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;
296 * the counter to account for mem+swap usage.
298 struct res_counter memsw;
301 * the counter to account for kernel memory usage.
303 struct res_counter kmem;
305 * Should the accounting and control be hierarchical, per subtree?
308 unsigned long kmem_account_flags; /* See KMEM_ACCOUNTED_*, below */
312 atomic_t oom_wakeups;
315 /* OOM-Killer disable */
316 int oom_kill_disable;
318 /* set when res.limit == memsw.limit */
319 bool memsw_is_minimum;
321 /* protect arrays of thresholds */
322 struct mutex thresholds_lock;
324 /* thresholds for memory usage. RCU-protected */
325 struct mem_cgroup_thresholds thresholds;
327 /* thresholds for mem+swap usage. RCU-protected */
328 struct mem_cgroup_thresholds memsw_thresholds;
330 /* For oom notifier event fd */
331 struct list_head oom_notify;
334 * Should we move charges of a task when a task is moved into this
335 * mem_cgroup ? And what type of charges should we move ?
337 unsigned long move_charge_at_immigrate;
339 * set > 0 if pages under this cgroup are moving to other cgroup.
341 atomic_t moving_account;
342 /* taken only while moving_account > 0 */
343 spinlock_t move_lock;
347 struct mem_cgroup_stat_cpu __percpu *stat;
349 * used when a cpu is offlined or other synchronizations
350 * See mem_cgroup_read_stat().
352 struct mem_cgroup_stat_cpu nocpu_base;
353 spinlock_t pcp_counter_lock;
356 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
357 struct cg_proto tcp_mem;
359 #if defined(CONFIG_MEMCG_KMEM)
360 /* analogous to slab_common's slab_caches list. per-memcg */
361 struct list_head memcg_slab_caches;
362 /* Not a spinlock, we can take a lot of time walking the list */
363 struct mutex slab_caches_mutex;
364 /* Index in the kmem_cache->memcg_params->memcg_caches array */
368 int last_scanned_node;
370 nodemask_t scan_nodes;
371 atomic_t numainfo_events;
372 atomic_t numainfo_updating;
375 /* List of events which userspace want to receive */
376 struct list_head event_list;
377 spinlock_t event_list_lock;
379 struct mem_cgroup_per_node *nodeinfo[0];
380 /* WARNING: nodeinfo must be the last member here */
383 /* internal only representation about the status of kmem accounting. */
385 KMEM_ACCOUNTED_ACTIVE, /* accounted by this cgroup itself */
386 KMEM_ACCOUNTED_DEAD, /* dead memcg with pending kmem charges */
389 #ifdef CONFIG_MEMCG_KMEM
390 static inline void memcg_kmem_set_active(struct mem_cgroup *memcg)
392 set_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
395 static bool memcg_kmem_is_active(struct mem_cgroup *memcg)
397 return test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
400 static void memcg_kmem_mark_dead(struct mem_cgroup *memcg)
403 * Our caller must use css_get() first, because memcg_uncharge_kmem()
404 * will call css_put() if it sees the memcg is dead.
407 if (test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags))
408 set_bit(KMEM_ACCOUNTED_DEAD, &memcg->kmem_account_flags);
411 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup *memcg)
413 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD,
414 &memcg->kmem_account_flags);
418 /* Stuffs for move charges at task migration. */
420 * Types of charges to be moved. "move_charge_at_immitgrate" and
421 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
424 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
425 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
429 /* "mc" and its members are protected by cgroup_mutex */
430 static struct move_charge_struct {
431 spinlock_t lock; /* for from, to */
432 struct mem_cgroup *from;
433 struct mem_cgroup *to;
434 unsigned long immigrate_flags;
435 unsigned long precharge;
436 unsigned long moved_charge;
437 unsigned long moved_swap;
438 struct task_struct *moving_task; /* a task moving charges */
439 wait_queue_head_t waitq; /* a waitq for other context */
441 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
442 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
445 static bool move_anon(void)
447 return test_bit(MOVE_CHARGE_TYPE_ANON, &mc.immigrate_flags);
450 static bool move_file(void)
452 return test_bit(MOVE_CHARGE_TYPE_FILE, &mc.immigrate_flags);
456 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
457 * limit reclaim to prevent infinite loops, if they ever occur.
459 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
460 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
463 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
464 MEM_CGROUP_CHARGE_TYPE_ANON,
465 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
466 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
470 /* for encoding cft->private value on file */
478 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
479 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
480 #define MEMFILE_ATTR(val) ((val) & 0xffff)
481 /* Used for OOM nofiier */
482 #define OOM_CONTROL (0)
485 * Reclaim flags for mem_cgroup_hierarchical_reclaim
487 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
488 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
489 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
490 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
493 * The memcg_create_mutex will be held whenever a new cgroup is created.
494 * As a consequence, any change that needs to protect against new child cgroups
495 * appearing has to hold it as well.
497 static DEFINE_MUTEX(memcg_create_mutex);
499 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
501 return s ? container_of(s, struct mem_cgroup, css) : NULL;
504 /* Some nice accessors for the vmpressure. */
505 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
508 memcg = root_mem_cgroup;
509 return &memcg->vmpressure;
512 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
514 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
517 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
519 return (memcg == root_mem_cgroup);
523 * We restrict the id in the range of [1, 65535], so it can fit into
526 #define MEM_CGROUP_ID_MAX USHRT_MAX
528 static inline unsigned short mem_cgroup_id(struct mem_cgroup *memcg)
531 * The ID of the root cgroup is 0, but memcg treat 0 as an
532 * invalid ID, so we return (cgroup_id + 1).
534 return memcg->css.cgroup->id + 1;
537 static inline struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
539 struct cgroup_subsys_state *css;
541 css = css_from_id(id - 1, &mem_cgroup_subsys);
542 return mem_cgroup_from_css(css);
545 /* Writing them here to avoid exposing memcg's inner layout */
546 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
548 void sock_update_memcg(struct sock *sk)
550 if (mem_cgroup_sockets_enabled) {
551 struct mem_cgroup *memcg;
552 struct cg_proto *cg_proto;
554 BUG_ON(!sk->sk_prot->proto_cgroup);
556 /* Socket cloning can throw us here with sk_cgrp already
557 * filled. It won't however, necessarily happen from
558 * process context. So the test for root memcg given
559 * the current task's memcg won't help us in this case.
561 * Respecting the original socket's memcg is a better
562 * decision in this case.
565 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
566 css_get(&sk->sk_cgrp->memcg->css);
571 memcg = mem_cgroup_from_task(current);
572 cg_proto = sk->sk_prot->proto_cgroup(memcg);
573 if (!mem_cgroup_is_root(memcg) &&
574 memcg_proto_active(cg_proto) && css_tryget(&memcg->css)) {
575 sk->sk_cgrp = cg_proto;
580 EXPORT_SYMBOL(sock_update_memcg);
582 void sock_release_memcg(struct sock *sk)
584 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
585 struct mem_cgroup *memcg;
586 WARN_ON(!sk->sk_cgrp->memcg);
587 memcg = sk->sk_cgrp->memcg;
588 css_put(&sk->sk_cgrp->memcg->css);
592 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
594 if (!memcg || mem_cgroup_is_root(memcg))
597 return &memcg->tcp_mem;
599 EXPORT_SYMBOL(tcp_proto_cgroup);
601 static void disarm_sock_keys(struct mem_cgroup *memcg)
603 if (!memcg_proto_activated(&memcg->tcp_mem))
605 static_key_slow_dec(&memcg_socket_limit_enabled);
608 static void disarm_sock_keys(struct mem_cgroup *memcg)
613 #ifdef CONFIG_MEMCG_KMEM
615 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
616 * The main reason for not using cgroup id for this:
617 * this works better in sparse environments, where we have a lot of memcgs,
618 * but only a few kmem-limited. Or also, if we have, for instance, 200
619 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
620 * 200 entry array for that.
622 * The current size of the caches array is stored in
623 * memcg_limited_groups_array_size. It will double each time we have to
626 static DEFINE_IDA(kmem_limited_groups);
627 int memcg_limited_groups_array_size;
630 * MIN_SIZE is different than 1, because we would like to avoid going through
631 * the alloc/free process all the time. In a small machine, 4 kmem-limited
632 * cgroups is a reasonable guess. In the future, it could be a parameter or
633 * tunable, but that is strictly not necessary.
635 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
636 * this constant directly from cgroup, but it is understandable that this is
637 * better kept as an internal representation in cgroup.c. In any case, the
638 * cgrp_id space is not getting any smaller, and we don't have to necessarily
639 * increase ours as well if it increases.
641 #define MEMCG_CACHES_MIN_SIZE 4
642 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
645 * A lot of the calls to the cache allocation functions are expected to be
646 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
647 * conditional to this static branch, we'll have to allow modules that does
648 * kmem_cache_alloc and the such to see this symbol as well
650 struct static_key memcg_kmem_enabled_key;
651 EXPORT_SYMBOL(memcg_kmem_enabled_key);
653 static void disarm_kmem_keys(struct mem_cgroup *memcg)
655 if (memcg_kmem_is_active(memcg)) {
656 static_key_slow_dec(&memcg_kmem_enabled_key);
657 ida_simple_remove(&kmem_limited_groups, memcg->kmemcg_id);
660 * This check can't live in kmem destruction function,
661 * since the charges will outlive the cgroup
663 WARN_ON(res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0);
666 static void disarm_kmem_keys(struct mem_cgroup *memcg)
669 #endif /* CONFIG_MEMCG_KMEM */
671 static void disarm_static_keys(struct mem_cgroup *memcg)
673 disarm_sock_keys(memcg);
674 disarm_kmem_keys(memcg);
677 static void drain_all_stock_async(struct mem_cgroup *memcg);
679 static struct mem_cgroup_per_zone *
680 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
682 VM_BUG_ON((unsigned)nid >= nr_node_ids);
683 return &memcg->nodeinfo[nid]->zoneinfo[zid];
686 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
691 static struct mem_cgroup_per_zone *
692 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
694 int nid = page_to_nid(page);
695 int zid = page_zonenum(page);
697 return mem_cgroup_zoneinfo(memcg, nid, zid);
700 static struct mem_cgroup_tree_per_zone *
701 soft_limit_tree_node_zone(int nid, int zid)
703 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
706 static struct mem_cgroup_tree_per_zone *
707 soft_limit_tree_from_page(struct page *page)
709 int nid = page_to_nid(page);
710 int zid = page_zonenum(page);
712 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
716 __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
717 struct mem_cgroup_per_zone *mz,
718 struct mem_cgroup_tree_per_zone *mctz,
719 unsigned long long new_usage_in_excess)
721 struct rb_node **p = &mctz->rb_root.rb_node;
722 struct rb_node *parent = NULL;
723 struct mem_cgroup_per_zone *mz_node;
728 mz->usage_in_excess = new_usage_in_excess;
729 if (!mz->usage_in_excess)
733 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
735 if (mz->usage_in_excess < mz_node->usage_in_excess)
738 * We can't avoid mem cgroups that are over their soft
739 * limit by the same amount
741 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
744 rb_link_node(&mz->tree_node, parent, p);
745 rb_insert_color(&mz->tree_node, &mctz->rb_root);
750 __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
751 struct mem_cgroup_per_zone *mz,
752 struct mem_cgroup_tree_per_zone *mctz)
756 rb_erase(&mz->tree_node, &mctz->rb_root);
761 mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
762 struct mem_cgroup_per_zone *mz,
763 struct mem_cgroup_tree_per_zone *mctz)
765 spin_lock(&mctz->lock);
766 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
767 spin_unlock(&mctz->lock);
771 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
773 unsigned long long excess;
774 struct mem_cgroup_per_zone *mz;
775 struct mem_cgroup_tree_per_zone *mctz;
776 int nid = page_to_nid(page);
777 int zid = page_zonenum(page);
778 mctz = soft_limit_tree_from_page(page);
781 * Necessary to update all ancestors when hierarchy is used.
782 * because their event counter is not touched.
784 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
785 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
786 excess = res_counter_soft_limit_excess(&memcg->res);
788 * We have to update the tree if mz is on RB-tree or
789 * mem is over its softlimit.
791 if (excess || mz->on_tree) {
792 spin_lock(&mctz->lock);
793 /* if on-tree, remove it */
795 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
797 * Insert again. mz->usage_in_excess will be updated.
798 * If excess is 0, no tree ops.
800 __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
801 spin_unlock(&mctz->lock);
806 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
809 struct mem_cgroup_per_zone *mz;
810 struct mem_cgroup_tree_per_zone *mctz;
812 for_each_node(node) {
813 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
814 mz = mem_cgroup_zoneinfo(memcg, node, zone);
815 mctz = soft_limit_tree_node_zone(node, zone);
816 mem_cgroup_remove_exceeded(memcg, mz, mctz);
821 static struct mem_cgroup_per_zone *
822 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
824 struct rb_node *rightmost = NULL;
825 struct mem_cgroup_per_zone *mz;
829 rightmost = rb_last(&mctz->rb_root);
831 goto done; /* Nothing to reclaim from */
833 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
835 * Remove the node now but someone else can add it back,
836 * we will to add it back at the end of reclaim to its correct
837 * position in the tree.
839 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
840 if (!res_counter_soft_limit_excess(&mz->memcg->res) ||
841 !css_tryget(&mz->memcg->css))
847 static struct mem_cgroup_per_zone *
848 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
850 struct mem_cgroup_per_zone *mz;
852 spin_lock(&mctz->lock);
853 mz = __mem_cgroup_largest_soft_limit_node(mctz);
854 spin_unlock(&mctz->lock);
859 * Implementation Note: reading percpu statistics for memcg.
861 * Both of vmstat[] and percpu_counter has threshold and do periodic
862 * synchronization to implement "quick" read. There are trade-off between
863 * reading cost and precision of value. Then, we may have a chance to implement
864 * a periodic synchronizion of counter in memcg's counter.
866 * But this _read() function is used for user interface now. The user accounts
867 * memory usage by memory cgroup and he _always_ requires exact value because
868 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
869 * have to visit all online cpus and make sum. So, for now, unnecessary
870 * synchronization is not implemented. (just implemented for cpu hotplug)
872 * If there are kernel internal actions which can make use of some not-exact
873 * value, and reading all cpu value can be performance bottleneck in some
874 * common workload, threashold and synchonization as vmstat[] should be
877 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
878 enum mem_cgroup_stat_index idx)
884 for_each_online_cpu(cpu)
885 val += per_cpu(memcg->stat->count[idx], cpu);
886 #ifdef CONFIG_HOTPLUG_CPU
887 spin_lock(&memcg->pcp_counter_lock);
888 val += memcg->nocpu_base.count[idx];
889 spin_unlock(&memcg->pcp_counter_lock);
895 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
898 int val = (charge) ? 1 : -1;
899 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
902 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
903 enum mem_cgroup_events_index idx)
905 unsigned long val = 0;
909 for_each_online_cpu(cpu)
910 val += per_cpu(memcg->stat->events[idx], cpu);
911 #ifdef CONFIG_HOTPLUG_CPU
912 spin_lock(&memcg->pcp_counter_lock);
913 val += memcg->nocpu_base.events[idx];
914 spin_unlock(&memcg->pcp_counter_lock);
920 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
922 bool anon, int nr_pages)
927 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
928 * counted as CACHE even if it's on ANON LRU.
931 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
934 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
937 if (PageTransHuge(page))
938 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
941 /* pagein of a big page is an event. So, ignore page size */
943 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
945 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
946 nr_pages = -nr_pages; /* for event */
949 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
955 mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
957 struct mem_cgroup_per_zone *mz;
959 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
960 return mz->lru_size[lru];
964 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
965 unsigned int lru_mask)
967 struct mem_cgroup_per_zone *mz;
969 unsigned long ret = 0;
971 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
974 if (BIT(lru) & lru_mask)
975 ret += mz->lru_size[lru];
981 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
982 int nid, unsigned int lru_mask)
987 for (zid = 0; zid < MAX_NR_ZONES; zid++)
988 total += mem_cgroup_zone_nr_lru_pages(memcg,
994 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
995 unsigned int lru_mask)
1000 for_each_node_state(nid, N_MEMORY)
1001 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
1005 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
1006 enum mem_cgroup_events_target target)
1008 unsigned long val, next;
1010 val = __this_cpu_read(memcg->stat->nr_page_events);
1011 next = __this_cpu_read(memcg->stat->targets[target]);
1012 /* from time_after() in jiffies.h */
1013 if ((long)next - (long)val < 0) {
1015 case MEM_CGROUP_TARGET_THRESH:
1016 next = val + THRESHOLDS_EVENTS_TARGET;
1018 case MEM_CGROUP_TARGET_SOFTLIMIT:
1019 next = val + SOFTLIMIT_EVENTS_TARGET;
1021 case MEM_CGROUP_TARGET_NUMAINFO:
1022 next = val + NUMAINFO_EVENTS_TARGET;
1027 __this_cpu_write(memcg->stat->targets[target], next);
1034 * Check events in order.
1037 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
1040 /* threshold event is triggered in finer grain than soft limit */
1041 if (unlikely(mem_cgroup_event_ratelimit(memcg,
1042 MEM_CGROUP_TARGET_THRESH))) {
1044 bool do_numainfo __maybe_unused;
1046 do_softlimit = mem_cgroup_event_ratelimit(memcg,
1047 MEM_CGROUP_TARGET_SOFTLIMIT);
1048 #if MAX_NUMNODES > 1
1049 do_numainfo = mem_cgroup_event_ratelimit(memcg,
1050 MEM_CGROUP_TARGET_NUMAINFO);
1054 mem_cgroup_threshold(memcg);
1055 if (unlikely(do_softlimit))
1056 mem_cgroup_update_tree(memcg, page);
1057 #if MAX_NUMNODES > 1
1058 if (unlikely(do_numainfo))
1059 atomic_inc(&memcg->numainfo_events);
1065 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1068 * mm_update_next_owner() may clear mm->owner to NULL
1069 * if it races with swapoff, page migration, etc.
1070 * So this can be called with p == NULL.
1075 return mem_cgroup_from_css(task_css(p, mem_cgroup_subsys_id));
1078 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
1080 struct mem_cgroup *memcg = NULL;
1085 * Because we have no locks, mm->owner's may be being moved to other
1086 * cgroup. We use css_tryget() here even if this looks
1087 * pessimistic (rather than adding locks here).
1091 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1092 if (unlikely(!memcg))
1094 } while (!css_tryget(&memcg->css));
1100 * Returns a next (in a pre-order walk) alive memcg (with elevated css
1101 * ref. count) or NULL if the whole root's subtree has been visited.
1103 * helper function to be used by mem_cgroup_iter
1105 static struct mem_cgroup *__mem_cgroup_iter_next(struct mem_cgroup *root,
1106 struct mem_cgroup *last_visited)
1108 struct cgroup_subsys_state *prev_css, *next_css;
1110 prev_css = last_visited ? &last_visited->css : NULL;
1112 next_css = css_next_descendant_pre(prev_css, &root->css);
1115 * Even if we found a group we have to make sure it is
1116 * alive. css && !memcg means that the groups should be
1117 * skipped and we should continue the tree walk.
1118 * last_visited css is safe to use because it is
1119 * protected by css_get and the tree walk is rcu safe.
1122 if ((next_css->flags & CSS_ONLINE) && css_tryget(next_css))
1123 return mem_cgroup_from_css(next_css);
1125 prev_css = next_css;
1133 static void mem_cgroup_iter_invalidate(struct mem_cgroup *root)
1136 * When a group in the hierarchy below root is destroyed, the
1137 * hierarchy iterator can no longer be trusted since it might
1138 * have pointed to the destroyed group. Invalidate it.
1140 atomic_inc(&root->dead_count);
1143 static struct mem_cgroup *
1144 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter *iter,
1145 struct mem_cgroup *root,
1148 struct mem_cgroup *position = NULL;
1150 * A cgroup destruction happens in two stages: offlining and
1151 * release. They are separated by a RCU grace period.
1153 * If the iterator is valid, we may still race with an
1154 * offlining. The RCU lock ensures the object won't be
1155 * released, tryget will fail if we lost the race.
1157 *sequence = atomic_read(&root->dead_count);
1158 if (iter->last_dead_count == *sequence) {
1160 position = iter->last_visited;
1161 if (position && !css_tryget(&position->css))
1167 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter *iter,
1168 struct mem_cgroup *last_visited,
1169 struct mem_cgroup *new_position,
1173 css_put(&last_visited->css);
1175 * We store the sequence count from the time @last_visited was
1176 * loaded successfully instead of rereading it here so that we
1177 * don't lose destruction events in between. We could have
1178 * raced with the destruction of @new_position after all.
1180 iter->last_visited = new_position;
1182 iter->last_dead_count = sequence;
1186 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1187 * @root: hierarchy root
1188 * @prev: previously returned memcg, NULL on first invocation
1189 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1191 * Returns references to children of the hierarchy below @root, or
1192 * @root itself, or %NULL after a full round-trip.
1194 * Caller must pass the return value in @prev on subsequent
1195 * invocations for reference counting, or use mem_cgroup_iter_break()
1196 * to cancel a hierarchy walk before the round-trip is complete.
1198 * Reclaimers can specify a zone and a priority level in @reclaim to
1199 * divide up the memcgs in the hierarchy among all concurrent
1200 * reclaimers operating on the same zone and priority.
1202 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1203 struct mem_cgroup *prev,
1204 struct mem_cgroup_reclaim_cookie *reclaim)
1206 struct mem_cgroup *memcg = NULL;
1207 struct mem_cgroup *last_visited = NULL;
1209 if (mem_cgroup_disabled())
1213 root = root_mem_cgroup;
1215 if (prev && !reclaim)
1216 last_visited = prev;
1218 if (!root->use_hierarchy && root != root_mem_cgroup) {
1226 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1227 int uninitialized_var(seq);
1230 int nid = zone_to_nid(reclaim->zone);
1231 int zid = zone_idx(reclaim->zone);
1232 struct mem_cgroup_per_zone *mz;
1234 mz = mem_cgroup_zoneinfo(root, nid, zid);
1235 iter = &mz->reclaim_iter[reclaim->priority];
1236 if (prev && reclaim->generation != iter->generation) {
1237 iter->last_visited = NULL;
1241 last_visited = mem_cgroup_iter_load(iter, root, &seq);
1244 memcg = __mem_cgroup_iter_next(root, last_visited);
1247 mem_cgroup_iter_update(iter, last_visited, memcg, seq);
1251 else if (!prev && memcg)
1252 reclaim->generation = iter->generation;
1261 if (prev && prev != root)
1262 css_put(&prev->css);
1268 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1269 * @root: hierarchy root
1270 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1272 void mem_cgroup_iter_break(struct mem_cgroup *root,
1273 struct mem_cgroup *prev)
1276 root = root_mem_cgroup;
1277 if (prev && prev != root)
1278 css_put(&prev->css);
1282 * Iteration constructs for visiting all cgroups (under a tree). If
1283 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1284 * be used for reference counting.
1286 #define for_each_mem_cgroup_tree(iter, root) \
1287 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1289 iter = mem_cgroup_iter(root, iter, NULL))
1291 #define for_each_mem_cgroup(iter) \
1292 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1294 iter = mem_cgroup_iter(NULL, iter, NULL))
1296 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1298 struct mem_cgroup *memcg;
1301 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1302 if (unlikely(!memcg))
1307 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1310 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1318 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1321 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1322 * @zone: zone of the wanted lruvec
1323 * @memcg: memcg of the wanted lruvec
1325 * Returns the lru list vector holding pages for the given @zone and
1326 * @mem. This can be the global zone lruvec, if the memory controller
1329 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1330 struct mem_cgroup *memcg)
1332 struct mem_cgroup_per_zone *mz;
1333 struct lruvec *lruvec;
1335 if (mem_cgroup_disabled()) {
1336 lruvec = &zone->lruvec;
1340 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1341 lruvec = &mz->lruvec;
1344 * Since a node can be onlined after the mem_cgroup was created,
1345 * we have to be prepared to initialize lruvec->zone here;
1346 * and if offlined then reonlined, we need to reinitialize it.
1348 if (unlikely(lruvec->zone != zone))
1349 lruvec->zone = zone;
1354 * Following LRU functions are allowed to be used without PCG_LOCK.
1355 * Operations are called by routine of global LRU independently from memcg.
1356 * What we have to take care of here is validness of pc->mem_cgroup.
1358 * Changes to pc->mem_cgroup happens when
1361 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1362 * It is added to LRU before charge.
1363 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1364 * When moving account, the page is not on LRU. It's isolated.
1368 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1370 * @zone: zone of the page
1372 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1374 struct mem_cgroup_per_zone *mz;
1375 struct mem_cgroup *memcg;
1376 struct page_cgroup *pc;
1377 struct lruvec *lruvec;
1379 if (mem_cgroup_disabled()) {
1380 lruvec = &zone->lruvec;
1384 pc = lookup_page_cgroup(page);
1385 memcg = pc->mem_cgroup;
1388 * Surreptitiously switch any uncharged offlist page to root:
1389 * an uncharged page off lru does nothing to secure
1390 * its former mem_cgroup from sudden removal.
1392 * Our caller holds lru_lock, and PageCgroupUsed is updated
1393 * under page_cgroup lock: between them, they make all uses
1394 * of pc->mem_cgroup safe.
1396 if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1397 pc->mem_cgroup = memcg = root_mem_cgroup;
1399 mz = page_cgroup_zoneinfo(memcg, page);
1400 lruvec = &mz->lruvec;
1403 * Since a node can be onlined after the mem_cgroup was created,
1404 * we have to be prepared to initialize lruvec->zone here;
1405 * and if offlined then reonlined, we need to reinitialize it.
1407 if (unlikely(lruvec->zone != zone))
1408 lruvec->zone = zone;
1413 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1414 * @lruvec: mem_cgroup per zone lru vector
1415 * @lru: index of lru list the page is sitting on
1416 * @nr_pages: positive when adding or negative when removing
1418 * This function must be called when a page is added to or removed from an
1421 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1424 struct mem_cgroup_per_zone *mz;
1425 unsigned long *lru_size;
1427 if (mem_cgroup_disabled())
1430 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1431 lru_size = mz->lru_size + lru;
1432 *lru_size += nr_pages;
1433 VM_BUG_ON((long)(*lru_size) < 0);
1437 * Checks whether given mem is same or in the root_mem_cgroup's
1440 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1441 struct mem_cgroup *memcg)
1443 if (root_memcg == memcg)
1445 if (!root_memcg->use_hierarchy || !memcg)
1447 return cgroup_is_descendant(memcg->css.cgroup, root_memcg->css.cgroup);
1450 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1451 struct mem_cgroup *memcg)
1456 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1461 bool task_in_mem_cgroup(struct task_struct *task,
1462 const struct mem_cgroup *memcg)
1464 struct mem_cgroup *curr = NULL;
1465 struct task_struct *p;
1468 p = find_lock_task_mm(task);
1470 curr = try_get_mem_cgroup_from_mm(p->mm);
1474 * All threads may have already detached their mm's, but the oom
1475 * killer still needs to detect if they have already been oom
1476 * killed to prevent needlessly killing additional tasks.
1479 curr = mem_cgroup_from_task(task);
1481 css_get(&curr->css);
1487 * We should check use_hierarchy of "memcg" not "curr". Because checking
1488 * use_hierarchy of "curr" here make this function true if hierarchy is
1489 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1490 * hierarchy(even if use_hierarchy is disabled in "memcg").
1492 ret = mem_cgroup_same_or_subtree(memcg, curr);
1493 css_put(&curr->css);
1497 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1499 unsigned long inactive_ratio;
1500 unsigned long inactive;
1501 unsigned long active;
1504 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1505 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1507 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1509 inactive_ratio = int_sqrt(10 * gb);
1513 return inactive * inactive_ratio < active;
1516 #define mem_cgroup_from_res_counter(counter, member) \
1517 container_of(counter, struct mem_cgroup, member)
1520 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1521 * @memcg: the memory cgroup
1523 * Returns the maximum amount of memory @mem can be charged with, in
1526 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1528 unsigned long long margin;
1530 margin = res_counter_margin(&memcg->res);
1531 if (do_swap_account)
1532 margin = min(margin, res_counter_margin(&memcg->memsw));
1533 return margin >> PAGE_SHIFT;
1536 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1539 if (!css_parent(&memcg->css))
1540 return vm_swappiness;
1542 return memcg->swappiness;
1546 * memcg->moving_account is used for checking possibility that some thread is
1547 * calling move_account(). When a thread on CPU-A starts moving pages under
1548 * a memcg, other threads should check memcg->moving_account under
1549 * rcu_read_lock(), like this:
1553 * memcg->moving_account+1 if (memcg->mocing_account)
1555 * synchronize_rcu() update something.
1560 /* for quick checking without looking up memcg */
1561 atomic_t memcg_moving __read_mostly;
1563 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1565 atomic_inc(&memcg_moving);
1566 atomic_inc(&memcg->moving_account);
1570 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1573 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1574 * We check NULL in callee rather than caller.
1577 atomic_dec(&memcg_moving);
1578 atomic_dec(&memcg->moving_account);
1583 * 2 routines for checking "mem" is under move_account() or not.
1585 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1586 * is used for avoiding races in accounting. If true,
1587 * pc->mem_cgroup may be overwritten.
1589 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1590 * under hierarchy of moving cgroups. This is for
1591 * waiting at hith-memory prressure caused by "move".
1594 static bool mem_cgroup_stolen(struct mem_cgroup *memcg)
1596 VM_BUG_ON(!rcu_read_lock_held());
1597 return atomic_read(&memcg->moving_account) > 0;
1600 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1602 struct mem_cgroup *from;
1603 struct mem_cgroup *to;
1606 * Unlike task_move routines, we access mc.to, mc.from not under
1607 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1609 spin_lock(&mc.lock);
1615 ret = mem_cgroup_same_or_subtree(memcg, from)
1616 || mem_cgroup_same_or_subtree(memcg, to);
1618 spin_unlock(&mc.lock);
1622 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1624 if (mc.moving_task && current != mc.moving_task) {
1625 if (mem_cgroup_under_move(memcg)) {
1627 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1628 /* moving charge context might have finished. */
1631 finish_wait(&mc.waitq, &wait);
1639 * Take this lock when
1640 * - a code tries to modify page's memcg while it's USED.
1641 * - a code tries to modify page state accounting in a memcg.
1642 * see mem_cgroup_stolen(), too.
1644 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1645 unsigned long *flags)
1647 spin_lock_irqsave(&memcg->move_lock, *flags);
1650 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1651 unsigned long *flags)
1653 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1656 #define K(x) ((x) << (PAGE_SHIFT-10))
1658 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1659 * @memcg: The memory cgroup that went over limit
1660 * @p: Task that is going to be killed
1662 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1665 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1668 * protects memcg_name and makes sure that parallel ooms do not
1671 static DEFINE_SPINLOCK(oom_info_lock);
1672 struct cgroup *task_cgrp;
1673 struct cgroup *mem_cgrp;
1674 static char memcg_name[PATH_MAX];
1676 struct mem_cgroup *iter;
1682 spin_lock(&oom_info_lock);
1685 mem_cgrp = memcg->css.cgroup;
1686 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1688 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1691 * Unfortunately, we are unable to convert to a useful name
1692 * But we'll still print out the usage information
1699 pr_info("Task in %s killed", memcg_name);
1702 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1710 * Continues from above, so we don't need an KERN_ level
1712 pr_cont(" as a result of limit of %s\n", memcg_name);
1715 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1716 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1717 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1718 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1719 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1720 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1721 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1722 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1723 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1724 res_counter_read_u64(&memcg->kmem, RES_USAGE) >> 10,
1725 res_counter_read_u64(&memcg->kmem, RES_LIMIT) >> 10,
1726 res_counter_read_u64(&memcg->kmem, RES_FAILCNT));
1728 for_each_mem_cgroup_tree(iter, memcg) {
1729 pr_info("Memory cgroup stats");
1732 ret = cgroup_path(iter->css.cgroup, memcg_name, PATH_MAX);
1734 pr_cont(" for %s", memcg_name);
1738 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1739 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1741 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1742 K(mem_cgroup_read_stat(iter, i)));
1745 for (i = 0; i < NR_LRU_LISTS; i++)
1746 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1747 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1751 spin_unlock(&oom_info_lock);
1755 * This function returns the number of memcg under hierarchy tree. Returns
1756 * 1(self count) if no children.
1758 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1761 struct mem_cgroup *iter;
1763 for_each_mem_cgroup_tree(iter, memcg)
1769 * Return the memory (and swap, if configured) limit for a memcg.
1771 static u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1775 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1778 * Do not consider swap space if we cannot swap due to swappiness
1780 if (mem_cgroup_swappiness(memcg)) {
1783 limit += total_swap_pages << PAGE_SHIFT;
1784 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1787 * If memsw is finite and limits the amount of swap space
1788 * available to this memcg, return that limit.
1790 limit = min(limit, memsw);
1796 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1799 struct mem_cgroup *iter;
1800 unsigned long chosen_points = 0;
1801 unsigned long totalpages;
1802 unsigned int points = 0;
1803 struct task_struct *chosen = NULL;
1806 * If current has a pending SIGKILL or is exiting, then automatically
1807 * select it. The goal is to allow it to allocate so that it may
1808 * quickly exit and free its memory.
1810 if (fatal_signal_pending(current) || current->flags & PF_EXITING) {
1811 set_thread_flag(TIF_MEMDIE);
1815 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1816 totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
1817 for_each_mem_cgroup_tree(iter, memcg) {
1818 struct css_task_iter it;
1819 struct task_struct *task;
1821 css_task_iter_start(&iter->css, &it);
1822 while ((task = css_task_iter_next(&it))) {
1823 switch (oom_scan_process_thread(task, totalpages, NULL,
1825 case OOM_SCAN_SELECT:
1827 put_task_struct(chosen);
1829 chosen_points = ULONG_MAX;
1830 get_task_struct(chosen);
1832 case OOM_SCAN_CONTINUE:
1834 case OOM_SCAN_ABORT:
1835 css_task_iter_end(&it);
1836 mem_cgroup_iter_break(memcg, iter);
1838 put_task_struct(chosen);
1843 points = oom_badness(task, memcg, NULL, totalpages);
1844 if (points > chosen_points) {
1846 put_task_struct(chosen);
1848 chosen_points = points;
1849 get_task_struct(chosen);
1852 css_task_iter_end(&it);
1857 points = chosen_points * 1000 / totalpages;
1858 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1859 NULL, "Memory cgroup out of memory");
1862 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1864 unsigned long flags)
1866 unsigned long total = 0;
1867 bool noswap = false;
1870 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1872 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1875 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1877 drain_all_stock_async(memcg);
1878 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1880 * Allow limit shrinkers, which are triggered directly
1881 * by userspace, to catch signals and stop reclaim
1882 * after minimal progress, regardless of the margin.
1884 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1886 if (mem_cgroup_margin(memcg))
1889 * If nothing was reclaimed after two attempts, there
1890 * may be no reclaimable pages in this hierarchy.
1899 * test_mem_cgroup_node_reclaimable
1900 * @memcg: the target memcg
1901 * @nid: the node ID to be checked.
1902 * @noswap : specify true here if the user wants flle only information.
1904 * This function returns whether the specified memcg contains any
1905 * reclaimable pages on a node. Returns true if there are any reclaimable
1906 * pages in the node.
1908 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1909 int nid, bool noswap)
1911 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1913 if (noswap || !total_swap_pages)
1915 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1920 #if MAX_NUMNODES > 1
1923 * Always updating the nodemask is not very good - even if we have an empty
1924 * list or the wrong list here, we can start from some node and traverse all
1925 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1928 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1932 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1933 * pagein/pageout changes since the last update.
1935 if (!atomic_read(&memcg->numainfo_events))
1937 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1940 /* make a nodemask where this memcg uses memory from */
1941 memcg->scan_nodes = node_states[N_MEMORY];
1943 for_each_node_mask(nid, node_states[N_MEMORY]) {
1945 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1946 node_clear(nid, memcg->scan_nodes);
1949 atomic_set(&memcg->numainfo_events, 0);
1950 atomic_set(&memcg->numainfo_updating, 0);
1954 * Selecting a node where we start reclaim from. Because what we need is just
1955 * reducing usage counter, start from anywhere is O,K. Considering
1956 * memory reclaim from current node, there are pros. and cons.
1958 * Freeing memory from current node means freeing memory from a node which
1959 * we'll use or we've used. So, it may make LRU bad. And if several threads
1960 * hit limits, it will see a contention on a node. But freeing from remote
1961 * node means more costs for memory reclaim because of memory latency.
1963 * Now, we use round-robin. Better algorithm is welcomed.
1965 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1969 mem_cgroup_may_update_nodemask(memcg);
1970 node = memcg->last_scanned_node;
1972 node = next_node(node, memcg->scan_nodes);
1973 if (node == MAX_NUMNODES)
1974 node = first_node(memcg->scan_nodes);
1976 * We call this when we hit limit, not when pages are added to LRU.
1977 * No LRU may hold pages because all pages are UNEVICTABLE or
1978 * memcg is too small and all pages are not on LRU. In that case,
1979 * we use curret node.
1981 if (unlikely(node == MAX_NUMNODES))
1982 node = numa_node_id();
1984 memcg->last_scanned_node = node;
1989 * Check all nodes whether it contains reclaimable pages or not.
1990 * For quick scan, we make use of scan_nodes. This will allow us to skip
1991 * unused nodes. But scan_nodes is lazily updated and may not cotain
1992 * enough new information. We need to do double check.
1994 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1999 * quick check...making use of scan_node.
2000 * We can skip unused nodes.
2002 if (!nodes_empty(memcg->scan_nodes)) {
2003 for (nid = first_node(memcg->scan_nodes);
2005 nid = next_node(nid, memcg->scan_nodes)) {
2007 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
2012 * Check rest of nodes.
2014 for_each_node_state(nid, N_MEMORY) {
2015 if (node_isset(nid, memcg->scan_nodes))
2017 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
2024 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
2029 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
2031 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
2035 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
2038 unsigned long *total_scanned)
2040 struct mem_cgroup *victim = NULL;
2043 unsigned long excess;
2044 unsigned long nr_scanned;
2045 struct mem_cgroup_reclaim_cookie reclaim = {
2050 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
2053 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
2058 * If we have not been able to reclaim
2059 * anything, it might because there are
2060 * no reclaimable pages under this hierarchy
2065 * We want to do more targeted reclaim.
2066 * excess >> 2 is not to excessive so as to
2067 * reclaim too much, nor too less that we keep
2068 * coming back to reclaim from this cgroup
2070 if (total >= (excess >> 2) ||
2071 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
2076 if (!mem_cgroup_reclaimable(victim, false))
2078 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
2080 *total_scanned += nr_scanned;
2081 if (!res_counter_soft_limit_excess(&root_memcg->res))
2084 mem_cgroup_iter_break(root_memcg, victim);
2088 #ifdef CONFIG_LOCKDEP
2089 static struct lockdep_map memcg_oom_lock_dep_map = {
2090 .name = "memcg_oom_lock",
2094 static DEFINE_SPINLOCK(memcg_oom_lock);
2097 * Check OOM-Killer is already running under our hierarchy.
2098 * If someone is running, return false.
2100 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
2102 struct mem_cgroup *iter, *failed = NULL;
2104 spin_lock(&memcg_oom_lock);
2106 for_each_mem_cgroup_tree(iter, memcg) {
2107 if (iter->oom_lock) {
2109 * this subtree of our hierarchy is already locked
2110 * so we cannot give a lock.
2113 mem_cgroup_iter_break(memcg, iter);
2116 iter->oom_lock = true;
2121 * OK, we failed to lock the whole subtree so we have
2122 * to clean up what we set up to the failing subtree
2124 for_each_mem_cgroup_tree(iter, memcg) {
2125 if (iter == failed) {
2126 mem_cgroup_iter_break(memcg, iter);
2129 iter->oom_lock = false;
2132 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
2134 spin_unlock(&memcg_oom_lock);
2139 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
2141 struct mem_cgroup *iter;
2143 spin_lock(&memcg_oom_lock);
2144 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
2145 for_each_mem_cgroup_tree(iter, memcg)
2146 iter->oom_lock = false;
2147 spin_unlock(&memcg_oom_lock);
2150 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
2152 struct mem_cgroup *iter;
2154 for_each_mem_cgroup_tree(iter, memcg)
2155 atomic_inc(&iter->under_oom);
2158 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
2160 struct mem_cgroup *iter;
2163 * When a new child is created while the hierarchy is under oom,
2164 * mem_cgroup_oom_lock() may not be called. We have to use
2165 * atomic_add_unless() here.
2167 for_each_mem_cgroup_tree(iter, memcg)
2168 atomic_add_unless(&iter->under_oom, -1, 0);
2171 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
2173 struct oom_wait_info {
2174 struct mem_cgroup *memcg;
2178 static int memcg_oom_wake_function(wait_queue_t *wait,
2179 unsigned mode, int sync, void *arg)
2181 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
2182 struct mem_cgroup *oom_wait_memcg;
2183 struct oom_wait_info *oom_wait_info;
2185 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
2186 oom_wait_memcg = oom_wait_info->memcg;
2189 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2190 * Then we can use css_is_ancestor without taking care of RCU.
2192 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
2193 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
2195 return autoremove_wake_function(wait, mode, sync, arg);
2198 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
2200 atomic_inc(&memcg->oom_wakeups);
2201 /* for filtering, pass "memcg" as argument. */
2202 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
2205 static void memcg_oom_recover(struct mem_cgroup *memcg)
2207 if (memcg && atomic_read(&memcg->under_oom))
2208 memcg_wakeup_oom(memcg);
2211 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
2213 if (!current->memcg_oom.may_oom)
2216 * We are in the middle of the charge context here, so we
2217 * don't want to block when potentially sitting on a callstack
2218 * that holds all kinds of filesystem and mm locks.
2220 * Also, the caller may handle a failed allocation gracefully
2221 * (like optional page cache readahead) and so an OOM killer
2222 * invocation might not even be necessary.
2224 * That's why we don't do anything here except remember the
2225 * OOM context and then deal with it at the end of the page
2226 * fault when the stack is unwound, the locks are released,
2227 * and when we know whether the fault was overall successful.
2229 css_get(&memcg->css);
2230 current->memcg_oom.memcg = memcg;
2231 current->memcg_oom.gfp_mask = mask;
2232 current->memcg_oom.order = order;
2236 * mem_cgroup_oom_synchronize - complete memcg OOM handling
2237 * @handle: actually kill/wait or just clean up the OOM state
2239 * This has to be called at the end of a page fault if the memcg OOM
2240 * handler was enabled.
2242 * Memcg supports userspace OOM handling where failed allocations must
2243 * sleep on a waitqueue until the userspace task resolves the
2244 * situation. Sleeping directly in the charge context with all kinds
2245 * of locks held is not a good idea, instead we remember an OOM state
2246 * in the task and mem_cgroup_oom_synchronize() has to be called at
2247 * the end of the page fault to complete the OOM handling.
2249 * Returns %true if an ongoing memcg OOM situation was detected and
2250 * completed, %false otherwise.
2252 bool mem_cgroup_oom_synchronize(bool handle)
2254 struct mem_cgroup *memcg = current->memcg_oom.memcg;
2255 struct oom_wait_info owait;
2258 /* OOM is global, do not handle */
2265 owait.memcg = memcg;
2266 owait.wait.flags = 0;
2267 owait.wait.func = memcg_oom_wake_function;
2268 owait.wait.private = current;
2269 INIT_LIST_HEAD(&owait.wait.task_list);
2271 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2272 mem_cgroup_mark_under_oom(memcg);
2274 locked = mem_cgroup_oom_trylock(memcg);
2277 mem_cgroup_oom_notify(memcg);
2279 if (locked && !memcg->oom_kill_disable) {
2280 mem_cgroup_unmark_under_oom(memcg);
2281 finish_wait(&memcg_oom_waitq, &owait.wait);
2282 mem_cgroup_out_of_memory(memcg, current->memcg_oom.gfp_mask,
2283 current->memcg_oom.order);
2286 mem_cgroup_unmark_under_oom(memcg);
2287 finish_wait(&memcg_oom_waitq, &owait.wait);
2291 mem_cgroup_oom_unlock(memcg);
2293 * There is no guarantee that an OOM-lock contender
2294 * sees the wakeups triggered by the OOM kill
2295 * uncharges. Wake any sleepers explicitely.
2297 memcg_oom_recover(memcg);
2300 current->memcg_oom.memcg = NULL;
2301 css_put(&memcg->css);
2306 * Currently used to update mapped file statistics, but the routine can be
2307 * generalized to update other statistics as well.
2309 * Notes: Race condition
2311 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2312 * it tends to be costly. But considering some conditions, we doesn't need
2313 * to do so _always_.
2315 * Considering "charge", lock_page_cgroup() is not required because all
2316 * file-stat operations happen after a page is attached to radix-tree. There
2317 * are no race with "charge".
2319 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2320 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2321 * if there are race with "uncharge". Statistics itself is properly handled
2324 * Considering "move", this is an only case we see a race. To make the race
2325 * small, we check mm->moving_account and detect there are possibility of race
2326 * If there is, we take a lock.
2329 void __mem_cgroup_begin_update_page_stat(struct page *page,
2330 bool *locked, unsigned long *flags)
2332 struct mem_cgroup *memcg;
2333 struct page_cgroup *pc;
2335 pc = lookup_page_cgroup(page);
2337 memcg = pc->mem_cgroup;
2338 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2341 * If this memory cgroup is not under account moving, we don't
2342 * need to take move_lock_mem_cgroup(). Because we already hold
2343 * rcu_read_lock(), any calls to move_account will be delayed until
2344 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2346 if (!mem_cgroup_stolen(memcg))
2349 move_lock_mem_cgroup(memcg, flags);
2350 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2351 move_unlock_mem_cgroup(memcg, flags);
2357 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
2359 struct page_cgroup *pc = lookup_page_cgroup(page);
2362 * It's guaranteed that pc->mem_cgroup never changes while
2363 * lock is held because a routine modifies pc->mem_cgroup
2364 * should take move_lock_mem_cgroup().
2366 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
2369 void mem_cgroup_update_page_stat(struct page *page,
2370 enum mem_cgroup_stat_index idx, int val)
2372 struct mem_cgroup *memcg;
2373 struct page_cgroup *pc = lookup_page_cgroup(page);
2374 unsigned long uninitialized_var(flags);
2376 if (mem_cgroup_disabled())
2379 VM_BUG_ON(!rcu_read_lock_held());
2380 memcg = pc->mem_cgroup;
2381 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2384 this_cpu_add(memcg->stat->count[idx], val);
2388 * size of first charge trial. "32" comes from vmscan.c's magic value.
2389 * TODO: maybe necessary to use big numbers in big irons.
2391 #define CHARGE_BATCH 32U
2392 struct memcg_stock_pcp {
2393 struct mem_cgroup *cached; /* this never be root cgroup */
2394 unsigned int nr_pages;
2395 struct work_struct work;
2396 unsigned long flags;
2397 #define FLUSHING_CACHED_CHARGE 0
2399 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2400 static DEFINE_MUTEX(percpu_charge_mutex);
2403 * consume_stock: Try to consume stocked charge on this cpu.
2404 * @memcg: memcg to consume from.
2405 * @nr_pages: how many pages to charge.
2407 * The charges will only happen if @memcg matches the current cpu's memcg
2408 * stock, and at least @nr_pages are available in that stock. Failure to
2409 * service an allocation will refill the stock.
2411 * returns true if successful, false otherwise.
2413 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2415 struct memcg_stock_pcp *stock;
2418 if (nr_pages > CHARGE_BATCH)
2421 stock = &get_cpu_var(memcg_stock);
2422 if (memcg == stock->cached && stock->nr_pages >= nr_pages)
2423 stock->nr_pages -= nr_pages;
2424 else /* need to call res_counter_charge */
2426 put_cpu_var(memcg_stock);
2431 * Returns stocks cached in percpu to res_counter and reset cached information.
2433 static void drain_stock(struct memcg_stock_pcp *stock)
2435 struct mem_cgroup *old = stock->cached;
2437 if (stock->nr_pages) {
2438 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2440 res_counter_uncharge(&old->res, bytes);
2441 if (do_swap_account)
2442 res_counter_uncharge(&old->memsw, bytes);
2443 stock->nr_pages = 0;
2445 stock->cached = NULL;
2449 * This must be called under preempt disabled or must be called by
2450 * a thread which is pinned to local cpu.
2452 static void drain_local_stock(struct work_struct *dummy)
2454 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2456 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2459 static void __init memcg_stock_init(void)
2463 for_each_possible_cpu(cpu) {
2464 struct memcg_stock_pcp *stock =
2465 &per_cpu(memcg_stock, cpu);
2466 INIT_WORK(&stock->work, drain_local_stock);
2471 * Cache charges(val) which is from res_counter, to local per_cpu area.
2472 * This will be consumed by consume_stock() function, later.
2474 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2476 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2478 if (stock->cached != memcg) { /* reset if necessary */
2480 stock->cached = memcg;
2482 stock->nr_pages += nr_pages;
2483 put_cpu_var(memcg_stock);
2487 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2488 * of the hierarchy under it. sync flag says whether we should block
2489 * until the work is done.
2491 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2495 /* Notify other cpus that system-wide "drain" is running */
2498 for_each_online_cpu(cpu) {
2499 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2500 struct mem_cgroup *memcg;
2502 memcg = stock->cached;
2503 if (!memcg || !stock->nr_pages)
2505 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2507 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2509 drain_local_stock(&stock->work);
2511 schedule_work_on(cpu, &stock->work);
2519 for_each_online_cpu(cpu) {
2520 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2521 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2522 flush_work(&stock->work);
2529 * Tries to drain stocked charges in other cpus. This function is asynchronous
2530 * and just put a work per cpu for draining localy on each cpu. Caller can
2531 * expects some charges will be back to res_counter later but cannot wait for
2534 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2537 * If someone calls draining, avoid adding more kworker runs.
2539 if (!mutex_trylock(&percpu_charge_mutex))
2541 drain_all_stock(root_memcg, false);
2542 mutex_unlock(&percpu_charge_mutex);
2545 /* This is a synchronous drain interface. */
2546 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2548 /* called when force_empty is called */
2549 mutex_lock(&percpu_charge_mutex);
2550 drain_all_stock(root_memcg, true);
2551 mutex_unlock(&percpu_charge_mutex);
2555 * This function drains percpu counter value from DEAD cpu and
2556 * move it to local cpu. Note that this function can be preempted.
2558 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2562 spin_lock(&memcg->pcp_counter_lock);
2563 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2564 long x = per_cpu(memcg->stat->count[i], cpu);
2566 per_cpu(memcg->stat->count[i], cpu) = 0;
2567 memcg->nocpu_base.count[i] += x;
2569 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2570 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2572 per_cpu(memcg->stat->events[i], cpu) = 0;
2573 memcg->nocpu_base.events[i] += x;
2575 spin_unlock(&memcg->pcp_counter_lock);
2578 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
2579 unsigned long action,
2582 int cpu = (unsigned long)hcpu;
2583 struct memcg_stock_pcp *stock;
2584 struct mem_cgroup *iter;
2586 if (action == CPU_ONLINE)
2589 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2592 for_each_mem_cgroup(iter)
2593 mem_cgroup_drain_pcp_counter(iter, cpu);
2595 stock = &per_cpu(memcg_stock, cpu);
2601 /* See __mem_cgroup_try_charge() for details */
2603 CHARGE_OK, /* success */
2604 CHARGE_RETRY, /* need to retry but retry is not bad */
2605 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2606 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2609 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2610 unsigned int nr_pages, unsigned int min_pages,
2613 unsigned long csize = nr_pages * PAGE_SIZE;
2614 struct mem_cgroup *mem_over_limit;
2615 struct res_counter *fail_res;
2616 unsigned long flags = 0;
2619 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2622 if (!do_swap_account)
2624 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2628 res_counter_uncharge(&memcg->res, csize);
2629 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2630 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2632 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2634 * Never reclaim on behalf of optional batching, retry with a
2635 * single page instead.
2637 if (nr_pages > min_pages)
2638 return CHARGE_RETRY;
2640 if (!(gfp_mask & __GFP_WAIT))
2641 return CHARGE_WOULDBLOCK;
2643 if (gfp_mask & __GFP_NORETRY)
2644 return CHARGE_NOMEM;
2646 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2647 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2648 return CHARGE_RETRY;
2650 * Even though the limit is exceeded at this point, reclaim
2651 * may have been able to free some pages. Retry the charge
2652 * before killing the task.
2654 * Only for regular pages, though: huge pages are rather
2655 * unlikely to succeed so close to the limit, and we fall back
2656 * to regular pages anyway in case of failure.
2658 if (nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER) && ret)
2659 return CHARGE_RETRY;
2662 * At task move, charge accounts can be doubly counted. So, it's
2663 * better to wait until the end of task_move if something is going on.
2665 if (mem_cgroup_wait_acct_move(mem_over_limit))
2666 return CHARGE_RETRY;
2669 mem_cgroup_oom(mem_over_limit, gfp_mask, get_order(csize));
2671 return CHARGE_NOMEM;
2675 * __mem_cgroup_try_charge() does
2676 * 1. detect memcg to be charged against from passed *mm and *ptr,
2677 * 2. update res_counter
2678 * 3. call memory reclaim if necessary.
2680 * In some special case, if the task is fatal, fatal_signal_pending() or
2681 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2682 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2683 * as possible without any hazards. 2: all pages should have a valid
2684 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2685 * pointer, that is treated as a charge to root_mem_cgroup.
2687 * So __mem_cgroup_try_charge() will return
2688 * 0 ... on success, filling *ptr with a valid memcg pointer.
2689 * -ENOMEM ... charge failure because of resource limits.
2690 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2692 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2693 * the oom-killer can be invoked.
2695 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2697 unsigned int nr_pages,
2698 struct mem_cgroup **ptr,
2701 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2702 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2703 struct mem_cgroup *memcg = NULL;
2707 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2708 * in system level. So, allow to go ahead dying process in addition to
2711 if (unlikely(test_thread_flag(TIF_MEMDIE)
2712 || fatal_signal_pending(current)))
2715 if (unlikely(task_in_memcg_oom(current)))
2718 if (gfp_mask & __GFP_NOFAIL)
2722 * We always charge the cgroup the mm_struct belongs to.
2723 * The mm_struct's mem_cgroup changes on task migration if the
2724 * thread group leader migrates. It's possible that mm is not
2725 * set, if so charge the root memcg (happens for pagecache usage).
2728 *ptr = root_mem_cgroup;
2730 if (*ptr) { /* css should be a valid one */
2732 if (mem_cgroup_is_root(memcg))
2734 if (consume_stock(memcg, nr_pages))
2736 css_get(&memcg->css);
2738 struct task_struct *p;
2741 p = rcu_dereference(mm->owner);
2743 * Because we don't have task_lock(), "p" can exit.
2744 * In that case, "memcg" can point to root or p can be NULL with
2745 * race with swapoff. Then, we have small risk of mis-accouning.
2746 * But such kind of mis-account by race always happens because
2747 * we don't have cgroup_mutex(). It's overkill and we allo that
2749 * (*) swapoff at el will charge against mm-struct not against
2750 * task-struct. So, mm->owner can be NULL.
2752 memcg = mem_cgroup_from_task(p);
2754 memcg = root_mem_cgroup;
2755 if (mem_cgroup_is_root(memcg)) {
2759 if (consume_stock(memcg, nr_pages)) {
2761 * It seems dagerous to access memcg without css_get().
2762 * But considering how consume_stok works, it's not
2763 * necessary. If consume_stock success, some charges
2764 * from this memcg are cached on this cpu. So, we
2765 * don't need to call css_get()/css_tryget() before
2766 * calling consume_stock().
2771 /* after here, we may be blocked. we need to get refcnt */
2772 if (!css_tryget(&memcg->css)) {
2780 bool invoke_oom = oom && !nr_oom_retries;
2782 /* If killed, bypass charge */
2783 if (fatal_signal_pending(current)) {
2784 css_put(&memcg->css);
2788 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch,
2789 nr_pages, invoke_oom);
2793 case CHARGE_RETRY: /* not in OOM situation but retry */
2795 css_put(&memcg->css);
2798 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2799 css_put(&memcg->css);
2801 case CHARGE_NOMEM: /* OOM routine works */
2802 if (!oom || invoke_oom) {
2803 css_put(&memcg->css);
2809 } while (ret != CHARGE_OK);
2811 if (batch > nr_pages)
2812 refill_stock(memcg, batch - nr_pages);
2813 css_put(&memcg->css);
2818 if (!(gfp_mask & __GFP_NOFAIL)) {
2823 *ptr = root_mem_cgroup;
2828 * Somemtimes we have to undo a charge we got by try_charge().
2829 * This function is for that and do uncharge, put css's refcnt.
2830 * gotten by try_charge().
2832 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2833 unsigned int nr_pages)
2835 if (!mem_cgroup_is_root(memcg)) {
2836 unsigned long bytes = nr_pages * PAGE_SIZE;
2838 res_counter_uncharge(&memcg->res, bytes);
2839 if (do_swap_account)
2840 res_counter_uncharge(&memcg->memsw, bytes);
2845 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2846 * This is useful when moving usage to parent cgroup.
2848 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2849 unsigned int nr_pages)
2851 unsigned long bytes = nr_pages * PAGE_SIZE;
2853 if (mem_cgroup_is_root(memcg))
2856 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2857 if (do_swap_account)
2858 res_counter_uncharge_until(&memcg->memsw,
2859 memcg->memsw.parent, bytes);
2863 * A helper function to get mem_cgroup from ID. must be called under
2864 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2865 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2866 * called against removed memcg.)
2868 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2870 /* ID 0 is unused ID */
2873 return mem_cgroup_from_id(id);
2876 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2878 struct mem_cgroup *memcg = NULL;
2879 struct page_cgroup *pc;
2883 VM_BUG_ON_PAGE(!PageLocked(page), page);
2885 pc = lookup_page_cgroup(page);
2886 lock_page_cgroup(pc);
2887 if (PageCgroupUsed(pc)) {
2888 memcg = pc->mem_cgroup;
2889 if (memcg && !css_tryget(&memcg->css))
2891 } else if (PageSwapCache(page)) {
2892 ent.val = page_private(page);
2893 id = lookup_swap_cgroup_id(ent);
2895 memcg = mem_cgroup_lookup(id);
2896 if (memcg && !css_tryget(&memcg->css))
2900 unlock_page_cgroup(pc);
2904 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2906 unsigned int nr_pages,
2907 enum charge_type ctype,
2910 struct page_cgroup *pc = lookup_page_cgroup(page);
2911 struct zone *uninitialized_var(zone);
2912 struct lruvec *lruvec;
2913 bool was_on_lru = false;
2916 lock_page_cgroup(pc);
2917 VM_BUG_ON_PAGE(PageCgroupUsed(pc), page);
2919 * we don't need page_cgroup_lock about tail pages, becase they are not
2920 * accessed by any other context at this point.
2924 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2925 * may already be on some other mem_cgroup's LRU. Take care of it.
2928 zone = page_zone(page);
2929 spin_lock_irq(&zone->lru_lock);
2930 if (PageLRU(page)) {
2931 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2933 del_page_from_lru_list(page, lruvec, page_lru(page));
2938 pc->mem_cgroup = memcg;
2940 * We access a page_cgroup asynchronously without lock_page_cgroup().
2941 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2942 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2943 * before USED bit, we need memory barrier here.
2944 * See mem_cgroup_add_lru_list(), etc.
2947 SetPageCgroupUsed(pc);
2951 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2952 VM_BUG_ON_PAGE(PageLRU(page), page);
2954 add_page_to_lru_list(page, lruvec, page_lru(page));
2956 spin_unlock_irq(&zone->lru_lock);
2959 if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
2964 mem_cgroup_charge_statistics(memcg, page, anon, nr_pages);
2965 unlock_page_cgroup(pc);
2968 * "charge_statistics" updated event counter. Then, check it.
2969 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2970 * if they exceeds softlimit.
2972 memcg_check_events(memcg, page);
2975 static DEFINE_MUTEX(set_limit_mutex);
2977 #ifdef CONFIG_MEMCG_KMEM
2978 static DEFINE_MUTEX(activate_kmem_mutex);
2980 static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg)
2982 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) &&
2983 memcg_kmem_is_active(memcg);
2987 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2988 * in the memcg_cache_params struct.
2990 static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
2992 struct kmem_cache *cachep;
2994 VM_BUG_ON(p->is_root_cache);
2995 cachep = p->root_cache;
2996 return cache_from_memcg_idx(cachep, memcg_cache_id(p->memcg));
2999 #ifdef CONFIG_SLABINFO
3000 static int mem_cgroup_slabinfo_read(struct seq_file *m, void *v)
3002 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3003 struct memcg_cache_params *params;
3005 if (!memcg_can_account_kmem(memcg))
3008 print_slabinfo_header(m);
3010 mutex_lock(&memcg->slab_caches_mutex);
3011 list_for_each_entry(params, &memcg->memcg_slab_caches, list)
3012 cache_show(memcg_params_to_cache(params), m);
3013 mutex_unlock(&memcg->slab_caches_mutex);
3019 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size)
3021 struct res_counter *fail_res;
3022 struct mem_cgroup *_memcg;
3025 ret = res_counter_charge(&memcg->kmem, size, &fail_res);
3030 ret = __mem_cgroup_try_charge(NULL, gfp, size >> PAGE_SHIFT,
3031 &_memcg, oom_gfp_allowed(gfp));
3033 if (ret == -EINTR) {
3035 * __mem_cgroup_try_charge() chosed to bypass to root due to
3036 * OOM kill or fatal signal. Since our only options are to
3037 * either fail the allocation or charge it to this cgroup, do
3038 * it as a temporary condition. But we can't fail. From a
3039 * kmem/slab perspective, the cache has already been selected,
3040 * by mem_cgroup_kmem_get_cache(), so it is too late to change
3043 * This condition will only trigger if the task entered
3044 * memcg_charge_kmem in a sane state, but was OOM-killed during
3045 * __mem_cgroup_try_charge() above. Tasks that were already
3046 * dying when the allocation triggers should have been already
3047 * directed to the root cgroup in memcontrol.h
3049 res_counter_charge_nofail(&memcg->res, size, &fail_res);
3050 if (do_swap_account)
3051 res_counter_charge_nofail(&memcg->memsw, size,
3055 res_counter_uncharge(&memcg->kmem, size);
3060 static void memcg_uncharge_kmem(struct mem_cgroup *memcg, u64 size)
3062 res_counter_uncharge(&memcg->res, size);
3063 if (do_swap_account)
3064 res_counter_uncharge(&memcg->memsw, size);
3067 if (res_counter_uncharge(&memcg->kmem, size))
3071 * Releases a reference taken in kmem_cgroup_css_offline in case
3072 * this last uncharge is racing with the offlining code or it is
3073 * outliving the memcg existence.
3075 * The memory barrier imposed by test&clear is paired with the
3076 * explicit one in memcg_kmem_mark_dead().
3078 if (memcg_kmem_test_and_clear_dead(memcg))
3079 css_put(&memcg->css);
3083 * helper for acessing a memcg's index. It will be used as an index in the
3084 * child cache array in kmem_cache, and also to derive its name. This function
3085 * will return -1 when this is not a kmem-limited memcg.
3087 int memcg_cache_id(struct mem_cgroup *memcg)
3089 return memcg ? memcg->kmemcg_id : -1;
3092 static size_t memcg_caches_array_size(int num_groups)
3095 if (num_groups <= 0)
3098 size = 2 * num_groups;
3099 if (size < MEMCG_CACHES_MIN_SIZE)
3100 size = MEMCG_CACHES_MIN_SIZE;
3101 else if (size > MEMCG_CACHES_MAX_SIZE)
3102 size = MEMCG_CACHES_MAX_SIZE;
3108 * We should update the current array size iff all caches updates succeed. This
3109 * can only be done from the slab side. The slab mutex needs to be held when
3112 void memcg_update_array_size(int num)
3114 if (num > memcg_limited_groups_array_size)
3115 memcg_limited_groups_array_size = memcg_caches_array_size(num);
3118 static void kmem_cache_destroy_work_func(struct work_struct *w);
3120 int memcg_update_cache_size(struct kmem_cache *s, int num_groups)
3122 struct memcg_cache_params *cur_params = s->memcg_params;
3124 VM_BUG_ON(!is_root_cache(s));
3126 if (num_groups > memcg_limited_groups_array_size) {
3128 struct memcg_cache_params *new_params;
3129 ssize_t size = memcg_caches_array_size(num_groups);
3131 size *= sizeof(void *);
3132 size += offsetof(struct memcg_cache_params, memcg_caches);
3134 new_params = kzalloc(size, GFP_KERNEL);
3138 new_params->is_root_cache = true;
3141 * There is the chance it will be bigger than
3142 * memcg_limited_groups_array_size, if we failed an allocation
3143 * in a cache, in which case all caches updated before it, will
3144 * have a bigger array.
3146 * But if that is the case, the data after
3147 * memcg_limited_groups_array_size is certainly unused
3149 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3150 if (!cur_params->memcg_caches[i])
3152 new_params->memcg_caches[i] =
3153 cur_params->memcg_caches[i];
3157 * Ideally, we would wait until all caches succeed, and only
3158 * then free the old one. But this is not worth the extra
3159 * pointer per-cache we'd have to have for this.
3161 * It is not a big deal if some caches are left with a size
3162 * bigger than the others. And all updates will reset this
3165 rcu_assign_pointer(s->memcg_params, new_params);
3167 kfree_rcu(cur_params, rcu_head);
3172 int memcg_alloc_cache_params(struct mem_cgroup *memcg, struct kmem_cache *s,
3173 struct kmem_cache *root_cache)
3177 if (!memcg_kmem_enabled())
3181 size = offsetof(struct memcg_cache_params, memcg_caches);
3182 size += memcg_limited_groups_array_size * sizeof(void *);
3184 size = sizeof(struct memcg_cache_params);
3186 s->memcg_params = kzalloc(size, GFP_KERNEL);
3187 if (!s->memcg_params)
3191 s->memcg_params->memcg = memcg;
3192 s->memcg_params->root_cache = root_cache;
3193 INIT_WORK(&s->memcg_params->destroy,
3194 kmem_cache_destroy_work_func);
3196 s->memcg_params->is_root_cache = true;
3201 void memcg_free_cache_params(struct kmem_cache *s)
3203 kfree(s->memcg_params);
3206 void memcg_register_cache(struct kmem_cache *s)
3208 struct kmem_cache *root;
3209 struct mem_cgroup *memcg;
3212 if (is_root_cache(s))
3216 * Holding the slab_mutex assures nobody will touch the memcg_caches
3217 * array while we are modifying it.
3219 lockdep_assert_held(&slab_mutex);
3221 root = s->memcg_params->root_cache;
3222 memcg = s->memcg_params->memcg;
3223 id = memcg_cache_id(memcg);
3225 css_get(&memcg->css);
3229 * Since readers won't lock (see cache_from_memcg_idx()), we need a
3230 * barrier here to ensure nobody will see the kmem_cache partially
3236 * Initialize the pointer to this cache in its parent's memcg_params
3237 * before adding it to the memcg_slab_caches list, otherwise we can
3238 * fail to convert memcg_params_to_cache() while traversing the list.
3240 VM_BUG_ON(root->memcg_params->memcg_caches[id]);
3241 root->memcg_params->memcg_caches[id] = s;
3243 mutex_lock(&memcg->slab_caches_mutex);
3244 list_add(&s->memcg_params->list, &memcg->memcg_slab_caches);
3245 mutex_unlock(&memcg->slab_caches_mutex);
3248 void memcg_unregister_cache(struct kmem_cache *s)
3250 struct kmem_cache *root;
3251 struct mem_cgroup *memcg;
3254 if (is_root_cache(s))
3258 * Holding the slab_mutex assures nobody will touch the memcg_caches
3259 * array while we are modifying it.
3261 lockdep_assert_held(&slab_mutex);
3263 root = s->memcg_params->root_cache;
3264 memcg = s->memcg_params->memcg;
3265 id = memcg_cache_id(memcg);
3267 mutex_lock(&memcg->slab_caches_mutex);
3268 list_del(&s->memcg_params->list);
3269 mutex_unlock(&memcg->slab_caches_mutex);
3272 * Clear the pointer to this cache in its parent's memcg_params only
3273 * after removing it from the memcg_slab_caches list, otherwise we can
3274 * fail to convert memcg_params_to_cache() while traversing the list.
3276 VM_BUG_ON(!root->memcg_params->memcg_caches[id]);
3277 root->memcg_params->memcg_caches[id] = NULL;
3279 css_put(&memcg->css);
3283 * During the creation a new cache, we need to disable our accounting mechanism
3284 * altogether. This is true even if we are not creating, but rather just
3285 * enqueing new caches to be created.
3287 * This is because that process will trigger allocations; some visible, like
3288 * explicit kmallocs to auxiliary data structures, name strings and internal
3289 * cache structures; some well concealed, like INIT_WORK() that can allocate
3290 * objects during debug.
3292 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3293 * to it. This may not be a bounded recursion: since the first cache creation
3294 * failed to complete (waiting on the allocation), we'll just try to create the
3295 * cache again, failing at the same point.
3297 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3298 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3299 * inside the following two functions.
3301 static inline void memcg_stop_kmem_account(void)
3303 VM_BUG_ON(!current->mm);
3304 current->memcg_kmem_skip_account++;
3307 static inline void memcg_resume_kmem_account(void)
3309 VM_BUG_ON(!current->mm);
3310 current->memcg_kmem_skip_account--;
3313 static void kmem_cache_destroy_work_func(struct work_struct *w)
3315 struct kmem_cache *cachep;
3316 struct memcg_cache_params *p;
3318 p = container_of(w, struct memcg_cache_params, destroy);
3320 cachep = memcg_params_to_cache(p);
3323 * If we get down to 0 after shrink, we could delete right away.
3324 * However, memcg_release_pages() already puts us back in the workqueue
3325 * in that case. If we proceed deleting, we'll get a dangling
3326 * reference, and removing the object from the workqueue in that case
3327 * is unnecessary complication. We are not a fast path.
3329 * Note that this case is fundamentally different from racing with
3330 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3331 * kmem_cache_shrink, not only we would be reinserting a dead cache
3332 * into the queue, but doing so from inside the worker racing to
3335 * So if we aren't down to zero, we'll just schedule a worker and try
3338 if (atomic_read(&cachep->memcg_params->nr_pages) != 0) {
3339 kmem_cache_shrink(cachep);
3340 if (atomic_read(&cachep->memcg_params->nr_pages) == 0)
3343 kmem_cache_destroy(cachep);
3346 void mem_cgroup_destroy_cache(struct kmem_cache *cachep)
3348 if (!cachep->memcg_params->dead)
3352 * There are many ways in which we can get here.
3354 * We can get to a memory-pressure situation while the delayed work is
3355 * still pending to run. The vmscan shrinkers can then release all
3356 * cache memory and get us to destruction. If this is the case, we'll
3357 * be executed twice, which is a bug (the second time will execute over
3358 * bogus data). In this case, cancelling the work should be fine.
3360 * But we can also get here from the worker itself, if
3361 * kmem_cache_shrink is enough to shake all the remaining objects and
3362 * get the page count to 0. In this case, we'll deadlock if we try to
3363 * cancel the work (the worker runs with an internal lock held, which
3364 * is the same lock we would hold for cancel_work_sync().)
3366 * Since we can't possibly know who got us here, just refrain from
3367 * running if there is already work pending
3369 if (work_pending(&cachep->memcg_params->destroy))
3372 * We have to defer the actual destroying to a workqueue, because
3373 * we might currently be in a context that cannot sleep.
3375 schedule_work(&cachep->memcg_params->destroy);
3378 static struct kmem_cache *memcg_create_kmem_cache(struct mem_cgroup *memcg,
3379 struct kmem_cache *s)
3381 struct kmem_cache *new;
3382 static char *tmp_name = NULL;
3383 static DEFINE_MUTEX(mutex); /* protects tmp_name */
3385 BUG_ON(!memcg_can_account_kmem(memcg));
3389 * kmem_cache_create_memcg duplicates the given name and
3390 * cgroup_name for this name requires RCU context.
3391 * This static temporary buffer is used to prevent from
3392 * pointless shortliving allocation.
3395 tmp_name = kmalloc(PATH_MAX, GFP_KERNEL);
3401 snprintf(tmp_name, PATH_MAX, "%s(%d:%s)", s->name,
3402 memcg_cache_id(memcg), cgroup_name(memcg->css.cgroup));
3405 new = kmem_cache_create_memcg(memcg, tmp_name, s->object_size, s->align,
3406 (s->flags & ~SLAB_PANIC), s->ctor, s);
3409 new->allocflags |= __GFP_KMEMCG;
3413 mutex_unlock(&mutex);
3417 void kmem_cache_destroy_memcg_children(struct kmem_cache *s)
3419 struct kmem_cache *c;
3422 if (!s->memcg_params)
3424 if (!s->memcg_params->is_root_cache)
3428 * If the cache is being destroyed, we trust that there is no one else
3429 * requesting objects from it. Even if there are, the sanity checks in
3430 * kmem_cache_destroy should caught this ill-case.
3432 * Still, we don't want anyone else freeing memcg_caches under our
3433 * noses, which can happen if a new memcg comes to life. As usual,
3434 * we'll take the activate_kmem_mutex to protect ourselves against
3437 mutex_lock(&activate_kmem_mutex);
3438 for_each_memcg_cache_index(i) {
3439 c = cache_from_memcg_idx(s, i);
3444 * We will now manually delete the caches, so to avoid races
3445 * we need to cancel all pending destruction workers and
3446 * proceed with destruction ourselves.
3448 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3449 * and that could spawn the workers again: it is likely that
3450 * the cache still have active pages until this very moment.
3451 * This would lead us back to mem_cgroup_destroy_cache.
3453 * But that will not execute at all if the "dead" flag is not
3454 * set, so flip it down to guarantee we are in control.
3456 c->memcg_params->dead = false;
3457 cancel_work_sync(&c->memcg_params->destroy);
3458 kmem_cache_destroy(c);
3460 mutex_unlock(&activate_kmem_mutex);
3463 struct create_work {
3464 struct mem_cgroup *memcg;
3465 struct kmem_cache *cachep;
3466 struct work_struct work;
3469 static void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3471 struct kmem_cache *cachep;
3472 struct memcg_cache_params *params;
3474 if (!memcg_kmem_is_active(memcg))
3477 mutex_lock(&memcg->slab_caches_mutex);
3478 list_for_each_entry(params, &memcg->memcg_slab_caches, list) {
3479 cachep = memcg_params_to_cache(params);
3480 cachep->memcg_params->dead = true;
3481 schedule_work(&cachep->memcg_params->destroy);
3483 mutex_unlock(&memcg->slab_caches_mutex);
3486 static void memcg_create_cache_work_func(struct work_struct *w)
3488 struct create_work *cw;
3490 cw = container_of(w, struct create_work, work);
3491 memcg_create_kmem_cache(cw->memcg, cw->cachep);
3492 css_put(&cw->memcg->css);
3497 * Enqueue the creation of a per-memcg kmem_cache.
3499 static void __memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3500 struct kmem_cache *cachep)
3502 struct create_work *cw;
3504 cw = kmalloc(sizeof(struct create_work), GFP_NOWAIT);
3506 css_put(&memcg->css);
3511 cw->cachep = cachep;
3513 INIT_WORK(&cw->work, memcg_create_cache_work_func);
3514 schedule_work(&cw->work);
3517 static void memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3518 struct kmem_cache *cachep)
3521 * We need to stop accounting when we kmalloc, because if the
3522 * corresponding kmalloc cache is not yet created, the first allocation
3523 * in __memcg_create_cache_enqueue will recurse.
3525 * However, it is better to enclose the whole function. Depending on
3526 * the debugging options enabled, INIT_WORK(), for instance, can
3527 * trigger an allocation. This too, will make us recurse. Because at
3528 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3529 * the safest choice is to do it like this, wrapping the whole function.
3531 memcg_stop_kmem_account();
3532 __memcg_create_cache_enqueue(memcg, cachep);
3533 memcg_resume_kmem_account();
3536 * Return the kmem_cache we're supposed to use for a slab allocation.
3537 * We try to use the current memcg's version of the cache.
3539 * If the cache does not exist yet, if we are the first user of it,
3540 * we either create it immediately, if possible, or create it asynchronously
3542 * In the latter case, we will let the current allocation go through with
3543 * the original cache.
3545 * Can't be called in interrupt context or from kernel threads.
3546 * This function needs to be called with rcu_read_lock() held.
3548 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
3551 struct mem_cgroup *memcg;
3552 struct kmem_cache *memcg_cachep;
3554 VM_BUG_ON(!cachep->memcg_params);
3555 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
3557 if (!current->mm || current->memcg_kmem_skip_account)
3561 memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
3563 if (!memcg_can_account_kmem(memcg))
3566 memcg_cachep = cache_from_memcg_idx(cachep, memcg_cache_id(memcg));
3567 if (likely(memcg_cachep)) {
3568 cachep = memcg_cachep;
3572 /* The corresponding put will be done in the workqueue. */
3573 if (!css_tryget(&memcg->css))
3578 * If we are in a safe context (can wait, and not in interrupt
3579 * context), we could be be predictable and return right away.
3580 * This would guarantee that the allocation being performed
3581 * already belongs in the new cache.
3583 * However, there are some clashes that can arrive from locking.
3584 * For instance, because we acquire the slab_mutex while doing
3585 * kmem_cache_dup, this means no further allocation could happen
3586 * with the slab_mutex held.
3588 * Also, because cache creation issue get_online_cpus(), this
3589 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3590 * that ends up reversed during cpu hotplug. (cpuset allocates
3591 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3592 * better to defer everything.
3594 memcg_create_cache_enqueue(memcg, cachep);
3600 EXPORT_SYMBOL(__memcg_kmem_get_cache);
3603 * We need to verify if the allocation against current->mm->owner's memcg is
3604 * possible for the given order. But the page is not allocated yet, so we'll
3605 * need a further commit step to do the final arrangements.
3607 * It is possible for the task to switch cgroups in this mean time, so at
3608 * commit time, we can't rely on task conversion any longer. We'll then use
3609 * the handle argument to return to the caller which cgroup we should commit
3610 * against. We could also return the memcg directly and avoid the pointer
3611 * passing, but a boolean return value gives better semantics considering
3612 * the compiled-out case as well.
3614 * Returning true means the allocation is possible.
3617 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3619 struct mem_cgroup *memcg;
3625 * Disabling accounting is only relevant for some specific memcg
3626 * internal allocations. Therefore we would initially not have such
3627 * check here, since direct calls to the page allocator that are marked
3628 * with GFP_KMEMCG only happen outside memcg core. We are mostly
3629 * concerned with cache allocations, and by having this test at
3630 * memcg_kmem_get_cache, we are already able to relay the allocation to
3631 * the root cache and bypass the memcg cache altogether.
3633 * There is one exception, though: the SLUB allocator does not create
3634 * large order caches, but rather service large kmallocs directly from
3635 * the page allocator. Therefore, the following sequence when backed by
3636 * the SLUB allocator:
3638 * memcg_stop_kmem_account();
3639 * kmalloc(<large_number>)
3640 * memcg_resume_kmem_account();
3642 * would effectively ignore the fact that we should skip accounting,
3643 * since it will drive us directly to this function without passing
3644 * through the cache selector memcg_kmem_get_cache. Such large
3645 * allocations are extremely rare but can happen, for instance, for the
3646 * cache arrays. We bring this test here.
3648 if (!current->mm || current->memcg_kmem_skip_account)
3651 memcg = try_get_mem_cgroup_from_mm(current->mm);
3654 * very rare case described in mem_cgroup_from_task. Unfortunately there
3655 * isn't much we can do without complicating this too much, and it would
3656 * be gfp-dependent anyway. Just let it go
3658 if (unlikely(!memcg))
3661 if (!memcg_can_account_kmem(memcg)) {
3662 css_put(&memcg->css);
3666 ret = memcg_charge_kmem(memcg, gfp, PAGE_SIZE << order);
3670 css_put(&memcg->css);
3674 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3677 struct page_cgroup *pc;
3679 VM_BUG_ON(mem_cgroup_is_root(memcg));
3681 /* The page allocation failed. Revert */
3683 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3687 pc = lookup_page_cgroup(page);
3688 lock_page_cgroup(pc);
3689 pc->mem_cgroup = memcg;
3690 SetPageCgroupUsed(pc);
3691 unlock_page_cgroup(pc);
3694 void __memcg_kmem_uncharge_pages(struct page *page, int order)
3696 struct mem_cgroup *memcg = NULL;
3697 struct page_cgroup *pc;
3700 pc = lookup_page_cgroup(page);
3702 * Fast unlocked return. Theoretically might have changed, have to
3703 * check again after locking.
3705 if (!PageCgroupUsed(pc))
3708 lock_page_cgroup(pc);
3709 if (PageCgroupUsed(pc)) {
3710 memcg = pc->mem_cgroup;
3711 ClearPageCgroupUsed(pc);
3713 unlock_page_cgroup(pc);
3716 * We trust that only if there is a memcg associated with the page, it
3717 * is a valid allocation
3722 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
3723 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3726 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3729 #endif /* CONFIG_MEMCG_KMEM */
3731 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3733 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3735 * Because tail pages are not marked as "used", set it. We're under
3736 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3737 * charge/uncharge will be never happen and move_account() is done under
3738 * compound_lock(), so we don't have to take care of races.
3740 void mem_cgroup_split_huge_fixup(struct page *head)
3742 struct page_cgroup *head_pc = lookup_page_cgroup(head);
3743 struct page_cgroup *pc;
3744 struct mem_cgroup *memcg;
3747 if (mem_cgroup_disabled())
3750 memcg = head_pc->mem_cgroup;
3751 for (i = 1; i < HPAGE_PMD_NR; i++) {
3753 pc->mem_cgroup = memcg;
3754 smp_wmb();/* see __commit_charge() */
3755 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
3757 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
3760 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3763 void mem_cgroup_move_account_page_stat(struct mem_cgroup *from,
3764 struct mem_cgroup *to,
3765 unsigned int nr_pages,
3766 enum mem_cgroup_stat_index idx)
3768 /* Update stat data for mem_cgroup */
3770 __this_cpu_sub(from->stat->count[idx], nr_pages);
3771 __this_cpu_add(to->stat->count[idx], nr_pages);
3776 * mem_cgroup_move_account - move account of the page
3778 * @nr_pages: number of regular pages (>1 for huge pages)
3779 * @pc: page_cgroup of the page.
3780 * @from: mem_cgroup which the page is moved from.
3781 * @to: mem_cgroup which the page is moved to. @from != @to.
3783 * The caller must confirm following.
3784 * - page is not on LRU (isolate_page() is useful.)
3785 * - compound_lock is held when nr_pages > 1
3787 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3790 static int mem_cgroup_move_account(struct page *page,
3791 unsigned int nr_pages,
3792 struct page_cgroup *pc,
3793 struct mem_cgroup *from,
3794 struct mem_cgroup *to)
3796 unsigned long flags;
3798 bool anon = PageAnon(page);
3800 VM_BUG_ON(from == to);
3801 VM_BUG_ON_PAGE(PageLRU(page), page);
3803 * The page is isolated from LRU. So, collapse function
3804 * will not handle this page. But page splitting can happen.
3805 * Do this check under compound_page_lock(). The caller should
3809 if (nr_pages > 1 && !PageTransHuge(page))
3812 lock_page_cgroup(pc);
3815 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3818 move_lock_mem_cgroup(from, &flags);
3820 if (!anon && page_mapped(page))
3821 mem_cgroup_move_account_page_stat(from, to, nr_pages,
3822 MEM_CGROUP_STAT_FILE_MAPPED);
3824 if (PageWriteback(page))
3825 mem_cgroup_move_account_page_stat(from, to, nr_pages,
3826 MEM_CGROUP_STAT_WRITEBACK);
3828 mem_cgroup_charge_statistics(from, page, anon, -nr_pages);
3830 /* caller should have done css_get */
3831 pc->mem_cgroup = to;
3832 mem_cgroup_charge_statistics(to, page, anon, nr_pages);
3833 move_unlock_mem_cgroup(from, &flags);
3836 unlock_page_cgroup(pc);
3840 memcg_check_events(to, page);
3841 memcg_check_events(from, page);
3847 * mem_cgroup_move_parent - moves page to the parent group
3848 * @page: the page to move
3849 * @pc: page_cgroup of the page
3850 * @child: page's cgroup
3852 * move charges to its parent or the root cgroup if the group has no
3853 * parent (aka use_hierarchy==0).
3854 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3855 * mem_cgroup_move_account fails) the failure is always temporary and
3856 * it signals a race with a page removal/uncharge or migration. In the
3857 * first case the page is on the way out and it will vanish from the LRU
3858 * on the next attempt and the call should be retried later.
3859 * Isolation from the LRU fails only if page has been isolated from
3860 * the LRU since we looked at it and that usually means either global
3861 * reclaim or migration going on. The page will either get back to the
3863 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3864 * (!PageCgroupUsed) or moved to a different group. The page will
3865 * disappear in the next attempt.
3867 static int mem_cgroup_move_parent(struct page *page,
3868 struct page_cgroup *pc,
3869 struct mem_cgroup *child)
3871 struct mem_cgroup *parent;
3872 unsigned int nr_pages;
3873 unsigned long uninitialized_var(flags);
3876 VM_BUG_ON(mem_cgroup_is_root(child));
3879 if (!get_page_unless_zero(page))
3881 if (isolate_lru_page(page))
3884 nr_pages = hpage_nr_pages(page);
3886 parent = parent_mem_cgroup(child);
3888 * If no parent, move charges to root cgroup.
3891 parent = root_mem_cgroup;
3894 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
3895 flags = compound_lock_irqsave(page);
3898 ret = mem_cgroup_move_account(page, nr_pages,
3901 __mem_cgroup_cancel_local_charge(child, nr_pages);
3904 compound_unlock_irqrestore(page, flags);
3905 putback_lru_page(page);
3913 * Charge the memory controller for page usage.
3915 * 0 if the charge was successful
3916 * < 0 if the cgroup is over its limit
3918 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
3919 gfp_t gfp_mask, enum charge_type ctype)
3921 struct mem_cgroup *memcg = NULL;
3922 unsigned int nr_pages = 1;
3926 if (PageTransHuge(page)) {
3927 nr_pages <<= compound_order(page);
3928 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
3930 * Never OOM-kill a process for a huge page. The
3931 * fault handler will fall back to regular pages.
3936 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
3939 __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false);
3943 int mem_cgroup_newpage_charge(struct page *page,
3944 struct mm_struct *mm, gfp_t gfp_mask)
3946 if (mem_cgroup_disabled())
3948 VM_BUG_ON_PAGE(page_mapped(page), page);
3949 VM_BUG_ON_PAGE(page->mapping && !PageAnon(page), page);
3951 return mem_cgroup_charge_common(page, mm, gfp_mask,
3952 MEM_CGROUP_CHARGE_TYPE_ANON);
3956 * While swap-in, try_charge -> commit or cancel, the page is locked.
3957 * And when try_charge() successfully returns, one refcnt to memcg without
3958 * struct page_cgroup is acquired. This refcnt will be consumed by
3959 * "commit()" or removed by "cancel()"
3961 static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm,
3964 struct mem_cgroup **memcgp)
3966 struct mem_cgroup *memcg;
3967 struct page_cgroup *pc;
3970 pc = lookup_page_cgroup(page);
3972 * Every swap fault against a single page tries to charge the
3973 * page, bail as early as possible. shmem_unuse() encounters
3974 * already charged pages, too. The USED bit is protected by
3975 * the page lock, which serializes swap cache removal, which
3976 * in turn serializes uncharging.
3978 if (PageCgroupUsed(pc))
3980 if (!do_swap_account)
3982 memcg = try_get_mem_cgroup_from_page(page);
3986 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
3987 css_put(&memcg->css);
3992 ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
3998 int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page,
3999 gfp_t gfp_mask, struct mem_cgroup **memcgp)
4002 if (mem_cgroup_disabled())
4005 * A racing thread's fault, or swapoff, may have already
4006 * updated the pte, and even removed page from swap cache: in
4007 * those cases unuse_pte()'s pte_same() test will fail; but
4008 * there's also a KSM case which does need to charge the page.
4010 if (!PageSwapCache(page)) {
4013 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, memcgp, true);
4018 return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp);
4021 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
4023 if (mem_cgroup_disabled())
4027 __mem_cgroup_cancel_charge(memcg, 1);
4031 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
4032 enum charge_type ctype)
4034 if (mem_cgroup_disabled())
4039 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
4041 * Now swap is on-memory. This means this page may be
4042 * counted both as mem and swap....double count.
4043 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
4044 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
4045 * may call delete_from_swap_cache() before reach here.
4047 if (do_swap_account && PageSwapCache(page)) {
4048 swp_entry_t ent = {.val = page_private(page)};
4049 mem_cgroup_uncharge_swap(ent);
4053 void mem_cgroup_commit_charge_swapin(struct page *page,
4054 struct mem_cgroup *memcg)
4056 __mem_cgroup_commit_charge_swapin(page, memcg,
4057 MEM_CGROUP_CHARGE_TYPE_ANON);
4060 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
4063 struct mem_cgroup *memcg = NULL;
4064 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4067 if (mem_cgroup_disabled())
4069 if (PageCompound(page))
4072 if (!PageSwapCache(page))
4073 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
4074 else { /* page is swapcache/shmem */
4075 ret = __mem_cgroup_try_charge_swapin(mm, page,
4078 __mem_cgroup_commit_charge_swapin(page, memcg, type);
4083 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
4084 unsigned int nr_pages,
4085 const enum charge_type ctype)
4087 struct memcg_batch_info *batch = NULL;
4088 bool uncharge_memsw = true;
4090 /* If swapout, usage of swap doesn't decrease */
4091 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
4092 uncharge_memsw = false;
4094 batch = ¤t->memcg_batch;
4096 * In usual, we do css_get() when we remember memcg pointer.
4097 * But in this case, we keep res->usage until end of a series of
4098 * uncharges. Then, it's ok to ignore memcg's refcnt.
4101 batch->memcg = memcg;
4103 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
4104 * In those cases, all pages freed continuously can be expected to be in
4105 * the same cgroup and we have chance to coalesce uncharges.
4106 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
4107 * because we want to do uncharge as soon as possible.
4110 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
4111 goto direct_uncharge;
4114 goto direct_uncharge;
4117 * In typical case, batch->memcg == mem. This means we can
4118 * merge a series of uncharges to an uncharge of res_counter.
4119 * If not, we uncharge res_counter ony by one.
4121 if (batch->memcg != memcg)
4122 goto direct_uncharge;
4123 /* remember freed charge and uncharge it later */
4126 batch->memsw_nr_pages++;
4129 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
4131 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
4132 if (unlikely(batch->memcg != memcg))
4133 memcg_oom_recover(memcg);
4137 * uncharge if !page_mapped(page)
4139 static struct mem_cgroup *
4140 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
4143 struct mem_cgroup *memcg = NULL;
4144 unsigned int nr_pages = 1;
4145 struct page_cgroup *pc;
4148 if (mem_cgroup_disabled())
4151 if (PageTransHuge(page)) {
4152 nr_pages <<= compound_order(page);
4153 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
4156 * Check if our page_cgroup is valid
4158 pc = lookup_page_cgroup(page);
4159 if (unlikely(!PageCgroupUsed(pc)))
4162 lock_page_cgroup(pc);
4164 memcg = pc->mem_cgroup;
4166 if (!PageCgroupUsed(pc))
4169 anon = PageAnon(page);
4172 case MEM_CGROUP_CHARGE_TYPE_ANON:
4174 * Generally PageAnon tells if it's the anon statistics to be
4175 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
4176 * used before page reached the stage of being marked PageAnon.
4180 case MEM_CGROUP_CHARGE_TYPE_DROP:
4181 /* See mem_cgroup_prepare_migration() */
4182 if (page_mapped(page))
4185 * Pages under migration may not be uncharged. But
4186 * end_migration() /must/ be the one uncharging the
4187 * unused post-migration page and so it has to call
4188 * here with the migration bit still set. See the
4189 * res_counter handling below.
4191 if (!end_migration && PageCgroupMigration(pc))
4194 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
4195 if (!PageAnon(page)) { /* Shared memory */
4196 if (page->mapping && !page_is_file_cache(page))
4198 } else if (page_mapped(page)) /* Anon */
4205 mem_cgroup_charge_statistics(memcg, page, anon, -nr_pages);
4207 ClearPageCgroupUsed(pc);
4209 * pc->mem_cgroup is not cleared here. It will be accessed when it's
4210 * freed from LRU. This is safe because uncharged page is expected not
4211 * to be reused (freed soon). Exception is SwapCache, it's handled by
4212 * special functions.
4215 unlock_page_cgroup(pc);
4217 * even after unlock, we have memcg->res.usage here and this memcg
4218 * will never be freed, so it's safe to call css_get().
4220 memcg_check_events(memcg, page);
4221 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
4222 mem_cgroup_swap_statistics(memcg, true);
4223 css_get(&memcg->css);
4226 * Migration does not charge the res_counter for the
4227 * replacement page, so leave it alone when phasing out the
4228 * page that is unused after the migration.
4230 if (!end_migration && !mem_cgroup_is_root(memcg))
4231 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
4236 unlock_page_cgroup(pc);
4240 void mem_cgroup_uncharge_page(struct page *page)
4243 if (page_mapped(page))
4245 VM_BUG_ON_PAGE(page->mapping && !PageAnon(page), page);
4247 * If the page is in swap cache, uncharge should be deferred
4248 * to the swap path, which also properly accounts swap usage
4249 * and handles memcg lifetime.
4251 * Note that this check is not stable and reclaim may add the
4252 * page to swap cache at any time after this. However, if the
4253 * page is not in swap cache by the time page->mapcount hits
4254 * 0, there won't be any page table references to the swap
4255 * slot, and reclaim will free it and not actually write the
4258 if (PageSwapCache(page))
4260 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
4263 void mem_cgroup_uncharge_cache_page(struct page *page)
4265 VM_BUG_ON_PAGE(page_mapped(page), page);
4266 VM_BUG_ON_PAGE(page->mapping, page);
4267 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
4271 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4272 * In that cases, pages are freed continuously and we can expect pages
4273 * are in the same memcg. All these calls itself limits the number of
4274 * pages freed at once, then uncharge_start/end() is called properly.
4275 * This may be called prural(2) times in a context,
4278 void mem_cgroup_uncharge_start(void)
4280 current->memcg_batch.do_batch++;
4281 /* We can do nest. */
4282 if (current->memcg_batch.do_batch == 1) {
4283 current->memcg_batch.memcg = NULL;
4284 current->memcg_batch.nr_pages = 0;
4285 current->memcg_batch.memsw_nr_pages = 0;
4289 void mem_cgroup_uncharge_end(void)
4291 struct memcg_batch_info *batch = ¤t->memcg_batch;
4293 if (!batch->do_batch)
4297 if (batch->do_batch) /* If stacked, do nothing. */
4303 * This "batch->memcg" is valid without any css_get/put etc...
4304 * bacause we hide charges behind us.
4306 if (batch->nr_pages)
4307 res_counter_uncharge(&batch->memcg->res,
4308 batch->nr_pages * PAGE_SIZE);
4309 if (batch->memsw_nr_pages)
4310 res_counter_uncharge(&batch->memcg->memsw,
4311 batch->memsw_nr_pages * PAGE_SIZE);
4312 memcg_oom_recover(batch->memcg);
4313 /* forget this pointer (for sanity check) */
4314 batch->memcg = NULL;
4319 * called after __delete_from_swap_cache() and drop "page" account.
4320 * memcg information is recorded to swap_cgroup of "ent"
4323 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
4325 struct mem_cgroup *memcg;
4326 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
4328 if (!swapout) /* this was a swap cache but the swap is unused ! */
4329 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
4331 memcg = __mem_cgroup_uncharge_common(page, ctype, false);
4334 * record memcg information, if swapout && memcg != NULL,
4335 * css_get() was called in uncharge().
4337 if (do_swap_account && swapout && memcg)
4338 swap_cgroup_record(ent, mem_cgroup_id(memcg));
4342 #ifdef CONFIG_MEMCG_SWAP
4344 * called from swap_entry_free(). remove record in swap_cgroup and
4345 * uncharge "memsw" account.
4347 void mem_cgroup_uncharge_swap(swp_entry_t ent)
4349 struct mem_cgroup *memcg;
4352 if (!do_swap_account)
4355 id = swap_cgroup_record(ent, 0);
4357 memcg = mem_cgroup_lookup(id);
4360 * We uncharge this because swap is freed.
4361 * This memcg can be obsolete one. We avoid calling css_tryget
4363 if (!mem_cgroup_is_root(memcg))
4364 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
4365 mem_cgroup_swap_statistics(memcg, false);
4366 css_put(&memcg->css);
4372 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4373 * @entry: swap entry to be moved
4374 * @from: mem_cgroup which the entry is moved from
4375 * @to: mem_cgroup which the entry is moved to
4377 * It succeeds only when the swap_cgroup's record for this entry is the same
4378 * as the mem_cgroup's id of @from.
4380 * Returns 0 on success, -EINVAL on failure.
4382 * The caller must have charged to @to, IOW, called res_counter_charge() about
4383 * both res and memsw, and called css_get().
4385 static int mem_cgroup_move_swap_account(swp_entry_t entry,
4386 struct mem_cgroup *from, struct mem_cgroup *to)
4388 unsigned short old_id, new_id;
4390 old_id = mem_cgroup_id(from);
4391 new_id = mem_cgroup_id(to);
4393 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
4394 mem_cgroup_swap_statistics(from, false);
4395 mem_cgroup_swap_statistics(to, true);
4397 * This function is only called from task migration context now.
4398 * It postpones res_counter and refcount handling till the end
4399 * of task migration(mem_cgroup_clear_mc()) for performance
4400 * improvement. But we cannot postpone css_get(to) because if
4401 * the process that has been moved to @to does swap-in, the
4402 * refcount of @to might be decreased to 0.
4404 * We are in attach() phase, so the cgroup is guaranteed to be
4405 * alive, so we can just call css_get().
4413 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
4414 struct mem_cgroup *from, struct mem_cgroup *to)
4421 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4424 void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
4425 struct mem_cgroup **memcgp)
4427 struct mem_cgroup *memcg = NULL;
4428 unsigned int nr_pages = 1;
4429 struct page_cgroup *pc;
4430 enum charge_type ctype;
4434 if (mem_cgroup_disabled())
4437 if (PageTransHuge(page))
4438 nr_pages <<= compound_order(page);
4440 pc = lookup_page_cgroup(page);
4441 lock_page_cgroup(pc);
4442 if (PageCgroupUsed(pc)) {
4443 memcg = pc->mem_cgroup;
4444 css_get(&memcg->css);
4446 * At migrating an anonymous page, its mapcount goes down
4447 * to 0 and uncharge() will be called. But, even if it's fully
4448 * unmapped, migration may fail and this page has to be
4449 * charged again. We set MIGRATION flag here and delay uncharge
4450 * until end_migration() is called
4452 * Corner Case Thinking
4454 * When the old page was mapped as Anon and it's unmap-and-freed
4455 * while migration was ongoing.
4456 * If unmap finds the old page, uncharge() of it will be delayed
4457 * until end_migration(). If unmap finds a new page, it's
4458 * uncharged when it make mapcount to be 1->0. If unmap code
4459 * finds swap_migration_entry, the new page will not be mapped
4460 * and end_migration() will find it(mapcount==0).
4463 * When the old page was mapped but migraion fails, the kernel
4464 * remaps it. A charge for it is kept by MIGRATION flag even
4465 * if mapcount goes down to 0. We can do remap successfully
4466 * without charging it again.
4469 * The "old" page is under lock_page() until the end of
4470 * migration, so, the old page itself will not be swapped-out.
4471 * If the new page is swapped out before end_migraton, our
4472 * hook to usual swap-out path will catch the event.
4475 SetPageCgroupMigration(pc);
4477 unlock_page_cgroup(pc);
4479 * If the page is not charged at this point,
4487 * We charge new page before it's used/mapped. So, even if unlock_page()
4488 * is called before end_migration, we can catch all events on this new
4489 * page. In the case new page is migrated but not remapped, new page's
4490 * mapcount will be finally 0 and we call uncharge in end_migration().
4493 ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
4495 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
4497 * The page is committed to the memcg, but it's not actually
4498 * charged to the res_counter since we plan on replacing the
4499 * old one and only one page is going to be left afterwards.
4501 __mem_cgroup_commit_charge(memcg, newpage, nr_pages, ctype, false);
4504 /* remove redundant charge if migration failed*/
4505 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
4506 struct page *oldpage, struct page *newpage, bool migration_ok)
4508 struct page *used, *unused;
4509 struct page_cgroup *pc;
4515 if (!migration_ok) {
4522 anon = PageAnon(used);
4523 __mem_cgroup_uncharge_common(unused,
4524 anon ? MEM_CGROUP_CHARGE_TYPE_ANON
4525 : MEM_CGROUP_CHARGE_TYPE_CACHE,
4527 css_put(&memcg->css);
4529 * We disallowed uncharge of pages under migration because mapcount
4530 * of the page goes down to zero, temporarly.
4531 * Clear the flag and check the page should be charged.
4533 pc = lookup_page_cgroup(oldpage);
4534 lock_page_cgroup(pc);
4535 ClearPageCgroupMigration(pc);
4536 unlock_page_cgroup(pc);
4539 * If a page is a file cache, radix-tree replacement is very atomic
4540 * and we can skip this check. When it was an Anon page, its mapcount
4541 * goes down to 0. But because we added MIGRATION flage, it's not
4542 * uncharged yet. There are several case but page->mapcount check
4543 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4544 * check. (see prepare_charge() also)
4547 mem_cgroup_uncharge_page(used);
4551 * At replace page cache, newpage is not under any memcg but it's on
4552 * LRU. So, this function doesn't touch res_counter but handles LRU
4553 * in correct way. Both pages are locked so we cannot race with uncharge.
4555 void mem_cgroup_replace_page_cache(struct page *oldpage,
4556 struct page *newpage)
4558 struct mem_cgroup *memcg = NULL;
4559 struct page_cgroup *pc;
4560 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4562 if (mem_cgroup_disabled())
4565 pc = lookup_page_cgroup(oldpage);
4566 /* fix accounting on old pages */
4567 lock_page_cgroup(pc);
4568 if (PageCgroupUsed(pc)) {
4569 memcg = pc->mem_cgroup;
4570 mem_cgroup_charge_statistics(memcg, oldpage, false, -1);
4571 ClearPageCgroupUsed(pc);
4573 unlock_page_cgroup(pc);
4576 * When called from shmem_replace_page(), in some cases the
4577 * oldpage has already been charged, and in some cases not.
4582 * Even if newpage->mapping was NULL before starting replacement,
4583 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4584 * LRU while we overwrite pc->mem_cgroup.
4586 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
4589 #ifdef CONFIG_DEBUG_VM
4590 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
4592 struct page_cgroup *pc;
4594 pc = lookup_page_cgroup(page);
4596 * Can be NULL while feeding pages into the page allocator for
4597 * the first time, i.e. during boot or memory hotplug;
4598 * or when mem_cgroup_disabled().
4600 if (likely(pc) && PageCgroupUsed(pc))
4605 bool mem_cgroup_bad_page_check(struct page *page)
4607 if (mem_cgroup_disabled())
4610 return lookup_page_cgroup_used(page) != NULL;
4613 void mem_cgroup_print_bad_page(struct page *page)
4615 struct page_cgroup *pc;
4617 pc = lookup_page_cgroup_used(page);
4619 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4620 pc, pc->flags, pc->mem_cgroup);
4625 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
4626 unsigned long long val)
4629 u64 memswlimit, memlimit;
4631 int children = mem_cgroup_count_children(memcg);
4632 u64 curusage, oldusage;
4636 * For keeping hierarchical_reclaim simple, how long we should retry
4637 * is depends on callers. We set our retry-count to be function
4638 * of # of children which we should visit in this loop.
4640 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
4642 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4645 while (retry_count) {
4646 if (signal_pending(current)) {
4651 * Rather than hide all in some function, I do this in
4652 * open coded manner. You see what this really does.
4653 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4655 mutex_lock(&set_limit_mutex);
4656 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4657 if (memswlimit < val) {
4659 mutex_unlock(&set_limit_mutex);
4663 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4667 ret = res_counter_set_limit(&memcg->res, val);
4669 if (memswlimit == val)
4670 memcg->memsw_is_minimum = true;
4672 memcg->memsw_is_minimum = false;
4674 mutex_unlock(&set_limit_mutex);
4679 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4680 MEM_CGROUP_RECLAIM_SHRINK);
4681 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4682 /* Usage is reduced ? */
4683 if (curusage >= oldusage)
4686 oldusage = curusage;
4688 if (!ret && enlarge)
4689 memcg_oom_recover(memcg);
4694 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
4695 unsigned long long val)
4698 u64 memlimit, memswlimit, oldusage, curusage;
4699 int children = mem_cgroup_count_children(memcg);
4703 /* see mem_cgroup_resize_res_limit */
4704 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
4705 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4706 while (retry_count) {
4707 if (signal_pending(current)) {
4712 * Rather than hide all in some function, I do this in
4713 * open coded manner. You see what this really does.
4714 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4716 mutex_lock(&set_limit_mutex);
4717 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4718 if (memlimit > val) {
4720 mutex_unlock(&set_limit_mutex);
4723 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4724 if (memswlimit < val)
4726 ret = res_counter_set_limit(&memcg->memsw, val);
4728 if (memlimit == val)
4729 memcg->memsw_is_minimum = true;
4731 memcg->memsw_is_minimum = false;
4733 mutex_unlock(&set_limit_mutex);
4738 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4739 MEM_CGROUP_RECLAIM_NOSWAP |
4740 MEM_CGROUP_RECLAIM_SHRINK);
4741 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4742 /* Usage is reduced ? */
4743 if (curusage >= oldusage)
4746 oldusage = curusage;
4748 if (!ret && enlarge)
4749 memcg_oom_recover(memcg);
4753 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
4755 unsigned long *total_scanned)
4757 unsigned long nr_reclaimed = 0;
4758 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
4759 unsigned long reclaimed;
4761 struct mem_cgroup_tree_per_zone *mctz;
4762 unsigned long long excess;
4763 unsigned long nr_scanned;
4768 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
4770 * This loop can run a while, specially if mem_cgroup's continuously
4771 * keep exceeding their soft limit and putting the system under
4778 mz = mem_cgroup_largest_soft_limit_node(mctz);
4783 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
4784 gfp_mask, &nr_scanned);
4785 nr_reclaimed += reclaimed;
4786 *total_scanned += nr_scanned;
4787 spin_lock(&mctz->lock);
4790 * If we failed to reclaim anything from this memory cgroup
4791 * it is time to move on to the next cgroup
4797 * Loop until we find yet another one.
4799 * By the time we get the soft_limit lock
4800 * again, someone might have aded the
4801 * group back on the RB tree. Iterate to
4802 * make sure we get a different mem.
4803 * mem_cgroup_largest_soft_limit_node returns
4804 * NULL if no other cgroup is present on
4808 __mem_cgroup_largest_soft_limit_node(mctz);
4810 css_put(&next_mz->memcg->css);
4811 else /* next_mz == NULL or other memcg */
4815 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
4816 excess = res_counter_soft_limit_excess(&mz->memcg->res);
4818 * One school of thought says that we should not add
4819 * back the node to the tree if reclaim returns 0.
4820 * But our reclaim could return 0, simply because due
4821 * to priority we are exposing a smaller subset of
4822 * memory to reclaim from. Consider this as a longer
4825 /* If excess == 0, no tree ops */
4826 __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess);
4827 spin_unlock(&mctz->lock);
4828 css_put(&mz->memcg->css);
4831 * Could not reclaim anything and there are no more
4832 * mem cgroups to try or we seem to be looping without
4833 * reclaiming anything.
4835 if (!nr_reclaimed &&
4837 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
4839 } while (!nr_reclaimed);
4841 css_put(&next_mz->memcg->css);
4842 return nr_reclaimed;
4846 * mem_cgroup_force_empty_list - clears LRU of a group
4847 * @memcg: group to clear
4850 * @lru: lru to to clear
4852 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4853 * reclaim the pages page themselves - pages are moved to the parent (or root)
4856 static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
4857 int node, int zid, enum lru_list lru)
4859 struct lruvec *lruvec;
4860 unsigned long flags;
4861 struct list_head *list;
4865 zone = &NODE_DATA(node)->node_zones[zid];
4866 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
4867 list = &lruvec->lists[lru];
4871 struct page_cgroup *pc;
4874 spin_lock_irqsave(&zone->lru_lock, flags);
4875 if (list_empty(list)) {
4876 spin_unlock_irqrestore(&zone->lru_lock, flags);
4879 page = list_entry(list->prev, struct page, lru);
4881 list_move(&page->lru, list);
4883 spin_unlock_irqrestore(&zone->lru_lock, flags);
4886 spin_unlock_irqrestore(&zone->lru_lock, flags);
4888 pc = lookup_page_cgroup(page);
4890 if (mem_cgroup_move_parent(page, pc, memcg)) {
4891 /* found lock contention or "pc" is obsolete. */
4896 } while (!list_empty(list));
4900 * make mem_cgroup's charge to be 0 if there is no task by moving
4901 * all the charges and pages to the parent.
4902 * This enables deleting this mem_cgroup.
4904 * Caller is responsible for holding css reference on the memcg.
4906 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
4912 /* This is for making all *used* pages to be on LRU. */
4913 lru_add_drain_all();
4914 drain_all_stock_sync(memcg);
4915 mem_cgroup_start_move(memcg);
4916 for_each_node_state(node, N_MEMORY) {
4917 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4920 mem_cgroup_force_empty_list(memcg,
4925 mem_cgroup_end_move(memcg);
4926 memcg_oom_recover(memcg);
4930 * Kernel memory may not necessarily be trackable to a specific
4931 * process. So they are not migrated, and therefore we can't
4932 * expect their value to drop to 0 here.
4933 * Having res filled up with kmem only is enough.
4935 * This is a safety check because mem_cgroup_force_empty_list
4936 * could have raced with mem_cgroup_replace_page_cache callers
4937 * so the lru seemed empty but the page could have been added
4938 * right after the check. RES_USAGE should be safe as we always
4939 * charge before adding to the LRU.
4941 usage = res_counter_read_u64(&memcg->res, RES_USAGE) -
4942 res_counter_read_u64(&memcg->kmem, RES_USAGE);
4943 } while (usage > 0);
4946 static inline bool memcg_has_children(struct mem_cgroup *memcg)
4948 lockdep_assert_held(&memcg_create_mutex);
4950 * The lock does not prevent addition or deletion to the list
4951 * of children, but it prevents a new child from being
4952 * initialized based on this parent in css_online(), so it's
4953 * enough to decide whether hierarchically inherited
4954 * attributes can still be changed or not.
4956 return memcg->use_hierarchy &&
4957 !list_empty(&memcg->css.cgroup->children);
4961 * Reclaims as many pages from the given memcg as possible and moves
4962 * the rest to the parent.
4964 * Caller is responsible for holding css reference for memcg.
4966 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
4968 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
4969 struct cgroup *cgrp = memcg->css.cgroup;
4971 /* returns EBUSY if there is a task or if we come here twice. */
4972 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
4975 /* we call try-to-free pages for make this cgroup empty */
4976 lru_add_drain_all();
4977 /* try to free all pages in this cgroup */
4978 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
4981 if (signal_pending(current))
4984 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
4988 /* maybe some writeback is necessary */
4989 congestion_wait(BLK_RW_ASYNC, HZ/10);
4994 mem_cgroup_reparent_charges(memcg);
4999 static int mem_cgroup_force_empty_write(struct cgroup_subsys_state *css,
5002 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5004 if (mem_cgroup_is_root(memcg))
5006 return mem_cgroup_force_empty(memcg);
5009 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
5012 return mem_cgroup_from_css(css)->use_hierarchy;
5015 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
5016 struct cftype *cft, u64 val)
5019 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5020 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(css_parent(&memcg->css));
5022 mutex_lock(&memcg_create_mutex);
5024 if (memcg->use_hierarchy == val)
5028 * If parent's use_hierarchy is set, we can't make any modifications
5029 * in the child subtrees. If it is unset, then the change can
5030 * occur, provided the current cgroup has no children.
5032 * For the root cgroup, parent_mem is NULL, we allow value to be
5033 * set if there are no children.
5035 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
5036 (val == 1 || val == 0)) {
5037 if (list_empty(&memcg->css.cgroup->children))
5038 memcg->use_hierarchy = val;
5045 mutex_unlock(&memcg_create_mutex);
5051 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
5052 enum mem_cgroup_stat_index idx)
5054 struct mem_cgroup *iter;
5057 /* Per-cpu values can be negative, use a signed accumulator */
5058 for_each_mem_cgroup_tree(iter, memcg)
5059 val += mem_cgroup_read_stat(iter, idx);
5061 if (val < 0) /* race ? */
5066 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
5070 if (!mem_cgroup_is_root(memcg)) {
5072 return res_counter_read_u64(&memcg->res, RES_USAGE);
5074 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
5078 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
5079 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
5081 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
5082 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
5085 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
5087 return val << PAGE_SHIFT;
5090 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
5093 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5098 type = MEMFILE_TYPE(cft->private);
5099 name = MEMFILE_ATTR(cft->private);
5103 if (name == RES_USAGE)
5104 val = mem_cgroup_usage(memcg, false);
5106 val = res_counter_read_u64(&memcg->res, name);
5109 if (name == RES_USAGE)
5110 val = mem_cgroup_usage(memcg, true);
5112 val = res_counter_read_u64(&memcg->memsw, name);
5115 val = res_counter_read_u64(&memcg->kmem, name);
5124 #ifdef CONFIG_MEMCG_KMEM
5125 /* should be called with activate_kmem_mutex held */
5126 static int __memcg_activate_kmem(struct mem_cgroup *memcg,
5127 unsigned long long limit)
5132 if (memcg_kmem_is_active(memcg))
5136 * We are going to allocate memory for data shared by all memory
5137 * cgroups so let's stop accounting here.
5139 memcg_stop_kmem_account();
5142 * For simplicity, we won't allow this to be disabled. It also can't
5143 * be changed if the cgroup has children already, or if tasks had
5146 * If tasks join before we set the limit, a person looking at
5147 * kmem.usage_in_bytes will have no way to determine when it took
5148 * place, which makes the value quite meaningless.
5150 * After it first became limited, changes in the value of the limit are
5151 * of course permitted.
5153 mutex_lock(&memcg_create_mutex);
5154 if (cgroup_task_count(memcg->css.cgroup) || memcg_has_children(memcg))
5156 mutex_unlock(&memcg_create_mutex);
5160 memcg_id = ida_simple_get(&kmem_limited_groups,
5161 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
5168 * Make sure we have enough space for this cgroup in each root cache's
5171 err = memcg_update_all_caches(memcg_id + 1);
5175 memcg->kmemcg_id = memcg_id;
5176 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
5177 mutex_init(&memcg->slab_caches_mutex);
5180 * We couldn't have accounted to this cgroup, because it hasn't got the
5181 * active bit set yet, so this should succeed.
5183 err = res_counter_set_limit(&memcg->kmem, limit);
5186 static_key_slow_inc(&memcg_kmem_enabled_key);
5188 * Setting the active bit after enabling static branching will
5189 * guarantee no one starts accounting before all call sites are
5192 memcg_kmem_set_active(memcg);
5194 memcg_resume_kmem_account();
5198 ida_simple_remove(&kmem_limited_groups, memcg_id);
5202 static int memcg_activate_kmem(struct mem_cgroup *memcg,
5203 unsigned long long limit)
5207 mutex_lock(&activate_kmem_mutex);
5208 ret = __memcg_activate_kmem(memcg, limit);
5209 mutex_unlock(&activate_kmem_mutex);
5213 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
5214 unsigned long long val)
5218 if (!memcg_kmem_is_active(memcg))
5219 ret = memcg_activate_kmem(memcg, val);
5221 ret = res_counter_set_limit(&memcg->kmem, val);
5225 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
5228 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5233 mutex_lock(&activate_kmem_mutex);
5235 * If the parent cgroup is not kmem-active now, it cannot be activated
5236 * after this point, because it has at least one child already.
5238 if (memcg_kmem_is_active(parent))
5239 ret = __memcg_activate_kmem(memcg, RES_COUNTER_MAX);
5240 mutex_unlock(&activate_kmem_mutex);
5244 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
5245 unsigned long long val)
5249 #endif /* CONFIG_MEMCG_KMEM */
5252 * The user of this function is...
5255 static int mem_cgroup_write(struct cgroup_subsys_state *css, struct cftype *cft,
5258 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5261 unsigned long long val;
5264 type = MEMFILE_TYPE(cft->private);
5265 name = MEMFILE_ATTR(cft->private);
5269 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
5273 /* This function does all necessary parse...reuse it */
5274 ret = res_counter_memparse_write_strategy(buffer, &val);
5278 ret = mem_cgroup_resize_limit(memcg, val);
5279 else if (type == _MEMSWAP)
5280 ret = mem_cgroup_resize_memsw_limit(memcg, val);
5281 else if (type == _KMEM)
5282 ret = memcg_update_kmem_limit(memcg, val);
5286 case RES_SOFT_LIMIT:
5287 ret = res_counter_memparse_write_strategy(buffer, &val);
5291 * For memsw, soft limits are hard to implement in terms
5292 * of semantics, for now, we support soft limits for
5293 * control without swap
5296 ret = res_counter_set_soft_limit(&memcg->res, val);
5301 ret = -EINVAL; /* should be BUG() ? */
5307 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
5308 unsigned long long *mem_limit, unsigned long long *memsw_limit)
5310 unsigned long long min_limit, min_memsw_limit, tmp;
5312 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
5313 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5314 if (!memcg->use_hierarchy)
5317 while (css_parent(&memcg->css)) {
5318 memcg = mem_cgroup_from_css(css_parent(&memcg->css));
5319 if (!memcg->use_hierarchy)
5321 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
5322 min_limit = min(min_limit, tmp);
5323 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5324 min_memsw_limit = min(min_memsw_limit, tmp);
5327 *mem_limit = min_limit;
5328 *memsw_limit = min_memsw_limit;
5331 static int mem_cgroup_reset(struct cgroup_subsys_state *css, unsigned int event)
5333 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5337 type = MEMFILE_TYPE(event);
5338 name = MEMFILE_ATTR(event);
5343 res_counter_reset_max(&memcg->res);
5344 else if (type == _MEMSWAP)
5345 res_counter_reset_max(&memcg->memsw);
5346 else if (type == _KMEM)
5347 res_counter_reset_max(&memcg->kmem);
5353 res_counter_reset_failcnt(&memcg->res);
5354 else if (type == _MEMSWAP)
5355 res_counter_reset_failcnt(&memcg->memsw);
5356 else if (type == _KMEM)
5357 res_counter_reset_failcnt(&memcg->kmem);
5366 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
5369 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
5373 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5374 struct cftype *cft, u64 val)
5376 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5378 if (val >= (1 << NR_MOVE_TYPE))
5382 * No kind of locking is needed in here, because ->can_attach() will
5383 * check this value once in the beginning of the process, and then carry
5384 * on with stale data. This means that changes to this value will only
5385 * affect task migrations starting after the change.
5387 memcg->move_charge_at_immigrate = val;
5391 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5392 struct cftype *cft, u64 val)
5399 static int memcg_numa_stat_show(struct seq_file *m, void *v)
5403 unsigned int lru_mask;
5406 static const struct numa_stat stats[] = {
5407 { "total", LRU_ALL },
5408 { "file", LRU_ALL_FILE },
5409 { "anon", LRU_ALL_ANON },
5410 { "unevictable", BIT(LRU_UNEVICTABLE) },
5412 const struct numa_stat *stat;
5415 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5417 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
5418 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
5419 seq_printf(m, "%s=%lu", stat->name, nr);
5420 for_each_node_state(nid, N_MEMORY) {
5421 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5423 seq_printf(m, " N%d=%lu", nid, nr);
5428 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
5429 struct mem_cgroup *iter;
5432 for_each_mem_cgroup_tree(iter, memcg)
5433 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
5434 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
5435 for_each_node_state(nid, N_MEMORY) {
5437 for_each_mem_cgroup_tree(iter, memcg)
5438 nr += mem_cgroup_node_nr_lru_pages(
5439 iter, nid, stat->lru_mask);
5440 seq_printf(m, " N%d=%lu", nid, nr);
5447 #endif /* CONFIG_NUMA */
5449 static inline void mem_cgroup_lru_names_not_uptodate(void)
5451 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
5454 static int memcg_stat_show(struct seq_file *m, void *v)
5456 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5457 struct mem_cgroup *mi;
5460 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5461 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5463 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
5464 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
5467 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
5468 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
5469 mem_cgroup_read_events(memcg, i));
5471 for (i = 0; i < NR_LRU_LISTS; i++)
5472 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
5473 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
5475 /* Hierarchical information */
5477 unsigned long long limit, memsw_limit;
5478 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
5479 seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
5480 if (do_swap_account)
5481 seq_printf(m, "hierarchical_memsw_limit %llu\n",
5485 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5488 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5490 for_each_mem_cgroup_tree(mi, memcg)
5491 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
5492 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
5495 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
5496 unsigned long long val = 0;
5498 for_each_mem_cgroup_tree(mi, memcg)
5499 val += mem_cgroup_read_events(mi, i);
5500 seq_printf(m, "total_%s %llu\n",
5501 mem_cgroup_events_names[i], val);
5504 for (i = 0; i < NR_LRU_LISTS; i++) {
5505 unsigned long long val = 0;
5507 for_each_mem_cgroup_tree(mi, memcg)
5508 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
5509 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
5512 #ifdef CONFIG_DEBUG_VM
5515 struct mem_cgroup_per_zone *mz;
5516 struct zone_reclaim_stat *rstat;
5517 unsigned long recent_rotated[2] = {0, 0};
5518 unsigned long recent_scanned[2] = {0, 0};
5520 for_each_online_node(nid)
5521 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
5522 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
5523 rstat = &mz->lruvec.reclaim_stat;
5525 recent_rotated[0] += rstat->recent_rotated[0];
5526 recent_rotated[1] += rstat->recent_rotated[1];
5527 recent_scanned[0] += rstat->recent_scanned[0];
5528 recent_scanned[1] += rstat->recent_scanned[1];
5530 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
5531 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
5532 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
5533 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
5540 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
5543 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5545 return mem_cgroup_swappiness(memcg);
5548 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
5549 struct cftype *cft, u64 val)
5551 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5552 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(&memcg->css));
5554 if (val > 100 || !parent)
5557 mutex_lock(&memcg_create_mutex);
5559 /* If under hierarchy, only empty-root can set this value */
5560 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5561 mutex_unlock(&memcg_create_mutex);
5565 memcg->swappiness = val;
5567 mutex_unlock(&memcg_create_mutex);
5572 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
5574 struct mem_cgroup_threshold_ary *t;
5580 t = rcu_dereference(memcg->thresholds.primary);
5582 t = rcu_dereference(memcg->memsw_thresholds.primary);
5587 usage = mem_cgroup_usage(memcg, swap);
5590 * current_threshold points to threshold just below or equal to usage.
5591 * If it's not true, a threshold was crossed after last
5592 * call of __mem_cgroup_threshold().
5594 i = t->current_threshold;
5597 * Iterate backward over array of thresholds starting from
5598 * current_threshold and check if a threshold is crossed.
5599 * If none of thresholds below usage is crossed, we read
5600 * only one element of the array here.
5602 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
5603 eventfd_signal(t->entries[i].eventfd, 1);
5605 /* i = current_threshold + 1 */
5609 * Iterate forward over array of thresholds starting from
5610 * current_threshold+1 and check if a threshold is crossed.
5611 * If none of thresholds above usage is crossed, we read
5612 * only one element of the array here.
5614 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
5615 eventfd_signal(t->entries[i].eventfd, 1);
5617 /* Update current_threshold */
5618 t->current_threshold = i - 1;
5623 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
5626 __mem_cgroup_threshold(memcg, false);
5627 if (do_swap_account)
5628 __mem_cgroup_threshold(memcg, true);
5630 memcg = parent_mem_cgroup(memcg);
5634 static int compare_thresholds(const void *a, const void *b)
5636 const struct mem_cgroup_threshold *_a = a;
5637 const struct mem_cgroup_threshold *_b = b;
5639 if (_a->threshold > _b->threshold)
5642 if (_a->threshold < _b->threshold)
5648 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
5650 struct mem_cgroup_eventfd_list *ev;
5652 list_for_each_entry(ev, &memcg->oom_notify, list)
5653 eventfd_signal(ev->eventfd, 1);
5657 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
5659 struct mem_cgroup *iter;
5661 for_each_mem_cgroup_tree(iter, memcg)
5662 mem_cgroup_oom_notify_cb(iter);
5665 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
5666 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
5668 struct mem_cgroup_thresholds *thresholds;
5669 struct mem_cgroup_threshold_ary *new;
5670 u64 threshold, usage;
5673 ret = res_counter_memparse_write_strategy(args, &threshold);
5677 mutex_lock(&memcg->thresholds_lock);
5680 thresholds = &memcg->thresholds;
5681 else if (type == _MEMSWAP)
5682 thresholds = &memcg->memsw_thresholds;
5686 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5688 /* Check if a threshold crossed before adding a new one */
5689 if (thresholds->primary)
5690 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5692 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
5694 /* Allocate memory for new array of thresholds */
5695 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
5703 /* Copy thresholds (if any) to new array */
5704 if (thresholds->primary) {
5705 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
5706 sizeof(struct mem_cgroup_threshold));
5709 /* Add new threshold */
5710 new->entries[size - 1].eventfd = eventfd;
5711 new->entries[size - 1].threshold = threshold;
5713 /* Sort thresholds. Registering of new threshold isn't time-critical */
5714 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
5715 compare_thresholds, NULL);
5717 /* Find current threshold */
5718 new->current_threshold = -1;
5719 for (i = 0; i < size; i++) {
5720 if (new->entries[i].threshold <= usage) {
5722 * new->current_threshold will not be used until
5723 * rcu_assign_pointer(), so it's safe to increment
5726 ++new->current_threshold;
5731 /* Free old spare buffer and save old primary buffer as spare */
5732 kfree(thresholds->spare);
5733 thresholds->spare = thresholds->primary;
5735 rcu_assign_pointer(thresholds->primary, new);
5737 /* To be sure that nobody uses thresholds */
5741 mutex_unlock(&memcg->thresholds_lock);
5746 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
5747 struct eventfd_ctx *eventfd, const char *args)
5749 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
5752 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
5753 struct eventfd_ctx *eventfd, const char *args)
5755 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
5758 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
5759 struct eventfd_ctx *eventfd, enum res_type type)
5761 struct mem_cgroup_thresholds *thresholds;
5762 struct mem_cgroup_threshold_ary *new;
5766 mutex_lock(&memcg->thresholds_lock);
5768 thresholds = &memcg->thresholds;
5769 else if (type == _MEMSWAP)
5770 thresholds = &memcg->memsw_thresholds;
5774 if (!thresholds->primary)
5777 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5779 /* Check if a threshold crossed before removing */
5780 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5782 /* Calculate new number of threshold */
5784 for (i = 0; i < thresholds->primary->size; i++) {
5785 if (thresholds->primary->entries[i].eventfd != eventfd)
5789 new = thresholds->spare;
5791 /* Set thresholds array to NULL if we don't have thresholds */
5800 /* Copy thresholds and find current threshold */
5801 new->current_threshold = -1;
5802 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
5803 if (thresholds->primary->entries[i].eventfd == eventfd)
5806 new->entries[j] = thresholds->primary->entries[i];
5807 if (new->entries[j].threshold <= usage) {
5809 * new->current_threshold will not be used
5810 * until rcu_assign_pointer(), so it's safe to increment
5813 ++new->current_threshold;
5819 /* Swap primary and spare array */
5820 thresholds->spare = thresholds->primary;
5821 /* If all events are unregistered, free the spare array */
5823 kfree(thresholds->spare);
5824 thresholds->spare = NULL;
5827 rcu_assign_pointer(thresholds->primary, new);
5829 /* To be sure that nobody uses thresholds */
5832 mutex_unlock(&memcg->thresholds_lock);
5835 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
5836 struct eventfd_ctx *eventfd)
5838 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
5841 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
5842 struct eventfd_ctx *eventfd)
5844 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
5847 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
5848 struct eventfd_ctx *eventfd, const char *args)
5850 struct mem_cgroup_eventfd_list *event;
5852 event = kmalloc(sizeof(*event), GFP_KERNEL);
5856 spin_lock(&memcg_oom_lock);
5858 event->eventfd = eventfd;
5859 list_add(&event->list, &memcg->oom_notify);
5861 /* already in OOM ? */
5862 if (atomic_read(&memcg->under_oom))
5863 eventfd_signal(eventfd, 1);
5864 spin_unlock(&memcg_oom_lock);
5869 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
5870 struct eventfd_ctx *eventfd)
5872 struct mem_cgroup_eventfd_list *ev, *tmp;
5874 spin_lock(&memcg_oom_lock);
5876 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
5877 if (ev->eventfd == eventfd) {
5878 list_del(&ev->list);
5883 spin_unlock(&memcg_oom_lock);
5886 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
5888 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
5890 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
5891 seq_printf(sf, "under_oom %d\n", (bool)atomic_read(&memcg->under_oom));
5895 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
5896 struct cftype *cft, u64 val)
5898 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5899 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(&memcg->css));
5901 /* cannot set to root cgroup and only 0 and 1 are allowed */
5902 if (!parent || !((val == 0) || (val == 1)))
5905 mutex_lock(&memcg_create_mutex);
5906 /* oom-kill-disable is a flag for subhierarchy. */
5907 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5908 mutex_unlock(&memcg_create_mutex);
5911 memcg->oom_kill_disable = val;
5913 memcg_oom_recover(memcg);
5914 mutex_unlock(&memcg_create_mutex);
5918 #ifdef CONFIG_MEMCG_KMEM
5919 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5923 memcg->kmemcg_id = -1;
5924 ret = memcg_propagate_kmem(memcg);
5928 return mem_cgroup_sockets_init(memcg, ss);
5931 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5933 mem_cgroup_sockets_destroy(memcg);
5936 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5938 if (!memcg_kmem_is_active(memcg))
5942 * kmem charges can outlive the cgroup. In the case of slab
5943 * pages, for instance, a page contain objects from various
5944 * processes. As we prevent from taking a reference for every
5945 * such allocation we have to be careful when doing uncharge
5946 * (see memcg_uncharge_kmem) and here during offlining.
5948 * The idea is that that only the _last_ uncharge which sees
5949 * the dead memcg will drop the last reference. An additional
5950 * reference is taken here before the group is marked dead
5951 * which is then paired with css_put during uncharge resp. here.
5953 * Although this might sound strange as this path is called from
5954 * css_offline() when the referencemight have dropped down to 0
5955 * and shouldn't be incremented anymore (css_tryget would fail)
5956 * we do not have other options because of the kmem allocations
5959 css_get(&memcg->css);
5961 memcg_kmem_mark_dead(memcg);
5963 if (res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0)
5966 if (memcg_kmem_test_and_clear_dead(memcg))
5967 css_put(&memcg->css);
5970 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5975 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5979 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5985 * DO NOT USE IN NEW FILES.
5987 * "cgroup.event_control" implementation.
5989 * This is way over-engineered. It tries to support fully configurable
5990 * events for each user. Such level of flexibility is completely
5991 * unnecessary especially in the light of the planned unified hierarchy.
5993 * Please deprecate this and replace with something simpler if at all
5998 * Unregister event and free resources.
6000 * Gets called from workqueue.
6002 static void memcg_event_remove(struct work_struct *work)
6004 struct mem_cgroup_event *event =
6005 container_of(work, struct mem_cgroup_event, remove);
6006 struct mem_cgroup *memcg = event->memcg;
6008 remove_wait_queue(event->wqh, &event->wait);
6010 event->unregister_event(memcg, event->eventfd);
6012 /* Notify userspace the event is going away. */
6013 eventfd_signal(event->eventfd, 1);
6015 eventfd_ctx_put(event->eventfd);
6017 css_put(&memcg->css);
6021 * Gets called on POLLHUP on eventfd when user closes it.
6023 * Called with wqh->lock held and interrupts disabled.
6025 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
6026 int sync, void *key)
6028 struct mem_cgroup_event *event =
6029 container_of(wait, struct mem_cgroup_event, wait);
6030 struct mem_cgroup *memcg = event->memcg;
6031 unsigned long flags = (unsigned long)key;
6033 if (flags & POLLHUP) {
6035 * If the event has been detached at cgroup removal, we
6036 * can simply return knowing the other side will cleanup
6039 * We can't race against event freeing since the other
6040 * side will require wqh->lock via remove_wait_queue(),
6043 spin_lock(&memcg->event_list_lock);
6044 if (!list_empty(&event->list)) {
6045 list_del_init(&event->list);
6047 * We are in atomic context, but cgroup_event_remove()
6048 * may sleep, so we have to call it in workqueue.
6050 schedule_work(&event->remove);
6052 spin_unlock(&memcg->event_list_lock);
6058 static void memcg_event_ptable_queue_proc(struct file *file,
6059 wait_queue_head_t *wqh, poll_table *pt)
6061 struct mem_cgroup_event *event =
6062 container_of(pt, struct mem_cgroup_event, pt);
6065 add_wait_queue(wqh, &event->wait);
6069 * DO NOT USE IN NEW FILES.
6071 * Parse input and register new cgroup event handler.
6073 * Input must be in format '<event_fd> <control_fd> <args>'.
6074 * Interpretation of args is defined by control file implementation.
6076 static int memcg_write_event_control(struct cgroup_subsys_state *css,
6077 struct cftype *cft, const char *buffer)
6079 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6080 struct mem_cgroup_event *event;
6081 struct cgroup_subsys_state *cfile_css;
6082 unsigned int efd, cfd;
6089 efd = simple_strtoul(buffer, &endp, 10);
6094 cfd = simple_strtoul(buffer, &endp, 10);
6095 if ((*endp != ' ') && (*endp != '\0'))
6099 event = kzalloc(sizeof(*event), GFP_KERNEL);
6103 event->memcg = memcg;
6104 INIT_LIST_HEAD(&event->list);
6105 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
6106 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
6107 INIT_WORK(&event->remove, memcg_event_remove);
6115 event->eventfd = eventfd_ctx_fileget(efile.file);
6116 if (IS_ERR(event->eventfd)) {
6117 ret = PTR_ERR(event->eventfd);
6124 goto out_put_eventfd;
6127 /* the process need read permission on control file */
6128 /* AV: shouldn't we check that it's been opened for read instead? */
6129 ret = inode_permission(file_inode(cfile.file), MAY_READ);
6134 * Determine the event callbacks and set them in @event. This used
6135 * to be done via struct cftype but cgroup core no longer knows
6136 * about these events. The following is crude but the whole thing
6137 * is for compatibility anyway.
6139 * DO NOT ADD NEW FILES.
6141 name = cfile.file->f_dentry->d_name.name;
6143 if (!strcmp(name, "memory.usage_in_bytes")) {
6144 event->register_event = mem_cgroup_usage_register_event;
6145 event->unregister_event = mem_cgroup_usage_unregister_event;
6146 } else if (!strcmp(name, "memory.oom_control")) {
6147 event->register_event = mem_cgroup_oom_register_event;
6148 event->unregister_event = mem_cgroup_oom_unregister_event;
6149 } else if (!strcmp(name, "memory.pressure_level")) {
6150 event->register_event = vmpressure_register_event;
6151 event->unregister_event = vmpressure_unregister_event;
6152 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
6153 event->register_event = memsw_cgroup_usage_register_event;
6154 event->unregister_event = memsw_cgroup_usage_unregister_event;
6161 * Verify @cfile should belong to @css. Also, remaining events are
6162 * automatically removed on cgroup destruction but the removal is
6163 * asynchronous, so take an extra ref on @css.
6168 cfile_css = css_from_dir(cfile.file->f_dentry->d_parent,
6169 &mem_cgroup_subsys);
6170 if (cfile_css == css && css_tryget(css))
6177 ret = event->register_event(memcg, event->eventfd, buffer);
6181 efile.file->f_op->poll(efile.file, &event->pt);
6183 spin_lock(&memcg->event_list_lock);
6184 list_add(&event->list, &memcg->event_list);
6185 spin_unlock(&memcg->event_list_lock);
6197 eventfd_ctx_put(event->eventfd);
6206 static struct cftype mem_cgroup_files[] = {
6208 .name = "usage_in_bytes",
6209 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
6210 .read_u64 = mem_cgroup_read_u64,
6213 .name = "max_usage_in_bytes",
6214 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
6215 .trigger = mem_cgroup_reset,
6216 .read_u64 = mem_cgroup_read_u64,
6219 .name = "limit_in_bytes",
6220 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
6221 .write_string = mem_cgroup_write,
6222 .read_u64 = mem_cgroup_read_u64,
6225 .name = "soft_limit_in_bytes",
6226 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
6227 .write_string = mem_cgroup_write,
6228 .read_u64 = mem_cgroup_read_u64,
6232 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
6233 .trigger = mem_cgroup_reset,
6234 .read_u64 = mem_cgroup_read_u64,
6238 .seq_show = memcg_stat_show,
6241 .name = "force_empty",
6242 .trigger = mem_cgroup_force_empty_write,
6245 .name = "use_hierarchy",
6246 .flags = CFTYPE_INSANE,
6247 .write_u64 = mem_cgroup_hierarchy_write,
6248 .read_u64 = mem_cgroup_hierarchy_read,
6251 .name = "cgroup.event_control", /* XXX: for compat */
6252 .write_string = memcg_write_event_control,
6253 .flags = CFTYPE_NO_PREFIX,
6257 .name = "swappiness",
6258 .read_u64 = mem_cgroup_swappiness_read,
6259 .write_u64 = mem_cgroup_swappiness_write,
6262 .name = "move_charge_at_immigrate",
6263 .read_u64 = mem_cgroup_move_charge_read,
6264 .write_u64 = mem_cgroup_move_charge_write,
6267 .name = "oom_control",
6268 .seq_show = mem_cgroup_oom_control_read,
6269 .write_u64 = mem_cgroup_oom_control_write,
6270 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
6273 .name = "pressure_level",
6277 .name = "numa_stat",
6278 .seq_show = memcg_numa_stat_show,
6281 #ifdef CONFIG_MEMCG_KMEM
6283 .name = "kmem.limit_in_bytes",
6284 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
6285 .write_string = mem_cgroup_write,
6286 .read_u64 = mem_cgroup_read_u64,
6289 .name = "kmem.usage_in_bytes",
6290 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
6291 .read_u64 = mem_cgroup_read_u64,
6294 .name = "kmem.failcnt",
6295 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
6296 .trigger = mem_cgroup_reset,
6297 .read_u64 = mem_cgroup_read_u64,
6300 .name = "kmem.max_usage_in_bytes",
6301 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
6302 .trigger = mem_cgroup_reset,
6303 .read_u64 = mem_cgroup_read_u64,
6305 #ifdef CONFIG_SLABINFO
6307 .name = "kmem.slabinfo",
6308 .seq_show = mem_cgroup_slabinfo_read,
6312 { }, /* terminate */
6315 #ifdef CONFIG_MEMCG_SWAP
6316 static struct cftype memsw_cgroup_files[] = {
6318 .name = "memsw.usage_in_bytes",
6319 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6320 .read_u64 = mem_cgroup_read_u64,
6323 .name = "memsw.max_usage_in_bytes",
6324 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6325 .trigger = mem_cgroup_reset,
6326 .read_u64 = mem_cgroup_read_u64,
6329 .name = "memsw.limit_in_bytes",
6330 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6331 .write_string = mem_cgroup_write,
6332 .read_u64 = mem_cgroup_read_u64,
6335 .name = "memsw.failcnt",
6336 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6337 .trigger = mem_cgroup_reset,
6338 .read_u64 = mem_cgroup_read_u64,
6340 { }, /* terminate */
6343 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6345 struct mem_cgroup_per_node *pn;
6346 struct mem_cgroup_per_zone *mz;
6347 int zone, tmp = node;
6349 * This routine is called against possible nodes.
6350 * But it's BUG to call kmalloc() against offline node.
6352 * TODO: this routine can waste much memory for nodes which will
6353 * never be onlined. It's better to use memory hotplug callback
6356 if (!node_state(node, N_NORMAL_MEMORY))
6358 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
6362 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6363 mz = &pn->zoneinfo[zone];
6364 lruvec_init(&mz->lruvec);
6365 mz->usage_in_excess = 0;
6366 mz->on_tree = false;
6369 memcg->nodeinfo[node] = pn;
6373 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6375 kfree(memcg->nodeinfo[node]);
6378 static struct mem_cgroup *mem_cgroup_alloc(void)
6380 struct mem_cgroup *memcg;
6383 size = sizeof(struct mem_cgroup);
6384 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
6386 memcg = kzalloc(size, GFP_KERNEL);
6390 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
6393 spin_lock_init(&memcg->pcp_counter_lock);
6402 * At destroying mem_cgroup, references from swap_cgroup can remain.
6403 * (scanning all at force_empty is too costly...)
6405 * Instead of clearing all references at force_empty, we remember
6406 * the number of reference from swap_cgroup and free mem_cgroup when
6407 * it goes down to 0.
6409 * Removal of cgroup itself succeeds regardless of refs from swap.
6412 static void __mem_cgroup_free(struct mem_cgroup *memcg)
6416 mem_cgroup_remove_from_trees(memcg);
6419 free_mem_cgroup_per_zone_info(memcg, node);
6421 free_percpu(memcg->stat);
6424 * We need to make sure that (at least for now), the jump label
6425 * destruction code runs outside of the cgroup lock. This is because
6426 * get_online_cpus(), which is called from the static_branch update,
6427 * can't be called inside the cgroup_lock. cpusets are the ones
6428 * enforcing this dependency, so if they ever change, we might as well.
6430 * schedule_work() will guarantee this happens. Be careful if you need
6431 * to move this code around, and make sure it is outside
6434 disarm_static_keys(memcg);
6439 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6441 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
6443 if (!memcg->res.parent)
6445 return mem_cgroup_from_res_counter(memcg->res.parent, res);
6447 EXPORT_SYMBOL(parent_mem_cgroup);
6449 static void __init mem_cgroup_soft_limit_tree_init(void)
6451 struct mem_cgroup_tree_per_node *rtpn;
6452 struct mem_cgroup_tree_per_zone *rtpz;
6453 int tmp, node, zone;
6455 for_each_node(node) {
6457 if (!node_state(node, N_NORMAL_MEMORY))
6459 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
6462 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6464 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6465 rtpz = &rtpn->rb_tree_per_zone[zone];
6466 rtpz->rb_root = RB_ROOT;
6467 spin_lock_init(&rtpz->lock);
6472 static struct cgroup_subsys_state * __ref
6473 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
6475 struct mem_cgroup *memcg;
6476 long error = -ENOMEM;
6479 memcg = mem_cgroup_alloc();
6481 return ERR_PTR(error);
6484 if (alloc_mem_cgroup_per_zone_info(memcg, node))
6488 if (parent_css == NULL) {
6489 root_mem_cgroup = memcg;
6490 res_counter_init(&memcg->res, NULL);
6491 res_counter_init(&memcg->memsw, NULL);
6492 res_counter_init(&memcg->kmem, NULL);
6495 memcg->last_scanned_node = MAX_NUMNODES;
6496 INIT_LIST_HEAD(&memcg->oom_notify);
6497 memcg->move_charge_at_immigrate = 0;
6498 mutex_init(&memcg->thresholds_lock);
6499 spin_lock_init(&memcg->move_lock);
6500 vmpressure_init(&memcg->vmpressure);
6501 INIT_LIST_HEAD(&memcg->event_list);
6502 spin_lock_init(&memcg->event_list_lock);
6507 __mem_cgroup_free(memcg);
6508 return ERR_PTR(error);
6512 mem_cgroup_css_online(struct cgroup_subsys_state *css)
6514 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6515 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(css));
6517 if (css->cgroup->id > MEM_CGROUP_ID_MAX)
6523 mutex_lock(&memcg_create_mutex);
6525 memcg->use_hierarchy = parent->use_hierarchy;
6526 memcg->oom_kill_disable = parent->oom_kill_disable;
6527 memcg->swappiness = mem_cgroup_swappiness(parent);
6529 if (parent->use_hierarchy) {
6530 res_counter_init(&memcg->res, &parent->res);
6531 res_counter_init(&memcg->memsw, &parent->memsw);
6532 res_counter_init(&memcg->kmem, &parent->kmem);
6535 * No need to take a reference to the parent because cgroup
6536 * core guarantees its existence.
6539 res_counter_init(&memcg->res, NULL);
6540 res_counter_init(&memcg->memsw, NULL);
6541 res_counter_init(&memcg->kmem, NULL);
6543 * Deeper hierachy with use_hierarchy == false doesn't make
6544 * much sense so let cgroup subsystem know about this
6545 * unfortunate state in our controller.
6547 if (parent != root_mem_cgroup)
6548 mem_cgroup_subsys.broken_hierarchy = true;
6550 mutex_unlock(&memcg_create_mutex);
6552 return memcg_init_kmem(memcg, &mem_cgroup_subsys);
6556 * Announce all parents that a group from their hierarchy is gone.
6558 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup *memcg)
6560 struct mem_cgroup *parent = memcg;
6562 while ((parent = parent_mem_cgroup(parent)))
6563 mem_cgroup_iter_invalidate(parent);
6566 * if the root memcg is not hierarchical we have to check it
6569 if (!root_mem_cgroup->use_hierarchy)
6570 mem_cgroup_iter_invalidate(root_mem_cgroup);
6573 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
6575 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6576 struct mem_cgroup_event *event, *tmp;
6579 * Unregister events and notify userspace.
6580 * Notify userspace about cgroup removing only after rmdir of cgroup
6581 * directory to avoid race between userspace and kernelspace.
6583 spin_lock(&memcg->event_list_lock);
6584 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
6585 list_del_init(&event->list);
6586 schedule_work(&event->remove);
6588 spin_unlock(&memcg->event_list_lock);
6590 kmem_cgroup_css_offline(memcg);
6592 mem_cgroup_invalidate_reclaim_iterators(memcg);
6593 mem_cgroup_reparent_charges(memcg);
6594 mem_cgroup_destroy_all_caches(memcg);
6595 vmpressure_cleanup(&memcg->vmpressure);
6598 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
6600 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6602 * XXX: css_offline() would be where we should reparent all
6603 * memory to prepare the cgroup for destruction. However,
6604 * memcg does not do css_tryget() and res_counter charging
6605 * under the same RCU lock region, which means that charging
6606 * could race with offlining. Offlining only happens to
6607 * cgroups with no tasks in them but charges can show up
6608 * without any tasks from the swapin path when the target
6609 * memcg is looked up from the swapout record and not from the
6610 * current task as it usually is. A race like this can leak
6611 * charges and put pages with stale cgroup pointers into
6615 * lookup_swap_cgroup_id()
6617 * mem_cgroup_lookup()
6620 * disable css_tryget()
6623 * reparent_charges()
6624 * res_counter_charge()
6627 * pc->mem_cgroup = dead memcg
6630 * The bulk of the charges are still moved in offline_css() to
6631 * avoid pinning a lot of pages in case a long-term reference
6632 * like a swapout record is deferring the css_free() to long
6633 * after offlining. But this makes sure we catch any charges
6634 * made after offlining:
6636 mem_cgroup_reparent_charges(memcg);
6638 memcg_destroy_kmem(memcg);
6639 __mem_cgroup_free(memcg);
6643 /* Handlers for move charge at task migration. */
6644 #define PRECHARGE_COUNT_AT_ONCE 256
6645 static int mem_cgroup_do_precharge(unsigned long count)
6648 int batch_count = PRECHARGE_COUNT_AT_ONCE;
6649 struct mem_cgroup *memcg = mc.to;
6651 if (mem_cgroup_is_root(memcg)) {
6652 mc.precharge += count;
6653 /* we don't need css_get for root */
6656 /* try to charge at once */
6658 struct res_counter *dummy;
6660 * "memcg" cannot be under rmdir() because we've already checked
6661 * by cgroup_lock_live_cgroup() that it is not removed and we
6662 * are still under the same cgroup_mutex. So we can postpone
6665 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
6667 if (do_swap_account && res_counter_charge(&memcg->memsw,
6668 PAGE_SIZE * count, &dummy)) {
6669 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
6672 mc.precharge += count;
6676 /* fall back to one by one charge */
6678 if (signal_pending(current)) {
6682 if (!batch_count--) {
6683 batch_count = PRECHARGE_COUNT_AT_ONCE;
6686 ret = __mem_cgroup_try_charge(NULL,
6687 GFP_KERNEL, 1, &memcg, false);
6689 /* mem_cgroup_clear_mc() will do uncharge later */
6697 * get_mctgt_type - get target type of moving charge
6698 * @vma: the vma the pte to be checked belongs
6699 * @addr: the address corresponding to the pte to be checked
6700 * @ptent: the pte to be checked
6701 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6704 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6705 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6706 * move charge. if @target is not NULL, the page is stored in target->page
6707 * with extra refcnt got(Callers should handle it).
6708 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6709 * target for charge migration. if @target is not NULL, the entry is stored
6712 * Called with pte lock held.
6719 enum mc_target_type {
6725 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
6726 unsigned long addr, pte_t ptent)
6728 struct page *page = vm_normal_page(vma, addr, ptent);
6730 if (!page || !page_mapped(page))
6732 if (PageAnon(page)) {
6733 /* we don't move shared anon */
6736 } else if (!move_file())
6737 /* we ignore mapcount for file pages */
6739 if (!get_page_unless_zero(page))
6746 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6747 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6749 struct page *page = NULL;
6750 swp_entry_t ent = pte_to_swp_entry(ptent);
6752 if (!move_anon() || non_swap_entry(ent))
6755 * Because lookup_swap_cache() updates some statistics counter,
6756 * we call find_get_page() with swapper_space directly.
6758 page = find_get_page(swap_address_space(ent), ent.val);
6759 if (do_swap_account)
6760 entry->val = ent.val;
6765 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6766 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6772 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
6773 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6775 struct page *page = NULL;
6776 struct address_space *mapping;
6779 if (!vma->vm_file) /* anonymous vma */
6784 mapping = vma->vm_file->f_mapping;
6785 if (pte_none(ptent))
6786 pgoff = linear_page_index(vma, addr);
6787 else /* pte_file(ptent) is true */
6788 pgoff = pte_to_pgoff(ptent);
6790 /* page is moved even if it's not RSS of this task(page-faulted). */
6791 page = find_get_page(mapping, pgoff);
6794 /* shmem/tmpfs may report page out on swap: account for that too. */
6795 if (radix_tree_exceptional_entry(page)) {
6796 swp_entry_t swap = radix_to_swp_entry(page);
6797 if (do_swap_account)
6799 page = find_get_page(swap_address_space(swap), swap.val);
6805 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
6806 unsigned long addr, pte_t ptent, union mc_target *target)
6808 struct page *page = NULL;
6809 struct page_cgroup *pc;
6810 enum mc_target_type ret = MC_TARGET_NONE;
6811 swp_entry_t ent = { .val = 0 };
6813 if (pte_present(ptent))
6814 page = mc_handle_present_pte(vma, addr, ptent);
6815 else if (is_swap_pte(ptent))
6816 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
6817 else if (pte_none(ptent) || pte_file(ptent))
6818 page = mc_handle_file_pte(vma, addr, ptent, &ent);
6820 if (!page && !ent.val)
6823 pc = lookup_page_cgroup(page);
6825 * Do only loose check w/o page_cgroup lock.
6826 * mem_cgroup_move_account() checks the pc is valid or not under
6829 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6830 ret = MC_TARGET_PAGE;
6832 target->page = page;
6834 if (!ret || !target)
6837 /* There is a swap entry and a page doesn't exist or isn't charged */
6838 if (ent.val && !ret &&
6839 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
6840 ret = MC_TARGET_SWAP;
6847 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6849 * We don't consider swapping or file mapped pages because THP does not
6850 * support them for now.
6851 * Caller should make sure that pmd_trans_huge(pmd) is true.
6853 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6854 unsigned long addr, pmd_t pmd, union mc_target *target)
6856 struct page *page = NULL;
6857 struct page_cgroup *pc;
6858 enum mc_target_type ret = MC_TARGET_NONE;
6860 page = pmd_page(pmd);
6861 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
6864 pc = lookup_page_cgroup(page);
6865 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6866 ret = MC_TARGET_PAGE;
6869 target->page = page;
6875 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6876 unsigned long addr, pmd_t pmd, union mc_target *target)
6878 return MC_TARGET_NONE;
6882 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
6883 unsigned long addr, unsigned long end,
6884 struct mm_walk *walk)
6886 struct vm_area_struct *vma = walk->private;
6890 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
6891 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
6892 mc.precharge += HPAGE_PMD_NR;
6897 if (pmd_trans_unstable(pmd))
6899 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6900 for (; addr != end; pte++, addr += PAGE_SIZE)
6901 if (get_mctgt_type(vma, addr, *pte, NULL))
6902 mc.precharge++; /* increment precharge temporarily */
6903 pte_unmap_unlock(pte - 1, ptl);
6909 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
6911 unsigned long precharge;
6912 struct vm_area_struct *vma;
6914 down_read(&mm->mmap_sem);
6915 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6916 struct mm_walk mem_cgroup_count_precharge_walk = {
6917 .pmd_entry = mem_cgroup_count_precharge_pte_range,
6921 if (is_vm_hugetlb_page(vma))
6923 walk_page_range(vma->vm_start, vma->vm_end,
6924 &mem_cgroup_count_precharge_walk);
6926 up_read(&mm->mmap_sem);
6928 precharge = mc.precharge;
6934 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
6936 unsigned long precharge = mem_cgroup_count_precharge(mm);
6938 VM_BUG_ON(mc.moving_task);
6939 mc.moving_task = current;
6940 return mem_cgroup_do_precharge(precharge);
6943 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6944 static void __mem_cgroup_clear_mc(void)
6946 struct mem_cgroup *from = mc.from;
6947 struct mem_cgroup *to = mc.to;
6950 /* we must uncharge all the leftover precharges from mc.to */
6952 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
6956 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6957 * we must uncharge here.
6959 if (mc.moved_charge) {
6960 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
6961 mc.moved_charge = 0;
6963 /* we must fixup refcnts and charges */
6964 if (mc.moved_swap) {
6965 /* uncharge swap account from the old cgroup */
6966 if (!mem_cgroup_is_root(mc.from))
6967 res_counter_uncharge(&mc.from->memsw,
6968 PAGE_SIZE * mc.moved_swap);
6970 for (i = 0; i < mc.moved_swap; i++)
6971 css_put(&mc.from->css);
6973 if (!mem_cgroup_is_root(mc.to)) {
6975 * we charged both to->res and to->memsw, so we should
6978 res_counter_uncharge(&mc.to->res,
6979 PAGE_SIZE * mc.moved_swap);
6981 /* we've already done css_get(mc.to) */
6984 memcg_oom_recover(from);
6985 memcg_oom_recover(to);
6986 wake_up_all(&mc.waitq);
6989 static void mem_cgroup_clear_mc(void)
6991 struct mem_cgroup *from = mc.from;
6994 * we must clear moving_task before waking up waiters at the end of
6997 mc.moving_task = NULL;
6998 __mem_cgroup_clear_mc();
6999 spin_lock(&mc.lock);
7002 spin_unlock(&mc.lock);
7003 mem_cgroup_end_move(from);
7006 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
7007 struct cgroup_taskset *tset)
7009 struct task_struct *p = cgroup_taskset_first(tset);
7011 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7012 unsigned long move_charge_at_immigrate;
7015 * We are now commited to this value whatever it is. Changes in this
7016 * tunable will only affect upcoming migrations, not the current one.
7017 * So we need to save it, and keep it going.
7019 move_charge_at_immigrate = memcg->move_charge_at_immigrate;
7020 if (move_charge_at_immigrate) {
7021 struct mm_struct *mm;
7022 struct mem_cgroup *from = mem_cgroup_from_task(p);
7024 VM_BUG_ON(from == memcg);
7026 mm = get_task_mm(p);
7029 /* We move charges only when we move a owner of the mm */
7030 if (mm->owner == p) {
7033 VM_BUG_ON(mc.precharge);
7034 VM_BUG_ON(mc.moved_charge);
7035 VM_BUG_ON(mc.moved_swap);
7036 mem_cgroup_start_move(from);
7037 spin_lock(&mc.lock);
7040 mc.immigrate_flags = move_charge_at_immigrate;
7041 spin_unlock(&mc.lock);
7042 /* We set mc.moving_task later */
7044 ret = mem_cgroup_precharge_mc(mm);
7046 mem_cgroup_clear_mc();
7053 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
7054 struct cgroup_taskset *tset)
7056 mem_cgroup_clear_mc();
7059 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
7060 unsigned long addr, unsigned long end,
7061 struct mm_walk *walk)
7064 struct vm_area_struct *vma = walk->private;
7067 enum mc_target_type target_type;
7068 union mc_target target;
7070 struct page_cgroup *pc;
7073 * We don't take compound_lock() here but no race with splitting thp
7075 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
7076 * under splitting, which means there's no concurrent thp split,
7077 * - if another thread runs into split_huge_page() just after we
7078 * entered this if-block, the thread must wait for page table lock
7079 * to be unlocked in __split_huge_page_splitting(), where the main
7080 * part of thp split is not executed yet.
7082 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
7083 if (mc.precharge < HPAGE_PMD_NR) {
7087 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
7088 if (target_type == MC_TARGET_PAGE) {
7090 if (!isolate_lru_page(page)) {
7091 pc = lookup_page_cgroup(page);
7092 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
7093 pc, mc.from, mc.to)) {
7094 mc.precharge -= HPAGE_PMD_NR;
7095 mc.moved_charge += HPAGE_PMD_NR;
7097 putback_lru_page(page);
7105 if (pmd_trans_unstable(pmd))
7108 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
7109 for (; addr != end; addr += PAGE_SIZE) {
7110 pte_t ptent = *(pte++);
7116 switch (get_mctgt_type(vma, addr, ptent, &target)) {
7117 case MC_TARGET_PAGE:
7119 if (isolate_lru_page(page))
7121 pc = lookup_page_cgroup(page);
7122 if (!mem_cgroup_move_account(page, 1, pc,
7125 /* we uncharge from mc.from later. */
7128 putback_lru_page(page);
7129 put: /* get_mctgt_type() gets the page */
7132 case MC_TARGET_SWAP:
7134 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
7136 /* we fixup refcnts and charges later. */
7144 pte_unmap_unlock(pte - 1, ptl);
7149 * We have consumed all precharges we got in can_attach().
7150 * We try charge one by one, but don't do any additional
7151 * charges to mc.to if we have failed in charge once in attach()
7154 ret = mem_cgroup_do_precharge(1);
7162 static void mem_cgroup_move_charge(struct mm_struct *mm)
7164 struct vm_area_struct *vma;
7166 lru_add_drain_all();
7168 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
7170 * Someone who are holding the mmap_sem might be waiting in
7171 * waitq. So we cancel all extra charges, wake up all waiters,
7172 * and retry. Because we cancel precharges, we might not be able
7173 * to move enough charges, but moving charge is a best-effort
7174 * feature anyway, so it wouldn't be a big problem.
7176 __mem_cgroup_clear_mc();
7180 for (vma = mm->mmap; vma; vma = vma->vm_next) {
7182 struct mm_walk mem_cgroup_move_charge_walk = {
7183 .pmd_entry = mem_cgroup_move_charge_pte_range,
7187 if (is_vm_hugetlb_page(vma))
7189 ret = walk_page_range(vma->vm_start, vma->vm_end,
7190 &mem_cgroup_move_charge_walk);
7193 * means we have consumed all precharges and failed in
7194 * doing additional charge. Just abandon here.
7198 up_read(&mm->mmap_sem);
7201 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
7202 struct cgroup_taskset *tset)
7204 struct task_struct *p = cgroup_taskset_first(tset);
7205 struct mm_struct *mm = get_task_mm(p);
7209 mem_cgroup_move_charge(mm);
7213 mem_cgroup_clear_mc();
7215 #else /* !CONFIG_MMU */
7216 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
7217 struct cgroup_taskset *tset)
7221 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
7222 struct cgroup_taskset *tset)
7225 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
7226 struct cgroup_taskset *tset)
7232 * Cgroup retains root cgroups across [un]mount cycles making it necessary
7233 * to verify sane_behavior flag on each mount attempt.
7235 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
7238 * use_hierarchy is forced with sane_behavior. cgroup core
7239 * guarantees that @root doesn't have any children, so turning it
7240 * on for the root memcg is enough.
7242 if (cgroup_sane_behavior(root_css->cgroup))
7243 mem_cgroup_from_css(root_css)->use_hierarchy = true;
7246 struct cgroup_subsys mem_cgroup_subsys = {
7248 .subsys_id = mem_cgroup_subsys_id,
7249 .css_alloc = mem_cgroup_css_alloc,
7250 .css_online = mem_cgroup_css_online,
7251 .css_offline = mem_cgroup_css_offline,
7252 .css_free = mem_cgroup_css_free,
7253 .can_attach = mem_cgroup_can_attach,
7254 .cancel_attach = mem_cgroup_cancel_attach,
7255 .attach = mem_cgroup_move_task,
7256 .bind = mem_cgroup_bind,
7257 .base_cftypes = mem_cgroup_files,
7261 #ifdef CONFIG_MEMCG_SWAP
7262 static int __init enable_swap_account(char *s)
7264 if (!strcmp(s, "1"))
7265 really_do_swap_account = 1;
7266 else if (!strcmp(s, "0"))
7267 really_do_swap_account = 0;
7270 __setup("swapaccount=", enable_swap_account);
7272 static void __init memsw_file_init(void)
7274 WARN_ON(cgroup_add_cftypes(&mem_cgroup_subsys, memsw_cgroup_files));
7277 static void __init enable_swap_cgroup(void)
7279 if (!mem_cgroup_disabled() && really_do_swap_account) {
7280 do_swap_account = 1;
7286 static void __init enable_swap_cgroup(void)
7292 * subsys_initcall() for memory controller.
7294 * Some parts like hotcpu_notifier() have to be initialized from this context
7295 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
7296 * everything that doesn't depend on a specific mem_cgroup structure should
7297 * be initialized from here.
7299 static int __init mem_cgroup_init(void)
7301 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
7302 enable_swap_cgroup();
7303 mem_cgroup_soft_limit_tree_init();
7307 subsys_initcall(mem_cgroup_init);