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 mem_cgroup_subsys __read_mostly;
70 EXPORT_SYMBOL(mem_cgroup_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 = 0;
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. per-memcg */
364 struct list_head memcg_slab_caches;
365 /* Not a spinlock, we can take a lot of time walking the list */
366 struct mutex slab_caches_mutex;
367 /* Index in the kmem_cache->memcg_params->memcg_caches array */
371 int last_scanned_node;
373 nodemask_t scan_nodes;
374 atomic_t numainfo_events;
375 atomic_t numainfo_updating;
378 /* List of events which userspace want to receive */
379 struct list_head event_list;
380 spinlock_t event_list_lock;
382 struct mem_cgroup_per_node *nodeinfo[0];
383 /* WARNING: nodeinfo must be the last member here */
386 /* internal only representation about the status of kmem accounting. */
388 KMEM_ACCOUNTED_ACTIVE, /* accounted by this cgroup itself */
389 KMEM_ACCOUNTED_DEAD, /* dead memcg with pending kmem charges */
392 #ifdef CONFIG_MEMCG_KMEM
393 static inline void memcg_kmem_set_active(struct mem_cgroup *memcg)
395 set_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
398 static bool memcg_kmem_is_active(struct mem_cgroup *memcg)
400 return test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
403 static void memcg_kmem_mark_dead(struct mem_cgroup *memcg)
406 * Our caller must use css_get() first, because memcg_uncharge_kmem()
407 * will call css_put() if it sees the memcg is dead.
410 if (test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags))
411 set_bit(KMEM_ACCOUNTED_DEAD, &memcg->kmem_account_flags);
414 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup *memcg)
416 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD,
417 &memcg->kmem_account_flags);
421 /* Stuffs for move charges at task migration. */
423 * Types of charges to be moved. "move_charge_at_immitgrate" and
424 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
427 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
428 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
432 /* "mc" and its members are protected by cgroup_mutex */
433 static struct move_charge_struct {
434 spinlock_t lock; /* for from, to */
435 struct mem_cgroup *from;
436 struct mem_cgroup *to;
437 unsigned long immigrate_flags;
438 unsigned long precharge;
439 unsigned long moved_charge;
440 unsigned long moved_swap;
441 struct task_struct *moving_task; /* a task moving charges */
442 wait_queue_head_t waitq; /* a waitq for other context */
444 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
445 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
448 static bool move_anon(void)
450 return test_bit(MOVE_CHARGE_TYPE_ANON, &mc.immigrate_flags);
453 static bool move_file(void)
455 return test_bit(MOVE_CHARGE_TYPE_FILE, &mc.immigrate_flags);
459 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
460 * limit reclaim to prevent infinite loops, if they ever occur.
462 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
463 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
466 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
467 MEM_CGROUP_CHARGE_TYPE_ANON,
468 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
469 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
473 /* for encoding cft->private value on file */
481 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
482 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
483 #define MEMFILE_ATTR(val) ((val) & 0xffff)
484 /* Used for OOM nofiier */
485 #define OOM_CONTROL (0)
488 * Reclaim flags for mem_cgroup_hierarchical_reclaim
490 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
491 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
492 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
493 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
496 * The memcg_create_mutex will be held whenever a new cgroup is created.
497 * As a consequence, any change that needs to protect against new child cgroups
498 * appearing has to hold it as well.
500 static DEFINE_MUTEX(memcg_create_mutex);
502 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
504 return s ? container_of(s, struct mem_cgroup, css) : NULL;
507 /* Some nice accessors for the vmpressure. */
508 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
511 memcg = root_mem_cgroup;
512 return &memcg->vmpressure;
515 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
517 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
520 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
522 return (memcg == root_mem_cgroup);
526 * We restrict the id in the range of [1, 65535], so it can fit into
529 #define MEM_CGROUP_ID_MAX USHRT_MAX
531 static inline unsigned short mem_cgroup_id(struct mem_cgroup *memcg)
534 * The ID of the root cgroup is 0, but memcg treat 0 as an
535 * invalid ID, so we return (cgroup_id + 1).
537 return memcg->css.cgroup->id + 1;
540 static inline struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
542 struct cgroup_subsys_state *css;
544 css = css_from_id(id - 1, &mem_cgroup_subsys);
545 return mem_cgroup_from_css(css);
548 /* Writing them here to avoid exposing memcg's inner layout */
549 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
551 void sock_update_memcg(struct sock *sk)
553 if (mem_cgroup_sockets_enabled) {
554 struct mem_cgroup *memcg;
555 struct cg_proto *cg_proto;
557 BUG_ON(!sk->sk_prot->proto_cgroup);
559 /* Socket cloning can throw us here with sk_cgrp already
560 * filled. It won't however, necessarily happen from
561 * process context. So the test for root memcg given
562 * the current task's memcg won't help us in this case.
564 * Respecting the original socket's memcg is a better
565 * decision in this case.
568 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
569 css_get(&sk->sk_cgrp->memcg->css);
574 memcg = mem_cgroup_from_task(current);
575 cg_proto = sk->sk_prot->proto_cgroup(memcg);
576 if (!mem_cgroup_is_root(memcg) &&
577 memcg_proto_active(cg_proto) && css_tryget(&memcg->css)) {
578 sk->sk_cgrp = cg_proto;
583 EXPORT_SYMBOL(sock_update_memcg);
585 void sock_release_memcg(struct sock *sk)
587 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
588 struct mem_cgroup *memcg;
589 WARN_ON(!sk->sk_cgrp->memcg);
590 memcg = sk->sk_cgrp->memcg;
591 css_put(&sk->sk_cgrp->memcg->css);
595 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
597 if (!memcg || mem_cgroup_is_root(memcg))
600 return &memcg->tcp_mem;
602 EXPORT_SYMBOL(tcp_proto_cgroup);
604 static void disarm_sock_keys(struct mem_cgroup *memcg)
606 if (!memcg_proto_activated(&memcg->tcp_mem))
608 static_key_slow_dec(&memcg_socket_limit_enabled);
611 static void disarm_sock_keys(struct mem_cgroup *memcg)
616 #ifdef CONFIG_MEMCG_KMEM
618 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
619 * The main reason for not using cgroup id for this:
620 * this works better in sparse environments, where we have a lot of memcgs,
621 * but only a few kmem-limited. Or also, if we have, for instance, 200
622 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
623 * 200 entry array for that.
625 * The current size of the caches array is stored in
626 * memcg_limited_groups_array_size. It will double each time we have to
629 static DEFINE_IDA(kmem_limited_groups);
630 int memcg_limited_groups_array_size;
633 * MIN_SIZE is different than 1, because we would like to avoid going through
634 * the alloc/free process all the time. In a small machine, 4 kmem-limited
635 * cgroups is a reasonable guess. In the future, it could be a parameter or
636 * tunable, but that is strictly not necessary.
638 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
639 * this constant directly from cgroup, but it is understandable that this is
640 * better kept as an internal representation in cgroup.c. In any case, the
641 * cgrp_id space is not getting any smaller, and we don't have to necessarily
642 * increase ours as well if it increases.
644 #define MEMCG_CACHES_MIN_SIZE 4
645 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
648 * A lot of the calls to the cache allocation functions are expected to be
649 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
650 * conditional to this static branch, we'll have to allow modules that does
651 * kmem_cache_alloc and the such to see this symbol as well
653 struct static_key memcg_kmem_enabled_key;
654 EXPORT_SYMBOL(memcg_kmem_enabled_key);
656 static void disarm_kmem_keys(struct mem_cgroup *memcg)
658 if (memcg_kmem_is_active(memcg)) {
659 static_key_slow_dec(&memcg_kmem_enabled_key);
660 ida_simple_remove(&kmem_limited_groups, memcg->kmemcg_id);
663 * This check can't live in kmem destruction function,
664 * since the charges will outlive the cgroup
666 WARN_ON(res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0);
669 static void disarm_kmem_keys(struct mem_cgroup *memcg)
672 #endif /* CONFIG_MEMCG_KMEM */
674 static void disarm_static_keys(struct mem_cgroup *memcg)
676 disarm_sock_keys(memcg);
677 disarm_kmem_keys(memcg);
680 static void drain_all_stock_async(struct mem_cgroup *memcg);
682 static struct mem_cgroup_per_zone *
683 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
685 VM_BUG_ON((unsigned)nid >= nr_node_ids);
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 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
697 int nid = page_to_nid(page);
698 int zid = page_zonenum(page);
700 return mem_cgroup_zoneinfo(memcg, nid, 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];
719 __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
720 struct mem_cgroup_per_zone *mz,
721 struct mem_cgroup_tree_per_zone *mctz,
722 unsigned long long new_usage_in_excess)
724 struct rb_node **p = &mctz->rb_root.rb_node;
725 struct rb_node *parent = NULL;
726 struct mem_cgroup_per_zone *mz_node;
731 mz->usage_in_excess = new_usage_in_excess;
732 if (!mz->usage_in_excess)
736 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
738 if (mz->usage_in_excess < mz_node->usage_in_excess)
741 * We can't avoid mem cgroups that are over their soft
742 * limit by the same amount
744 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
747 rb_link_node(&mz->tree_node, parent, p);
748 rb_insert_color(&mz->tree_node, &mctz->rb_root);
753 __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
754 struct mem_cgroup_per_zone *mz,
755 struct mem_cgroup_tree_per_zone *mctz)
759 rb_erase(&mz->tree_node, &mctz->rb_root);
764 mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
765 struct mem_cgroup_per_zone *mz,
766 struct mem_cgroup_tree_per_zone *mctz)
768 spin_lock(&mctz->lock);
769 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
770 spin_unlock(&mctz->lock);
774 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
776 unsigned long long excess;
777 struct mem_cgroup_per_zone *mz;
778 struct mem_cgroup_tree_per_zone *mctz;
779 int nid = page_to_nid(page);
780 int zid = page_zonenum(page);
781 mctz = soft_limit_tree_from_page(page);
784 * Necessary to update all ancestors when hierarchy is used.
785 * because their event counter is not touched.
787 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
788 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
789 excess = res_counter_soft_limit_excess(&memcg->res);
791 * We have to update the tree if mz is on RB-tree or
792 * mem is over its softlimit.
794 if (excess || mz->on_tree) {
795 spin_lock(&mctz->lock);
796 /* if on-tree, remove it */
798 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
800 * Insert again. mz->usage_in_excess will be updated.
801 * If excess is 0, no tree ops.
803 __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
804 spin_unlock(&mctz->lock);
809 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
812 struct mem_cgroup_per_zone *mz;
813 struct mem_cgroup_tree_per_zone *mctz;
815 for_each_node(node) {
816 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
817 mz = mem_cgroup_zoneinfo(memcg, node, zone);
818 mctz = soft_limit_tree_node_zone(node, zone);
819 mem_cgroup_remove_exceeded(memcg, mz, mctz);
824 static struct mem_cgroup_per_zone *
825 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
827 struct rb_node *rightmost = NULL;
828 struct mem_cgroup_per_zone *mz;
832 rightmost = rb_last(&mctz->rb_root);
834 goto done; /* Nothing to reclaim from */
836 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
838 * Remove the node now but someone else can add it back,
839 * we will to add it back at the end of reclaim to its correct
840 * position in the tree.
842 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
843 if (!res_counter_soft_limit_excess(&mz->memcg->res) ||
844 !css_tryget(&mz->memcg->css))
850 static struct mem_cgroup_per_zone *
851 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
853 struct mem_cgroup_per_zone *mz;
855 spin_lock(&mctz->lock);
856 mz = __mem_cgroup_largest_soft_limit_node(mctz);
857 spin_unlock(&mctz->lock);
862 * Implementation Note: reading percpu statistics for memcg.
864 * Both of vmstat[] and percpu_counter has threshold and do periodic
865 * synchronization to implement "quick" read. There are trade-off between
866 * reading cost and precision of value. Then, we may have a chance to implement
867 * a periodic synchronizion of counter in memcg's counter.
869 * But this _read() function is used for user interface now. The user accounts
870 * memory usage by memory cgroup and he _always_ requires exact value because
871 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
872 * have to visit all online cpus and make sum. So, for now, unnecessary
873 * synchronization is not implemented. (just implemented for cpu hotplug)
875 * If there are kernel internal actions which can make use of some not-exact
876 * value, and reading all cpu value can be performance bottleneck in some
877 * common workload, threashold and synchonization as vmstat[] should be
880 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
881 enum mem_cgroup_stat_index idx)
887 for_each_online_cpu(cpu)
888 val += per_cpu(memcg->stat->count[idx], cpu);
889 #ifdef CONFIG_HOTPLUG_CPU
890 spin_lock(&memcg->pcp_counter_lock);
891 val += memcg->nocpu_base.count[idx];
892 spin_unlock(&memcg->pcp_counter_lock);
898 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
901 int val = (charge) ? 1 : -1;
902 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
905 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
906 enum mem_cgroup_events_index idx)
908 unsigned long val = 0;
912 for_each_online_cpu(cpu)
913 val += per_cpu(memcg->stat->events[idx], cpu);
914 #ifdef CONFIG_HOTPLUG_CPU
915 spin_lock(&memcg->pcp_counter_lock);
916 val += memcg->nocpu_base.events[idx];
917 spin_unlock(&memcg->pcp_counter_lock);
923 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
925 bool anon, int nr_pages)
930 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
931 * counted as CACHE even if it's on ANON LRU.
934 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
937 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
940 if (PageTransHuge(page))
941 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
944 /* pagein of a big page is an event. So, ignore page size */
946 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
948 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
949 nr_pages = -nr_pages; /* for event */
952 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
958 mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
960 struct mem_cgroup_per_zone *mz;
962 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
963 return mz->lru_size[lru];
967 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
968 unsigned int lru_mask)
970 struct mem_cgroup_per_zone *mz;
972 unsigned long ret = 0;
974 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
977 if (BIT(lru) & lru_mask)
978 ret += mz->lru_size[lru];
984 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
985 int nid, unsigned int lru_mask)
990 for (zid = 0; zid < MAX_NR_ZONES; zid++)
991 total += mem_cgroup_zone_nr_lru_pages(memcg,
997 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
998 unsigned int lru_mask)
1003 for_each_node_state(nid, N_MEMORY)
1004 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
1008 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
1009 enum mem_cgroup_events_target target)
1011 unsigned long val, next;
1013 val = __this_cpu_read(memcg->stat->nr_page_events);
1014 next = __this_cpu_read(memcg->stat->targets[target]);
1015 /* from time_after() in jiffies.h */
1016 if ((long)next - (long)val < 0) {
1018 case MEM_CGROUP_TARGET_THRESH:
1019 next = val + THRESHOLDS_EVENTS_TARGET;
1021 case MEM_CGROUP_TARGET_SOFTLIMIT:
1022 next = val + SOFTLIMIT_EVENTS_TARGET;
1024 case MEM_CGROUP_TARGET_NUMAINFO:
1025 next = val + NUMAINFO_EVENTS_TARGET;
1030 __this_cpu_write(memcg->stat->targets[target], next);
1037 * Check events in order.
1040 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
1043 /* threshold event is triggered in finer grain than soft limit */
1044 if (unlikely(mem_cgroup_event_ratelimit(memcg,
1045 MEM_CGROUP_TARGET_THRESH))) {
1047 bool do_numainfo __maybe_unused;
1049 do_softlimit = mem_cgroup_event_ratelimit(memcg,
1050 MEM_CGROUP_TARGET_SOFTLIMIT);
1051 #if MAX_NUMNODES > 1
1052 do_numainfo = mem_cgroup_event_ratelimit(memcg,
1053 MEM_CGROUP_TARGET_NUMAINFO);
1057 mem_cgroup_threshold(memcg);
1058 if (unlikely(do_softlimit))
1059 mem_cgroup_update_tree(memcg, page);
1060 #if MAX_NUMNODES > 1
1061 if (unlikely(do_numainfo))
1062 atomic_inc(&memcg->numainfo_events);
1068 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1071 * mm_update_next_owner() may clear mm->owner to NULL
1072 * if it races with swapoff, page migration, etc.
1073 * So this can be called with p == NULL.
1078 return mem_cgroup_from_css(task_css(p, mem_cgroup_subsys_id));
1081 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
1083 struct mem_cgroup *memcg = NULL;
1088 * Because we have no locks, mm->owner's may be being moved to other
1089 * cgroup. We use css_tryget() here even if this looks
1090 * pessimistic (rather than adding locks here).
1094 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1095 if (unlikely(!memcg))
1097 } while (!css_tryget(&memcg->css));
1103 * Returns a next (in a pre-order walk) alive memcg (with elevated css
1104 * ref. count) or NULL if the whole root's subtree has been visited.
1106 * helper function to be used by mem_cgroup_iter
1108 static struct mem_cgroup *__mem_cgroup_iter_next(struct mem_cgroup *root,
1109 struct mem_cgroup *last_visited)
1111 struct cgroup_subsys_state *prev_css, *next_css;
1113 prev_css = last_visited ? &last_visited->css : NULL;
1115 next_css = css_next_descendant_pre(prev_css, &root->css);
1118 * Even if we found a group we have to make sure it is
1119 * alive. css && !memcg means that the groups should be
1120 * skipped and we should continue the tree walk.
1121 * last_visited css is safe to use because it is
1122 * protected by css_get and the tree walk is rcu safe.
1124 * We do not take a reference on the root of the tree walk
1125 * because we might race with the root removal when it would
1126 * be the only node in the iterated hierarchy and mem_cgroup_iter
1127 * would end up in an endless loop because it expects that at
1128 * least one valid node will be returned. Root cannot disappear
1129 * because caller of the iterator should hold it already so
1130 * skipping css reference should be safe.
1133 struct mem_cgroup *memcg = mem_cgroup_from_css(next_css);
1135 if (next_css == &root->css)
1138 if (css_tryget(next_css)) {
1140 * Make sure the memcg is initialized:
1141 * mem_cgroup_css_online() orders the the
1142 * initialization against setting the flag.
1144 if (smp_load_acquire(&memcg->initialized))
1149 prev_css = next_css;
1156 static void mem_cgroup_iter_invalidate(struct mem_cgroup *root)
1159 * When a group in the hierarchy below root is destroyed, the
1160 * hierarchy iterator can no longer be trusted since it might
1161 * have pointed to the destroyed group. Invalidate it.
1163 atomic_inc(&root->dead_count);
1166 static struct mem_cgroup *
1167 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter *iter,
1168 struct mem_cgroup *root,
1171 struct mem_cgroup *position = NULL;
1173 * A cgroup destruction happens in two stages: offlining and
1174 * release. They are separated by a RCU grace period.
1176 * If the iterator is valid, we may still race with an
1177 * offlining. The RCU lock ensures the object won't be
1178 * released, tryget will fail if we lost the race.
1180 *sequence = atomic_read(&root->dead_count);
1181 if (iter->last_dead_count == *sequence) {
1183 position = iter->last_visited;
1186 * We cannot take a reference to root because we might race
1187 * with root removal and returning NULL would end up in
1188 * an endless loop on the iterator user level when root
1189 * would be returned all the time.
1191 if (position && position != root &&
1192 !css_tryget(&position->css))
1198 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter *iter,
1199 struct mem_cgroup *last_visited,
1200 struct mem_cgroup *new_position,
1201 struct mem_cgroup *root,
1204 /* root reference counting symmetric to mem_cgroup_iter_load */
1205 if (last_visited && last_visited != root)
1206 css_put(&last_visited->css);
1208 * We store the sequence count from the time @last_visited was
1209 * loaded successfully instead of rereading it here so that we
1210 * don't lose destruction events in between. We could have
1211 * raced with the destruction of @new_position after all.
1213 iter->last_visited = new_position;
1215 iter->last_dead_count = sequence;
1219 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1220 * @root: hierarchy root
1221 * @prev: previously returned memcg, NULL on first invocation
1222 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1224 * Returns references to children of the hierarchy below @root, or
1225 * @root itself, or %NULL after a full round-trip.
1227 * Caller must pass the return value in @prev on subsequent
1228 * invocations for reference counting, or use mem_cgroup_iter_break()
1229 * to cancel a hierarchy walk before the round-trip is complete.
1231 * Reclaimers can specify a zone and a priority level in @reclaim to
1232 * divide up the memcgs in the hierarchy among all concurrent
1233 * reclaimers operating on the same zone and priority.
1235 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1236 struct mem_cgroup *prev,
1237 struct mem_cgroup_reclaim_cookie *reclaim)
1239 struct mem_cgroup *memcg = NULL;
1240 struct mem_cgroup *last_visited = NULL;
1242 if (mem_cgroup_disabled())
1246 root = root_mem_cgroup;
1248 if (prev && !reclaim)
1249 last_visited = prev;
1251 if (!root->use_hierarchy && root != root_mem_cgroup) {
1259 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1260 int uninitialized_var(seq);
1263 int nid = zone_to_nid(reclaim->zone);
1264 int zid = zone_idx(reclaim->zone);
1265 struct mem_cgroup_per_zone *mz;
1267 mz = mem_cgroup_zoneinfo(root, nid, zid);
1268 iter = &mz->reclaim_iter[reclaim->priority];
1269 if (prev && reclaim->generation != iter->generation) {
1270 iter->last_visited = NULL;
1274 last_visited = mem_cgroup_iter_load(iter, root, &seq);
1277 memcg = __mem_cgroup_iter_next(root, last_visited);
1280 mem_cgroup_iter_update(iter, last_visited, memcg, root,
1285 else if (!prev && memcg)
1286 reclaim->generation = iter->generation;
1295 if (prev && prev != root)
1296 css_put(&prev->css);
1302 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1303 * @root: hierarchy root
1304 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1306 void mem_cgroup_iter_break(struct mem_cgroup *root,
1307 struct mem_cgroup *prev)
1310 root = root_mem_cgroup;
1311 if (prev && prev != root)
1312 css_put(&prev->css);
1316 * Iteration constructs for visiting all cgroups (under a tree). If
1317 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1318 * be used for reference counting.
1320 #define for_each_mem_cgroup_tree(iter, root) \
1321 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1323 iter = mem_cgroup_iter(root, iter, NULL))
1325 #define for_each_mem_cgroup(iter) \
1326 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1328 iter = mem_cgroup_iter(NULL, iter, NULL))
1330 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1332 struct mem_cgroup *memcg;
1335 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1336 if (unlikely(!memcg))
1341 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1344 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1352 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1355 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1356 * @zone: zone of the wanted lruvec
1357 * @memcg: memcg of the wanted lruvec
1359 * Returns the lru list vector holding pages for the given @zone and
1360 * @mem. This can be the global zone lruvec, if the memory controller
1363 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1364 struct mem_cgroup *memcg)
1366 struct mem_cgroup_per_zone *mz;
1367 struct lruvec *lruvec;
1369 if (mem_cgroup_disabled()) {
1370 lruvec = &zone->lruvec;
1374 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1375 lruvec = &mz->lruvec;
1378 * Since a node can be onlined after the mem_cgroup was created,
1379 * we have to be prepared to initialize lruvec->zone here;
1380 * and if offlined then reonlined, we need to reinitialize it.
1382 if (unlikely(lruvec->zone != zone))
1383 lruvec->zone = zone;
1388 * Following LRU functions are allowed to be used without PCG_LOCK.
1389 * Operations are called by routine of global LRU independently from memcg.
1390 * What we have to take care of here is validness of pc->mem_cgroup.
1392 * Changes to pc->mem_cgroup happens when
1395 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1396 * It is added to LRU before charge.
1397 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1398 * When moving account, the page is not on LRU. It's isolated.
1402 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1404 * @zone: zone of the page
1406 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1408 struct mem_cgroup_per_zone *mz;
1409 struct mem_cgroup *memcg;
1410 struct page_cgroup *pc;
1411 struct lruvec *lruvec;
1413 if (mem_cgroup_disabled()) {
1414 lruvec = &zone->lruvec;
1418 pc = lookup_page_cgroup(page);
1419 memcg = pc->mem_cgroup;
1422 * Surreptitiously switch any uncharged offlist page to root:
1423 * an uncharged page off lru does nothing to secure
1424 * its former mem_cgroup from sudden removal.
1426 * Our caller holds lru_lock, and PageCgroupUsed is updated
1427 * under page_cgroup lock: between them, they make all uses
1428 * of pc->mem_cgroup safe.
1430 if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1431 pc->mem_cgroup = memcg = root_mem_cgroup;
1433 mz = page_cgroup_zoneinfo(memcg, page);
1434 lruvec = &mz->lruvec;
1437 * Since a node can be onlined after the mem_cgroup was created,
1438 * we have to be prepared to initialize lruvec->zone here;
1439 * and if offlined then reonlined, we need to reinitialize it.
1441 if (unlikely(lruvec->zone != zone))
1442 lruvec->zone = zone;
1447 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1448 * @lruvec: mem_cgroup per zone lru vector
1449 * @lru: index of lru list the page is sitting on
1450 * @nr_pages: positive when adding or negative when removing
1452 * This function must be called when a page is added to or removed from an
1455 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1458 struct mem_cgroup_per_zone *mz;
1459 unsigned long *lru_size;
1461 if (mem_cgroup_disabled())
1464 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1465 lru_size = mz->lru_size + lru;
1466 *lru_size += nr_pages;
1467 VM_BUG_ON((long)(*lru_size) < 0);
1471 * Checks whether given mem is same or in the root_mem_cgroup's
1474 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1475 struct mem_cgroup *memcg)
1477 if (root_memcg == memcg)
1479 if (!root_memcg->use_hierarchy || !memcg)
1481 return cgroup_is_descendant(memcg->css.cgroup, root_memcg->css.cgroup);
1484 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1485 struct mem_cgroup *memcg)
1490 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1495 bool task_in_mem_cgroup(struct task_struct *task,
1496 const struct mem_cgroup *memcg)
1498 struct mem_cgroup *curr = NULL;
1499 struct task_struct *p;
1502 p = find_lock_task_mm(task);
1504 curr = try_get_mem_cgroup_from_mm(p->mm);
1508 * All threads may have already detached their mm's, but the oom
1509 * killer still needs to detect if they have already been oom
1510 * killed to prevent needlessly killing additional tasks.
1513 curr = mem_cgroup_from_task(task);
1515 css_get(&curr->css);
1521 * We should check use_hierarchy of "memcg" not "curr". Because checking
1522 * use_hierarchy of "curr" here make this function true if hierarchy is
1523 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1524 * hierarchy(even if use_hierarchy is disabled in "memcg").
1526 ret = mem_cgroup_same_or_subtree(memcg, curr);
1527 css_put(&curr->css);
1531 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1533 unsigned long inactive_ratio;
1534 unsigned long inactive;
1535 unsigned long active;
1538 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1539 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1541 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1543 inactive_ratio = int_sqrt(10 * gb);
1547 return inactive * inactive_ratio < active;
1550 #define mem_cgroup_from_res_counter(counter, member) \
1551 container_of(counter, struct mem_cgroup, member)
1554 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1555 * @memcg: the memory cgroup
1557 * Returns the maximum amount of memory @mem can be charged with, in
1560 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1562 unsigned long long margin;
1564 margin = res_counter_margin(&memcg->res);
1565 if (do_swap_account)
1566 margin = min(margin, res_counter_margin(&memcg->memsw));
1567 return margin >> PAGE_SHIFT;
1570 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1573 if (!css_parent(&memcg->css))
1574 return vm_swappiness;
1576 return memcg->swappiness;
1580 * memcg->moving_account is used for checking possibility that some thread is
1581 * calling move_account(). When a thread on CPU-A starts moving pages under
1582 * a memcg, other threads should check memcg->moving_account under
1583 * rcu_read_lock(), like this:
1587 * memcg->moving_account+1 if (memcg->mocing_account)
1589 * synchronize_rcu() update something.
1594 /* for quick checking without looking up memcg */
1595 atomic_t memcg_moving __read_mostly;
1597 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1599 atomic_inc(&memcg_moving);
1600 atomic_inc(&memcg->moving_account);
1604 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1607 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1608 * We check NULL in callee rather than caller.
1611 atomic_dec(&memcg_moving);
1612 atomic_dec(&memcg->moving_account);
1617 * 2 routines for checking "mem" is under move_account() or not.
1619 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1620 * is used for avoiding races in accounting. If true,
1621 * pc->mem_cgroup may be overwritten.
1623 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1624 * under hierarchy of moving cgroups. This is for
1625 * waiting at hith-memory prressure caused by "move".
1628 static bool mem_cgroup_stolen(struct mem_cgroup *memcg)
1630 VM_BUG_ON(!rcu_read_lock_held());
1631 return atomic_read(&memcg->moving_account) > 0;
1634 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1636 struct mem_cgroup *from;
1637 struct mem_cgroup *to;
1640 * Unlike task_move routines, we access mc.to, mc.from not under
1641 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1643 spin_lock(&mc.lock);
1649 ret = mem_cgroup_same_or_subtree(memcg, from)
1650 || mem_cgroup_same_or_subtree(memcg, to);
1652 spin_unlock(&mc.lock);
1656 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1658 if (mc.moving_task && current != mc.moving_task) {
1659 if (mem_cgroup_under_move(memcg)) {
1661 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1662 /* moving charge context might have finished. */
1665 finish_wait(&mc.waitq, &wait);
1673 * Take this lock when
1674 * - a code tries to modify page's memcg while it's USED.
1675 * - a code tries to modify page state accounting in a memcg.
1676 * see mem_cgroup_stolen(), too.
1678 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1679 unsigned long *flags)
1681 spin_lock_irqsave(&memcg->move_lock, *flags);
1684 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1685 unsigned long *flags)
1687 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1690 #define K(x) ((x) << (PAGE_SHIFT-10))
1692 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1693 * @memcg: The memory cgroup that went over limit
1694 * @p: Task that is going to be killed
1696 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1699 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1702 * protects memcg_name and makes sure that parallel ooms do not
1705 static DEFINE_MUTEX(oom_info_lock);
1706 struct cgroup *task_cgrp;
1707 struct cgroup *mem_cgrp;
1708 static char memcg_name[PATH_MAX];
1710 struct mem_cgroup *iter;
1716 mutex_lock(&oom_info_lock);
1719 mem_cgrp = memcg->css.cgroup;
1720 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1722 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1725 * Unfortunately, we are unable to convert to a useful name
1726 * But we'll still print out the usage information
1733 pr_info("Task in %s killed", memcg_name);
1736 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1744 * Continues from above, so we don't need an KERN_ level
1746 pr_cont(" as a result of limit of %s\n", memcg_name);
1749 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1750 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1751 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1752 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1753 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1754 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1755 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1756 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1757 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1758 res_counter_read_u64(&memcg->kmem, RES_USAGE) >> 10,
1759 res_counter_read_u64(&memcg->kmem, RES_LIMIT) >> 10,
1760 res_counter_read_u64(&memcg->kmem, RES_FAILCNT));
1762 for_each_mem_cgroup_tree(iter, memcg) {
1763 pr_info("Memory cgroup stats");
1766 ret = cgroup_path(iter->css.cgroup, memcg_name, PATH_MAX);
1768 pr_cont(" for %s", memcg_name);
1772 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1773 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1775 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1776 K(mem_cgroup_read_stat(iter, i)));
1779 for (i = 0; i < NR_LRU_LISTS; i++)
1780 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1781 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1785 mutex_unlock(&oom_info_lock);
1789 * This function returns the number of memcg under hierarchy tree. Returns
1790 * 1(self count) if no children.
1792 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1795 struct mem_cgroup *iter;
1797 for_each_mem_cgroup_tree(iter, memcg)
1803 * Return the memory (and swap, if configured) limit for a memcg.
1805 static u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1809 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1812 * Do not consider swap space if we cannot swap due to swappiness
1814 if (mem_cgroup_swappiness(memcg)) {
1817 limit += total_swap_pages << PAGE_SHIFT;
1818 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1821 * If memsw is finite and limits the amount of swap space
1822 * available to this memcg, return that limit.
1824 limit = min(limit, memsw);
1830 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1833 struct mem_cgroup *iter;
1834 unsigned long chosen_points = 0;
1835 unsigned long totalpages;
1836 unsigned int points = 0;
1837 struct task_struct *chosen = NULL;
1840 * If current has a pending SIGKILL or is exiting, then automatically
1841 * select it. The goal is to allow it to allocate so that it may
1842 * quickly exit and free its memory.
1844 if (fatal_signal_pending(current) || current->flags & PF_EXITING) {
1845 set_thread_flag(TIF_MEMDIE);
1849 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1850 totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
1851 for_each_mem_cgroup_tree(iter, memcg) {
1852 struct css_task_iter it;
1853 struct task_struct *task;
1855 css_task_iter_start(&iter->css, &it);
1856 while ((task = css_task_iter_next(&it))) {
1857 switch (oom_scan_process_thread(task, totalpages, NULL,
1859 case OOM_SCAN_SELECT:
1861 put_task_struct(chosen);
1863 chosen_points = ULONG_MAX;
1864 get_task_struct(chosen);
1866 case OOM_SCAN_CONTINUE:
1868 case OOM_SCAN_ABORT:
1869 css_task_iter_end(&it);
1870 mem_cgroup_iter_break(memcg, iter);
1872 put_task_struct(chosen);
1877 points = oom_badness(task, memcg, NULL, totalpages);
1878 if (!points || points < chosen_points)
1880 /* Prefer thread group leaders for display purposes */
1881 if (points == chosen_points &&
1882 thread_group_leader(chosen))
1886 put_task_struct(chosen);
1888 chosen_points = points;
1889 get_task_struct(chosen);
1891 css_task_iter_end(&it);
1896 points = chosen_points * 1000 / totalpages;
1897 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1898 NULL, "Memory cgroup out of memory");
1901 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1903 unsigned long flags)
1905 unsigned long total = 0;
1906 bool noswap = false;
1909 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1911 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1914 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1916 drain_all_stock_async(memcg);
1917 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1919 * Allow limit shrinkers, which are triggered directly
1920 * by userspace, to catch signals and stop reclaim
1921 * after minimal progress, regardless of the margin.
1923 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1925 if (mem_cgroup_margin(memcg))
1928 * If nothing was reclaimed after two attempts, there
1929 * may be no reclaimable pages in this hierarchy.
1938 * test_mem_cgroup_node_reclaimable
1939 * @memcg: the target memcg
1940 * @nid: the node ID to be checked.
1941 * @noswap : specify true here if the user wants flle only information.
1943 * This function returns whether the specified memcg contains any
1944 * reclaimable pages on a node. Returns true if there are any reclaimable
1945 * pages in the node.
1947 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1948 int nid, bool noswap)
1950 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1952 if (noswap || !total_swap_pages)
1954 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1959 #if MAX_NUMNODES > 1
1962 * Always updating the nodemask is not very good - even if we have an empty
1963 * list or the wrong list here, we can start from some node and traverse all
1964 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1967 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1971 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1972 * pagein/pageout changes since the last update.
1974 if (!atomic_read(&memcg->numainfo_events))
1976 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1979 /* make a nodemask where this memcg uses memory from */
1980 memcg->scan_nodes = node_states[N_MEMORY];
1982 for_each_node_mask(nid, node_states[N_MEMORY]) {
1984 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1985 node_clear(nid, memcg->scan_nodes);
1988 atomic_set(&memcg->numainfo_events, 0);
1989 atomic_set(&memcg->numainfo_updating, 0);
1993 * Selecting a node where we start reclaim from. Because what we need is just
1994 * reducing usage counter, start from anywhere is O,K. Considering
1995 * memory reclaim from current node, there are pros. and cons.
1997 * Freeing memory from current node means freeing memory from a node which
1998 * we'll use or we've used. So, it may make LRU bad. And if several threads
1999 * hit limits, it will see a contention on a node. But freeing from remote
2000 * node means more costs for memory reclaim because of memory latency.
2002 * Now, we use round-robin. Better algorithm is welcomed.
2004 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
2008 mem_cgroup_may_update_nodemask(memcg);
2009 node = memcg->last_scanned_node;
2011 node = next_node(node, memcg->scan_nodes);
2012 if (node == MAX_NUMNODES)
2013 node = first_node(memcg->scan_nodes);
2015 * We call this when we hit limit, not when pages are added to LRU.
2016 * No LRU may hold pages because all pages are UNEVICTABLE or
2017 * memcg is too small and all pages are not on LRU. In that case,
2018 * we use curret node.
2020 if (unlikely(node == MAX_NUMNODES))
2021 node = numa_node_id();
2023 memcg->last_scanned_node = node;
2028 * Check all nodes whether it contains reclaimable pages or not.
2029 * For quick scan, we make use of scan_nodes. This will allow us to skip
2030 * unused nodes. But scan_nodes is lazily updated and may not cotain
2031 * enough new information. We need to do double check.
2033 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
2038 * quick check...making use of scan_node.
2039 * We can skip unused nodes.
2041 if (!nodes_empty(memcg->scan_nodes)) {
2042 for (nid = first_node(memcg->scan_nodes);
2044 nid = next_node(nid, memcg->scan_nodes)) {
2046 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
2051 * Check rest of nodes.
2053 for_each_node_state(nid, N_MEMORY) {
2054 if (node_isset(nid, memcg->scan_nodes))
2056 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
2063 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
2068 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
2070 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
2074 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
2077 unsigned long *total_scanned)
2079 struct mem_cgroup *victim = NULL;
2082 unsigned long excess;
2083 unsigned long nr_scanned;
2084 struct mem_cgroup_reclaim_cookie reclaim = {
2089 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
2092 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
2097 * If we have not been able to reclaim
2098 * anything, it might because there are
2099 * no reclaimable pages under this hierarchy
2104 * We want to do more targeted reclaim.
2105 * excess >> 2 is not to excessive so as to
2106 * reclaim too much, nor too less that we keep
2107 * coming back to reclaim from this cgroup
2109 if (total >= (excess >> 2) ||
2110 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
2115 if (!mem_cgroup_reclaimable(victim, false))
2117 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
2119 *total_scanned += nr_scanned;
2120 if (!res_counter_soft_limit_excess(&root_memcg->res))
2123 mem_cgroup_iter_break(root_memcg, victim);
2127 #ifdef CONFIG_LOCKDEP
2128 static struct lockdep_map memcg_oom_lock_dep_map = {
2129 .name = "memcg_oom_lock",
2133 static DEFINE_SPINLOCK(memcg_oom_lock);
2136 * Check OOM-Killer is already running under our hierarchy.
2137 * If someone is running, return false.
2139 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
2141 struct mem_cgroup *iter, *failed = NULL;
2143 spin_lock(&memcg_oom_lock);
2145 for_each_mem_cgroup_tree(iter, memcg) {
2146 if (iter->oom_lock) {
2148 * this subtree of our hierarchy is already locked
2149 * so we cannot give a lock.
2152 mem_cgroup_iter_break(memcg, iter);
2155 iter->oom_lock = true;
2160 * OK, we failed to lock the whole subtree so we have
2161 * to clean up what we set up to the failing subtree
2163 for_each_mem_cgroup_tree(iter, memcg) {
2164 if (iter == failed) {
2165 mem_cgroup_iter_break(memcg, iter);
2168 iter->oom_lock = false;
2171 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
2173 spin_unlock(&memcg_oom_lock);
2178 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
2180 struct mem_cgroup *iter;
2182 spin_lock(&memcg_oom_lock);
2183 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
2184 for_each_mem_cgroup_tree(iter, memcg)
2185 iter->oom_lock = false;
2186 spin_unlock(&memcg_oom_lock);
2189 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
2191 struct mem_cgroup *iter;
2193 for_each_mem_cgroup_tree(iter, memcg)
2194 atomic_inc(&iter->under_oom);
2197 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
2199 struct mem_cgroup *iter;
2202 * When a new child is created while the hierarchy is under oom,
2203 * mem_cgroup_oom_lock() may not be called. We have to use
2204 * atomic_add_unless() here.
2206 for_each_mem_cgroup_tree(iter, memcg)
2207 atomic_add_unless(&iter->under_oom, -1, 0);
2210 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
2212 struct oom_wait_info {
2213 struct mem_cgroup *memcg;
2217 static int memcg_oom_wake_function(wait_queue_t *wait,
2218 unsigned mode, int sync, void *arg)
2220 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
2221 struct mem_cgroup *oom_wait_memcg;
2222 struct oom_wait_info *oom_wait_info;
2224 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
2225 oom_wait_memcg = oom_wait_info->memcg;
2228 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2229 * Then we can use css_is_ancestor without taking care of RCU.
2231 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
2232 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
2234 return autoremove_wake_function(wait, mode, sync, arg);
2237 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
2239 atomic_inc(&memcg->oom_wakeups);
2240 /* for filtering, pass "memcg" as argument. */
2241 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
2244 static void memcg_oom_recover(struct mem_cgroup *memcg)
2246 if (memcg && atomic_read(&memcg->under_oom))
2247 memcg_wakeup_oom(memcg);
2250 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
2252 if (!current->memcg_oom.may_oom)
2255 * We are in the middle of the charge context here, so we
2256 * don't want to block when potentially sitting on a callstack
2257 * that holds all kinds of filesystem and mm locks.
2259 * Also, the caller may handle a failed allocation gracefully
2260 * (like optional page cache readahead) and so an OOM killer
2261 * invocation might not even be necessary.
2263 * That's why we don't do anything here except remember the
2264 * OOM context and then deal with it at the end of the page
2265 * fault when the stack is unwound, the locks are released,
2266 * and when we know whether the fault was overall successful.
2268 css_get(&memcg->css);
2269 current->memcg_oom.memcg = memcg;
2270 current->memcg_oom.gfp_mask = mask;
2271 current->memcg_oom.order = order;
2275 * mem_cgroup_oom_synchronize - complete memcg OOM handling
2276 * @handle: actually kill/wait or just clean up the OOM state
2278 * This has to be called at the end of a page fault if the memcg OOM
2279 * handler was enabled.
2281 * Memcg supports userspace OOM handling where failed allocations must
2282 * sleep on a waitqueue until the userspace task resolves the
2283 * situation. Sleeping directly in the charge context with all kinds
2284 * of locks held is not a good idea, instead we remember an OOM state
2285 * in the task and mem_cgroup_oom_synchronize() has to be called at
2286 * the end of the page fault to complete the OOM handling.
2288 * Returns %true if an ongoing memcg OOM situation was detected and
2289 * completed, %false otherwise.
2291 bool mem_cgroup_oom_synchronize(bool handle)
2293 struct mem_cgroup *memcg = current->memcg_oom.memcg;
2294 struct oom_wait_info owait;
2297 /* OOM is global, do not handle */
2304 owait.memcg = memcg;
2305 owait.wait.flags = 0;
2306 owait.wait.func = memcg_oom_wake_function;
2307 owait.wait.private = current;
2308 INIT_LIST_HEAD(&owait.wait.task_list);
2310 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2311 mem_cgroup_mark_under_oom(memcg);
2313 locked = mem_cgroup_oom_trylock(memcg);
2316 mem_cgroup_oom_notify(memcg);
2318 if (locked && !memcg->oom_kill_disable) {
2319 mem_cgroup_unmark_under_oom(memcg);
2320 finish_wait(&memcg_oom_waitq, &owait.wait);
2321 mem_cgroup_out_of_memory(memcg, current->memcg_oom.gfp_mask,
2322 current->memcg_oom.order);
2325 mem_cgroup_unmark_under_oom(memcg);
2326 finish_wait(&memcg_oom_waitq, &owait.wait);
2330 mem_cgroup_oom_unlock(memcg);
2332 * There is no guarantee that an OOM-lock contender
2333 * sees the wakeups triggered by the OOM kill
2334 * uncharges. Wake any sleepers explicitely.
2336 memcg_oom_recover(memcg);
2339 current->memcg_oom.memcg = NULL;
2340 css_put(&memcg->css);
2345 * Currently used to update mapped file statistics, but the routine can be
2346 * generalized to update other statistics as well.
2348 * Notes: Race condition
2350 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2351 * it tends to be costly. But considering some conditions, we doesn't need
2352 * to do so _always_.
2354 * Considering "charge", lock_page_cgroup() is not required because all
2355 * file-stat operations happen after a page is attached to radix-tree. There
2356 * are no race with "charge".
2358 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2359 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2360 * if there are race with "uncharge". Statistics itself is properly handled
2363 * Considering "move", this is an only case we see a race. To make the race
2364 * small, we check mm->moving_account and detect there are possibility of race
2365 * If there is, we take a lock.
2368 void __mem_cgroup_begin_update_page_stat(struct page *page,
2369 bool *locked, unsigned long *flags)
2371 struct mem_cgroup *memcg;
2372 struct page_cgroup *pc;
2374 pc = lookup_page_cgroup(page);
2376 memcg = pc->mem_cgroup;
2377 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2380 * If this memory cgroup is not under account moving, we don't
2381 * need to take move_lock_mem_cgroup(). Because we already hold
2382 * rcu_read_lock(), any calls to move_account will be delayed until
2383 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2385 if (!mem_cgroup_stolen(memcg))
2388 move_lock_mem_cgroup(memcg, flags);
2389 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2390 move_unlock_mem_cgroup(memcg, flags);
2396 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
2398 struct page_cgroup *pc = lookup_page_cgroup(page);
2401 * It's guaranteed that pc->mem_cgroup never changes while
2402 * lock is held because a routine modifies pc->mem_cgroup
2403 * should take move_lock_mem_cgroup().
2405 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
2408 void mem_cgroup_update_page_stat(struct page *page,
2409 enum mem_cgroup_stat_index idx, int val)
2411 struct mem_cgroup *memcg;
2412 struct page_cgroup *pc = lookup_page_cgroup(page);
2413 unsigned long uninitialized_var(flags);
2415 if (mem_cgroup_disabled())
2418 VM_BUG_ON(!rcu_read_lock_held());
2419 memcg = pc->mem_cgroup;
2420 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2423 this_cpu_add(memcg->stat->count[idx], val);
2427 * size of first charge trial. "32" comes from vmscan.c's magic value.
2428 * TODO: maybe necessary to use big numbers in big irons.
2430 #define CHARGE_BATCH 32U
2431 struct memcg_stock_pcp {
2432 struct mem_cgroup *cached; /* this never be root cgroup */
2433 unsigned int nr_pages;
2434 struct work_struct work;
2435 unsigned long flags;
2436 #define FLUSHING_CACHED_CHARGE 0
2438 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2439 static DEFINE_MUTEX(percpu_charge_mutex);
2442 * consume_stock: Try to consume stocked charge on this cpu.
2443 * @memcg: memcg to consume from.
2444 * @nr_pages: how many pages to charge.
2446 * The charges will only happen if @memcg matches the current cpu's memcg
2447 * stock, and at least @nr_pages are available in that stock. Failure to
2448 * service an allocation will refill the stock.
2450 * returns true if successful, false otherwise.
2452 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2454 struct memcg_stock_pcp *stock;
2457 if (nr_pages > CHARGE_BATCH)
2460 stock = &get_cpu_var(memcg_stock);
2461 if (memcg == stock->cached && stock->nr_pages >= nr_pages)
2462 stock->nr_pages -= nr_pages;
2463 else /* need to call res_counter_charge */
2465 put_cpu_var(memcg_stock);
2470 * Returns stocks cached in percpu to res_counter and reset cached information.
2472 static void drain_stock(struct memcg_stock_pcp *stock)
2474 struct mem_cgroup *old = stock->cached;
2476 if (stock->nr_pages) {
2477 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2479 res_counter_uncharge(&old->res, bytes);
2480 if (do_swap_account)
2481 res_counter_uncharge(&old->memsw, bytes);
2482 stock->nr_pages = 0;
2484 stock->cached = NULL;
2488 * This must be called under preempt disabled or must be called by
2489 * a thread which is pinned to local cpu.
2491 static void drain_local_stock(struct work_struct *dummy)
2493 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2495 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2498 static void __init memcg_stock_init(void)
2502 for_each_possible_cpu(cpu) {
2503 struct memcg_stock_pcp *stock =
2504 &per_cpu(memcg_stock, cpu);
2505 INIT_WORK(&stock->work, drain_local_stock);
2510 * Cache charges(val) which is from res_counter, to local per_cpu area.
2511 * This will be consumed by consume_stock() function, later.
2513 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2515 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2517 if (stock->cached != memcg) { /* reset if necessary */
2519 stock->cached = memcg;
2521 stock->nr_pages += nr_pages;
2522 put_cpu_var(memcg_stock);
2526 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2527 * of the hierarchy under it. sync flag says whether we should block
2528 * until the work is done.
2530 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2534 /* Notify other cpus that system-wide "drain" is running */
2537 for_each_online_cpu(cpu) {
2538 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2539 struct mem_cgroup *memcg;
2541 memcg = stock->cached;
2542 if (!memcg || !stock->nr_pages)
2544 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2546 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2548 drain_local_stock(&stock->work);
2550 schedule_work_on(cpu, &stock->work);
2558 for_each_online_cpu(cpu) {
2559 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2560 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2561 flush_work(&stock->work);
2568 * Tries to drain stocked charges in other cpus. This function is asynchronous
2569 * and just put a work per cpu for draining localy on each cpu. Caller can
2570 * expects some charges will be back to res_counter later but cannot wait for
2573 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2576 * If someone calls draining, avoid adding more kworker runs.
2578 if (!mutex_trylock(&percpu_charge_mutex))
2580 drain_all_stock(root_memcg, false);
2581 mutex_unlock(&percpu_charge_mutex);
2584 /* This is a synchronous drain interface. */
2585 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2587 /* called when force_empty is called */
2588 mutex_lock(&percpu_charge_mutex);
2589 drain_all_stock(root_memcg, true);
2590 mutex_unlock(&percpu_charge_mutex);
2594 * This function drains percpu counter value from DEAD cpu and
2595 * move it to local cpu. Note that this function can be preempted.
2597 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2601 spin_lock(&memcg->pcp_counter_lock);
2602 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2603 long x = per_cpu(memcg->stat->count[i], cpu);
2605 per_cpu(memcg->stat->count[i], cpu) = 0;
2606 memcg->nocpu_base.count[i] += x;
2608 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2609 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2611 per_cpu(memcg->stat->events[i], cpu) = 0;
2612 memcg->nocpu_base.events[i] += x;
2614 spin_unlock(&memcg->pcp_counter_lock);
2617 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
2618 unsigned long action,
2621 int cpu = (unsigned long)hcpu;
2622 struct memcg_stock_pcp *stock;
2623 struct mem_cgroup *iter;
2625 if (action == CPU_ONLINE)
2628 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2631 for_each_mem_cgroup(iter)
2632 mem_cgroup_drain_pcp_counter(iter, cpu);
2634 stock = &per_cpu(memcg_stock, cpu);
2640 /* See __mem_cgroup_try_charge() for details */
2642 CHARGE_OK, /* success */
2643 CHARGE_RETRY, /* need to retry but retry is not bad */
2644 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2645 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2648 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2649 unsigned int nr_pages, unsigned int min_pages,
2652 unsigned long csize = nr_pages * PAGE_SIZE;
2653 struct mem_cgroup *mem_over_limit;
2654 struct res_counter *fail_res;
2655 unsigned long flags = 0;
2658 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2661 if (!do_swap_account)
2663 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2667 res_counter_uncharge(&memcg->res, csize);
2668 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2669 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2671 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2673 * Never reclaim on behalf of optional batching, retry with a
2674 * single page instead.
2676 if (nr_pages > min_pages)
2677 return CHARGE_RETRY;
2679 if (!(gfp_mask & __GFP_WAIT))
2680 return CHARGE_WOULDBLOCK;
2682 if (gfp_mask & __GFP_NORETRY)
2683 return CHARGE_NOMEM;
2685 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2686 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2687 return CHARGE_RETRY;
2689 * Even though the limit is exceeded at this point, reclaim
2690 * may have been able to free some pages. Retry the charge
2691 * before killing the task.
2693 * Only for regular pages, though: huge pages are rather
2694 * unlikely to succeed so close to the limit, and we fall back
2695 * to regular pages anyway in case of failure.
2697 if (nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER) && ret)
2698 return CHARGE_RETRY;
2701 * At task move, charge accounts can be doubly counted. So, it's
2702 * better to wait until the end of task_move if something is going on.
2704 if (mem_cgroup_wait_acct_move(mem_over_limit))
2705 return CHARGE_RETRY;
2708 mem_cgroup_oom(mem_over_limit, gfp_mask, get_order(csize));
2710 return CHARGE_NOMEM;
2714 * __mem_cgroup_try_charge() does
2715 * 1. detect memcg to be charged against from passed *mm and *ptr,
2716 * 2. update res_counter
2717 * 3. call memory reclaim if necessary.
2719 * In some special case, if the task is fatal, fatal_signal_pending() or
2720 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2721 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2722 * as possible without any hazards. 2: all pages should have a valid
2723 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2724 * pointer, that is treated as a charge to root_mem_cgroup.
2726 * So __mem_cgroup_try_charge() will return
2727 * 0 ... on success, filling *ptr with a valid memcg pointer.
2728 * -ENOMEM ... charge failure because of resource limits.
2729 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2731 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2732 * the oom-killer can be invoked.
2734 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2736 unsigned int nr_pages,
2737 struct mem_cgroup **ptr,
2740 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2741 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2742 struct mem_cgroup *memcg = NULL;
2746 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2747 * in system level. So, allow to go ahead dying process in addition to
2750 if (unlikely(test_thread_flag(TIF_MEMDIE)
2751 || fatal_signal_pending(current)))
2754 if (unlikely(task_in_memcg_oom(current)))
2757 if (gfp_mask & __GFP_NOFAIL)
2761 * We always charge the cgroup the mm_struct belongs to.
2762 * The mm_struct's mem_cgroup changes on task migration if the
2763 * thread group leader migrates. It's possible that mm is not
2764 * set, if so charge the root memcg (happens for pagecache usage).
2767 *ptr = root_mem_cgroup;
2769 if (*ptr) { /* css should be a valid one */
2771 if (mem_cgroup_is_root(memcg))
2773 if (consume_stock(memcg, nr_pages))
2775 css_get(&memcg->css);
2777 struct task_struct *p;
2780 p = rcu_dereference(mm->owner);
2782 * Because we don't have task_lock(), "p" can exit.
2783 * In that case, "memcg" can point to root or p can be NULL with
2784 * race with swapoff. Then, we have small risk of mis-accouning.
2785 * But such kind of mis-account by race always happens because
2786 * we don't have cgroup_mutex(). It's overkill and we allo that
2788 * (*) swapoff at el will charge against mm-struct not against
2789 * task-struct. So, mm->owner can be NULL.
2791 memcg = mem_cgroup_from_task(p);
2793 memcg = root_mem_cgroup;
2794 if (mem_cgroup_is_root(memcg)) {
2798 if (consume_stock(memcg, nr_pages)) {
2800 * It seems dagerous to access memcg without css_get().
2801 * But considering how consume_stok works, it's not
2802 * necessary. If consume_stock success, some charges
2803 * from this memcg are cached on this cpu. So, we
2804 * don't need to call css_get()/css_tryget() before
2805 * calling consume_stock().
2810 /* after here, we may be blocked. we need to get refcnt */
2811 if (!css_tryget(&memcg->css)) {
2819 bool invoke_oom = oom && !nr_oom_retries;
2821 /* If killed, bypass charge */
2822 if (fatal_signal_pending(current)) {
2823 css_put(&memcg->css);
2827 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch,
2828 nr_pages, invoke_oom);
2832 case CHARGE_RETRY: /* not in OOM situation but retry */
2834 css_put(&memcg->css);
2837 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2838 css_put(&memcg->css);
2840 case CHARGE_NOMEM: /* OOM routine works */
2841 if (!oom || invoke_oom) {
2842 css_put(&memcg->css);
2848 } while (ret != CHARGE_OK);
2850 if (batch > nr_pages)
2851 refill_stock(memcg, batch - nr_pages);
2852 css_put(&memcg->css);
2857 if (!(gfp_mask & __GFP_NOFAIL)) {
2862 *ptr = root_mem_cgroup;
2867 * Somemtimes we have to undo a charge we got by try_charge().
2868 * This function is for that and do uncharge, put css's refcnt.
2869 * gotten by try_charge().
2871 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2872 unsigned int nr_pages)
2874 if (!mem_cgroup_is_root(memcg)) {
2875 unsigned long bytes = nr_pages * PAGE_SIZE;
2877 res_counter_uncharge(&memcg->res, bytes);
2878 if (do_swap_account)
2879 res_counter_uncharge(&memcg->memsw, bytes);
2884 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2885 * This is useful when moving usage to parent cgroup.
2887 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2888 unsigned int nr_pages)
2890 unsigned long bytes = nr_pages * PAGE_SIZE;
2892 if (mem_cgroup_is_root(memcg))
2895 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2896 if (do_swap_account)
2897 res_counter_uncharge_until(&memcg->memsw,
2898 memcg->memsw.parent, bytes);
2902 * A helper function to get mem_cgroup from ID. must be called under
2903 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2904 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2905 * called against removed memcg.)
2907 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2909 /* ID 0 is unused ID */
2912 return mem_cgroup_from_id(id);
2915 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2917 struct mem_cgroup *memcg = NULL;
2918 struct page_cgroup *pc;
2922 VM_BUG_ON_PAGE(!PageLocked(page), page);
2924 pc = lookup_page_cgroup(page);
2925 lock_page_cgroup(pc);
2926 if (PageCgroupUsed(pc)) {
2927 memcg = pc->mem_cgroup;
2928 if (memcg && !css_tryget(&memcg->css))
2930 } else if (PageSwapCache(page)) {
2931 ent.val = page_private(page);
2932 id = lookup_swap_cgroup_id(ent);
2934 memcg = mem_cgroup_lookup(id);
2935 if (memcg && !css_tryget(&memcg->css))
2939 unlock_page_cgroup(pc);
2943 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2945 unsigned int nr_pages,
2946 enum charge_type ctype,
2949 struct page_cgroup *pc = lookup_page_cgroup(page);
2950 struct zone *uninitialized_var(zone);
2951 struct lruvec *lruvec;
2952 bool was_on_lru = false;
2955 lock_page_cgroup(pc);
2956 VM_BUG_ON_PAGE(PageCgroupUsed(pc), page);
2958 * we don't need page_cgroup_lock about tail pages, becase they are not
2959 * accessed by any other context at this point.
2963 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2964 * may already be on some other mem_cgroup's LRU. Take care of it.
2967 zone = page_zone(page);
2968 spin_lock_irq(&zone->lru_lock);
2969 if (PageLRU(page)) {
2970 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2972 del_page_from_lru_list(page, lruvec, page_lru(page));
2977 pc->mem_cgroup = memcg;
2979 * We access a page_cgroup asynchronously without lock_page_cgroup().
2980 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2981 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2982 * before USED bit, we need memory barrier here.
2983 * See mem_cgroup_add_lru_list(), etc.
2986 SetPageCgroupUsed(pc);
2990 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2991 VM_BUG_ON_PAGE(PageLRU(page), page);
2993 add_page_to_lru_list(page, lruvec, page_lru(page));
2995 spin_unlock_irq(&zone->lru_lock);
2998 if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
3003 mem_cgroup_charge_statistics(memcg, page, anon, nr_pages);
3004 unlock_page_cgroup(pc);
3007 * "charge_statistics" updated event counter. Then, check it.
3008 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
3009 * if they exceeds softlimit.
3011 memcg_check_events(memcg, page);
3014 static DEFINE_MUTEX(set_limit_mutex);
3016 #ifdef CONFIG_MEMCG_KMEM
3017 static DEFINE_MUTEX(activate_kmem_mutex);
3019 static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg)
3021 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) &&
3022 memcg_kmem_is_active(memcg);
3026 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
3027 * in the memcg_cache_params struct.
3029 static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
3031 struct kmem_cache *cachep;
3033 VM_BUG_ON(p->is_root_cache);
3034 cachep = p->root_cache;
3035 return cache_from_memcg_idx(cachep, memcg_cache_id(p->memcg));
3038 #ifdef CONFIG_SLABINFO
3039 static int mem_cgroup_slabinfo_read(struct seq_file *m, void *v)
3041 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3042 struct memcg_cache_params *params;
3044 if (!memcg_can_account_kmem(memcg))
3047 print_slabinfo_header(m);
3049 mutex_lock(&memcg->slab_caches_mutex);
3050 list_for_each_entry(params, &memcg->memcg_slab_caches, list)
3051 cache_show(memcg_params_to_cache(params), m);
3052 mutex_unlock(&memcg->slab_caches_mutex);
3058 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size)
3060 struct res_counter *fail_res;
3061 struct mem_cgroup *_memcg;
3064 ret = res_counter_charge(&memcg->kmem, size, &fail_res);
3069 ret = __mem_cgroup_try_charge(NULL, gfp, size >> PAGE_SHIFT,
3070 &_memcg, oom_gfp_allowed(gfp));
3072 if (ret == -EINTR) {
3074 * __mem_cgroup_try_charge() chosed to bypass to root due to
3075 * OOM kill or fatal signal. Since our only options are to
3076 * either fail the allocation or charge it to this cgroup, do
3077 * it as a temporary condition. But we can't fail. From a
3078 * kmem/slab perspective, the cache has already been selected,
3079 * by mem_cgroup_kmem_get_cache(), so it is too late to change
3082 * This condition will only trigger if the task entered
3083 * memcg_charge_kmem in a sane state, but was OOM-killed during
3084 * __mem_cgroup_try_charge() above. Tasks that were already
3085 * dying when the allocation triggers should have been already
3086 * directed to the root cgroup in memcontrol.h
3088 res_counter_charge_nofail(&memcg->res, size, &fail_res);
3089 if (do_swap_account)
3090 res_counter_charge_nofail(&memcg->memsw, size,
3094 res_counter_uncharge(&memcg->kmem, size);
3099 static void memcg_uncharge_kmem(struct mem_cgroup *memcg, u64 size)
3101 res_counter_uncharge(&memcg->res, size);
3102 if (do_swap_account)
3103 res_counter_uncharge(&memcg->memsw, size);
3106 if (res_counter_uncharge(&memcg->kmem, size))
3110 * Releases a reference taken in kmem_cgroup_css_offline in case
3111 * this last uncharge is racing with the offlining code or it is
3112 * outliving the memcg existence.
3114 * The memory barrier imposed by test&clear is paired with the
3115 * explicit one in memcg_kmem_mark_dead().
3117 if (memcg_kmem_test_and_clear_dead(memcg))
3118 css_put(&memcg->css);
3122 * helper for acessing a memcg's index. It will be used as an index in the
3123 * child cache array in kmem_cache, and also to derive its name. This function
3124 * will return -1 when this is not a kmem-limited memcg.
3126 int memcg_cache_id(struct mem_cgroup *memcg)
3128 return memcg ? memcg->kmemcg_id : -1;
3131 static size_t memcg_caches_array_size(int num_groups)
3134 if (num_groups <= 0)
3137 size = 2 * num_groups;
3138 if (size < MEMCG_CACHES_MIN_SIZE)
3139 size = MEMCG_CACHES_MIN_SIZE;
3140 else if (size > MEMCG_CACHES_MAX_SIZE)
3141 size = MEMCG_CACHES_MAX_SIZE;
3147 * We should update the current array size iff all caches updates succeed. This
3148 * can only be done from the slab side. The slab mutex needs to be held when
3151 void memcg_update_array_size(int num)
3153 if (num > memcg_limited_groups_array_size)
3154 memcg_limited_groups_array_size = memcg_caches_array_size(num);
3157 static void kmem_cache_destroy_work_func(struct work_struct *w);
3159 int memcg_update_cache_size(struct kmem_cache *s, int num_groups)
3161 struct memcg_cache_params *cur_params = s->memcg_params;
3163 VM_BUG_ON(!is_root_cache(s));
3165 if (num_groups > memcg_limited_groups_array_size) {
3167 struct memcg_cache_params *new_params;
3168 ssize_t size = memcg_caches_array_size(num_groups);
3170 size *= sizeof(void *);
3171 size += offsetof(struct memcg_cache_params, memcg_caches);
3173 new_params = kzalloc(size, GFP_KERNEL);
3177 new_params->is_root_cache = true;
3180 * There is the chance it will be bigger than
3181 * memcg_limited_groups_array_size, if we failed an allocation
3182 * in a cache, in which case all caches updated before it, will
3183 * have a bigger array.
3185 * But if that is the case, the data after
3186 * memcg_limited_groups_array_size is certainly unused
3188 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3189 if (!cur_params->memcg_caches[i])
3191 new_params->memcg_caches[i] =
3192 cur_params->memcg_caches[i];
3196 * Ideally, we would wait until all caches succeed, and only
3197 * then free the old one. But this is not worth the extra
3198 * pointer per-cache we'd have to have for this.
3200 * It is not a big deal if some caches are left with a size
3201 * bigger than the others. And all updates will reset this
3204 rcu_assign_pointer(s->memcg_params, new_params);
3206 kfree_rcu(cur_params, rcu_head);
3211 int memcg_alloc_cache_params(struct mem_cgroup *memcg, struct kmem_cache *s,
3212 struct kmem_cache *root_cache)
3216 if (!memcg_kmem_enabled())
3220 size = offsetof(struct memcg_cache_params, memcg_caches);
3221 size += memcg_limited_groups_array_size * sizeof(void *);
3223 size = sizeof(struct memcg_cache_params);
3225 s->memcg_params = kzalloc(size, GFP_KERNEL);
3226 if (!s->memcg_params)
3230 s->memcg_params->memcg = memcg;
3231 s->memcg_params->root_cache = root_cache;
3232 INIT_WORK(&s->memcg_params->destroy,
3233 kmem_cache_destroy_work_func);
3235 s->memcg_params->is_root_cache = true;
3240 void memcg_free_cache_params(struct kmem_cache *s)
3242 kfree(s->memcg_params);
3245 void memcg_register_cache(struct kmem_cache *s)
3247 struct kmem_cache *root;
3248 struct mem_cgroup *memcg;
3251 if (is_root_cache(s))
3255 * Holding the slab_mutex assures nobody will touch the memcg_caches
3256 * array while we are modifying it.
3258 lockdep_assert_held(&slab_mutex);
3260 root = s->memcg_params->root_cache;
3261 memcg = s->memcg_params->memcg;
3262 id = memcg_cache_id(memcg);
3264 css_get(&memcg->css);
3268 * Since readers won't lock (see cache_from_memcg_idx()), we need a
3269 * barrier here to ensure nobody will see the kmem_cache partially
3275 * Initialize the pointer to this cache in its parent's memcg_params
3276 * before adding it to the memcg_slab_caches list, otherwise we can
3277 * fail to convert memcg_params_to_cache() while traversing the list.
3279 VM_BUG_ON(root->memcg_params->memcg_caches[id]);
3280 root->memcg_params->memcg_caches[id] = s;
3282 mutex_lock(&memcg->slab_caches_mutex);
3283 list_add(&s->memcg_params->list, &memcg->memcg_slab_caches);
3284 mutex_unlock(&memcg->slab_caches_mutex);
3287 void memcg_unregister_cache(struct kmem_cache *s)
3289 struct kmem_cache *root;
3290 struct mem_cgroup *memcg;
3293 if (is_root_cache(s))
3297 * Holding the slab_mutex assures nobody will touch the memcg_caches
3298 * array while we are modifying it.
3300 lockdep_assert_held(&slab_mutex);
3302 root = s->memcg_params->root_cache;
3303 memcg = s->memcg_params->memcg;
3304 id = memcg_cache_id(memcg);
3306 mutex_lock(&memcg->slab_caches_mutex);
3307 list_del(&s->memcg_params->list);
3308 mutex_unlock(&memcg->slab_caches_mutex);
3311 * Clear the pointer to this cache in its parent's memcg_params only
3312 * after removing it from the memcg_slab_caches list, otherwise we can
3313 * fail to convert memcg_params_to_cache() while traversing the list.
3315 VM_BUG_ON(!root->memcg_params->memcg_caches[id]);
3316 root->memcg_params->memcg_caches[id] = NULL;
3318 css_put(&memcg->css);
3322 * During the creation a new cache, we need to disable our accounting mechanism
3323 * altogether. This is true even if we are not creating, but rather just
3324 * enqueing new caches to be created.
3326 * This is because that process will trigger allocations; some visible, like
3327 * explicit kmallocs to auxiliary data structures, name strings and internal
3328 * cache structures; some well concealed, like INIT_WORK() that can allocate
3329 * objects during debug.
3331 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3332 * to it. This may not be a bounded recursion: since the first cache creation
3333 * failed to complete (waiting on the allocation), we'll just try to create the
3334 * cache again, failing at the same point.
3336 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3337 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3338 * inside the following two functions.
3340 static inline void memcg_stop_kmem_account(void)
3342 VM_BUG_ON(!current->mm);
3343 current->memcg_kmem_skip_account++;
3346 static inline void memcg_resume_kmem_account(void)
3348 VM_BUG_ON(!current->mm);
3349 current->memcg_kmem_skip_account--;
3352 static void kmem_cache_destroy_work_func(struct work_struct *w)
3354 struct kmem_cache *cachep;
3355 struct memcg_cache_params *p;
3357 p = container_of(w, struct memcg_cache_params, destroy);
3359 cachep = memcg_params_to_cache(p);
3362 * If we get down to 0 after shrink, we could delete right away.
3363 * However, memcg_release_pages() already puts us back in the workqueue
3364 * in that case. If we proceed deleting, we'll get a dangling
3365 * reference, and removing the object from the workqueue in that case
3366 * is unnecessary complication. We are not a fast path.
3368 * Note that this case is fundamentally different from racing with
3369 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3370 * kmem_cache_shrink, not only we would be reinserting a dead cache
3371 * into the queue, but doing so from inside the worker racing to
3374 * So if we aren't down to zero, we'll just schedule a worker and try
3377 if (atomic_read(&cachep->memcg_params->nr_pages) != 0)
3378 kmem_cache_shrink(cachep);
3380 kmem_cache_destroy(cachep);
3383 void mem_cgroup_destroy_cache(struct kmem_cache *cachep)
3385 if (!cachep->memcg_params->dead)
3389 * There are many ways in which we can get here.
3391 * We can get to a memory-pressure situation while the delayed work is
3392 * still pending to run. The vmscan shrinkers can then release all
3393 * cache memory and get us to destruction. If this is the case, we'll
3394 * be executed twice, which is a bug (the second time will execute over
3395 * bogus data). In this case, cancelling the work should be fine.
3397 * But we can also get here from the worker itself, if
3398 * kmem_cache_shrink is enough to shake all the remaining objects and
3399 * get the page count to 0. In this case, we'll deadlock if we try to
3400 * cancel the work (the worker runs with an internal lock held, which
3401 * is the same lock we would hold for cancel_work_sync().)
3403 * Since we can't possibly know who got us here, just refrain from
3404 * running if there is already work pending
3406 if (work_pending(&cachep->memcg_params->destroy))
3409 * We have to defer the actual destroying to a workqueue, because
3410 * we might currently be in a context that cannot sleep.
3412 schedule_work(&cachep->memcg_params->destroy);
3415 static struct kmem_cache *memcg_create_kmem_cache(struct mem_cgroup *memcg,
3416 struct kmem_cache *s)
3418 struct kmem_cache *new = NULL;
3419 static char *tmp_name = NULL;
3420 static DEFINE_MUTEX(mutex); /* protects tmp_name */
3422 BUG_ON(!memcg_can_account_kmem(memcg));
3426 * kmem_cache_create_memcg duplicates the given name and
3427 * cgroup_name for this name requires RCU context.
3428 * This static temporary buffer is used to prevent from
3429 * pointless shortliving allocation.
3432 tmp_name = kmalloc(PATH_MAX, GFP_KERNEL);
3438 snprintf(tmp_name, PATH_MAX, "%s(%d:%s)", s->name,
3439 memcg_cache_id(memcg), cgroup_name(memcg->css.cgroup));
3442 new = kmem_cache_create_memcg(memcg, tmp_name, s->object_size, s->align,
3443 (s->flags & ~SLAB_PANIC), s->ctor, s);
3445 new->allocflags |= __GFP_KMEMCG;
3449 mutex_unlock(&mutex);
3453 void kmem_cache_destroy_memcg_children(struct kmem_cache *s)
3455 struct kmem_cache *c;
3458 if (!s->memcg_params)
3460 if (!s->memcg_params->is_root_cache)
3464 * If the cache is being destroyed, we trust that there is no one else
3465 * requesting objects from it. Even if there are, the sanity checks in
3466 * kmem_cache_destroy should caught this ill-case.
3468 * Still, we don't want anyone else freeing memcg_caches under our
3469 * noses, which can happen if a new memcg comes to life. As usual,
3470 * we'll take the activate_kmem_mutex to protect ourselves against
3473 mutex_lock(&activate_kmem_mutex);
3474 for_each_memcg_cache_index(i) {
3475 c = cache_from_memcg_idx(s, i);
3480 * We will now manually delete the caches, so to avoid races
3481 * we need to cancel all pending destruction workers and
3482 * proceed with destruction ourselves.
3484 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3485 * and that could spawn the workers again: it is likely that
3486 * the cache still have active pages until this very moment.
3487 * This would lead us back to mem_cgroup_destroy_cache.
3489 * But that will not execute at all if the "dead" flag is not
3490 * set, so flip it down to guarantee we are in control.
3492 c->memcg_params->dead = false;
3493 cancel_work_sync(&c->memcg_params->destroy);
3494 kmem_cache_destroy(c);
3496 mutex_unlock(&activate_kmem_mutex);
3499 struct create_work {
3500 struct mem_cgroup *memcg;
3501 struct kmem_cache *cachep;
3502 struct work_struct work;
3505 static void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3507 struct kmem_cache *cachep;
3508 struct memcg_cache_params *params;
3510 if (!memcg_kmem_is_active(memcg))
3513 mutex_lock(&memcg->slab_caches_mutex);
3514 list_for_each_entry(params, &memcg->memcg_slab_caches, list) {
3515 cachep = memcg_params_to_cache(params);
3516 cachep->memcg_params->dead = true;
3517 schedule_work(&cachep->memcg_params->destroy);
3519 mutex_unlock(&memcg->slab_caches_mutex);
3522 static void memcg_create_cache_work_func(struct work_struct *w)
3524 struct create_work *cw;
3526 cw = container_of(w, struct create_work, work);
3527 memcg_create_kmem_cache(cw->memcg, cw->cachep);
3528 css_put(&cw->memcg->css);
3533 * Enqueue the creation of a per-memcg kmem_cache.
3535 static void __memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3536 struct kmem_cache *cachep)
3538 struct create_work *cw;
3540 cw = kmalloc(sizeof(struct create_work), GFP_NOWAIT);
3542 css_put(&memcg->css);
3547 cw->cachep = cachep;
3549 INIT_WORK(&cw->work, memcg_create_cache_work_func);
3550 schedule_work(&cw->work);
3553 static void memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3554 struct kmem_cache *cachep)
3557 * We need to stop accounting when we kmalloc, because if the
3558 * corresponding kmalloc cache is not yet created, the first allocation
3559 * in __memcg_create_cache_enqueue will recurse.
3561 * However, it is better to enclose the whole function. Depending on
3562 * the debugging options enabled, INIT_WORK(), for instance, can
3563 * trigger an allocation. This too, will make us recurse. Because at
3564 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3565 * the safest choice is to do it like this, wrapping the whole function.
3567 memcg_stop_kmem_account();
3568 __memcg_create_cache_enqueue(memcg, cachep);
3569 memcg_resume_kmem_account();
3572 * Return the kmem_cache we're supposed to use for a slab allocation.
3573 * We try to use the current memcg's version of the cache.
3575 * If the cache does not exist yet, if we are the first user of it,
3576 * we either create it immediately, if possible, or create it asynchronously
3578 * In the latter case, we will let the current allocation go through with
3579 * the original cache.
3581 * Can't be called in interrupt context or from kernel threads.
3582 * This function needs to be called with rcu_read_lock() held.
3584 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
3587 struct mem_cgroup *memcg;
3588 struct kmem_cache *memcg_cachep;
3590 VM_BUG_ON(!cachep->memcg_params);
3591 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
3593 if (!current->mm || current->memcg_kmem_skip_account)
3597 memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
3599 if (!memcg_can_account_kmem(memcg))
3602 memcg_cachep = cache_from_memcg_idx(cachep, memcg_cache_id(memcg));
3603 if (likely(memcg_cachep)) {
3604 cachep = memcg_cachep;
3608 /* The corresponding put will be done in the workqueue. */
3609 if (!css_tryget(&memcg->css))
3614 * If we are in a safe context (can wait, and not in interrupt
3615 * context), we could be be predictable and return right away.
3616 * This would guarantee that the allocation being performed
3617 * already belongs in the new cache.
3619 * However, there are some clashes that can arrive from locking.
3620 * For instance, because we acquire the slab_mutex while doing
3621 * kmem_cache_dup, this means no further allocation could happen
3622 * with the slab_mutex held.
3624 * Also, because cache creation issue get_online_cpus(), this
3625 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3626 * that ends up reversed during cpu hotplug. (cpuset allocates
3627 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3628 * better to defer everything.
3630 memcg_create_cache_enqueue(memcg, cachep);
3636 EXPORT_SYMBOL(__memcg_kmem_get_cache);
3639 * We need to verify if the allocation against current->mm->owner's memcg is
3640 * possible for the given order. But the page is not allocated yet, so we'll
3641 * need a further commit step to do the final arrangements.
3643 * It is possible for the task to switch cgroups in this mean time, so at
3644 * commit time, we can't rely on task conversion any longer. We'll then use
3645 * the handle argument to return to the caller which cgroup we should commit
3646 * against. We could also return the memcg directly and avoid the pointer
3647 * passing, but a boolean return value gives better semantics considering
3648 * the compiled-out case as well.
3650 * Returning true means the allocation is possible.
3653 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3655 struct mem_cgroup *memcg;
3661 * Disabling accounting is only relevant for some specific memcg
3662 * internal allocations. Therefore we would initially not have such
3663 * check here, since direct calls to the page allocator that are marked
3664 * with GFP_KMEMCG only happen outside memcg core. We are mostly
3665 * concerned with cache allocations, and by having this test at
3666 * memcg_kmem_get_cache, we are already able to relay the allocation to
3667 * the root cache and bypass the memcg cache altogether.
3669 * There is one exception, though: the SLUB allocator does not create
3670 * large order caches, but rather service large kmallocs directly from
3671 * the page allocator. Therefore, the following sequence when backed by
3672 * the SLUB allocator:
3674 * memcg_stop_kmem_account();
3675 * kmalloc(<large_number>)
3676 * memcg_resume_kmem_account();
3678 * would effectively ignore the fact that we should skip accounting,
3679 * since it will drive us directly to this function without passing
3680 * through the cache selector memcg_kmem_get_cache. Such large
3681 * allocations are extremely rare but can happen, for instance, for the
3682 * cache arrays. We bring this test here.
3684 if (!current->mm || current->memcg_kmem_skip_account)
3687 memcg = try_get_mem_cgroup_from_mm(current->mm);
3690 * very rare case described in mem_cgroup_from_task. Unfortunately there
3691 * isn't much we can do without complicating this too much, and it would
3692 * be gfp-dependent anyway. Just let it go
3694 if (unlikely(!memcg))
3697 if (!memcg_can_account_kmem(memcg)) {
3698 css_put(&memcg->css);
3702 ret = memcg_charge_kmem(memcg, gfp, PAGE_SIZE << order);
3706 css_put(&memcg->css);
3710 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3713 struct page_cgroup *pc;
3715 VM_BUG_ON(mem_cgroup_is_root(memcg));
3717 /* The page allocation failed. Revert */
3719 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3723 pc = lookup_page_cgroup(page);
3724 lock_page_cgroup(pc);
3725 pc->mem_cgroup = memcg;
3726 SetPageCgroupUsed(pc);
3727 unlock_page_cgroup(pc);
3730 void __memcg_kmem_uncharge_pages(struct page *page, int order)
3732 struct mem_cgroup *memcg = NULL;
3733 struct page_cgroup *pc;
3736 pc = lookup_page_cgroup(page);
3738 * Fast unlocked return. Theoretically might have changed, have to
3739 * check again after locking.
3741 if (!PageCgroupUsed(pc))
3744 lock_page_cgroup(pc);
3745 if (PageCgroupUsed(pc)) {
3746 memcg = pc->mem_cgroup;
3747 ClearPageCgroupUsed(pc);
3749 unlock_page_cgroup(pc);
3752 * We trust that only if there is a memcg associated with the page, it
3753 * is a valid allocation
3758 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
3759 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3762 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3765 #endif /* CONFIG_MEMCG_KMEM */
3767 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3769 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3771 * Because tail pages are not marked as "used", set it. We're under
3772 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3773 * charge/uncharge will be never happen and move_account() is done under
3774 * compound_lock(), so we don't have to take care of races.
3776 void mem_cgroup_split_huge_fixup(struct page *head)
3778 struct page_cgroup *head_pc = lookup_page_cgroup(head);
3779 struct page_cgroup *pc;
3780 struct mem_cgroup *memcg;
3783 if (mem_cgroup_disabled())
3786 memcg = head_pc->mem_cgroup;
3787 for (i = 1; i < HPAGE_PMD_NR; i++) {
3789 pc->mem_cgroup = memcg;
3790 smp_wmb();/* see __commit_charge() */
3791 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
3793 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
3796 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3799 void mem_cgroup_move_account_page_stat(struct mem_cgroup *from,
3800 struct mem_cgroup *to,
3801 unsigned int nr_pages,
3802 enum mem_cgroup_stat_index idx)
3804 /* Update stat data for mem_cgroup */
3806 __this_cpu_sub(from->stat->count[idx], nr_pages);
3807 __this_cpu_add(to->stat->count[idx], nr_pages);
3812 * mem_cgroup_move_account - move account of the page
3814 * @nr_pages: number of regular pages (>1 for huge pages)
3815 * @pc: page_cgroup of the page.
3816 * @from: mem_cgroup which the page is moved from.
3817 * @to: mem_cgroup which the page is moved to. @from != @to.
3819 * The caller must confirm following.
3820 * - page is not on LRU (isolate_page() is useful.)
3821 * - compound_lock is held when nr_pages > 1
3823 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3826 static int mem_cgroup_move_account(struct page *page,
3827 unsigned int nr_pages,
3828 struct page_cgroup *pc,
3829 struct mem_cgroup *from,
3830 struct mem_cgroup *to)
3832 unsigned long flags;
3834 bool anon = PageAnon(page);
3836 VM_BUG_ON(from == to);
3837 VM_BUG_ON_PAGE(PageLRU(page), page);
3839 * The page is isolated from LRU. So, collapse function
3840 * will not handle this page. But page splitting can happen.
3841 * Do this check under compound_page_lock(). The caller should
3845 if (nr_pages > 1 && !PageTransHuge(page))
3848 lock_page_cgroup(pc);
3851 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3854 move_lock_mem_cgroup(from, &flags);
3856 if (!anon && page_mapped(page))
3857 mem_cgroup_move_account_page_stat(from, to, nr_pages,
3858 MEM_CGROUP_STAT_FILE_MAPPED);
3860 if (PageWriteback(page))
3861 mem_cgroup_move_account_page_stat(from, to, nr_pages,
3862 MEM_CGROUP_STAT_WRITEBACK);
3864 mem_cgroup_charge_statistics(from, page, anon, -nr_pages);
3866 /* caller should have done css_get */
3867 pc->mem_cgroup = to;
3868 mem_cgroup_charge_statistics(to, page, anon, nr_pages);
3869 move_unlock_mem_cgroup(from, &flags);
3872 unlock_page_cgroup(pc);
3876 memcg_check_events(to, page);
3877 memcg_check_events(from, page);
3883 * mem_cgroup_move_parent - moves page to the parent group
3884 * @page: the page to move
3885 * @pc: page_cgroup of the page
3886 * @child: page's cgroup
3888 * move charges to its parent or the root cgroup if the group has no
3889 * parent (aka use_hierarchy==0).
3890 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3891 * mem_cgroup_move_account fails) the failure is always temporary and
3892 * it signals a race with a page removal/uncharge or migration. In the
3893 * first case the page is on the way out and it will vanish from the LRU
3894 * on the next attempt and the call should be retried later.
3895 * Isolation from the LRU fails only if page has been isolated from
3896 * the LRU since we looked at it and that usually means either global
3897 * reclaim or migration going on. The page will either get back to the
3899 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3900 * (!PageCgroupUsed) or moved to a different group. The page will
3901 * disappear in the next attempt.
3903 static int mem_cgroup_move_parent(struct page *page,
3904 struct page_cgroup *pc,
3905 struct mem_cgroup *child)
3907 struct mem_cgroup *parent;
3908 unsigned int nr_pages;
3909 unsigned long uninitialized_var(flags);
3912 VM_BUG_ON(mem_cgroup_is_root(child));
3915 if (!get_page_unless_zero(page))
3917 if (isolate_lru_page(page))
3920 nr_pages = hpage_nr_pages(page);
3922 parent = parent_mem_cgroup(child);
3924 * If no parent, move charges to root cgroup.
3927 parent = root_mem_cgroup;
3930 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
3931 flags = compound_lock_irqsave(page);
3934 ret = mem_cgroup_move_account(page, nr_pages,
3937 __mem_cgroup_cancel_local_charge(child, nr_pages);
3940 compound_unlock_irqrestore(page, flags);
3941 putback_lru_page(page);
3949 * Charge the memory controller for page usage.
3951 * 0 if the charge was successful
3952 * < 0 if the cgroup is over its limit
3954 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
3955 gfp_t gfp_mask, enum charge_type ctype)
3957 struct mem_cgroup *memcg = NULL;
3958 unsigned int nr_pages = 1;
3962 if (PageTransHuge(page)) {
3963 nr_pages <<= compound_order(page);
3964 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
3966 * Never OOM-kill a process for a huge page. The
3967 * fault handler will fall back to regular pages.
3972 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
3975 __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false);
3979 int mem_cgroup_newpage_charge(struct page *page,
3980 struct mm_struct *mm, gfp_t gfp_mask)
3982 if (mem_cgroup_disabled())
3984 VM_BUG_ON_PAGE(page_mapped(page), page);
3985 VM_BUG_ON_PAGE(page->mapping && !PageAnon(page), page);
3987 return mem_cgroup_charge_common(page, mm, gfp_mask,
3988 MEM_CGROUP_CHARGE_TYPE_ANON);
3992 * While swap-in, try_charge -> commit or cancel, the page is locked.
3993 * And when try_charge() successfully returns, one refcnt to memcg without
3994 * struct page_cgroup is acquired. This refcnt will be consumed by
3995 * "commit()" or removed by "cancel()"
3997 static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm,
4000 struct mem_cgroup **memcgp)
4002 struct mem_cgroup *memcg;
4003 struct page_cgroup *pc;
4006 pc = lookup_page_cgroup(page);
4008 * Every swap fault against a single page tries to charge the
4009 * page, bail as early as possible. shmem_unuse() encounters
4010 * already charged pages, too. The USED bit is protected by
4011 * the page lock, which serializes swap cache removal, which
4012 * in turn serializes uncharging.
4014 if (PageCgroupUsed(pc))
4016 if (!do_swap_account)
4018 memcg = try_get_mem_cgroup_from_page(page);
4022 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
4023 css_put(&memcg->css);
4028 ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
4034 int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page,
4035 gfp_t gfp_mask, struct mem_cgroup **memcgp)
4038 if (mem_cgroup_disabled())
4041 * A racing thread's fault, or swapoff, may have already
4042 * updated the pte, and even removed page from swap cache: in
4043 * those cases unuse_pte()'s pte_same() test will fail; but
4044 * there's also a KSM case which does need to charge the page.
4046 if (!PageSwapCache(page)) {
4049 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, memcgp, true);
4054 return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp);
4057 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
4059 if (mem_cgroup_disabled())
4063 __mem_cgroup_cancel_charge(memcg, 1);
4067 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
4068 enum charge_type ctype)
4070 if (mem_cgroup_disabled())
4075 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
4077 * Now swap is on-memory. This means this page may be
4078 * counted both as mem and swap....double count.
4079 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
4080 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
4081 * may call delete_from_swap_cache() before reach here.
4083 if (do_swap_account && PageSwapCache(page)) {
4084 swp_entry_t ent = {.val = page_private(page)};
4085 mem_cgroup_uncharge_swap(ent);
4089 void mem_cgroup_commit_charge_swapin(struct page *page,
4090 struct mem_cgroup *memcg)
4092 __mem_cgroup_commit_charge_swapin(page, memcg,
4093 MEM_CGROUP_CHARGE_TYPE_ANON);
4096 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
4099 struct mem_cgroup *memcg = NULL;
4100 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4103 if (mem_cgroup_disabled())
4105 if (PageCompound(page))
4108 if (!PageSwapCache(page))
4109 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
4110 else { /* page is swapcache/shmem */
4111 ret = __mem_cgroup_try_charge_swapin(mm, page,
4114 __mem_cgroup_commit_charge_swapin(page, memcg, type);
4119 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
4120 unsigned int nr_pages,
4121 const enum charge_type ctype)
4123 struct memcg_batch_info *batch = NULL;
4124 bool uncharge_memsw = true;
4126 /* If swapout, usage of swap doesn't decrease */
4127 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
4128 uncharge_memsw = false;
4130 batch = ¤t->memcg_batch;
4132 * In usual, we do css_get() when we remember memcg pointer.
4133 * But in this case, we keep res->usage until end of a series of
4134 * uncharges. Then, it's ok to ignore memcg's refcnt.
4137 batch->memcg = memcg;
4139 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
4140 * In those cases, all pages freed continuously can be expected to be in
4141 * the same cgroup and we have chance to coalesce uncharges.
4142 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
4143 * because we want to do uncharge as soon as possible.
4146 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
4147 goto direct_uncharge;
4150 goto direct_uncharge;
4153 * In typical case, batch->memcg == mem. This means we can
4154 * merge a series of uncharges to an uncharge of res_counter.
4155 * If not, we uncharge res_counter ony by one.
4157 if (batch->memcg != memcg)
4158 goto direct_uncharge;
4159 /* remember freed charge and uncharge it later */
4162 batch->memsw_nr_pages++;
4165 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
4167 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
4168 if (unlikely(batch->memcg != memcg))
4169 memcg_oom_recover(memcg);
4173 * uncharge if !page_mapped(page)
4175 static struct mem_cgroup *
4176 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
4179 struct mem_cgroup *memcg = NULL;
4180 unsigned int nr_pages = 1;
4181 struct page_cgroup *pc;
4184 if (mem_cgroup_disabled())
4187 if (PageTransHuge(page)) {
4188 nr_pages <<= compound_order(page);
4189 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
4192 * Check if our page_cgroup is valid
4194 pc = lookup_page_cgroup(page);
4195 if (unlikely(!PageCgroupUsed(pc)))
4198 lock_page_cgroup(pc);
4200 memcg = pc->mem_cgroup;
4202 if (!PageCgroupUsed(pc))
4205 anon = PageAnon(page);
4208 case MEM_CGROUP_CHARGE_TYPE_ANON:
4210 * Generally PageAnon tells if it's the anon statistics to be
4211 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
4212 * used before page reached the stage of being marked PageAnon.
4216 case MEM_CGROUP_CHARGE_TYPE_DROP:
4217 /* See mem_cgroup_prepare_migration() */
4218 if (page_mapped(page))
4221 * Pages under migration may not be uncharged. But
4222 * end_migration() /must/ be the one uncharging the
4223 * unused post-migration page and so it has to call
4224 * here with the migration bit still set. See the
4225 * res_counter handling below.
4227 if (!end_migration && PageCgroupMigration(pc))
4230 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
4231 if (!PageAnon(page)) { /* Shared memory */
4232 if (page->mapping && !page_is_file_cache(page))
4234 } else if (page_mapped(page)) /* Anon */
4241 mem_cgroup_charge_statistics(memcg, page, anon, -nr_pages);
4243 ClearPageCgroupUsed(pc);
4245 * pc->mem_cgroup is not cleared here. It will be accessed when it's
4246 * freed from LRU. This is safe because uncharged page is expected not
4247 * to be reused (freed soon). Exception is SwapCache, it's handled by
4248 * special functions.
4251 unlock_page_cgroup(pc);
4253 * even after unlock, we have memcg->res.usage here and this memcg
4254 * will never be freed, so it's safe to call css_get().
4256 memcg_check_events(memcg, page);
4257 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
4258 mem_cgroup_swap_statistics(memcg, true);
4259 css_get(&memcg->css);
4262 * Migration does not charge the res_counter for the
4263 * replacement page, so leave it alone when phasing out the
4264 * page that is unused after the migration.
4266 if (!end_migration && !mem_cgroup_is_root(memcg))
4267 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
4272 unlock_page_cgroup(pc);
4276 void mem_cgroup_uncharge_page(struct page *page)
4279 if (page_mapped(page))
4281 VM_BUG_ON_PAGE(page->mapping && !PageAnon(page), page);
4283 * If the page is in swap cache, uncharge should be deferred
4284 * to the swap path, which also properly accounts swap usage
4285 * and handles memcg lifetime.
4287 * Note that this check is not stable and reclaim may add the
4288 * page to swap cache at any time after this. However, if the
4289 * page is not in swap cache by the time page->mapcount hits
4290 * 0, there won't be any page table references to the swap
4291 * slot, and reclaim will free it and not actually write the
4294 if (PageSwapCache(page))
4296 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
4299 void mem_cgroup_uncharge_cache_page(struct page *page)
4301 VM_BUG_ON_PAGE(page_mapped(page), page);
4302 VM_BUG_ON_PAGE(page->mapping, page);
4303 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
4307 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4308 * In that cases, pages are freed continuously and we can expect pages
4309 * are in the same memcg. All these calls itself limits the number of
4310 * pages freed at once, then uncharge_start/end() is called properly.
4311 * This may be called prural(2) times in a context,
4314 void mem_cgroup_uncharge_start(void)
4316 current->memcg_batch.do_batch++;
4317 /* We can do nest. */
4318 if (current->memcg_batch.do_batch == 1) {
4319 current->memcg_batch.memcg = NULL;
4320 current->memcg_batch.nr_pages = 0;
4321 current->memcg_batch.memsw_nr_pages = 0;
4325 void mem_cgroup_uncharge_end(void)
4327 struct memcg_batch_info *batch = ¤t->memcg_batch;
4329 if (!batch->do_batch)
4333 if (batch->do_batch) /* If stacked, do nothing. */
4339 * This "batch->memcg" is valid without any css_get/put etc...
4340 * bacause we hide charges behind us.
4342 if (batch->nr_pages)
4343 res_counter_uncharge(&batch->memcg->res,
4344 batch->nr_pages * PAGE_SIZE);
4345 if (batch->memsw_nr_pages)
4346 res_counter_uncharge(&batch->memcg->memsw,
4347 batch->memsw_nr_pages * PAGE_SIZE);
4348 memcg_oom_recover(batch->memcg);
4349 /* forget this pointer (for sanity check) */
4350 batch->memcg = NULL;
4355 * called after __delete_from_swap_cache() and drop "page" account.
4356 * memcg information is recorded to swap_cgroup of "ent"
4359 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
4361 struct mem_cgroup *memcg;
4362 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
4364 if (!swapout) /* this was a swap cache but the swap is unused ! */
4365 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
4367 memcg = __mem_cgroup_uncharge_common(page, ctype, false);
4370 * record memcg information, if swapout && memcg != NULL,
4371 * css_get() was called in uncharge().
4373 if (do_swap_account && swapout && memcg)
4374 swap_cgroup_record(ent, mem_cgroup_id(memcg));
4378 #ifdef CONFIG_MEMCG_SWAP
4380 * called from swap_entry_free(). remove record in swap_cgroup and
4381 * uncharge "memsw" account.
4383 void mem_cgroup_uncharge_swap(swp_entry_t ent)
4385 struct mem_cgroup *memcg;
4388 if (!do_swap_account)
4391 id = swap_cgroup_record(ent, 0);
4393 memcg = mem_cgroup_lookup(id);
4396 * We uncharge this because swap is freed.
4397 * This memcg can be obsolete one. We avoid calling css_tryget
4399 if (!mem_cgroup_is_root(memcg))
4400 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
4401 mem_cgroup_swap_statistics(memcg, false);
4402 css_put(&memcg->css);
4408 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4409 * @entry: swap entry to be moved
4410 * @from: mem_cgroup which the entry is moved from
4411 * @to: mem_cgroup which the entry is moved to
4413 * It succeeds only when the swap_cgroup's record for this entry is the same
4414 * as the mem_cgroup's id of @from.
4416 * Returns 0 on success, -EINVAL on failure.
4418 * The caller must have charged to @to, IOW, called res_counter_charge() about
4419 * both res and memsw, and called css_get().
4421 static int mem_cgroup_move_swap_account(swp_entry_t entry,
4422 struct mem_cgroup *from, struct mem_cgroup *to)
4424 unsigned short old_id, new_id;
4426 old_id = mem_cgroup_id(from);
4427 new_id = mem_cgroup_id(to);
4429 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
4430 mem_cgroup_swap_statistics(from, false);
4431 mem_cgroup_swap_statistics(to, true);
4433 * This function is only called from task migration context now.
4434 * It postpones res_counter and refcount handling till the end
4435 * of task migration(mem_cgroup_clear_mc()) for performance
4436 * improvement. But we cannot postpone css_get(to) because if
4437 * the process that has been moved to @to does swap-in, the
4438 * refcount of @to might be decreased to 0.
4440 * We are in attach() phase, so the cgroup is guaranteed to be
4441 * alive, so we can just call css_get().
4449 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
4450 struct mem_cgroup *from, struct mem_cgroup *to)
4457 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4460 void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
4461 struct mem_cgroup **memcgp)
4463 struct mem_cgroup *memcg = NULL;
4464 unsigned int nr_pages = 1;
4465 struct page_cgroup *pc;
4466 enum charge_type ctype;
4470 if (mem_cgroup_disabled())
4473 if (PageTransHuge(page))
4474 nr_pages <<= compound_order(page);
4476 pc = lookup_page_cgroup(page);
4477 lock_page_cgroup(pc);
4478 if (PageCgroupUsed(pc)) {
4479 memcg = pc->mem_cgroup;
4480 css_get(&memcg->css);
4482 * At migrating an anonymous page, its mapcount goes down
4483 * to 0 and uncharge() will be called. But, even if it's fully
4484 * unmapped, migration may fail and this page has to be
4485 * charged again. We set MIGRATION flag here and delay uncharge
4486 * until end_migration() is called
4488 * Corner Case Thinking
4490 * When the old page was mapped as Anon and it's unmap-and-freed
4491 * while migration was ongoing.
4492 * If unmap finds the old page, uncharge() of it will be delayed
4493 * until end_migration(). If unmap finds a new page, it's
4494 * uncharged when it make mapcount to be 1->0. If unmap code
4495 * finds swap_migration_entry, the new page will not be mapped
4496 * and end_migration() will find it(mapcount==0).
4499 * When the old page was mapped but migraion fails, the kernel
4500 * remaps it. A charge for it is kept by MIGRATION flag even
4501 * if mapcount goes down to 0. We can do remap successfully
4502 * without charging it again.
4505 * The "old" page is under lock_page() until the end of
4506 * migration, so, the old page itself will not be swapped-out.
4507 * If the new page is swapped out before end_migraton, our
4508 * hook to usual swap-out path will catch the event.
4511 SetPageCgroupMigration(pc);
4513 unlock_page_cgroup(pc);
4515 * If the page is not charged at this point,
4523 * We charge new page before it's used/mapped. So, even if unlock_page()
4524 * is called before end_migration, we can catch all events on this new
4525 * page. In the case new page is migrated but not remapped, new page's
4526 * mapcount will be finally 0 and we call uncharge in end_migration().
4529 ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
4531 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
4533 * The page is committed to the memcg, but it's not actually
4534 * charged to the res_counter since we plan on replacing the
4535 * old one and only one page is going to be left afterwards.
4537 __mem_cgroup_commit_charge(memcg, newpage, nr_pages, ctype, false);
4540 /* remove redundant charge if migration failed*/
4541 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
4542 struct page *oldpage, struct page *newpage, bool migration_ok)
4544 struct page *used, *unused;
4545 struct page_cgroup *pc;
4551 if (!migration_ok) {
4558 anon = PageAnon(used);
4559 __mem_cgroup_uncharge_common(unused,
4560 anon ? MEM_CGROUP_CHARGE_TYPE_ANON
4561 : MEM_CGROUP_CHARGE_TYPE_CACHE,
4563 css_put(&memcg->css);
4565 * We disallowed uncharge of pages under migration because mapcount
4566 * of the page goes down to zero, temporarly.
4567 * Clear the flag and check the page should be charged.
4569 pc = lookup_page_cgroup(oldpage);
4570 lock_page_cgroup(pc);
4571 ClearPageCgroupMigration(pc);
4572 unlock_page_cgroup(pc);
4575 * If a page is a file cache, radix-tree replacement is very atomic
4576 * and we can skip this check. When it was an Anon page, its mapcount
4577 * goes down to 0. But because we added MIGRATION flage, it's not
4578 * uncharged yet. There are several case but page->mapcount check
4579 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4580 * check. (see prepare_charge() also)
4583 mem_cgroup_uncharge_page(used);
4587 * At replace page cache, newpage is not under any memcg but it's on
4588 * LRU. So, this function doesn't touch res_counter but handles LRU
4589 * in correct way. Both pages are locked so we cannot race with uncharge.
4591 void mem_cgroup_replace_page_cache(struct page *oldpage,
4592 struct page *newpage)
4594 struct mem_cgroup *memcg = NULL;
4595 struct page_cgroup *pc;
4596 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4598 if (mem_cgroup_disabled())
4601 pc = lookup_page_cgroup(oldpage);
4602 /* fix accounting on old pages */
4603 lock_page_cgroup(pc);
4604 if (PageCgroupUsed(pc)) {
4605 memcg = pc->mem_cgroup;
4606 mem_cgroup_charge_statistics(memcg, oldpage, false, -1);
4607 ClearPageCgroupUsed(pc);
4609 unlock_page_cgroup(pc);
4612 * When called from shmem_replace_page(), in some cases the
4613 * oldpage has already been charged, and in some cases not.
4618 * Even if newpage->mapping was NULL before starting replacement,
4619 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4620 * LRU while we overwrite pc->mem_cgroup.
4622 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
4625 #ifdef CONFIG_DEBUG_VM
4626 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
4628 struct page_cgroup *pc;
4630 pc = lookup_page_cgroup(page);
4632 * Can be NULL while feeding pages into the page allocator for
4633 * the first time, i.e. during boot or memory hotplug;
4634 * or when mem_cgroup_disabled().
4636 if (likely(pc) && PageCgroupUsed(pc))
4641 bool mem_cgroup_bad_page_check(struct page *page)
4643 if (mem_cgroup_disabled())
4646 return lookup_page_cgroup_used(page) != NULL;
4649 void mem_cgroup_print_bad_page(struct page *page)
4651 struct page_cgroup *pc;
4653 pc = lookup_page_cgroup_used(page);
4655 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4656 pc, pc->flags, pc->mem_cgroup);
4661 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
4662 unsigned long long val)
4665 u64 memswlimit, memlimit;
4667 int children = mem_cgroup_count_children(memcg);
4668 u64 curusage, oldusage;
4672 * For keeping hierarchical_reclaim simple, how long we should retry
4673 * is depends on callers. We set our retry-count to be function
4674 * of # of children which we should visit in this loop.
4676 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
4678 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4681 while (retry_count) {
4682 if (signal_pending(current)) {
4687 * Rather than hide all in some function, I do this in
4688 * open coded manner. You see what this really does.
4689 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4691 mutex_lock(&set_limit_mutex);
4692 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4693 if (memswlimit < val) {
4695 mutex_unlock(&set_limit_mutex);
4699 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4703 ret = res_counter_set_limit(&memcg->res, val);
4705 if (memswlimit == val)
4706 memcg->memsw_is_minimum = true;
4708 memcg->memsw_is_minimum = false;
4710 mutex_unlock(&set_limit_mutex);
4715 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4716 MEM_CGROUP_RECLAIM_SHRINK);
4717 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4718 /* Usage is reduced ? */
4719 if (curusage >= oldusage)
4722 oldusage = curusage;
4724 if (!ret && enlarge)
4725 memcg_oom_recover(memcg);
4730 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
4731 unsigned long long val)
4734 u64 memlimit, memswlimit, oldusage, curusage;
4735 int children = mem_cgroup_count_children(memcg);
4739 /* see mem_cgroup_resize_res_limit */
4740 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
4741 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4742 while (retry_count) {
4743 if (signal_pending(current)) {
4748 * Rather than hide all in some function, I do this in
4749 * open coded manner. You see what this really does.
4750 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4752 mutex_lock(&set_limit_mutex);
4753 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4754 if (memlimit > val) {
4756 mutex_unlock(&set_limit_mutex);
4759 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4760 if (memswlimit < val)
4762 ret = res_counter_set_limit(&memcg->memsw, val);
4764 if (memlimit == val)
4765 memcg->memsw_is_minimum = true;
4767 memcg->memsw_is_minimum = false;
4769 mutex_unlock(&set_limit_mutex);
4774 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4775 MEM_CGROUP_RECLAIM_NOSWAP |
4776 MEM_CGROUP_RECLAIM_SHRINK);
4777 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4778 /* Usage is reduced ? */
4779 if (curusage >= oldusage)
4782 oldusage = curusage;
4784 if (!ret && enlarge)
4785 memcg_oom_recover(memcg);
4789 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
4791 unsigned long *total_scanned)
4793 unsigned long nr_reclaimed = 0;
4794 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
4795 unsigned long reclaimed;
4797 struct mem_cgroup_tree_per_zone *mctz;
4798 unsigned long long excess;
4799 unsigned long nr_scanned;
4804 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
4806 * This loop can run a while, specially if mem_cgroup's continuously
4807 * keep exceeding their soft limit and putting the system under
4814 mz = mem_cgroup_largest_soft_limit_node(mctz);
4819 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
4820 gfp_mask, &nr_scanned);
4821 nr_reclaimed += reclaimed;
4822 *total_scanned += nr_scanned;
4823 spin_lock(&mctz->lock);
4826 * If we failed to reclaim anything from this memory cgroup
4827 * it is time to move on to the next cgroup
4833 * Loop until we find yet another one.
4835 * By the time we get the soft_limit lock
4836 * again, someone might have aded the
4837 * group back on the RB tree. Iterate to
4838 * make sure we get a different mem.
4839 * mem_cgroup_largest_soft_limit_node returns
4840 * NULL if no other cgroup is present on
4844 __mem_cgroup_largest_soft_limit_node(mctz);
4846 css_put(&next_mz->memcg->css);
4847 else /* next_mz == NULL or other memcg */
4851 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
4852 excess = res_counter_soft_limit_excess(&mz->memcg->res);
4854 * One school of thought says that we should not add
4855 * back the node to the tree if reclaim returns 0.
4856 * But our reclaim could return 0, simply because due
4857 * to priority we are exposing a smaller subset of
4858 * memory to reclaim from. Consider this as a longer
4861 /* If excess == 0, no tree ops */
4862 __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess);
4863 spin_unlock(&mctz->lock);
4864 css_put(&mz->memcg->css);
4867 * Could not reclaim anything and there are no more
4868 * mem cgroups to try or we seem to be looping without
4869 * reclaiming anything.
4871 if (!nr_reclaimed &&
4873 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
4875 } while (!nr_reclaimed);
4877 css_put(&next_mz->memcg->css);
4878 return nr_reclaimed;
4882 * mem_cgroup_force_empty_list - clears LRU of a group
4883 * @memcg: group to clear
4886 * @lru: lru to to clear
4888 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4889 * reclaim the pages page themselves - pages are moved to the parent (or root)
4892 static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
4893 int node, int zid, enum lru_list lru)
4895 struct lruvec *lruvec;
4896 unsigned long flags;
4897 struct list_head *list;
4901 zone = &NODE_DATA(node)->node_zones[zid];
4902 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
4903 list = &lruvec->lists[lru];
4907 struct page_cgroup *pc;
4910 spin_lock_irqsave(&zone->lru_lock, flags);
4911 if (list_empty(list)) {
4912 spin_unlock_irqrestore(&zone->lru_lock, flags);
4915 page = list_entry(list->prev, struct page, lru);
4917 list_move(&page->lru, list);
4919 spin_unlock_irqrestore(&zone->lru_lock, flags);
4922 spin_unlock_irqrestore(&zone->lru_lock, flags);
4924 pc = lookup_page_cgroup(page);
4926 if (mem_cgroup_move_parent(page, pc, memcg)) {
4927 /* found lock contention or "pc" is obsolete. */
4932 } while (!list_empty(list));
4936 * make mem_cgroup's charge to be 0 if there is no task by moving
4937 * all the charges and pages to the parent.
4938 * This enables deleting this mem_cgroup.
4940 * Caller is responsible for holding css reference on the memcg.
4942 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
4948 /* This is for making all *used* pages to be on LRU. */
4949 lru_add_drain_all();
4950 drain_all_stock_sync(memcg);
4951 mem_cgroup_start_move(memcg);
4952 for_each_node_state(node, N_MEMORY) {
4953 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4956 mem_cgroup_force_empty_list(memcg,
4961 mem_cgroup_end_move(memcg);
4962 memcg_oom_recover(memcg);
4966 * Kernel memory may not necessarily be trackable to a specific
4967 * process. So they are not migrated, and therefore we can't
4968 * expect their value to drop to 0 here.
4969 * Having res filled up with kmem only is enough.
4971 * This is a safety check because mem_cgroup_force_empty_list
4972 * could have raced with mem_cgroup_replace_page_cache callers
4973 * so the lru seemed empty but the page could have been added
4974 * right after the check. RES_USAGE should be safe as we always
4975 * charge before adding to the LRU.
4977 usage = res_counter_read_u64(&memcg->res, RES_USAGE) -
4978 res_counter_read_u64(&memcg->kmem, RES_USAGE);
4979 } while (usage > 0);
4982 static inline bool memcg_has_children(struct mem_cgroup *memcg)
4984 lockdep_assert_held(&memcg_create_mutex);
4986 * The lock does not prevent addition or deletion to the list
4987 * of children, but it prevents a new child from being
4988 * initialized based on this parent in css_online(), so it's
4989 * enough to decide whether hierarchically inherited
4990 * attributes can still be changed or not.
4992 return memcg->use_hierarchy &&
4993 !list_empty(&memcg->css.cgroup->children);
4997 * Reclaims as many pages from the given memcg as possible and moves
4998 * the rest to the parent.
5000 * Caller is responsible for holding css reference for memcg.
5002 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
5004 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
5005 struct cgroup *cgrp = memcg->css.cgroup;
5007 /* returns EBUSY if there is a task or if we come here twice. */
5008 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
5011 /* we call try-to-free pages for make this cgroup empty */
5012 lru_add_drain_all();
5013 /* try to free all pages in this cgroup */
5014 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
5017 if (signal_pending(current))
5020 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
5024 /* maybe some writeback is necessary */
5025 congestion_wait(BLK_RW_ASYNC, HZ/10);
5030 mem_cgroup_reparent_charges(memcg);
5035 static int mem_cgroup_force_empty_write(struct cgroup_subsys_state *css,
5038 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5040 if (mem_cgroup_is_root(memcg))
5042 return mem_cgroup_force_empty(memcg);
5045 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
5048 return mem_cgroup_from_css(css)->use_hierarchy;
5051 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
5052 struct cftype *cft, u64 val)
5055 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5056 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(css_parent(&memcg->css));
5058 mutex_lock(&memcg_create_mutex);
5060 if (memcg->use_hierarchy == val)
5064 * If parent's use_hierarchy is set, we can't make any modifications
5065 * in the child subtrees. If it is unset, then the change can
5066 * occur, provided the current cgroup has no children.
5068 * For the root cgroup, parent_mem is NULL, we allow value to be
5069 * set if there are no children.
5071 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
5072 (val == 1 || val == 0)) {
5073 if (list_empty(&memcg->css.cgroup->children))
5074 memcg->use_hierarchy = val;
5081 mutex_unlock(&memcg_create_mutex);
5087 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
5088 enum mem_cgroup_stat_index idx)
5090 struct mem_cgroup *iter;
5093 /* Per-cpu values can be negative, use a signed accumulator */
5094 for_each_mem_cgroup_tree(iter, memcg)
5095 val += mem_cgroup_read_stat(iter, idx);
5097 if (val < 0) /* race ? */
5102 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
5106 if (!mem_cgroup_is_root(memcg)) {
5108 return res_counter_read_u64(&memcg->res, RES_USAGE);
5110 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
5114 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
5115 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
5117 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
5118 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
5121 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
5123 return val << PAGE_SHIFT;
5126 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
5129 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5134 type = MEMFILE_TYPE(cft->private);
5135 name = MEMFILE_ATTR(cft->private);
5139 if (name == RES_USAGE)
5140 val = mem_cgroup_usage(memcg, false);
5142 val = res_counter_read_u64(&memcg->res, name);
5145 if (name == RES_USAGE)
5146 val = mem_cgroup_usage(memcg, true);
5148 val = res_counter_read_u64(&memcg->memsw, name);
5151 val = res_counter_read_u64(&memcg->kmem, name);
5160 #ifdef CONFIG_MEMCG_KMEM
5161 /* should be called with activate_kmem_mutex held */
5162 static int __memcg_activate_kmem(struct mem_cgroup *memcg,
5163 unsigned long long limit)
5168 if (memcg_kmem_is_active(memcg))
5172 * We are going to allocate memory for data shared by all memory
5173 * cgroups so let's stop accounting here.
5175 memcg_stop_kmem_account();
5178 * For simplicity, we won't allow this to be disabled. It also can't
5179 * be changed if the cgroup has children already, or if tasks had
5182 * If tasks join before we set the limit, a person looking at
5183 * kmem.usage_in_bytes will have no way to determine when it took
5184 * place, which makes the value quite meaningless.
5186 * After it first became limited, changes in the value of the limit are
5187 * of course permitted.
5189 mutex_lock(&memcg_create_mutex);
5190 if (cgroup_task_count(memcg->css.cgroup) || memcg_has_children(memcg))
5192 mutex_unlock(&memcg_create_mutex);
5196 memcg_id = ida_simple_get(&kmem_limited_groups,
5197 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
5204 * Make sure we have enough space for this cgroup in each root cache's
5207 err = memcg_update_all_caches(memcg_id + 1);
5211 memcg->kmemcg_id = memcg_id;
5212 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
5213 mutex_init(&memcg->slab_caches_mutex);
5216 * We couldn't have accounted to this cgroup, because it hasn't got the
5217 * active bit set yet, so this should succeed.
5219 err = res_counter_set_limit(&memcg->kmem, limit);
5222 static_key_slow_inc(&memcg_kmem_enabled_key);
5224 * Setting the active bit after enabling static branching will
5225 * guarantee no one starts accounting before all call sites are
5228 memcg_kmem_set_active(memcg);
5230 memcg_resume_kmem_account();
5234 ida_simple_remove(&kmem_limited_groups, memcg_id);
5238 static int memcg_activate_kmem(struct mem_cgroup *memcg,
5239 unsigned long long limit)
5243 mutex_lock(&activate_kmem_mutex);
5244 ret = __memcg_activate_kmem(memcg, limit);
5245 mutex_unlock(&activate_kmem_mutex);
5249 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
5250 unsigned long long val)
5254 if (!memcg_kmem_is_active(memcg))
5255 ret = memcg_activate_kmem(memcg, val);
5257 ret = res_counter_set_limit(&memcg->kmem, val);
5261 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
5264 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5269 mutex_lock(&activate_kmem_mutex);
5271 * If the parent cgroup is not kmem-active now, it cannot be activated
5272 * after this point, because it has at least one child already.
5274 if (memcg_kmem_is_active(parent))
5275 ret = __memcg_activate_kmem(memcg, RES_COUNTER_MAX);
5276 mutex_unlock(&activate_kmem_mutex);
5280 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
5281 unsigned long long val)
5285 #endif /* CONFIG_MEMCG_KMEM */
5288 * The user of this function is...
5291 static int mem_cgroup_write(struct cgroup_subsys_state *css, struct cftype *cft,
5294 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5297 unsigned long long val;
5300 type = MEMFILE_TYPE(cft->private);
5301 name = MEMFILE_ATTR(cft->private);
5305 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
5309 /* This function does all necessary parse...reuse it */
5310 ret = res_counter_memparse_write_strategy(buffer, &val);
5314 ret = mem_cgroup_resize_limit(memcg, val);
5315 else if (type == _MEMSWAP)
5316 ret = mem_cgroup_resize_memsw_limit(memcg, val);
5317 else if (type == _KMEM)
5318 ret = memcg_update_kmem_limit(memcg, val);
5322 case RES_SOFT_LIMIT:
5323 ret = res_counter_memparse_write_strategy(buffer, &val);
5327 * For memsw, soft limits are hard to implement in terms
5328 * of semantics, for now, we support soft limits for
5329 * control without swap
5332 ret = res_counter_set_soft_limit(&memcg->res, val);
5337 ret = -EINVAL; /* should be BUG() ? */
5343 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
5344 unsigned long long *mem_limit, unsigned long long *memsw_limit)
5346 unsigned long long min_limit, min_memsw_limit, tmp;
5348 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
5349 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5350 if (!memcg->use_hierarchy)
5353 while (css_parent(&memcg->css)) {
5354 memcg = mem_cgroup_from_css(css_parent(&memcg->css));
5355 if (!memcg->use_hierarchy)
5357 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
5358 min_limit = min(min_limit, tmp);
5359 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5360 min_memsw_limit = min(min_memsw_limit, tmp);
5363 *mem_limit = min_limit;
5364 *memsw_limit = min_memsw_limit;
5367 static int mem_cgroup_reset(struct cgroup_subsys_state *css, unsigned int event)
5369 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5373 type = MEMFILE_TYPE(event);
5374 name = MEMFILE_ATTR(event);
5379 res_counter_reset_max(&memcg->res);
5380 else if (type == _MEMSWAP)
5381 res_counter_reset_max(&memcg->memsw);
5382 else if (type == _KMEM)
5383 res_counter_reset_max(&memcg->kmem);
5389 res_counter_reset_failcnt(&memcg->res);
5390 else if (type == _MEMSWAP)
5391 res_counter_reset_failcnt(&memcg->memsw);
5392 else if (type == _KMEM)
5393 res_counter_reset_failcnt(&memcg->kmem);
5402 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
5405 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
5409 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5410 struct cftype *cft, u64 val)
5412 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5414 if (val >= (1 << NR_MOVE_TYPE))
5418 * No kind of locking is needed in here, because ->can_attach() will
5419 * check this value once in the beginning of the process, and then carry
5420 * on with stale data. This means that changes to this value will only
5421 * affect task migrations starting after the change.
5423 memcg->move_charge_at_immigrate = val;
5427 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5428 struct cftype *cft, u64 val)
5435 static int memcg_numa_stat_show(struct seq_file *m, void *v)
5439 unsigned int lru_mask;
5442 static const struct numa_stat stats[] = {
5443 { "total", LRU_ALL },
5444 { "file", LRU_ALL_FILE },
5445 { "anon", LRU_ALL_ANON },
5446 { "unevictable", BIT(LRU_UNEVICTABLE) },
5448 const struct numa_stat *stat;
5451 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5453 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
5454 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
5455 seq_printf(m, "%s=%lu", stat->name, nr);
5456 for_each_node_state(nid, N_MEMORY) {
5457 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5459 seq_printf(m, " N%d=%lu", nid, nr);
5464 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
5465 struct mem_cgroup *iter;
5468 for_each_mem_cgroup_tree(iter, memcg)
5469 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
5470 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
5471 for_each_node_state(nid, N_MEMORY) {
5473 for_each_mem_cgroup_tree(iter, memcg)
5474 nr += mem_cgroup_node_nr_lru_pages(
5475 iter, nid, stat->lru_mask);
5476 seq_printf(m, " N%d=%lu", nid, nr);
5483 #endif /* CONFIG_NUMA */
5485 static inline void mem_cgroup_lru_names_not_uptodate(void)
5487 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
5490 static int memcg_stat_show(struct seq_file *m, void *v)
5492 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5493 struct mem_cgroup *mi;
5496 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5497 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5499 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
5500 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
5503 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
5504 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
5505 mem_cgroup_read_events(memcg, i));
5507 for (i = 0; i < NR_LRU_LISTS; i++)
5508 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
5509 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
5511 /* Hierarchical information */
5513 unsigned long long limit, memsw_limit;
5514 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
5515 seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
5516 if (do_swap_account)
5517 seq_printf(m, "hierarchical_memsw_limit %llu\n",
5521 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5524 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5526 for_each_mem_cgroup_tree(mi, memcg)
5527 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
5528 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
5531 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
5532 unsigned long long val = 0;
5534 for_each_mem_cgroup_tree(mi, memcg)
5535 val += mem_cgroup_read_events(mi, i);
5536 seq_printf(m, "total_%s %llu\n",
5537 mem_cgroup_events_names[i], val);
5540 for (i = 0; i < NR_LRU_LISTS; i++) {
5541 unsigned long long val = 0;
5543 for_each_mem_cgroup_tree(mi, memcg)
5544 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
5545 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
5548 #ifdef CONFIG_DEBUG_VM
5551 struct mem_cgroup_per_zone *mz;
5552 struct zone_reclaim_stat *rstat;
5553 unsigned long recent_rotated[2] = {0, 0};
5554 unsigned long recent_scanned[2] = {0, 0};
5556 for_each_online_node(nid)
5557 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
5558 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
5559 rstat = &mz->lruvec.reclaim_stat;
5561 recent_rotated[0] += rstat->recent_rotated[0];
5562 recent_rotated[1] += rstat->recent_rotated[1];
5563 recent_scanned[0] += rstat->recent_scanned[0];
5564 recent_scanned[1] += rstat->recent_scanned[1];
5566 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
5567 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
5568 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
5569 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
5576 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
5579 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5581 return mem_cgroup_swappiness(memcg);
5584 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
5585 struct cftype *cft, u64 val)
5587 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5588 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(&memcg->css));
5590 if (val > 100 || !parent)
5593 mutex_lock(&memcg_create_mutex);
5595 /* If under hierarchy, only empty-root can set this value */
5596 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5597 mutex_unlock(&memcg_create_mutex);
5601 memcg->swappiness = val;
5603 mutex_unlock(&memcg_create_mutex);
5608 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
5610 struct mem_cgroup_threshold_ary *t;
5616 t = rcu_dereference(memcg->thresholds.primary);
5618 t = rcu_dereference(memcg->memsw_thresholds.primary);
5623 usage = mem_cgroup_usage(memcg, swap);
5626 * current_threshold points to threshold just below or equal to usage.
5627 * If it's not true, a threshold was crossed after last
5628 * call of __mem_cgroup_threshold().
5630 i = t->current_threshold;
5633 * Iterate backward over array of thresholds starting from
5634 * current_threshold and check if a threshold is crossed.
5635 * If none of thresholds below usage is crossed, we read
5636 * only one element of the array here.
5638 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
5639 eventfd_signal(t->entries[i].eventfd, 1);
5641 /* i = current_threshold + 1 */
5645 * Iterate forward over array of thresholds starting from
5646 * current_threshold+1 and check if a threshold is crossed.
5647 * If none of thresholds above usage is crossed, we read
5648 * only one element of the array here.
5650 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
5651 eventfd_signal(t->entries[i].eventfd, 1);
5653 /* Update current_threshold */
5654 t->current_threshold = i - 1;
5659 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
5662 __mem_cgroup_threshold(memcg, false);
5663 if (do_swap_account)
5664 __mem_cgroup_threshold(memcg, true);
5666 memcg = parent_mem_cgroup(memcg);
5670 static int compare_thresholds(const void *a, const void *b)
5672 const struct mem_cgroup_threshold *_a = a;
5673 const struct mem_cgroup_threshold *_b = b;
5675 if (_a->threshold > _b->threshold)
5678 if (_a->threshold < _b->threshold)
5684 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
5686 struct mem_cgroup_eventfd_list *ev;
5688 spin_lock(&memcg_oom_lock);
5690 list_for_each_entry(ev, &memcg->oom_notify, list)
5691 eventfd_signal(ev->eventfd, 1);
5693 spin_unlock(&memcg_oom_lock);
5697 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
5699 struct mem_cgroup *iter;
5701 for_each_mem_cgroup_tree(iter, memcg)
5702 mem_cgroup_oom_notify_cb(iter);
5705 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
5706 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
5708 struct mem_cgroup_thresholds *thresholds;
5709 struct mem_cgroup_threshold_ary *new;
5710 u64 threshold, usage;
5713 ret = res_counter_memparse_write_strategy(args, &threshold);
5717 mutex_lock(&memcg->thresholds_lock);
5720 thresholds = &memcg->thresholds;
5721 else if (type == _MEMSWAP)
5722 thresholds = &memcg->memsw_thresholds;
5726 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5728 /* Check if a threshold crossed before adding a new one */
5729 if (thresholds->primary)
5730 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5732 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
5734 /* Allocate memory for new array of thresholds */
5735 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
5743 /* Copy thresholds (if any) to new array */
5744 if (thresholds->primary) {
5745 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
5746 sizeof(struct mem_cgroup_threshold));
5749 /* Add new threshold */
5750 new->entries[size - 1].eventfd = eventfd;
5751 new->entries[size - 1].threshold = threshold;
5753 /* Sort thresholds. Registering of new threshold isn't time-critical */
5754 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
5755 compare_thresholds, NULL);
5757 /* Find current threshold */
5758 new->current_threshold = -1;
5759 for (i = 0; i < size; i++) {
5760 if (new->entries[i].threshold <= usage) {
5762 * new->current_threshold will not be used until
5763 * rcu_assign_pointer(), so it's safe to increment
5766 ++new->current_threshold;
5771 /* Free old spare buffer and save old primary buffer as spare */
5772 kfree(thresholds->spare);
5773 thresholds->spare = thresholds->primary;
5775 rcu_assign_pointer(thresholds->primary, new);
5777 /* To be sure that nobody uses thresholds */
5781 mutex_unlock(&memcg->thresholds_lock);
5786 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
5787 struct eventfd_ctx *eventfd, const char *args)
5789 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
5792 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
5793 struct eventfd_ctx *eventfd, const char *args)
5795 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
5798 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
5799 struct eventfd_ctx *eventfd, enum res_type type)
5801 struct mem_cgroup_thresholds *thresholds;
5802 struct mem_cgroup_threshold_ary *new;
5806 mutex_lock(&memcg->thresholds_lock);
5808 thresholds = &memcg->thresholds;
5809 else if (type == _MEMSWAP)
5810 thresholds = &memcg->memsw_thresholds;
5814 if (!thresholds->primary)
5817 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5819 /* Check if a threshold crossed before removing */
5820 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5822 /* Calculate new number of threshold */
5824 for (i = 0; i < thresholds->primary->size; i++) {
5825 if (thresholds->primary->entries[i].eventfd != eventfd)
5829 new = thresholds->spare;
5831 /* Set thresholds array to NULL if we don't have thresholds */
5840 /* Copy thresholds and find current threshold */
5841 new->current_threshold = -1;
5842 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
5843 if (thresholds->primary->entries[i].eventfd == eventfd)
5846 new->entries[j] = thresholds->primary->entries[i];
5847 if (new->entries[j].threshold <= usage) {
5849 * new->current_threshold will not be used
5850 * until rcu_assign_pointer(), so it's safe to increment
5853 ++new->current_threshold;
5859 /* Swap primary and spare array */
5860 thresholds->spare = thresholds->primary;
5861 /* If all events are unregistered, free the spare array */
5863 kfree(thresholds->spare);
5864 thresholds->spare = NULL;
5867 rcu_assign_pointer(thresholds->primary, new);
5869 /* To be sure that nobody uses thresholds */
5872 mutex_unlock(&memcg->thresholds_lock);
5875 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
5876 struct eventfd_ctx *eventfd)
5878 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
5881 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
5882 struct eventfd_ctx *eventfd)
5884 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
5887 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
5888 struct eventfd_ctx *eventfd, const char *args)
5890 struct mem_cgroup_eventfd_list *event;
5892 event = kmalloc(sizeof(*event), GFP_KERNEL);
5896 spin_lock(&memcg_oom_lock);
5898 event->eventfd = eventfd;
5899 list_add(&event->list, &memcg->oom_notify);
5901 /* already in OOM ? */
5902 if (atomic_read(&memcg->under_oom))
5903 eventfd_signal(eventfd, 1);
5904 spin_unlock(&memcg_oom_lock);
5909 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
5910 struct eventfd_ctx *eventfd)
5912 struct mem_cgroup_eventfd_list *ev, *tmp;
5914 spin_lock(&memcg_oom_lock);
5916 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
5917 if (ev->eventfd == eventfd) {
5918 list_del(&ev->list);
5923 spin_unlock(&memcg_oom_lock);
5926 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
5928 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
5930 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
5931 seq_printf(sf, "under_oom %d\n", (bool)atomic_read(&memcg->under_oom));
5935 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
5936 struct cftype *cft, u64 val)
5938 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5939 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(&memcg->css));
5941 /* cannot set to root cgroup and only 0 and 1 are allowed */
5942 if (!parent || !((val == 0) || (val == 1)))
5945 mutex_lock(&memcg_create_mutex);
5946 /* oom-kill-disable is a flag for subhierarchy. */
5947 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5948 mutex_unlock(&memcg_create_mutex);
5951 memcg->oom_kill_disable = val;
5953 memcg_oom_recover(memcg);
5954 mutex_unlock(&memcg_create_mutex);
5958 #ifdef CONFIG_MEMCG_KMEM
5959 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5963 memcg->kmemcg_id = -1;
5964 ret = memcg_propagate_kmem(memcg);
5968 return mem_cgroup_sockets_init(memcg, ss);
5971 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5973 mem_cgroup_sockets_destroy(memcg);
5976 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5978 if (!memcg_kmem_is_active(memcg))
5982 * kmem charges can outlive the cgroup. In the case of slab
5983 * pages, for instance, a page contain objects from various
5984 * processes. As we prevent from taking a reference for every
5985 * such allocation we have to be careful when doing uncharge
5986 * (see memcg_uncharge_kmem) and here during offlining.
5988 * The idea is that that only the _last_ uncharge which sees
5989 * the dead memcg will drop the last reference. An additional
5990 * reference is taken here before the group is marked dead
5991 * which is then paired with css_put during uncharge resp. here.
5993 * Although this might sound strange as this path is called from
5994 * css_offline() when the referencemight have dropped down to 0
5995 * and shouldn't be incremented anymore (css_tryget would fail)
5996 * we do not have other options because of the kmem allocations
5999 css_get(&memcg->css);
6001 memcg_kmem_mark_dead(memcg);
6003 if (res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0)
6006 if (memcg_kmem_test_and_clear_dead(memcg))
6007 css_put(&memcg->css);
6010 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
6015 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
6019 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
6025 * DO NOT USE IN NEW FILES.
6027 * "cgroup.event_control" implementation.
6029 * This is way over-engineered. It tries to support fully configurable
6030 * events for each user. Such level of flexibility is completely
6031 * unnecessary especially in the light of the planned unified hierarchy.
6033 * Please deprecate this and replace with something simpler if at all
6038 * Unregister event and free resources.
6040 * Gets called from workqueue.
6042 static void memcg_event_remove(struct work_struct *work)
6044 struct mem_cgroup_event *event =
6045 container_of(work, struct mem_cgroup_event, remove);
6046 struct mem_cgroup *memcg = event->memcg;
6048 remove_wait_queue(event->wqh, &event->wait);
6050 event->unregister_event(memcg, event->eventfd);
6052 /* Notify userspace the event is going away. */
6053 eventfd_signal(event->eventfd, 1);
6055 eventfd_ctx_put(event->eventfd);
6057 css_put(&memcg->css);
6061 * Gets called on POLLHUP on eventfd when user closes it.
6063 * Called with wqh->lock held and interrupts disabled.
6065 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
6066 int sync, void *key)
6068 struct mem_cgroup_event *event =
6069 container_of(wait, struct mem_cgroup_event, wait);
6070 struct mem_cgroup *memcg = event->memcg;
6071 unsigned long flags = (unsigned long)key;
6073 if (flags & POLLHUP) {
6075 * If the event has been detached at cgroup removal, we
6076 * can simply return knowing the other side will cleanup
6079 * We can't race against event freeing since the other
6080 * side will require wqh->lock via remove_wait_queue(),
6083 spin_lock(&memcg->event_list_lock);
6084 if (!list_empty(&event->list)) {
6085 list_del_init(&event->list);
6087 * We are in atomic context, but cgroup_event_remove()
6088 * may sleep, so we have to call it in workqueue.
6090 schedule_work(&event->remove);
6092 spin_unlock(&memcg->event_list_lock);
6098 static void memcg_event_ptable_queue_proc(struct file *file,
6099 wait_queue_head_t *wqh, poll_table *pt)
6101 struct mem_cgroup_event *event =
6102 container_of(pt, struct mem_cgroup_event, pt);
6105 add_wait_queue(wqh, &event->wait);
6109 * DO NOT USE IN NEW FILES.
6111 * Parse input and register new cgroup event handler.
6113 * Input must be in format '<event_fd> <control_fd> <args>'.
6114 * Interpretation of args is defined by control file implementation.
6116 static int memcg_write_event_control(struct cgroup_subsys_state *css,
6117 struct cftype *cft, const char *buffer)
6119 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6120 struct mem_cgroup_event *event;
6121 struct cgroup_subsys_state *cfile_css;
6122 unsigned int efd, cfd;
6129 efd = simple_strtoul(buffer, &endp, 10);
6134 cfd = simple_strtoul(buffer, &endp, 10);
6135 if ((*endp != ' ') && (*endp != '\0'))
6139 event = kzalloc(sizeof(*event), GFP_KERNEL);
6143 event->memcg = memcg;
6144 INIT_LIST_HEAD(&event->list);
6145 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
6146 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
6147 INIT_WORK(&event->remove, memcg_event_remove);
6155 event->eventfd = eventfd_ctx_fileget(efile.file);
6156 if (IS_ERR(event->eventfd)) {
6157 ret = PTR_ERR(event->eventfd);
6164 goto out_put_eventfd;
6167 /* the process need read permission on control file */
6168 /* AV: shouldn't we check that it's been opened for read instead? */
6169 ret = inode_permission(file_inode(cfile.file), MAY_READ);
6174 * Determine the event callbacks and set them in @event. This used
6175 * to be done via struct cftype but cgroup core no longer knows
6176 * about these events. The following is crude but the whole thing
6177 * is for compatibility anyway.
6179 * DO NOT ADD NEW FILES.
6181 name = cfile.file->f_dentry->d_name.name;
6183 if (!strcmp(name, "memory.usage_in_bytes")) {
6184 event->register_event = mem_cgroup_usage_register_event;
6185 event->unregister_event = mem_cgroup_usage_unregister_event;
6186 } else if (!strcmp(name, "memory.oom_control")) {
6187 event->register_event = mem_cgroup_oom_register_event;
6188 event->unregister_event = mem_cgroup_oom_unregister_event;
6189 } else if (!strcmp(name, "memory.pressure_level")) {
6190 event->register_event = vmpressure_register_event;
6191 event->unregister_event = vmpressure_unregister_event;
6192 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
6193 event->register_event = memsw_cgroup_usage_register_event;
6194 event->unregister_event = memsw_cgroup_usage_unregister_event;
6201 * Verify @cfile should belong to @css. Also, remaining events are
6202 * automatically removed on cgroup destruction but the removal is
6203 * asynchronous, so take an extra ref on @css.
6208 cfile_css = css_from_dir(cfile.file->f_dentry->d_parent,
6209 &mem_cgroup_subsys);
6210 if (cfile_css == css && css_tryget(css))
6217 ret = event->register_event(memcg, event->eventfd, buffer);
6221 efile.file->f_op->poll(efile.file, &event->pt);
6223 spin_lock(&memcg->event_list_lock);
6224 list_add(&event->list, &memcg->event_list);
6225 spin_unlock(&memcg->event_list_lock);
6237 eventfd_ctx_put(event->eventfd);
6246 static struct cftype mem_cgroup_files[] = {
6248 .name = "usage_in_bytes",
6249 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
6250 .read_u64 = mem_cgroup_read_u64,
6253 .name = "max_usage_in_bytes",
6254 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
6255 .trigger = mem_cgroup_reset,
6256 .read_u64 = mem_cgroup_read_u64,
6259 .name = "limit_in_bytes",
6260 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
6261 .write_string = mem_cgroup_write,
6262 .read_u64 = mem_cgroup_read_u64,
6265 .name = "soft_limit_in_bytes",
6266 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
6267 .write_string = mem_cgroup_write,
6268 .read_u64 = mem_cgroup_read_u64,
6272 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
6273 .trigger = mem_cgroup_reset,
6274 .read_u64 = mem_cgroup_read_u64,
6278 .seq_show = memcg_stat_show,
6281 .name = "force_empty",
6282 .trigger = mem_cgroup_force_empty_write,
6285 .name = "use_hierarchy",
6286 .flags = CFTYPE_INSANE,
6287 .write_u64 = mem_cgroup_hierarchy_write,
6288 .read_u64 = mem_cgroup_hierarchy_read,
6291 .name = "cgroup.event_control", /* XXX: for compat */
6292 .write_string = memcg_write_event_control,
6293 .flags = CFTYPE_NO_PREFIX,
6297 .name = "swappiness",
6298 .read_u64 = mem_cgroup_swappiness_read,
6299 .write_u64 = mem_cgroup_swappiness_write,
6302 .name = "move_charge_at_immigrate",
6303 .read_u64 = mem_cgroup_move_charge_read,
6304 .write_u64 = mem_cgroup_move_charge_write,
6307 .name = "oom_control",
6308 .seq_show = mem_cgroup_oom_control_read,
6309 .write_u64 = mem_cgroup_oom_control_write,
6310 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
6313 .name = "pressure_level",
6317 .name = "numa_stat",
6318 .seq_show = memcg_numa_stat_show,
6321 #ifdef CONFIG_MEMCG_KMEM
6323 .name = "kmem.limit_in_bytes",
6324 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
6325 .write_string = mem_cgroup_write,
6326 .read_u64 = mem_cgroup_read_u64,
6329 .name = "kmem.usage_in_bytes",
6330 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
6331 .read_u64 = mem_cgroup_read_u64,
6334 .name = "kmem.failcnt",
6335 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
6336 .trigger = mem_cgroup_reset,
6337 .read_u64 = mem_cgroup_read_u64,
6340 .name = "kmem.max_usage_in_bytes",
6341 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
6342 .trigger = mem_cgroup_reset,
6343 .read_u64 = mem_cgroup_read_u64,
6345 #ifdef CONFIG_SLABINFO
6347 .name = "kmem.slabinfo",
6348 .seq_show = mem_cgroup_slabinfo_read,
6352 { }, /* terminate */
6355 #ifdef CONFIG_MEMCG_SWAP
6356 static struct cftype memsw_cgroup_files[] = {
6358 .name = "memsw.usage_in_bytes",
6359 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6360 .read_u64 = mem_cgroup_read_u64,
6363 .name = "memsw.max_usage_in_bytes",
6364 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6365 .trigger = mem_cgroup_reset,
6366 .read_u64 = mem_cgroup_read_u64,
6369 .name = "memsw.limit_in_bytes",
6370 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6371 .write_string = mem_cgroup_write,
6372 .read_u64 = mem_cgroup_read_u64,
6375 .name = "memsw.failcnt",
6376 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6377 .trigger = mem_cgroup_reset,
6378 .read_u64 = mem_cgroup_read_u64,
6380 { }, /* terminate */
6383 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6385 struct mem_cgroup_per_node *pn;
6386 struct mem_cgroup_per_zone *mz;
6387 int zone, tmp = node;
6389 * This routine is called against possible nodes.
6390 * But it's BUG to call kmalloc() against offline node.
6392 * TODO: this routine can waste much memory for nodes which will
6393 * never be onlined. It's better to use memory hotplug callback
6396 if (!node_state(node, N_NORMAL_MEMORY))
6398 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
6402 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6403 mz = &pn->zoneinfo[zone];
6404 lruvec_init(&mz->lruvec);
6405 mz->usage_in_excess = 0;
6406 mz->on_tree = false;
6409 memcg->nodeinfo[node] = pn;
6413 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6415 kfree(memcg->nodeinfo[node]);
6418 static struct mem_cgroup *mem_cgroup_alloc(void)
6420 struct mem_cgroup *memcg;
6423 size = sizeof(struct mem_cgroup);
6424 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
6426 memcg = kzalloc(size, GFP_KERNEL);
6430 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
6433 spin_lock_init(&memcg->pcp_counter_lock);
6442 * At destroying mem_cgroup, references from swap_cgroup can remain.
6443 * (scanning all at force_empty is too costly...)
6445 * Instead of clearing all references at force_empty, we remember
6446 * the number of reference from swap_cgroup and free mem_cgroup when
6447 * it goes down to 0.
6449 * Removal of cgroup itself succeeds regardless of refs from swap.
6452 static void __mem_cgroup_free(struct mem_cgroup *memcg)
6456 mem_cgroup_remove_from_trees(memcg);
6459 free_mem_cgroup_per_zone_info(memcg, node);
6461 free_percpu(memcg->stat);
6464 * We need to make sure that (at least for now), the jump label
6465 * destruction code runs outside of the cgroup lock. This is because
6466 * get_online_cpus(), which is called from the static_branch update,
6467 * can't be called inside the cgroup_lock. cpusets are the ones
6468 * enforcing this dependency, so if they ever change, we might as well.
6470 * schedule_work() will guarantee this happens. Be careful if you need
6471 * to move this code around, and make sure it is outside
6474 disarm_static_keys(memcg);
6479 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6481 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
6483 if (!memcg->res.parent)
6485 return mem_cgroup_from_res_counter(memcg->res.parent, res);
6487 EXPORT_SYMBOL(parent_mem_cgroup);
6489 static void __init mem_cgroup_soft_limit_tree_init(void)
6491 struct mem_cgroup_tree_per_node *rtpn;
6492 struct mem_cgroup_tree_per_zone *rtpz;
6493 int tmp, node, zone;
6495 for_each_node(node) {
6497 if (!node_state(node, N_NORMAL_MEMORY))
6499 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
6502 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6504 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6505 rtpz = &rtpn->rb_tree_per_zone[zone];
6506 rtpz->rb_root = RB_ROOT;
6507 spin_lock_init(&rtpz->lock);
6512 static struct cgroup_subsys_state * __ref
6513 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
6515 struct mem_cgroup *memcg;
6516 long error = -ENOMEM;
6519 memcg = mem_cgroup_alloc();
6521 return ERR_PTR(error);
6524 if (alloc_mem_cgroup_per_zone_info(memcg, node))
6528 if (parent_css == NULL) {
6529 root_mem_cgroup = memcg;
6530 res_counter_init(&memcg->res, NULL);
6531 res_counter_init(&memcg->memsw, NULL);
6532 res_counter_init(&memcg->kmem, NULL);
6535 memcg->last_scanned_node = MAX_NUMNODES;
6536 INIT_LIST_HEAD(&memcg->oom_notify);
6537 memcg->move_charge_at_immigrate = 0;
6538 mutex_init(&memcg->thresholds_lock);
6539 spin_lock_init(&memcg->move_lock);
6540 vmpressure_init(&memcg->vmpressure);
6541 INIT_LIST_HEAD(&memcg->event_list);
6542 spin_lock_init(&memcg->event_list_lock);
6547 __mem_cgroup_free(memcg);
6548 return ERR_PTR(error);
6552 mem_cgroup_css_online(struct cgroup_subsys_state *css)
6554 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6555 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(css));
6558 if (css->cgroup->id > MEM_CGROUP_ID_MAX)
6564 mutex_lock(&memcg_create_mutex);
6566 memcg->use_hierarchy = parent->use_hierarchy;
6567 memcg->oom_kill_disable = parent->oom_kill_disable;
6568 memcg->swappiness = mem_cgroup_swappiness(parent);
6570 if (parent->use_hierarchy) {
6571 res_counter_init(&memcg->res, &parent->res);
6572 res_counter_init(&memcg->memsw, &parent->memsw);
6573 res_counter_init(&memcg->kmem, &parent->kmem);
6576 * No need to take a reference to the parent because cgroup
6577 * core guarantees its existence.
6580 res_counter_init(&memcg->res, NULL);
6581 res_counter_init(&memcg->memsw, NULL);
6582 res_counter_init(&memcg->kmem, NULL);
6584 * Deeper hierachy with use_hierarchy == false doesn't make
6585 * much sense so let cgroup subsystem know about this
6586 * unfortunate state in our controller.
6588 if (parent != root_mem_cgroup)
6589 mem_cgroup_subsys.broken_hierarchy = true;
6591 mutex_unlock(&memcg_create_mutex);
6593 ret = memcg_init_kmem(memcg, &mem_cgroup_subsys);
6598 * Make sure the memcg is initialized: mem_cgroup_iter()
6599 * orders reading memcg->initialized against its callers
6600 * reading the memcg members.
6602 smp_store_release(&memcg->initialized, 1);
6608 * Announce all parents that a group from their hierarchy is gone.
6610 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup *memcg)
6612 struct mem_cgroup *parent = memcg;
6614 while ((parent = parent_mem_cgroup(parent)))
6615 mem_cgroup_iter_invalidate(parent);
6618 * if the root memcg is not hierarchical we have to check it
6621 if (!root_mem_cgroup->use_hierarchy)
6622 mem_cgroup_iter_invalidate(root_mem_cgroup);
6625 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
6627 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6628 struct mem_cgroup_event *event, *tmp;
6629 struct cgroup_subsys_state *iter;
6632 * Unregister events and notify userspace.
6633 * Notify userspace about cgroup removing only after rmdir of cgroup
6634 * directory to avoid race between userspace and kernelspace.
6636 spin_lock(&memcg->event_list_lock);
6637 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
6638 list_del_init(&event->list);
6639 schedule_work(&event->remove);
6641 spin_unlock(&memcg->event_list_lock);
6643 kmem_cgroup_css_offline(memcg);
6645 mem_cgroup_invalidate_reclaim_iterators(memcg);
6648 * This requires that offlining is serialized. Right now that is
6649 * guaranteed because css_killed_work_fn() holds the cgroup_mutex.
6651 css_for_each_descendant_post(iter, css)
6652 mem_cgroup_reparent_charges(mem_cgroup_from_css(iter));
6654 mem_cgroup_destroy_all_caches(memcg);
6655 vmpressure_cleanup(&memcg->vmpressure);
6658 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
6660 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6662 * XXX: css_offline() would be where we should reparent all
6663 * memory to prepare the cgroup for destruction. However,
6664 * memcg does not do css_tryget() and res_counter charging
6665 * under the same RCU lock region, which means that charging
6666 * could race with offlining. Offlining only happens to
6667 * cgroups with no tasks in them but charges can show up
6668 * without any tasks from the swapin path when the target
6669 * memcg is looked up from the swapout record and not from the
6670 * current task as it usually is. A race like this can leak
6671 * charges and put pages with stale cgroup pointers into
6675 * lookup_swap_cgroup_id()
6677 * mem_cgroup_lookup()
6680 * disable css_tryget()
6683 * reparent_charges()
6684 * res_counter_charge()
6687 * pc->mem_cgroup = dead memcg
6690 * The bulk of the charges are still moved in offline_css() to
6691 * avoid pinning a lot of pages in case a long-term reference
6692 * like a swapout record is deferring the css_free() to long
6693 * after offlining. But this makes sure we catch any charges
6694 * made after offlining:
6696 mem_cgroup_reparent_charges(memcg);
6698 memcg_destroy_kmem(memcg);
6699 __mem_cgroup_free(memcg);
6703 /* Handlers for move charge at task migration. */
6704 #define PRECHARGE_COUNT_AT_ONCE 256
6705 static int mem_cgroup_do_precharge(unsigned long count)
6708 int batch_count = PRECHARGE_COUNT_AT_ONCE;
6709 struct mem_cgroup *memcg = mc.to;
6711 if (mem_cgroup_is_root(memcg)) {
6712 mc.precharge += count;
6713 /* we don't need css_get for root */
6716 /* try to charge at once */
6718 struct res_counter *dummy;
6720 * "memcg" cannot be under rmdir() because we've already checked
6721 * by cgroup_lock_live_cgroup() that it is not removed and we
6722 * are still under the same cgroup_mutex. So we can postpone
6725 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
6727 if (do_swap_account && res_counter_charge(&memcg->memsw,
6728 PAGE_SIZE * count, &dummy)) {
6729 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
6732 mc.precharge += count;
6736 /* fall back to one by one charge */
6738 if (signal_pending(current)) {
6742 if (!batch_count--) {
6743 batch_count = PRECHARGE_COUNT_AT_ONCE;
6746 ret = __mem_cgroup_try_charge(NULL,
6747 GFP_KERNEL, 1, &memcg, false);
6749 /* mem_cgroup_clear_mc() will do uncharge later */
6757 * get_mctgt_type - get target type of moving charge
6758 * @vma: the vma the pte to be checked belongs
6759 * @addr: the address corresponding to the pte to be checked
6760 * @ptent: the pte to be checked
6761 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6764 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6765 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6766 * move charge. if @target is not NULL, the page is stored in target->page
6767 * with extra refcnt got(Callers should handle it).
6768 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6769 * target for charge migration. if @target is not NULL, the entry is stored
6772 * Called with pte lock held.
6779 enum mc_target_type {
6785 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
6786 unsigned long addr, pte_t ptent)
6788 struct page *page = vm_normal_page(vma, addr, ptent);
6790 if (!page || !page_mapped(page))
6792 if (PageAnon(page)) {
6793 /* we don't move shared anon */
6796 } else if (!move_file())
6797 /* we ignore mapcount for file pages */
6799 if (!get_page_unless_zero(page))
6806 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6807 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6809 struct page *page = NULL;
6810 swp_entry_t ent = pte_to_swp_entry(ptent);
6812 if (!move_anon() || non_swap_entry(ent))
6815 * Because lookup_swap_cache() updates some statistics counter,
6816 * we call find_get_page() with swapper_space directly.
6818 page = find_get_page(swap_address_space(ent), ent.val);
6819 if (do_swap_account)
6820 entry->val = ent.val;
6825 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6826 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6832 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
6833 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6835 struct page *page = NULL;
6836 struct address_space *mapping;
6839 if (!vma->vm_file) /* anonymous vma */
6844 mapping = vma->vm_file->f_mapping;
6845 if (pte_none(ptent))
6846 pgoff = linear_page_index(vma, addr);
6847 else /* pte_file(ptent) is true */
6848 pgoff = pte_to_pgoff(ptent);
6850 /* page is moved even if it's not RSS of this task(page-faulted). */
6851 page = find_get_page(mapping, pgoff);
6854 /* shmem/tmpfs may report page out on swap: account for that too. */
6855 if (radix_tree_exceptional_entry(page)) {
6856 swp_entry_t swap = radix_to_swp_entry(page);
6857 if (do_swap_account)
6859 page = find_get_page(swap_address_space(swap), swap.val);
6865 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
6866 unsigned long addr, pte_t ptent, union mc_target *target)
6868 struct page *page = NULL;
6869 struct page_cgroup *pc;
6870 enum mc_target_type ret = MC_TARGET_NONE;
6871 swp_entry_t ent = { .val = 0 };
6873 if (pte_present(ptent))
6874 page = mc_handle_present_pte(vma, addr, ptent);
6875 else if (is_swap_pte(ptent))
6876 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
6877 else if (pte_none(ptent) || pte_file(ptent))
6878 page = mc_handle_file_pte(vma, addr, ptent, &ent);
6880 if (!page && !ent.val)
6883 pc = lookup_page_cgroup(page);
6885 * Do only loose check w/o page_cgroup lock.
6886 * mem_cgroup_move_account() checks the pc is valid or not under
6889 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6890 ret = MC_TARGET_PAGE;
6892 target->page = page;
6894 if (!ret || !target)
6897 /* There is a swap entry and a page doesn't exist or isn't charged */
6898 if (ent.val && !ret &&
6899 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
6900 ret = MC_TARGET_SWAP;
6907 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6909 * We don't consider swapping or file mapped pages because THP does not
6910 * support them for now.
6911 * Caller should make sure that pmd_trans_huge(pmd) is true.
6913 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6914 unsigned long addr, pmd_t pmd, union mc_target *target)
6916 struct page *page = NULL;
6917 struct page_cgroup *pc;
6918 enum mc_target_type ret = MC_TARGET_NONE;
6920 page = pmd_page(pmd);
6921 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
6924 pc = lookup_page_cgroup(page);
6925 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6926 ret = MC_TARGET_PAGE;
6929 target->page = page;
6935 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6936 unsigned long addr, pmd_t pmd, union mc_target *target)
6938 return MC_TARGET_NONE;
6942 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
6943 unsigned long addr, unsigned long end,
6944 struct mm_walk *walk)
6946 struct vm_area_struct *vma = walk->private;
6950 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
6951 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
6952 mc.precharge += HPAGE_PMD_NR;
6957 if (pmd_trans_unstable(pmd))
6959 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6960 for (; addr != end; pte++, addr += PAGE_SIZE)
6961 if (get_mctgt_type(vma, addr, *pte, NULL))
6962 mc.precharge++; /* increment precharge temporarily */
6963 pte_unmap_unlock(pte - 1, ptl);
6969 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
6971 unsigned long precharge;
6972 struct vm_area_struct *vma;
6974 down_read(&mm->mmap_sem);
6975 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6976 struct mm_walk mem_cgroup_count_precharge_walk = {
6977 .pmd_entry = mem_cgroup_count_precharge_pte_range,
6981 if (is_vm_hugetlb_page(vma))
6983 walk_page_range(vma->vm_start, vma->vm_end,
6984 &mem_cgroup_count_precharge_walk);
6986 up_read(&mm->mmap_sem);
6988 precharge = mc.precharge;
6994 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
6996 unsigned long precharge = mem_cgroup_count_precharge(mm);
6998 VM_BUG_ON(mc.moving_task);
6999 mc.moving_task = current;
7000 return mem_cgroup_do_precharge(precharge);
7003 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
7004 static void __mem_cgroup_clear_mc(void)
7006 struct mem_cgroup *from = mc.from;
7007 struct mem_cgroup *to = mc.to;
7010 /* we must uncharge all the leftover precharges from mc.to */
7012 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
7016 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
7017 * we must uncharge here.
7019 if (mc.moved_charge) {
7020 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
7021 mc.moved_charge = 0;
7023 /* we must fixup refcnts and charges */
7024 if (mc.moved_swap) {
7025 /* uncharge swap account from the old cgroup */
7026 if (!mem_cgroup_is_root(mc.from))
7027 res_counter_uncharge(&mc.from->memsw,
7028 PAGE_SIZE * mc.moved_swap);
7030 for (i = 0; i < mc.moved_swap; i++)
7031 css_put(&mc.from->css);
7033 if (!mem_cgroup_is_root(mc.to)) {
7035 * we charged both to->res and to->memsw, so we should
7038 res_counter_uncharge(&mc.to->res,
7039 PAGE_SIZE * mc.moved_swap);
7041 /* we've already done css_get(mc.to) */
7044 memcg_oom_recover(from);
7045 memcg_oom_recover(to);
7046 wake_up_all(&mc.waitq);
7049 static void mem_cgroup_clear_mc(void)
7051 struct mem_cgroup *from = mc.from;
7054 * we must clear moving_task before waking up waiters at the end of
7057 mc.moving_task = NULL;
7058 __mem_cgroup_clear_mc();
7059 spin_lock(&mc.lock);
7062 spin_unlock(&mc.lock);
7063 mem_cgroup_end_move(from);
7066 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
7067 struct cgroup_taskset *tset)
7069 struct task_struct *p = cgroup_taskset_first(tset);
7071 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7072 unsigned long move_charge_at_immigrate;
7075 * We are now commited to this value whatever it is. Changes in this
7076 * tunable will only affect upcoming migrations, not the current one.
7077 * So we need to save it, and keep it going.
7079 move_charge_at_immigrate = memcg->move_charge_at_immigrate;
7080 if (move_charge_at_immigrate) {
7081 struct mm_struct *mm;
7082 struct mem_cgroup *from = mem_cgroup_from_task(p);
7084 VM_BUG_ON(from == memcg);
7086 mm = get_task_mm(p);
7089 /* We move charges only when we move a owner of the mm */
7090 if (mm->owner == p) {
7093 VM_BUG_ON(mc.precharge);
7094 VM_BUG_ON(mc.moved_charge);
7095 VM_BUG_ON(mc.moved_swap);
7096 mem_cgroup_start_move(from);
7097 spin_lock(&mc.lock);
7100 mc.immigrate_flags = move_charge_at_immigrate;
7101 spin_unlock(&mc.lock);
7102 /* We set mc.moving_task later */
7104 ret = mem_cgroup_precharge_mc(mm);
7106 mem_cgroup_clear_mc();
7113 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
7114 struct cgroup_taskset *tset)
7116 mem_cgroup_clear_mc();
7119 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
7120 unsigned long addr, unsigned long end,
7121 struct mm_walk *walk)
7124 struct vm_area_struct *vma = walk->private;
7127 enum mc_target_type target_type;
7128 union mc_target target;
7130 struct page_cgroup *pc;
7133 * We don't take compound_lock() here but no race with splitting thp
7135 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
7136 * under splitting, which means there's no concurrent thp split,
7137 * - if another thread runs into split_huge_page() just after we
7138 * entered this if-block, the thread must wait for page table lock
7139 * to be unlocked in __split_huge_page_splitting(), where the main
7140 * part of thp split is not executed yet.
7142 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
7143 if (mc.precharge < HPAGE_PMD_NR) {
7147 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
7148 if (target_type == MC_TARGET_PAGE) {
7150 if (!isolate_lru_page(page)) {
7151 pc = lookup_page_cgroup(page);
7152 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
7153 pc, mc.from, mc.to)) {
7154 mc.precharge -= HPAGE_PMD_NR;
7155 mc.moved_charge += HPAGE_PMD_NR;
7157 putback_lru_page(page);
7165 if (pmd_trans_unstable(pmd))
7168 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
7169 for (; addr != end; addr += PAGE_SIZE) {
7170 pte_t ptent = *(pte++);
7176 switch (get_mctgt_type(vma, addr, ptent, &target)) {
7177 case MC_TARGET_PAGE:
7179 if (isolate_lru_page(page))
7181 pc = lookup_page_cgroup(page);
7182 if (!mem_cgroup_move_account(page, 1, pc,
7185 /* we uncharge from mc.from later. */
7188 putback_lru_page(page);
7189 put: /* get_mctgt_type() gets the page */
7192 case MC_TARGET_SWAP:
7194 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
7196 /* we fixup refcnts and charges later. */
7204 pte_unmap_unlock(pte - 1, ptl);
7209 * We have consumed all precharges we got in can_attach().
7210 * We try charge one by one, but don't do any additional
7211 * charges to mc.to if we have failed in charge once in attach()
7214 ret = mem_cgroup_do_precharge(1);
7222 static void mem_cgroup_move_charge(struct mm_struct *mm)
7224 struct vm_area_struct *vma;
7226 lru_add_drain_all();
7228 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
7230 * Someone who are holding the mmap_sem might be waiting in
7231 * waitq. So we cancel all extra charges, wake up all waiters,
7232 * and retry. Because we cancel precharges, we might not be able
7233 * to move enough charges, but moving charge is a best-effort
7234 * feature anyway, so it wouldn't be a big problem.
7236 __mem_cgroup_clear_mc();
7240 for (vma = mm->mmap; vma; vma = vma->vm_next) {
7242 struct mm_walk mem_cgroup_move_charge_walk = {
7243 .pmd_entry = mem_cgroup_move_charge_pte_range,
7247 if (is_vm_hugetlb_page(vma))
7249 ret = walk_page_range(vma->vm_start, vma->vm_end,
7250 &mem_cgroup_move_charge_walk);
7253 * means we have consumed all precharges and failed in
7254 * doing additional charge. Just abandon here.
7258 up_read(&mm->mmap_sem);
7261 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
7262 struct cgroup_taskset *tset)
7264 struct task_struct *p = cgroup_taskset_first(tset);
7265 struct mm_struct *mm = get_task_mm(p);
7269 mem_cgroup_move_charge(mm);
7273 mem_cgroup_clear_mc();
7275 #else /* !CONFIG_MMU */
7276 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
7277 struct cgroup_taskset *tset)
7281 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
7282 struct cgroup_taskset *tset)
7285 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
7286 struct cgroup_taskset *tset)
7292 * Cgroup retains root cgroups across [un]mount cycles making it necessary
7293 * to verify sane_behavior flag on each mount attempt.
7295 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
7298 * use_hierarchy is forced with sane_behavior. cgroup core
7299 * guarantees that @root doesn't have any children, so turning it
7300 * on for the root memcg is enough.
7302 if (cgroup_sane_behavior(root_css->cgroup))
7303 mem_cgroup_from_css(root_css)->use_hierarchy = true;
7306 struct cgroup_subsys mem_cgroup_subsys = {
7308 .subsys_id = mem_cgroup_subsys_id,
7309 .css_alloc = mem_cgroup_css_alloc,
7310 .css_online = mem_cgroup_css_online,
7311 .css_offline = mem_cgroup_css_offline,
7312 .css_free = mem_cgroup_css_free,
7313 .can_attach = mem_cgroup_can_attach,
7314 .cancel_attach = mem_cgroup_cancel_attach,
7315 .attach = mem_cgroup_move_task,
7316 .bind = mem_cgroup_bind,
7317 .base_cftypes = mem_cgroup_files,
7321 #ifdef CONFIG_MEMCG_SWAP
7322 static int __init enable_swap_account(char *s)
7324 if (!strcmp(s, "1"))
7325 really_do_swap_account = 1;
7326 else if (!strcmp(s, "0"))
7327 really_do_swap_account = 0;
7330 __setup("swapaccount=", enable_swap_account);
7332 static void __init memsw_file_init(void)
7334 WARN_ON(cgroup_add_cftypes(&mem_cgroup_subsys, memsw_cgroup_files));
7337 static void __init enable_swap_cgroup(void)
7339 if (!mem_cgroup_disabled() && really_do_swap_account) {
7340 do_swap_account = 1;
7346 static void __init enable_swap_cgroup(void)
7352 * subsys_initcall() for memory controller.
7354 * Some parts like hotcpu_notifier() have to be initialized from this context
7355 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
7356 * everything that doesn't depend on a specific mem_cgroup structure should
7357 * be initialized from here.
7359 static int __init mem_cgroup_init(void)
7361 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
7362 enable_swap_cgroup();
7363 mem_cgroup_soft_limit_tree_init();
7367 subsys_initcall(mem_cgroup_init);