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;
295 /* css_online() has been completed */
299 * the counter to account for mem+swap usage.
301 struct res_counter memsw;
304 * the counter to account for kernel memory usage.
306 struct res_counter kmem;
308 * Should the accounting and control be hierarchical, per subtree?
311 unsigned long kmem_account_flags; /* See KMEM_ACCOUNTED_*, below */
315 atomic_t oom_wakeups;
318 /* OOM-Killer disable */
319 int oom_kill_disable;
321 /* set when res.limit == memsw.limit */
322 bool memsw_is_minimum;
324 /* protect arrays of thresholds */
325 struct mutex thresholds_lock;
327 /* thresholds for memory usage. RCU-protected */
328 struct mem_cgroup_thresholds thresholds;
330 /* thresholds for mem+swap usage. RCU-protected */
331 struct mem_cgroup_thresholds memsw_thresholds;
333 /* For oom notifier event fd */
334 struct list_head oom_notify;
337 * Should we move charges of a task when a task is moved into this
338 * mem_cgroup ? And what type of charges should we move ?
340 unsigned long move_charge_at_immigrate;
342 * set > 0 if pages under this cgroup are moving to other cgroup.
344 atomic_t moving_account;
345 /* taken only while moving_account > 0 */
346 spinlock_t move_lock;
350 struct mem_cgroup_stat_cpu __percpu *stat;
352 * used when a cpu is offlined or other synchronizations
353 * See mem_cgroup_read_stat().
355 struct mem_cgroup_stat_cpu nocpu_base;
356 spinlock_t pcp_counter_lock;
359 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
360 struct cg_proto tcp_mem;
362 #if defined(CONFIG_MEMCG_KMEM)
363 /* analogous to slab_common's slab_caches list, but per-memcg;
364 * protected by memcg_slab_mutex */
365 struct list_head memcg_slab_caches;
366 /* Index in the kmem_cache->memcg_params->memcg_caches array */
370 int last_scanned_node;
372 nodemask_t scan_nodes;
373 atomic_t numainfo_events;
374 atomic_t numainfo_updating;
377 /* List of events which userspace want to receive */
378 struct list_head event_list;
379 spinlock_t event_list_lock;
381 struct mem_cgroup_per_node *nodeinfo[0];
382 /* WARNING: nodeinfo must be the last member here */
385 /* internal only representation about the status of kmem accounting. */
387 KMEM_ACCOUNTED_ACTIVE, /* accounted by this cgroup itself */
388 KMEM_ACCOUNTED_DEAD, /* dead memcg with pending kmem charges */
391 #ifdef CONFIG_MEMCG_KMEM
392 static inline void memcg_kmem_set_active(struct mem_cgroup *memcg)
394 set_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
397 static bool memcg_kmem_is_active(struct mem_cgroup *memcg)
399 return test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
402 static void memcg_kmem_mark_dead(struct mem_cgroup *memcg)
405 * Our caller must use css_get() first, because memcg_uncharge_kmem()
406 * will call css_put() if it sees the memcg is dead.
409 if (test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags))
410 set_bit(KMEM_ACCOUNTED_DEAD, &memcg->kmem_account_flags);
413 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup *memcg)
415 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD,
416 &memcg->kmem_account_flags);
420 /* Stuffs for move charges at task migration. */
422 * Types of charges to be moved. "move_charge_at_immitgrate" and
423 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
426 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
427 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
431 /* "mc" and its members are protected by cgroup_mutex */
432 static struct move_charge_struct {
433 spinlock_t lock; /* for from, to */
434 struct mem_cgroup *from;
435 struct mem_cgroup *to;
436 unsigned long immigrate_flags;
437 unsigned long precharge;
438 unsigned long moved_charge;
439 unsigned long moved_swap;
440 struct task_struct *moving_task; /* a task moving charges */
441 wait_queue_head_t waitq; /* a waitq for other context */
443 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
444 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
447 static bool move_anon(void)
449 return test_bit(MOVE_CHARGE_TYPE_ANON, &mc.immigrate_flags);
452 static bool move_file(void)
454 return test_bit(MOVE_CHARGE_TYPE_FILE, &mc.immigrate_flags);
458 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
459 * limit reclaim to prevent infinite loops, if they ever occur.
461 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
462 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
465 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
466 MEM_CGROUP_CHARGE_TYPE_ANON,
467 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
468 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
472 /* for encoding cft->private value on file */
480 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
481 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
482 #define MEMFILE_ATTR(val) ((val) & 0xffff)
483 /* Used for OOM nofiier */
484 #define OOM_CONTROL (0)
487 * Reclaim flags for mem_cgroup_hierarchical_reclaim
489 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
490 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
491 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
492 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
495 * The memcg_create_mutex will be held whenever a new cgroup is created.
496 * As a consequence, any change that needs to protect against new child cgroups
497 * appearing has to hold it as well.
499 static DEFINE_MUTEX(memcg_create_mutex);
501 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
503 return s ? container_of(s, struct mem_cgroup, css) : NULL;
506 /* Some nice accessors for the vmpressure. */
507 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
510 memcg = root_mem_cgroup;
511 return &memcg->vmpressure;
514 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
516 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
519 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
521 return (memcg == root_mem_cgroup);
525 * We restrict the id in the range of [1, 65535], so it can fit into
528 #define MEM_CGROUP_ID_MAX USHRT_MAX
530 static inline unsigned short mem_cgroup_id(struct mem_cgroup *memcg)
532 return memcg->css.id;
535 static inline struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
537 struct cgroup_subsys_state *css;
539 css = css_from_id(id, &memory_cgrp_subsys);
540 return mem_cgroup_from_css(css);
543 /* Writing them here to avoid exposing memcg's inner layout */
544 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
546 void sock_update_memcg(struct sock *sk)
548 if (mem_cgroup_sockets_enabled) {
549 struct mem_cgroup *memcg;
550 struct cg_proto *cg_proto;
552 BUG_ON(!sk->sk_prot->proto_cgroup);
554 /* Socket cloning can throw us here with sk_cgrp already
555 * filled. It won't however, necessarily happen from
556 * process context. So the test for root memcg given
557 * the current task's memcg won't help us in this case.
559 * Respecting the original socket's memcg is a better
560 * decision in this case.
563 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
564 css_get(&sk->sk_cgrp->memcg->css);
569 memcg = mem_cgroup_from_task(current);
570 cg_proto = sk->sk_prot->proto_cgroup(memcg);
571 if (!mem_cgroup_is_root(memcg) &&
572 memcg_proto_active(cg_proto) &&
573 css_tryget_online(&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 memcg_free_cache_id(int id);
654 static void disarm_kmem_keys(struct mem_cgroup *memcg)
656 if (memcg_kmem_is_active(memcg)) {
657 static_key_slow_dec(&memcg_kmem_enabled_key);
658 memcg_free_cache_id(memcg->kmemcg_id);
661 * This check can't live in kmem destruction function,
662 * since the charges will outlive the cgroup
664 WARN_ON(res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0);
667 static void disarm_kmem_keys(struct mem_cgroup *memcg)
670 #endif /* CONFIG_MEMCG_KMEM */
672 static void disarm_static_keys(struct mem_cgroup *memcg)
674 disarm_sock_keys(memcg);
675 disarm_kmem_keys(memcg);
678 static void drain_all_stock_async(struct mem_cgroup *memcg);
680 static struct mem_cgroup_per_zone *
681 mem_cgroup_zone_zoneinfo(struct mem_cgroup *memcg, struct zone *zone)
683 int nid = zone_to_nid(zone);
684 int zid = zone_idx(zone);
686 return &memcg->nodeinfo[nid]->zoneinfo[zid];
689 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
694 static struct mem_cgroup_per_zone *
695 mem_cgroup_page_zoneinfo(struct mem_cgroup *memcg, struct page *page)
697 int nid = page_to_nid(page);
698 int zid = page_zonenum(page);
700 return &memcg->nodeinfo[nid]->zoneinfo[zid];
703 static struct mem_cgroup_tree_per_zone *
704 soft_limit_tree_node_zone(int nid, int zid)
706 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
709 static struct mem_cgroup_tree_per_zone *
710 soft_limit_tree_from_page(struct page *page)
712 int nid = page_to_nid(page);
713 int zid = page_zonenum(page);
715 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
718 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone *mz,
719 struct mem_cgroup_tree_per_zone *mctz,
720 unsigned long long new_usage_in_excess)
722 struct rb_node **p = &mctz->rb_root.rb_node;
723 struct rb_node *parent = NULL;
724 struct mem_cgroup_per_zone *mz_node;
729 mz->usage_in_excess = new_usage_in_excess;
730 if (!mz->usage_in_excess)
734 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
736 if (mz->usage_in_excess < mz_node->usage_in_excess)
739 * We can't avoid mem cgroups that are over their soft
740 * limit by the same amount
742 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
745 rb_link_node(&mz->tree_node, parent, p);
746 rb_insert_color(&mz->tree_node, &mctz->rb_root);
750 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
751 struct mem_cgroup_tree_per_zone *mctz)
755 rb_erase(&mz->tree_node, &mctz->rb_root);
759 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
760 struct mem_cgroup_tree_per_zone *mctz)
764 spin_lock_irqsave(&mctz->lock, flags);
765 __mem_cgroup_remove_exceeded(mz, mctz);
766 spin_unlock_irqrestore(&mctz->lock, flags);
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;
776 mctz = soft_limit_tree_from_page(page);
778 * Necessary to update all ancestors when hierarchy is used.
779 * because their event counter is not touched.
781 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
782 mz = mem_cgroup_page_zoneinfo(memcg, page);
783 excess = res_counter_soft_limit_excess(&memcg->res);
785 * We have to update the tree if mz is on RB-tree or
786 * mem is over its softlimit.
788 if (excess || mz->on_tree) {
791 spin_lock_irqsave(&mctz->lock, flags);
792 /* if on-tree, remove it */
794 __mem_cgroup_remove_exceeded(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(mz, mctz, excess);
800 spin_unlock_irqrestore(&mctz->lock, flags);
805 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
807 struct mem_cgroup_tree_per_zone *mctz;
808 struct mem_cgroup_per_zone *mz;
812 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
813 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
814 mctz = soft_limit_tree_node_zone(nid, zid);
815 mem_cgroup_remove_exceeded(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, mctz);
839 if (!res_counter_soft_limit_excess(&mz->memcg->res) ||
840 !css_tryget_online(&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_irq(&mctz->lock);
852 mz = __mem_cgroup_largest_soft_limit_node(mctz);
853 spin_unlock_irq(&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 unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
895 enum mem_cgroup_events_index idx)
897 unsigned long val = 0;
901 for_each_online_cpu(cpu)
902 val += per_cpu(memcg->stat->events[idx], cpu);
903 #ifdef CONFIG_HOTPLUG_CPU
904 spin_lock(&memcg->pcp_counter_lock);
905 val += memcg->nocpu_base.events[idx];
906 spin_unlock(&memcg->pcp_counter_lock);
912 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
917 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
918 * counted as CACHE even if it's on ANON LRU.
921 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
924 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
927 if (PageTransHuge(page))
928 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
931 /* pagein of a big page is an event. So, ignore page size */
933 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
935 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
936 nr_pages = -nr_pages; /* for event */
939 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
942 unsigned long mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
944 struct mem_cgroup_per_zone *mz;
946 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
947 return mz->lru_size[lru];
950 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
952 unsigned int lru_mask)
954 unsigned long nr = 0;
957 VM_BUG_ON((unsigned)nid >= nr_node_ids);
959 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
960 struct mem_cgroup_per_zone *mz;
964 if (!(BIT(lru) & lru_mask))
966 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
967 nr += mz->lru_size[lru];
973 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
974 unsigned int lru_mask)
976 unsigned long nr = 0;
979 for_each_node_state(nid, N_MEMORY)
980 nr += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
984 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
985 enum mem_cgroup_events_target target)
987 unsigned long val, next;
989 val = __this_cpu_read(memcg->stat->nr_page_events);
990 next = __this_cpu_read(memcg->stat->targets[target]);
991 /* from time_after() in jiffies.h */
992 if ((long)next - (long)val < 0) {
994 case MEM_CGROUP_TARGET_THRESH:
995 next = val + THRESHOLDS_EVENTS_TARGET;
997 case MEM_CGROUP_TARGET_SOFTLIMIT:
998 next = val + SOFTLIMIT_EVENTS_TARGET;
1000 case MEM_CGROUP_TARGET_NUMAINFO:
1001 next = val + NUMAINFO_EVENTS_TARGET;
1006 __this_cpu_write(memcg->stat->targets[target], next);
1013 * Check events in order.
1016 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
1018 /* threshold event is triggered in finer grain than soft limit */
1019 if (unlikely(mem_cgroup_event_ratelimit(memcg,
1020 MEM_CGROUP_TARGET_THRESH))) {
1022 bool do_numainfo __maybe_unused;
1024 do_softlimit = mem_cgroup_event_ratelimit(memcg,
1025 MEM_CGROUP_TARGET_SOFTLIMIT);
1026 #if MAX_NUMNODES > 1
1027 do_numainfo = mem_cgroup_event_ratelimit(memcg,
1028 MEM_CGROUP_TARGET_NUMAINFO);
1030 mem_cgroup_threshold(memcg);
1031 if (unlikely(do_softlimit))
1032 mem_cgroup_update_tree(memcg, page);
1033 #if MAX_NUMNODES > 1
1034 if (unlikely(do_numainfo))
1035 atomic_inc(&memcg->numainfo_events);
1040 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1043 * mm_update_next_owner() may clear mm->owner to NULL
1044 * if it races with swapoff, page migration, etc.
1045 * So this can be called with p == NULL.
1050 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
1053 static struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
1055 struct mem_cgroup *memcg = NULL;
1060 * Page cache insertions can happen withou an
1061 * actual mm context, e.g. during disk probing
1062 * on boot, loopback IO, acct() writes etc.
1065 memcg = root_mem_cgroup;
1067 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1068 if (unlikely(!memcg))
1069 memcg = root_mem_cgroup;
1071 } while (!css_tryget_online(&memcg->css));
1077 * Returns a next (in a pre-order walk) alive memcg (with elevated css
1078 * ref. count) or NULL if the whole root's subtree has been visited.
1080 * helper function to be used by mem_cgroup_iter
1082 static struct mem_cgroup *__mem_cgroup_iter_next(struct mem_cgroup *root,
1083 struct mem_cgroup *last_visited)
1085 struct cgroup_subsys_state *prev_css, *next_css;
1087 prev_css = last_visited ? &last_visited->css : NULL;
1089 next_css = css_next_descendant_pre(prev_css, &root->css);
1092 * Even if we found a group we have to make sure it is
1093 * alive. css && !memcg means that the groups should be
1094 * skipped and we should continue the tree walk.
1095 * last_visited css is safe to use because it is
1096 * protected by css_get and the tree walk is rcu safe.
1098 * We do not take a reference on the root of the tree walk
1099 * because we might race with the root removal when it would
1100 * be the only node in the iterated hierarchy and mem_cgroup_iter
1101 * would end up in an endless loop because it expects that at
1102 * least one valid node will be returned. Root cannot disappear
1103 * because caller of the iterator should hold it already so
1104 * skipping css reference should be safe.
1107 struct mem_cgroup *memcg = mem_cgroup_from_css(next_css);
1109 if (next_css == &root->css)
1112 if (css_tryget_online(next_css)) {
1114 * Make sure the memcg is initialized:
1115 * mem_cgroup_css_online() orders the the
1116 * initialization against setting the flag.
1118 if (smp_load_acquire(&memcg->initialized))
1123 prev_css = next_css;
1130 static void mem_cgroup_iter_invalidate(struct mem_cgroup *root)
1133 * When a group in the hierarchy below root is destroyed, the
1134 * hierarchy iterator can no longer be trusted since it might
1135 * have pointed to the destroyed group. Invalidate it.
1137 atomic_inc(&root->dead_count);
1140 static struct mem_cgroup *
1141 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter *iter,
1142 struct mem_cgroup *root,
1145 struct mem_cgroup *position = NULL;
1147 * A cgroup destruction happens in two stages: offlining and
1148 * release. They are separated by a RCU grace period.
1150 * If the iterator is valid, we may still race with an
1151 * offlining. The RCU lock ensures the object won't be
1152 * released, tryget will fail if we lost the race.
1154 *sequence = atomic_read(&root->dead_count);
1155 if (iter->last_dead_count == *sequence) {
1157 position = iter->last_visited;
1160 * We cannot take a reference to root because we might race
1161 * with root removal and returning NULL would end up in
1162 * an endless loop on the iterator user level when root
1163 * would be returned all the time.
1165 if (position && position != root &&
1166 !css_tryget_online(&position->css))
1172 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter *iter,
1173 struct mem_cgroup *last_visited,
1174 struct mem_cgroup *new_position,
1175 struct mem_cgroup *root,
1178 /* root reference counting symmetric to mem_cgroup_iter_load */
1179 if (last_visited && last_visited != root)
1180 css_put(&last_visited->css);
1182 * We store the sequence count from the time @last_visited was
1183 * loaded successfully instead of rereading it here so that we
1184 * don't lose destruction events in between. We could have
1185 * raced with the destruction of @new_position after all.
1187 iter->last_visited = new_position;
1189 iter->last_dead_count = sequence;
1193 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1194 * @root: hierarchy root
1195 * @prev: previously returned memcg, NULL on first invocation
1196 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1198 * Returns references to children of the hierarchy below @root, or
1199 * @root itself, or %NULL after a full round-trip.
1201 * Caller must pass the return value in @prev on subsequent
1202 * invocations for reference counting, or use mem_cgroup_iter_break()
1203 * to cancel a hierarchy walk before the round-trip is complete.
1205 * Reclaimers can specify a zone and a priority level in @reclaim to
1206 * divide up the memcgs in the hierarchy among all concurrent
1207 * reclaimers operating on the same zone and priority.
1209 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1210 struct mem_cgroup *prev,
1211 struct mem_cgroup_reclaim_cookie *reclaim)
1213 struct mem_cgroup *memcg = NULL;
1214 struct mem_cgroup *last_visited = NULL;
1216 if (mem_cgroup_disabled())
1220 root = root_mem_cgroup;
1222 if (prev && !reclaim)
1223 last_visited = prev;
1225 if (!root->use_hierarchy && root != root_mem_cgroup) {
1233 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1234 int uninitialized_var(seq);
1237 struct mem_cgroup_per_zone *mz;
1239 mz = mem_cgroup_zone_zoneinfo(root, reclaim->zone);
1240 iter = &mz->reclaim_iter[reclaim->priority];
1241 if (prev && reclaim->generation != iter->generation) {
1242 iter->last_visited = NULL;
1246 last_visited = mem_cgroup_iter_load(iter, root, &seq);
1249 memcg = __mem_cgroup_iter_next(root, last_visited);
1252 mem_cgroup_iter_update(iter, last_visited, memcg, root,
1257 else if (!prev && memcg)
1258 reclaim->generation = iter->generation;
1267 if (prev && prev != root)
1268 css_put(&prev->css);
1274 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1275 * @root: hierarchy root
1276 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1278 void mem_cgroup_iter_break(struct mem_cgroup *root,
1279 struct mem_cgroup *prev)
1282 root = root_mem_cgroup;
1283 if (prev && prev != root)
1284 css_put(&prev->css);
1288 * Iteration constructs for visiting all cgroups (under a tree). If
1289 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1290 * be used for reference counting.
1292 #define for_each_mem_cgroup_tree(iter, root) \
1293 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1295 iter = mem_cgroup_iter(root, iter, NULL))
1297 #define for_each_mem_cgroup(iter) \
1298 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1300 iter = mem_cgroup_iter(NULL, iter, NULL))
1302 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1304 struct mem_cgroup *memcg;
1307 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1308 if (unlikely(!memcg))
1313 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1316 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1324 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1327 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1328 * @zone: zone of the wanted lruvec
1329 * @memcg: memcg of the wanted lruvec
1331 * Returns the lru list vector holding pages for the given @zone and
1332 * @mem. This can be the global zone lruvec, if the memory controller
1335 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1336 struct mem_cgroup *memcg)
1338 struct mem_cgroup_per_zone *mz;
1339 struct lruvec *lruvec;
1341 if (mem_cgroup_disabled()) {
1342 lruvec = &zone->lruvec;
1346 mz = mem_cgroup_zone_zoneinfo(memcg, zone);
1347 lruvec = &mz->lruvec;
1350 * Since a node can be onlined after the mem_cgroup was created,
1351 * we have to be prepared to initialize lruvec->zone here;
1352 * and if offlined then reonlined, we need to reinitialize it.
1354 if (unlikely(lruvec->zone != zone))
1355 lruvec->zone = zone;
1360 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1362 * @zone: zone of the page
1364 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1366 struct mem_cgroup_per_zone *mz;
1367 struct mem_cgroup *memcg;
1368 struct page_cgroup *pc;
1369 struct lruvec *lruvec;
1371 if (mem_cgroup_disabled()) {
1372 lruvec = &zone->lruvec;
1376 pc = lookup_page_cgroup(page);
1377 memcg = pc->mem_cgroup;
1380 * Surreptitiously switch any uncharged offlist page to root:
1381 * an uncharged page off lru does nothing to secure
1382 * its former mem_cgroup from sudden removal.
1384 * Our caller holds lru_lock, and PageCgroupUsed is updated
1385 * under page_cgroup lock: between them, they make all uses
1386 * of pc->mem_cgroup safe.
1388 if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1389 pc->mem_cgroup = memcg = root_mem_cgroup;
1391 mz = mem_cgroup_page_zoneinfo(memcg, page);
1392 lruvec = &mz->lruvec;
1395 * Since a node can be onlined after the mem_cgroup was created,
1396 * we have to be prepared to initialize lruvec->zone here;
1397 * and if offlined then reonlined, we need to reinitialize it.
1399 if (unlikely(lruvec->zone != zone))
1400 lruvec->zone = zone;
1405 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1406 * @lruvec: mem_cgroup per zone lru vector
1407 * @lru: index of lru list the page is sitting on
1408 * @nr_pages: positive when adding or negative when removing
1410 * This function must be called when a page is added to or removed from an
1413 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1416 struct mem_cgroup_per_zone *mz;
1417 unsigned long *lru_size;
1419 if (mem_cgroup_disabled())
1422 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1423 lru_size = mz->lru_size + lru;
1424 *lru_size += nr_pages;
1425 VM_BUG_ON((long)(*lru_size) < 0);
1429 * Checks whether given mem is same or in the root_mem_cgroup's
1432 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1433 struct mem_cgroup *memcg)
1435 if (root_memcg == memcg)
1437 if (!root_memcg->use_hierarchy || !memcg)
1439 return cgroup_is_descendant(memcg->css.cgroup, root_memcg->css.cgroup);
1442 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1443 struct mem_cgroup *memcg)
1448 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1453 bool task_in_mem_cgroup(struct task_struct *task,
1454 const struct mem_cgroup *memcg)
1456 struct mem_cgroup *curr = NULL;
1457 struct task_struct *p;
1460 p = find_lock_task_mm(task);
1462 curr = get_mem_cgroup_from_mm(p->mm);
1466 * All threads may have already detached their mm's, but the oom
1467 * killer still needs to detect if they have already been oom
1468 * killed to prevent needlessly killing additional tasks.
1471 curr = mem_cgroup_from_task(task);
1473 css_get(&curr->css);
1477 * We should check use_hierarchy of "memcg" not "curr". Because checking
1478 * use_hierarchy of "curr" here make this function true if hierarchy is
1479 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1480 * hierarchy(even if use_hierarchy is disabled in "memcg").
1482 ret = mem_cgroup_same_or_subtree(memcg, curr);
1483 css_put(&curr->css);
1487 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1489 unsigned long inactive_ratio;
1490 unsigned long inactive;
1491 unsigned long active;
1494 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1495 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1497 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1499 inactive_ratio = int_sqrt(10 * gb);
1503 return inactive * inactive_ratio < active;
1506 #define mem_cgroup_from_res_counter(counter, member) \
1507 container_of(counter, struct mem_cgroup, member)
1510 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1511 * @memcg: the memory cgroup
1513 * Returns the maximum amount of memory @mem can be charged with, in
1516 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1518 unsigned long long margin;
1520 margin = res_counter_margin(&memcg->res);
1521 if (do_swap_account)
1522 margin = min(margin, res_counter_margin(&memcg->memsw));
1523 return margin >> PAGE_SHIFT;
1526 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1529 if (mem_cgroup_disabled() || !memcg->css.parent)
1530 return vm_swappiness;
1532 return memcg->swappiness;
1536 * memcg->moving_account is used for checking possibility that some thread is
1537 * calling move_account(). When a thread on CPU-A starts moving pages under
1538 * a memcg, other threads should check memcg->moving_account under
1539 * rcu_read_lock(), like this:
1543 * memcg->moving_account+1 if (memcg->mocing_account)
1545 * synchronize_rcu() update something.
1550 /* for quick checking without looking up memcg */
1551 atomic_t memcg_moving __read_mostly;
1553 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1555 atomic_inc(&memcg_moving);
1556 atomic_inc(&memcg->moving_account);
1560 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1563 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1564 * We check NULL in callee rather than caller.
1567 atomic_dec(&memcg_moving);
1568 atomic_dec(&memcg->moving_account);
1573 * A routine for checking "mem" is under move_account() or not.
1575 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1576 * moving cgroups. This is for waiting at high-memory pressure
1579 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1581 struct mem_cgroup *from;
1582 struct mem_cgroup *to;
1585 * Unlike task_move routines, we access mc.to, mc.from not under
1586 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1588 spin_lock(&mc.lock);
1594 ret = mem_cgroup_same_or_subtree(memcg, from)
1595 || mem_cgroup_same_or_subtree(memcg, to);
1597 spin_unlock(&mc.lock);
1601 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1603 if (mc.moving_task && current != mc.moving_task) {
1604 if (mem_cgroup_under_move(memcg)) {
1606 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1607 /* moving charge context might have finished. */
1610 finish_wait(&mc.waitq, &wait);
1618 * Take this lock when
1619 * - a code tries to modify page's memcg while it's USED.
1620 * - a code tries to modify page state accounting in a memcg.
1622 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1623 unsigned long *flags)
1625 spin_lock_irqsave(&memcg->move_lock, *flags);
1628 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1629 unsigned long *flags)
1631 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1634 #define K(x) ((x) << (PAGE_SHIFT-10))
1636 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1637 * @memcg: The memory cgroup that went over limit
1638 * @p: Task that is going to be killed
1640 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1643 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1645 /* oom_info_lock ensures that parallel ooms do not interleave */
1646 static DEFINE_MUTEX(oom_info_lock);
1647 struct mem_cgroup *iter;
1653 mutex_lock(&oom_info_lock);
1656 pr_info("Task in ");
1657 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1658 pr_info(" killed as a result of limit of ");
1659 pr_cont_cgroup_path(memcg->css.cgroup);
1664 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1665 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1666 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1667 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1668 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1669 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1670 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1671 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1672 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1673 res_counter_read_u64(&memcg->kmem, RES_USAGE) >> 10,
1674 res_counter_read_u64(&memcg->kmem, RES_LIMIT) >> 10,
1675 res_counter_read_u64(&memcg->kmem, RES_FAILCNT));
1677 for_each_mem_cgroup_tree(iter, memcg) {
1678 pr_info("Memory cgroup stats for ");
1679 pr_cont_cgroup_path(iter->css.cgroup);
1682 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1683 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1685 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1686 K(mem_cgroup_read_stat(iter, i)));
1689 for (i = 0; i < NR_LRU_LISTS; i++)
1690 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1691 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1695 mutex_unlock(&oom_info_lock);
1699 * This function returns the number of memcg under hierarchy tree. Returns
1700 * 1(self count) if no children.
1702 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1705 struct mem_cgroup *iter;
1707 for_each_mem_cgroup_tree(iter, memcg)
1713 * Return the memory (and swap, if configured) limit for a memcg.
1715 static u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1719 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1722 * Do not consider swap space if we cannot swap due to swappiness
1724 if (mem_cgroup_swappiness(memcg)) {
1727 limit += total_swap_pages << PAGE_SHIFT;
1728 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1731 * If memsw is finite and limits the amount of swap space
1732 * available to this memcg, return that limit.
1734 limit = min(limit, memsw);
1740 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1743 struct mem_cgroup *iter;
1744 unsigned long chosen_points = 0;
1745 unsigned long totalpages;
1746 unsigned int points = 0;
1747 struct task_struct *chosen = NULL;
1750 * If current has a pending SIGKILL or is exiting, then automatically
1751 * select it. The goal is to allow it to allocate so that it may
1752 * quickly exit and free its memory.
1754 if (fatal_signal_pending(current) || current->flags & PF_EXITING) {
1755 set_thread_flag(TIF_MEMDIE);
1759 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1760 totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
1761 for_each_mem_cgroup_tree(iter, memcg) {
1762 struct css_task_iter it;
1763 struct task_struct *task;
1765 css_task_iter_start(&iter->css, &it);
1766 while ((task = css_task_iter_next(&it))) {
1767 switch (oom_scan_process_thread(task, totalpages, NULL,
1769 case OOM_SCAN_SELECT:
1771 put_task_struct(chosen);
1773 chosen_points = ULONG_MAX;
1774 get_task_struct(chosen);
1776 case OOM_SCAN_CONTINUE:
1778 case OOM_SCAN_ABORT:
1779 css_task_iter_end(&it);
1780 mem_cgroup_iter_break(memcg, iter);
1782 put_task_struct(chosen);
1787 points = oom_badness(task, memcg, NULL, totalpages);
1788 if (!points || points < chosen_points)
1790 /* Prefer thread group leaders for display purposes */
1791 if (points == chosen_points &&
1792 thread_group_leader(chosen))
1796 put_task_struct(chosen);
1798 chosen_points = points;
1799 get_task_struct(chosen);
1801 css_task_iter_end(&it);
1806 points = chosen_points * 1000 / totalpages;
1807 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1808 NULL, "Memory cgroup out of memory");
1811 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1813 unsigned long flags)
1815 unsigned long total = 0;
1816 bool noswap = false;
1819 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1821 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1824 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1826 drain_all_stock_async(memcg);
1827 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1829 * Allow limit shrinkers, which are triggered directly
1830 * by userspace, to catch signals and stop reclaim
1831 * after minimal progress, regardless of the margin.
1833 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1835 if (mem_cgroup_margin(memcg))
1838 * If nothing was reclaimed after two attempts, there
1839 * may be no reclaimable pages in this hierarchy.
1848 * test_mem_cgroup_node_reclaimable
1849 * @memcg: the target memcg
1850 * @nid: the node ID to be checked.
1851 * @noswap : specify true here if the user wants flle only information.
1853 * This function returns whether the specified memcg contains any
1854 * reclaimable pages on a node. Returns true if there are any reclaimable
1855 * pages in the node.
1857 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1858 int nid, bool noswap)
1860 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1862 if (noswap || !total_swap_pages)
1864 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1869 #if MAX_NUMNODES > 1
1872 * Always updating the nodemask is not very good - even if we have an empty
1873 * list or the wrong list here, we can start from some node and traverse all
1874 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1877 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1881 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1882 * pagein/pageout changes since the last update.
1884 if (!atomic_read(&memcg->numainfo_events))
1886 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1889 /* make a nodemask where this memcg uses memory from */
1890 memcg->scan_nodes = node_states[N_MEMORY];
1892 for_each_node_mask(nid, node_states[N_MEMORY]) {
1894 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1895 node_clear(nid, memcg->scan_nodes);
1898 atomic_set(&memcg->numainfo_events, 0);
1899 atomic_set(&memcg->numainfo_updating, 0);
1903 * Selecting a node where we start reclaim from. Because what we need is just
1904 * reducing usage counter, start from anywhere is O,K. Considering
1905 * memory reclaim from current node, there are pros. and cons.
1907 * Freeing memory from current node means freeing memory from a node which
1908 * we'll use or we've used. So, it may make LRU bad. And if several threads
1909 * hit limits, it will see a contention on a node. But freeing from remote
1910 * node means more costs for memory reclaim because of memory latency.
1912 * Now, we use round-robin. Better algorithm is welcomed.
1914 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1918 mem_cgroup_may_update_nodemask(memcg);
1919 node = memcg->last_scanned_node;
1921 node = next_node(node, memcg->scan_nodes);
1922 if (node == MAX_NUMNODES)
1923 node = first_node(memcg->scan_nodes);
1925 * We call this when we hit limit, not when pages are added to LRU.
1926 * No LRU may hold pages because all pages are UNEVICTABLE or
1927 * memcg is too small and all pages are not on LRU. In that case,
1928 * we use curret node.
1930 if (unlikely(node == MAX_NUMNODES))
1931 node = numa_node_id();
1933 memcg->last_scanned_node = node;
1938 * Check all nodes whether it contains reclaimable pages or not.
1939 * For quick scan, we make use of scan_nodes. This will allow us to skip
1940 * unused nodes. But scan_nodes is lazily updated and may not cotain
1941 * enough new information. We need to do double check.
1943 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1948 * quick check...making use of scan_node.
1949 * We can skip unused nodes.
1951 if (!nodes_empty(memcg->scan_nodes)) {
1952 for (nid = first_node(memcg->scan_nodes);
1954 nid = next_node(nid, memcg->scan_nodes)) {
1956 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1961 * Check rest of nodes.
1963 for_each_node_state(nid, N_MEMORY) {
1964 if (node_isset(nid, memcg->scan_nodes))
1966 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1973 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1978 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1980 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
1984 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1987 unsigned long *total_scanned)
1989 struct mem_cgroup *victim = NULL;
1992 unsigned long excess;
1993 unsigned long nr_scanned;
1994 struct mem_cgroup_reclaim_cookie reclaim = {
1999 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
2002 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
2007 * If we have not been able to reclaim
2008 * anything, it might because there are
2009 * no reclaimable pages under this hierarchy
2014 * We want to do more targeted reclaim.
2015 * excess >> 2 is not to excessive so as to
2016 * reclaim too much, nor too less that we keep
2017 * coming back to reclaim from this cgroup
2019 if (total >= (excess >> 2) ||
2020 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
2025 if (!mem_cgroup_reclaimable(victim, false))
2027 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
2029 *total_scanned += nr_scanned;
2030 if (!res_counter_soft_limit_excess(&root_memcg->res))
2033 mem_cgroup_iter_break(root_memcg, victim);
2037 #ifdef CONFIG_LOCKDEP
2038 static struct lockdep_map memcg_oom_lock_dep_map = {
2039 .name = "memcg_oom_lock",
2043 static DEFINE_SPINLOCK(memcg_oom_lock);
2046 * Check OOM-Killer is already running under our hierarchy.
2047 * If someone is running, return false.
2049 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
2051 struct mem_cgroup *iter, *failed = NULL;
2053 spin_lock(&memcg_oom_lock);
2055 for_each_mem_cgroup_tree(iter, memcg) {
2056 if (iter->oom_lock) {
2058 * this subtree of our hierarchy is already locked
2059 * so we cannot give a lock.
2062 mem_cgroup_iter_break(memcg, iter);
2065 iter->oom_lock = true;
2070 * OK, we failed to lock the whole subtree so we have
2071 * to clean up what we set up to the failing subtree
2073 for_each_mem_cgroup_tree(iter, memcg) {
2074 if (iter == failed) {
2075 mem_cgroup_iter_break(memcg, iter);
2078 iter->oom_lock = false;
2081 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
2083 spin_unlock(&memcg_oom_lock);
2088 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
2090 struct mem_cgroup *iter;
2092 spin_lock(&memcg_oom_lock);
2093 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
2094 for_each_mem_cgroup_tree(iter, memcg)
2095 iter->oom_lock = false;
2096 spin_unlock(&memcg_oom_lock);
2099 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
2101 struct mem_cgroup *iter;
2103 for_each_mem_cgroup_tree(iter, memcg)
2104 atomic_inc(&iter->under_oom);
2107 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
2109 struct mem_cgroup *iter;
2112 * When a new child is created while the hierarchy is under oom,
2113 * mem_cgroup_oom_lock() may not be called. We have to use
2114 * atomic_add_unless() here.
2116 for_each_mem_cgroup_tree(iter, memcg)
2117 atomic_add_unless(&iter->under_oom, -1, 0);
2120 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
2122 struct oom_wait_info {
2123 struct mem_cgroup *memcg;
2127 static int memcg_oom_wake_function(wait_queue_t *wait,
2128 unsigned mode, int sync, void *arg)
2130 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
2131 struct mem_cgroup *oom_wait_memcg;
2132 struct oom_wait_info *oom_wait_info;
2134 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
2135 oom_wait_memcg = oom_wait_info->memcg;
2138 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2139 * Then we can use css_is_ancestor without taking care of RCU.
2141 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
2142 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
2144 return autoremove_wake_function(wait, mode, sync, arg);
2147 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
2149 atomic_inc(&memcg->oom_wakeups);
2150 /* for filtering, pass "memcg" as argument. */
2151 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
2154 static void memcg_oom_recover(struct mem_cgroup *memcg)
2156 if (memcg && atomic_read(&memcg->under_oom))
2157 memcg_wakeup_oom(memcg);
2160 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
2162 if (!current->memcg_oom.may_oom)
2165 * We are in the middle of the charge context here, so we
2166 * don't want to block when potentially sitting on a callstack
2167 * that holds all kinds of filesystem and mm locks.
2169 * Also, the caller may handle a failed allocation gracefully
2170 * (like optional page cache readahead) and so an OOM killer
2171 * invocation might not even be necessary.
2173 * That's why we don't do anything here except remember the
2174 * OOM context and then deal with it at the end of the page
2175 * fault when the stack is unwound, the locks are released,
2176 * and when we know whether the fault was overall successful.
2178 css_get(&memcg->css);
2179 current->memcg_oom.memcg = memcg;
2180 current->memcg_oom.gfp_mask = mask;
2181 current->memcg_oom.order = order;
2185 * mem_cgroup_oom_synchronize - complete memcg OOM handling
2186 * @handle: actually kill/wait or just clean up the OOM state
2188 * This has to be called at the end of a page fault if the memcg OOM
2189 * handler was enabled.
2191 * Memcg supports userspace OOM handling where failed allocations must
2192 * sleep on a waitqueue until the userspace task resolves the
2193 * situation. Sleeping directly in the charge context with all kinds
2194 * of locks held is not a good idea, instead we remember an OOM state
2195 * in the task and mem_cgroup_oom_synchronize() has to be called at
2196 * the end of the page fault to complete the OOM handling.
2198 * Returns %true if an ongoing memcg OOM situation was detected and
2199 * completed, %false otherwise.
2201 bool mem_cgroup_oom_synchronize(bool handle)
2203 struct mem_cgroup *memcg = current->memcg_oom.memcg;
2204 struct oom_wait_info owait;
2207 /* OOM is global, do not handle */
2214 owait.memcg = memcg;
2215 owait.wait.flags = 0;
2216 owait.wait.func = memcg_oom_wake_function;
2217 owait.wait.private = current;
2218 INIT_LIST_HEAD(&owait.wait.task_list);
2220 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2221 mem_cgroup_mark_under_oom(memcg);
2223 locked = mem_cgroup_oom_trylock(memcg);
2226 mem_cgroup_oom_notify(memcg);
2228 if (locked && !memcg->oom_kill_disable) {
2229 mem_cgroup_unmark_under_oom(memcg);
2230 finish_wait(&memcg_oom_waitq, &owait.wait);
2231 mem_cgroup_out_of_memory(memcg, current->memcg_oom.gfp_mask,
2232 current->memcg_oom.order);
2235 mem_cgroup_unmark_under_oom(memcg);
2236 finish_wait(&memcg_oom_waitq, &owait.wait);
2240 mem_cgroup_oom_unlock(memcg);
2242 * There is no guarantee that an OOM-lock contender
2243 * sees the wakeups triggered by the OOM kill
2244 * uncharges. Wake any sleepers explicitely.
2246 memcg_oom_recover(memcg);
2249 current->memcg_oom.memcg = NULL;
2250 css_put(&memcg->css);
2255 * Used to update mapped file or writeback or other statistics.
2257 * Notes: Race condition
2259 * Charging occurs during page instantiation, while the page is
2260 * unmapped and locked in page migration, or while the page table is
2261 * locked in THP migration. No race is possible.
2263 * Uncharge happens to pages with zero references, no race possible.
2265 * Charge moving between groups is protected by checking mm->moving
2266 * account and taking the move_lock in the slowpath.
2269 void __mem_cgroup_begin_update_page_stat(struct page *page,
2270 bool *locked, unsigned long *flags)
2272 struct mem_cgroup *memcg;
2273 struct page_cgroup *pc;
2275 pc = lookup_page_cgroup(page);
2277 memcg = pc->mem_cgroup;
2278 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2281 * If this memory cgroup is not under account moving, we don't
2282 * need to take move_lock_mem_cgroup(). Because we already hold
2283 * rcu_read_lock(), any calls to move_account will be delayed until
2284 * rcu_read_unlock().
2286 VM_BUG_ON(!rcu_read_lock_held());
2287 if (atomic_read(&memcg->moving_account) <= 0)
2290 move_lock_mem_cgroup(memcg, flags);
2291 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2292 move_unlock_mem_cgroup(memcg, flags);
2298 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
2300 struct page_cgroup *pc = lookup_page_cgroup(page);
2303 * It's guaranteed that pc->mem_cgroup never changes while
2304 * lock is held because a routine modifies pc->mem_cgroup
2305 * should take move_lock_mem_cgroup().
2307 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
2310 void mem_cgroup_update_page_stat(struct page *page,
2311 enum mem_cgroup_stat_index idx, int val)
2313 struct mem_cgroup *memcg;
2314 struct page_cgroup *pc = lookup_page_cgroup(page);
2315 unsigned long uninitialized_var(flags);
2317 if (mem_cgroup_disabled())
2320 VM_BUG_ON(!rcu_read_lock_held());
2321 memcg = pc->mem_cgroup;
2322 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2325 this_cpu_add(memcg->stat->count[idx], val);
2329 * size of first charge trial. "32" comes from vmscan.c's magic value.
2330 * TODO: maybe necessary to use big numbers in big irons.
2332 #define CHARGE_BATCH 32U
2333 struct memcg_stock_pcp {
2334 struct mem_cgroup *cached; /* this never be root cgroup */
2335 unsigned int nr_pages;
2336 struct work_struct work;
2337 unsigned long flags;
2338 #define FLUSHING_CACHED_CHARGE 0
2340 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2341 static DEFINE_MUTEX(percpu_charge_mutex);
2344 * consume_stock: Try to consume stocked charge on this cpu.
2345 * @memcg: memcg to consume from.
2346 * @nr_pages: how many pages to charge.
2348 * The charges will only happen if @memcg matches the current cpu's memcg
2349 * stock, and at least @nr_pages are available in that stock. Failure to
2350 * service an allocation will refill the stock.
2352 * returns true if successful, false otherwise.
2354 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2356 struct memcg_stock_pcp *stock;
2359 if (nr_pages > CHARGE_BATCH)
2362 stock = &get_cpu_var(memcg_stock);
2363 if (memcg == stock->cached && stock->nr_pages >= nr_pages)
2364 stock->nr_pages -= nr_pages;
2365 else /* need to call res_counter_charge */
2367 put_cpu_var(memcg_stock);
2372 * Returns stocks cached in percpu to res_counter and reset cached information.
2374 static void drain_stock(struct memcg_stock_pcp *stock)
2376 struct mem_cgroup *old = stock->cached;
2378 if (stock->nr_pages) {
2379 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2381 res_counter_uncharge(&old->res, bytes);
2382 if (do_swap_account)
2383 res_counter_uncharge(&old->memsw, bytes);
2384 stock->nr_pages = 0;
2386 stock->cached = NULL;
2390 * This must be called under preempt disabled or must be called by
2391 * a thread which is pinned to local cpu.
2393 static void drain_local_stock(struct work_struct *dummy)
2395 struct memcg_stock_pcp *stock = this_cpu_ptr(&memcg_stock);
2397 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2400 static void __init memcg_stock_init(void)
2404 for_each_possible_cpu(cpu) {
2405 struct memcg_stock_pcp *stock =
2406 &per_cpu(memcg_stock, cpu);
2407 INIT_WORK(&stock->work, drain_local_stock);
2412 * Cache charges(val) which is from res_counter, to local per_cpu area.
2413 * This will be consumed by consume_stock() function, later.
2415 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2417 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2419 if (stock->cached != memcg) { /* reset if necessary */
2421 stock->cached = memcg;
2423 stock->nr_pages += nr_pages;
2424 put_cpu_var(memcg_stock);
2428 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2429 * of the hierarchy under it. sync flag says whether we should block
2430 * until the work is done.
2432 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2436 /* Notify other cpus that system-wide "drain" is running */
2439 for_each_online_cpu(cpu) {
2440 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2441 struct mem_cgroup *memcg;
2443 memcg = stock->cached;
2444 if (!memcg || !stock->nr_pages)
2446 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2448 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2450 drain_local_stock(&stock->work);
2452 schedule_work_on(cpu, &stock->work);
2460 for_each_online_cpu(cpu) {
2461 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2462 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2463 flush_work(&stock->work);
2470 * Tries to drain stocked charges in other cpus. This function is asynchronous
2471 * and just put a work per cpu for draining localy on each cpu. Caller can
2472 * expects some charges will be back to res_counter later but cannot wait for
2475 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2478 * If someone calls draining, avoid adding more kworker runs.
2480 if (!mutex_trylock(&percpu_charge_mutex))
2482 drain_all_stock(root_memcg, false);
2483 mutex_unlock(&percpu_charge_mutex);
2486 /* This is a synchronous drain interface. */
2487 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2489 /* called when force_empty is called */
2490 mutex_lock(&percpu_charge_mutex);
2491 drain_all_stock(root_memcg, true);
2492 mutex_unlock(&percpu_charge_mutex);
2496 * This function drains percpu counter value from DEAD cpu and
2497 * move it to local cpu. Note that this function can be preempted.
2499 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2503 spin_lock(&memcg->pcp_counter_lock);
2504 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2505 long x = per_cpu(memcg->stat->count[i], cpu);
2507 per_cpu(memcg->stat->count[i], cpu) = 0;
2508 memcg->nocpu_base.count[i] += x;
2510 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2511 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2513 per_cpu(memcg->stat->events[i], cpu) = 0;
2514 memcg->nocpu_base.events[i] += x;
2516 spin_unlock(&memcg->pcp_counter_lock);
2519 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
2520 unsigned long action,
2523 int cpu = (unsigned long)hcpu;
2524 struct memcg_stock_pcp *stock;
2525 struct mem_cgroup *iter;
2527 if (action == CPU_ONLINE)
2530 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2533 for_each_mem_cgroup(iter)
2534 mem_cgroup_drain_pcp_counter(iter, cpu);
2536 stock = &per_cpu(memcg_stock, cpu);
2541 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2542 unsigned int nr_pages)
2544 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2545 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2546 struct mem_cgroup *mem_over_limit;
2547 struct res_counter *fail_res;
2548 unsigned long nr_reclaimed;
2549 unsigned long flags = 0;
2550 unsigned long long size;
2553 if (mem_cgroup_is_root(memcg))
2556 if (consume_stock(memcg, nr_pages))
2559 size = batch * PAGE_SIZE;
2560 if (!res_counter_charge(&memcg->res, size, &fail_res)) {
2561 if (!do_swap_account)
2563 if (!res_counter_charge(&memcg->memsw, size, &fail_res))
2565 res_counter_uncharge(&memcg->res, size);
2566 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2567 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2569 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2571 if (batch > nr_pages) {
2577 * Unlike in global OOM situations, memcg is not in a physical
2578 * memory shortage. Allow dying and OOM-killed tasks to
2579 * bypass the last charges so that they can exit quickly and
2580 * free their memory.
2582 if (unlikely(test_thread_flag(TIF_MEMDIE) ||
2583 fatal_signal_pending(current) ||
2584 current->flags & PF_EXITING))
2587 if (unlikely(task_in_memcg_oom(current)))
2590 if (!(gfp_mask & __GFP_WAIT))
2593 nr_reclaimed = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2595 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2598 if (gfp_mask & __GFP_NORETRY)
2601 * Even though the limit is exceeded at this point, reclaim
2602 * may have been able to free some pages. Retry the charge
2603 * before killing the task.
2605 * Only for regular pages, though: huge pages are rather
2606 * unlikely to succeed so close to the limit, and we fall back
2607 * to regular pages anyway in case of failure.
2609 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2612 * At task move, charge accounts can be doubly counted. So, it's
2613 * better to wait until the end of task_move if something is going on.
2615 if (mem_cgroup_wait_acct_move(mem_over_limit))
2621 if (gfp_mask & __GFP_NOFAIL)
2624 if (fatal_signal_pending(current))
2627 mem_cgroup_oom(mem_over_limit, gfp_mask, get_order(nr_pages));
2629 if (!(gfp_mask & __GFP_NOFAIL))
2635 if (batch > nr_pages)
2636 refill_stock(memcg, batch - nr_pages);
2641 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2643 unsigned long bytes = nr_pages * PAGE_SIZE;
2645 if (mem_cgroup_is_root(memcg))
2648 res_counter_uncharge(&memcg->res, bytes);
2649 if (do_swap_account)
2650 res_counter_uncharge(&memcg->memsw, bytes);
2654 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2655 * This is useful when moving usage to parent cgroup.
2657 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2658 unsigned int nr_pages)
2660 unsigned long bytes = nr_pages * PAGE_SIZE;
2662 if (mem_cgroup_is_root(memcg))
2665 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2666 if (do_swap_account)
2667 res_counter_uncharge_until(&memcg->memsw,
2668 memcg->memsw.parent, bytes);
2672 * A helper function to get mem_cgroup from ID. must be called under
2673 * rcu_read_lock(). The caller is responsible for calling
2674 * css_tryget_online() if the mem_cgroup is used for charging. (dropping
2675 * refcnt from swap can be called against removed memcg.)
2677 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2679 /* ID 0 is unused ID */
2682 return mem_cgroup_from_id(id);
2686 * try_get_mem_cgroup_from_page - look up page's memcg association
2689 * Look up, get a css reference, and return the memcg that owns @page.
2691 * The page must be locked to prevent racing with swap-in and page
2692 * cache charges. If coming from an unlocked page table, the caller
2693 * must ensure the page is on the LRU or this can race with charging.
2695 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2697 struct mem_cgroup *memcg = NULL;
2698 struct page_cgroup *pc;
2702 VM_BUG_ON_PAGE(!PageLocked(page), page);
2704 pc = lookup_page_cgroup(page);
2705 if (PageCgroupUsed(pc)) {
2706 memcg = pc->mem_cgroup;
2707 if (memcg && !css_tryget_online(&memcg->css))
2709 } else if (PageSwapCache(page)) {
2710 ent.val = page_private(page);
2711 id = lookup_swap_cgroup_id(ent);
2713 memcg = mem_cgroup_lookup(id);
2714 if (memcg && !css_tryget_online(&memcg->css))
2721 static void lock_page_lru(struct page *page, int *isolated)
2723 struct zone *zone = page_zone(page);
2725 spin_lock_irq(&zone->lru_lock);
2726 if (PageLRU(page)) {
2727 struct lruvec *lruvec;
2729 lruvec = mem_cgroup_page_lruvec(page, zone);
2731 del_page_from_lru_list(page, lruvec, page_lru(page));
2737 static void unlock_page_lru(struct page *page, int isolated)
2739 struct zone *zone = page_zone(page);
2742 struct lruvec *lruvec;
2744 lruvec = mem_cgroup_page_lruvec(page, zone);
2745 VM_BUG_ON_PAGE(PageLRU(page), page);
2747 add_page_to_lru_list(page, lruvec, page_lru(page));
2749 spin_unlock_irq(&zone->lru_lock);
2752 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2755 struct page_cgroup *pc = lookup_page_cgroup(page);
2758 VM_BUG_ON_PAGE(PageCgroupUsed(pc), page);
2760 * we don't need page_cgroup_lock about tail pages, becase they are not
2761 * accessed by any other context at this point.
2765 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2766 * may already be on some other mem_cgroup's LRU. Take care of it.
2769 lock_page_lru(page, &isolated);
2772 * Nobody should be changing or seriously looking at
2773 * pc->mem_cgroup and pc->flags at this point:
2775 * - the page is uncharged
2777 * - the page is off-LRU
2779 * - an anonymous fault has exclusive page access, except for
2780 * a locked page table
2782 * - a page cache insertion, a swapin fault, or a migration
2783 * have the page locked
2785 pc->mem_cgroup = memcg;
2786 pc->flags = PCG_USED | PCG_MEM | (do_swap_account ? PCG_MEMSW : 0);
2789 unlock_page_lru(page, isolated);
2792 static DEFINE_MUTEX(set_limit_mutex);
2794 #ifdef CONFIG_MEMCG_KMEM
2796 * The memcg_slab_mutex is held whenever a per memcg kmem cache is created or
2797 * destroyed. It protects memcg_caches arrays and memcg_slab_caches lists.
2799 static DEFINE_MUTEX(memcg_slab_mutex);
2801 static DEFINE_MUTEX(activate_kmem_mutex);
2803 static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg)
2805 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) &&
2806 memcg_kmem_is_active(memcg);
2810 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2811 * in the memcg_cache_params struct.
2813 static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
2815 struct kmem_cache *cachep;
2817 VM_BUG_ON(p->is_root_cache);
2818 cachep = p->root_cache;
2819 return cache_from_memcg_idx(cachep, memcg_cache_id(p->memcg));
2822 #ifdef CONFIG_SLABINFO
2823 static int mem_cgroup_slabinfo_read(struct seq_file *m, void *v)
2825 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
2826 struct memcg_cache_params *params;
2828 if (!memcg_can_account_kmem(memcg))
2831 print_slabinfo_header(m);
2833 mutex_lock(&memcg_slab_mutex);
2834 list_for_each_entry(params, &memcg->memcg_slab_caches, list)
2835 cache_show(memcg_params_to_cache(params), m);
2836 mutex_unlock(&memcg_slab_mutex);
2842 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size)
2844 struct res_counter *fail_res;
2847 ret = res_counter_charge(&memcg->kmem, size, &fail_res);
2851 ret = try_charge(memcg, gfp, size >> PAGE_SHIFT);
2852 if (ret == -EINTR) {
2854 * try_charge() chose to bypass to root due to OOM kill or
2855 * fatal signal. Since our only options are to either fail
2856 * the allocation or charge it to this cgroup, do it as a
2857 * temporary condition. But we can't fail. From a kmem/slab
2858 * perspective, the cache has already been selected, by
2859 * mem_cgroup_kmem_get_cache(), so it is too late to change
2862 * This condition will only trigger if the task entered
2863 * memcg_charge_kmem in a sane state, but was OOM-killed
2864 * during try_charge() above. Tasks that were already dying
2865 * when the allocation triggers should have been already
2866 * directed to the root cgroup in memcontrol.h
2868 res_counter_charge_nofail(&memcg->res, size, &fail_res);
2869 if (do_swap_account)
2870 res_counter_charge_nofail(&memcg->memsw, size,
2874 res_counter_uncharge(&memcg->kmem, size);
2879 static void memcg_uncharge_kmem(struct mem_cgroup *memcg, u64 size)
2881 res_counter_uncharge(&memcg->res, size);
2882 if (do_swap_account)
2883 res_counter_uncharge(&memcg->memsw, size);
2886 if (res_counter_uncharge(&memcg->kmem, size))
2890 * Releases a reference taken in kmem_cgroup_css_offline in case
2891 * this last uncharge is racing with the offlining code or it is
2892 * outliving the memcg existence.
2894 * The memory barrier imposed by test&clear is paired with the
2895 * explicit one in memcg_kmem_mark_dead().
2897 if (memcg_kmem_test_and_clear_dead(memcg))
2898 css_put(&memcg->css);
2902 * helper for acessing a memcg's index. It will be used as an index in the
2903 * child cache array in kmem_cache, and also to derive its name. This function
2904 * will return -1 when this is not a kmem-limited memcg.
2906 int memcg_cache_id(struct mem_cgroup *memcg)
2908 return memcg ? memcg->kmemcg_id : -1;
2911 static int memcg_alloc_cache_id(void)
2916 id = ida_simple_get(&kmem_limited_groups,
2917 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2921 if (id < memcg_limited_groups_array_size)
2925 * There's no space for the new id in memcg_caches arrays,
2926 * so we have to grow them.
2929 size = 2 * (id + 1);
2930 if (size < MEMCG_CACHES_MIN_SIZE)
2931 size = MEMCG_CACHES_MIN_SIZE;
2932 else if (size > MEMCG_CACHES_MAX_SIZE)
2933 size = MEMCG_CACHES_MAX_SIZE;
2935 mutex_lock(&memcg_slab_mutex);
2936 err = memcg_update_all_caches(size);
2937 mutex_unlock(&memcg_slab_mutex);
2940 ida_simple_remove(&kmem_limited_groups, id);
2946 static void memcg_free_cache_id(int id)
2948 ida_simple_remove(&kmem_limited_groups, id);
2952 * We should update the current array size iff all caches updates succeed. This
2953 * can only be done from the slab side. The slab mutex needs to be held when
2956 void memcg_update_array_size(int num)
2958 memcg_limited_groups_array_size = num;
2961 int memcg_update_cache_size(struct kmem_cache *s, int num_groups)
2963 struct memcg_cache_params *cur_params = s->memcg_params;
2964 struct memcg_cache_params *new_params;
2968 VM_BUG_ON(!is_root_cache(s));
2970 size = num_groups * sizeof(void *);
2971 size += offsetof(struct memcg_cache_params, memcg_caches);
2973 new_params = kzalloc(size, GFP_KERNEL);
2977 new_params->is_root_cache = true;
2980 * There is the chance it will be bigger than
2981 * memcg_limited_groups_array_size, if we failed an allocation
2982 * in a cache, in which case all caches updated before it, will
2983 * have a bigger array.
2985 * But if that is the case, the data after
2986 * memcg_limited_groups_array_size is certainly unused
2988 for (i = 0; i < memcg_limited_groups_array_size; i++) {
2989 if (!cur_params->memcg_caches[i])
2991 new_params->memcg_caches[i] =
2992 cur_params->memcg_caches[i];
2996 * Ideally, we would wait until all caches succeed, and only
2997 * then free the old one. But this is not worth the extra
2998 * pointer per-cache we'd have to have for this.
3000 * It is not a big deal if some caches are left with a size
3001 * bigger than the others. And all updates will reset this
3004 rcu_assign_pointer(s->memcg_params, new_params);
3006 kfree_rcu(cur_params, rcu_head);
3010 static void memcg_register_cache(struct mem_cgroup *memcg,
3011 struct kmem_cache *root_cache)
3013 static char memcg_name_buf[NAME_MAX + 1]; /* protected by
3015 struct kmem_cache *cachep;
3018 lockdep_assert_held(&memcg_slab_mutex);
3020 id = memcg_cache_id(memcg);
3023 * Since per-memcg caches are created asynchronously on first
3024 * allocation (see memcg_kmem_get_cache()), several threads can try to
3025 * create the same cache, but only one of them may succeed.
3027 if (cache_from_memcg_idx(root_cache, id))
3030 cgroup_name(memcg->css.cgroup, memcg_name_buf, NAME_MAX + 1);
3031 cachep = memcg_create_kmem_cache(memcg, root_cache, memcg_name_buf);
3033 * If we could not create a memcg cache, do not complain, because
3034 * that's not critical at all as we can always proceed with the root
3040 css_get(&memcg->css);
3041 list_add(&cachep->memcg_params->list, &memcg->memcg_slab_caches);
3044 * Since readers won't lock (see cache_from_memcg_idx()), we need a
3045 * barrier here to ensure nobody will see the kmem_cache partially
3050 BUG_ON(root_cache->memcg_params->memcg_caches[id]);
3051 root_cache->memcg_params->memcg_caches[id] = cachep;
3054 static void memcg_unregister_cache(struct kmem_cache *cachep)
3056 struct kmem_cache *root_cache;
3057 struct mem_cgroup *memcg;
3060 lockdep_assert_held(&memcg_slab_mutex);
3062 BUG_ON(is_root_cache(cachep));
3064 root_cache = cachep->memcg_params->root_cache;
3065 memcg = cachep->memcg_params->memcg;
3066 id = memcg_cache_id(memcg);
3068 BUG_ON(root_cache->memcg_params->memcg_caches[id] != cachep);
3069 root_cache->memcg_params->memcg_caches[id] = NULL;
3071 list_del(&cachep->memcg_params->list);
3073 kmem_cache_destroy(cachep);
3075 /* drop the reference taken in memcg_register_cache */
3076 css_put(&memcg->css);
3080 * During the creation a new cache, we need to disable our accounting mechanism
3081 * altogether. This is true even if we are not creating, but rather just
3082 * enqueing new caches to be created.
3084 * This is because that process will trigger allocations; some visible, like
3085 * explicit kmallocs to auxiliary data structures, name strings and internal
3086 * cache structures; some well concealed, like INIT_WORK() that can allocate
3087 * objects during debug.
3089 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3090 * to it. This may not be a bounded recursion: since the first cache creation
3091 * failed to complete (waiting on the allocation), we'll just try to create the
3092 * cache again, failing at the same point.
3094 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3095 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3096 * inside the following two functions.
3098 static inline void memcg_stop_kmem_account(void)
3100 VM_BUG_ON(!current->mm);
3101 current->memcg_kmem_skip_account++;
3104 static inline void memcg_resume_kmem_account(void)
3106 VM_BUG_ON(!current->mm);
3107 current->memcg_kmem_skip_account--;
3110 int __memcg_cleanup_cache_params(struct kmem_cache *s)
3112 struct kmem_cache *c;
3115 mutex_lock(&memcg_slab_mutex);
3116 for_each_memcg_cache_index(i) {
3117 c = cache_from_memcg_idx(s, i);
3121 memcg_unregister_cache(c);
3123 if (cache_from_memcg_idx(s, i))
3126 mutex_unlock(&memcg_slab_mutex);
3130 static void memcg_unregister_all_caches(struct mem_cgroup *memcg)
3132 struct kmem_cache *cachep;
3133 struct memcg_cache_params *params, *tmp;
3135 if (!memcg_kmem_is_active(memcg))
3138 mutex_lock(&memcg_slab_mutex);
3139 list_for_each_entry_safe(params, tmp, &memcg->memcg_slab_caches, list) {
3140 cachep = memcg_params_to_cache(params);
3141 kmem_cache_shrink(cachep);
3142 if (atomic_read(&cachep->memcg_params->nr_pages) == 0)
3143 memcg_unregister_cache(cachep);
3145 mutex_unlock(&memcg_slab_mutex);
3148 struct memcg_register_cache_work {
3149 struct mem_cgroup *memcg;
3150 struct kmem_cache *cachep;
3151 struct work_struct work;
3154 static void memcg_register_cache_func(struct work_struct *w)
3156 struct memcg_register_cache_work *cw =
3157 container_of(w, struct memcg_register_cache_work, work);
3158 struct mem_cgroup *memcg = cw->memcg;
3159 struct kmem_cache *cachep = cw->cachep;
3161 mutex_lock(&memcg_slab_mutex);
3162 memcg_register_cache(memcg, cachep);
3163 mutex_unlock(&memcg_slab_mutex);
3165 css_put(&memcg->css);
3170 * Enqueue the creation of a per-memcg kmem_cache.
3172 static void __memcg_schedule_register_cache(struct mem_cgroup *memcg,
3173 struct kmem_cache *cachep)
3175 struct memcg_register_cache_work *cw;
3177 cw = kmalloc(sizeof(*cw), GFP_NOWAIT);
3179 css_put(&memcg->css);
3184 cw->cachep = cachep;
3186 INIT_WORK(&cw->work, memcg_register_cache_func);
3187 schedule_work(&cw->work);
3190 static void memcg_schedule_register_cache(struct mem_cgroup *memcg,
3191 struct kmem_cache *cachep)
3194 * We need to stop accounting when we kmalloc, because if the
3195 * corresponding kmalloc cache is not yet created, the first allocation
3196 * in __memcg_schedule_register_cache will recurse.
3198 * However, it is better to enclose the whole function. Depending on
3199 * the debugging options enabled, INIT_WORK(), for instance, can
3200 * trigger an allocation. This too, will make us recurse. Because at
3201 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3202 * the safest choice is to do it like this, wrapping the whole function.
3204 memcg_stop_kmem_account();
3205 __memcg_schedule_register_cache(memcg, cachep);
3206 memcg_resume_kmem_account();
3209 int __memcg_charge_slab(struct kmem_cache *cachep, gfp_t gfp, int order)
3213 res = memcg_charge_kmem(cachep->memcg_params->memcg, gfp,
3214 PAGE_SIZE << order);
3216 atomic_add(1 << order, &cachep->memcg_params->nr_pages);
3220 void __memcg_uncharge_slab(struct kmem_cache *cachep, int order)
3222 memcg_uncharge_kmem(cachep->memcg_params->memcg, PAGE_SIZE << order);
3223 atomic_sub(1 << order, &cachep->memcg_params->nr_pages);
3227 * Return the kmem_cache we're supposed to use for a slab allocation.
3228 * We try to use the current memcg's version of the cache.
3230 * If the cache does not exist yet, if we are the first user of it,
3231 * we either create it immediately, if possible, or create it asynchronously
3233 * In the latter case, we will let the current allocation go through with
3234 * the original cache.
3236 * Can't be called in interrupt context or from kernel threads.
3237 * This function needs to be called with rcu_read_lock() held.
3239 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
3242 struct mem_cgroup *memcg;
3243 struct kmem_cache *memcg_cachep;
3245 VM_BUG_ON(!cachep->memcg_params);
3246 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
3248 if (!current->mm || current->memcg_kmem_skip_account)
3252 memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
3254 if (!memcg_can_account_kmem(memcg))
3257 memcg_cachep = cache_from_memcg_idx(cachep, memcg_cache_id(memcg));
3258 if (likely(memcg_cachep)) {
3259 cachep = memcg_cachep;
3263 /* The corresponding put will be done in the workqueue. */
3264 if (!css_tryget_online(&memcg->css))
3269 * If we are in a safe context (can wait, and not in interrupt
3270 * context), we could be be predictable and return right away.
3271 * This would guarantee that the allocation being performed
3272 * already belongs in the new cache.
3274 * However, there are some clashes that can arrive from locking.
3275 * For instance, because we acquire the slab_mutex while doing
3276 * memcg_create_kmem_cache, this means no further allocation
3277 * could happen with the slab_mutex held. So it's better to
3280 memcg_schedule_register_cache(memcg, cachep);
3288 * We need to verify if the allocation against current->mm->owner's memcg is
3289 * possible for the given order. But the page is not allocated yet, so we'll
3290 * need a further commit step to do the final arrangements.
3292 * It is possible for the task to switch cgroups in this mean time, so at
3293 * commit time, we can't rely on task conversion any longer. We'll then use
3294 * the handle argument to return to the caller which cgroup we should commit
3295 * against. We could also return the memcg directly and avoid the pointer
3296 * passing, but a boolean return value gives better semantics considering
3297 * the compiled-out case as well.
3299 * Returning true means the allocation is possible.
3302 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3304 struct mem_cgroup *memcg;
3310 * Disabling accounting is only relevant for some specific memcg
3311 * internal allocations. Therefore we would initially not have such
3312 * check here, since direct calls to the page allocator that are
3313 * accounted to kmemcg (alloc_kmem_pages and friends) only happen
3314 * outside memcg core. We are mostly concerned with cache allocations,
3315 * and by having this test at memcg_kmem_get_cache, we are already able
3316 * to relay the allocation to the root cache and bypass the memcg cache
3319 * There is one exception, though: the SLUB allocator does not create
3320 * large order caches, but rather service large kmallocs directly from
3321 * the page allocator. Therefore, the following sequence when backed by
3322 * the SLUB allocator:
3324 * memcg_stop_kmem_account();
3325 * kmalloc(<large_number>)
3326 * memcg_resume_kmem_account();
3328 * would effectively ignore the fact that we should skip accounting,
3329 * since it will drive us directly to this function without passing
3330 * through the cache selector memcg_kmem_get_cache. Such large
3331 * allocations are extremely rare but can happen, for instance, for the
3332 * cache arrays. We bring this test here.
3334 if (!current->mm || current->memcg_kmem_skip_account)
3337 memcg = get_mem_cgroup_from_mm(current->mm);
3339 if (!memcg_can_account_kmem(memcg)) {
3340 css_put(&memcg->css);
3344 ret = memcg_charge_kmem(memcg, gfp, PAGE_SIZE << order);
3348 css_put(&memcg->css);
3352 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3355 struct page_cgroup *pc;
3357 VM_BUG_ON(mem_cgroup_is_root(memcg));
3359 /* The page allocation failed. Revert */
3361 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3365 * The page is freshly allocated and not visible to any
3366 * outside callers yet. Set up pc non-atomically.
3368 pc = lookup_page_cgroup(page);
3369 pc->mem_cgroup = memcg;
3370 pc->flags = PCG_USED;
3373 void __memcg_kmem_uncharge_pages(struct page *page, int order)
3375 struct mem_cgroup *memcg = NULL;
3376 struct page_cgroup *pc;
3379 pc = lookup_page_cgroup(page);
3380 if (!PageCgroupUsed(pc))
3383 memcg = pc->mem_cgroup;
3387 * We trust that only if there is a memcg associated with the page, it
3388 * is a valid allocation
3393 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
3394 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3397 static inline void memcg_unregister_all_caches(struct mem_cgroup *memcg)
3400 #endif /* CONFIG_MEMCG_KMEM */
3402 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3405 * Because tail pages are not marked as "used", set it. We're under
3406 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3407 * charge/uncharge will be never happen and move_account() is done under
3408 * compound_lock(), so we don't have to take care of races.
3410 void mem_cgroup_split_huge_fixup(struct page *head)
3412 struct page_cgroup *head_pc = lookup_page_cgroup(head);
3413 struct page_cgroup *pc;
3414 struct mem_cgroup *memcg;
3417 if (mem_cgroup_disabled())
3420 memcg = head_pc->mem_cgroup;
3421 for (i = 1; i < HPAGE_PMD_NR; i++) {
3423 pc->mem_cgroup = memcg;
3424 pc->flags = head_pc->flags;
3426 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
3429 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3432 * mem_cgroup_move_account - move account of the page
3434 * @nr_pages: number of regular pages (>1 for huge pages)
3435 * @pc: page_cgroup of the page.
3436 * @from: mem_cgroup which the page is moved from.
3437 * @to: mem_cgroup which the page is moved to. @from != @to.
3439 * The caller must confirm following.
3440 * - page is not on LRU (isolate_page() is useful.)
3441 * - compound_lock is held when nr_pages > 1
3443 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3446 static int mem_cgroup_move_account(struct page *page,
3447 unsigned int nr_pages,
3448 struct page_cgroup *pc,
3449 struct mem_cgroup *from,
3450 struct mem_cgroup *to)
3452 unsigned long flags;
3455 VM_BUG_ON(from == to);
3456 VM_BUG_ON_PAGE(PageLRU(page), page);
3458 * The page is isolated from LRU. So, collapse function
3459 * will not handle this page. But page splitting can happen.
3460 * Do this check under compound_page_lock(). The caller should
3464 if (nr_pages > 1 && !PageTransHuge(page))
3468 * Prevent mem_cgroup_migrate() from looking at pc->mem_cgroup
3469 * of its source page while we change it: page migration takes
3470 * both pages off the LRU, but page cache replacement doesn't.
3472 if (!trylock_page(page))
3476 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3479 move_lock_mem_cgroup(from, &flags);
3481 if (!PageAnon(page) && page_mapped(page)) {
3482 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
3484 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
3488 if (PageWriteback(page)) {
3489 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK],
3491 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK],
3496 * It is safe to change pc->mem_cgroup here because the page
3497 * is referenced, charged, and isolated - we can't race with
3498 * uncharging, charging, migration, or LRU putback.
3501 /* caller should have done css_get */
3502 pc->mem_cgroup = to;
3503 move_unlock_mem_cgroup(from, &flags);
3506 local_irq_disable();
3507 mem_cgroup_charge_statistics(to, page, nr_pages);
3508 memcg_check_events(to, page);
3509 mem_cgroup_charge_statistics(from, page, -nr_pages);
3510 memcg_check_events(from, page);
3519 * mem_cgroup_move_parent - moves page to the parent group
3520 * @page: the page to move
3521 * @pc: page_cgroup of the page
3522 * @child: page's cgroup
3524 * move charges to its parent or the root cgroup if the group has no
3525 * parent (aka use_hierarchy==0).
3526 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3527 * mem_cgroup_move_account fails) the failure is always temporary and
3528 * it signals a race with a page removal/uncharge or migration. In the
3529 * first case the page is on the way out and it will vanish from the LRU
3530 * on the next attempt and the call should be retried later.
3531 * Isolation from the LRU fails only if page has been isolated from
3532 * the LRU since we looked at it and that usually means either global
3533 * reclaim or migration going on. The page will either get back to the
3535 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3536 * (!PageCgroupUsed) or moved to a different group. The page will
3537 * disappear in the next attempt.
3539 static int mem_cgroup_move_parent(struct page *page,
3540 struct page_cgroup *pc,
3541 struct mem_cgroup *child)
3543 struct mem_cgroup *parent;
3544 unsigned int nr_pages;
3545 unsigned long uninitialized_var(flags);
3548 VM_BUG_ON(mem_cgroup_is_root(child));
3551 if (!get_page_unless_zero(page))
3553 if (isolate_lru_page(page))
3556 nr_pages = hpage_nr_pages(page);
3558 parent = parent_mem_cgroup(child);
3560 * If no parent, move charges to root cgroup.
3563 parent = root_mem_cgroup;
3566 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
3567 flags = compound_lock_irqsave(page);
3570 ret = mem_cgroup_move_account(page, nr_pages,
3573 __mem_cgroup_cancel_local_charge(child, nr_pages);
3576 compound_unlock_irqrestore(page, flags);
3577 putback_lru_page(page);
3584 #ifdef CONFIG_MEMCG_SWAP
3585 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
3588 int val = (charge) ? 1 : -1;
3589 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
3593 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3594 * @entry: swap entry to be moved
3595 * @from: mem_cgroup which the entry is moved from
3596 * @to: mem_cgroup which the entry is moved to
3598 * It succeeds only when the swap_cgroup's record for this entry is the same
3599 * as the mem_cgroup's id of @from.
3601 * Returns 0 on success, -EINVAL on failure.
3603 * The caller must have charged to @to, IOW, called res_counter_charge() about
3604 * both res and memsw, and called css_get().
3606 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3607 struct mem_cgroup *from, struct mem_cgroup *to)
3609 unsigned short old_id, new_id;
3611 old_id = mem_cgroup_id(from);
3612 new_id = mem_cgroup_id(to);
3614 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3615 mem_cgroup_swap_statistics(from, false);
3616 mem_cgroup_swap_statistics(to, true);
3618 * This function is only called from task migration context now.
3619 * It postpones res_counter and refcount handling till the end
3620 * of task migration(mem_cgroup_clear_mc()) for performance
3621 * improvement. But we cannot postpone css_get(to) because if
3622 * the process that has been moved to @to does swap-in, the
3623 * refcount of @to might be decreased to 0.
3625 * We are in attach() phase, so the cgroup is guaranteed to be
3626 * alive, so we can just call css_get().
3634 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3635 struct mem_cgroup *from, struct mem_cgroup *to)
3641 #ifdef CONFIG_DEBUG_VM
3642 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3644 struct page_cgroup *pc;
3646 pc = lookup_page_cgroup(page);
3648 * Can be NULL while feeding pages into the page allocator for
3649 * the first time, i.e. during boot or memory hotplug;
3650 * or when mem_cgroup_disabled().
3652 if (likely(pc) && PageCgroupUsed(pc))
3657 bool mem_cgroup_bad_page_check(struct page *page)
3659 if (mem_cgroup_disabled())
3662 return lookup_page_cgroup_used(page) != NULL;
3665 void mem_cgroup_print_bad_page(struct page *page)
3667 struct page_cgroup *pc;
3669 pc = lookup_page_cgroup_used(page);
3671 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
3672 pc, pc->flags, pc->mem_cgroup);
3677 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3678 unsigned long long val)
3681 u64 memswlimit, memlimit;
3683 int children = mem_cgroup_count_children(memcg);
3684 u64 curusage, oldusage;
3688 * For keeping hierarchical_reclaim simple, how long we should retry
3689 * is depends on callers. We set our retry-count to be function
3690 * of # of children which we should visit in this loop.
3692 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3694 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3697 while (retry_count) {
3698 if (signal_pending(current)) {
3703 * Rather than hide all in some function, I do this in
3704 * open coded manner. You see what this really does.
3705 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
3707 mutex_lock(&set_limit_mutex);
3708 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3709 if (memswlimit < val) {
3711 mutex_unlock(&set_limit_mutex);
3715 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3719 ret = res_counter_set_limit(&memcg->res, val);
3721 if (memswlimit == val)
3722 memcg->memsw_is_minimum = true;
3724 memcg->memsw_is_minimum = false;
3726 mutex_unlock(&set_limit_mutex);
3731 mem_cgroup_reclaim(memcg, GFP_KERNEL,
3732 MEM_CGROUP_RECLAIM_SHRINK);
3733 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3734 /* Usage is reduced ? */
3735 if (curusage >= oldusage)
3738 oldusage = curusage;
3740 if (!ret && enlarge)
3741 memcg_oom_recover(memcg);
3746 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3747 unsigned long long val)
3750 u64 memlimit, memswlimit, oldusage, curusage;
3751 int children = mem_cgroup_count_children(memcg);
3755 /* see mem_cgroup_resize_res_limit */
3756 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3757 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3758 while (retry_count) {
3759 if (signal_pending(current)) {
3764 * Rather than hide all in some function, I do this in
3765 * open coded manner. You see what this really does.
3766 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
3768 mutex_lock(&set_limit_mutex);
3769 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3770 if (memlimit > val) {
3772 mutex_unlock(&set_limit_mutex);
3775 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3776 if (memswlimit < val)
3778 ret = res_counter_set_limit(&memcg->memsw, val);
3780 if (memlimit == val)
3781 memcg->memsw_is_minimum = true;
3783 memcg->memsw_is_minimum = false;
3785 mutex_unlock(&set_limit_mutex);
3790 mem_cgroup_reclaim(memcg, GFP_KERNEL,
3791 MEM_CGROUP_RECLAIM_NOSWAP |
3792 MEM_CGROUP_RECLAIM_SHRINK);
3793 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3794 /* Usage is reduced ? */
3795 if (curusage >= oldusage)
3798 oldusage = curusage;
3800 if (!ret && enlarge)
3801 memcg_oom_recover(memcg);
3805 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3807 unsigned long *total_scanned)
3809 unsigned long nr_reclaimed = 0;
3810 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3811 unsigned long reclaimed;
3813 struct mem_cgroup_tree_per_zone *mctz;
3814 unsigned long long excess;
3815 unsigned long nr_scanned;
3820 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3822 * This loop can run a while, specially if mem_cgroup's continuously
3823 * keep exceeding their soft limit and putting the system under
3830 mz = mem_cgroup_largest_soft_limit_node(mctz);
3835 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
3836 gfp_mask, &nr_scanned);
3837 nr_reclaimed += reclaimed;
3838 *total_scanned += nr_scanned;
3839 spin_lock_irq(&mctz->lock);
3842 * If we failed to reclaim anything from this memory cgroup
3843 * it is time to move on to the next cgroup
3849 * Loop until we find yet another one.
3851 * By the time we get the soft_limit lock
3852 * again, someone might have aded the
3853 * group back on the RB tree. Iterate to
3854 * make sure we get a different mem.
3855 * mem_cgroup_largest_soft_limit_node returns
3856 * NULL if no other cgroup is present on
3860 __mem_cgroup_largest_soft_limit_node(mctz);
3862 css_put(&next_mz->memcg->css);
3863 else /* next_mz == NULL or other memcg */
3867 __mem_cgroup_remove_exceeded(mz, mctz);
3868 excess = res_counter_soft_limit_excess(&mz->memcg->res);
3870 * One school of thought says that we should not add
3871 * back the node to the tree if reclaim returns 0.
3872 * But our reclaim could return 0, simply because due
3873 * to priority we are exposing a smaller subset of
3874 * memory to reclaim from. Consider this as a longer
3877 /* If excess == 0, no tree ops */
3878 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3879 spin_unlock_irq(&mctz->lock);
3880 css_put(&mz->memcg->css);
3883 * Could not reclaim anything and there are no more
3884 * mem cgroups to try or we seem to be looping without
3885 * reclaiming anything.
3887 if (!nr_reclaimed &&
3889 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3891 } while (!nr_reclaimed);
3893 css_put(&next_mz->memcg->css);
3894 return nr_reclaimed;
3898 * mem_cgroup_force_empty_list - clears LRU of a group
3899 * @memcg: group to clear
3902 * @lru: lru to to clear
3904 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
3905 * reclaim the pages page themselves - pages are moved to the parent (or root)
3908 static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
3909 int node, int zid, enum lru_list lru)
3911 struct lruvec *lruvec;
3912 unsigned long flags;
3913 struct list_head *list;
3917 zone = &NODE_DATA(node)->node_zones[zid];
3918 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
3919 list = &lruvec->lists[lru];
3923 struct page_cgroup *pc;
3926 spin_lock_irqsave(&zone->lru_lock, flags);
3927 if (list_empty(list)) {
3928 spin_unlock_irqrestore(&zone->lru_lock, flags);
3931 page = list_entry(list->prev, struct page, lru);
3933 list_move(&page->lru, list);
3935 spin_unlock_irqrestore(&zone->lru_lock, flags);
3938 spin_unlock_irqrestore(&zone->lru_lock, flags);
3940 pc = lookup_page_cgroup(page);
3942 if (mem_cgroup_move_parent(page, pc, memcg)) {
3943 /* found lock contention or "pc" is obsolete. */
3948 } while (!list_empty(list));
3952 * make mem_cgroup's charge to be 0 if there is no task by moving
3953 * all the charges and pages to the parent.
3954 * This enables deleting this mem_cgroup.
3956 * Caller is responsible for holding css reference on the memcg.
3958 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
3964 /* This is for making all *used* pages to be on LRU. */
3965 lru_add_drain_all();
3966 drain_all_stock_sync(memcg);
3967 mem_cgroup_start_move(memcg);
3968 for_each_node_state(node, N_MEMORY) {
3969 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3972 mem_cgroup_force_empty_list(memcg,
3977 mem_cgroup_end_move(memcg);
3978 memcg_oom_recover(memcg);
3982 * Kernel memory may not necessarily be trackable to a specific
3983 * process. So they are not migrated, and therefore we can't
3984 * expect their value to drop to 0 here.
3985 * Having res filled up with kmem only is enough.
3987 * This is a safety check because mem_cgroup_force_empty_list
3988 * could have raced with mem_cgroup_replace_page_cache callers
3989 * so the lru seemed empty but the page could have been added
3990 * right after the check. RES_USAGE should be safe as we always
3991 * charge before adding to the LRU.
3993 usage = res_counter_read_u64(&memcg->res, RES_USAGE) -
3994 res_counter_read_u64(&memcg->kmem, RES_USAGE);
3995 } while (usage > 0);
3999 * Test whether @memcg has children, dead or alive. Note that this
4000 * function doesn't care whether @memcg has use_hierarchy enabled and
4001 * returns %true if there are child csses according to the cgroup
4002 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
4004 static inline bool memcg_has_children(struct mem_cgroup *memcg)
4009 * The lock does not prevent addition or deletion of children, but
4010 * it prevents a new child from being initialized based on this
4011 * parent in css_online(), so it's enough to decide whether
4012 * hierarchically inherited attributes can still be changed or not.
4014 lockdep_assert_held(&memcg_create_mutex);
4017 ret = css_next_child(NULL, &memcg->css);
4023 * Reclaims as many pages from the given memcg as possible and moves
4024 * the rest to the parent.
4026 * Caller is responsible for holding css reference for memcg.
4028 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
4030 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
4032 /* we call try-to-free pages for make this cgroup empty */
4033 lru_add_drain_all();
4034 /* try to free all pages in this cgroup */
4035 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
4038 if (signal_pending(current))
4041 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
4045 /* maybe some writeback is necessary */
4046 congestion_wait(BLK_RW_ASYNC, HZ/10);
4054 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
4055 char *buf, size_t nbytes,
4058 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
4060 if (mem_cgroup_is_root(memcg))
4062 return mem_cgroup_force_empty(memcg) ?: nbytes;
4065 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
4068 return mem_cgroup_from_css(css)->use_hierarchy;
4071 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
4072 struct cftype *cft, u64 val)
4075 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4076 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
4078 mutex_lock(&memcg_create_mutex);
4080 if (memcg->use_hierarchy == val)
4084 * If parent's use_hierarchy is set, we can't make any modifications
4085 * in the child subtrees. If it is unset, then the change can
4086 * occur, provided the current cgroup has no children.
4088 * For the root cgroup, parent_mem is NULL, we allow value to be
4089 * set if there are no children.
4091 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
4092 (val == 1 || val == 0)) {
4093 if (!memcg_has_children(memcg))
4094 memcg->use_hierarchy = val;
4101 mutex_unlock(&memcg_create_mutex);
4106 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
4107 enum mem_cgroup_stat_index idx)
4109 struct mem_cgroup *iter;
4112 /* Per-cpu values can be negative, use a signed accumulator */
4113 for_each_mem_cgroup_tree(iter, memcg)
4114 val += mem_cgroup_read_stat(iter, idx);
4116 if (val < 0) /* race ? */
4121 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
4125 if (!mem_cgroup_is_root(memcg)) {
4127 return res_counter_read_u64(&memcg->res, RES_USAGE);
4129 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
4133 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
4134 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
4136 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
4137 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
4140 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
4142 return val << PAGE_SHIFT;
4146 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
4149 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4150 enum res_type type = MEMFILE_TYPE(cft->private);
4151 int name = MEMFILE_ATTR(cft->private);
4155 if (name == RES_USAGE)
4156 return mem_cgroup_usage(memcg, false);
4157 return res_counter_read_u64(&memcg->res, name);
4159 if (name == RES_USAGE)
4160 return mem_cgroup_usage(memcg, true);
4161 return res_counter_read_u64(&memcg->memsw, name);
4163 return res_counter_read_u64(&memcg->kmem, name);
4170 #ifdef CONFIG_MEMCG_KMEM
4171 /* should be called with activate_kmem_mutex held */
4172 static int __memcg_activate_kmem(struct mem_cgroup *memcg,
4173 unsigned long long limit)
4178 if (memcg_kmem_is_active(memcg))
4182 * We are going to allocate memory for data shared by all memory
4183 * cgroups so let's stop accounting here.
4185 memcg_stop_kmem_account();
4188 * For simplicity, we won't allow this to be disabled. It also can't
4189 * be changed if the cgroup has children already, or if tasks had
4192 * If tasks join before we set the limit, a person looking at
4193 * kmem.usage_in_bytes will have no way to determine when it took
4194 * place, which makes the value quite meaningless.
4196 * After it first became limited, changes in the value of the limit are
4197 * of course permitted.
4199 mutex_lock(&memcg_create_mutex);
4200 if (cgroup_has_tasks(memcg->css.cgroup) ||
4201 (memcg->use_hierarchy && memcg_has_children(memcg)))
4203 mutex_unlock(&memcg_create_mutex);
4207 memcg_id = memcg_alloc_cache_id();
4213 memcg->kmemcg_id = memcg_id;
4214 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
4217 * We couldn't have accounted to this cgroup, because it hasn't got the
4218 * active bit set yet, so this should succeed.
4220 err = res_counter_set_limit(&memcg->kmem, limit);
4223 static_key_slow_inc(&memcg_kmem_enabled_key);
4225 * Setting the active bit after enabling static branching will
4226 * guarantee no one starts accounting before all call sites are
4229 memcg_kmem_set_active(memcg);
4231 memcg_resume_kmem_account();
4235 static int memcg_activate_kmem(struct mem_cgroup *memcg,
4236 unsigned long long limit)
4240 mutex_lock(&activate_kmem_mutex);
4241 ret = __memcg_activate_kmem(memcg, limit);
4242 mutex_unlock(&activate_kmem_mutex);
4246 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
4247 unsigned long long val)
4251 if (!memcg_kmem_is_active(memcg))
4252 ret = memcg_activate_kmem(memcg, val);
4254 ret = res_counter_set_limit(&memcg->kmem, val);
4258 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
4261 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4266 mutex_lock(&activate_kmem_mutex);
4268 * If the parent cgroup is not kmem-active now, it cannot be activated
4269 * after this point, because it has at least one child already.
4271 if (memcg_kmem_is_active(parent))
4272 ret = __memcg_activate_kmem(memcg, RES_COUNTER_MAX);
4273 mutex_unlock(&activate_kmem_mutex);
4277 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
4278 unsigned long long val)
4282 #endif /* CONFIG_MEMCG_KMEM */
4285 * The user of this function is...
4288 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
4289 char *buf, size_t nbytes, loff_t off)
4291 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
4294 unsigned long long val;
4297 buf = strstrip(buf);
4298 type = MEMFILE_TYPE(of_cft(of)->private);
4299 name = MEMFILE_ATTR(of_cft(of)->private);
4303 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
4307 /* This function does all necessary parse...reuse it */
4308 ret = res_counter_memparse_write_strategy(buf, &val);
4312 ret = mem_cgroup_resize_limit(memcg, val);
4313 else if (type == _MEMSWAP)
4314 ret = mem_cgroup_resize_memsw_limit(memcg, val);
4315 else if (type == _KMEM)
4316 ret = memcg_update_kmem_limit(memcg, val);
4320 case RES_SOFT_LIMIT:
4321 ret = res_counter_memparse_write_strategy(buf, &val);
4325 * For memsw, soft limits are hard to implement in terms
4326 * of semantics, for now, we support soft limits for
4327 * control without swap
4330 ret = res_counter_set_soft_limit(&memcg->res, val);
4335 ret = -EINVAL; /* should be BUG() ? */
4338 return ret ?: nbytes;
4341 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
4342 unsigned long long *mem_limit, unsigned long long *memsw_limit)
4344 unsigned long long min_limit, min_memsw_limit, tmp;
4346 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4347 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4348 if (!memcg->use_hierarchy)
4351 while (memcg->css.parent) {
4352 memcg = mem_cgroup_from_css(memcg->css.parent);
4353 if (!memcg->use_hierarchy)
4355 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
4356 min_limit = min(min_limit, tmp);
4357 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4358 min_memsw_limit = min(min_memsw_limit, tmp);
4361 *mem_limit = min_limit;
4362 *memsw_limit = min_memsw_limit;
4365 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
4366 size_t nbytes, loff_t off)
4368 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
4372 type = MEMFILE_TYPE(of_cft(of)->private);
4373 name = MEMFILE_ATTR(of_cft(of)->private);
4378 res_counter_reset_max(&memcg->res);
4379 else if (type == _MEMSWAP)
4380 res_counter_reset_max(&memcg->memsw);
4381 else if (type == _KMEM)
4382 res_counter_reset_max(&memcg->kmem);
4388 res_counter_reset_failcnt(&memcg->res);
4389 else if (type == _MEMSWAP)
4390 res_counter_reset_failcnt(&memcg->memsw);
4391 else if (type == _KMEM)
4392 res_counter_reset_failcnt(&memcg->kmem);
4401 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
4404 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
4408 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
4409 struct cftype *cft, u64 val)
4411 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4413 if (val >= (1 << NR_MOVE_TYPE))
4417 * No kind of locking is needed in here, because ->can_attach() will
4418 * check this value once in the beginning of the process, and then carry
4419 * on with stale data. This means that changes to this value will only
4420 * affect task migrations starting after the change.
4422 memcg->move_charge_at_immigrate = val;
4426 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
4427 struct cftype *cft, u64 val)
4434 static int memcg_numa_stat_show(struct seq_file *m, void *v)
4438 unsigned int lru_mask;
4441 static const struct numa_stat stats[] = {
4442 { "total", LRU_ALL },
4443 { "file", LRU_ALL_FILE },
4444 { "anon", LRU_ALL_ANON },
4445 { "unevictable", BIT(LRU_UNEVICTABLE) },
4447 const struct numa_stat *stat;
4450 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
4452 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
4453 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
4454 seq_printf(m, "%s=%lu", stat->name, nr);
4455 for_each_node_state(nid, N_MEMORY) {
4456 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4458 seq_printf(m, " N%d=%lu", nid, nr);
4463 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
4464 struct mem_cgroup *iter;
4467 for_each_mem_cgroup_tree(iter, memcg)
4468 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
4469 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
4470 for_each_node_state(nid, N_MEMORY) {
4472 for_each_mem_cgroup_tree(iter, memcg)
4473 nr += mem_cgroup_node_nr_lru_pages(
4474 iter, nid, stat->lru_mask);
4475 seq_printf(m, " N%d=%lu", nid, nr);
4482 #endif /* CONFIG_NUMA */
4484 static inline void mem_cgroup_lru_names_not_uptodate(void)
4486 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
4489 static int memcg_stat_show(struct seq_file *m, void *v)
4491 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
4492 struct mem_cgroup *mi;
4495 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
4496 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
4498 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
4499 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
4502 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
4503 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
4504 mem_cgroup_read_events(memcg, i));
4506 for (i = 0; i < NR_LRU_LISTS; i++)
4507 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
4508 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
4510 /* Hierarchical information */
4512 unsigned long long limit, memsw_limit;
4513 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
4514 seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
4515 if (do_swap_account)
4516 seq_printf(m, "hierarchical_memsw_limit %llu\n",
4520 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
4523 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
4525 for_each_mem_cgroup_tree(mi, memcg)
4526 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
4527 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
4530 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
4531 unsigned long long val = 0;
4533 for_each_mem_cgroup_tree(mi, memcg)
4534 val += mem_cgroup_read_events(mi, i);
4535 seq_printf(m, "total_%s %llu\n",
4536 mem_cgroup_events_names[i], val);
4539 for (i = 0; i < NR_LRU_LISTS; i++) {
4540 unsigned long long val = 0;
4542 for_each_mem_cgroup_tree(mi, memcg)
4543 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
4544 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
4547 #ifdef CONFIG_DEBUG_VM
4550 struct mem_cgroup_per_zone *mz;
4551 struct zone_reclaim_stat *rstat;
4552 unsigned long recent_rotated[2] = {0, 0};
4553 unsigned long recent_scanned[2] = {0, 0};
4555 for_each_online_node(nid)
4556 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4557 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
4558 rstat = &mz->lruvec.reclaim_stat;
4560 recent_rotated[0] += rstat->recent_rotated[0];
4561 recent_rotated[1] += rstat->recent_rotated[1];
4562 recent_scanned[0] += rstat->recent_scanned[0];
4563 recent_scanned[1] += rstat->recent_scanned[1];
4565 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
4566 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
4567 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
4568 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
4575 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
4578 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4580 return mem_cgroup_swappiness(memcg);
4583 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
4584 struct cftype *cft, u64 val)
4586 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4592 memcg->swappiness = val;
4594 vm_swappiness = val;
4599 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4601 struct mem_cgroup_threshold_ary *t;
4607 t = rcu_dereference(memcg->thresholds.primary);
4609 t = rcu_dereference(memcg->memsw_thresholds.primary);
4614 usage = mem_cgroup_usage(memcg, swap);
4617 * current_threshold points to threshold just below or equal to usage.
4618 * If it's not true, a threshold was crossed after last
4619 * call of __mem_cgroup_threshold().
4621 i = t->current_threshold;
4624 * Iterate backward over array of thresholds starting from
4625 * current_threshold and check if a threshold is crossed.
4626 * If none of thresholds below usage is crossed, we read
4627 * only one element of the array here.
4629 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4630 eventfd_signal(t->entries[i].eventfd, 1);
4632 /* i = current_threshold + 1 */
4636 * Iterate forward over array of thresholds starting from
4637 * current_threshold+1 and check if a threshold is crossed.
4638 * If none of thresholds above usage is crossed, we read
4639 * only one element of the array here.
4641 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4642 eventfd_signal(t->entries[i].eventfd, 1);
4644 /* Update current_threshold */
4645 t->current_threshold = i - 1;
4650 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4653 __mem_cgroup_threshold(memcg, false);
4654 if (do_swap_account)
4655 __mem_cgroup_threshold(memcg, true);
4657 memcg = parent_mem_cgroup(memcg);
4661 static int compare_thresholds(const void *a, const void *b)
4663 const struct mem_cgroup_threshold *_a = a;
4664 const struct mem_cgroup_threshold *_b = b;
4666 if (_a->threshold > _b->threshold)
4669 if (_a->threshold < _b->threshold)
4675 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4677 struct mem_cgroup_eventfd_list *ev;
4679 spin_lock(&memcg_oom_lock);
4681 list_for_each_entry(ev, &memcg->oom_notify, list)
4682 eventfd_signal(ev->eventfd, 1);
4684 spin_unlock(&memcg_oom_lock);
4688 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4690 struct mem_cgroup *iter;
4692 for_each_mem_cgroup_tree(iter, memcg)
4693 mem_cgroup_oom_notify_cb(iter);
4696 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4697 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4699 struct mem_cgroup_thresholds *thresholds;
4700 struct mem_cgroup_threshold_ary *new;
4701 u64 threshold, usage;
4704 ret = res_counter_memparse_write_strategy(args, &threshold);
4708 mutex_lock(&memcg->thresholds_lock);
4711 thresholds = &memcg->thresholds;
4712 usage = mem_cgroup_usage(memcg, false);
4713 } else if (type == _MEMSWAP) {
4714 thresholds = &memcg->memsw_thresholds;
4715 usage = mem_cgroup_usage(memcg, true);
4719 /* Check if a threshold crossed before adding a new one */
4720 if (thresholds->primary)
4721 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4723 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4725 /* Allocate memory for new array of thresholds */
4726 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4734 /* Copy thresholds (if any) to new array */
4735 if (thresholds->primary) {
4736 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4737 sizeof(struct mem_cgroup_threshold));
4740 /* Add new threshold */
4741 new->entries[size - 1].eventfd = eventfd;
4742 new->entries[size - 1].threshold = threshold;
4744 /* Sort thresholds. Registering of new threshold isn't time-critical */
4745 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4746 compare_thresholds, NULL);
4748 /* Find current threshold */
4749 new->current_threshold = -1;
4750 for (i = 0; i < size; i++) {
4751 if (new->entries[i].threshold <= usage) {
4753 * new->current_threshold will not be used until
4754 * rcu_assign_pointer(), so it's safe to increment
4757 ++new->current_threshold;
4762 /* Free old spare buffer and save old primary buffer as spare */
4763 kfree(thresholds->spare);
4764 thresholds->spare = thresholds->primary;
4766 rcu_assign_pointer(thresholds->primary, new);
4768 /* To be sure that nobody uses thresholds */
4772 mutex_unlock(&memcg->thresholds_lock);
4777 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4778 struct eventfd_ctx *eventfd, const char *args)
4780 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4783 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4784 struct eventfd_ctx *eventfd, const char *args)
4786 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4789 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4790 struct eventfd_ctx *eventfd, enum res_type type)
4792 struct mem_cgroup_thresholds *thresholds;
4793 struct mem_cgroup_threshold_ary *new;
4797 mutex_lock(&memcg->thresholds_lock);
4800 thresholds = &memcg->thresholds;
4801 usage = mem_cgroup_usage(memcg, false);
4802 } else if (type == _MEMSWAP) {
4803 thresholds = &memcg->memsw_thresholds;
4804 usage = mem_cgroup_usage(memcg, true);
4808 if (!thresholds->primary)
4811 /* Check if a threshold crossed before removing */
4812 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4814 /* Calculate new number of threshold */
4816 for (i = 0; i < thresholds->primary->size; i++) {
4817 if (thresholds->primary->entries[i].eventfd != eventfd)
4821 new = thresholds->spare;
4823 /* Set thresholds array to NULL if we don't have thresholds */
4832 /* Copy thresholds and find current threshold */
4833 new->current_threshold = -1;
4834 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4835 if (thresholds->primary->entries[i].eventfd == eventfd)
4838 new->entries[j] = thresholds->primary->entries[i];
4839 if (new->entries[j].threshold <= usage) {
4841 * new->current_threshold will not be used
4842 * until rcu_assign_pointer(), so it's safe to increment
4845 ++new->current_threshold;
4851 /* Swap primary and spare array */
4852 thresholds->spare = thresholds->primary;
4853 /* If all events are unregistered, free the spare array */
4855 kfree(thresholds->spare);
4856 thresholds->spare = NULL;
4859 rcu_assign_pointer(thresholds->primary, new);
4861 /* To be sure that nobody uses thresholds */
4864 mutex_unlock(&memcg->thresholds_lock);
4867 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4868 struct eventfd_ctx *eventfd)
4870 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4873 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4874 struct eventfd_ctx *eventfd)
4876 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4879 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4880 struct eventfd_ctx *eventfd, const char *args)
4882 struct mem_cgroup_eventfd_list *event;
4884 event = kmalloc(sizeof(*event), GFP_KERNEL);
4888 spin_lock(&memcg_oom_lock);
4890 event->eventfd = eventfd;
4891 list_add(&event->list, &memcg->oom_notify);
4893 /* already in OOM ? */
4894 if (atomic_read(&memcg->under_oom))
4895 eventfd_signal(eventfd, 1);
4896 spin_unlock(&memcg_oom_lock);
4901 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4902 struct eventfd_ctx *eventfd)
4904 struct mem_cgroup_eventfd_list *ev, *tmp;
4906 spin_lock(&memcg_oom_lock);
4908 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4909 if (ev->eventfd == eventfd) {
4910 list_del(&ev->list);
4915 spin_unlock(&memcg_oom_lock);
4918 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4920 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
4922 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4923 seq_printf(sf, "under_oom %d\n", (bool)atomic_read(&memcg->under_oom));
4927 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4928 struct cftype *cft, u64 val)
4930 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4932 /* cannot set to root cgroup and only 0 and 1 are allowed */
4933 if (!css->parent || !((val == 0) || (val == 1)))
4936 memcg->oom_kill_disable = val;
4938 memcg_oom_recover(memcg);
4943 #ifdef CONFIG_MEMCG_KMEM
4944 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
4948 memcg->kmemcg_id = -1;
4949 ret = memcg_propagate_kmem(memcg);
4953 return mem_cgroup_sockets_init(memcg, ss);
4956 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
4958 mem_cgroup_sockets_destroy(memcg);
4961 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
4963 if (!memcg_kmem_is_active(memcg))
4967 * kmem charges can outlive the cgroup. In the case of slab
4968 * pages, for instance, a page contain objects from various
4969 * processes. As we prevent from taking a reference for every
4970 * such allocation we have to be careful when doing uncharge
4971 * (see memcg_uncharge_kmem) and here during offlining.
4973 * The idea is that that only the _last_ uncharge which sees
4974 * the dead memcg will drop the last reference. An additional
4975 * reference is taken here before the group is marked dead
4976 * which is then paired with css_put during uncharge resp. here.
4978 * Although this might sound strange as this path is called from
4979 * css_offline() when the referencemight have dropped down to 0 and
4980 * shouldn't be incremented anymore (css_tryget_online() would
4981 * fail) we do not have other options because of the kmem
4982 * allocations lifetime.
4984 css_get(&memcg->css);
4986 memcg_kmem_mark_dead(memcg);
4988 if (res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0)
4991 if (memcg_kmem_test_and_clear_dead(memcg))
4992 css_put(&memcg->css);
4995 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5000 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5004 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5010 * DO NOT USE IN NEW FILES.
5012 * "cgroup.event_control" implementation.
5014 * This is way over-engineered. It tries to support fully configurable
5015 * events for each user. Such level of flexibility is completely
5016 * unnecessary especially in the light of the planned unified hierarchy.
5018 * Please deprecate this and replace with something simpler if at all
5023 * Unregister event and free resources.
5025 * Gets called from workqueue.
5027 static void memcg_event_remove(struct work_struct *work)
5029 struct mem_cgroup_event *event =
5030 container_of(work, struct mem_cgroup_event, remove);
5031 struct mem_cgroup *memcg = event->memcg;
5033 remove_wait_queue(event->wqh, &event->wait);
5035 event->unregister_event(memcg, event->eventfd);
5037 /* Notify userspace the event is going away. */
5038 eventfd_signal(event->eventfd, 1);
5040 eventfd_ctx_put(event->eventfd);
5042 css_put(&memcg->css);
5046 * Gets called on POLLHUP on eventfd when user closes it.
5048 * Called with wqh->lock held and interrupts disabled.
5050 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
5051 int sync, void *key)
5053 struct mem_cgroup_event *event =
5054 container_of(wait, struct mem_cgroup_event, wait);
5055 struct mem_cgroup *memcg = event->memcg;
5056 unsigned long flags = (unsigned long)key;
5058 if (flags & POLLHUP) {
5060 * If the event has been detached at cgroup removal, we
5061 * can simply return knowing the other side will cleanup
5064 * We can't race against event freeing since the other
5065 * side will require wqh->lock via remove_wait_queue(),
5068 spin_lock(&memcg->event_list_lock);
5069 if (!list_empty(&event->list)) {
5070 list_del_init(&event->list);
5072 * We are in atomic context, but cgroup_event_remove()
5073 * may sleep, so we have to call it in workqueue.
5075 schedule_work(&event->remove);
5077 spin_unlock(&memcg->event_list_lock);
5083 static void memcg_event_ptable_queue_proc(struct file *file,
5084 wait_queue_head_t *wqh, poll_table *pt)
5086 struct mem_cgroup_event *event =
5087 container_of(pt, struct mem_cgroup_event, pt);
5090 add_wait_queue(wqh, &event->wait);
5094 * DO NOT USE IN NEW FILES.
5096 * Parse input and register new cgroup event handler.
5098 * Input must be in format '<event_fd> <control_fd> <args>'.
5099 * Interpretation of args is defined by control file implementation.
5101 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
5102 char *buf, size_t nbytes, loff_t off)
5104 struct cgroup_subsys_state *css = of_css(of);
5105 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5106 struct mem_cgroup_event *event;
5107 struct cgroup_subsys_state *cfile_css;
5108 unsigned int efd, cfd;
5115 buf = strstrip(buf);
5117 efd = simple_strtoul(buf, &endp, 10);
5122 cfd = simple_strtoul(buf, &endp, 10);
5123 if ((*endp != ' ') && (*endp != '\0'))
5127 event = kzalloc(sizeof(*event), GFP_KERNEL);
5131 event->memcg = memcg;
5132 INIT_LIST_HEAD(&event->list);
5133 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
5134 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
5135 INIT_WORK(&event->remove, memcg_event_remove);
5143 event->eventfd = eventfd_ctx_fileget(efile.file);
5144 if (IS_ERR(event->eventfd)) {
5145 ret = PTR_ERR(event->eventfd);
5152 goto out_put_eventfd;
5155 /* the process need read permission on control file */
5156 /* AV: shouldn't we check that it's been opened for read instead? */
5157 ret = inode_permission(file_inode(cfile.file), MAY_READ);
5162 * Determine the event callbacks and set them in @event. This used
5163 * to be done via struct cftype but cgroup core no longer knows
5164 * about these events. The following is crude but the whole thing
5165 * is for compatibility anyway.
5167 * DO NOT ADD NEW FILES.
5169 name = cfile.file->f_dentry->d_name.name;
5171 if (!strcmp(name, "memory.usage_in_bytes")) {
5172 event->register_event = mem_cgroup_usage_register_event;
5173 event->unregister_event = mem_cgroup_usage_unregister_event;
5174 } else if (!strcmp(name, "memory.oom_control")) {
5175 event->register_event = mem_cgroup_oom_register_event;
5176 event->unregister_event = mem_cgroup_oom_unregister_event;
5177 } else if (!strcmp(name, "memory.pressure_level")) {
5178 event->register_event = vmpressure_register_event;
5179 event->unregister_event = vmpressure_unregister_event;
5180 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
5181 event->register_event = memsw_cgroup_usage_register_event;
5182 event->unregister_event = memsw_cgroup_usage_unregister_event;
5189 * Verify @cfile should belong to @css. Also, remaining events are
5190 * automatically removed on cgroup destruction but the removal is
5191 * asynchronous, so take an extra ref on @css.
5193 cfile_css = css_tryget_online_from_dir(cfile.file->f_dentry->d_parent,
5194 &memory_cgrp_subsys);
5196 if (IS_ERR(cfile_css))
5198 if (cfile_css != css) {
5203 ret = event->register_event(memcg, event->eventfd, buf);
5207 efile.file->f_op->poll(efile.file, &event->pt);
5209 spin_lock(&memcg->event_list_lock);
5210 list_add(&event->list, &memcg->event_list);
5211 spin_unlock(&memcg->event_list_lock);
5223 eventfd_ctx_put(event->eventfd);
5232 static struct cftype mem_cgroup_files[] = {
5234 .name = "usage_in_bytes",
5235 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
5236 .read_u64 = mem_cgroup_read_u64,
5239 .name = "max_usage_in_bytes",
5240 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
5241 .write = mem_cgroup_reset,
5242 .read_u64 = mem_cgroup_read_u64,
5245 .name = "limit_in_bytes",
5246 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
5247 .write = mem_cgroup_write,
5248 .read_u64 = mem_cgroup_read_u64,
5251 .name = "soft_limit_in_bytes",
5252 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
5253 .write = mem_cgroup_write,
5254 .read_u64 = mem_cgroup_read_u64,
5258 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
5259 .write = mem_cgroup_reset,
5260 .read_u64 = mem_cgroup_read_u64,
5264 .seq_show = memcg_stat_show,
5267 .name = "force_empty",
5268 .write = mem_cgroup_force_empty_write,
5271 .name = "use_hierarchy",
5272 .write_u64 = mem_cgroup_hierarchy_write,
5273 .read_u64 = mem_cgroup_hierarchy_read,
5276 .name = "cgroup.event_control", /* XXX: for compat */
5277 .write = memcg_write_event_control,
5278 .flags = CFTYPE_NO_PREFIX,
5282 .name = "swappiness",
5283 .read_u64 = mem_cgroup_swappiness_read,
5284 .write_u64 = mem_cgroup_swappiness_write,
5287 .name = "move_charge_at_immigrate",
5288 .read_u64 = mem_cgroup_move_charge_read,
5289 .write_u64 = mem_cgroup_move_charge_write,
5292 .name = "oom_control",
5293 .seq_show = mem_cgroup_oom_control_read,
5294 .write_u64 = mem_cgroup_oom_control_write,
5295 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
5298 .name = "pressure_level",
5302 .name = "numa_stat",
5303 .seq_show = memcg_numa_stat_show,
5306 #ifdef CONFIG_MEMCG_KMEM
5308 .name = "kmem.limit_in_bytes",
5309 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
5310 .write = mem_cgroup_write,
5311 .read_u64 = mem_cgroup_read_u64,
5314 .name = "kmem.usage_in_bytes",
5315 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5316 .read_u64 = mem_cgroup_read_u64,
5319 .name = "kmem.failcnt",
5320 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5321 .write = mem_cgroup_reset,
5322 .read_u64 = mem_cgroup_read_u64,
5325 .name = "kmem.max_usage_in_bytes",
5326 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5327 .write = mem_cgroup_reset,
5328 .read_u64 = mem_cgroup_read_u64,
5330 #ifdef CONFIG_SLABINFO
5332 .name = "kmem.slabinfo",
5333 .seq_show = mem_cgroup_slabinfo_read,
5337 { }, /* terminate */
5340 #ifdef CONFIG_MEMCG_SWAP
5341 static struct cftype memsw_cgroup_files[] = {
5343 .name = "memsw.usage_in_bytes",
5344 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
5345 .read_u64 = mem_cgroup_read_u64,
5348 .name = "memsw.max_usage_in_bytes",
5349 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
5350 .write = mem_cgroup_reset,
5351 .read_u64 = mem_cgroup_read_u64,
5354 .name = "memsw.limit_in_bytes",
5355 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
5356 .write = mem_cgroup_write,
5357 .read_u64 = mem_cgroup_read_u64,
5360 .name = "memsw.failcnt",
5361 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
5362 .write = mem_cgroup_reset,
5363 .read_u64 = mem_cgroup_read_u64,
5365 { }, /* terminate */
5368 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
5370 struct mem_cgroup_per_node *pn;
5371 struct mem_cgroup_per_zone *mz;
5372 int zone, tmp = node;
5374 * This routine is called against possible nodes.
5375 * But it's BUG to call kmalloc() against offline node.
5377 * TODO: this routine can waste much memory for nodes which will
5378 * never be onlined. It's better to use memory hotplug callback
5381 if (!node_state(node, N_NORMAL_MEMORY))
5383 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5387 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5388 mz = &pn->zoneinfo[zone];
5389 lruvec_init(&mz->lruvec);
5390 mz->usage_in_excess = 0;
5391 mz->on_tree = false;
5394 memcg->nodeinfo[node] = pn;
5398 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
5400 kfree(memcg->nodeinfo[node]);
5403 static struct mem_cgroup *mem_cgroup_alloc(void)
5405 struct mem_cgroup *memcg;
5408 size = sizeof(struct mem_cgroup);
5409 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
5411 memcg = kzalloc(size, GFP_KERNEL);
5415 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
5418 spin_lock_init(&memcg->pcp_counter_lock);
5427 * At destroying mem_cgroup, references from swap_cgroup can remain.
5428 * (scanning all at force_empty is too costly...)
5430 * Instead of clearing all references at force_empty, we remember
5431 * the number of reference from swap_cgroup and free mem_cgroup when
5432 * it goes down to 0.
5434 * Removal of cgroup itself succeeds regardless of refs from swap.
5437 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5441 mem_cgroup_remove_from_trees(memcg);
5444 free_mem_cgroup_per_zone_info(memcg, node);
5446 free_percpu(memcg->stat);
5449 * We need to make sure that (at least for now), the jump label
5450 * destruction code runs outside of the cgroup lock. This is because
5451 * get_online_cpus(), which is called from the static_branch update,
5452 * can't be called inside the cgroup_lock. cpusets are the ones
5453 * enforcing this dependency, so if they ever change, we might as well.
5455 * schedule_work() will guarantee this happens. Be careful if you need
5456 * to move this code around, and make sure it is outside
5459 disarm_static_keys(memcg);
5464 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
5466 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
5468 if (!memcg->res.parent)
5470 return mem_cgroup_from_res_counter(memcg->res.parent, res);
5472 EXPORT_SYMBOL(parent_mem_cgroup);
5474 static void __init mem_cgroup_soft_limit_tree_init(void)
5476 struct mem_cgroup_tree_per_node *rtpn;
5477 struct mem_cgroup_tree_per_zone *rtpz;
5478 int tmp, node, zone;
5480 for_each_node(node) {
5482 if (!node_state(node, N_NORMAL_MEMORY))
5484 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
5487 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5489 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5490 rtpz = &rtpn->rb_tree_per_zone[zone];
5491 rtpz->rb_root = RB_ROOT;
5492 spin_lock_init(&rtpz->lock);
5497 static struct cgroup_subsys_state * __ref
5498 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5500 struct mem_cgroup *memcg;
5501 long error = -ENOMEM;
5504 memcg = mem_cgroup_alloc();
5506 return ERR_PTR(error);
5509 if (alloc_mem_cgroup_per_zone_info(memcg, node))
5513 if (parent_css == NULL) {
5514 root_mem_cgroup = memcg;
5515 res_counter_init(&memcg->res, NULL);
5516 res_counter_init(&memcg->memsw, NULL);
5517 res_counter_init(&memcg->kmem, NULL);
5520 memcg->last_scanned_node = MAX_NUMNODES;
5521 INIT_LIST_HEAD(&memcg->oom_notify);
5522 memcg->move_charge_at_immigrate = 0;
5523 mutex_init(&memcg->thresholds_lock);
5524 spin_lock_init(&memcg->move_lock);
5525 vmpressure_init(&memcg->vmpressure);
5526 INIT_LIST_HEAD(&memcg->event_list);
5527 spin_lock_init(&memcg->event_list_lock);
5532 __mem_cgroup_free(memcg);
5533 return ERR_PTR(error);
5537 mem_cgroup_css_online(struct cgroup_subsys_state *css)
5539 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5540 struct mem_cgroup *parent = mem_cgroup_from_css(css->parent);
5543 if (css->id > MEM_CGROUP_ID_MAX)
5549 mutex_lock(&memcg_create_mutex);
5551 memcg->use_hierarchy = parent->use_hierarchy;
5552 memcg->oom_kill_disable = parent->oom_kill_disable;
5553 memcg->swappiness = mem_cgroup_swappiness(parent);
5555 if (parent->use_hierarchy) {
5556 res_counter_init(&memcg->res, &parent->res);
5557 res_counter_init(&memcg->memsw, &parent->memsw);
5558 res_counter_init(&memcg->kmem, &parent->kmem);
5561 * No need to take a reference to the parent because cgroup
5562 * core guarantees its existence.
5565 res_counter_init(&memcg->res, NULL);
5566 res_counter_init(&memcg->memsw, NULL);
5567 res_counter_init(&memcg->kmem, NULL);
5569 * Deeper hierachy with use_hierarchy == false doesn't make
5570 * much sense so let cgroup subsystem know about this
5571 * unfortunate state in our controller.
5573 if (parent != root_mem_cgroup)
5574 memory_cgrp_subsys.broken_hierarchy = true;
5576 mutex_unlock(&memcg_create_mutex);
5578 ret = memcg_init_kmem(memcg, &memory_cgrp_subsys);
5583 * Make sure the memcg is initialized: mem_cgroup_iter()
5584 * orders reading memcg->initialized against its callers
5585 * reading the memcg members.
5587 smp_store_release(&memcg->initialized, 1);
5593 * Announce all parents that a group from their hierarchy is gone.
5595 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup *memcg)
5597 struct mem_cgroup *parent = memcg;
5599 while ((parent = parent_mem_cgroup(parent)))
5600 mem_cgroup_iter_invalidate(parent);
5603 * if the root memcg is not hierarchical we have to check it
5606 if (!root_mem_cgroup->use_hierarchy)
5607 mem_cgroup_iter_invalidate(root_mem_cgroup);
5610 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5612 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5613 struct mem_cgroup_event *event, *tmp;
5614 struct cgroup_subsys_state *iter;
5617 * Unregister events and notify userspace.
5618 * Notify userspace about cgroup removing only after rmdir of cgroup
5619 * directory to avoid race between userspace and kernelspace.
5621 spin_lock(&memcg->event_list_lock);
5622 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5623 list_del_init(&event->list);
5624 schedule_work(&event->remove);
5626 spin_unlock(&memcg->event_list_lock);
5628 kmem_cgroup_css_offline(memcg);
5630 mem_cgroup_invalidate_reclaim_iterators(memcg);
5633 * This requires that offlining is serialized. Right now that is
5634 * guaranteed because css_killed_work_fn() holds the cgroup_mutex.
5636 css_for_each_descendant_post(iter, css)
5637 mem_cgroup_reparent_charges(mem_cgroup_from_css(iter));
5639 memcg_unregister_all_caches(memcg);
5640 vmpressure_cleanup(&memcg->vmpressure);
5643 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5645 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5647 * XXX: css_offline() would be where we should reparent all
5648 * memory to prepare the cgroup for destruction. However,
5649 * memcg does not do css_tryget_online() and res_counter charging
5650 * under the same RCU lock region, which means that charging
5651 * could race with offlining. Offlining only happens to
5652 * cgroups with no tasks in them but charges can show up
5653 * without any tasks from the swapin path when the target
5654 * memcg is looked up from the swapout record and not from the
5655 * current task as it usually is. A race like this can leak
5656 * charges and put pages with stale cgroup pointers into
5660 * lookup_swap_cgroup_id()
5662 * mem_cgroup_lookup()
5663 * css_tryget_online()
5665 * disable css_tryget_online()
5668 * reparent_charges()
5669 * res_counter_charge()
5672 * pc->mem_cgroup = dead memcg
5675 * The bulk of the charges are still moved in offline_css() to
5676 * avoid pinning a lot of pages in case a long-term reference
5677 * like a swapout record is deferring the css_free() to long
5678 * after offlining. But this makes sure we catch any charges
5679 * made after offlining:
5681 mem_cgroup_reparent_charges(memcg);
5683 memcg_destroy_kmem(memcg);
5684 __mem_cgroup_free(memcg);
5688 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5689 * @css: the target css
5691 * Reset the states of the mem_cgroup associated with @css. This is
5692 * invoked when the userland requests disabling on the default hierarchy
5693 * but the memcg is pinned through dependency. The memcg should stop
5694 * applying policies and should revert to the vanilla state as it may be
5695 * made visible again.
5697 * The current implementation only resets the essential configurations.
5698 * This needs to be expanded to cover all the visible parts.
5700 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5702 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5704 mem_cgroup_resize_limit(memcg, ULLONG_MAX);
5705 mem_cgroup_resize_memsw_limit(memcg, ULLONG_MAX);
5706 memcg_update_kmem_limit(memcg, ULLONG_MAX);
5707 res_counter_set_soft_limit(&memcg->res, ULLONG_MAX);
5711 /* Handlers for move charge at task migration. */
5712 static int mem_cgroup_do_precharge(unsigned long count)
5716 /* Try a single bulk charge without reclaim first */
5717 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_WAIT, count);
5719 mc.precharge += count;
5722 if (ret == -EINTR) {
5723 cancel_charge(root_mem_cgroup, count);
5727 /* Try charges one by one with reclaim */
5729 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_NORETRY, 1);
5731 * In case of failure, any residual charges against
5732 * mc.to will be dropped by mem_cgroup_clear_mc()
5733 * later on. However, cancel any charges that are
5734 * bypassed to root right away or they'll be lost.
5737 cancel_charge(root_mem_cgroup, 1);
5747 * get_mctgt_type - get target type of moving charge
5748 * @vma: the vma the pte to be checked belongs
5749 * @addr: the address corresponding to the pte to be checked
5750 * @ptent: the pte to be checked
5751 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5754 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5755 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5756 * move charge. if @target is not NULL, the page is stored in target->page
5757 * with extra refcnt got(Callers should handle it).
5758 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5759 * target for charge migration. if @target is not NULL, the entry is stored
5762 * Called with pte lock held.
5769 enum mc_target_type {
5775 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5776 unsigned long addr, pte_t ptent)
5778 struct page *page = vm_normal_page(vma, addr, ptent);
5780 if (!page || !page_mapped(page))
5782 if (PageAnon(page)) {
5783 /* we don't move shared anon */
5786 } else if (!move_file())
5787 /* we ignore mapcount for file pages */
5789 if (!get_page_unless_zero(page))
5796 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5797 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5799 struct page *page = NULL;
5800 swp_entry_t ent = pte_to_swp_entry(ptent);
5802 if (!move_anon() || non_swap_entry(ent))
5805 * Because lookup_swap_cache() updates some statistics counter,
5806 * we call find_get_page() with swapper_space directly.
5808 page = find_get_page(swap_address_space(ent), ent.val);
5809 if (do_swap_account)
5810 entry->val = ent.val;
5815 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5816 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5822 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5823 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5825 struct page *page = NULL;
5826 struct address_space *mapping;
5829 if (!vma->vm_file) /* anonymous vma */
5834 mapping = vma->vm_file->f_mapping;
5835 if (pte_none(ptent))
5836 pgoff = linear_page_index(vma, addr);
5837 else /* pte_file(ptent) is true */
5838 pgoff = pte_to_pgoff(ptent);
5840 /* page is moved even if it's not RSS of this task(page-faulted). */
5842 /* shmem/tmpfs may report page out on swap: account for that too. */
5843 if (shmem_mapping(mapping)) {
5844 page = find_get_entry(mapping, pgoff);
5845 if (radix_tree_exceptional_entry(page)) {
5846 swp_entry_t swp = radix_to_swp_entry(page);
5847 if (do_swap_account)
5849 page = find_get_page(swap_address_space(swp), swp.val);
5852 page = find_get_page(mapping, pgoff);
5854 page = find_get_page(mapping, pgoff);
5859 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5860 unsigned long addr, pte_t ptent, union mc_target *target)
5862 struct page *page = NULL;
5863 struct page_cgroup *pc;
5864 enum mc_target_type ret = MC_TARGET_NONE;
5865 swp_entry_t ent = { .val = 0 };
5867 if (pte_present(ptent))
5868 page = mc_handle_present_pte(vma, addr, ptent);
5869 else if (is_swap_pte(ptent))
5870 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5871 else if (pte_none(ptent) || pte_file(ptent))
5872 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5874 if (!page && !ent.val)
5877 pc = lookup_page_cgroup(page);
5879 * Do only loose check w/o serialization.
5880 * mem_cgroup_move_account() checks the pc is valid or
5881 * not under LRU exclusion.
5883 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5884 ret = MC_TARGET_PAGE;
5886 target->page = page;
5888 if (!ret || !target)
5891 /* There is a swap entry and a page doesn't exist or isn't charged */
5892 if (ent.val && !ret &&
5893 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5894 ret = MC_TARGET_SWAP;
5901 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5903 * We don't consider swapping or file mapped pages because THP does not
5904 * support them for now.
5905 * Caller should make sure that pmd_trans_huge(pmd) is true.
5907 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5908 unsigned long addr, pmd_t pmd, union mc_target *target)
5910 struct page *page = NULL;
5911 struct page_cgroup *pc;
5912 enum mc_target_type ret = MC_TARGET_NONE;
5914 page = pmd_page(pmd);
5915 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5918 pc = lookup_page_cgroup(page);
5919 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5920 ret = MC_TARGET_PAGE;
5923 target->page = page;
5929 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5930 unsigned long addr, pmd_t pmd, union mc_target *target)
5932 return MC_TARGET_NONE;
5936 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5937 unsigned long addr, unsigned long end,
5938 struct mm_walk *walk)
5940 struct vm_area_struct *vma = walk->private;
5944 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
5945 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5946 mc.precharge += HPAGE_PMD_NR;
5951 if (pmd_trans_unstable(pmd))
5953 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5954 for (; addr != end; pte++, addr += PAGE_SIZE)
5955 if (get_mctgt_type(vma, addr, *pte, NULL))
5956 mc.precharge++; /* increment precharge temporarily */
5957 pte_unmap_unlock(pte - 1, ptl);
5963 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5965 unsigned long precharge;
5966 struct vm_area_struct *vma;
5968 down_read(&mm->mmap_sem);
5969 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5970 struct mm_walk mem_cgroup_count_precharge_walk = {
5971 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5975 if (is_vm_hugetlb_page(vma))
5977 walk_page_range(vma->vm_start, vma->vm_end,
5978 &mem_cgroup_count_precharge_walk);
5980 up_read(&mm->mmap_sem);
5982 precharge = mc.precharge;
5988 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5990 unsigned long precharge = mem_cgroup_count_precharge(mm);
5992 VM_BUG_ON(mc.moving_task);
5993 mc.moving_task = current;
5994 return mem_cgroup_do_precharge(precharge);
5997 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5998 static void __mem_cgroup_clear_mc(void)
6000 struct mem_cgroup *from = mc.from;
6001 struct mem_cgroup *to = mc.to;
6004 /* we must uncharge all the leftover precharges from mc.to */
6006 cancel_charge(mc.to, mc.precharge);
6010 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6011 * we must uncharge here.
6013 if (mc.moved_charge) {
6014 cancel_charge(mc.from, mc.moved_charge);
6015 mc.moved_charge = 0;
6017 /* we must fixup refcnts and charges */
6018 if (mc.moved_swap) {
6019 /* uncharge swap account from the old cgroup */
6020 if (!mem_cgroup_is_root(mc.from))
6021 res_counter_uncharge(&mc.from->memsw,
6022 PAGE_SIZE * mc.moved_swap);
6024 for (i = 0; i < mc.moved_swap; i++)
6025 css_put(&mc.from->css);
6028 * we charged both to->res and to->memsw, so we should
6031 if (!mem_cgroup_is_root(mc.to))
6032 res_counter_uncharge(&mc.to->res,
6033 PAGE_SIZE * mc.moved_swap);
6034 /* we've already done css_get(mc.to) */
6037 memcg_oom_recover(from);
6038 memcg_oom_recover(to);
6039 wake_up_all(&mc.waitq);
6042 static void mem_cgroup_clear_mc(void)
6044 struct mem_cgroup *from = mc.from;
6047 * we must clear moving_task before waking up waiters at the end of
6050 mc.moving_task = NULL;
6051 __mem_cgroup_clear_mc();
6052 spin_lock(&mc.lock);
6055 spin_unlock(&mc.lock);
6056 mem_cgroup_end_move(from);
6059 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
6060 struct cgroup_taskset *tset)
6062 struct task_struct *p = cgroup_taskset_first(tset);
6064 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6065 unsigned long move_charge_at_immigrate;
6068 * We are now commited to this value whatever it is. Changes in this
6069 * tunable will only affect upcoming migrations, not the current one.
6070 * So we need to save it, and keep it going.
6072 move_charge_at_immigrate = memcg->move_charge_at_immigrate;
6073 if (move_charge_at_immigrate) {
6074 struct mm_struct *mm;
6075 struct mem_cgroup *from = mem_cgroup_from_task(p);
6077 VM_BUG_ON(from == memcg);
6079 mm = get_task_mm(p);
6082 /* We move charges only when we move a owner of the mm */
6083 if (mm->owner == p) {
6086 VM_BUG_ON(mc.precharge);
6087 VM_BUG_ON(mc.moved_charge);
6088 VM_BUG_ON(mc.moved_swap);
6089 mem_cgroup_start_move(from);
6090 spin_lock(&mc.lock);
6093 mc.immigrate_flags = move_charge_at_immigrate;
6094 spin_unlock(&mc.lock);
6095 /* We set mc.moving_task later */
6097 ret = mem_cgroup_precharge_mc(mm);
6099 mem_cgroup_clear_mc();
6106 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
6107 struct cgroup_taskset *tset)
6109 mem_cgroup_clear_mc();
6112 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6113 unsigned long addr, unsigned long end,
6114 struct mm_walk *walk)
6117 struct vm_area_struct *vma = walk->private;
6120 enum mc_target_type target_type;
6121 union mc_target target;
6123 struct page_cgroup *pc;
6126 * We don't take compound_lock() here but no race with splitting thp
6128 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6129 * under splitting, which means there's no concurrent thp split,
6130 * - if another thread runs into split_huge_page() just after we
6131 * entered this if-block, the thread must wait for page table lock
6132 * to be unlocked in __split_huge_page_splitting(), where the main
6133 * part of thp split is not executed yet.
6135 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
6136 if (mc.precharge < HPAGE_PMD_NR) {
6140 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6141 if (target_type == MC_TARGET_PAGE) {
6143 if (!isolate_lru_page(page)) {
6144 pc = lookup_page_cgroup(page);
6145 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
6146 pc, mc.from, mc.to)) {
6147 mc.precharge -= HPAGE_PMD_NR;
6148 mc.moved_charge += HPAGE_PMD_NR;
6150 putback_lru_page(page);
6158 if (pmd_trans_unstable(pmd))
6161 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6162 for (; addr != end; addr += PAGE_SIZE) {
6163 pte_t ptent = *(pte++);
6169 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6170 case MC_TARGET_PAGE:
6172 if (isolate_lru_page(page))
6174 pc = lookup_page_cgroup(page);
6175 if (!mem_cgroup_move_account(page, 1, pc,
6178 /* we uncharge from mc.from later. */
6181 putback_lru_page(page);
6182 put: /* get_mctgt_type() gets the page */
6185 case MC_TARGET_SWAP:
6187 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6189 /* we fixup refcnts and charges later. */
6197 pte_unmap_unlock(pte - 1, ptl);
6202 * We have consumed all precharges we got in can_attach().
6203 * We try charge one by one, but don't do any additional
6204 * charges to mc.to if we have failed in charge once in attach()
6207 ret = mem_cgroup_do_precharge(1);
6215 static void mem_cgroup_move_charge(struct mm_struct *mm)
6217 struct vm_area_struct *vma;
6219 lru_add_drain_all();
6221 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
6223 * Someone who are holding the mmap_sem might be waiting in
6224 * waitq. So we cancel all extra charges, wake up all waiters,
6225 * and retry. Because we cancel precharges, we might not be able
6226 * to move enough charges, but moving charge is a best-effort
6227 * feature anyway, so it wouldn't be a big problem.
6229 __mem_cgroup_clear_mc();
6233 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6235 struct mm_walk mem_cgroup_move_charge_walk = {
6236 .pmd_entry = mem_cgroup_move_charge_pte_range,
6240 if (is_vm_hugetlb_page(vma))
6242 ret = walk_page_range(vma->vm_start, vma->vm_end,
6243 &mem_cgroup_move_charge_walk);
6246 * means we have consumed all precharges and failed in
6247 * doing additional charge. Just abandon here.
6251 up_read(&mm->mmap_sem);
6254 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
6255 struct cgroup_taskset *tset)
6257 struct task_struct *p = cgroup_taskset_first(tset);
6258 struct mm_struct *mm = get_task_mm(p);
6262 mem_cgroup_move_charge(mm);
6266 mem_cgroup_clear_mc();
6268 #else /* !CONFIG_MMU */
6269 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
6270 struct cgroup_taskset *tset)
6274 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
6275 struct cgroup_taskset *tset)
6278 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
6279 struct cgroup_taskset *tset)
6285 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6286 * to verify whether we're attached to the default hierarchy on each mount
6289 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
6292 * use_hierarchy is forced on the default hierarchy. cgroup core
6293 * guarantees that @root doesn't have any children, so turning it
6294 * on for the root memcg is enough.
6296 if (cgroup_on_dfl(root_css->cgroup))
6297 mem_cgroup_from_css(root_css)->use_hierarchy = true;
6300 struct cgroup_subsys memory_cgrp_subsys = {
6301 .css_alloc = mem_cgroup_css_alloc,
6302 .css_online = mem_cgroup_css_online,
6303 .css_offline = mem_cgroup_css_offline,
6304 .css_free = mem_cgroup_css_free,
6305 .css_reset = mem_cgroup_css_reset,
6306 .can_attach = mem_cgroup_can_attach,
6307 .cancel_attach = mem_cgroup_cancel_attach,
6308 .attach = mem_cgroup_move_task,
6309 .bind = mem_cgroup_bind,
6310 .legacy_cftypes = mem_cgroup_files,
6314 #ifdef CONFIG_MEMCG_SWAP
6315 static int __init enable_swap_account(char *s)
6317 if (!strcmp(s, "1"))
6318 really_do_swap_account = 1;
6319 else if (!strcmp(s, "0"))
6320 really_do_swap_account = 0;
6323 __setup("swapaccount=", enable_swap_account);
6325 static void __init memsw_file_init(void)
6327 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
6328 memsw_cgroup_files));
6331 static void __init enable_swap_cgroup(void)
6333 if (!mem_cgroup_disabled() && really_do_swap_account) {
6334 do_swap_account = 1;
6340 static void __init enable_swap_cgroup(void)
6345 #ifdef CONFIG_MEMCG_SWAP
6347 * mem_cgroup_swapout - transfer a memsw charge to swap
6348 * @page: page whose memsw charge to transfer
6349 * @entry: swap entry to move the charge to
6351 * Transfer the memsw charge of @page to @entry.
6353 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
6355 struct page_cgroup *pc;
6356 unsigned short oldid;
6358 VM_BUG_ON_PAGE(PageLRU(page), page);
6359 VM_BUG_ON_PAGE(page_count(page), page);
6361 if (!do_swap_account)
6364 pc = lookup_page_cgroup(page);
6366 /* Readahead page, never charged */
6367 if (!PageCgroupUsed(pc))
6370 VM_BUG_ON_PAGE(!(pc->flags & PCG_MEMSW), page);
6372 oldid = swap_cgroup_record(entry, mem_cgroup_id(pc->mem_cgroup));
6373 VM_BUG_ON_PAGE(oldid, page);
6375 pc->flags &= ~PCG_MEMSW;
6376 css_get(&pc->mem_cgroup->css);
6377 mem_cgroup_swap_statistics(pc->mem_cgroup, true);
6381 * mem_cgroup_uncharge_swap - uncharge a swap entry
6382 * @entry: swap entry to uncharge
6384 * Drop the memsw charge associated with @entry.
6386 void mem_cgroup_uncharge_swap(swp_entry_t entry)
6388 struct mem_cgroup *memcg;
6391 if (!do_swap_account)
6394 id = swap_cgroup_record(entry, 0);
6396 memcg = mem_cgroup_lookup(id);
6398 if (!mem_cgroup_is_root(memcg))
6399 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
6400 mem_cgroup_swap_statistics(memcg, false);
6401 css_put(&memcg->css);
6408 * mem_cgroup_try_charge - try charging a page
6409 * @page: page to charge
6410 * @mm: mm context of the victim
6411 * @gfp_mask: reclaim mode
6412 * @memcgp: charged memcg return
6414 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6415 * pages according to @gfp_mask if necessary.
6417 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
6418 * Otherwise, an error code is returned.
6420 * After page->mapping has been set up, the caller must finalize the
6421 * charge with mem_cgroup_commit_charge(). Or abort the transaction
6422 * with mem_cgroup_cancel_charge() in case page instantiation fails.
6424 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
6425 gfp_t gfp_mask, struct mem_cgroup **memcgp)
6427 struct mem_cgroup *memcg = NULL;
6428 unsigned int nr_pages = 1;
6431 if (mem_cgroup_disabled())
6434 if (PageSwapCache(page)) {
6435 struct page_cgroup *pc = lookup_page_cgroup(page);
6437 * Every swap fault against a single page tries to charge the
6438 * page, bail as early as possible. shmem_unuse() encounters
6439 * already charged pages, too. The USED bit is protected by
6440 * the page lock, which serializes swap cache removal, which
6441 * in turn serializes uncharging.
6443 if (PageCgroupUsed(pc))
6447 if (PageTransHuge(page)) {
6448 nr_pages <<= compound_order(page);
6449 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
6452 if (do_swap_account && PageSwapCache(page))
6453 memcg = try_get_mem_cgroup_from_page(page);
6455 memcg = get_mem_cgroup_from_mm(mm);
6457 ret = try_charge(memcg, gfp_mask, nr_pages);
6459 css_put(&memcg->css);
6461 if (ret == -EINTR) {
6462 memcg = root_mem_cgroup;
6471 * mem_cgroup_commit_charge - commit a page charge
6472 * @page: page to charge
6473 * @memcg: memcg to charge the page to
6474 * @lrucare: page might be on LRU already
6476 * Finalize a charge transaction started by mem_cgroup_try_charge(),
6477 * after page->mapping has been set up. This must happen atomically
6478 * as part of the page instantiation, i.e. under the page table lock
6479 * for anonymous pages, under the page lock for page and swap cache.
6481 * In addition, the page must not be on the LRU during the commit, to
6482 * prevent racing with task migration. If it might be, use @lrucare.
6484 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
6486 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
6489 unsigned int nr_pages = 1;
6491 VM_BUG_ON_PAGE(!page->mapping, page);
6492 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
6494 if (mem_cgroup_disabled())
6497 * Swap faults will attempt to charge the same page multiple
6498 * times. But reuse_swap_page() might have removed the page
6499 * from swapcache already, so we can't check PageSwapCache().
6504 commit_charge(page, memcg, lrucare);
6506 if (PageTransHuge(page)) {
6507 nr_pages <<= compound_order(page);
6508 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
6511 local_irq_disable();
6512 mem_cgroup_charge_statistics(memcg, page, nr_pages);
6513 memcg_check_events(memcg, page);
6516 if (do_swap_account && PageSwapCache(page)) {
6517 swp_entry_t entry = { .val = page_private(page) };
6519 * The swap entry might not get freed for a long time,
6520 * let's not wait for it. The page already received a
6521 * memory+swap charge, drop the swap entry duplicate.
6523 mem_cgroup_uncharge_swap(entry);
6528 * mem_cgroup_cancel_charge - cancel a page charge
6529 * @page: page to charge
6530 * @memcg: memcg to charge the page to
6532 * Cancel a charge transaction started by mem_cgroup_try_charge().
6534 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg)
6536 unsigned int nr_pages = 1;
6538 if (mem_cgroup_disabled())
6541 * Swap faults will attempt to charge the same page multiple
6542 * times. But reuse_swap_page() might have removed the page
6543 * from swapcache already, so we can't check PageSwapCache().
6548 if (PageTransHuge(page)) {
6549 nr_pages <<= compound_order(page);
6550 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
6553 cancel_charge(memcg, nr_pages);
6556 static void uncharge_batch(struct mem_cgroup *memcg, unsigned long pgpgout,
6557 unsigned long nr_mem, unsigned long nr_memsw,
6558 unsigned long nr_anon, unsigned long nr_file,
6559 unsigned long nr_huge, struct page *dummy_page)
6561 unsigned long flags;
6563 if (!mem_cgroup_is_root(memcg)) {
6565 res_counter_uncharge(&memcg->res,
6566 nr_mem * PAGE_SIZE);
6568 res_counter_uncharge(&memcg->memsw,
6569 nr_memsw * PAGE_SIZE);
6570 memcg_oom_recover(memcg);
6573 local_irq_save(flags);
6574 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS], nr_anon);
6575 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_CACHE], nr_file);
6576 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE], nr_huge);
6577 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT], pgpgout);
6578 __this_cpu_add(memcg->stat->nr_page_events, nr_anon + nr_file);
6579 memcg_check_events(memcg, dummy_page);
6580 local_irq_restore(flags);
6583 static void uncharge_list(struct list_head *page_list)
6585 struct mem_cgroup *memcg = NULL;
6586 unsigned long nr_memsw = 0;
6587 unsigned long nr_anon = 0;
6588 unsigned long nr_file = 0;
6589 unsigned long nr_huge = 0;
6590 unsigned long pgpgout = 0;
6591 unsigned long nr_mem = 0;
6592 struct list_head *next;
6595 next = page_list->next;
6597 unsigned int nr_pages = 1;
6598 struct page_cgroup *pc;
6600 page = list_entry(next, struct page, lru);
6601 next = page->lru.next;
6603 VM_BUG_ON_PAGE(PageLRU(page), page);
6604 VM_BUG_ON_PAGE(page_count(page), page);
6606 pc = lookup_page_cgroup(page);
6607 if (!PageCgroupUsed(pc))
6611 * Nobody should be changing or seriously looking at
6612 * pc->mem_cgroup and pc->flags at this point, we have
6613 * fully exclusive access to the page.
6616 if (memcg != pc->mem_cgroup) {
6618 uncharge_batch(memcg, pgpgout, nr_mem, nr_memsw,
6619 nr_anon, nr_file, nr_huge, page);
6620 pgpgout = nr_mem = nr_memsw = 0;
6621 nr_anon = nr_file = nr_huge = 0;
6623 memcg = pc->mem_cgroup;
6626 if (PageTransHuge(page)) {
6627 nr_pages <<= compound_order(page);
6628 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
6629 nr_huge += nr_pages;
6633 nr_anon += nr_pages;
6635 nr_file += nr_pages;
6637 if (pc->flags & PCG_MEM)
6639 if (pc->flags & PCG_MEMSW)
6640 nr_memsw += nr_pages;
6644 } while (next != page_list);
6647 uncharge_batch(memcg, pgpgout, nr_mem, nr_memsw,
6648 nr_anon, nr_file, nr_huge, page);
6652 * mem_cgroup_uncharge - uncharge a page
6653 * @page: page to uncharge
6655 * Uncharge a page previously charged with mem_cgroup_try_charge() and
6656 * mem_cgroup_commit_charge().
6658 void mem_cgroup_uncharge(struct page *page)
6660 struct page_cgroup *pc;
6662 if (mem_cgroup_disabled())
6665 /* Don't touch page->lru of any random page, pre-check: */
6666 pc = lookup_page_cgroup(page);
6667 if (!PageCgroupUsed(pc))
6670 INIT_LIST_HEAD(&page->lru);
6671 uncharge_list(&page->lru);
6675 * mem_cgroup_uncharge_list - uncharge a list of page
6676 * @page_list: list of pages to uncharge
6678 * Uncharge a list of pages previously charged with
6679 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
6681 void mem_cgroup_uncharge_list(struct list_head *page_list)
6683 if (mem_cgroup_disabled())
6686 if (!list_empty(page_list))
6687 uncharge_list(page_list);
6691 * mem_cgroup_migrate - migrate a charge to another page
6692 * @oldpage: currently charged page
6693 * @newpage: page to transfer the charge to
6694 * @lrucare: both pages might be on the LRU already
6696 * Migrate the charge from @oldpage to @newpage.
6698 * Both pages must be locked, @newpage->mapping must be set up.
6700 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage,
6703 struct page_cgroup *pc;
6706 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6707 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6708 VM_BUG_ON_PAGE(!lrucare && PageLRU(oldpage), oldpage);
6709 VM_BUG_ON_PAGE(!lrucare && PageLRU(newpage), newpage);
6710 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6711 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6714 if (mem_cgroup_disabled())
6717 /* Page cache replacement: new page already charged? */
6718 pc = lookup_page_cgroup(newpage);
6719 if (PageCgroupUsed(pc))
6722 /* Re-entrant migration: old page already uncharged? */
6723 pc = lookup_page_cgroup(oldpage);
6724 if (!PageCgroupUsed(pc))
6727 VM_BUG_ON_PAGE(!(pc->flags & PCG_MEM), oldpage);
6728 VM_BUG_ON_PAGE(do_swap_account && !(pc->flags & PCG_MEMSW), oldpage);
6731 lock_page_lru(oldpage, &isolated);
6736 unlock_page_lru(oldpage, isolated);
6738 commit_charge(newpage, pc->mem_cgroup, lrucare);
6742 * subsys_initcall() for memory controller.
6744 * Some parts like hotcpu_notifier() have to be initialized from this context
6745 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
6746 * everything that doesn't depend on a specific mem_cgroup structure should
6747 * be initialized from here.
6749 static int __init mem_cgroup_init(void)
6751 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
6752 enable_swap_cgroup();
6753 mem_cgroup_soft_limit_tree_init();
6757 subsys_initcall(mem_cgroup_init);