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 * 2 routines for checking "mem" is under move_account() or not.
1590 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1591 * is used for avoiding races in accounting. If true,
1592 * pc->mem_cgroup may be overwritten.
1594 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1595 * under hierarchy of moving cgroups. This is for
1596 * waiting at hith-memory prressure caused by "move".
1599 static bool mem_cgroup_stolen(struct mem_cgroup *memcg)
1601 VM_BUG_ON(!rcu_read_lock_held());
1602 return atomic_read(&memcg->moving_account) > 0;
1605 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1607 struct mem_cgroup *from;
1608 struct mem_cgroup *to;
1611 * Unlike task_move routines, we access mc.to, mc.from not under
1612 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1614 spin_lock(&mc.lock);
1620 ret = mem_cgroup_same_or_subtree(memcg, from)
1621 || mem_cgroup_same_or_subtree(memcg, to);
1623 spin_unlock(&mc.lock);
1627 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1629 if (mc.moving_task && current != mc.moving_task) {
1630 if (mem_cgroup_under_move(memcg)) {
1632 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1633 /* moving charge context might have finished. */
1636 finish_wait(&mc.waitq, &wait);
1644 * Take this lock when
1645 * - a code tries to modify page's memcg while it's USED.
1646 * - a code tries to modify page state accounting in a memcg.
1647 * see mem_cgroup_stolen(), too.
1649 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1650 unsigned long *flags)
1652 spin_lock_irqsave(&memcg->move_lock, *flags);
1655 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1656 unsigned long *flags)
1658 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1661 #define K(x) ((x) << (PAGE_SHIFT-10))
1663 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1664 * @memcg: The memory cgroup that went over limit
1665 * @p: Task that is going to be killed
1667 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1670 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1672 /* oom_info_lock ensures that parallel ooms do not interleave */
1673 static DEFINE_MUTEX(oom_info_lock);
1674 struct mem_cgroup *iter;
1680 mutex_lock(&oom_info_lock);
1683 pr_info("Task in ");
1684 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1685 pr_info(" killed as a result of limit of ");
1686 pr_cont_cgroup_path(memcg->css.cgroup);
1691 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1692 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1693 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1694 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1695 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1696 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1697 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1698 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1699 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1700 res_counter_read_u64(&memcg->kmem, RES_USAGE) >> 10,
1701 res_counter_read_u64(&memcg->kmem, RES_LIMIT) >> 10,
1702 res_counter_read_u64(&memcg->kmem, RES_FAILCNT));
1704 for_each_mem_cgroup_tree(iter, memcg) {
1705 pr_info("Memory cgroup stats for ");
1706 pr_cont_cgroup_path(iter->css.cgroup);
1709 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1710 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1712 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1713 K(mem_cgroup_read_stat(iter, i)));
1716 for (i = 0; i < NR_LRU_LISTS; i++)
1717 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1718 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1722 mutex_unlock(&oom_info_lock);
1726 * This function returns the number of memcg under hierarchy tree. Returns
1727 * 1(self count) if no children.
1729 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1732 struct mem_cgroup *iter;
1734 for_each_mem_cgroup_tree(iter, memcg)
1740 * Return the memory (and swap, if configured) limit for a memcg.
1742 static u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1746 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1749 * Do not consider swap space if we cannot swap due to swappiness
1751 if (mem_cgroup_swappiness(memcg)) {
1754 limit += total_swap_pages << PAGE_SHIFT;
1755 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1758 * If memsw is finite and limits the amount of swap space
1759 * available to this memcg, return that limit.
1761 limit = min(limit, memsw);
1767 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1770 struct mem_cgroup *iter;
1771 unsigned long chosen_points = 0;
1772 unsigned long totalpages;
1773 unsigned int points = 0;
1774 struct task_struct *chosen = NULL;
1777 * If current has a pending SIGKILL or is exiting, then automatically
1778 * select it. The goal is to allow it to allocate so that it may
1779 * quickly exit and free its memory.
1781 if (fatal_signal_pending(current) || current->flags & PF_EXITING) {
1782 set_thread_flag(TIF_MEMDIE);
1786 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1787 totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
1788 for_each_mem_cgroup_tree(iter, memcg) {
1789 struct css_task_iter it;
1790 struct task_struct *task;
1792 css_task_iter_start(&iter->css, &it);
1793 while ((task = css_task_iter_next(&it))) {
1794 switch (oom_scan_process_thread(task, totalpages, NULL,
1796 case OOM_SCAN_SELECT:
1798 put_task_struct(chosen);
1800 chosen_points = ULONG_MAX;
1801 get_task_struct(chosen);
1803 case OOM_SCAN_CONTINUE:
1805 case OOM_SCAN_ABORT:
1806 css_task_iter_end(&it);
1807 mem_cgroup_iter_break(memcg, iter);
1809 put_task_struct(chosen);
1814 points = oom_badness(task, memcg, NULL, totalpages);
1815 if (!points || points < chosen_points)
1817 /* Prefer thread group leaders for display purposes */
1818 if (points == chosen_points &&
1819 thread_group_leader(chosen))
1823 put_task_struct(chosen);
1825 chosen_points = points;
1826 get_task_struct(chosen);
1828 css_task_iter_end(&it);
1833 points = chosen_points * 1000 / totalpages;
1834 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1835 NULL, "Memory cgroup out of memory");
1838 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1840 unsigned long flags)
1842 unsigned long total = 0;
1843 bool noswap = false;
1846 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1848 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1851 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1853 drain_all_stock_async(memcg);
1854 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1856 * Allow limit shrinkers, which are triggered directly
1857 * by userspace, to catch signals and stop reclaim
1858 * after minimal progress, regardless of the margin.
1860 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1862 if (mem_cgroup_margin(memcg))
1865 * If nothing was reclaimed after two attempts, there
1866 * may be no reclaimable pages in this hierarchy.
1875 * test_mem_cgroup_node_reclaimable
1876 * @memcg: the target memcg
1877 * @nid: the node ID to be checked.
1878 * @noswap : specify true here if the user wants flle only information.
1880 * This function returns whether the specified memcg contains any
1881 * reclaimable pages on a node. Returns true if there are any reclaimable
1882 * pages in the node.
1884 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1885 int nid, bool noswap)
1887 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1889 if (noswap || !total_swap_pages)
1891 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1896 #if MAX_NUMNODES > 1
1899 * Always updating the nodemask is not very good - even if we have an empty
1900 * list or the wrong list here, we can start from some node and traverse all
1901 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1904 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1908 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1909 * pagein/pageout changes since the last update.
1911 if (!atomic_read(&memcg->numainfo_events))
1913 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1916 /* make a nodemask where this memcg uses memory from */
1917 memcg->scan_nodes = node_states[N_MEMORY];
1919 for_each_node_mask(nid, node_states[N_MEMORY]) {
1921 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1922 node_clear(nid, memcg->scan_nodes);
1925 atomic_set(&memcg->numainfo_events, 0);
1926 atomic_set(&memcg->numainfo_updating, 0);
1930 * Selecting a node where we start reclaim from. Because what we need is just
1931 * reducing usage counter, start from anywhere is O,K. Considering
1932 * memory reclaim from current node, there are pros. and cons.
1934 * Freeing memory from current node means freeing memory from a node which
1935 * we'll use or we've used. So, it may make LRU bad. And if several threads
1936 * hit limits, it will see a contention on a node. But freeing from remote
1937 * node means more costs for memory reclaim because of memory latency.
1939 * Now, we use round-robin. Better algorithm is welcomed.
1941 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1945 mem_cgroup_may_update_nodemask(memcg);
1946 node = memcg->last_scanned_node;
1948 node = next_node(node, memcg->scan_nodes);
1949 if (node == MAX_NUMNODES)
1950 node = first_node(memcg->scan_nodes);
1952 * We call this when we hit limit, not when pages are added to LRU.
1953 * No LRU may hold pages because all pages are UNEVICTABLE or
1954 * memcg is too small and all pages are not on LRU. In that case,
1955 * we use curret node.
1957 if (unlikely(node == MAX_NUMNODES))
1958 node = numa_node_id();
1960 memcg->last_scanned_node = node;
1965 * Check all nodes whether it contains reclaimable pages or not.
1966 * For quick scan, we make use of scan_nodes. This will allow us to skip
1967 * unused nodes. But scan_nodes is lazily updated and may not cotain
1968 * enough new information. We need to do double check.
1970 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1975 * quick check...making use of scan_node.
1976 * We can skip unused nodes.
1978 if (!nodes_empty(memcg->scan_nodes)) {
1979 for (nid = first_node(memcg->scan_nodes);
1981 nid = next_node(nid, memcg->scan_nodes)) {
1983 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1988 * Check rest of nodes.
1990 for_each_node_state(nid, N_MEMORY) {
1991 if (node_isset(nid, memcg->scan_nodes))
1993 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
2000 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
2005 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
2007 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
2011 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
2014 unsigned long *total_scanned)
2016 struct mem_cgroup *victim = NULL;
2019 unsigned long excess;
2020 unsigned long nr_scanned;
2021 struct mem_cgroup_reclaim_cookie reclaim = {
2026 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
2029 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
2034 * If we have not been able to reclaim
2035 * anything, it might because there are
2036 * no reclaimable pages under this hierarchy
2041 * We want to do more targeted reclaim.
2042 * excess >> 2 is not to excessive so as to
2043 * reclaim too much, nor too less that we keep
2044 * coming back to reclaim from this cgroup
2046 if (total >= (excess >> 2) ||
2047 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
2052 if (!mem_cgroup_reclaimable(victim, false))
2054 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
2056 *total_scanned += nr_scanned;
2057 if (!res_counter_soft_limit_excess(&root_memcg->res))
2060 mem_cgroup_iter_break(root_memcg, victim);
2064 #ifdef CONFIG_LOCKDEP
2065 static struct lockdep_map memcg_oom_lock_dep_map = {
2066 .name = "memcg_oom_lock",
2070 static DEFINE_SPINLOCK(memcg_oom_lock);
2073 * Check OOM-Killer is already running under our hierarchy.
2074 * If someone is running, return false.
2076 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
2078 struct mem_cgroup *iter, *failed = NULL;
2080 spin_lock(&memcg_oom_lock);
2082 for_each_mem_cgroup_tree(iter, memcg) {
2083 if (iter->oom_lock) {
2085 * this subtree of our hierarchy is already locked
2086 * so we cannot give a lock.
2089 mem_cgroup_iter_break(memcg, iter);
2092 iter->oom_lock = true;
2097 * OK, we failed to lock the whole subtree so we have
2098 * to clean up what we set up to the failing subtree
2100 for_each_mem_cgroup_tree(iter, memcg) {
2101 if (iter == failed) {
2102 mem_cgroup_iter_break(memcg, iter);
2105 iter->oom_lock = false;
2108 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
2110 spin_unlock(&memcg_oom_lock);
2115 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
2117 struct mem_cgroup *iter;
2119 spin_lock(&memcg_oom_lock);
2120 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
2121 for_each_mem_cgroup_tree(iter, memcg)
2122 iter->oom_lock = false;
2123 spin_unlock(&memcg_oom_lock);
2126 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
2128 struct mem_cgroup *iter;
2130 for_each_mem_cgroup_tree(iter, memcg)
2131 atomic_inc(&iter->under_oom);
2134 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
2136 struct mem_cgroup *iter;
2139 * When a new child is created while the hierarchy is under oom,
2140 * mem_cgroup_oom_lock() may not be called. We have to use
2141 * atomic_add_unless() here.
2143 for_each_mem_cgroup_tree(iter, memcg)
2144 atomic_add_unless(&iter->under_oom, -1, 0);
2147 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
2149 struct oom_wait_info {
2150 struct mem_cgroup *memcg;
2154 static int memcg_oom_wake_function(wait_queue_t *wait,
2155 unsigned mode, int sync, void *arg)
2157 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
2158 struct mem_cgroup *oom_wait_memcg;
2159 struct oom_wait_info *oom_wait_info;
2161 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
2162 oom_wait_memcg = oom_wait_info->memcg;
2165 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2166 * Then we can use css_is_ancestor without taking care of RCU.
2168 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
2169 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
2171 return autoremove_wake_function(wait, mode, sync, arg);
2174 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
2176 atomic_inc(&memcg->oom_wakeups);
2177 /* for filtering, pass "memcg" as argument. */
2178 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
2181 static void memcg_oom_recover(struct mem_cgroup *memcg)
2183 if (memcg && atomic_read(&memcg->under_oom))
2184 memcg_wakeup_oom(memcg);
2187 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
2189 if (!current->memcg_oom.may_oom)
2192 * We are in the middle of the charge context here, so we
2193 * don't want to block when potentially sitting on a callstack
2194 * that holds all kinds of filesystem and mm locks.
2196 * Also, the caller may handle a failed allocation gracefully
2197 * (like optional page cache readahead) and so an OOM killer
2198 * invocation might not even be necessary.
2200 * That's why we don't do anything here except remember the
2201 * OOM context and then deal with it at the end of the page
2202 * fault when the stack is unwound, the locks are released,
2203 * and when we know whether the fault was overall successful.
2205 css_get(&memcg->css);
2206 current->memcg_oom.memcg = memcg;
2207 current->memcg_oom.gfp_mask = mask;
2208 current->memcg_oom.order = order;
2212 * mem_cgroup_oom_synchronize - complete memcg OOM handling
2213 * @handle: actually kill/wait or just clean up the OOM state
2215 * This has to be called at the end of a page fault if the memcg OOM
2216 * handler was enabled.
2218 * Memcg supports userspace OOM handling where failed allocations must
2219 * sleep on a waitqueue until the userspace task resolves the
2220 * situation. Sleeping directly in the charge context with all kinds
2221 * of locks held is not a good idea, instead we remember an OOM state
2222 * in the task and mem_cgroup_oom_synchronize() has to be called at
2223 * the end of the page fault to complete the OOM handling.
2225 * Returns %true if an ongoing memcg OOM situation was detected and
2226 * completed, %false otherwise.
2228 bool mem_cgroup_oom_synchronize(bool handle)
2230 struct mem_cgroup *memcg = current->memcg_oom.memcg;
2231 struct oom_wait_info owait;
2234 /* OOM is global, do not handle */
2241 owait.memcg = memcg;
2242 owait.wait.flags = 0;
2243 owait.wait.func = memcg_oom_wake_function;
2244 owait.wait.private = current;
2245 INIT_LIST_HEAD(&owait.wait.task_list);
2247 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2248 mem_cgroup_mark_under_oom(memcg);
2250 locked = mem_cgroup_oom_trylock(memcg);
2253 mem_cgroup_oom_notify(memcg);
2255 if (locked && !memcg->oom_kill_disable) {
2256 mem_cgroup_unmark_under_oom(memcg);
2257 finish_wait(&memcg_oom_waitq, &owait.wait);
2258 mem_cgroup_out_of_memory(memcg, current->memcg_oom.gfp_mask,
2259 current->memcg_oom.order);
2262 mem_cgroup_unmark_under_oom(memcg);
2263 finish_wait(&memcg_oom_waitq, &owait.wait);
2267 mem_cgroup_oom_unlock(memcg);
2269 * There is no guarantee that an OOM-lock contender
2270 * sees the wakeups triggered by the OOM kill
2271 * uncharges. Wake any sleepers explicitely.
2273 memcg_oom_recover(memcg);
2276 current->memcg_oom.memcg = NULL;
2277 css_put(&memcg->css);
2282 * Currently used to update mapped file statistics, but the routine can be
2283 * generalized to update other statistics as well.
2285 * Notes: Race condition
2287 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2288 * it tends to be costly. But considering some conditions, we doesn't need
2289 * to do so _always_.
2291 * Considering "charge", lock_page_cgroup() is not required because all
2292 * file-stat operations happen after a page is attached to radix-tree. There
2293 * are no race with "charge".
2295 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2296 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2297 * if there are race with "uncharge". Statistics itself is properly handled
2300 * Considering "move", this is an only case we see a race. To make the race
2301 * small, we check mm->moving_account and detect there are possibility of race
2302 * If there is, we take a lock.
2305 void __mem_cgroup_begin_update_page_stat(struct page *page,
2306 bool *locked, unsigned long *flags)
2308 struct mem_cgroup *memcg;
2309 struct page_cgroup *pc;
2311 pc = lookup_page_cgroup(page);
2313 memcg = pc->mem_cgroup;
2314 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2317 * If this memory cgroup is not under account moving, we don't
2318 * need to take move_lock_mem_cgroup(). Because we already hold
2319 * rcu_read_lock(), any calls to move_account will be delayed until
2320 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2322 if (!mem_cgroup_stolen(memcg))
2325 move_lock_mem_cgroup(memcg, flags);
2326 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2327 move_unlock_mem_cgroup(memcg, flags);
2333 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
2335 struct page_cgroup *pc = lookup_page_cgroup(page);
2338 * It's guaranteed that pc->mem_cgroup never changes while
2339 * lock is held because a routine modifies pc->mem_cgroup
2340 * should take move_lock_mem_cgroup().
2342 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
2345 void mem_cgroup_update_page_stat(struct page *page,
2346 enum mem_cgroup_stat_index idx, int val)
2348 struct mem_cgroup *memcg;
2349 struct page_cgroup *pc = lookup_page_cgroup(page);
2350 unsigned long uninitialized_var(flags);
2352 if (mem_cgroup_disabled())
2355 VM_BUG_ON(!rcu_read_lock_held());
2356 memcg = pc->mem_cgroup;
2357 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2360 this_cpu_add(memcg->stat->count[idx], val);
2364 * size of first charge trial. "32" comes from vmscan.c's magic value.
2365 * TODO: maybe necessary to use big numbers in big irons.
2367 #define CHARGE_BATCH 32U
2368 struct memcg_stock_pcp {
2369 struct mem_cgroup *cached; /* this never be root cgroup */
2370 unsigned int nr_pages;
2371 struct work_struct work;
2372 unsigned long flags;
2373 #define FLUSHING_CACHED_CHARGE 0
2375 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2376 static DEFINE_MUTEX(percpu_charge_mutex);
2379 * consume_stock: Try to consume stocked charge on this cpu.
2380 * @memcg: memcg to consume from.
2381 * @nr_pages: how many pages to charge.
2383 * The charges will only happen if @memcg matches the current cpu's memcg
2384 * stock, and at least @nr_pages are available in that stock. Failure to
2385 * service an allocation will refill the stock.
2387 * returns true if successful, false otherwise.
2389 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2391 struct memcg_stock_pcp *stock;
2394 if (nr_pages > CHARGE_BATCH)
2397 stock = &get_cpu_var(memcg_stock);
2398 if (memcg == stock->cached && stock->nr_pages >= nr_pages)
2399 stock->nr_pages -= nr_pages;
2400 else /* need to call res_counter_charge */
2402 put_cpu_var(memcg_stock);
2407 * Returns stocks cached in percpu to res_counter and reset cached information.
2409 static void drain_stock(struct memcg_stock_pcp *stock)
2411 struct mem_cgroup *old = stock->cached;
2413 if (stock->nr_pages) {
2414 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2416 res_counter_uncharge(&old->res, bytes);
2417 if (do_swap_account)
2418 res_counter_uncharge(&old->memsw, bytes);
2419 stock->nr_pages = 0;
2421 stock->cached = NULL;
2425 * This must be called under preempt disabled or must be called by
2426 * a thread which is pinned to local cpu.
2428 static void drain_local_stock(struct work_struct *dummy)
2430 struct memcg_stock_pcp *stock = this_cpu_ptr(&memcg_stock);
2432 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2435 static void __init memcg_stock_init(void)
2439 for_each_possible_cpu(cpu) {
2440 struct memcg_stock_pcp *stock =
2441 &per_cpu(memcg_stock, cpu);
2442 INIT_WORK(&stock->work, drain_local_stock);
2447 * Cache charges(val) which is from res_counter, to local per_cpu area.
2448 * This will be consumed by consume_stock() function, later.
2450 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2452 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2454 if (stock->cached != memcg) { /* reset if necessary */
2456 stock->cached = memcg;
2458 stock->nr_pages += nr_pages;
2459 put_cpu_var(memcg_stock);
2463 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2464 * of the hierarchy under it. sync flag says whether we should block
2465 * until the work is done.
2467 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2471 /* Notify other cpus that system-wide "drain" is running */
2474 for_each_online_cpu(cpu) {
2475 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2476 struct mem_cgroup *memcg;
2478 memcg = stock->cached;
2479 if (!memcg || !stock->nr_pages)
2481 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2483 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2485 drain_local_stock(&stock->work);
2487 schedule_work_on(cpu, &stock->work);
2495 for_each_online_cpu(cpu) {
2496 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2497 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2498 flush_work(&stock->work);
2505 * Tries to drain stocked charges in other cpus. This function is asynchronous
2506 * and just put a work per cpu for draining localy on each cpu. Caller can
2507 * expects some charges will be back to res_counter later but cannot wait for
2510 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2513 * If someone calls draining, avoid adding more kworker runs.
2515 if (!mutex_trylock(&percpu_charge_mutex))
2517 drain_all_stock(root_memcg, false);
2518 mutex_unlock(&percpu_charge_mutex);
2521 /* This is a synchronous drain interface. */
2522 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2524 /* called when force_empty is called */
2525 mutex_lock(&percpu_charge_mutex);
2526 drain_all_stock(root_memcg, true);
2527 mutex_unlock(&percpu_charge_mutex);
2531 * This function drains percpu counter value from DEAD cpu and
2532 * move it to local cpu. Note that this function can be preempted.
2534 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2538 spin_lock(&memcg->pcp_counter_lock);
2539 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2540 long x = per_cpu(memcg->stat->count[i], cpu);
2542 per_cpu(memcg->stat->count[i], cpu) = 0;
2543 memcg->nocpu_base.count[i] += x;
2545 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2546 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2548 per_cpu(memcg->stat->events[i], cpu) = 0;
2549 memcg->nocpu_base.events[i] += x;
2551 spin_unlock(&memcg->pcp_counter_lock);
2554 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
2555 unsigned long action,
2558 int cpu = (unsigned long)hcpu;
2559 struct memcg_stock_pcp *stock;
2560 struct mem_cgroup *iter;
2562 if (action == CPU_ONLINE)
2565 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2568 for_each_mem_cgroup(iter)
2569 mem_cgroup_drain_pcp_counter(iter, cpu);
2571 stock = &per_cpu(memcg_stock, cpu);
2577 /* See mem_cgroup_try_charge() for details */
2579 CHARGE_OK, /* success */
2580 CHARGE_RETRY, /* need to retry but retry is not bad */
2581 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2582 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2585 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2586 unsigned int nr_pages, unsigned int min_pages,
2589 unsigned long csize = nr_pages * PAGE_SIZE;
2590 struct mem_cgroup *mem_over_limit;
2591 struct res_counter *fail_res;
2592 unsigned long flags = 0;
2595 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2598 if (!do_swap_account)
2600 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2604 res_counter_uncharge(&memcg->res, csize);
2605 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2606 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2608 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2610 * Never reclaim on behalf of optional batching, retry with a
2611 * single page instead.
2613 if (nr_pages > min_pages)
2614 return CHARGE_RETRY;
2616 if (!(gfp_mask & __GFP_WAIT))
2617 return CHARGE_WOULDBLOCK;
2619 if (gfp_mask & __GFP_NORETRY)
2620 return CHARGE_NOMEM;
2622 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2623 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2624 return CHARGE_RETRY;
2626 * Even though the limit is exceeded at this point, reclaim
2627 * may have been able to free some pages. Retry the charge
2628 * before killing the task.
2630 * Only for regular pages, though: huge pages are rather
2631 * unlikely to succeed so close to the limit, and we fall back
2632 * to regular pages anyway in case of failure.
2634 if (nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER) && ret)
2635 return CHARGE_RETRY;
2638 * At task move, charge accounts can be doubly counted. So, it's
2639 * better to wait until the end of task_move if something is going on.
2641 if (mem_cgroup_wait_acct_move(mem_over_limit))
2642 return CHARGE_RETRY;
2645 mem_cgroup_oom(mem_over_limit, gfp_mask, get_order(csize));
2647 return CHARGE_NOMEM;
2651 * mem_cgroup_try_charge - try charging a memcg
2652 * @memcg: memcg to charge
2653 * @nr_pages: number of pages to charge
2654 * @oom: trigger OOM if reclaim fails
2656 * Returns 0 if @memcg was charged successfully, -EINTR if the charge
2657 * was bypassed to root_mem_cgroup, and -ENOMEM if the charge failed.
2659 static int mem_cgroup_try_charge(struct mem_cgroup *memcg,
2661 unsigned int nr_pages,
2664 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2665 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2668 if (mem_cgroup_is_root(memcg))
2671 * Unlike in global OOM situations, memcg is not in a physical
2672 * memory shortage. Allow dying and OOM-killed tasks to
2673 * bypass the last charges so that they can exit quickly and
2674 * free their memory.
2676 if (unlikely(test_thread_flag(TIF_MEMDIE) ||
2677 fatal_signal_pending(current) ||
2678 current->flags & PF_EXITING))
2681 if (unlikely(task_in_memcg_oom(current)))
2684 if (gfp_mask & __GFP_NOFAIL)
2687 if (consume_stock(memcg, nr_pages))
2691 bool invoke_oom = oom && !nr_oom_retries;
2693 /* If killed, bypass charge */
2694 if (fatal_signal_pending(current))
2697 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch,
2698 nr_pages, invoke_oom);
2702 case CHARGE_RETRY: /* not in OOM situation but retry */
2705 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2707 case CHARGE_NOMEM: /* OOM routine works */
2708 if (!oom || invoke_oom)
2713 } while (ret != CHARGE_OK);
2715 if (batch > nr_pages)
2716 refill_stock(memcg, batch - nr_pages);
2720 if (!(gfp_mask & __GFP_NOFAIL))
2727 * mem_cgroup_try_charge_mm - try charging a mm
2728 * @mm: mm_struct to charge
2729 * @nr_pages: number of pages to charge
2730 * @oom: trigger OOM if reclaim fails
2732 * Returns the charged mem_cgroup associated with the given mm_struct or
2733 * NULL the charge failed.
2735 static struct mem_cgroup *mem_cgroup_try_charge_mm(struct mm_struct *mm,
2737 unsigned int nr_pages,
2741 struct mem_cgroup *memcg;
2744 memcg = get_mem_cgroup_from_mm(mm);
2745 ret = mem_cgroup_try_charge(memcg, gfp_mask, nr_pages, oom);
2746 css_put(&memcg->css);
2748 memcg = root_mem_cgroup;
2756 * Somemtimes we have to undo a charge we got by try_charge().
2757 * This function is for that and do uncharge, put css's refcnt.
2758 * gotten by try_charge().
2760 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2761 unsigned int nr_pages)
2763 if (!mem_cgroup_is_root(memcg)) {
2764 unsigned long bytes = nr_pages * PAGE_SIZE;
2766 res_counter_uncharge(&memcg->res, bytes);
2767 if (do_swap_account)
2768 res_counter_uncharge(&memcg->memsw, bytes);
2773 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2774 * This is useful when moving usage to parent cgroup.
2776 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2777 unsigned int nr_pages)
2779 unsigned long bytes = nr_pages * PAGE_SIZE;
2781 if (mem_cgroup_is_root(memcg))
2784 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2785 if (do_swap_account)
2786 res_counter_uncharge_until(&memcg->memsw,
2787 memcg->memsw.parent, bytes);
2791 * A helper function to get mem_cgroup from ID. must be called under
2792 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2793 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2794 * called against removed memcg.)
2796 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2798 /* ID 0 is unused ID */
2801 return mem_cgroup_from_id(id);
2804 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2806 struct mem_cgroup *memcg = NULL;
2807 struct page_cgroup *pc;
2811 VM_BUG_ON_PAGE(!PageLocked(page), page);
2813 pc = lookup_page_cgroup(page);
2814 lock_page_cgroup(pc);
2815 if (PageCgroupUsed(pc)) {
2816 memcg = pc->mem_cgroup;
2817 if (memcg && !css_tryget(&memcg->css))
2819 } else if (PageSwapCache(page)) {
2820 ent.val = page_private(page);
2821 id = lookup_swap_cgroup_id(ent);
2823 memcg = mem_cgroup_lookup(id);
2824 if (memcg && !css_tryget(&memcg->css))
2828 unlock_page_cgroup(pc);
2832 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2834 unsigned int nr_pages,
2835 enum charge_type ctype,
2838 struct page_cgroup *pc = lookup_page_cgroup(page);
2839 struct zone *uninitialized_var(zone);
2840 struct lruvec *lruvec;
2841 bool was_on_lru = false;
2844 lock_page_cgroup(pc);
2845 VM_BUG_ON_PAGE(PageCgroupUsed(pc), page);
2847 * we don't need page_cgroup_lock about tail pages, becase they are not
2848 * accessed by any other context at this point.
2852 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2853 * may already be on some other mem_cgroup's LRU. Take care of it.
2856 zone = page_zone(page);
2857 spin_lock_irq(&zone->lru_lock);
2858 if (PageLRU(page)) {
2859 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2861 del_page_from_lru_list(page, lruvec, page_lru(page));
2866 pc->mem_cgroup = memcg;
2868 * We access a page_cgroup asynchronously without lock_page_cgroup().
2869 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2870 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2871 * before USED bit, we need memory barrier here.
2872 * See mem_cgroup_add_lru_list(), etc.
2875 SetPageCgroupUsed(pc);
2879 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2880 VM_BUG_ON_PAGE(PageLRU(page), page);
2882 add_page_to_lru_list(page, lruvec, page_lru(page));
2884 spin_unlock_irq(&zone->lru_lock);
2887 if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
2892 mem_cgroup_charge_statistics(memcg, page, anon, nr_pages);
2893 unlock_page_cgroup(pc);
2896 * "charge_statistics" updated event counter. Then, check it.
2897 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2898 * if they exceeds softlimit.
2900 memcg_check_events(memcg, page);
2903 static DEFINE_MUTEX(set_limit_mutex);
2905 #ifdef CONFIG_MEMCG_KMEM
2907 * The memcg_slab_mutex is held whenever a per memcg kmem cache is created or
2908 * destroyed. It protects memcg_caches arrays and memcg_slab_caches lists.
2910 static DEFINE_MUTEX(memcg_slab_mutex);
2912 static DEFINE_MUTEX(activate_kmem_mutex);
2914 static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg)
2916 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) &&
2917 memcg_kmem_is_active(memcg);
2921 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2922 * in the memcg_cache_params struct.
2924 static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
2926 struct kmem_cache *cachep;
2928 VM_BUG_ON(p->is_root_cache);
2929 cachep = p->root_cache;
2930 return cache_from_memcg_idx(cachep, memcg_cache_id(p->memcg));
2933 #ifdef CONFIG_SLABINFO
2934 static int mem_cgroup_slabinfo_read(struct seq_file *m, void *v)
2936 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
2937 struct memcg_cache_params *params;
2939 if (!memcg_can_account_kmem(memcg))
2942 print_slabinfo_header(m);
2944 mutex_lock(&memcg_slab_mutex);
2945 list_for_each_entry(params, &memcg->memcg_slab_caches, list)
2946 cache_show(memcg_params_to_cache(params), m);
2947 mutex_unlock(&memcg_slab_mutex);
2953 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size)
2955 struct res_counter *fail_res;
2958 ret = res_counter_charge(&memcg->kmem, size, &fail_res);
2962 ret = mem_cgroup_try_charge(memcg, gfp, size >> PAGE_SHIFT,
2963 oom_gfp_allowed(gfp));
2964 if (ret == -EINTR) {
2966 * mem_cgroup_try_charge() chosed to bypass to root due to
2967 * OOM kill or fatal signal. Since our only options are to
2968 * either fail the allocation or charge it to this cgroup, do
2969 * it as a temporary condition. But we can't fail. From a
2970 * kmem/slab perspective, the cache has already been selected,
2971 * by mem_cgroup_kmem_get_cache(), so it is too late to change
2974 * This condition will only trigger if the task entered
2975 * memcg_charge_kmem in a sane state, but was OOM-killed during
2976 * mem_cgroup_try_charge() above. Tasks that were already
2977 * dying when the allocation triggers should have been already
2978 * directed to the root cgroup in memcontrol.h
2980 res_counter_charge_nofail(&memcg->res, size, &fail_res);
2981 if (do_swap_account)
2982 res_counter_charge_nofail(&memcg->memsw, size,
2986 res_counter_uncharge(&memcg->kmem, size);
2991 static void memcg_uncharge_kmem(struct mem_cgroup *memcg, u64 size)
2993 res_counter_uncharge(&memcg->res, size);
2994 if (do_swap_account)
2995 res_counter_uncharge(&memcg->memsw, size);
2998 if (res_counter_uncharge(&memcg->kmem, size))
3002 * Releases a reference taken in kmem_cgroup_css_offline in case
3003 * this last uncharge is racing with the offlining code or it is
3004 * outliving the memcg existence.
3006 * The memory barrier imposed by test&clear is paired with the
3007 * explicit one in memcg_kmem_mark_dead().
3009 if (memcg_kmem_test_and_clear_dead(memcg))
3010 css_put(&memcg->css);
3014 * helper for acessing a memcg's index. It will be used as an index in the
3015 * child cache array in kmem_cache, and also to derive its name. This function
3016 * will return -1 when this is not a kmem-limited memcg.
3018 int memcg_cache_id(struct mem_cgroup *memcg)
3020 return memcg ? memcg->kmemcg_id : -1;
3023 static size_t memcg_caches_array_size(int num_groups)
3026 if (num_groups <= 0)
3029 size = 2 * num_groups;
3030 if (size < MEMCG_CACHES_MIN_SIZE)
3031 size = MEMCG_CACHES_MIN_SIZE;
3032 else if (size > MEMCG_CACHES_MAX_SIZE)
3033 size = MEMCG_CACHES_MAX_SIZE;
3039 * We should update the current array size iff all caches updates succeed. This
3040 * can only be done from the slab side. The slab mutex needs to be held when
3043 void memcg_update_array_size(int num)
3045 if (num > memcg_limited_groups_array_size)
3046 memcg_limited_groups_array_size = memcg_caches_array_size(num);
3049 int memcg_update_cache_size(struct kmem_cache *s, int num_groups)
3051 struct memcg_cache_params *cur_params = s->memcg_params;
3053 VM_BUG_ON(!is_root_cache(s));
3055 if (num_groups > memcg_limited_groups_array_size) {
3057 struct memcg_cache_params *new_params;
3058 ssize_t size = memcg_caches_array_size(num_groups);
3060 size *= sizeof(void *);
3061 size += offsetof(struct memcg_cache_params, memcg_caches);
3063 new_params = kzalloc(size, GFP_KERNEL);
3067 new_params->is_root_cache = true;
3070 * There is the chance it will be bigger than
3071 * memcg_limited_groups_array_size, if we failed an allocation
3072 * in a cache, in which case all caches updated before it, will
3073 * have a bigger array.
3075 * But if that is the case, the data after
3076 * memcg_limited_groups_array_size is certainly unused
3078 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3079 if (!cur_params->memcg_caches[i])
3081 new_params->memcg_caches[i] =
3082 cur_params->memcg_caches[i];
3086 * Ideally, we would wait until all caches succeed, and only
3087 * then free the old one. But this is not worth the extra
3088 * pointer per-cache we'd have to have for this.
3090 * It is not a big deal if some caches are left with a size
3091 * bigger than the others. And all updates will reset this
3094 rcu_assign_pointer(s->memcg_params, new_params);
3096 kfree_rcu(cur_params, rcu_head);
3101 char *memcg_create_cache_name(struct mem_cgroup *memcg,
3102 struct kmem_cache *root_cache)
3107 * We need a mutex here to protect the shared buffer. Since this is
3108 * expected to be called only on cache creation, we can employ the
3109 * slab_mutex for that purpose.
3111 lockdep_assert_held(&slab_mutex);
3114 buf = kmalloc(NAME_MAX + 1, GFP_KERNEL);
3119 cgroup_name(memcg->css.cgroup, buf, NAME_MAX + 1);
3120 return kasprintf(GFP_KERNEL, "%s(%d:%s)", root_cache->name,
3121 memcg_cache_id(memcg), buf);
3124 int memcg_alloc_cache_params(struct mem_cgroup *memcg, struct kmem_cache *s,
3125 struct kmem_cache *root_cache)
3129 if (!memcg_kmem_enabled())
3133 size = offsetof(struct memcg_cache_params, memcg_caches);
3134 size += memcg_limited_groups_array_size * sizeof(void *);
3136 size = sizeof(struct memcg_cache_params);
3138 s->memcg_params = kzalloc(size, GFP_KERNEL);
3139 if (!s->memcg_params)
3143 s->memcg_params->memcg = memcg;
3144 s->memcg_params->root_cache = root_cache;
3145 css_get(&memcg->css);
3147 s->memcg_params->is_root_cache = true;
3152 void memcg_free_cache_params(struct kmem_cache *s)
3154 if (!s->memcg_params)
3156 if (!s->memcg_params->is_root_cache)
3157 css_put(&s->memcg_params->memcg->css);
3158 kfree(s->memcg_params);
3161 static void memcg_kmem_create_cache(struct mem_cgroup *memcg,
3162 struct kmem_cache *root_cache)
3164 struct kmem_cache *cachep;
3167 lockdep_assert_held(&memcg_slab_mutex);
3169 id = memcg_cache_id(memcg);
3172 * Since per-memcg caches are created asynchronously on first
3173 * allocation (see memcg_kmem_get_cache()), several threads can try to
3174 * create the same cache, but only one of them may succeed.
3176 if (cache_from_memcg_idx(root_cache, id))
3179 cachep = kmem_cache_create_memcg(memcg, root_cache);
3181 * If we could not create a memcg cache, do not complain, because
3182 * that's not critical at all as we can always proceed with the root
3188 list_add(&cachep->memcg_params->list, &memcg->memcg_slab_caches);
3191 * Since readers won't lock (see cache_from_memcg_idx()), we need a
3192 * barrier here to ensure nobody will see the kmem_cache partially
3197 BUG_ON(root_cache->memcg_params->memcg_caches[id]);
3198 root_cache->memcg_params->memcg_caches[id] = cachep;
3201 static void memcg_kmem_destroy_cache(struct kmem_cache *cachep)
3203 struct kmem_cache *root_cache;
3204 struct mem_cgroup *memcg;
3207 lockdep_assert_held(&memcg_slab_mutex);
3209 BUG_ON(is_root_cache(cachep));
3211 root_cache = cachep->memcg_params->root_cache;
3212 memcg = cachep->memcg_params->memcg;
3213 id = memcg_cache_id(memcg);
3215 BUG_ON(root_cache->memcg_params->memcg_caches[id] != cachep);
3216 root_cache->memcg_params->memcg_caches[id] = NULL;
3218 list_del(&cachep->memcg_params->list);
3220 kmem_cache_destroy(cachep);
3224 * During the creation a new cache, we need to disable our accounting mechanism
3225 * altogether. This is true even if we are not creating, but rather just
3226 * enqueing new caches to be created.
3228 * This is because that process will trigger allocations; some visible, like
3229 * explicit kmallocs to auxiliary data structures, name strings and internal
3230 * cache structures; some well concealed, like INIT_WORK() that can allocate
3231 * objects during debug.
3233 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3234 * to it. This may not be a bounded recursion: since the first cache creation
3235 * failed to complete (waiting on the allocation), we'll just try to create the
3236 * cache again, failing at the same point.
3238 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3239 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3240 * inside the following two functions.
3242 static inline void memcg_stop_kmem_account(void)
3244 VM_BUG_ON(!current->mm);
3245 current->memcg_kmem_skip_account++;
3248 static inline void memcg_resume_kmem_account(void)
3250 VM_BUG_ON(!current->mm);
3251 current->memcg_kmem_skip_account--;
3254 int __kmem_cache_destroy_memcg_children(struct kmem_cache *s)
3256 struct kmem_cache *c;
3259 mutex_lock(&memcg_slab_mutex);
3260 for_each_memcg_cache_index(i) {
3261 c = cache_from_memcg_idx(s, i);
3265 memcg_kmem_destroy_cache(c);
3267 if (cache_from_memcg_idx(s, i))
3270 mutex_unlock(&memcg_slab_mutex);
3274 static void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3276 struct kmem_cache *cachep;
3277 struct memcg_cache_params *params, *tmp;
3279 if (!memcg_kmem_is_active(memcg))
3282 mutex_lock(&memcg_slab_mutex);
3283 list_for_each_entry_safe(params, tmp, &memcg->memcg_slab_caches, list) {
3284 cachep = memcg_params_to_cache(params);
3285 kmem_cache_shrink(cachep);
3286 if (atomic_read(&cachep->memcg_params->nr_pages) == 0)
3287 memcg_kmem_destroy_cache(cachep);
3289 mutex_unlock(&memcg_slab_mutex);
3292 struct create_work {
3293 struct mem_cgroup *memcg;
3294 struct kmem_cache *cachep;
3295 struct work_struct work;
3298 static void memcg_create_cache_work_func(struct work_struct *w)
3300 struct create_work *cw = container_of(w, struct create_work, work);
3301 struct mem_cgroup *memcg = cw->memcg;
3302 struct kmem_cache *cachep = cw->cachep;
3304 mutex_lock(&memcg_slab_mutex);
3305 memcg_kmem_create_cache(memcg, cachep);
3306 mutex_unlock(&memcg_slab_mutex);
3308 css_put(&memcg->css);
3313 * Enqueue the creation of a per-memcg kmem_cache.
3315 static void __memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3316 struct kmem_cache *cachep)
3318 struct create_work *cw;
3320 cw = kmalloc(sizeof(struct create_work), GFP_NOWAIT);
3322 css_put(&memcg->css);
3327 cw->cachep = cachep;
3329 INIT_WORK(&cw->work, memcg_create_cache_work_func);
3330 schedule_work(&cw->work);
3333 static void memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3334 struct kmem_cache *cachep)
3337 * We need to stop accounting when we kmalloc, because if the
3338 * corresponding kmalloc cache is not yet created, the first allocation
3339 * in __memcg_create_cache_enqueue will recurse.
3341 * However, it is better to enclose the whole function. Depending on
3342 * the debugging options enabled, INIT_WORK(), for instance, can
3343 * trigger an allocation. This too, will make us recurse. Because at
3344 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3345 * the safest choice is to do it like this, wrapping the whole function.
3347 memcg_stop_kmem_account();
3348 __memcg_create_cache_enqueue(memcg, cachep);
3349 memcg_resume_kmem_account();
3352 int __memcg_charge_slab(struct kmem_cache *cachep, gfp_t gfp, int order)
3356 res = memcg_charge_kmem(cachep->memcg_params->memcg, gfp,
3357 PAGE_SIZE << order);
3359 atomic_add(1 << order, &cachep->memcg_params->nr_pages);
3363 void __memcg_uncharge_slab(struct kmem_cache *cachep, int order)
3365 memcg_uncharge_kmem(cachep->memcg_params->memcg, PAGE_SIZE << order);
3366 atomic_sub(1 << order, &cachep->memcg_params->nr_pages);
3370 * Return the kmem_cache we're supposed to use for a slab allocation.
3371 * We try to use the current memcg's version of the cache.
3373 * If the cache does not exist yet, if we are the first user of it,
3374 * we either create it immediately, if possible, or create it asynchronously
3376 * In the latter case, we will let the current allocation go through with
3377 * the original cache.
3379 * Can't be called in interrupt context or from kernel threads.
3380 * This function needs to be called with rcu_read_lock() held.
3382 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
3385 struct mem_cgroup *memcg;
3386 struct kmem_cache *memcg_cachep;
3388 VM_BUG_ON(!cachep->memcg_params);
3389 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
3391 if (!current->mm || current->memcg_kmem_skip_account)
3395 memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
3397 if (!memcg_can_account_kmem(memcg))
3400 memcg_cachep = cache_from_memcg_idx(cachep, memcg_cache_id(memcg));
3401 if (likely(memcg_cachep)) {
3402 cachep = memcg_cachep;
3406 /* The corresponding put will be done in the workqueue. */
3407 if (!css_tryget(&memcg->css))
3412 * If we are in a safe context (can wait, and not in interrupt
3413 * context), we could be be predictable and return right away.
3414 * This would guarantee that the allocation being performed
3415 * already belongs in the new cache.
3417 * However, there are some clashes that can arrive from locking.
3418 * For instance, because we acquire the slab_mutex while doing
3419 * kmem_cache_dup, this means no further allocation could happen
3420 * with the slab_mutex held.
3422 * Also, because cache creation issue get_online_cpus(), this
3423 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3424 * that ends up reversed during cpu hotplug. (cpuset allocates
3425 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3426 * better to defer everything.
3428 memcg_create_cache_enqueue(memcg, cachep);
3436 * We need to verify if the allocation against current->mm->owner's memcg is
3437 * possible for the given order. But the page is not allocated yet, so we'll
3438 * need a further commit step to do the final arrangements.
3440 * It is possible for the task to switch cgroups in this mean time, so at
3441 * commit time, we can't rely on task conversion any longer. We'll then use
3442 * the handle argument to return to the caller which cgroup we should commit
3443 * against. We could also return the memcg directly and avoid the pointer
3444 * passing, but a boolean return value gives better semantics considering
3445 * the compiled-out case as well.
3447 * Returning true means the allocation is possible.
3450 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3452 struct mem_cgroup *memcg;
3458 * Disabling accounting is only relevant for some specific memcg
3459 * internal allocations. Therefore we would initially not have such
3460 * check here, since direct calls to the page allocator that are
3461 * accounted to kmemcg (alloc_kmem_pages and friends) only happen
3462 * outside memcg core. We are mostly concerned with cache allocations,
3463 * and by having this test at memcg_kmem_get_cache, we are already able
3464 * to relay the allocation to the root cache and bypass the memcg cache
3467 * There is one exception, though: the SLUB allocator does not create
3468 * large order caches, but rather service large kmallocs directly from
3469 * the page allocator. Therefore, the following sequence when backed by
3470 * the SLUB allocator:
3472 * memcg_stop_kmem_account();
3473 * kmalloc(<large_number>)
3474 * memcg_resume_kmem_account();
3476 * would effectively ignore the fact that we should skip accounting,
3477 * since it will drive us directly to this function without passing
3478 * through the cache selector memcg_kmem_get_cache. Such large
3479 * allocations are extremely rare but can happen, for instance, for the
3480 * cache arrays. We bring this test here.
3482 if (!current->mm || current->memcg_kmem_skip_account)
3485 memcg = get_mem_cgroup_from_mm(current->mm);
3487 if (!memcg_can_account_kmem(memcg)) {
3488 css_put(&memcg->css);
3492 ret = memcg_charge_kmem(memcg, gfp, PAGE_SIZE << order);
3496 css_put(&memcg->css);
3500 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3503 struct page_cgroup *pc;
3505 VM_BUG_ON(mem_cgroup_is_root(memcg));
3507 /* The page allocation failed. Revert */
3509 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3513 pc = lookup_page_cgroup(page);
3514 lock_page_cgroup(pc);
3515 pc->mem_cgroup = memcg;
3516 SetPageCgroupUsed(pc);
3517 unlock_page_cgroup(pc);
3520 void __memcg_kmem_uncharge_pages(struct page *page, int order)
3522 struct mem_cgroup *memcg = NULL;
3523 struct page_cgroup *pc;
3526 pc = lookup_page_cgroup(page);
3528 * Fast unlocked return. Theoretically might have changed, have to
3529 * check again after locking.
3531 if (!PageCgroupUsed(pc))
3534 lock_page_cgroup(pc);
3535 if (PageCgroupUsed(pc)) {
3536 memcg = pc->mem_cgroup;
3537 ClearPageCgroupUsed(pc);
3539 unlock_page_cgroup(pc);
3542 * We trust that only if there is a memcg associated with the page, it
3543 * is a valid allocation
3548 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
3549 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3552 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3555 #endif /* CONFIG_MEMCG_KMEM */
3557 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3559 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3561 * Because tail pages are not marked as "used", set it. We're under
3562 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3563 * charge/uncharge will be never happen and move_account() is done under
3564 * compound_lock(), so we don't have to take care of races.
3566 void mem_cgroup_split_huge_fixup(struct page *head)
3568 struct page_cgroup *head_pc = lookup_page_cgroup(head);
3569 struct page_cgroup *pc;
3570 struct mem_cgroup *memcg;
3573 if (mem_cgroup_disabled())
3576 memcg = head_pc->mem_cgroup;
3577 for (i = 1; i < HPAGE_PMD_NR; i++) {
3579 pc->mem_cgroup = memcg;
3580 smp_wmb();/* see __commit_charge() */
3581 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
3583 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
3586 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3589 * mem_cgroup_move_account - move account of the page
3591 * @nr_pages: number of regular pages (>1 for huge pages)
3592 * @pc: page_cgroup of the page.
3593 * @from: mem_cgroup which the page is moved from.
3594 * @to: mem_cgroup which the page is moved to. @from != @to.
3596 * The caller must confirm following.
3597 * - page is not on LRU (isolate_page() is useful.)
3598 * - compound_lock is held when nr_pages > 1
3600 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3603 static int mem_cgroup_move_account(struct page *page,
3604 unsigned int nr_pages,
3605 struct page_cgroup *pc,
3606 struct mem_cgroup *from,
3607 struct mem_cgroup *to)
3609 unsigned long flags;
3611 bool anon = PageAnon(page);
3613 VM_BUG_ON(from == to);
3614 VM_BUG_ON_PAGE(PageLRU(page), page);
3616 * The page is isolated from LRU. So, collapse function
3617 * will not handle this page. But page splitting can happen.
3618 * Do this check under compound_page_lock(). The caller should
3622 if (nr_pages > 1 && !PageTransHuge(page))
3625 lock_page_cgroup(pc);
3628 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3631 move_lock_mem_cgroup(from, &flags);
3633 if (!anon && page_mapped(page)) {
3634 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
3636 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
3640 if (PageWriteback(page)) {
3641 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK],
3643 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK],
3647 mem_cgroup_charge_statistics(from, page, anon, -nr_pages);
3649 /* caller should have done css_get */
3650 pc->mem_cgroup = to;
3651 mem_cgroup_charge_statistics(to, page, anon, nr_pages);
3652 move_unlock_mem_cgroup(from, &flags);
3655 unlock_page_cgroup(pc);
3659 memcg_check_events(to, page);
3660 memcg_check_events(from, page);
3666 * mem_cgroup_move_parent - moves page to the parent group
3667 * @page: the page to move
3668 * @pc: page_cgroup of the page
3669 * @child: page's cgroup
3671 * move charges to its parent or the root cgroup if the group has no
3672 * parent (aka use_hierarchy==0).
3673 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3674 * mem_cgroup_move_account fails) the failure is always temporary and
3675 * it signals a race with a page removal/uncharge or migration. In the
3676 * first case the page is on the way out and it will vanish from the LRU
3677 * on the next attempt and the call should be retried later.
3678 * Isolation from the LRU fails only if page has been isolated from
3679 * the LRU since we looked at it and that usually means either global
3680 * reclaim or migration going on. The page will either get back to the
3682 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3683 * (!PageCgroupUsed) or moved to a different group. The page will
3684 * disappear in the next attempt.
3686 static int mem_cgroup_move_parent(struct page *page,
3687 struct page_cgroup *pc,
3688 struct mem_cgroup *child)
3690 struct mem_cgroup *parent;
3691 unsigned int nr_pages;
3692 unsigned long uninitialized_var(flags);
3695 VM_BUG_ON(mem_cgroup_is_root(child));
3698 if (!get_page_unless_zero(page))
3700 if (isolate_lru_page(page))
3703 nr_pages = hpage_nr_pages(page);
3705 parent = parent_mem_cgroup(child);
3707 * If no parent, move charges to root cgroup.
3710 parent = root_mem_cgroup;
3713 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
3714 flags = compound_lock_irqsave(page);
3717 ret = mem_cgroup_move_account(page, nr_pages,
3720 __mem_cgroup_cancel_local_charge(child, nr_pages);
3723 compound_unlock_irqrestore(page, flags);
3724 putback_lru_page(page);
3731 int mem_cgroup_charge_anon(struct page *page,
3732 struct mm_struct *mm, gfp_t gfp_mask)
3734 unsigned int nr_pages = 1;
3735 struct mem_cgroup *memcg;
3738 if (mem_cgroup_disabled())
3741 VM_BUG_ON_PAGE(page_mapped(page), page);
3742 VM_BUG_ON_PAGE(page->mapping && !PageAnon(page), page);
3745 if (PageTransHuge(page)) {
3746 nr_pages <<= compound_order(page);
3747 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
3749 * Never OOM-kill a process for a huge page. The
3750 * fault handler will fall back to regular pages.
3755 memcg = mem_cgroup_try_charge_mm(mm, gfp_mask, nr_pages, oom);
3758 __mem_cgroup_commit_charge(memcg, page, nr_pages,
3759 MEM_CGROUP_CHARGE_TYPE_ANON, false);
3764 * While swap-in, try_charge -> commit or cancel, the page is locked.
3765 * And when try_charge() successfully returns, one refcnt to memcg without
3766 * struct page_cgroup is acquired. This refcnt will be consumed by
3767 * "commit()" or removed by "cancel()"
3769 static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm,
3772 struct mem_cgroup **memcgp)
3774 struct mem_cgroup *memcg = NULL;
3775 struct page_cgroup *pc;
3778 pc = lookup_page_cgroup(page);
3780 * Every swap fault against a single page tries to charge the
3781 * page, bail as early as possible. shmem_unuse() encounters
3782 * already charged pages, too. The USED bit is protected by
3783 * the page lock, which serializes swap cache removal, which
3784 * in turn serializes uncharging.
3786 if (PageCgroupUsed(pc))
3788 if (do_swap_account)
3789 memcg = try_get_mem_cgroup_from_page(page);
3791 memcg = get_mem_cgroup_from_mm(mm);
3792 ret = mem_cgroup_try_charge(memcg, mask, 1, true);
3793 css_put(&memcg->css);
3795 memcg = root_mem_cgroup;
3803 int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page,
3804 gfp_t gfp_mask, struct mem_cgroup **memcgp)
3806 if (mem_cgroup_disabled()) {
3811 * A racing thread's fault, or swapoff, may have already
3812 * updated the pte, and even removed page from swap cache: in
3813 * those cases unuse_pte()'s pte_same() test will fail; but
3814 * there's also a KSM case which does need to charge the page.
3816 if (!PageSwapCache(page)) {
3817 struct mem_cgroup *memcg;
3819 memcg = mem_cgroup_try_charge_mm(mm, gfp_mask, 1, true);
3825 return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp);
3828 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
3830 if (mem_cgroup_disabled())
3834 __mem_cgroup_cancel_charge(memcg, 1);
3838 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
3839 enum charge_type ctype)
3841 if (mem_cgroup_disabled())
3846 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
3848 * Now swap is on-memory. This means this page may be
3849 * counted both as mem and swap....double count.
3850 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
3851 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
3852 * may call delete_from_swap_cache() before reach here.
3854 if (do_swap_account && PageSwapCache(page)) {
3855 swp_entry_t ent = {.val = page_private(page)};
3856 mem_cgroup_uncharge_swap(ent);
3860 void mem_cgroup_commit_charge_swapin(struct page *page,
3861 struct mem_cgroup *memcg)
3863 __mem_cgroup_commit_charge_swapin(page, memcg,
3864 MEM_CGROUP_CHARGE_TYPE_ANON);
3867 int mem_cgroup_charge_file(struct page *page, struct mm_struct *mm,
3870 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
3871 struct mem_cgroup *memcg;
3874 if (mem_cgroup_disabled())
3876 if (PageCompound(page))
3879 if (PageSwapCache(page)) { /* shmem */
3880 ret = __mem_cgroup_try_charge_swapin(mm, page,
3884 __mem_cgroup_commit_charge_swapin(page, memcg, type);
3889 * Page cache insertions can happen without an actual mm
3890 * context, e.g. during disk probing on boot.
3893 memcg = root_mem_cgroup;
3895 memcg = mem_cgroup_try_charge_mm(mm, gfp_mask, 1, true);
3899 __mem_cgroup_commit_charge(memcg, page, 1, type, false);
3903 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
3904 unsigned int nr_pages,
3905 const enum charge_type ctype)
3907 struct memcg_batch_info *batch = NULL;
3908 bool uncharge_memsw = true;
3910 /* If swapout, usage of swap doesn't decrease */
3911 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
3912 uncharge_memsw = false;
3914 batch = ¤t->memcg_batch;
3916 * In usual, we do css_get() when we remember memcg pointer.
3917 * But in this case, we keep res->usage until end of a series of
3918 * uncharges. Then, it's ok to ignore memcg's refcnt.
3921 batch->memcg = memcg;
3923 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
3924 * In those cases, all pages freed continuously can be expected to be in
3925 * the same cgroup and we have chance to coalesce uncharges.
3926 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
3927 * because we want to do uncharge as soon as possible.
3930 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
3931 goto direct_uncharge;
3934 goto direct_uncharge;
3937 * In typical case, batch->memcg == mem. This means we can
3938 * merge a series of uncharges to an uncharge of res_counter.
3939 * If not, we uncharge res_counter ony by one.
3941 if (batch->memcg != memcg)
3942 goto direct_uncharge;
3943 /* remember freed charge and uncharge it later */
3946 batch->memsw_nr_pages++;
3949 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
3951 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
3952 if (unlikely(batch->memcg != memcg))
3953 memcg_oom_recover(memcg);
3957 * uncharge if !page_mapped(page)
3959 static struct mem_cgroup *
3960 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
3963 struct mem_cgroup *memcg = NULL;
3964 unsigned int nr_pages = 1;
3965 struct page_cgroup *pc;
3968 if (mem_cgroup_disabled())
3971 if (PageTransHuge(page)) {
3972 nr_pages <<= compound_order(page);
3973 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
3976 * Check if our page_cgroup is valid
3978 pc = lookup_page_cgroup(page);
3979 if (unlikely(!PageCgroupUsed(pc)))
3982 lock_page_cgroup(pc);
3984 memcg = pc->mem_cgroup;
3986 if (!PageCgroupUsed(pc))
3989 anon = PageAnon(page);
3992 case MEM_CGROUP_CHARGE_TYPE_ANON:
3994 * Generally PageAnon tells if it's the anon statistics to be
3995 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
3996 * used before page reached the stage of being marked PageAnon.
4000 case MEM_CGROUP_CHARGE_TYPE_DROP:
4001 /* See mem_cgroup_prepare_migration() */
4002 if (page_mapped(page))
4005 * Pages under migration may not be uncharged. But
4006 * end_migration() /must/ be the one uncharging the
4007 * unused post-migration page and so it has to call
4008 * here with the migration bit still set. See the
4009 * res_counter handling below.
4011 if (!end_migration && PageCgroupMigration(pc))
4014 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
4015 if (!PageAnon(page)) { /* Shared memory */
4016 if (page->mapping && !page_is_file_cache(page))
4018 } else if (page_mapped(page)) /* Anon */
4025 mem_cgroup_charge_statistics(memcg, page, anon, -nr_pages);
4027 ClearPageCgroupUsed(pc);
4029 * pc->mem_cgroup is not cleared here. It will be accessed when it's
4030 * freed from LRU. This is safe because uncharged page is expected not
4031 * to be reused (freed soon). Exception is SwapCache, it's handled by
4032 * special functions.
4035 unlock_page_cgroup(pc);
4037 * even after unlock, we have memcg->res.usage here and this memcg
4038 * will never be freed, so it's safe to call css_get().
4040 memcg_check_events(memcg, page);
4041 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
4042 mem_cgroup_swap_statistics(memcg, true);
4043 css_get(&memcg->css);
4046 * Migration does not charge the res_counter for the
4047 * replacement page, so leave it alone when phasing out the
4048 * page that is unused after the migration.
4050 if (!end_migration && !mem_cgroup_is_root(memcg))
4051 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
4056 unlock_page_cgroup(pc);
4060 void mem_cgroup_uncharge_page(struct page *page)
4063 if (page_mapped(page))
4065 VM_BUG_ON_PAGE(page->mapping && !PageAnon(page), page);
4067 * If the page is in swap cache, uncharge should be deferred
4068 * to the swap path, which also properly accounts swap usage
4069 * and handles memcg lifetime.
4071 * Note that this check is not stable and reclaim may add the
4072 * page to swap cache at any time after this. However, if the
4073 * page is not in swap cache by the time page->mapcount hits
4074 * 0, there won't be any page table references to the swap
4075 * slot, and reclaim will free it and not actually write the
4078 if (PageSwapCache(page))
4080 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
4083 void mem_cgroup_uncharge_cache_page(struct page *page)
4085 VM_BUG_ON_PAGE(page_mapped(page), page);
4086 VM_BUG_ON_PAGE(page->mapping, page);
4087 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
4091 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4092 * In that cases, pages are freed continuously and we can expect pages
4093 * are in the same memcg. All these calls itself limits the number of
4094 * pages freed at once, then uncharge_start/end() is called properly.
4095 * This may be called prural(2) times in a context,
4098 void mem_cgroup_uncharge_start(void)
4100 current->memcg_batch.do_batch++;
4101 /* We can do nest. */
4102 if (current->memcg_batch.do_batch == 1) {
4103 current->memcg_batch.memcg = NULL;
4104 current->memcg_batch.nr_pages = 0;
4105 current->memcg_batch.memsw_nr_pages = 0;
4109 void mem_cgroup_uncharge_end(void)
4111 struct memcg_batch_info *batch = ¤t->memcg_batch;
4113 if (!batch->do_batch)
4117 if (batch->do_batch) /* If stacked, do nothing. */
4123 * This "batch->memcg" is valid without any css_get/put etc...
4124 * bacause we hide charges behind us.
4126 if (batch->nr_pages)
4127 res_counter_uncharge(&batch->memcg->res,
4128 batch->nr_pages * PAGE_SIZE);
4129 if (batch->memsw_nr_pages)
4130 res_counter_uncharge(&batch->memcg->memsw,
4131 batch->memsw_nr_pages * PAGE_SIZE);
4132 memcg_oom_recover(batch->memcg);
4133 /* forget this pointer (for sanity check) */
4134 batch->memcg = NULL;
4139 * called after __delete_from_swap_cache() and drop "page" account.
4140 * memcg information is recorded to swap_cgroup of "ent"
4143 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
4145 struct mem_cgroup *memcg;
4146 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
4148 if (!swapout) /* this was a swap cache but the swap is unused ! */
4149 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
4151 memcg = __mem_cgroup_uncharge_common(page, ctype, false);
4154 * record memcg information, if swapout && memcg != NULL,
4155 * css_get() was called in uncharge().
4157 if (do_swap_account && swapout && memcg)
4158 swap_cgroup_record(ent, mem_cgroup_id(memcg));
4162 #ifdef CONFIG_MEMCG_SWAP
4164 * called from swap_entry_free(). remove record in swap_cgroup and
4165 * uncharge "memsw" account.
4167 void mem_cgroup_uncharge_swap(swp_entry_t ent)
4169 struct mem_cgroup *memcg;
4172 if (!do_swap_account)
4175 id = swap_cgroup_record(ent, 0);
4177 memcg = mem_cgroup_lookup(id);
4180 * We uncharge this because swap is freed.
4181 * This memcg can be obsolete one. We avoid calling css_tryget
4183 if (!mem_cgroup_is_root(memcg))
4184 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
4185 mem_cgroup_swap_statistics(memcg, false);
4186 css_put(&memcg->css);
4192 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4193 * @entry: swap entry to be moved
4194 * @from: mem_cgroup which the entry is moved from
4195 * @to: mem_cgroup which the entry is moved to
4197 * It succeeds only when the swap_cgroup's record for this entry is the same
4198 * as the mem_cgroup's id of @from.
4200 * Returns 0 on success, -EINVAL on failure.
4202 * The caller must have charged to @to, IOW, called res_counter_charge() about
4203 * both res and memsw, and called css_get().
4205 static int mem_cgroup_move_swap_account(swp_entry_t entry,
4206 struct mem_cgroup *from, struct mem_cgroup *to)
4208 unsigned short old_id, new_id;
4210 old_id = mem_cgroup_id(from);
4211 new_id = mem_cgroup_id(to);
4213 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
4214 mem_cgroup_swap_statistics(from, false);
4215 mem_cgroup_swap_statistics(to, true);
4217 * This function is only called from task migration context now.
4218 * It postpones res_counter and refcount handling till the end
4219 * of task migration(mem_cgroup_clear_mc()) for performance
4220 * improvement. But we cannot postpone css_get(to) because if
4221 * the process that has been moved to @to does swap-in, the
4222 * refcount of @to might be decreased to 0.
4224 * We are in attach() phase, so the cgroup is guaranteed to be
4225 * alive, so we can just call css_get().
4233 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
4234 struct mem_cgroup *from, struct mem_cgroup *to)
4241 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4244 void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
4245 struct mem_cgroup **memcgp)
4247 struct mem_cgroup *memcg = NULL;
4248 unsigned int nr_pages = 1;
4249 struct page_cgroup *pc;
4250 enum charge_type ctype;
4254 if (mem_cgroup_disabled())
4257 if (PageTransHuge(page))
4258 nr_pages <<= compound_order(page);
4260 pc = lookup_page_cgroup(page);
4261 lock_page_cgroup(pc);
4262 if (PageCgroupUsed(pc)) {
4263 memcg = pc->mem_cgroup;
4264 css_get(&memcg->css);
4266 * At migrating an anonymous page, its mapcount goes down
4267 * to 0 and uncharge() will be called. But, even if it's fully
4268 * unmapped, migration may fail and this page has to be
4269 * charged again. We set MIGRATION flag here and delay uncharge
4270 * until end_migration() is called
4272 * Corner Case Thinking
4274 * When the old page was mapped as Anon and it's unmap-and-freed
4275 * while migration was ongoing.
4276 * If unmap finds the old page, uncharge() of it will be delayed
4277 * until end_migration(). If unmap finds a new page, it's
4278 * uncharged when it make mapcount to be 1->0. If unmap code
4279 * finds swap_migration_entry, the new page will not be mapped
4280 * and end_migration() will find it(mapcount==0).
4283 * When the old page was mapped but migraion fails, the kernel
4284 * remaps it. A charge for it is kept by MIGRATION flag even
4285 * if mapcount goes down to 0. We can do remap successfully
4286 * without charging it again.
4289 * The "old" page is under lock_page() until the end of
4290 * migration, so, the old page itself will not be swapped-out.
4291 * If the new page is swapped out before end_migraton, our
4292 * hook to usual swap-out path will catch the event.
4295 SetPageCgroupMigration(pc);
4297 unlock_page_cgroup(pc);
4299 * If the page is not charged at this point,
4307 * We charge new page before it's used/mapped. So, even if unlock_page()
4308 * is called before end_migration, we can catch all events on this new
4309 * page. In the case new page is migrated but not remapped, new page's
4310 * mapcount will be finally 0 and we call uncharge in end_migration().
4313 ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
4315 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
4317 * The page is committed to the memcg, but it's not actually
4318 * charged to the res_counter since we plan on replacing the
4319 * old one and only one page is going to be left afterwards.
4321 __mem_cgroup_commit_charge(memcg, newpage, nr_pages, ctype, false);
4324 /* remove redundant charge if migration failed*/
4325 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
4326 struct page *oldpage, struct page *newpage, bool migration_ok)
4328 struct page *used, *unused;
4329 struct page_cgroup *pc;
4335 if (!migration_ok) {
4342 anon = PageAnon(used);
4343 __mem_cgroup_uncharge_common(unused,
4344 anon ? MEM_CGROUP_CHARGE_TYPE_ANON
4345 : MEM_CGROUP_CHARGE_TYPE_CACHE,
4347 css_put(&memcg->css);
4349 * We disallowed uncharge of pages under migration because mapcount
4350 * of the page goes down to zero, temporarly.
4351 * Clear the flag and check the page should be charged.
4353 pc = lookup_page_cgroup(oldpage);
4354 lock_page_cgroup(pc);
4355 ClearPageCgroupMigration(pc);
4356 unlock_page_cgroup(pc);
4359 * If a page is a file cache, radix-tree replacement is very atomic
4360 * and we can skip this check. When it was an Anon page, its mapcount
4361 * goes down to 0. But because we added MIGRATION flage, it's not
4362 * uncharged yet. There are several case but page->mapcount check
4363 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4364 * check. (see prepare_charge() also)
4367 mem_cgroup_uncharge_page(used);
4371 * At replace page cache, newpage is not under any memcg but it's on
4372 * LRU. So, this function doesn't touch res_counter but handles LRU
4373 * in correct way. Both pages are locked so we cannot race with uncharge.
4375 void mem_cgroup_replace_page_cache(struct page *oldpage,
4376 struct page *newpage)
4378 struct mem_cgroup *memcg = NULL;
4379 struct page_cgroup *pc;
4380 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4382 if (mem_cgroup_disabled())
4385 pc = lookup_page_cgroup(oldpage);
4386 /* fix accounting on old pages */
4387 lock_page_cgroup(pc);
4388 if (PageCgroupUsed(pc)) {
4389 memcg = pc->mem_cgroup;
4390 mem_cgroup_charge_statistics(memcg, oldpage, false, -1);
4391 ClearPageCgroupUsed(pc);
4393 unlock_page_cgroup(pc);
4396 * When called from shmem_replace_page(), in some cases the
4397 * oldpage has already been charged, and in some cases not.
4402 * Even if newpage->mapping was NULL before starting replacement,
4403 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4404 * LRU while we overwrite pc->mem_cgroup.
4406 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
4409 #ifdef CONFIG_DEBUG_VM
4410 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
4412 struct page_cgroup *pc;
4414 pc = lookup_page_cgroup(page);
4416 * Can be NULL while feeding pages into the page allocator for
4417 * the first time, i.e. during boot or memory hotplug;
4418 * or when mem_cgroup_disabled().
4420 if (likely(pc) && PageCgroupUsed(pc))
4425 bool mem_cgroup_bad_page_check(struct page *page)
4427 if (mem_cgroup_disabled())
4430 return lookup_page_cgroup_used(page) != NULL;
4433 void mem_cgroup_print_bad_page(struct page *page)
4435 struct page_cgroup *pc;
4437 pc = lookup_page_cgroup_used(page);
4439 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4440 pc, pc->flags, pc->mem_cgroup);
4445 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
4446 unsigned long long val)
4449 u64 memswlimit, memlimit;
4451 int children = mem_cgroup_count_children(memcg);
4452 u64 curusage, oldusage;
4456 * For keeping hierarchical_reclaim simple, how long we should retry
4457 * is depends on callers. We set our retry-count to be function
4458 * of # of children which we should visit in this loop.
4460 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
4462 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4465 while (retry_count) {
4466 if (signal_pending(current)) {
4471 * Rather than hide all in some function, I do this in
4472 * open coded manner. You see what this really does.
4473 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4475 mutex_lock(&set_limit_mutex);
4476 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4477 if (memswlimit < val) {
4479 mutex_unlock(&set_limit_mutex);
4483 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4487 ret = res_counter_set_limit(&memcg->res, val);
4489 if (memswlimit == val)
4490 memcg->memsw_is_minimum = true;
4492 memcg->memsw_is_minimum = false;
4494 mutex_unlock(&set_limit_mutex);
4499 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4500 MEM_CGROUP_RECLAIM_SHRINK);
4501 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4502 /* Usage is reduced ? */
4503 if (curusage >= oldusage)
4506 oldusage = curusage;
4508 if (!ret && enlarge)
4509 memcg_oom_recover(memcg);
4514 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
4515 unsigned long long val)
4518 u64 memlimit, memswlimit, oldusage, curusage;
4519 int children = mem_cgroup_count_children(memcg);
4523 /* see mem_cgroup_resize_res_limit */
4524 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
4525 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4526 while (retry_count) {
4527 if (signal_pending(current)) {
4532 * Rather than hide all in some function, I do this in
4533 * open coded manner. You see what this really does.
4534 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4536 mutex_lock(&set_limit_mutex);
4537 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4538 if (memlimit > val) {
4540 mutex_unlock(&set_limit_mutex);
4543 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4544 if (memswlimit < val)
4546 ret = res_counter_set_limit(&memcg->memsw, val);
4548 if (memlimit == val)
4549 memcg->memsw_is_minimum = true;
4551 memcg->memsw_is_minimum = false;
4553 mutex_unlock(&set_limit_mutex);
4558 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4559 MEM_CGROUP_RECLAIM_NOSWAP |
4560 MEM_CGROUP_RECLAIM_SHRINK);
4561 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4562 /* Usage is reduced ? */
4563 if (curusage >= oldusage)
4566 oldusage = curusage;
4568 if (!ret && enlarge)
4569 memcg_oom_recover(memcg);
4573 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
4575 unsigned long *total_scanned)
4577 unsigned long nr_reclaimed = 0;
4578 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
4579 unsigned long reclaimed;
4581 struct mem_cgroup_tree_per_zone *mctz;
4582 unsigned long long excess;
4583 unsigned long nr_scanned;
4588 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
4590 * This loop can run a while, specially if mem_cgroup's continuously
4591 * keep exceeding their soft limit and putting the system under
4598 mz = mem_cgroup_largest_soft_limit_node(mctz);
4603 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
4604 gfp_mask, &nr_scanned);
4605 nr_reclaimed += reclaimed;
4606 *total_scanned += nr_scanned;
4607 spin_lock(&mctz->lock);
4610 * If we failed to reclaim anything from this memory cgroup
4611 * it is time to move on to the next cgroup
4617 * Loop until we find yet another one.
4619 * By the time we get the soft_limit lock
4620 * again, someone might have aded the
4621 * group back on the RB tree. Iterate to
4622 * make sure we get a different mem.
4623 * mem_cgroup_largest_soft_limit_node returns
4624 * NULL if no other cgroup is present on
4628 __mem_cgroup_largest_soft_limit_node(mctz);
4630 css_put(&next_mz->memcg->css);
4631 else /* next_mz == NULL or other memcg */
4635 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
4636 excess = res_counter_soft_limit_excess(&mz->memcg->res);
4638 * One school of thought says that we should not add
4639 * back the node to the tree if reclaim returns 0.
4640 * But our reclaim could return 0, simply because due
4641 * to priority we are exposing a smaller subset of
4642 * memory to reclaim from. Consider this as a longer
4645 /* If excess == 0, no tree ops */
4646 __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess);
4647 spin_unlock(&mctz->lock);
4648 css_put(&mz->memcg->css);
4651 * Could not reclaim anything and there are no more
4652 * mem cgroups to try or we seem to be looping without
4653 * reclaiming anything.
4655 if (!nr_reclaimed &&
4657 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
4659 } while (!nr_reclaimed);
4661 css_put(&next_mz->memcg->css);
4662 return nr_reclaimed;
4666 * mem_cgroup_force_empty_list - clears LRU of a group
4667 * @memcg: group to clear
4670 * @lru: lru to to clear
4672 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4673 * reclaim the pages page themselves - pages are moved to the parent (or root)
4676 static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
4677 int node, int zid, enum lru_list lru)
4679 struct lruvec *lruvec;
4680 unsigned long flags;
4681 struct list_head *list;
4685 zone = &NODE_DATA(node)->node_zones[zid];
4686 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
4687 list = &lruvec->lists[lru];
4691 struct page_cgroup *pc;
4694 spin_lock_irqsave(&zone->lru_lock, flags);
4695 if (list_empty(list)) {
4696 spin_unlock_irqrestore(&zone->lru_lock, flags);
4699 page = list_entry(list->prev, struct page, lru);
4701 list_move(&page->lru, list);
4703 spin_unlock_irqrestore(&zone->lru_lock, flags);
4706 spin_unlock_irqrestore(&zone->lru_lock, flags);
4708 pc = lookup_page_cgroup(page);
4710 if (mem_cgroup_move_parent(page, pc, memcg)) {
4711 /* found lock contention or "pc" is obsolete. */
4716 } while (!list_empty(list));
4720 * make mem_cgroup's charge to be 0 if there is no task by moving
4721 * all the charges and pages to the parent.
4722 * This enables deleting this mem_cgroup.
4724 * Caller is responsible for holding css reference on the memcg.
4726 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
4732 /* This is for making all *used* pages to be on LRU. */
4733 lru_add_drain_all();
4734 drain_all_stock_sync(memcg);
4735 mem_cgroup_start_move(memcg);
4736 for_each_node_state(node, N_MEMORY) {
4737 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4740 mem_cgroup_force_empty_list(memcg,
4745 mem_cgroup_end_move(memcg);
4746 memcg_oom_recover(memcg);
4750 * Kernel memory may not necessarily be trackable to a specific
4751 * process. So they are not migrated, and therefore we can't
4752 * expect their value to drop to 0 here.
4753 * Having res filled up with kmem only is enough.
4755 * This is a safety check because mem_cgroup_force_empty_list
4756 * could have raced with mem_cgroup_replace_page_cache callers
4757 * so the lru seemed empty but the page could have been added
4758 * right after the check. RES_USAGE should be safe as we always
4759 * charge before adding to the LRU.
4761 usage = res_counter_read_u64(&memcg->res, RES_USAGE) -
4762 res_counter_read_u64(&memcg->kmem, RES_USAGE);
4763 } while (usage > 0);
4766 static inline bool memcg_has_children(struct mem_cgroup *memcg)
4768 lockdep_assert_held(&memcg_create_mutex);
4770 * The lock does not prevent addition or deletion to the list
4771 * of children, but it prevents a new child from being
4772 * initialized based on this parent in css_online(), so it's
4773 * enough to decide whether hierarchically inherited
4774 * attributes can still be changed or not.
4776 return memcg->use_hierarchy &&
4777 !list_empty(&memcg->css.cgroup->children);
4781 * Reclaims as many pages from the given memcg as possible and moves
4782 * the rest to the parent.
4784 * Caller is responsible for holding css reference for memcg.
4786 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
4788 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
4789 struct cgroup *cgrp = memcg->css.cgroup;
4791 /* returns EBUSY if there is a task or if we come here twice. */
4792 if (cgroup_has_tasks(cgrp) || !list_empty(&cgrp->children))
4795 /* we call try-to-free pages for make this cgroup empty */
4796 lru_add_drain_all();
4797 /* try to free all pages in this cgroup */
4798 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
4801 if (signal_pending(current))
4804 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
4808 /* maybe some writeback is necessary */
4809 congestion_wait(BLK_RW_ASYNC, HZ/10);
4814 mem_cgroup_reparent_charges(memcg);
4819 static int mem_cgroup_force_empty_write(struct cgroup_subsys_state *css,
4822 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4824 if (mem_cgroup_is_root(memcg))
4826 return mem_cgroup_force_empty(memcg);
4829 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
4832 return mem_cgroup_from_css(css)->use_hierarchy;
4835 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
4836 struct cftype *cft, u64 val)
4839 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4840 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(css_parent(&memcg->css));
4842 mutex_lock(&memcg_create_mutex);
4844 if (memcg->use_hierarchy == val)
4848 * If parent's use_hierarchy is set, we can't make any modifications
4849 * in the child subtrees. If it is unset, then the change can
4850 * occur, provided the current cgroup has no children.
4852 * For the root cgroup, parent_mem is NULL, we allow value to be
4853 * set if there are no children.
4855 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
4856 (val == 1 || val == 0)) {
4857 if (list_empty(&memcg->css.cgroup->children))
4858 memcg->use_hierarchy = val;
4865 mutex_unlock(&memcg_create_mutex);
4871 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
4872 enum mem_cgroup_stat_index idx)
4874 struct mem_cgroup *iter;
4877 /* Per-cpu values can be negative, use a signed accumulator */
4878 for_each_mem_cgroup_tree(iter, memcg)
4879 val += mem_cgroup_read_stat(iter, idx);
4881 if (val < 0) /* race ? */
4886 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
4890 if (!mem_cgroup_is_root(memcg)) {
4892 return res_counter_read_u64(&memcg->res, RES_USAGE);
4894 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
4898 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
4899 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
4901 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
4902 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
4905 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
4907 return val << PAGE_SHIFT;
4910 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
4913 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4918 type = MEMFILE_TYPE(cft->private);
4919 name = MEMFILE_ATTR(cft->private);
4923 if (name == RES_USAGE)
4924 val = mem_cgroup_usage(memcg, false);
4926 val = res_counter_read_u64(&memcg->res, name);
4929 if (name == RES_USAGE)
4930 val = mem_cgroup_usage(memcg, true);
4932 val = res_counter_read_u64(&memcg->memsw, name);
4935 val = res_counter_read_u64(&memcg->kmem, name);
4944 #ifdef CONFIG_MEMCG_KMEM
4945 /* should be called with activate_kmem_mutex held */
4946 static int __memcg_activate_kmem(struct mem_cgroup *memcg,
4947 unsigned long long limit)
4952 if (memcg_kmem_is_active(memcg))
4956 * We are going to allocate memory for data shared by all memory
4957 * cgroups so let's stop accounting here.
4959 memcg_stop_kmem_account();
4962 * For simplicity, we won't allow this to be disabled. It also can't
4963 * be changed if the cgroup has children already, or if tasks had
4966 * If tasks join before we set the limit, a person looking at
4967 * kmem.usage_in_bytes will have no way to determine when it took
4968 * place, which makes the value quite meaningless.
4970 * After it first became limited, changes in the value of the limit are
4971 * of course permitted.
4973 mutex_lock(&memcg_create_mutex);
4974 if (cgroup_has_tasks(memcg->css.cgroup) || memcg_has_children(memcg))
4976 mutex_unlock(&memcg_create_mutex);
4980 memcg_id = ida_simple_get(&kmem_limited_groups,
4981 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
4988 * Make sure we have enough space for this cgroup in each root cache's
4991 mutex_lock(&memcg_slab_mutex);
4992 err = memcg_update_all_caches(memcg_id + 1);
4993 mutex_unlock(&memcg_slab_mutex);
4997 memcg->kmemcg_id = memcg_id;
4998 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
5001 * We couldn't have accounted to this cgroup, because it hasn't got the
5002 * active bit set yet, so this should succeed.
5004 err = res_counter_set_limit(&memcg->kmem, limit);
5007 static_key_slow_inc(&memcg_kmem_enabled_key);
5009 * Setting the active bit after enabling static branching will
5010 * guarantee no one starts accounting before all call sites are
5013 memcg_kmem_set_active(memcg);
5015 memcg_resume_kmem_account();
5019 ida_simple_remove(&kmem_limited_groups, memcg_id);
5023 static int memcg_activate_kmem(struct mem_cgroup *memcg,
5024 unsigned long long limit)
5028 mutex_lock(&activate_kmem_mutex);
5029 ret = __memcg_activate_kmem(memcg, limit);
5030 mutex_unlock(&activate_kmem_mutex);
5034 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
5035 unsigned long long val)
5039 if (!memcg_kmem_is_active(memcg))
5040 ret = memcg_activate_kmem(memcg, val);
5042 ret = res_counter_set_limit(&memcg->kmem, val);
5046 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
5049 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5054 mutex_lock(&activate_kmem_mutex);
5056 * If the parent cgroup is not kmem-active now, it cannot be activated
5057 * after this point, because it has at least one child already.
5059 if (memcg_kmem_is_active(parent))
5060 ret = __memcg_activate_kmem(memcg, RES_COUNTER_MAX);
5061 mutex_unlock(&activate_kmem_mutex);
5065 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
5066 unsigned long long val)
5070 #endif /* CONFIG_MEMCG_KMEM */
5073 * The user of this function is...
5076 static int mem_cgroup_write(struct cgroup_subsys_state *css, struct cftype *cft,
5079 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5082 unsigned long long val;
5085 type = MEMFILE_TYPE(cft->private);
5086 name = MEMFILE_ATTR(cft->private);
5090 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
5094 /* This function does all necessary parse...reuse it */
5095 ret = res_counter_memparse_write_strategy(buffer, &val);
5099 ret = mem_cgroup_resize_limit(memcg, val);
5100 else if (type == _MEMSWAP)
5101 ret = mem_cgroup_resize_memsw_limit(memcg, val);
5102 else if (type == _KMEM)
5103 ret = memcg_update_kmem_limit(memcg, val);
5107 case RES_SOFT_LIMIT:
5108 ret = res_counter_memparse_write_strategy(buffer, &val);
5112 * For memsw, soft limits are hard to implement in terms
5113 * of semantics, for now, we support soft limits for
5114 * control without swap
5117 ret = res_counter_set_soft_limit(&memcg->res, val);
5122 ret = -EINVAL; /* should be BUG() ? */
5128 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
5129 unsigned long long *mem_limit, unsigned long long *memsw_limit)
5131 unsigned long long min_limit, min_memsw_limit, tmp;
5133 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
5134 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5135 if (!memcg->use_hierarchy)
5138 while (css_parent(&memcg->css)) {
5139 memcg = mem_cgroup_from_css(css_parent(&memcg->css));
5140 if (!memcg->use_hierarchy)
5142 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
5143 min_limit = min(min_limit, tmp);
5144 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5145 min_memsw_limit = min(min_memsw_limit, tmp);
5148 *mem_limit = min_limit;
5149 *memsw_limit = min_memsw_limit;
5152 static int mem_cgroup_reset(struct cgroup_subsys_state *css, unsigned int event)
5154 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5158 type = MEMFILE_TYPE(event);
5159 name = MEMFILE_ATTR(event);
5164 res_counter_reset_max(&memcg->res);
5165 else if (type == _MEMSWAP)
5166 res_counter_reset_max(&memcg->memsw);
5167 else if (type == _KMEM)
5168 res_counter_reset_max(&memcg->kmem);
5174 res_counter_reset_failcnt(&memcg->res);
5175 else if (type == _MEMSWAP)
5176 res_counter_reset_failcnt(&memcg->memsw);
5177 else if (type == _KMEM)
5178 res_counter_reset_failcnt(&memcg->kmem);
5187 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
5190 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
5194 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5195 struct cftype *cft, u64 val)
5197 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5199 if (val >= (1 << NR_MOVE_TYPE))
5203 * No kind of locking is needed in here, because ->can_attach() will
5204 * check this value once in the beginning of the process, and then carry
5205 * on with stale data. This means that changes to this value will only
5206 * affect task migrations starting after the change.
5208 memcg->move_charge_at_immigrate = val;
5212 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5213 struct cftype *cft, u64 val)
5220 static int memcg_numa_stat_show(struct seq_file *m, void *v)
5224 unsigned int lru_mask;
5227 static const struct numa_stat stats[] = {
5228 { "total", LRU_ALL },
5229 { "file", LRU_ALL_FILE },
5230 { "anon", LRU_ALL_ANON },
5231 { "unevictable", BIT(LRU_UNEVICTABLE) },
5233 const struct numa_stat *stat;
5236 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5238 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
5239 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
5240 seq_printf(m, "%s=%lu", stat->name, nr);
5241 for_each_node_state(nid, N_MEMORY) {
5242 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5244 seq_printf(m, " N%d=%lu", nid, nr);
5249 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
5250 struct mem_cgroup *iter;
5253 for_each_mem_cgroup_tree(iter, memcg)
5254 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
5255 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
5256 for_each_node_state(nid, N_MEMORY) {
5258 for_each_mem_cgroup_tree(iter, memcg)
5259 nr += mem_cgroup_node_nr_lru_pages(
5260 iter, nid, stat->lru_mask);
5261 seq_printf(m, " N%d=%lu", nid, nr);
5268 #endif /* CONFIG_NUMA */
5270 static inline void mem_cgroup_lru_names_not_uptodate(void)
5272 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
5275 static int memcg_stat_show(struct seq_file *m, void *v)
5277 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5278 struct mem_cgroup *mi;
5281 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5282 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5284 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
5285 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
5288 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
5289 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
5290 mem_cgroup_read_events(memcg, i));
5292 for (i = 0; i < NR_LRU_LISTS; i++)
5293 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
5294 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
5296 /* Hierarchical information */
5298 unsigned long long limit, memsw_limit;
5299 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
5300 seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
5301 if (do_swap_account)
5302 seq_printf(m, "hierarchical_memsw_limit %llu\n",
5306 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5309 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5311 for_each_mem_cgroup_tree(mi, memcg)
5312 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
5313 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
5316 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
5317 unsigned long long val = 0;
5319 for_each_mem_cgroup_tree(mi, memcg)
5320 val += mem_cgroup_read_events(mi, i);
5321 seq_printf(m, "total_%s %llu\n",
5322 mem_cgroup_events_names[i], val);
5325 for (i = 0; i < NR_LRU_LISTS; i++) {
5326 unsigned long long val = 0;
5328 for_each_mem_cgroup_tree(mi, memcg)
5329 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
5330 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
5333 #ifdef CONFIG_DEBUG_VM
5336 struct mem_cgroup_per_zone *mz;
5337 struct zone_reclaim_stat *rstat;
5338 unsigned long recent_rotated[2] = {0, 0};
5339 unsigned long recent_scanned[2] = {0, 0};
5341 for_each_online_node(nid)
5342 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
5343 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
5344 rstat = &mz->lruvec.reclaim_stat;
5346 recent_rotated[0] += rstat->recent_rotated[0];
5347 recent_rotated[1] += rstat->recent_rotated[1];
5348 recent_scanned[0] += rstat->recent_scanned[0];
5349 recent_scanned[1] += rstat->recent_scanned[1];
5351 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
5352 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
5353 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
5354 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
5361 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
5364 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5366 return mem_cgroup_swappiness(memcg);
5369 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
5370 struct cftype *cft, u64 val)
5372 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5377 if (css_parent(css))
5378 memcg->swappiness = val;
5380 vm_swappiness = val;
5385 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
5387 struct mem_cgroup_threshold_ary *t;
5393 t = rcu_dereference(memcg->thresholds.primary);
5395 t = rcu_dereference(memcg->memsw_thresholds.primary);
5400 usage = mem_cgroup_usage(memcg, swap);
5403 * current_threshold points to threshold just below or equal to usage.
5404 * If it's not true, a threshold was crossed after last
5405 * call of __mem_cgroup_threshold().
5407 i = t->current_threshold;
5410 * Iterate backward over array of thresholds starting from
5411 * current_threshold and check if a threshold is crossed.
5412 * If none of thresholds below usage is crossed, we read
5413 * only one element of the array here.
5415 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
5416 eventfd_signal(t->entries[i].eventfd, 1);
5418 /* i = current_threshold + 1 */
5422 * Iterate forward over array of thresholds starting from
5423 * current_threshold+1 and check if a threshold is crossed.
5424 * If none of thresholds above usage is crossed, we read
5425 * only one element of the array here.
5427 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
5428 eventfd_signal(t->entries[i].eventfd, 1);
5430 /* Update current_threshold */
5431 t->current_threshold = i - 1;
5436 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
5439 __mem_cgroup_threshold(memcg, false);
5440 if (do_swap_account)
5441 __mem_cgroup_threshold(memcg, true);
5443 memcg = parent_mem_cgroup(memcg);
5447 static int compare_thresholds(const void *a, const void *b)
5449 const struct mem_cgroup_threshold *_a = a;
5450 const struct mem_cgroup_threshold *_b = b;
5452 if (_a->threshold > _b->threshold)
5455 if (_a->threshold < _b->threshold)
5461 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
5463 struct mem_cgroup_eventfd_list *ev;
5465 list_for_each_entry(ev, &memcg->oom_notify, list)
5466 eventfd_signal(ev->eventfd, 1);
5470 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
5472 struct mem_cgroup *iter;
5474 for_each_mem_cgroup_tree(iter, memcg)
5475 mem_cgroup_oom_notify_cb(iter);
5478 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
5479 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
5481 struct mem_cgroup_thresholds *thresholds;
5482 struct mem_cgroup_threshold_ary *new;
5483 u64 threshold, usage;
5486 ret = res_counter_memparse_write_strategy(args, &threshold);
5490 mutex_lock(&memcg->thresholds_lock);
5493 thresholds = &memcg->thresholds;
5494 else if (type == _MEMSWAP)
5495 thresholds = &memcg->memsw_thresholds;
5499 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5501 /* Check if a threshold crossed before adding a new one */
5502 if (thresholds->primary)
5503 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5505 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
5507 /* Allocate memory for new array of thresholds */
5508 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
5516 /* Copy thresholds (if any) to new array */
5517 if (thresholds->primary) {
5518 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
5519 sizeof(struct mem_cgroup_threshold));
5522 /* Add new threshold */
5523 new->entries[size - 1].eventfd = eventfd;
5524 new->entries[size - 1].threshold = threshold;
5526 /* Sort thresholds. Registering of new threshold isn't time-critical */
5527 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
5528 compare_thresholds, NULL);
5530 /* Find current threshold */
5531 new->current_threshold = -1;
5532 for (i = 0; i < size; i++) {
5533 if (new->entries[i].threshold <= usage) {
5535 * new->current_threshold will not be used until
5536 * rcu_assign_pointer(), so it's safe to increment
5539 ++new->current_threshold;
5544 /* Free old spare buffer and save old primary buffer as spare */
5545 kfree(thresholds->spare);
5546 thresholds->spare = thresholds->primary;
5548 rcu_assign_pointer(thresholds->primary, new);
5550 /* To be sure that nobody uses thresholds */
5554 mutex_unlock(&memcg->thresholds_lock);
5559 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
5560 struct eventfd_ctx *eventfd, const char *args)
5562 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
5565 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
5566 struct eventfd_ctx *eventfd, const char *args)
5568 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
5571 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
5572 struct eventfd_ctx *eventfd, enum res_type type)
5574 struct mem_cgroup_thresholds *thresholds;
5575 struct mem_cgroup_threshold_ary *new;
5579 mutex_lock(&memcg->thresholds_lock);
5581 thresholds = &memcg->thresholds;
5582 else if (type == _MEMSWAP)
5583 thresholds = &memcg->memsw_thresholds;
5587 if (!thresholds->primary)
5590 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5592 /* Check if a threshold crossed before removing */
5593 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5595 /* Calculate new number of threshold */
5597 for (i = 0; i < thresholds->primary->size; i++) {
5598 if (thresholds->primary->entries[i].eventfd != eventfd)
5602 new = thresholds->spare;
5604 /* Set thresholds array to NULL if we don't have thresholds */
5613 /* Copy thresholds and find current threshold */
5614 new->current_threshold = -1;
5615 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
5616 if (thresholds->primary->entries[i].eventfd == eventfd)
5619 new->entries[j] = thresholds->primary->entries[i];
5620 if (new->entries[j].threshold <= usage) {
5622 * new->current_threshold will not be used
5623 * until rcu_assign_pointer(), so it's safe to increment
5626 ++new->current_threshold;
5632 /* Swap primary and spare array */
5633 thresholds->spare = thresholds->primary;
5634 /* If all events are unregistered, free the spare array */
5636 kfree(thresholds->spare);
5637 thresholds->spare = NULL;
5640 rcu_assign_pointer(thresholds->primary, new);
5642 /* To be sure that nobody uses thresholds */
5645 mutex_unlock(&memcg->thresholds_lock);
5648 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
5649 struct eventfd_ctx *eventfd)
5651 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
5654 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
5655 struct eventfd_ctx *eventfd)
5657 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
5660 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
5661 struct eventfd_ctx *eventfd, const char *args)
5663 struct mem_cgroup_eventfd_list *event;
5665 event = kmalloc(sizeof(*event), GFP_KERNEL);
5669 spin_lock(&memcg_oom_lock);
5671 event->eventfd = eventfd;
5672 list_add(&event->list, &memcg->oom_notify);
5674 /* already in OOM ? */
5675 if (atomic_read(&memcg->under_oom))
5676 eventfd_signal(eventfd, 1);
5677 spin_unlock(&memcg_oom_lock);
5682 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
5683 struct eventfd_ctx *eventfd)
5685 struct mem_cgroup_eventfd_list *ev, *tmp;
5687 spin_lock(&memcg_oom_lock);
5689 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
5690 if (ev->eventfd == eventfd) {
5691 list_del(&ev->list);
5696 spin_unlock(&memcg_oom_lock);
5699 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
5701 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
5703 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
5704 seq_printf(sf, "under_oom %d\n", (bool)atomic_read(&memcg->under_oom));
5708 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
5709 struct cftype *cft, u64 val)
5711 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5713 /* cannot set to root cgroup and only 0 and 1 are allowed */
5714 if (!css_parent(css) || !((val == 0) || (val == 1)))
5717 memcg->oom_kill_disable = val;
5719 memcg_oom_recover(memcg);
5724 #ifdef CONFIG_MEMCG_KMEM
5725 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5729 memcg->kmemcg_id = -1;
5730 ret = memcg_propagate_kmem(memcg);
5734 return mem_cgroup_sockets_init(memcg, ss);
5737 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5739 mem_cgroup_sockets_destroy(memcg);
5742 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5744 if (!memcg_kmem_is_active(memcg))
5748 * kmem charges can outlive the cgroup. In the case of slab
5749 * pages, for instance, a page contain objects from various
5750 * processes. As we prevent from taking a reference for every
5751 * such allocation we have to be careful when doing uncharge
5752 * (see memcg_uncharge_kmem) and here during offlining.
5754 * The idea is that that only the _last_ uncharge which sees
5755 * the dead memcg will drop the last reference. An additional
5756 * reference is taken here before the group is marked dead
5757 * which is then paired with css_put during uncharge resp. here.
5759 * Although this might sound strange as this path is called from
5760 * css_offline() when the referencemight have dropped down to 0
5761 * and shouldn't be incremented anymore (css_tryget would fail)
5762 * we do not have other options because of the kmem allocations
5765 css_get(&memcg->css);
5767 memcg_kmem_mark_dead(memcg);
5769 if (res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0)
5772 if (memcg_kmem_test_and_clear_dead(memcg))
5773 css_put(&memcg->css);
5776 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5781 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5785 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5791 * DO NOT USE IN NEW FILES.
5793 * "cgroup.event_control" implementation.
5795 * This is way over-engineered. It tries to support fully configurable
5796 * events for each user. Such level of flexibility is completely
5797 * unnecessary especially in the light of the planned unified hierarchy.
5799 * Please deprecate this and replace with something simpler if at all
5804 * Unregister event and free resources.
5806 * Gets called from workqueue.
5808 static void memcg_event_remove(struct work_struct *work)
5810 struct mem_cgroup_event *event =
5811 container_of(work, struct mem_cgroup_event, remove);
5812 struct mem_cgroup *memcg = event->memcg;
5814 remove_wait_queue(event->wqh, &event->wait);
5816 event->unregister_event(memcg, event->eventfd);
5818 /* Notify userspace the event is going away. */
5819 eventfd_signal(event->eventfd, 1);
5821 eventfd_ctx_put(event->eventfd);
5823 css_put(&memcg->css);
5827 * Gets called on POLLHUP on eventfd when user closes it.
5829 * Called with wqh->lock held and interrupts disabled.
5831 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
5832 int sync, void *key)
5834 struct mem_cgroup_event *event =
5835 container_of(wait, struct mem_cgroup_event, wait);
5836 struct mem_cgroup *memcg = event->memcg;
5837 unsigned long flags = (unsigned long)key;
5839 if (flags & POLLHUP) {
5841 * If the event has been detached at cgroup removal, we
5842 * can simply return knowing the other side will cleanup
5845 * We can't race against event freeing since the other
5846 * side will require wqh->lock via remove_wait_queue(),
5849 spin_lock(&memcg->event_list_lock);
5850 if (!list_empty(&event->list)) {
5851 list_del_init(&event->list);
5853 * We are in atomic context, but cgroup_event_remove()
5854 * may sleep, so we have to call it in workqueue.
5856 schedule_work(&event->remove);
5858 spin_unlock(&memcg->event_list_lock);
5864 static void memcg_event_ptable_queue_proc(struct file *file,
5865 wait_queue_head_t *wqh, poll_table *pt)
5867 struct mem_cgroup_event *event =
5868 container_of(pt, struct mem_cgroup_event, pt);
5871 add_wait_queue(wqh, &event->wait);
5875 * DO NOT USE IN NEW FILES.
5877 * Parse input and register new cgroup event handler.
5879 * Input must be in format '<event_fd> <control_fd> <args>'.
5880 * Interpretation of args is defined by control file implementation.
5882 static int memcg_write_event_control(struct cgroup_subsys_state *css,
5883 struct cftype *cft, char *buffer)
5885 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5886 struct mem_cgroup_event *event;
5887 struct cgroup_subsys_state *cfile_css;
5888 unsigned int efd, cfd;
5895 efd = simple_strtoul(buffer, &endp, 10);
5900 cfd = simple_strtoul(buffer, &endp, 10);
5901 if ((*endp != ' ') && (*endp != '\0'))
5905 event = kzalloc(sizeof(*event), GFP_KERNEL);
5909 event->memcg = memcg;
5910 INIT_LIST_HEAD(&event->list);
5911 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
5912 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
5913 INIT_WORK(&event->remove, memcg_event_remove);
5921 event->eventfd = eventfd_ctx_fileget(efile.file);
5922 if (IS_ERR(event->eventfd)) {
5923 ret = PTR_ERR(event->eventfd);
5930 goto out_put_eventfd;
5933 /* the process need read permission on control file */
5934 /* AV: shouldn't we check that it's been opened for read instead? */
5935 ret = inode_permission(file_inode(cfile.file), MAY_READ);
5940 * Determine the event callbacks and set them in @event. This used
5941 * to be done via struct cftype but cgroup core no longer knows
5942 * about these events. The following is crude but the whole thing
5943 * is for compatibility anyway.
5945 * DO NOT ADD NEW FILES.
5947 name = cfile.file->f_dentry->d_name.name;
5949 if (!strcmp(name, "memory.usage_in_bytes")) {
5950 event->register_event = mem_cgroup_usage_register_event;
5951 event->unregister_event = mem_cgroup_usage_unregister_event;
5952 } else if (!strcmp(name, "memory.oom_control")) {
5953 event->register_event = mem_cgroup_oom_register_event;
5954 event->unregister_event = mem_cgroup_oom_unregister_event;
5955 } else if (!strcmp(name, "memory.pressure_level")) {
5956 event->register_event = vmpressure_register_event;
5957 event->unregister_event = vmpressure_unregister_event;
5958 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
5959 event->register_event = memsw_cgroup_usage_register_event;
5960 event->unregister_event = memsw_cgroup_usage_unregister_event;
5967 * Verify @cfile should belong to @css. Also, remaining events are
5968 * automatically removed on cgroup destruction but the removal is
5969 * asynchronous, so take an extra ref on @css.
5971 cfile_css = css_tryget_from_dir(cfile.file->f_dentry->d_parent,
5972 &memory_cgrp_subsys);
5974 if (IS_ERR(cfile_css))
5976 if (cfile_css != css) {
5981 ret = event->register_event(memcg, event->eventfd, buffer);
5985 efile.file->f_op->poll(efile.file, &event->pt);
5987 spin_lock(&memcg->event_list_lock);
5988 list_add(&event->list, &memcg->event_list);
5989 spin_unlock(&memcg->event_list_lock);
6001 eventfd_ctx_put(event->eventfd);
6010 static struct cftype mem_cgroup_files[] = {
6012 .name = "usage_in_bytes",
6013 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
6014 .read_u64 = mem_cgroup_read_u64,
6017 .name = "max_usage_in_bytes",
6018 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
6019 .trigger = mem_cgroup_reset,
6020 .read_u64 = mem_cgroup_read_u64,
6023 .name = "limit_in_bytes",
6024 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
6025 .write_string = mem_cgroup_write,
6026 .read_u64 = mem_cgroup_read_u64,
6029 .name = "soft_limit_in_bytes",
6030 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
6031 .write_string = mem_cgroup_write,
6032 .read_u64 = mem_cgroup_read_u64,
6036 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
6037 .trigger = mem_cgroup_reset,
6038 .read_u64 = mem_cgroup_read_u64,
6042 .seq_show = memcg_stat_show,
6045 .name = "force_empty",
6046 .trigger = mem_cgroup_force_empty_write,
6049 .name = "use_hierarchy",
6050 .flags = CFTYPE_INSANE,
6051 .write_u64 = mem_cgroup_hierarchy_write,
6052 .read_u64 = mem_cgroup_hierarchy_read,
6055 .name = "cgroup.event_control", /* XXX: for compat */
6056 .write_string = memcg_write_event_control,
6057 .flags = CFTYPE_NO_PREFIX,
6061 .name = "swappiness",
6062 .read_u64 = mem_cgroup_swappiness_read,
6063 .write_u64 = mem_cgroup_swappiness_write,
6066 .name = "move_charge_at_immigrate",
6067 .read_u64 = mem_cgroup_move_charge_read,
6068 .write_u64 = mem_cgroup_move_charge_write,
6071 .name = "oom_control",
6072 .seq_show = mem_cgroup_oom_control_read,
6073 .write_u64 = mem_cgroup_oom_control_write,
6074 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
6077 .name = "pressure_level",
6081 .name = "numa_stat",
6082 .seq_show = memcg_numa_stat_show,
6085 #ifdef CONFIG_MEMCG_KMEM
6087 .name = "kmem.limit_in_bytes",
6088 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
6089 .write_string = mem_cgroup_write,
6090 .read_u64 = mem_cgroup_read_u64,
6093 .name = "kmem.usage_in_bytes",
6094 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
6095 .read_u64 = mem_cgroup_read_u64,
6098 .name = "kmem.failcnt",
6099 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
6100 .trigger = mem_cgroup_reset,
6101 .read_u64 = mem_cgroup_read_u64,
6104 .name = "kmem.max_usage_in_bytes",
6105 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
6106 .trigger = mem_cgroup_reset,
6107 .read_u64 = mem_cgroup_read_u64,
6109 #ifdef CONFIG_SLABINFO
6111 .name = "kmem.slabinfo",
6112 .seq_show = mem_cgroup_slabinfo_read,
6116 { }, /* terminate */
6119 #ifdef CONFIG_MEMCG_SWAP
6120 static struct cftype memsw_cgroup_files[] = {
6122 .name = "memsw.usage_in_bytes",
6123 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6124 .read_u64 = mem_cgroup_read_u64,
6127 .name = "memsw.max_usage_in_bytes",
6128 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6129 .trigger = mem_cgroup_reset,
6130 .read_u64 = mem_cgroup_read_u64,
6133 .name = "memsw.limit_in_bytes",
6134 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6135 .write_string = mem_cgroup_write,
6136 .read_u64 = mem_cgroup_read_u64,
6139 .name = "memsw.failcnt",
6140 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6141 .trigger = mem_cgroup_reset,
6142 .read_u64 = mem_cgroup_read_u64,
6144 { }, /* terminate */
6147 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6149 struct mem_cgroup_per_node *pn;
6150 struct mem_cgroup_per_zone *mz;
6151 int zone, tmp = node;
6153 * This routine is called against possible nodes.
6154 * But it's BUG to call kmalloc() against offline node.
6156 * TODO: this routine can waste much memory for nodes which will
6157 * never be onlined. It's better to use memory hotplug callback
6160 if (!node_state(node, N_NORMAL_MEMORY))
6162 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
6166 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6167 mz = &pn->zoneinfo[zone];
6168 lruvec_init(&mz->lruvec);
6169 mz->usage_in_excess = 0;
6170 mz->on_tree = false;
6173 memcg->nodeinfo[node] = pn;
6177 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6179 kfree(memcg->nodeinfo[node]);
6182 static struct mem_cgroup *mem_cgroup_alloc(void)
6184 struct mem_cgroup *memcg;
6187 size = sizeof(struct mem_cgroup);
6188 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
6190 memcg = kzalloc(size, GFP_KERNEL);
6194 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
6197 spin_lock_init(&memcg->pcp_counter_lock);
6206 * At destroying mem_cgroup, references from swap_cgroup can remain.
6207 * (scanning all at force_empty is too costly...)
6209 * Instead of clearing all references at force_empty, we remember
6210 * the number of reference from swap_cgroup and free mem_cgroup when
6211 * it goes down to 0.
6213 * Removal of cgroup itself succeeds regardless of refs from swap.
6216 static void __mem_cgroup_free(struct mem_cgroup *memcg)
6220 mem_cgroup_remove_from_trees(memcg);
6223 free_mem_cgroup_per_zone_info(memcg, node);
6225 free_percpu(memcg->stat);
6228 * We need to make sure that (at least for now), the jump label
6229 * destruction code runs outside of the cgroup lock. This is because
6230 * get_online_cpus(), which is called from the static_branch update,
6231 * can't be called inside the cgroup_lock. cpusets are the ones
6232 * enforcing this dependency, so if they ever change, we might as well.
6234 * schedule_work() will guarantee this happens. Be careful if you need
6235 * to move this code around, and make sure it is outside
6238 disarm_static_keys(memcg);
6243 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6245 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
6247 if (!memcg->res.parent)
6249 return mem_cgroup_from_res_counter(memcg->res.parent, res);
6251 EXPORT_SYMBOL(parent_mem_cgroup);
6253 static void __init mem_cgroup_soft_limit_tree_init(void)
6255 struct mem_cgroup_tree_per_node *rtpn;
6256 struct mem_cgroup_tree_per_zone *rtpz;
6257 int tmp, node, zone;
6259 for_each_node(node) {
6261 if (!node_state(node, N_NORMAL_MEMORY))
6263 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
6266 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6268 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6269 rtpz = &rtpn->rb_tree_per_zone[zone];
6270 rtpz->rb_root = RB_ROOT;
6271 spin_lock_init(&rtpz->lock);
6276 static struct cgroup_subsys_state * __ref
6277 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
6279 struct mem_cgroup *memcg;
6280 long error = -ENOMEM;
6283 memcg = mem_cgroup_alloc();
6285 return ERR_PTR(error);
6288 if (alloc_mem_cgroup_per_zone_info(memcg, node))
6292 if (parent_css == NULL) {
6293 root_mem_cgroup = memcg;
6294 res_counter_init(&memcg->res, NULL);
6295 res_counter_init(&memcg->memsw, NULL);
6296 res_counter_init(&memcg->kmem, NULL);
6299 memcg->last_scanned_node = MAX_NUMNODES;
6300 INIT_LIST_HEAD(&memcg->oom_notify);
6301 memcg->move_charge_at_immigrate = 0;
6302 mutex_init(&memcg->thresholds_lock);
6303 spin_lock_init(&memcg->move_lock);
6304 vmpressure_init(&memcg->vmpressure);
6305 INIT_LIST_HEAD(&memcg->event_list);
6306 spin_lock_init(&memcg->event_list_lock);
6311 __mem_cgroup_free(memcg);
6312 return ERR_PTR(error);
6316 mem_cgroup_css_online(struct cgroup_subsys_state *css)
6318 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6319 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(css));
6321 if (css->cgroup->id > MEM_CGROUP_ID_MAX)
6327 mutex_lock(&memcg_create_mutex);
6329 memcg->use_hierarchy = parent->use_hierarchy;
6330 memcg->oom_kill_disable = parent->oom_kill_disable;
6331 memcg->swappiness = mem_cgroup_swappiness(parent);
6333 if (parent->use_hierarchy) {
6334 res_counter_init(&memcg->res, &parent->res);
6335 res_counter_init(&memcg->memsw, &parent->memsw);
6336 res_counter_init(&memcg->kmem, &parent->kmem);
6339 * No need to take a reference to the parent because cgroup
6340 * core guarantees its existence.
6343 res_counter_init(&memcg->res, NULL);
6344 res_counter_init(&memcg->memsw, NULL);
6345 res_counter_init(&memcg->kmem, NULL);
6347 * Deeper hierachy with use_hierarchy == false doesn't make
6348 * much sense so let cgroup subsystem know about this
6349 * unfortunate state in our controller.
6351 if (parent != root_mem_cgroup)
6352 memory_cgrp_subsys.broken_hierarchy = true;
6354 mutex_unlock(&memcg_create_mutex);
6356 return memcg_init_kmem(memcg, &memory_cgrp_subsys);
6360 * Announce all parents that a group from their hierarchy is gone.
6362 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup *memcg)
6364 struct mem_cgroup *parent = memcg;
6366 while ((parent = parent_mem_cgroup(parent)))
6367 mem_cgroup_iter_invalidate(parent);
6370 * if the root memcg is not hierarchical we have to check it
6373 if (!root_mem_cgroup->use_hierarchy)
6374 mem_cgroup_iter_invalidate(root_mem_cgroup);
6377 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
6379 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6380 struct mem_cgroup_event *event, *tmp;
6381 struct cgroup_subsys_state *iter;
6384 * Unregister events and notify userspace.
6385 * Notify userspace about cgroup removing only after rmdir of cgroup
6386 * directory to avoid race between userspace and kernelspace.
6388 spin_lock(&memcg->event_list_lock);
6389 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
6390 list_del_init(&event->list);
6391 schedule_work(&event->remove);
6393 spin_unlock(&memcg->event_list_lock);
6395 kmem_cgroup_css_offline(memcg);
6397 mem_cgroup_invalidate_reclaim_iterators(memcg);
6400 * This requires that offlining is serialized. Right now that is
6401 * guaranteed because css_killed_work_fn() holds the cgroup_mutex.
6403 css_for_each_descendant_post(iter, css)
6404 mem_cgroup_reparent_charges(mem_cgroup_from_css(iter));
6406 mem_cgroup_destroy_all_caches(memcg);
6407 vmpressure_cleanup(&memcg->vmpressure);
6410 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
6412 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6414 * XXX: css_offline() would be where we should reparent all
6415 * memory to prepare the cgroup for destruction. However,
6416 * memcg does not do css_tryget() and res_counter charging
6417 * under the same RCU lock region, which means that charging
6418 * could race with offlining. Offlining only happens to
6419 * cgroups with no tasks in them but charges can show up
6420 * without any tasks from the swapin path when the target
6421 * memcg is looked up from the swapout record and not from the
6422 * current task as it usually is. A race like this can leak
6423 * charges and put pages with stale cgroup pointers into
6427 * lookup_swap_cgroup_id()
6429 * mem_cgroup_lookup()
6432 * disable css_tryget()
6435 * reparent_charges()
6436 * res_counter_charge()
6439 * pc->mem_cgroup = dead memcg
6442 * The bulk of the charges are still moved in offline_css() to
6443 * avoid pinning a lot of pages in case a long-term reference
6444 * like a swapout record is deferring the css_free() to long
6445 * after offlining. But this makes sure we catch any charges
6446 * made after offlining:
6448 mem_cgroup_reparent_charges(memcg);
6450 memcg_destroy_kmem(memcg);
6451 __mem_cgroup_free(memcg);
6455 /* Handlers for move charge at task migration. */
6456 #define PRECHARGE_COUNT_AT_ONCE 256
6457 static int mem_cgroup_do_precharge(unsigned long count)
6460 int batch_count = PRECHARGE_COUNT_AT_ONCE;
6461 struct mem_cgroup *memcg = mc.to;
6463 if (mem_cgroup_is_root(memcg)) {
6464 mc.precharge += count;
6465 /* we don't need css_get for root */
6468 /* try to charge at once */
6470 struct res_counter *dummy;
6472 * "memcg" cannot be under rmdir() because we've already checked
6473 * by cgroup_lock_live_cgroup() that it is not removed and we
6474 * are still under the same cgroup_mutex. So we can postpone
6477 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
6479 if (do_swap_account && res_counter_charge(&memcg->memsw,
6480 PAGE_SIZE * count, &dummy)) {
6481 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
6484 mc.precharge += count;
6488 /* fall back to one by one charge */
6490 if (signal_pending(current)) {
6494 if (!batch_count--) {
6495 batch_count = PRECHARGE_COUNT_AT_ONCE;
6498 ret = mem_cgroup_try_charge(memcg, GFP_KERNEL, 1, false);
6500 /* mem_cgroup_clear_mc() will do uncharge later */
6508 * get_mctgt_type - get target type of moving charge
6509 * @vma: the vma the pte to be checked belongs
6510 * @addr: the address corresponding to the pte to be checked
6511 * @ptent: the pte to be checked
6512 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6515 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6516 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6517 * move charge. if @target is not NULL, the page is stored in target->page
6518 * with extra refcnt got(Callers should handle it).
6519 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6520 * target for charge migration. if @target is not NULL, the entry is stored
6523 * Called with pte lock held.
6530 enum mc_target_type {
6536 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
6537 unsigned long addr, pte_t ptent)
6539 struct page *page = vm_normal_page(vma, addr, ptent);
6541 if (!page || !page_mapped(page))
6543 if (PageAnon(page)) {
6544 /* we don't move shared anon */
6547 } else if (!move_file())
6548 /* we ignore mapcount for file pages */
6550 if (!get_page_unless_zero(page))
6557 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6558 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6560 struct page *page = NULL;
6561 swp_entry_t ent = pte_to_swp_entry(ptent);
6563 if (!move_anon() || non_swap_entry(ent))
6566 * Because lookup_swap_cache() updates some statistics counter,
6567 * we call find_get_page() with swapper_space directly.
6569 page = find_get_page(swap_address_space(ent), ent.val);
6570 if (do_swap_account)
6571 entry->val = ent.val;
6576 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6577 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6583 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
6584 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6586 struct page *page = NULL;
6587 struct address_space *mapping;
6590 if (!vma->vm_file) /* anonymous vma */
6595 mapping = vma->vm_file->f_mapping;
6596 if (pte_none(ptent))
6597 pgoff = linear_page_index(vma, addr);
6598 else /* pte_file(ptent) is true */
6599 pgoff = pte_to_pgoff(ptent);
6601 /* page is moved even if it's not RSS of this task(page-faulted). */
6603 /* shmem/tmpfs may report page out on swap: account for that too. */
6604 if (shmem_mapping(mapping)) {
6605 page = find_get_entry(mapping, pgoff);
6606 if (radix_tree_exceptional_entry(page)) {
6607 swp_entry_t swp = radix_to_swp_entry(page);
6608 if (do_swap_account)
6610 page = find_get_page(swap_address_space(swp), swp.val);
6613 page = find_get_page(mapping, pgoff);
6615 page = find_get_page(mapping, pgoff);
6620 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
6621 unsigned long addr, pte_t ptent, union mc_target *target)
6623 struct page *page = NULL;
6624 struct page_cgroup *pc;
6625 enum mc_target_type ret = MC_TARGET_NONE;
6626 swp_entry_t ent = { .val = 0 };
6628 if (pte_present(ptent))
6629 page = mc_handle_present_pte(vma, addr, ptent);
6630 else if (is_swap_pte(ptent))
6631 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
6632 else if (pte_none(ptent) || pte_file(ptent))
6633 page = mc_handle_file_pte(vma, addr, ptent, &ent);
6635 if (!page && !ent.val)
6638 pc = lookup_page_cgroup(page);
6640 * Do only loose check w/o page_cgroup lock.
6641 * mem_cgroup_move_account() checks the pc is valid or not under
6644 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6645 ret = MC_TARGET_PAGE;
6647 target->page = page;
6649 if (!ret || !target)
6652 /* There is a swap entry and a page doesn't exist or isn't charged */
6653 if (ent.val && !ret &&
6654 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
6655 ret = MC_TARGET_SWAP;
6662 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6664 * We don't consider swapping or file mapped pages because THP does not
6665 * support them for now.
6666 * Caller should make sure that pmd_trans_huge(pmd) is true.
6668 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6669 unsigned long addr, pmd_t pmd, union mc_target *target)
6671 struct page *page = NULL;
6672 struct page_cgroup *pc;
6673 enum mc_target_type ret = MC_TARGET_NONE;
6675 page = pmd_page(pmd);
6676 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
6679 pc = lookup_page_cgroup(page);
6680 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6681 ret = MC_TARGET_PAGE;
6684 target->page = page;
6690 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6691 unsigned long addr, pmd_t pmd, union mc_target *target)
6693 return MC_TARGET_NONE;
6697 static int mem_cgroup_count_precharge_pte(pte_t *pte,
6698 unsigned long addr, unsigned long end,
6699 struct mm_walk *walk)
6701 if (get_mctgt_type(walk->vma, addr, *pte, NULL))
6702 mc.precharge++; /* increment precharge temporarily */
6706 static int mem_cgroup_count_precharge_pmd(pmd_t *pmd,
6707 unsigned long addr, unsigned long end,
6708 struct mm_walk *walk)
6710 struct vm_area_struct *vma = walk->vma;
6713 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
6714 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
6715 mc.precharge += HPAGE_PMD_NR;
6717 /* don't call mem_cgroup_count_precharge_pte() */
6723 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
6725 unsigned long precharge;
6726 struct vm_area_struct *vma;
6728 struct mm_walk mem_cgroup_count_precharge_walk = {
6729 .pmd_entry = mem_cgroup_count_precharge_pmd,
6730 .pte_entry = mem_cgroup_count_precharge_pte,
6733 down_read(&mm->mmap_sem);
6734 for (vma = mm->mmap; vma; vma = vma->vm_next)
6735 walk_page_vma(vma, &mem_cgroup_count_precharge_walk);
6736 up_read(&mm->mmap_sem);
6738 precharge = mc.precharge;
6744 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
6746 unsigned long precharge = mem_cgroup_count_precharge(mm);
6748 VM_BUG_ON(mc.moving_task);
6749 mc.moving_task = current;
6750 return mem_cgroup_do_precharge(precharge);
6753 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6754 static void __mem_cgroup_clear_mc(void)
6756 struct mem_cgroup *from = mc.from;
6757 struct mem_cgroup *to = mc.to;
6760 /* we must uncharge all the leftover precharges from mc.to */
6762 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
6766 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6767 * we must uncharge here.
6769 if (mc.moved_charge) {
6770 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
6771 mc.moved_charge = 0;
6773 /* we must fixup refcnts and charges */
6774 if (mc.moved_swap) {
6775 /* uncharge swap account from the old cgroup */
6776 if (!mem_cgroup_is_root(mc.from))
6777 res_counter_uncharge(&mc.from->memsw,
6778 PAGE_SIZE * mc.moved_swap);
6780 for (i = 0; i < mc.moved_swap; i++)
6781 css_put(&mc.from->css);
6783 if (!mem_cgroup_is_root(mc.to)) {
6785 * we charged both to->res and to->memsw, so we should
6788 res_counter_uncharge(&mc.to->res,
6789 PAGE_SIZE * mc.moved_swap);
6791 /* we've already done css_get(mc.to) */
6794 memcg_oom_recover(from);
6795 memcg_oom_recover(to);
6796 wake_up_all(&mc.waitq);
6799 static void mem_cgroup_clear_mc(void)
6801 struct mem_cgroup *from = mc.from;
6804 * we must clear moving_task before waking up waiters at the end of
6807 mc.moving_task = NULL;
6808 __mem_cgroup_clear_mc();
6809 spin_lock(&mc.lock);
6812 spin_unlock(&mc.lock);
6813 mem_cgroup_end_move(from);
6816 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
6817 struct cgroup_taskset *tset)
6819 struct task_struct *p = cgroup_taskset_first(tset);
6821 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6822 unsigned long move_charge_at_immigrate;
6825 * We are now commited to this value whatever it is. Changes in this
6826 * tunable will only affect upcoming migrations, not the current one.
6827 * So we need to save it, and keep it going.
6829 move_charge_at_immigrate = memcg->move_charge_at_immigrate;
6830 if (move_charge_at_immigrate) {
6831 struct mm_struct *mm;
6832 struct mem_cgroup *from = mem_cgroup_from_task(p);
6834 VM_BUG_ON(from == memcg);
6836 mm = get_task_mm(p);
6839 /* We move charges only when we move a owner of the mm */
6840 if (mm->owner == p) {
6843 VM_BUG_ON(mc.precharge);
6844 VM_BUG_ON(mc.moved_charge);
6845 VM_BUG_ON(mc.moved_swap);
6846 mem_cgroup_start_move(from);
6847 spin_lock(&mc.lock);
6850 mc.immigrate_flags = move_charge_at_immigrate;
6851 spin_unlock(&mc.lock);
6852 /* We set mc.moving_task later */
6854 ret = mem_cgroup_precharge_mc(mm);
6856 mem_cgroup_clear_mc();
6863 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
6864 struct cgroup_taskset *tset)
6866 mem_cgroup_clear_mc();
6869 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6870 unsigned long addr, unsigned long end,
6871 struct mm_walk *walk)
6874 struct vm_area_struct *vma = walk->vma;
6877 enum mc_target_type target_type;
6878 union mc_target target;
6880 struct page_cgroup *pc;
6883 * We don't take compound_lock() here but no race with splitting thp
6885 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6886 * under splitting, which means there's no concurrent thp split,
6887 * - if another thread runs into split_huge_page() just after we
6888 * entered this if-block, the thread must wait for page table lock
6889 * to be unlocked in __split_huge_page_splitting(), where the main
6890 * part of thp split is not executed yet.
6892 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
6893 if (mc.precharge < HPAGE_PMD_NR) {
6897 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6898 if (target_type == MC_TARGET_PAGE) {
6900 if (!isolate_lru_page(page)) {
6901 pc = lookup_page_cgroup(page);
6902 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
6903 pc, mc.from, mc.to)) {
6904 mc.precharge -= HPAGE_PMD_NR;
6905 mc.moved_charge += HPAGE_PMD_NR;
6907 putback_lru_page(page);
6915 if (pmd_trans_unstable(pmd))
6918 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6919 for (; addr != end; addr += PAGE_SIZE) {
6920 pte_t ptent = *(pte++);
6926 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6927 case MC_TARGET_PAGE:
6929 if (isolate_lru_page(page))
6931 pc = lookup_page_cgroup(page);
6932 if (!mem_cgroup_move_account(page, 1, pc,
6935 /* we uncharge from mc.from later. */
6938 putback_lru_page(page);
6939 put: /* get_mctgt_type() gets the page */
6942 case MC_TARGET_SWAP:
6944 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6946 /* we fixup refcnts and charges later. */
6954 pte_unmap_unlock(pte - 1, ptl);
6959 * We have consumed all precharges we got in can_attach().
6960 * We try charge one by one, but don't do any additional
6961 * charges to mc.to if we have failed in charge once in attach()
6964 ret = mem_cgroup_do_precharge(1);
6972 static void mem_cgroup_move_charge(struct mm_struct *mm)
6974 struct vm_area_struct *vma;
6975 struct mm_walk mem_cgroup_move_charge_walk = {
6976 .pmd_entry = mem_cgroup_move_charge_pte_range,
6980 lru_add_drain_all();
6982 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
6984 * Someone who are holding the mmap_sem might be waiting in
6985 * waitq. So we cancel all extra charges, wake up all waiters,
6986 * and retry. Because we cancel precharges, we might not be able
6987 * to move enough charges, but moving charge is a best-effort
6988 * feature anyway, so it wouldn't be a big problem.
6990 __mem_cgroup_clear_mc();
6994 for (vma = mm->mmap; vma; vma = vma->vm_next)
6995 walk_page_vma(vma, &mem_cgroup_move_charge_walk);
6996 up_read(&mm->mmap_sem);
6999 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
7000 struct cgroup_taskset *tset)
7002 struct task_struct *p = cgroup_taskset_first(tset);
7003 struct mm_struct *mm = get_task_mm(p);
7007 mem_cgroup_move_charge(mm);
7011 mem_cgroup_clear_mc();
7013 #else /* !CONFIG_MMU */
7014 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
7015 struct cgroup_taskset *tset)
7019 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
7020 struct cgroup_taskset *tset)
7023 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
7024 struct cgroup_taskset *tset)
7030 * Cgroup retains root cgroups across [un]mount cycles making it necessary
7031 * to verify sane_behavior flag on each mount attempt.
7033 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
7036 * use_hierarchy is forced with sane_behavior. cgroup core
7037 * guarantees that @root doesn't have any children, so turning it
7038 * on for the root memcg is enough.
7040 if (cgroup_sane_behavior(root_css->cgroup))
7041 mem_cgroup_from_css(root_css)->use_hierarchy = true;
7044 struct cgroup_subsys memory_cgrp_subsys = {
7045 .css_alloc = mem_cgroup_css_alloc,
7046 .css_online = mem_cgroup_css_online,
7047 .css_offline = mem_cgroup_css_offline,
7048 .css_free = mem_cgroup_css_free,
7049 .can_attach = mem_cgroup_can_attach,
7050 .cancel_attach = mem_cgroup_cancel_attach,
7051 .attach = mem_cgroup_move_task,
7052 .bind = mem_cgroup_bind,
7053 .base_cftypes = mem_cgroup_files,
7057 #ifdef CONFIG_MEMCG_SWAP
7058 static int __init enable_swap_account(char *s)
7060 if (!strcmp(s, "1"))
7061 really_do_swap_account = 1;
7062 else if (!strcmp(s, "0"))
7063 really_do_swap_account = 0;
7066 __setup("swapaccount=", enable_swap_account);
7068 static void __init memsw_file_init(void)
7070 WARN_ON(cgroup_add_cftypes(&memory_cgrp_subsys, memsw_cgroup_files));
7073 static void __init enable_swap_cgroup(void)
7075 if (!mem_cgroup_disabled() && really_do_swap_account) {
7076 do_swap_account = 1;
7082 static void __init enable_swap_cgroup(void)
7088 * subsys_initcall() for memory controller.
7090 * Some parts like hotcpu_notifier() have to be initialized from this context
7091 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
7092 * everything that doesn't depend on a specific mem_cgroup structure should
7093 * be initialized from here.
7095 static int __init mem_cgroup_init(void)
7097 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
7098 enable_swap_cgroup();
7099 mem_cgroup_soft_limit_tree_init();
7103 subsys_initcall(mem_cgroup_init);