3382 * context), we could be be predictable and return right away.
3383 * This would guarantee that the allocation being performed
3384 * already belongs in the new cache.
3386 * However, there are some clashes that can arrive from locking.
3387 * For instance, because we acquire the slab_mutex while doing
3388 * memcg_create_kmem_cache, this means no further allocation
3389 * could happen with the slab_mutex held. So it's better to
3392 memcg_schedule_register_cache(memcg, cachep);
3400 * We need to verify if the allocation against current->mm->owner's memcg is
3401 * possible for the given order. But the page is not allocated yet, so we'll
3402 * need a further commit step to do the final arrangements.
3404 * It is possible for the task to switch cgroups in this mean time, so at
3405 * commit time, we can't rely on task conversion any longer. We'll then use
3406 * the handle argument to return to the caller which cgroup we should commit
3407 * against. We could also return the memcg directly and avoid the pointer
3408 * passing, but a boolean return value gives better semantics considering
3409 * the compiled-out case as well.
3411 * Returning true means the allocation is possible.
3414 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3416 struct mem_cgroup *memcg;
3422 * Disabling accounting is only relevant for some specific memcg
3423 * internal allocations. Therefore we would initially not have such
3424 * check here, since direct calls to the page allocator that are
3425 * accounted to kmemcg (alloc_kmem_pages and friends) only happen
3426 * outside memcg core. We are mostly concerned with cache allocations,
3427 * and by having this test at memcg_kmem_get_cache, we are already able
3428 * to relay the allocation to the root cache and bypass the memcg cache
3431 * There is one exception, though: the SLUB allocator does not create
3432 * large order caches, but rather service large kmallocs directly from
3433 * the page allocator. Therefore, the following sequence when backed by
3434 * the SLUB allocator:
3436 * memcg_stop_kmem_account();
3437 * kmalloc(<large_number>)
3438 * memcg_resume_kmem_account();
3440 * would effectively ignore the fact that we should skip accounting,
3441 * since it will drive us directly to this function without passing
3442 * through the cache selector memcg_kmem_get_cache. Such large
3443 * allocations are extremely rare but can happen, for instance, for the
3444 * cache arrays. We bring this test here.
3446 if (!current->mm || current->memcg_kmem_skip_account)
3449 memcg = get_mem_cgroup_from_mm(current->mm);
3451 if (!memcg_can_account_kmem(memcg)) {
3452 css_put(&memcg->css);
3456 ret = memcg_charge_kmem(memcg, gfp, PAGE_SIZE << order);
3460 css_put(&memcg->css);
3464 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3467 struct page_cgroup *pc;
3469 VM_BUG_ON(mem_cgroup_is_root(memcg));
3471 /* The page allocation failed. Revert */
3473 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3477 pc = lookup_page_cgroup(page);
3478 lock_page_cgroup(pc);
3479 pc->mem_cgroup = memcg;
3480 SetPageCgroupUsed(pc);
3481 unlock_page_cgroup(pc);
3484 void __memcg_kmem_uncharge_pages(struct page *page, int order)
3486 struct mem_cgroup *memcg = NULL;
3487 struct page_cgroup *pc;
3490 pc = lookup_page_cgroup(page);
3492 * Fast unlocked return. Theoretically might have changed, have to
3493 * check again after locking.
3495 if (!PageCgroupUsed(pc))
3498 lock_page_cgroup(pc);
3499 if (PageCgroupUsed(pc)) {
3500 memcg = pc->mem_cgroup;
3501 ClearPageCgroupUsed(pc);
3503 unlock_page_cgroup(pc);
3506 * We trust that only if there is a memcg associated with the page, it
3507 * is a valid allocation
3512 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
3513 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3516 static inline void memcg_unregister_all_caches(struct mem_cgroup *memcg)
3519 #endif /* CONFIG_MEMCG_KMEM */
3521 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3523 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3525 * Because tail pages are not marked as "used", set it. We're under
3526 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3527 * charge/uncharge will be never happen and move_account() is done under
3528 * compound_lock(), so we don't have to take care of races.
3530 void mem_cgroup_split_huge_fixup(struct page *head)
3532 struct page_cgroup *head_pc = lookup_page_cgroup(head);
3533 struct page_cgroup *pc;
3534 struct mem_cgroup *memcg;
3537 if (mem_cgroup_disabled())
3540 memcg = head_pc->mem_cgroup;
3541 for (i = 1; i < HPAGE_PMD_NR; i++) {
3543 pc->mem_cgroup = memcg;
3544 smp_wmb();/* see __commit_charge() */
3545 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
3547 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
3550 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3553 * mem_cgroup_move_account - move account of the page
3555 * @nr_pages: number of regular pages (>1 for huge pages)
3556 * @pc: page_cgroup of the page.
3557 * @from: mem_cgroup which the page is moved from.
3558 * @to: mem_cgroup which the page is moved to. @from != @to.
3560 * The caller must confirm following.
3561 * - page is not on LRU (isolate_page() is useful.)
3562 * - compound_lock is held when nr_pages > 1
3564 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3567 static int mem_cgroup_move_account(struct page *page,
3568 unsigned int nr_pages,
3569 struct page_cgroup *pc,
3570 struct mem_cgroup *from,
3571 struct mem_cgroup *to)
3573 unsigned long flags;
3575 bool anon = PageAnon(page);
3577 VM_BUG_ON(from == to);
3578 VM_BUG_ON_PAGE(PageLRU(page), page);
3580 * The page is isolated from LRU. So, collapse function
3581 * will not handle this page. But page splitting can happen.
3582 * Do this check under compound_page_lock(). The caller should
3586 if (nr_pages > 1 && !PageTransHuge(page))
3589 lock_page_cgroup(pc);
3592 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3595 move_lock_mem_cgroup(from, &flags);
3597 if (!anon && page_mapped(page)) {
3598 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
3600 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
3604 if (PageWriteback(page)) {
3605 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK],
3607 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK],
3611 mem_cgroup_charge_statistics(from, page, anon, -nr_pages);
3613 /* caller should have done css_get */
3614 pc->mem_cgroup = to;
3615 mem_cgroup_charge_statistics(to, page, anon, nr_pages);
3616 move_unlock_mem_cgroup(from, &flags);
3619 unlock_page_cgroup(pc);
3623 memcg_check_events(to, page);
3624 memcg_check_events(from, page);
3630 * mem_cgroup_move_parent - moves page to the parent group
3631 * @page: the page to move
3632 * @pc: page_cgroup of the page
3633 * @child: page's cgroup
3635 * move charges to its parent or the root cgroup if the group has no
3636 * parent (aka use_hierarchy==0).
3637 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3638 * mem_cgroup_move_account fails) the failure is always temporary and
3639 * it signals a race with a page removal/uncharge or migration. In the
3640 * first case the page is on the way out and it will vanish from the LRU
3641 * on the next attempt and the call should be retried later.
3642 * Isolation from the LRU fails only if page has been isolated from
3643 * the LRU since we looked at it and that usually means either global
3644 * reclaim or migration going on. The page will either get back to the
3646 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3647 * (!PageCgroupUsed) or moved to a different group. The page will
3648 * disappear in the next attempt.
3650 static int mem_cgroup_move_parent(struct page *page,
3651 struct page_cgroup *pc,
3652 struct mem_cgroup *child)
3654 struct mem_cgroup *parent;
3655 unsigned int nr_pages;
3656 unsigned long uninitialized_var(flags);
3659 VM_BUG_ON(mem_cgroup_is_root(child));
3662 if (!get_page_unless_zero(page))
3664 if (isolate_lru_page(page))
3667 nr_pages = hpage_nr_pages(page);
3669 parent = parent_mem_cgroup(child);
3671 * If no parent, move charges to root cgroup.
3674 parent = root_mem_cgroup;
3677 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
3678 flags = compound_lock_irqsave(page);
3681 ret = mem_cgroup_move_account(page, nr_pages,
3684 __mem_cgroup_cancel_local_charge(child, nr_pages);
3687 compound_unlock_irqrestore(page, flags);
3688 putback_lru_page(page);
3695 int mem_cgroup_charge_anon(struct page *page,
3696 struct mm_struct *mm, gfp_t gfp_mask)
3698 unsigned int nr_pages = 1;
3699 struct mem_cgroup *memcg;
3702 if (mem_cgroup_disabled())
3705 VM_BUG_ON_PAGE(page_mapped(page), page);
3706 VM_BUG_ON_PAGE(page->mapping && !PageAnon(page), page);
3709 if (PageTransHuge(page)) {
3710 nr_pages <<= compound_order(page);
3711 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
3713 * Never OOM-kill a process for a huge page. The
3714 * fault handler will fall back to regular pages.
3719 memcg = mem_cgroup_try_charge_mm(mm, gfp_mask, nr_pages, oom);
3722 __mem_cgroup_commit_charge(memcg, page, nr_pages,
3723 MEM_CGROUP_CHARGE_TYPE_ANON, false);
3728 * While swap-in, try_charge -> commit or cancel, the page is locked.
3729 * And when try_charge() successfully returns, one refcnt to memcg without
3730 * struct page_cgroup is acquired. This refcnt will be consumed by
3731 * "commit()" or removed by "cancel()"
3733 static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm,
3736 struct mem_cgroup **memcgp)
3738 struct mem_cgroup *memcg = NULL;
3739 struct page_cgroup *pc;
3742 pc = lookup_page_cgroup(page);
3744 * Every swap fault against a single page tries to charge the
3745 * page, bail as early as possible. shmem_unuse() encounters
3746 * already charged pages, too. The USED bit is protected by
3747 * the page lock, which serializes swap cache removal, which
3748 * in turn serializes uncharging.
3750 if (PageCgroupUsed(pc))
3752 if (do_swap_account)
3753 memcg = try_get_mem_cgroup_from_page(page);
3755 memcg = get_mem_cgroup_from_mm(mm);
3756 ret = mem_cgroup_try_charge(memcg, mask, 1, true);
3757 css_put(&memcg->css);
3759 memcg = root_mem_cgroup;
3767 int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page,
3768 gfp_t gfp_mask, struct mem_cgroup **memcgp)
3770 if (mem_cgroup_disabled()) {
3775 * A racing thread's fault, or swapoff, may have already
3776 * updated the pte, and even removed page from swap cache: in
3777 * those cases unuse_pte()'s pte_same() test will fail; but
3778 * there's also a KSM case which does need to charge the page.
3780 if (!PageSwapCache(page)) {
3781 struct mem_cgroup *memcg;
3783 memcg = mem_cgroup_try_charge_mm(mm, gfp_mask, 1, true);
3789 return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp);
3792 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
3794 if (mem_cgroup_disabled())
3798 __mem_cgroup_cancel_charge(memcg, 1);
3802 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
3803 enum charge_type ctype)
3805 if (mem_cgroup_disabled())
3810 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
3812 * Now swap is on-memory. This means this page may be
3813 * counted both as mem and swap....double count.
3814 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
3815 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
3816 * may call delete_from_swap_cache() before reach here.
3818 if (do_swap_account && PageSwapCache(page)) {
3819 swp_entry_t ent = {.val = page_private(page)};
3820 mem_cgroup_uncharge_swap(ent);
3824 void mem_cgroup_commit_charge_swapin(struct page *page,
3825 struct mem_cgroup *memcg)
3827 __mem_cgroup_commit_charge_swapin(page, memcg,
3828 MEM_CGROUP_CHARGE_TYPE_ANON);
3831 int mem_cgroup_charge_file(struct page *page, struct mm_struct *mm,
3834 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
3835 struct mem_cgroup *memcg;
3838 if (mem_cgroup_disabled())
3840 if (PageCompound(page))
3843 if (PageSwapCache(page)) { /* shmem */
3844 ret = __mem_cgroup_try_charge_swapin(mm, page,
3848 __mem_cgroup_commit_charge_swapin(page, memcg, type);
3853 * Page cache insertions can happen without an actual mm
3854 * context, e.g. during disk probing on boot.
3857 memcg = root_mem_cgroup;
3859 memcg = mem_cgroup_try_charge_mm(mm, gfp_mask, 1, true);
3863 __mem_cgroup_commit_charge(memcg, page, 1, type, false);
3867 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
3868 unsigned int nr_pages,
3869 const enum charge_type ctype)
3871 struct memcg_batch_info *batch = NULL;
3872 bool uncharge_memsw = true;
3874 /* If swapout, usage of swap doesn't decrease */
3875 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
3876 uncharge_memsw = false;
3878 batch = ¤t->memcg_batch;
3880 * In usual, we do css_get() when we remember memcg pointer.
3881 * But in this case, we keep res->usage until end of a series of
3882 * uncharges. Then, it's ok to ignore memcg's refcnt.
3885 batch->memcg = memcg;
3887 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
3888 * In those cases, all pages freed continuously can be expected to be in
3889 * the same cgroup and we have chance to coalesce uncharges.
3890 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
3891 * because we want to do uncharge as soon as possible.
3894 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
3895 goto direct_uncharge;
3898 goto direct_uncharge;
3901 * In typical case, batch->memcg == mem. This means we can
3902 * merge a series of uncharges to an uncharge of res_counter.
3903 * If not, we uncharge res_counter ony by one.
3905 if (batch->memcg != memcg)
3906 goto direct_uncharge;
3907 /* remember freed charge and uncharge it later */
3910 batch->memsw_nr_pages++;
3913 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
3915 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
3916 if (unlikely(batch->memcg != memcg))
3917 memcg_oom_recover(memcg);
3921 * uncharge if !page_mapped(page)
3923 static struct mem_cgroup *
3924 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
3927 struct mem_cgroup *memcg = NULL;
3928 unsigned int nr_pages = 1;
3929 struct page_cgroup *pc;
3932 if (mem_cgroup_disabled())
3935 if (PageTransHuge(page)) {
3936 nr_pages <<= compound_order(page);
3937 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
3940 * Check if our page_cgroup is valid
3942 pc = lookup_page_cgroup(page);
3943 if (unlikely(!PageCgroupUsed(pc)))
3946 lock_page_cgroup(pc);
3948 memcg = pc->mem_cgroup;
3950 if (!PageCgroupUsed(pc))
3953 anon = PageAnon(page);
3956 case MEM_CGROUP_CHARGE_TYPE_ANON:
3958 * Generally PageAnon tells if it's the anon statistics to be
3959 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
3960 * used before page reached the stage of being marked PageAnon.
3964 case MEM_CGROUP_CHARGE_TYPE_DROP:
3965 /* See mem_cgroup_prepare_migration() */
3966 if (page_mapped(page))
3969 * Pages under migration may not be uncharged. But
3970 * end_migration() /must/ be the one uncharging the
3971 * unused post-migration page and so it has to call
3972 * here with the migration bit still set. See the
3973 * res_counter handling below.
3975 if (!end_migration && PageCgroupMigration(pc))
3978 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
3979 if (!PageAnon(page)) { /* Shared memory */
3980 if (page->mapping && !page_is_file_cache(page))
3982 } else if (page_mapped(page)) /* Anon */
3989 mem_cgroup_charge_statistics(memcg, page, anon, -nr_pages);
3991 ClearPageCgroupUsed(pc);
3993 * pc->mem_cgroup is not cleared here. It will be accessed when it's
3994 * freed from LRU. This is safe because uncharged page is expected not
3995 * to be reused (freed soon). Exception is SwapCache, it's handled by
3996 * special functions.
3999 unlock_page_cgroup(pc);
4001 * even after unlock, we have memcg->res.usage here and this memcg
4002 * will never be freed, so it's safe to call css_get().
4004 memcg_check_events(memcg, page);
4005 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
4006 mem_cgroup_swap_statistics(memcg, true);
4007 css_get(&memcg->css);
4010 * Migration does not charge the res_counter for the
4011 * replacement page, so leave it alone when phasing out the
4012 * page that is unused after the migration.
4014 if (!end_migration && !mem_cgroup_is_root(memcg))
4015 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
4020 unlock_page_cgroup(pc);
4024 void mem_cgroup_uncharge_page(struct page *page)
4027 if (page_mapped(page))
4029 VM_BUG_ON_PAGE(page->mapping && !PageAnon(page), page);
4031 * If the page is in swap cache, uncharge should be deferred
4032 * to the swap path, which also properly accounts swap usage
4033 * and handles memcg lifetime.
4035 * Note that this check is not stable and reclaim may add the
4036 * page to swap cache at any time after this. However, if the
4037 * page is not in swap cache by the time page->mapcount hits
4038 * 0, there won't be any page table references to the swap
4039 * slot, and reclaim will free it and not actually write the
4042 if (PageSwapCache(page))
4044 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
4047 void mem_cgroup_uncharge_cache_page(struct page *page)
4049 VM_BUG_ON_PAGE(page_mapped(page), page);
4050 VM_BUG_ON_PAGE(page->mapping, page);
4051 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
4055 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4056 * In that cases, pages are freed continuously and we can expect pages
4057 * are in the same memcg. All these calls itself limits the number of
4058 * pages freed at once, then uncharge_start/end() is called properly.
4059 * This may be called prural(2) times in a context,
4062 void mem_cgroup_uncharge_start(void)
4064 current->memcg_batch.do_batch++;
4065 /* We can do nest. */
4066 if (current->memcg_batch.do_batch == 1) {
4067 current->memcg_batch.memcg = NULL;
4068 current->memcg_batch.nr_pages = 0;
4069 current->memcg_batch.memsw_nr_pages = 0;
4073 void mem_cgroup_uncharge_end(void)
4075 struct memcg_batch_info *batch = ¤t->memcg_batch;
4077 if (!batch->do_batch)
4081 if (batch->do_batch) /* If stacked, do nothing. */
4087 * This "batch->memcg" is valid without any css_get/put etc...
4088 * bacause we hide charges behind us.
4090 if (batch->nr_pages)
4091 res_counter_uncharge(&batch->memcg->res,
4092 batch->nr_pages * PAGE_SIZE);
4093 if (batch->memsw_nr_pages)
4094 res_counter_uncharge(&batch->memcg->memsw,
4095 batch->memsw_nr_pages * PAGE_SIZE);
4096 memcg_oom_recover(batch->memcg);
4097 /* forget this pointer (for sanity check) */
4098 batch->memcg = NULL;
4103 * called after __delete_from_swap_cache() and drop "page" account.
4104 * memcg information is recorded to swap_cgroup of "ent"
4107 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
4109 struct mem_cgroup *memcg;
4110 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
4112 if (!swapout) /* this was a swap cache but the swap is unused ! */
4113 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
4115 memcg = __mem_cgroup_uncharge_common(page, ctype, false);
4118 * record memcg information, if swapout && memcg != NULL,
4119 * css_get() was called in uncharge().
4121 if (do_swap_account && swapout && memcg)
4122 swap_cgroup_record(ent, mem_cgroup_id(memcg));
4126 #ifdef CONFIG_MEMCG_SWAP
4128 * called from swap_entry_free(). remove record in swap_cgroup and
4129 * uncharge "memsw" account.
4131 void mem_cgroup_uncharge_swap(swp_entry_t ent)
4133 struct mem_cgroup *memcg;
4136 if (!do_swap_account)
4139 id = swap_cgroup_record(ent, 0);
4141 memcg = mem_cgroup_lookup(id);
4144 * We uncharge this because swap is freed.
4145 * This memcg can be obsolete one. We avoid calling css_tryget
4147 if (!mem_cgroup_is_root(memcg))
4148 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
4149 mem_cgroup_swap_statistics(memcg, false);
4150 css_put(&memcg->css);
4156 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4157 * @entry: swap entry to be moved
4158 * @from: mem_cgroup which the entry is moved from
4159 * @to: mem_cgroup which the entry is moved to
4161 * It succeeds only when the swap_cgroup's record for this entry is the same
4162 * as the mem_cgroup's id of @from.
4164 * Returns 0 on success, -EINVAL on failure.
4166 * The caller must have charged to @to, IOW, called res_counter_charge() about
4167 * both res and memsw, and called css_get().
4169 static int mem_cgroup_move_swap_account(swp_entry_t entry,
4170 struct mem_cgroup *from, struct mem_cgroup *to)
4172 unsigned short old_id, new_id;
4174 old_id = mem_cgroup_id(from);
4175 new_id = mem_cgroup_id(to);
4177 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
4178 mem_cgroup_swap_statistics(from, false);
4179 mem_cgroup_swap_statistics(to, true);
4181 * This function is only called from task migration context now.
4182 * It postpones res_counter and refcount handling till the end
4183 * of task migration(mem_cgroup_clear_mc()) for performance
4184 * improvement. But we cannot postpone css_get(to) because if
4185 * the process that has been moved to @to does swap-in, the
4186 * refcount of @to might be decreased to 0.
4188 * We are in attach() phase, so the cgroup is guaranteed to be
4189 * alive, so we can just call css_get().
4197 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
4198 struct mem_cgroup *from, struct mem_cgroup *to)
4205 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4208 void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
4209 struct mem_cgroup **memcgp)
4211 struct mem_cgroup *memcg = NULL;
4212 unsigned int nr_pages = 1;
4213 struct page_cgroup *pc;
4214 enum charge_type ctype;
4218 if (mem_cgroup_disabled())
4221 if (PageTransHuge(page))
4222 nr_pages <<= compound_order(page);
4224 pc = lookup_page_cgroup(page);
4225 lock_page_cgroup(pc);
4226 if (PageCgroupUsed(pc)) {
4227 memcg = pc->mem_cgroup;
4228 css_get(&memcg->css);
4230 * At migrating an anonymous page, its mapcount goes down
4231 * to 0 and uncharge() will be called. But, even if it's fully
4232 * unmapped, migration may fail and this page has to be
4233 * charged again. We set MIGRATION flag here and delay uncharge
4234 * until end_migration() is called
4236 * Corner Case Thinking
4238 * When the old page was mapped as Anon and it's unmap-and-freed
4239 * while migration was ongoing.
4240 * If unmap finds the old page, uncharge() of it will be delayed
4241 * until end_migration(). If unmap finds a new page, it's
4242 * uncharged when it make mapcount to be 1->0. If unmap code
4243 * finds swap_migration_entry, the new page will not be mapped
4244 * and end_migration() will find it(mapcount==0).
4247 * When the old page was mapped but migraion fails, the kernel
4248 * remaps it. A charge for it is kept by MIGRATION flag even
4249 * if mapcount goes down to 0. We can do remap successfully
4250 * without charging it again.
4253 * The "old" page is under lock_page() until the end of
4254 * migration, so, the old page itself will not be swapped-out.
4255 * If the new page is swapped out before end_migraton, our
4256 * hook to usual swap-out path will catch the event.
4259 SetPageCgroupMigration(pc);
4261 unlock_page_cgroup(pc);
4263 * If the page is not charged at this point,
4271 * We charge new page before it's used/mapped. So, even if unlock_page()
4272 * is called before end_migration, we can catch all events on this new
4273 * page. In the case new page is migrated but not remapped, new page's
4274 * mapcount will be finally 0 and we call uncharge in end_migration().
4277 ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
4279 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
4281 * The page is committed to the memcg, but it's not actually
4282 * charged to the res_counter since we plan on replacing the
4283 * old one and only one page is going to be left afterwards.
4285 __mem_cgroup_commit_charge(memcg, newpage, nr_pages, ctype, false);
4288 /* remove redundant charge if migration failed*/
4289 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
4290 struct page *oldpage, struct page *newpage, bool migration_ok)
4292 struct page *used, *unused;
4293 struct page_cgroup *pc;
4299 if (!migration_ok) {
4306 anon = PageAnon(used);
4307 __mem_cgroup_uncharge_common(unused,
4308 anon ? MEM_CGROUP_CHARGE_TYPE_ANON
4309 : MEM_CGROUP_CHARGE_TYPE_CACHE,
4311 css_put(&memcg->css);
4313 * We disallowed uncharge of pages under migration because mapcount
4314 * of the page goes down to zero, temporarly.
4315 * Clear the flag and check the page should be charged.
4317 pc = lookup_page_cgroup(oldpage);
4318 lock_page_cgroup(pc);
4319 ClearPageCgroupMigration(pc);
4320 unlock_page_cgroup(pc);
4323 * If a page is a file cache, radix-tree replacement is very atomic
4324 * and we can skip this check. When it was an Anon page, its mapcount
4325 * goes down to 0. But because we added MIGRATION flage, it's not
4326 * uncharged yet. There are several case but page->mapcount check
4327 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4328 * check. (see prepare_charge() also)
4331 mem_cgroup_uncharge_page(used);
4335 * At replace page cache, newpage is not under any memcg but it's on
4336 * LRU. So, this function doesn't touch res_counter but handles LRU
4337 * in correct way. Both pages are locked so we cannot race with uncharge.
4339 void mem_cgroup_replace_page_cache(struct page *oldpage,
4340 struct page *newpage)
4342 struct mem_cgroup *memcg = NULL;
4343 struct page_cgroup *pc;
4344 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4346 if (mem_cgroup_disabled())
4349 pc = lookup_page_cgroup(oldpage);
4350 /* fix accounting on old pages */
4351 lock_page_cgroup(pc);
4352 if (PageCgroupUsed(pc)) {
4353 memcg = pc->mem_cgroup;
4354 mem_cgroup_charge_statistics(memcg, oldpage, false, -1);
4355 ClearPageCgroupUsed(pc);
4357 unlock_page_cgroup(pc);
4360 * When called from shmem_replace_page(), in some cases the
4361 * oldpage has already been charged, and in some cases not.
4366 * Even if newpage->mapping was NULL before starting replacement,
4367 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4368 * LRU while we overwrite pc->mem_cgroup.
4370 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
4373 #ifdef CONFIG_DEBUG_VM
4374 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
4376 struct page_cgroup *pc;
4378 pc = lookup_page_cgroup(page);
4380 * Can be NULL while feeding pages into the page allocator for
4381 * the first time, i.e. during boot or memory hotplug;
4382 * or when mem_cgroup_disabled().
4384 if (likely(pc) && PageCgroupUsed(pc))
4389 bool mem_cgroup_bad_page_check(struct page *page)
4391 if (mem_cgroup_disabled())
4394 return lookup_page_cgroup_used(page) != NULL;
4397 void mem_cgroup_print_bad_page(struct page *page)
4399 struct page_cgroup *pc;
4401 pc = lookup_page_cgroup_used(page);
4403 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4404 pc, pc->flags, pc->mem_cgroup);
4409 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
4410 unsigned long long val)
4413 u64 memswlimit, memlimit;
4415 int children = mem_cgroup_count_children(memcg);
4416 u64 curusage, oldusage;
4420 * For keeping hierarchical_reclaim simple, how long we should retry
4421 * is depends on callers. We set our retry-count to be function
4422 * of # of children which we should visit in this loop.
4424 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
4426 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4429 while (retry_count) {
4430 if (signal_pending(current)) {
4435 * Rather than hide all in some function, I do this in
4436 * open coded manner. You see what this really does.
4437 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4439 mutex_lock(&set_limit_mutex);
4440 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4441 if (memswlimit < val) {
4443 mutex_unlock(&set_limit_mutex);
4447 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4451 ret = res_counter_set_limit(&memcg->res, val);
4453 if (memswlimit == val)
4454 memcg->memsw_is_minimum = true;
4456 memcg->memsw_is_minimum = false;
4458 mutex_unlock(&set_limit_mutex);
4463 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4464 MEM_CGROUP_RECLAIM_SHRINK);
4465 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4466 /* Usage is reduced ? */
4467 if (curusage >= oldusage)
4470 oldusage = curusage;
4472 if (!ret && enlarge)
4473 memcg_oom_recover(memcg);
4478 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
4479 unsigned long long val)
4482 u64 memlimit, memswlimit, oldusage, curusage;
4483 int children = mem_cgroup_count_children(memcg);
4487 /* see mem_cgroup_resize_res_limit */
4488 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
4489 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4490 while (retry_count) {
4491 if (signal_pending(current)) {
4496 * Rather than hide all in some function, I do this in
4497 * open coded manner. You see what this really does.
4498 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4500 mutex_lock(&set_limit_mutex);
4501 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4502 if (memlimit > val) {
4504 mutex_unlock(&set_limit_mutex);
4507 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4508 if (memswlimit < val)
4510 ret = res_counter_set_limit(&memcg->memsw, val);
4512 if (memlimit == val)
4513 memcg->memsw_is_minimum = true;
4515 memcg->memsw_is_minimum = false;
4517 mutex_unlock(&set_limit_mutex);
4522 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4523 MEM_CGROUP_RECLAIM_NOSWAP |
4524 MEM_CGROUP_RECLAIM_SHRINK);
4525 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4526 /* Usage is reduced ? */
4527 if (curusage >= oldusage)
4530 oldusage = curusage;
4532 if (!ret && enlarge)
4533 memcg_oom_recover(memcg);
4537 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
4539 unsigned long *total_scanned)
4541 unsigned long nr_reclaimed = 0;
4542 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
4543 unsigned long reclaimed;
4545 struct mem_cgroup_tree_per_zone *mctz;
4546 unsigned long long excess;
4547 unsigned long nr_scanned;
4552 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
4554 * This loop can run a while, specially if mem_cgroup's continuously
4555 * keep exceeding their soft limit and putting the system under
4562 mz = mem_cgroup_largest_soft_limit_node(mctz);
4567 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
4568 gfp_mask, &nr_scanned);
4569 nr_reclaimed += reclaimed;
4570 *total_scanned += nr_scanned;
4571 spin_lock(&mctz->lock);
4574 * If we failed to reclaim anything from this memory cgroup
4575 * it is time to move on to the next cgroup
4581 * Loop until we find yet another one.
4583 * By the time we get the soft_limit lock
4584 * again, someone might have aded the
4585 * group back on the RB tree. Iterate to
4586 * make sure we get a different mem.
4587 * mem_cgroup_largest_soft_limit_node returns
4588 * NULL if no other cgroup is present on
4592 __mem_cgroup_largest_soft_limit_node(mctz);
4594 css_put(&next_mz->memcg->css);
4595 else /* next_mz == NULL or other memcg */
4599 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
4600 excess = res_counter_soft_limit_excess(&mz->memcg->res);
4602 * One school of thought says that we should not add
4603 * back the node to the tree if reclaim returns 0.
4604 * But our reclaim could return 0, simply because due
4605 * to priority we are exposing a smaller subset of
4606 * memory to reclaim from. Consider this as a longer
4609 /* If excess == 0, no tree ops */
4610 __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess);
4611 spin_unlock(&mctz->lock);
4612 css_put(&mz->memcg->css);
4615 * Could not reclaim anything and there are no more
4616 * mem cgroups to try or we seem to be looping without
4617 * reclaiming anything.
4619 if (!nr_reclaimed &&
4621 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
4623 } while (!nr_reclaimed);
4625 css_put(&next_mz->memcg->css);
4626 return nr_reclaimed;
4630 * mem_cgroup_force_empty_list - clears LRU of a group
4631 * @memcg: group to clear
4634 * @lru: lru to to clear
4636 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4637 * reclaim the pages page themselves - pages are moved to the parent (or root)
4640 static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
4641 int node, int zid, enum lru_list lru)
4643 struct lruvec *lruvec;
4644 unsigned long flags;
4645 struct list_head *list;
4649 zone = &NODE_DATA(node)->node_zones[zid];
4650 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
4651 list = &lruvec->lists[lru];
4655 struct page_cgroup *pc;
4658 spin_lock_irqsave(&zone->lru_lock, flags);
4659 if (list_empty(list)) {
4660 spin_unlock_irqrestore(&zone->lru_lock, flags);
4663 page = list_entry(list->prev, struct page, lru);
4665 list_move(&page->lru, list);
4667 spin_unlock_irqrestore(&zone->lru_lock, flags);
4670 spin_unlock_irqrestore(&zone->lru_lock, flags);
4672 pc = lookup_page_cgroup(page);
4674 if (mem_cgroup_move_parent(page, pc, memcg)) {
4675 /* found lock contention or "pc" is obsolete. */
4680 } while (!list_empty(list));
4684 * make mem_cgroup's charge to be 0 if there is no task by moving
4685 * all the charges and pages to the parent.
4686 * This enables deleting this mem_cgroup.
4688 * Caller is responsible for holding css reference on the memcg.
4690 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
4696 /* This is for making all *used* pages to be on LRU. */
4697 lru_add_drain_all();
4698 drain_all_stock_sync(memcg);
4699 mem_cgroup_start_move(memcg);
4700 for_each_node_state(node, N_MEMORY) {
4701 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4704 mem_cgroup_force_empty_list(memcg,
4709 mem_cgroup_end_move(memcg);
4710 memcg_oom_recover(memcg);
4714 * Kernel memory may not necessarily be trackable to a specific
4715 * process. So they are not migrated, and therefore we can't
4716 * expect their value to drop to 0 here.
4717 * Having res filled up with kmem only is enough.
4719 * This is a safety check because mem_cgroup_force_empty_list
4720 * could have raced with mem_cgroup_replace_page_cache callers
4721 * so the lru seemed empty but the page could have been added
4722 * right after the check. RES_USAGE should be safe as we always
4723 * charge before adding to the LRU.
4725 usage = res_counter_read_u64(&memcg->res, RES_USAGE) -
4726 res_counter_read_u64(&memcg->kmem, RES_USAGE);
4727 } while (usage > 0);
4730 static inline bool memcg_has_children(struct mem_cgroup *memcg)
4732 lockdep_assert_held(&memcg_create_mutex);
4734 * The lock does not prevent addition or deletion to the list
4735 * of children, but it prevents a new child from being
4736 * initialized based on this parent in css_online(), so it's
4737 * enough to decide whether hierarchically inherited
4738 * attributes can still be changed or not.
4740 return memcg->use_hierarchy &&
4741 !list_empty(&memcg->css.cgroup->children);
4745 * Reclaims as many pages from the given memcg as possible and moves
4746 * the rest to the parent.
4748 * Caller is responsible for holding css reference for memcg.
4750 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
4752 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
4753 struct cgroup *cgrp = memcg->css.cgroup;
4755 /* returns EBUSY if there is a task or if we come here twice. */
4756 if (cgroup_has_tasks(cgrp) || !list_empty(&cgrp->children))
4759 /* we call try-to-free pages for make this cgroup empty */
4760 lru_add_drain_all();
4761 /* try to free all pages in this cgroup */
4762 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
4765 if (signal_pending(current))
4768 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
4772 /* maybe some writeback is necessary */
4773 congestion_wait(BLK_RW_ASYNC, HZ/10);
4778 mem_cgroup_reparent_charges(memcg);
4783 static int mem_cgroup_force_empty_write(struct cgroup_subsys_state *css,
4786 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4788 if (mem_cgroup_is_root(memcg))
4790 return mem_cgroup_force_empty(memcg);
4793 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
4796 return mem_cgroup_from_css(css)->use_hierarchy;
4799 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
4800 struct cftype *cft, u64 val)
4803 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4804 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(css_parent(&memcg->css));
4806 mutex_lock(&memcg_create_mutex);
4808 if (memcg->use_hierarchy == val)
4812 * If parent's use_hierarchy is set, we can't make any modifications
4813 * in the child subtrees. If it is unset, then the change can
4814 * occur, provided the current cgroup has no children.
4816 * For the root cgroup, parent_mem is NULL, we allow value to be
4817 * set if there are no children.
4819 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
4820 (val == 1 || val == 0)) {
4821 if (list_empty(&memcg->css.cgroup->children))
4822 memcg->use_hierarchy = val;
4829 mutex_unlock(&memcg_create_mutex);
4835 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
4836 enum mem_cgroup_stat_index idx)
4838 struct mem_cgroup *iter;
4841 /* Per-cpu values can be negative, use a signed accumulator */
4842 for_each_mem_cgroup_tree(iter, memcg)
4843 val += mem_cgroup_read_stat(iter, idx);
4845 if (val < 0) /* race ? */
4850 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
4854 if (!mem_cgroup_is_root(memcg)) {
4856 return res_counter_read_u64(&memcg->res, RES_USAGE);
4858 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
4862 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
4863 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
4865 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
4866 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
4869 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
4871 return val << PAGE_SHIFT;
4874 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
4877 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4882 type = MEMFILE_TYPE(cft->private);
4883 name = MEMFILE_ATTR(cft->private);
4887 if (name == RES_USAGE)
4888 val = mem_cgroup_usage(memcg, false);
4890 val = res_counter_read_u64(&memcg->res, name);
4893 if (name == RES_USAGE)
4894 val = mem_cgroup_usage(memcg, true);
4896 val = res_counter_read_u64(&memcg->memsw, name);
4899 val = res_counter_read_u64(&memcg->kmem, name);
4908 #ifdef CONFIG_MEMCG_KMEM
4909 /* should be called with activate_kmem_mutex held */
4910 static int __memcg_activate_kmem(struct mem_cgroup *memcg,
4911 unsigned long long limit)
4916 if (memcg_kmem_is_active(memcg))
4920 * We are going to allocate memory for data shared by all memory
4921 * cgroups so let's stop accounting here.
4923 memcg_stop_kmem_account();
4926 * For simplicity, we won't allow this to be disabled. It also can't
4927 * be changed if the cgroup has children already, or if tasks had
4930 * If tasks join before we set the limit, a person looking at
4931 * kmem.usage_in_bytes will have no way to determine when it took
4932 * place, which makes the value quite meaningless.
4934 * After it first became limited, changes in the value of the limit are
4935 * of course permitted.
4937 mutex_lock(&memcg_create_mutex);
4938 if (cgroup_has_tasks(memcg->css.cgroup) || memcg_has_children(memcg))
4940 mutex_unlock(&memcg_create_mutex);
4944 memcg_id = ida_simple_get(&kmem_limited_groups,
4945 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
4952 * Make sure we have enough space for this cgroup in each root cache's
4955 mutex_lock(&memcg_slab_mutex);
4956 err = memcg_update_all_caches(memcg_id + 1);
4957 mutex_unlock(&memcg_slab_mutex);
4961 memcg->kmemcg_id = memcg_id;
4962 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
4965 * We couldn't have accounted to this cgroup, because it hasn't got the
4966 * active bit set yet, so this should succeed.
4968 err = res_counter_set_limit(&memcg->kmem, limit);
4971 static_key_slow_inc(&memcg_kmem_enabled_key);
4973 * Setting the active bit after enabling static branching will
4974 * guarantee no one starts accounting before all call sites are
4977 memcg_kmem_set_active(memcg);
4979 memcg_resume_kmem_account();
4983 ida_simple_remove(&kmem_limited_groups, memcg_id);
4987 static int memcg_activate_kmem(struct mem_cgroup *memcg,
4988 unsigned long long limit)
4992 mutex_lock(&activate_kmem_mutex);
4993 ret = __memcg_activate_kmem(memcg, limit);
4994 mutex_unlock(&activate_kmem_mutex);
4998 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
4999 unsigned long long val)
5003 if (!memcg_kmem_is_active(memcg))
5004 ret = memcg_activate_kmem(memcg, val);
5006 ret = res_counter_set_limit(&memcg->kmem, val);
5010 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
5013 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5018 mutex_lock(&activate_kmem_mutex);
5020 * If the parent cgroup is not kmem-active now, it cannot be activated
5021 * after this point, because it has at least one child already.
5023 if (memcg_kmem_is_active(parent))
5024 ret = __memcg_activate_kmem(memcg, RES_COUNTER_MAX);
5025 mutex_unlock(&activate_kmem_mutex);
5029 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
5030 unsigned long long val)
5034 #endif /* CONFIG_MEMCG_KMEM */
5037 * The user of this function is...
5040 static int mem_cgroup_write(struct cgroup_subsys_state *css, struct cftype *cft,
5043 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5046 unsigned long long val;
5049 type = MEMFILE_TYPE(cft->private);
5050 name = MEMFILE_ATTR(cft->private);
5054 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
5058 /* This function does all necessary parse...reuse it */
5059 ret = res_counter_memparse_write_strategy(buffer, &val);
5063 ret = mem_cgroup_resize_limit(memcg, val);
5064 else if (type == _MEMSWAP)
5065 ret = mem_cgroup_resize_memsw_limit(memcg, val);
5066 else if (type == _KMEM)
5067 ret = memcg_update_kmem_limit(memcg, val);
5071 case RES_SOFT_LIMIT:
5072 ret = res_counter_memparse_write_strategy(buffer, &val);
5076 * For memsw, soft limits are hard to implement in terms
5077 * of semantics, for now, we support soft limits for
5078 * control without swap
5081 ret = res_counter_set_soft_limit(&memcg->res, val);
5086 ret = -EINVAL; /* should be BUG() ? */
5092 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
5093 unsigned long long *mem_limit, unsigned long long *memsw_limit)
5095 unsigned long long min_limit, min_memsw_limit, tmp;
5097 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
5098 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5099 if (!memcg->use_hierarchy)
5102 while (css_parent(&memcg->css)) {
5103 memcg = mem_cgroup_from_css(css_parent(&memcg->css));
5104 if (!memcg->use_hierarchy)
5106 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
5107 min_limit = min(min_limit, tmp);
5108 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5109 min_memsw_limit = min(min_memsw_limit, tmp);
5112 *mem_limit = min_limit;
5113 *memsw_limit = min_memsw_limit;
5116 static int mem_cgroup_reset(struct cgroup_subsys_state *css, unsigned int event)
5118 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5122 type = MEMFILE_TYPE(event);
5123 name = MEMFILE_ATTR(event);
5128 res_counter_reset_max(&memcg->res);
5129 else if (type == _MEMSWAP)
5130 res_counter_reset_max(&memcg->memsw);
5131 else if (type == _KMEM)
5132 res_counter_reset_max(&memcg->kmem);
5138 res_counter_reset_failcnt(&memcg->res);
5139 else if (type == _MEMSWAP)
5140 res_counter_reset_failcnt(&memcg->memsw);
5141 else if (type == _KMEM)
5142 res_counter_reset_failcnt(&memcg->kmem);
5151 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
5154 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
5158 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5159 struct cftype *cft, u64 val)
5161 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5163 if (val >= (1 << NR_MOVE_TYPE))
5167 * No kind of locking is needed in here, because ->can_attach() will
5168 * check this value once in the beginning of the process, and then carry
5169 * on with stale data. This means that changes to this value will only
5170 * affect task migrations starting after the change.
5172 memcg->move_charge_at_immigrate = val;
5176 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5177 struct cftype *cft, u64 val)
5184 static int memcg_numa_stat_show(struct seq_file *m, void *v)
5188 unsigned int lru_mask;
5191 static const struct numa_stat stats[] = {
5192 { "total", LRU_ALL },
5193 { "file", LRU_ALL_FILE },
5194 { "anon", LRU_ALL_ANON },
5195 { "unevictable", BIT(LRU_UNEVICTABLE) },
5197 const struct numa_stat *stat;
5200 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5202 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
5203 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
5204 seq_printf(m, "%s=%lu", stat->name, nr);
5205 for_each_node_state(nid, N_MEMORY) {
5206 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5208 seq_printf(m, " N%d=%lu", nid, nr);
5213 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
5214 struct mem_cgroup *iter;
5217 for_each_mem_cgroup_tree(iter, memcg)
5218 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
5219 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
5220 for_each_node_state(nid, N_MEMORY) {
5222 for_each_mem_cgroup_tree(iter, memcg)
5223 nr += mem_cgroup_node_nr_lru_pages(
5224 iter, nid, stat->lru_mask);
5225 seq_printf(m, " N%d=%lu", nid, nr);
5232 #endif /* CONFIG_NUMA */
5234 static inline void mem_cgroup_lru_names_not_uptodate(void)
5236 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
5239 static int memcg_stat_show(struct seq_file *m, void *v)
5241 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5242 struct mem_cgroup *mi;
5245 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5246 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5248 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
5249 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
5252 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
5253 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
5254 mem_cgroup_read_events(memcg, i));
5256 for (i = 0; i < NR_LRU_LISTS; i++)
5257 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
5258 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
5260 /* Hierarchical information */
5262 unsigned long long limit, memsw_limit;
5263 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
5264 seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
5265 if (do_swap_account)
5266 seq_printf(m, "hierarchical_memsw_limit %llu\n",
5270 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5273 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5275 for_each_mem_cgroup_tree(mi, memcg)
5276 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
5277 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
5280 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
5281 unsigned long long val = 0;
5283 for_each_mem_cgroup_tree(mi, memcg)
5284 val += mem_cgroup_read_events(mi, i);
5285 seq_printf(m, "total_%s %llu\n",
5286 mem_cgroup_events_names[i], val);
5289 for (i = 0; i < NR_LRU_LISTS; i++) {
5290 unsigned long long val = 0;
5292 for_each_mem_cgroup_tree(mi, memcg)
5293 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
5294 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
5297 #ifdef CONFIG_DEBUG_VM
5300 struct mem_cgroup_per_zone *mz;
5301 struct zone_reclaim_stat *rstat;
5302 unsigned long recent_rotated[2] = {0, 0};
5303 unsigned long recent_scanned[2] = {0, 0};
5305 for_each_online_node(nid)
5306 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
5307 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
5308 rstat = &mz->lruvec.reclaim_stat;
5310 recent_rotated[0] += rstat->recent_rotated[0];
5311 recent_rotated[1] += rstat->recent_rotated[1];
5312 recent_scanned[0] += rstat->recent_scanned[0];
5313 recent_scanned[1] += rstat->recent_scanned[1];
5315 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
5316 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
5317 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
5318 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
5325 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
5328 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5330 return mem_cgroup_swappiness(memcg);
5333 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
5334 struct cftype *cft, u64 val)
5336 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5341 if (css_parent(css))
5342 memcg->swappiness = val;
5344 vm_swappiness = val;
5349 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
5351 struct mem_cgroup_threshold_ary *t;
5357 t = rcu_dereference(memcg->thresholds.primary);
5359 t = rcu_dereference(memcg->memsw_thresholds.primary);
5364 usage = mem_cgroup_usage(memcg, swap);
5367 * current_threshold points to threshold just below or equal to usage.
5368 * If it's not true, a threshold was crossed after last
5369 * call of __mem_cgroup_threshold().
5371 i = t->current_threshold;
5374 * Iterate backward over array of thresholds starting from
5375 * current_threshold and check if a threshold is crossed.
5376 * If none of thresholds below usage is crossed, we read
5377 * only one element of the array here.
5379 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
5380 eventfd_signal(t->entries[i].eventfd, 1);
5382 /* i = current_threshold + 1 */
5386 * Iterate forward over array of thresholds starting from
5387 * current_threshold+1 and check if a threshold is crossed.
5388 * If none of thresholds above usage is crossed, we read
5389 * only one element of the array here.
5391 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
5392 eventfd_signal(t->entries[i].eventfd, 1);
5394 /* Update current_threshold */
5395 t->current_threshold = i - 1;
5400 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
5403 __mem_cgroup_threshold(memcg, false);
5404 if (do_swap_account)
5405 __mem_cgroup_threshold(memcg, true);
5407 memcg = parent_mem_cgroup(memcg);
5411 static int compare_thresholds(const void *a, const void *b)
5413 const struct mem_cgroup_threshold *_a = a;
5414 const struct mem_cgroup_threshold *_b = b;
5416 if (_a->threshold > _b->threshold)
5419 if (_a->threshold < _b->threshold)
5425 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
5427 struct mem_cgroup_eventfd_list *ev;
5429 list_for_each_entry(ev, &memcg->oom_notify, list)
5430 eventfd_signal(ev->eventfd, 1);
5434 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
5436 struct mem_cgroup *iter;
5438 for_each_mem_cgroup_tree(iter, memcg)
5439 mem_cgroup_oom_notify_cb(iter);
5442 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
5443 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
5445 struct mem_cgroup_thresholds *thresholds;
5446 struct mem_cgroup_threshold_ary *new;
5447 u64 threshold, usage;
5450 ret = res_counter_memparse_write_strategy(args, &threshold);
5454 mutex_lock(&memcg->thresholds_lock);
5457 thresholds = &memcg->thresholds;
5458 else if (type == _MEMSWAP)
5459 thresholds = &memcg->memsw_thresholds;
5463 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5465 /* Check if a threshold crossed before adding a new one */
5466 if (thresholds->primary)
5467 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5469 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
5471 /* Allocate memory for new array of thresholds */
5472 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
5480 /* Copy thresholds (if any) to new array */
5481 if (thresholds->primary) {
5482 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
5483 sizeof(struct mem_cgroup_threshold));
5486 /* Add new threshold */
5487 new->entries[size - 1].eventfd = eventfd;
5488 new->entries[size - 1].threshold = threshold;
5490 /* Sort thresholds. Registering of new threshold isn't time-critical */
5491 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
5492 compare_thresholds, NULL);
5494 /* Find current threshold */
5495 new->current_threshold = -1;
5496 for (i = 0; i < size; i++) {
5497 if (new->entries[i].threshold <= usage) {
5499 * new->current_threshold will not be used until
5500 * rcu_assign_pointer(), so it's safe to increment
5503 ++new->current_threshold;
5508 /* Free old spare buffer and save old primary buffer as spare */
5509 kfree(thresholds->spare);
5510 thresholds->spare = thresholds->primary;
5512 rcu_assign_pointer(thresholds->primary, new);
5514 /* To be sure that nobody uses thresholds */
5518 mutex_unlock(&memcg->thresholds_lock);
5523 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
5524 struct eventfd_ctx *eventfd, const char *args)
5526 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
5529 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
5530 struct eventfd_ctx *eventfd, const char *args)
5532 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
5535 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
5536 struct eventfd_ctx *eventfd, enum res_type type)
5538 struct mem_cgroup_thresholds *thresholds;
5539 struct mem_cgroup_threshold_ary *new;
5543 mutex_lock(&memcg->thresholds_lock);
5545 thresholds = &memcg->thresholds;
5546 else if (type == _MEMSWAP)
5547 thresholds = &memcg->memsw_thresholds;
5551 if (!thresholds->primary)
5554 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5556 /* Check if a threshold crossed before removing */
5557 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5559 /* Calculate new number of threshold */
5561 for (i = 0; i < thresholds->primary->size; i++) {
5562 if (thresholds->primary->entries[i].eventfd != eventfd)
5566 new = thresholds->spare;
5568 /* Set thresholds array to NULL if we don't have thresholds */
5577 /* Copy thresholds and find current threshold */
5578 new->current_threshold = -1;
5579 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
5580 if (thresholds->primary->entries[i].eventfd == eventfd)
5583 new->entries[j] = thresholds->primary->entries[i];
5584 if (new->entries[j].threshold <= usage) {
5586 * new->current_threshold will not be used
5587 * until rcu_assign_pointer(), so it's safe to increment
5590 ++new->current_threshold;
5596 /* Swap primary and spare array */
5597 thresholds->spare = thresholds->primary;
5598 /* If all events are unregistered, free the spare array */
5600 kfree(thresholds->spare);
5601 thresholds->spare = NULL;
5604 rcu_assign_pointer(thresholds->primary, new);
5606 /* To be sure that nobody uses thresholds */
5609 mutex_unlock(&memcg->thresholds_lock);
5612 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
5613 struct eventfd_ctx *eventfd)
5615 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
5618 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
5619 struct eventfd_ctx *eventfd)
5621 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
5624 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
5625 struct eventfd_ctx *eventfd, const char *args)
5627 struct mem_cgroup_eventfd_list *event;
5629 event = kmalloc(sizeof(*event), GFP_KERNEL);
5633 spin_lock(&memcg_oom_lock);
5635 event->eventfd = eventfd;
5636 list_add(&event->list, &memcg->oom_notify);
5638 /* already in OOM ? */
5639 if (atomic_read(&memcg->under_oom))
5640 eventfd_signal(eventfd, 1);
5641 spin_unlock(&memcg_oom_lock);
5646 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
5647 struct eventfd_ctx *eventfd)
5649 struct mem_cgroup_eventfd_list *ev, *tmp;
5651 spin_lock(&memcg_oom_lock);
5653 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
5654 if (ev->eventfd == eventfd) {
5655 list_del(&ev->list);
5660 spin_unlock(&memcg_oom_lock);
5663 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
5665 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
5667 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
5668 seq_printf(sf, "under_oom %d\n", (bool)atomic_read(&memcg->under_oom));
5672 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
5673 struct cftype *cft, u64 val)
5675 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5677 /* cannot set to root cgroup and only 0 and 1 are allowed */
5678 if (!css_parent(css) || !((val == 0) || (val == 1)))
5681 memcg->oom_kill_disable = val;
5683 memcg_oom_recover(memcg);
5688 #ifdef CONFIG_MEMCG_KMEM
5689 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5693 memcg->kmemcg_id = -1;
5694 ret = memcg_propagate_kmem(memcg);
5698 return mem_cgroup_sockets_init(memcg, ss);
5701 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5703 mem_cgroup_sockets_destroy(memcg);
5706 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5708 if (!memcg_kmem_is_active(memcg))
5712 * kmem charges can outlive the cgroup. In the case of slab
5713 * pages, for instance, a page contain objects from various
5714 * processes. As we prevent from taking a reference for every
5715 * such allocation we have to be careful when doing uncharge
5716 * (see memcg_uncharge_kmem) and here during offlining.
5718 * The idea is that that only the _last_ uncharge which sees
5719 * the dead memcg will drop the last reference. An additional
5720 * reference is taken here before the group is marked dead
5721 * which is then paired with css_put during uncharge resp. here.
5723 * Although this might sound strange as this path is called from
5724 * css_offline() when the referencemight have dropped down to 0
5725 * and shouldn't be incremented anymore (css_tryget would fail)
5726 * we do not have other options because of the kmem allocations
5729 css_get(&memcg->css);
5731 memcg_kmem_mark_dead(memcg);
5733 if (res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0)
5736 if (memcg_kmem_test_and_clear_dead(memcg))
5737 css_put(&memcg->css);
5740 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5745 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5749 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5755 * DO NOT USE IN NEW FILES.
5757 * "cgroup.event_control" implementation.
5759 * This is way over-engineered. It tries to support fully configurable
5760 * events for each user. Such level of flexibility is completely
5761 * unnecessary especially in the light of the planned unified hierarchy.
5763 * Please deprecate this and replace with something simpler if at all
5768 * Unregister event and free resources.
5770 * Gets called from workqueue.
5772 static void memcg_event_remove(struct work_struct *work)
5774 struct mem_cgroup_event *event =
5775 container_of(work, struct mem_cgroup_event, remove);
5776 struct mem_cgroup *memcg = event->memcg;
5778 remove_wait_queue(event->wqh, &event->wait);
5780 event->unregister_event(memcg, event->eventfd);
5782 /* Notify userspace the event is going away. */
5783 eventfd_signal(event->eventfd, 1);
5785 eventfd_ctx_put(event->eventfd);
5787 css_put(&memcg->css);
5791 * Gets called on POLLHUP on eventfd when user closes it.
5793 * Called with wqh->lock held and interrupts disabled.
5795 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
5796 int sync, void *key)
5798 struct mem_cgroup_event *event =
5799 container_of(wait, struct mem_cgroup_event, wait);
5800 struct mem_cgroup *memcg = event->memcg;
5801 unsigned long flags = (unsigned long)key;
5803 if (flags & POLLHUP) {
5805 * If the event has been detached at cgroup removal, we
5806 * can simply return knowing the other side will cleanup
5809 * We can't race against event freeing since the other
5810 * side will require wqh->lock via remove_wait_queue(),
5813 spin_lock(&memcg->event_list_lock);
5814 if (!list_empty(&event->list)) {
5815 list_del_init(&event->list);
5817 * We are in atomic context, but cgroup_event_remove()
5818 * may sleep, so we have to call it in workqueue.
5820 schedule_work(&event->remove);
5822 spin_unlock(&memcg->event_list_lock);
5828 static void memcg_event_ptable_queue_proc(struct file *file,
5829 wait_queue_head_t *wqh, poll_table *pt)
5831 struct mem_cgroup_event *event =
5832 container_of(pt, struct mem_cgroup_event, pt);
5835 add_wait_queue(wqh, &event->wait);
5839 * DO NOT USE IN NEW FILES.
5841 * Parse input and register new cgroup event handler.
5843 * Input must be in format '<event_fd> <control_fd> <args>'.
5844 * Interpretation of args is defined by control file implementation.
5846 static int memcg_write_event_control(struct cgroup_subsys_state *css,
5847 struct cftype *cft, char *buffer)
5849 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5850 struct mem_cgroup_event *event;
5851 struct cgroup_subsys_state *cfile_css;
5852 unsigned int efd, cfd;
5859 efd = simple_strtoul(buffer, &endp, 10);
5864 cfd = simple_strtoul(buffer, &endp, 10);
5865 if ((*endp != ' ') && (*endp != '\0'))
5869 event = kzalloc(sizeof(*event), GFP_KERNEL);
5873 event->memcg = memcg;
5874 INIT_LIST_HEAD(&event->list);
5875 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
5876 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
5877 INIT_WORK(&event->remove, memcg_event_remove);
5885 event->eventfd = eventfd_ctx_fileget(efile.file);
5886 if (IS_ERR(event->eventfd)) {
5887 ret = PTR_ERR(event->eventfd);
5894 goto out_put_eventfd;
5897 /* the process need read permission on control file */
5898 /* AV: shouldn't we check that it's been opened for read instead? */
5899 ret = inode_permission(file_inode(cfile.file), MAY_READ);
5904 * Determine the event callbacks and set them in @event. This used
5905 * to be done via struct cftype but cgroup core no longer knows
5906 * about these events. The following is crude but the whole thing
5907 * is for compatibility anyway.
5909 * DO NOT ADD NEW FILES.
5911 name = cfile.file->f_dentry->d_name.name;
5913 if (!strcmp(name, "memory.usage_in_bytes")) {
5914 event->register_event = mem_cgroup_usage_register_event;
5915 event->unregister_event = mem_cgroup_usage_unregister_event;
5916 } else if (!strcmp(name, "memory.oom_control")) {
5917 event->register_event = mem_cgroup_oom_register_event;
5918 event->unregister_event = mem_cgroup_oom_unregister_event;
5919 } else if (!strcmp(name, "memory.pressure_level")) {
5920 event->register_event = vmpressure_register_event;
5921 event->unregister_event = vmpressure_unregister_event;
5922 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
5923 event->register_event = memsw_cgroup_usage_register_event;
5924 event->unregister_event = memsw_cgroup_usage_unregister_event;
5931 * Verify @cfile should belong to @css. Also, remaining events are
5932 * automatically removed on cgroup destruction but the removal is
5933 * asynchronous, so take an extra ref on @css.
5935 cfile_css = css_tryget_from_dir(cfile.file->f_dentry->d_parent,
5936 &memory_cgrp_subsys);
5938 if (IS_ERR(cfile_css))
5940 if (cfile_css != css) {
5945 ret = event->register_event(memcg, event->eventfd, buffer);
5949 efile.file->f_op->poll(efile.file, &event->pt);
5951 spin_lock(&memcg->event_list_lock);
5952 list_add(&event->list, &memcg->event_list);
5953 spin_unlock(&memcg->event_list_lock);
5965 eventfd_ctx_put(event->eventfd);
5974 static struct cftype mem_cgroup_files[] = {
5976 .name = "usage_in_bytes",
5977 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
5978 .read_u64 = mem_cgroup_read_u64,
5981 .name = "max_usage_in_bytes",
5982 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
5983 .trigger = mem_cgroup_reset,
5984 .read_u64 = mem_cgroup_read_u64,
5987 .name = "limit_in_bytes",
5988 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
5989 .write_string = mem_cgroup_write,
5990 .read_u64 = mem_cgroup_read_u64,
5993 .name = "soft_limit_in_bytes",
5994 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
5995 .write_string = mem_cgroup_write,
5996 .read_u64 = mem_cgroup_read_u64,
6000 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
6001 .trigger = mem_cgroup_reset,
6002 .read_u64 = mem_cgroup_read_u64,
6006 .seq_show = memcg_stat_show,
6009 .name = "force_empty",
6010 .trigger = mem_cgroup_force_empty_write,
6013 .name = "use_hierarchy",
6014 .flags = CFTYPE_INSANE,
6015 .write_u64 = mem_cgroup_hierarchy_write,
6016 .read_u64 = mem_cgroup_hierarchy_read,
6019 .name = "cgroup.event_control", /* XXX: for compat */
6020 .write_string = memcg_write_event_control,
6021 .flags = CFTYPE_NO_PREFIX,
6025 .name = "swappiness",
6026 .read_u64 = mem_cgroup_swappiness_read,
6027 .write_u64 = mem_cgroup_swappiness_write,
6030 .name = "move_charge_at_immigrate",
6031 .read_u64 = mem_cgroup_move_charge_read,
6032 .write_u64 = mem_cgroup_move_charge_write,
6035 .name = "oom_control",
6036 .seq_show = mem_cgroup_oom_control_read,
6037 .write_u64 = mem_cgroup_oom_control_write,
6038 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
6041 .name = "pressure_level",
6045 .name = "numa_stat",
6046 .seq_show = memcg_numa_stat_show,
6049 #ifdef CONFIG_MEMCG_KMEM
6051 .name = "kmem.limit_in_bytes",
6052 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
6053 .write_string = mem_cgroup_write,
6054 .read_u64 = mem_cgroup_read_u64,
6057 .name = "kmem.usage_in_bytes",
6058 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
6059 .read_u64 = mem_cgroup_read_u64,
6062 .name = "kmem.failcnt",
6063 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
6064 .trigger = mem_cgroup_reset,
6065 .read_u64 = mem_cgroup_read_u64,
6068 .name = "kmem.max_usage_in_bytes",
6069 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
6070 .trigger = mem_cgroup_reset,
6071 .read_u64 = mem_cgroup_read_u64,
6073 #ifdef CONFIG_SLABINFO
6075 .name = "kmem.slabinfo",
6076 .seq_show = mem_cgroup_slabinfo_read,
6080 { }, /* terminate */
6083 #ifdef CONFIG_MEMCG_SWAP
6084 static struct cftype memsw_cgroup_files[] = {
6086 .name = "memsw.usage_in_bytes",
6087 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6088 .read_u64 = mem_cgroup_read_u64,
6091 .name = "memsw.max_usage_in_bytes",
6092 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6093 .trigger = mem_cgroup_reset,
6094 .read_u64 = mem_cgroup_read_u64,
6097 .name = "memsw.limit_in_bytes",
6098 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6099 .write_string = mem_cgroup_write,
6100 .read_u64 = mem_cgroup_read_u64,
6103 .name = "memsw.failcnt",
6104 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6105 .trigger = mem_cgroup_reset,
6106 .read_u64 = mem_cgroup_read_u64,
6108 { }, /* terminate */
6111 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6113 struct mem_cgroup_per_node *pn;
6114 struct mem_cgroup_per_zone *mz;
6115 int zone, tmp = node;
6117 * This routine is called against possible nodes.
6118 * But it's BUG to call kmalloc() against offline node.
6120 * TODO: this routine can waste much memory for nodes which will
6121 * never be onlined. It's better to use memory hotplug callback
6124 if (!node_state(node, N_NORMAL_MEMORY))
6126 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
6130 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6131 mz = &pn->zoneinfo[zone];
6132 lruvec_init(&mz->lruvec);
6133 mz->usage_in_excess = 0;
6134 mz->on_tree = false;
6137 memcg->nodeinfo[node] = pn;
6141 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6143 kfree(memcg->nodeinfo[node]);
6146 static struct mem_cgroup *mem_cgroup_alloc(void)
6148 struct mem_cgroup *memcg;
6151 size = sizeof(struct mem_cgroup);
6152 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
6154 memcg = kzalloc(size, GFP_KERNEL);
6158 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
6161 spin_lock_init(&memcg->pcp_counter_lock);
6170 * At destroying mem_cgroup, references from swap_cgroup can remain.
6171 * (scanning all at force_empty is too costly...)
6173 * Instead of clearing all references at force_empty, we remember
6174 * the number of reference from swap_cgroup and free mem_cgroup when
6175 * it goes down to 0.
6177 * Removal of cgroup itself succeeds regardless of refs from swap.
6180 static void __mem_cgroup_free(struct mem_cgroup *memcg)
6184 mem_cgroup_remove_from_trees(memcg);
6187 free_mem_cgroup_per_zone_info(memcg, node);
6189 free_percpu(memcg->stat);
6192 * We need to make sure that (at least for now), the jump label
6193 * destruction code runs outside of the cgroup lock. This is because
6194 * get_online_cpus(), which is called from the static_branch update,
6195 * can't be called inside the cgroup_lock. cpusets are the ones
6196 * enforcing this dependency, so if they ever change, we might as well.
6198 * schedule_work() will guarantee this happens. Be careful if you need
6199 * to move this code around, and make sure it is outside
6202 disarm_static_keys(memcg);
6207 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6209 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
6211 if (!memcg->res.parent)
6213 return mem_cgroup_from_res_counter(memcg->res.parent, res);
6215 EXPORT_SYMBOL(parent_mem_cgroup);
6217 static void __init mem_cgroup_soft_limit_tree_init(void)
6219 struct mem_cgroup_tree_per_node *rtpn;
6220 struct mem_cgroup_tree_per_zone *rtpz;
6221 int tmp, node, zone;
6223 for_each_node(node) {
6225 if (!node_state(node, N_NORMAL_MEMORY))
6227 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
6230 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6232 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6233 rtpz = &rtpn->rb_tree_per_zone[zone];
6234 rtpz->rb_root = RB_ROOT;
6235 spin_lock_init(&rtpz->lock);
6240 static struct cgroup_subsys_state * __ref
6241 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
6243 struct mem_cgroup *memcg;
6244 long error = -ENOMEM;
6247 memcg = mem_cgroup_alloc();
6249 return ERR_PTR(error);
6252 if (alloc_mem_cgroup_per_zone_info(memcg, node))
6256 if (parent_css == NULL) {
6257 root_mem_cgroup = memcg;
6258 res_counter_init(&memcg->res, NULL);
6259 res_counter_init(&memcg->memsw, NULL);
6260 res_counter_init(&memcg->kmem, NULL);
6263 memcg->last_scanned_node = MAX_NUMNODES;
6264 INIT_LIST_HEAD(&memcg->oom_notify);
6265 memcg->move_charge_at_immigrate = 0;
6266 mutex_init(&memcg->thresholds_lock);
6267 spin_lock_init(&memcg->move_lock);
6268 vmpressure_init(&memcg->vmpressure);
6269 INIT_LIST_HEAD(&memcg->event_list);
6270 spin_lock_init(&memcg->event_list_lock);
6275 __mem_cgroup_free(memcg);
6276 return ERR_PTR(error);
6280 mem_cgroup_css_online(struct cgroup_subsys_state *css)
6282 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6283 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(css));
6285 if (css->cgroup->id > MEM_CGROUP_ID_MAX)
6291 mutex_lock(&memcg_create_mutex);
6293 memcg->use_hierarchy = parent->use_hierarchy;
6294 memcg->oom_kill_disable = parent->oom_kill_disable;
6295 memcg->swappiness = mem_cgroup_swappiness(parent);
6297 if (parent->use_hierarchy) {
6298 res_counter_init(&memcg->res, &parent->res);
6299 res_counter_init(&memcg->memsw, &parent->memsw);
6300 res_counter_init(&memcg->kmem, &parent->kmem);
6303 * No need to take a reference to the parent because cgroup
6304 * core guarantees its existence.
6307 res_counter_init(&memcg->res, NULL);
6308 res_counter_init(&memcg->memsw, NULL);
6309 res_counter_init(&memcg->kmem, NULL);
6311 * Deeper hierachy with use_hierarchy == false doesn't make
6312 * much sense so let cgroup subsystem know about this
6313 * unfortunate state in our controller.
6315 if (parent != root_mem_cgroup)
6316 memory_cgrp_subsys.broken_hierarchy = true;
6318 mutex_unlock(&memcg_create_mutex);
6320 return memcg_init_kmem(memcg, &memory_cgrp_subsys);
6324 * Announce all parents that a group from their hierarchy is gone.
6326 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup *memcg)
6328 struct mem_cgroup *parent = memcg;
6330 while ((parent = parent_mem_cgroup(parent)))
6331 mem_cgroup_iter_invalidate(parent);
6334 * if the root memcg is not hierarchical we have to check it
6337 if (!root_mem_cgroup->use_hierarchy)
6338 mem_cgroup_iter_invalidate(root_mem_cgroup);
6341 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
6343 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6344 struct mem_cgroup_event *event, *tmp;
6345 struct cgroup_subsys_state *iter;
6348 * Unregister events and notify userspace.
6349 * Notify userspace about cgroup removing only after rmdir of cgroup
6350 * directory to avoid race between userspace and kernelspace.
6352 spin_lock(&memcg->event_list_lock);
6353 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
6354 list_del_init(&event->list);
6355 schedule_work(&event->remove);
6357 spin_unlock(&memcg->event_list_lock);
6359 kmem_cgroup_css_offline(memcg);
6361 mem_cgroup_invalidate_reclaim_iterators(memcg);
6364 * This requires that offlining is serialized. Right now that is
6365 * guaranteed because css_killed_work_fn() holds the cgroup_mutex.
6367 css_for_each_descendant_post(iter, css)
6368 mem_cgroup_reparent_charges(mem_cgroup_from_css(iter));
6370 memcg_unregister_all_caches(memcg);
6371 vmpressure_cleanup(&memcg->vmpressure);
6374 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
6376 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6378 * XXX: css_offline() would be where we should reparent all
6379 * memory to prepare the cgroup for destruction. However,
6380 * memcg does not do css_tryget() and res_counter charging
6381 * under the same RCU lock region, which means that charging
6382 * could race with offlining. Offlining only happens to
6383 * cgroups with no tasks in them but charges can show up
6384 * without any tasks from the swapin path when the target
6385 * memcg is looked up from the swapout record and not from the
6386 * current task as it usually is. A race like this can leak
6387 * charges and put pages with stale cgroup pointers into
6391 * lookup_swap_cgroup_id()
6393 * mem_cgroup_lookup()
6396 * disable css_tryget()
6399 * reparent_charges()
6400 * res_counter_charge()
6403 * pc->mem_cgroup = dead memcg
6406 * The bulk of the charges are still moved in offline_css() to
6407 * avoid pinning a lot of pages in case a long-term reference
6408 * like a swapout record is deferring the css_free() to long
6409 * after offlining. But this makes sure we catch any charges
6410 * made after offlining:
6412 mem_cgroup_reparent_charges(memcg);
6414 memcg_destroy_kmem(memcg);
6415 __mem_cgroup_free(memcg);
6419 /* Handlers for move charge at task migration. */
6420 #define PRECHARGE_COUNT_AT_ONCE 256
6421 static int mem_cgroup_do_precharge(unsigned long count)
6424 int batch_count = PRECHARGE_COUNT_AT_ONCE;
6425 struct mem_cgroup *memcg = mc.to;
6427 if (mem_cgroup_is_root(memcg)) {
6428 mc.precharge += count;
6429 /* we don't need css_get for root */
6432 /* try to charge at once */
6434 struct res_counter *dummy;
6436 * "memcg" cannot be under rmdir() because we've already checked
6437 * by cgroup_lock_live_cgroup() that it is not removed and we
6438 * are still under the same cgroup_mutex. So we can postpone
6441 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
6443 if (do_swap_account && res_counter_charge(&memcg->memsw,
6444 PAGE_SIZE * count, &dummy)) {
6445 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
6448 mc.precharge += count;
6452 /* fall back to one by one charge */
6454 if (signal_pending(current)) {
6458 if (!batch_count--) {
6459 batch_count = PRECHARGE_COUNT_AT_ONCE;
6462 ret = mem_cgroup_try_charge(memcg, GFP_KERNEL, 1, false);
6464 /* mem_cgroup_clear_mc() will do uncharge later */
6472 * get_mctgt_type - get target type of moving charge
6473 * @vma: the vma the pte to be checked belongs
6474 * @addr: the address corresponding to the pte to be checked
6475 * @ptent: the pte to be checked
6476 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6479 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6480 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6481 * move charge. if @target is not NULL, the page is stored in target->page
6482 * with extra refcnt got(Callers should handle it).
6483 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6484 * target for charge migration. if @target is not NULL, the entry is stored
6487 * Called with pte lock held.
6494 enum mc_target_type {
6500 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
6501 unsigned long addr, pte_t ptent)
6503 struct page *page = vm_normal_page(vma, addr, ptent);
6505 if (!page || !page_mapped(page))
6507 if (PageAnon(page)) {
6508 /* we don't move shared anon */
6511 } else if (!move_file())
6512 /* we ignore mapcount for file pages */
6514 if (!get_page_unless_zero(page))
6521 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6522 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6524 struct page *page = NULL;
6525 swp_entry_t ent = pte_to_swp_entry(ptent);
6527 if (!move_anon() || non_swap_entry(ent))
6530 * Because lookup_swap_cache() updates some statistics counter,
6531 * we call find_get_page() with swapper_space directly.
6533 page = find_get_page(swap_address_space(ent), ent.val);
6534 if (do_swap_account)
6535 entry->val = ent.val;
6540 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6541 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6547 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
6548 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6550 struct page *page = NULL;
6551 struct address_space *mapping;
6554 if (!vma->vm_file) /* anonymous vma */
6559 mapping = vma->vm_file->f_mapping;
6560 if (pte_none(ptent))
6561 pgoff = linear_page_index(vma, addr);
6562 else /* pte_file(ptent) is true */
6563 pgoff = pte_to_pgoff(ptent);
6565 /* page is moved even if it's not RSS of this task(page-faulted). */
6567 /* shmem/tmpfs may report page out on swap: account for that too. */
6568 if (shmem_mapping(mapping)) {
6569 page = find_get_entry(mapping, pgoff);
6570 if (radix_tree_exceptional_entry(page)) {
6571 swp_entry_t swp = radix_to_swp_entry(page);
6572 if (do_swap_account)
6574 page = find_get_page(swap_address_space(swp), swp.val);
6577 page = find_get_page(mapping, pgoff);
6579 page = find_get_page(mapping, pgoff);
6584 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
6585 unsigned long addr, pte_t ptent, union mc_target *target)
6587 struct page *page = NULL;
6588 struct page_cgroup *pc;
6589 enum mc_target_type ret = MC_TARGET_NONE;
6590 swp_entry_t ent = { .val = 0 };
6592 if (pte_present(ptent))
6593 page = mc_handle_present_pte(vma, addr, ptent);
6594 else if (is_swap_pte(ptent))
6595 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
6596 else if (pte_none(ptent) || pte_file(ptent))
6597 page = mc_handle_file_pte(vma, addr, ptent, &ent);
6599 if (!page && !ent.val)
6602 pc = lookup_page_cgroup(page);
6604 * Do only loose check w/o page_cgroup lock.
6605 * mem_cgroup_move_account() checks the pc is valid or not under
6608 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6609 ret = MC_TARGET_PAGE;
6611 target->page = page;
6613 if (!ret || !target)
6616 /* There is a swap entry and a page doesn't exist or isn't charged */
6617 if (ent.val && !ret &&
6618 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
6619 ret = MC_TARGET_SWAP;
6626 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6628 * We don't consider swapping or file mapped pages because THP does not
6629 * support them for now.
6630 * Caller should make sure that pmd_trans_huge(pmd) is true.
6632 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6633 unsigned long addr, pmd_t pmd, union mc_target *target)
6635 struct page *page = NULL;
6636 struct page_cgroup *pc;
6637 enum mc_target_type ret = MC_TARGET_NONE;
6639 page = pmd_page(pmd);
6640 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
6643 pc = lookup_page_cgroup(page);
6644 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6645 ret = MC_TARGET_PAGE;
6648 target->page = page;
6654 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6655 unsigned long addr, pmd_t pmd, union mc_target *target)
6657 return MC_TARGET_NONE;
6661 static int mem_cgroup_count_precharge_pte(pte_t *pte,
6662 unsigned long addr, unsigned long end,
6663 struct mm_walk *walk)
6665 if (get_mctgt_type(walk->vma, addr, *pte, NULL))
6666 mc.precharge++; /* increment precharge temporarily */
6670 static int mem_cgroup_count_precharge_pmd(pmd_t *pmd,
6671 unsigned long addr, unsigned long end,
6672 struct mm_walk *walk)
6674 struct vm_area_struct *vma = walk->vma;
6677 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
6678 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
6679 mc.precharge += HPAGE_PMD_NR;
6681 /* don't call mem_cgroup_count_precharge_pte() */
6687 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
6689 unsigned long precharge;
6690 struct vm_area_struct *vma;
6692 struct mm_walk mem_cgroup_count_precharge_walk = {
6693 .pmd_entry = mem_cgroup_count_precharge_pmd,
6694 .pte_entry = mem_cgroup_count_precharge_pte,
6697 down_read(&mm->mmap_sem);
6698 for (vma = mm->mmap; vma; vma = vma->vm_next)
6699 walk_page_vma(vma, &mem_cgroup_count_precharge_walk);
6700 up_read(&mm->mmap_sem);
6702 precharge = mc.precharge;
6708 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
6710 unsigned long precharge = mem_cgroup_count_precharge(mm);
6712 VM_BUG_ON(mc.moving_task);
6713 mc.moving_task = current;
6714 return mem_cgroup_do_precharge(precharge);
6717 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6718 static void __mem_cgroup_clear_mc(void)
6720 struct mem_cgroup *from = mc.from;
6721 struct mem_cgroup *to = mc.to;
6724 /* we must uncharge all the leftover precharges from mc.to */
6726 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
6730 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6731 * we must uncharge here.
6733 if (mc.moved_charge) {
6734 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
6735 mc.moved_charge = 0;
6737 /* we must fixup refcnts and charges */
6738 if (mc.moved_swap) {
6739 /* uncharge swap account from the old cgroup */
6740 if (!mem_cgroup_is_root(mc.from))
6741 res_counter_uncharge(&mc.from->memsw,
6742 PAGE_SIZE * mc.moved_swap);
6744 for (i = 0; i < mc.moved_swap; i++)
6745 css_put(&mc.from->css);
6747 if (!mem_cgroup_is_root(mc.to)) {
6749 * we charged both to->res and to->memsw, so we should
6752 res_counter_uncharge(&mc.to->res,
6753 PAGE_SIZE * mc.moved_swap);
6755 /* we've already done css_get(mc.to) */
6758 memcg_oom_recover(from);
6759 memcg_oom_recover(to);
6760 wake_up_all(&mc.waitq);
6763 static void mem_cgroup_clear_mc(void)
6765 struct mem_cgroup *from = mc.from;
6768 * we must clear moving_task before waking up waiters at the end of
6771 mc.moving_task = NULL;
6772 __mem_cgroup_clear_mc();
6773 spin_lock(&mc.lock);
6776 spin_unlock(&mc.lock);
6777 mem_cgroup_end_move(from);
6780 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
6781 struct cgroup_taskset *tset)
6783 struct task_struct *p = cgroup_taskset_first(tset);
6785 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6786 unsigned long move_charge_at_immigrate;
6789 * We are now commited to this value whatever it is. Changes in this
6790 * tunable will only affect upcoming migrations, not the current one.
6791 * So we need to save it, and keep it going.
6793 move_charge_at_immigrate = memcg->move_charge_at_immigrate;
6794 if (move_charge_at_immigrate) {
6795 struct mm_struct *mm;
6796 struct mem_cgroup *from = mem_cgroup_from_task(p);
6798 VM_BUG_ON(from == memcg);
6800 mm = get_task_mm(p);
6803 /* We move charges only when we move a owner of the mm */
6804 if (mm->owner == p) {
6807 VM_BUG_ON(mc.precharge);
6808 VM_BUG_ON(mc.moved_charge);
6809 VM_BUG_ON(mc.moved_swap);
6810 mem_cgroup_start_move(from);
6811 spin_lock(&mc.lock);
6814 mc.immigrate_flags = move_charge_at_immigrate;
6815 spin_unlock(&mc.lock);
6816 /* We set mc.moving_task later */
6818 ret = mem_cgroup_precharge_mc(mm);
6820 mem_cgroup_clear_mc();
6827 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
6828 struct cgroup_taskset *tset)
6830 mem_cgroup_clear_mc();
6833 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6834 unsigned long addr, unsigned long end,
6835 struct mm_walk *walk)
6838 struct vm_area_struct *vma = walk->vma;
6841 enum mc_target_type target_type;
6842 union mc_target target;
6844 struct page_cgroup *pc;
6847 * We don't take compound_lock() here but no race with splitting thp
6849 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6850 * under splitting, which means there's no concurrent thp split,
6851 * - if another thread runs into split_huge_page() just after we
6852 * entered this if-block, the thread must wait for page table lock
6853 * to be unlocked in __split_huge_page_splitting(), where the main
6854 * part of thp split is not executed yet.
6856 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
6857 if (mc.precharge < HPAGE_PMD_NR) {
6861 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6862 if (target_type == MC_TARGET_PAGE) {
6864 if (!isolate_lru_page(page)) {
6865 pc = lookup_page_cgroup(page);
6866 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
6867 pc, mc.from, mc.to)) {
6868 mc.precharge -= HPAGE_PMD_NR;
6869 mc.moved_charge += HPAGE_PMD_NR;
6871 putback_lru_page(page);
6879 if (pmd_trans_unstable(pmd))
6882 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6883 for (; addr != end; addr += PAGE_SIZE) {
6884 pte_t ptent = *(pte++);
6890 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6891 case MC_TARGET_PAGE:
6893 if (isolate_lru_page(page))
6895 pc = lookup_page_cgroup(page);
6896 if (!mem_cgroup_move_account(page, 1, pc,
6899 /* we uncharge from mc.from later. */
6902 putback_lru_page(page);
6903 put: /* get_mctgt_type() gets the page */
6906 case MC_TARGET_SWAP:
6908 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6910 /* we fixup refcnts and charges later. */
6918 pte_unmap_unlock(pte - 1, ptl);
6923 * We have consumed all precharges we got in can_attach().
6924 * We try charge one by one, but don't do any additional
6925 * charges to mc.to if we have failed in charge once in attach()
6928 ret = mem_cgroup_do_precharge(1);
6936 static void mem_cgroup_move_charge(struct mm_struct *mm)
6938 struct vm_area_struct *vma;
6939 struct mm_walk mem_cgroup_move_charge_walk = {
6940 .pmd_entry = mem_cgroup_move_charge_pte_range,
6944 lru_add_drain_all();
6946 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
6948 * Someone who are holding the mmap_sem might be waiting in
6949 * waitq. So we cancel all extra charges, wake up all waiters,
6950 * and retry. Because we cancel precharges, we might not be able
6951 * to move enough charges, but moving charge is a best-effort
6952 * feature anyway, so it wouldn't be a big problem.
6954 __mem_cgroup_clear_mc();
6958 for (vma = mm->mmap; vma; vma = vma->vm_next)
6959 walk_page_vma(vma, &mem_cgroup_move_charge_walk);
6960 up_read(&mm->mmap_sem);
6963 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
6964 struct cgroup_taskset *tset)
6966 struct task_struct *p = cgroup_taskset_first(tset);
6967 struct mm_struct *mm = get_task_mm(p);
6971 mem_cgroup_move_charge(mm);
6975 mem_cgroup_clear_mc();
6977 #else /* !CONFIG_MMU */
6978 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
6979 struct cgroup_taskset *tset)
6983 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
6984 struct cgroup_taskset *tset)
6987 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
6988 struct cgroup_taskset *tset)
6994 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6995 * to verify sane_behavior flag on each mount attempt.
6997 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
7000 * use_hierarchy is forced with sane_behavior. cgroup core
7001 * guarantees that @root doesn't have any children, so turning it
7002 * on for the root memcg is enough.
7004 if (cgroup_sane_behavior(root_css->cgroup))
7005 mem_cgroup_from_css(root_css)->use_hierarchy = true;
7008 struct cgroup_subsys memory_cgrp_subsys = {
7009 .css_alloc = mem_cgroup_css_alloc,
7010 .css_online = mem_cgroup_css_online,
7011 .css_offline = mem_cgroup_css_offline,
7012 .css_free = mem_cgroup_css_free,
7013 .can_attach = mem_cgroup_can_attach,
7014 .cancel_attach = mem_cgroup_cancel_attach,
7015 .attach = mem_cgroup_move_task,
7016 .bind = mem_cgroup_bind,
7017 .base_cftypes = mem_cgroup_files,
7021 #ifdef CONFIG_MEMCG_SWAP
7022 static int __init enable_swap_account(char *s)
7024 if (!strcmp(s, "1"))
7025 really_do_swap_account = 1;
7026 else if (!strcmp(s, "0"))
7027 really_do_swap_account = 0;
7030 __setup("swapaccount=", enable_swap_account);
7032 static void __init memsw_file_init(void)
7034 WARN_ON(cgroup_add_cftypes(&memory_cgrp_subsys, memsw_cgroup_files));
7037 static void __init enable_swap_cgroup(void)
7039 if (!mem_cgroup_disabled() && really_do_swap_account) {
7040 do_swap_account = 1;
7046 static void __init enable_swap_cgroup(void)
7052 * subsys_initcall() for memory controller.
7054 * Some parts like hotcpu_notifier() have to be initialized from this context
7055 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
7056 * everything that doesn't depend on a specific mem_cgroup structure should
7057 * be initialized from here.
7059 static int __init mem_cgroup_init(void)
7061 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
7062 enable_swap_cgroup();
7063 mem_cgroup_soft_limit_tree_init();
7067 subsys_initcall(mem_cgroup_init);