1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * Kernel Memory Controller
14 * Copyright (C) 2012 Parallels Inc. and Google Inc.
15 * Authors: Glauber Costa and Suleiman Souhlal
17 * This program is free software; you can redistribute it and/or modify
18 * it under the terms of the GNU General Public License as published by
19 * the Free Software Foundation; either version 2 of the License, or
20 * (at your option) any later version.
22 * This program is distributed in the hope that it will be useful,
23 * but WITHOUT ANY WARRANTY; without even the implied warranty of
24 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
25 * GNU General Public License for more details.
28 #include <linux/res_counter.h>
29 #include <linux/memcontrol.h>
30 #include <linux/cgroup.h>
32 #include <linux/hugetlb.h>
33 #include <linux/pagemap.h>
34 #include <linux/smp.h>
35 #include <linux/page-flags.h>
36 #include <linux/backing-dev.h>
37 #include <linux/bit_spinlock.h>
38 #include <linux/rcupdate.h>
39 #include <linux/limits.h>
40 #include <linux/export.h>
41 #include <linux/mutex.h>
42 #include <linux/rbtree.h>
43 #include <linux/slab.h>
44 #include <linux/swap.h>
45 #include <linux/swapops.h>
46 #include <linux/spinlock.h>
47 #include <linux/eventfd.h>
48 #include <linux/poll.h>
49 #include <linux/sort.h>
51 #include <linux/seq_file.h>
52 #include <linux/vmpressure.h>
53 #include <linux/mm_inline.h>
54 #include <linux/page_cgroup.h>
55 #include <linux/cpu.h>
56 #include <linux/oom.h>
57 #include <linux/lockdep.h>
58 #include <linux/file.h>
62 #include <net/tcp_memcontrol.h>
65 #include <asm/uaccess.h>
67 #include <trace/events/vmscan.h>
69 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
70 EXPORT_SYMBOL(memory_cgrp_subsys);
72 #define MEM_CGROUP_RECLAIM_RETRIES 5
73 static struct mem_cgroup *root_mem_cgroup __read_mostly;
75 #ifdef CONFIG_MEMCG_SWAP
76 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
77 int do_swap_account __read_mostly;
79 /* for remember boot option*/
80 #ifdef CONFIG_MEMCG_SWAP_ENABLED
81 static int really_do_swap_account __initdata = 1;
83 static int really_do_swap_account __initdata = 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, &memory_cgrp_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)
925 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
926 * counted as CACHE even if it's on ANON LRU.
929 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
932 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
935 if (PageTransHuge(page))
936 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
939 /* pagein of a big page is an event. So, ignore page size */
941 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
943 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
944 nr_pages = -nr_pages; /* for event */
947 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
951 mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
953 struct mem_cgroup_per_zone *mz;
955 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
956 return mz->lru_size[lru];
960 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
961 unsigned int lru_mask)
963 struct mem_cgroup_per_zone *mz;
965 unsigned long ret = 0;
967 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
970 if (BIT(lru) & lru_mask)
971 ret += mz->lru_size[lru];
977 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
978 int nid, unsigned int lru_mask)
983 for (zid = 0; zid < MAX_NR_ZONES; zid++)
984 total += mem_cgroup_zone_nr_lru_pages(memcg,
990 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
991 unsigned int lru_mask)
996 for_each_node_state(nid, N_MEMORY)
997 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
1001 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
1002 enum mem_cgroup_events_target target)
1004 unsigned long val, next;
1006 val = __this_cpu_read(memcg->stat->nr_page_events);
1007 next = __this_cpu_read(memcg->stat->targets[target]);
1008 /* from time_after() in jiffies.h */
1009 if ((long)next - (long)val < 0) {
1011 case MEM_CGROUP_TARGET_THRESH:
1012 next = val + THRESHOLDS_EVENTS_TARGET;
1014 case MEM_CGROUP_TARGET_SOFTLIMIT:
1015 next = val + SOFTLIMIT_EVENTS_TARGET;
1017 case MEM_CGROUP_TARGET_NUMAINFO:
1018 next = val + NUMAINFO_EVENTS_TARGET;
1023 __this_cpu_write(memcg->stat->targets[target], next);
1030 * Check events in order.
1033 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
1036 /* threshold event is triggered in finer grain than soft limit */
1037 if (unlikely(mem_cgroup_event_ratelimit(memcg,
1038 MEM_CGROUP_TARGET_THRESH))) {
1040 bool do_numainfo __maybe_unused;
1042 do_softlimit = mem_cgroup_event_ratelimit(memcg,
1043 MEM_CGROUP_TARGET_SOFTLIMIT);
1044 #if MAX_NUMNODES > 1
1045 do_numainfo = mem_cgroup_event_ratelimit(memcg,
1046 MEM_CGROUP_TARGET_NUMAINFO);
1050 mem_cgroup_threshold(memcg);
1051 if (unlikely(do_softlimit))
1052 mem_cgroup_update_tree(memcg, page);
1053 #if MAX_NUMNODES > 1
1054 if (unlikely(do_numainfo))
1055 atomic_inc(&memcg->numainfo_events);
1061 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1064 * mm_update_next_owner() may clear mm->owner to NULL
1065 * if it races with swapoff, page migration, etc.
1066 * So this can be called with p == NULL.
1071 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
1074 static struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
1076 struct mem_cgroup *memcg = NULL;
1080 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1081 if (unlikely(!memcg))
1082 memcg = root_mem_cgroup;
1083 } while (!css_tryget(&memcg->css));
1089 * Returns a next (in a pre-order walk) alive memcg (with elevated css
1090 * ref. count) or NULL if the whole root's subtree has been visited.
1092 * helper function to be used by mem_cgroup_iter
1094 static struct mem_cgroup *__mem_cgroup_iter_next(struct mem_cgroup *root,
1095 struct mem_cgroup *last_visited)
1097 struct cgroup_subsys_state *prev_css, *next_css;
1099 prev_css = last_visited ? &last_visited->css : NULL;
1101 next_css = css_next_descendant_pre(prev_css, &root->css);
1104 * Even if we found a group we have to make sure it is
1105 * alive. css && !memcg means that the groups should be
1106 * skipped and we should continue the tree walk.
1107 * last_visited css is safe to use because it is
1108 * protected by css_get and the tree walk is rcu safe.
1110 * We do not take a reference on the root of the tree walk
1111 * because we might race with the root removal when it would
1112 * be the only node in the iterated hierarchy and mem_cgroup_iter
1113 * would end up in an endless loop because it expects that at
1114 * least one valid node will be returned. Root cannot disappear
1115 * because caller of the iterator should hold it already so
1116 * skipping css reference should be safe.
1119 if ((next_css == &root->css) ||
1120 ((next_css->flags & CSS_ONLINE) && css_tryget(next_css)))
1121 return mem_cgroup_from_css(next_css);
1123 prev_css = next_css;
1130 static void mem_cgroup_iter_invalidate(struct mem_cgroup *root)
1133 * When a group in the hierarchy below root is destroyed, the
1134 * hierarchy iterator can no longer be trusted since it might
1135 * have pointed to the destroyed group. Invalidate it.
1137 atomic_inc(&root->dead_count);
1140 static struct mem_cgroup *
1141 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter *iter,
1142 struct mem_cgroup *root,
1145 struct mem_cgroup *position = NULL;
1147 * A cgroup destruction happens in two stages: offlining and
1148 * release. They are separated by a RCU grace period.
1150 * If the iterator is valid, we may still race with an
1151 * offlining. The RCU lock ensures the object won't be
1152 * released, tryget will fail if we lost the race.
1154 *sequence = atomic_read(&root->dead_count);
1155 if (iter->last_dead_count == *sequence) {
1157 position = iter->last_visited;
1160 * We cannot take a reference to root because we might race
1161 * with root removal and returning NULL would end up in
1162 * an endless loop on the iterator user level when root
1163 * would be returned all the time.
1165 if (position && position != root &&
1166 !css_tryget(&position->css))
1172 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter *iter,
1173 struct mem_cgroup *last_visited,
1174 struct mem_cgroup *new_position,
1175 struct mem_cgroup *root,
1178 /* root reference counting symmetric to mem_cgroup_iter_load */
1179 if (last_visited && last_visited != root)
1180 css_put(&last_visited->css);
1182 * We store the sequence count from the time @last_visited was
1183 * loaded successfully instead of rereading it here so that we
1184 * don't lose destruction events in between. We could have
1185 * raced with the destruction of @new_position after all.
1187 iter->last_visited = new_position;
1189 iter->last_dead_count = sequence;
1193 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1194 * @root: hierarchy root
1195 * @prev: previously returned memcg, NULL on first invocation
1196 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1198 * Returns references to children of the hierarchy below @root, or
1199 * @root itself, or %NULL after a full round-trip.
1201 * Caller must pass the return value in @prev on subsequent
1202 * invocations for reference counting, or use mem_cgroup_iter_break()
1203 * to cancel a hierarchy walk before the round-trip is complete.
1205 * Reclaimers can specify a zone and a priority level in @reclaim to
1206 * divide up the memcgs in the hierarchy among all concurrent
1207 * reclaimers operating on the same zone and priority.
1209 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1210 struct mem_cgroup *prev,
1211 struct mem_cgroup_reclaim_cookie *reclaim)
1213 struct mem_cgroup *memcg = NULL;
1214 struct mem_cgroup *last_visited = NULL;
1216 if (mem_cgroup_disabled())
1220 root = root_mem_cgroup;
1222 if (prev && !reclaim)
1223 last_visited = prev;
1225 if (!root->use_hierarchy && root != root_mem_cgroup) {
1233 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1234 int uninitialized_var(seq);
1237 int nid = zone_to_nid(reclaim->zone);
1238 int zid = zone_idx(reclaim->zone);
1239 struct mem_cgroup_per_zone *mz;
1241 mz = mem_cgroup_zoneinfo(root, nid, zid);
1242 iter = &mz->reclaim_iter[reclaim->priority];
1243 if (prev && reclaim->generation != iter->generation) {
1244 iter->last_visited = NULL;
1248 last_visited = mem_cgroup_iter_load(iter, root, &seq);
1251 memcg = __mem_cgroup_iter_next(root, last_visited);
1254 mem_cgroup_iter_update(iter, last_visited, memcg, root,
1259 else if (!prev && memcg)
1260 reclaim->generation = iter->generation;
1269 if (prev && prev != root)
1270 css_put(&prev->css);
1276 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1277 * @root: hierarchy root
1278 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1280 void mem_cgroup_iter_break(struct mem_cgroup *root,
1281 struct mem_cgroup *prev)
1284 root = root_mem_cgroup;
1285 if (prev && prev != root)
1286 css_put(&prev->css);
1290 * Iteration constructs for visiting all cgroups (under a tree). If
1291 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1292 * be used for reference counting.
1294 #define for_each_mem_cgroup_tree(iter, root) \
1295 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1297 iter = mem_cgroup_iter(root, iter, NULL))
1299 #define for_each_mem_cgroup(iter) \
1300 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1302 iter = mem_cgroup_iter(NULL, iter, NULL))
1304 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1306 struct mem_cgroup *memcg;
1309 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1310 if (unlikely(!memcg))
1315 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1318 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1326 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1329 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1330 * @zone: zone of the wanted lruvec
1331 * @memcg: memcg of the wanted lruvec
1333 * Returns the lru list vector holding pages for the given @zone and
1334 * @mem. This can be the global zone lruvec, if the memory controller
1337 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1338 struct mem_cgroup *memcg)
1340 struct mem_cgroup_per_zone *mz;
1341 struct lruvec *lruvec;
1343 if (mem_cgroup_disabled()) {
1344 lruvec = &zone->lruvec;
1348 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1349 lruvec = &mz->lruvec;
1352 * Since a node can be onlined after the mem_cgroup was created,
1353 * we have to be prepared to initialize lruvec->zone here;
1354 * and if offlined then reonlined, we need to reinitialize it.
1356 if (unlikely(lruvec->zone != zone))
1357 lruvec->zone = zone;
1362 * Following LRU functions are allowed to be used without PCG_LOCK.
1363 * Operations are called by routine of global LRU independently from memcg.
1364 * What we have to take care of here is validness of pc->mem_cgroup.
1366 * Changes to pc->mem_cgroup happens when
1369 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1370 * It is added to LRU before charge.
1371 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1372 * When moving account, the page is not on LRU. It's isolated.
1376 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1378 * @zone: zone of the page
1380 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1382 struct mem_cgroup_per_zone *mz;
1383 struct mem_cgroup *memcg;
1384 struct page_cgroup *pc;
1385 struct lruvec *lruvec;
1387 if (mem_cgroup_disabled()) {
1388 lruvec = &zone->lruvec;
1392 pc = lookup_page_cgroup(page);
1393 memcg = pc->mem_cgroup;
1396 * Surreptitiously switch any uncharged offlist page to root:
1397 * an uncharged page off lru does nothing to secure
1398 * its former mem_cgroup from sudden removal.
1400 * Our caller holds lru_lock, and PageCgroupUsed is updated
1401 * under page_cgroup lock: between them, they make all uses
1402 * of pc->mem_cgroup safe.
1404 if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1405 pc->mem_cgroup = memcg = root_mem_cgroup;
1407 mz = page_cgroup_zoneinfo(memcg, page);
1408 lruvec = &mz->lruvec;
1411 * Since a node can be onlined after the mem_cgroup was created,
1412 * we have to be prepared to initialize lruvec->zone here;
1413 * and if offlined then reonlined, we need to reinitialize it.
1415 if (unlikely(lruvec->zone != zone))
1416 lruvec->zone = zone;
1421 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1422 * @lruvec: mem_cgroup per zone lru vector
1423 * @lru: index of lru list the page is sitting on
1424 * @nr_pages: positive when adding or negative when removing
1426 * This function must be called when a page is added to or removed from an
1429 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1432 struct mem_cgroup_per_zone *mz;
1433 unsigned long *lru_size;
1435 if (mem_cgroup_disabled())
1438 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1439 lru_size = mz->lru_size + lru;
1440 *lru_size += nr_pages;
1441 VM_BUG_ON((long)(*lru_size) < 0);
1445 * Checks whether given mem is same or in the root_mem_cgroup's
1448 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1449 struct mem_cgroup *memcg)
1451 if (root_memcg == memcg)
1453 if (!root_memcg->use_hierarchy || !memcg)
1455 return cgroup_is_descendant(memcg->css.cgroup, root_memcg->css.cgroup);
1458 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1459 struct mem_cgroup *memcg)
1464 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1469 bool task_in_mem_cgroup(struct task_struct *task,
1470 const struct mem_cgroup *memcg)
1472 struct mem_cgroup *curr = NULL;
1473 struct task_struct *p;
1476 p = find_lock_task_mm(task);
1478 curr = get_mem_cgroup_from_mm(p->mm);
1482 * All threads may have already detached their mm's, but the oom
1483 * killer still needs to detect if they have already been oom
1484 * killed to prevent needlessly killing additional tasks.
1487 curr = mem_cgroup_from_task(task);
1489 css_get(&curr->css);
1493 * We should check use_hierarchy of "memcg" not "curr". Because checking
1494 * use_hierarchy of "curr" here make this function true if hierarchy is
1495 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1496 * hierarchy(even if use_hierarchy is disabled in "memcg").
1498 ret = mem_cgroup_same_or_subtree(memcg, curr);
1499 css_put(&curr->css);
1503 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1505 unsigned long inactive_ratio;
1506 unsigned long inactive;
1507 unsigned long active;
1510 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1511 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1513 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1515 inactive_ratio = int_sqrt(10 * gb);
1519 return inactive * inactive_ratio < active;
1522 #define mem_cgroup_from_res_counter(counter, member) \
1523 container_of(counter, struct mem_cgroup, member)
1526 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1527 * @memcg: the memory cgroup
1529 * Returns the maximum amount of memory @mem can be charged with, in
1532 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1534 unsigned long long margin;
1536 margin = res_counter_margin(&memcg->res);
1537 if (do_swap_account)
1538 margin = min(margin, res_counter_margin(&memcg->memsw));
1539 return margin >> PAGE_SHIFT;
1542 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1545 if (!css_parent(&memcg->css))
1546 return vm_swappiness;
1548 return memcg->swappiness;
1552 * memcg->moving_account is used for checking possibility that some thread is
1553 * calling move_account(). When a thread on CPU-A starts moving pages under
1554 * a memcg, other threads should check memcg->moving_account under
1555 * rcu_read_lock(), like this:
1559 * memcg->moving_account+1 if (memcg->mocing_account)
1561 * synchronize_rcu() update something.
1566 /* for quick checking without looking up memcg */
1567 atomic_t memcg_moving __read_mostly;
1569 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1571 atomic_inc(&memcg_moving);
1572 atomic_inc(&memcg->moving_account);
1576 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1579 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1580 * We check NULL in callee rather than caller.
1583 atomic_dec(&memcg_moving);
1584 atomic_dec(&memcg->moving_account);
1589 * 2 routines for checking "mem" is under move_account() or not.
1591 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1592 * is used for avoiding races in accounting. If true,
1593 * pc->mem_cgroup may be overwritten.
1595 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1596 * under hierarchy of moving cgroups. This is for
1597 * waiting at hith-memory prressure caused by "move".
1600 static bool mem_cgroup_stolen(struct mem_cgroup *memcg)
1602 VM_BUG_ON(!rcu_read_lock_held());
1603 return atomic_read(&memcg->moving_account) > 0;
1606 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1608 struct mem_cgroup *from;
1609 struct mem_cgroup *to;
1612 * Unlike task_move routines, we access mc.to, mc.from not under
1613 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1615 spin_lock(&mc.lock);
1621 ret = mem_cgroup_same_or_subtree(memcg, from)
1622 || mem_cgroup_same_or_subtree(memcg, to);
1624 spin_unlock(&mc.lock);
1628 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1630 if (mc.moving_task && current != mc.moving_task) {
1631 if (mem_cgroup_under_move(memcg)) {
1633 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1634 /* moving charge context might have finished. */
1637 finish_wait(&mc.waitq, &wait);
1645 * Take this lock when
1646 * - a code tries to modify page's memcg while it's USED.
1647 * - a code tries to modify page state accounting in a memcg.
1648 * see mem_cgroup_stolen(), too.
1650 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1651 unsigned long *flags)
1653 spin_lock_irqsave(&memcg->move_lock, *flags);
1656 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1657 unsigned long *flags)
1659 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1662 #define K(x) ((x) << (PAGE_SHIFT-10))
1664 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1665 * @memcg: The memory cgroup that went over limit
1666 * @p: Task that is going to be killed
1668 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1671 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1673 /* oom_info_lock ensures that parallel ooms do not interleave */
1674 static DEFINE_MUTEX(oom_info_lock);
1675 struct mem_cgroup *iter;
1681 mutex_lock(&oom_info_lock);
1684 pr_info("Task in ");
1685 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1686 pr_info(" killed as a result of limit of ");
1687 pr_cont_cgroup_path(memcg->css.cgroup);
1692 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1693 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1694 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1695 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1696 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1697 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1698 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1699 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1700 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1701 res_counter_read_u64(&memcg->kmem, RES_USAGE) >> 10,
1702 res_counter_read_u64(&memcg->kmem, RES_LIMIT) >> 10,
1703 res_counter_read_u64(&memcg->kmem, RES_FAILCNT));
1705 for_each_mem_cgroup_tree(iter, memcg) {
1706 pr_info("Memory cgroup stats for ");
1707 pr_cont_cgroup_path(iter->css.cgroup);
1710 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1711 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1713 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1714 K(mem_cgroup_read_stat(iter, i)));
1717 for (i = 0; i < NR_LRU_LISTS; i++)
1718 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1719 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1723 mutex_unlock(&oom_info_lock);
1727 * This function returns the number of memcg under hierarchy tree. Returns
1728 * 1(self count) if no children.
1730 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1733 struct mem_cgroup *iter;
1735 for_each_mem_cgroup_tree(iter, memcg)
1741 * Return the memory (and swap, if configured) limit for a memcg.
1743 static u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1747 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1750 * Do not consider swap space if we cannot swap due to swappiness
1752 if (mem_cgroup_swappiness(memcg)) {
1755 limit += total_swap_pages << PAGE_SHIFT;
1756 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1759 * If memsw is finite and limits the amount of swap space
1760 * available to this memcg, return that limit.
1762 limit = min(limit, memsw);
1768 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1771 struct mem_cgroup *iter;
1772 unsigned long chosen_points = 0;
1773 unsigned long totalpages;
1774 unsigned int points = 0;
1775 struct task_struct *chosen = NULL;
1778 * If current has a pending SIGKILL or is exiting, then automatically
1779 * select it. The goal is to allow it to allocate so that it may
1780 * quickly exit and free its memory.
1782 if (fatal_signal_pending(current) || current->flags & PF_EXITING) {
1783 set_thread_flag(TIF_MEMDIE);
1787 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1788 totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
1789 for_each_mem_cgroup_tree(iter, memcg) {
1790 struct css_task_iter it;
1791 struct task_struct *task;
1793 css_task_iter_start(&iter->css, &it);
1794 while ((task = css_task_iter_next(&it))) {
1795 switch (oom_scan_process_thread(task, totalpages, NULL,
1797 case OOM_SCAN_SELECT:
1799 put_task_struct(chosen);
1801 chosen_points = ULONG_MAX;
1802 get_task_struct(chosen);
1804 case OOM_SCAN_CONTINUE:
1806 case OOM_SCAN_ABORT:
1807 css_task_iter_end(&it);
1808 mem_cgroup_iter_break(memcg, iter);
1810 put_task_struct(chosen);
1815 points = oom_badness(task, memcg, NULL, totalpages);
1816 if (!points || points < chosen_points)
1818 /* Prefer thread group leaders for display purposes */
1819 if (points == chosen_points &&
1820 thread_group_leader(chosen))
1824 put_task_struct(chosen);
1826 chosen_points = points;
1827 get_task_struct(chosen);
1829 css_task_iter_end(&it);
1834 points = chosen_points * 1000 / totalpages;
1835 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1836 NULL, "Memory cgroup out of memory");
1839 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1841 unsigned long flags)
1843 unsigned long total = 0;
1844 bool noswap = false;
1847 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1849 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1852 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1854 drain_all_stock_async(memcg);
1855 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1857 * Allow limit shrinkers, which are triggered directly
1858 * by userspace, to catch signals and stop reclaim
1859 * after minimal progress, regardless of the margin.
1861 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1863 if (mem_cgroup_margin(memcg))
1866 * If nothing was reclaimed after two attempts, there
1867 * may be no reclaimable pages in this hierarchy.
1876 * test_mem_cgroup_node_reclaimable
1877 * @memcg: the target memcg
1878 * @nid: the node ID to be checked.
1879 * @noswap : specify true here if the user wants flle only information.
1881 * This function returns whether the specified memcg contains any
1882 * reclaimable pages on a node. Returns true if there are any reclaimable
1883 * pages in the node.
1885 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1886 int nid, bool noswap)
1888 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1890 if (noswap || !total_swap_pages)
1892 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1897 #if MAX_NUMNODES > 1
1900 * Always updating the nodemask is not very good - even if we have an empty
1901 * list or the wrong list here, we can start from some node and traverse all
1902 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1905 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1909 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1910 * pagein/pageout changes since the last update.
1912 if (!atomic_read(&memcg->numainfo_events))
1914 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1917 /* make a nodemask where this memcg uses memory from */
1918 memcg->scan_nodes = node_states[N_MEMORY];
1920 for_each_node_mask(nid, node_states[N_MEMORY]) {
1922 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1923 node_clear(nid, memcg->scan_nodes);
1926 atomic_set(&memcg->numainfo_events, 0);
1927 atomic_set(&memcg->numainfo_updating, 0);
1931 * Selecting a node where we start reclaim from. Because what we need is just
1932 * reducing usage counter, start from anywhere is O,K. Considering
1933 * memory reclaim from current node, there are pros. and cons.
1935 * Freeing memory from current node means freeing memory from a node which
1936 * we'll use or we've used. So, it may make LRU bad. And if several threads
1937 * hit limits, it will see a contention on a node. But freeing from remote
1938 * node means more costs for memory reclaim because of memory latency.
1940 * Now, we use round-robin. Better algorithm is welcomed.
1942 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1946 mem_cgroup_may_update_nodemask(memcg);
1947 node = memcg->last_scanned_node;
1949 node = next_node(node, memcg->scan_nodes);
1950 if (node == MAX_NUMNODES)
1951 node = first_node(memcg->scan_nodes);
1953 * We call this when we hit limit, not when pages are added to LRU.
1954 * No LRU may hold pages because all pages are UNEVICTABLE or
1955 * memcg is too small and all pages are not on LRU. In that case,
1956 * we use curret node.
1958 if (unlikely(node == MAX_NUMNODES))
1959 node = numa_node_id();
1961 memcg->last_scanned_node = node;
1966 * Check all nodes whether it contains reclaimable pages or not.
1967 * For quick scan, we make use of scan_nodes. This will allow us to skip
1968 * unused nodes. But scan_nodes is lazily updated and may not cotain
1969 * enough new information. We need to do double check.
1971 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1976 * quick check...making use of scan_node.
1977 * We can skip unused nodes.
1979 if (!nodes_empty(memcg->scan_nodes)) {
1980 for (nid = first_node(memcg->scan_nodes);
1982 nid = next_node(nid, memcg->scan_nodes)) {
1984 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1989 * Check rest of nodes.
1991 for_each_node_state(nid, N_MEMORY) {
1992 if (node_isset(nid, memcg->scan_nodes))
1994 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
2001 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
2006 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
2008 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
2012 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
2015 unsigned long *total_scanned)
2017 struct mem_cgroup *victim = NULL;
2020 unsigned long excess;
2021 unsigned long nr_scanned;
2022 struct mem_cgroup_reclaim_cookie reclaim = {
2027 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
2030 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
2035 * If we have not been able to reclaim
2036 * anything, it might because there are
2037 * no reclaimable pages under this hierarchy
2042 * We want to do more targeted reclaim.
2043 * excess >> 2 is not to excessive so as to
2044 * reclaim too much, nor too less that we keep
2045 * coming back to reclaim from this cgroup
2047 if (total >= (excess >> 2) ||
2048 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
2053 if (!mem_cgroup_reclaimable(victim, false))
2055 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
2057 *total_scanned += nr_scanned;
2058 if (!res_counter_soft_limit_excess(&root_memcg->res))
2061 mem_cgroup_iter_break(root_memcg, victim);
2065 #ifdef CONFIG_LOCKDEP
2066 static struct lockdep_map memcg_oom_lock_dep_map = {
2067 .name = "memcg_oom_lock",
2071 static DEFINE_SPINLOCK(memcg_oom_lock);
2074 * Check OOM-Killer is already running under our hierarchy.
2075 * If someone is running, return false.
2077 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
2079 struct mem_cgroup *iter, *failed = NULL;
2081 spin_lock(&memcg_oom_lock);
2083 for_each_mem_cgroup_tree(iter, memcg) {
2084 if (iter->oom_lock) {
2086 * this subtree of our hierarchy is already locked
2087 * so we cannot give a lock.
2090 mem_cgroup_iter_break(memcg, iter);
2093 iter->oom_lock = true;
2098 * OK, we failed to lock the whole subtree so we have
2099 * to clean up what we set up to the failing subtree
2101 for_each_mem_cgroup_tree(iter, memcg) {
2102 if (iter == failed) {
2103 mem_cgroup_iter_break(memcg, iter);
2106 iter->oom_lock = false;
2109 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
2111 spin_unlock(&memcg_oom_lock);
2116 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
2118 struct mem_cgroup *iter;
2120 spin_lock(&memcg_oom_lock);
2121 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
2122 for_each_mem_cgroup_tree(iter, memcg)
2123 iter->oom_lock = false;
2124 spin_unlock(&memcg_oom_lock);
2127 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
2129 struct mem_cgroup *iter;
2131 for_each_mem_cgroup_tree(iter, memcg)
2132 atomic_inc(&iter->under_oom);
2135 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
2137 struct mem_cgroup *iter;
2140 * When a new child is created while the hierarchy is under oom,
2141 * mem_cgroup_oom_lock() may not be called. We have to use
2142 * atomic_add_unless() here.
2144 for_each_mem_cgroup_tree(iter, memcg)
2145 atomic_add_unless(&iter->under_oom, -1, 0);
2148 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
2150 struct oom_wait_info {
2151 struct mem_cgroup *memcg;
2155 static int memcg_oom_wake_function(wait_queue_t *wait,
2156 unsigned mode, int sync, void *arg)
2158 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
2159 struct mem_cgroup *oom_wait_memcg;
2160 struct oom_wait_info *oom_wait_info;
2162 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
2163 oom_wait_memcg = oom_wait_info->memcg;
2166 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2167 * Then we can use css_is_ancestor without taking care of RCU.
2169 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
2170 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
2172 return autoremove_wake_function(wait, mode, sync, arg);
2175 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
2177 atomic_inc(&memcg->oom_wakeups);
2178 /* for filtering, pass "memcg" as argument. */
2179 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
2182 static void memcg_oom_recover(struct mem_cgroup *memcg)
2184 if (memcg && atomic_read(&memcg->under_oom))
2185 memcg_wakeup_oom(memcg);
2188 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
2190 if (!current->memcg_oom.may_oom)
2193 * We are in the middle of the charge context here, so we
2194 * don't want to block when potentially sitting on a callstack
2195 * that holds all kinds of filesystem and mm locks.
2197 * Also, the caller may handle a failed allocation gracefully
2198 * (like optional page cache readahead) and so an OOM killer
2199 * invocation might not even be necessary.
2201 * That's why we don't do anything here except remember the
2202 * OOM context and then deal with it at the end of the page
2203 * fault when the stack is unwound, the locks are released,
2204 * and when we know whether the fault was overall successful.
2206 css_get(&memcg->css);
2207 current->memcg_oom.memcg = memcg;
2208 current->memcg_oom.gfp_mask = mask;
2209 current->memcg_oom.order = order;
2213 * mem_cgroup_oom_synchronize - complete memcg OOM handling
2214 * @handle: actually kill/wait or just clean up the OOM state
2216 * This has to be called at the end of a page fault if the memcg OOM
2217 * handler was enabled.
2219 * Memcg supports userspace OOM handling where failed allocations must
2220 * sleep on a waitqueue until the userspace task resolves the
2221 * situation. Sleeping directly in the charge context with all kinds
2222 * of locks held is not a good idea, instead we remember an OOM state
2223 * in the task and mem_cgroup_oom_synchronize() has to be called at
2224 * the end of the page fault to complete the OOM handling.
2226 * Returns %true if an ongoing memcg OOM situation was detected and
2227 * completed, %false otherwise.
2229 bool mem_cgroup_oom_synchronize(bool handle)
2231 struct mem_cgroup *memcg = current->memcg_oom.memcg;
2232 struct oom_wait_info owait;
2235 /* OOM is global, do not handle */
2242 owait.memcg = memcg;
2243 owait.wait.flags = 0;
2244 owait.wait.func = memcg_oom_wake_function;
2245 owait.wait.private = current;
2246 INIT_LIST_HEAD(&owait.wait.task_list);
2248 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2249 mem_cgroup_mark_under_oom(memcg);
2251 locked = mem_cgroup_oom_trylock(memcg);
2254 mem_cgroup_oom_notify(memcg);
2256 if (locked && !memcg->oom_kill_disable) {
2257 mem_cgroup_unmark_under_oom(memcg);
2258 finish_wait(&memcg_oom_waitq, &owait.wait);
2259 mem_cgroup_out_of_memory(memcg, current->memcg_oom.gfp_mask,
2260 current->memcg_oom.order);
2263 mem_cgroup_unmark_under_oom(memcg);
2264 finish_wait(&memcg_oom_waitq, &owait.wait);
2268 mem_cgroup_oom_unlock(memcg);
2270 * There is no guarantee that an OOM-lock contender
2271 * sees the wakeups triggered by the OOM kill
2272 * uncharges. Wake any sleepers explicitely.
2274 memcg_oom_recover(memcg);
2277 current->memcg_oom.memcg = NULL;
2278 css_put(&memcg->css);
2283 * Currently used to update mapped file statistics, but the routine can be
2284 * generalized to update other statistics as well.
2286 * Notes: Race condition
2288 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2289 * it tends to be costly. But considering some conditions, we doesn't need
2290 * to do so _always_.
2292 * Considering "charge", lock_page_cgroup() is not required because all
2293 * file-stat operations happen after a page is attached to radix-tree. There
2294 * are no race with "charge".
2296 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2297 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2298 * if there are race with "uncharge". Statistics itself is properly handled
2301 * Considering "move", this is an only case we see a race. To make the race
2302 * small, we check mm->moving_account and detect there are possibility of race
2303 * If there is, we take a lock.
2306 void __mem_cgroup_begin_update_page_stat(struct page *page,
2307 bool *locked, unsigned long *flags)
2309 struct mem_cgroup *memcg;
2310 struct page_cgroup *pc;
2312 pc = lookup_page_cgroup(page);
2314 memcg = pc->mem_cgroup;
2315 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2318 * If this memory cgroup is not under account moving, we don't
2319 * need to take move_lock_mem_cgroup(). Because we already hold
2320 * rcu_read_lock(), any calls to move_account will be delayed until
2321 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2323 if (!mem_cgroup_stolen(memcg))
2326 move_lock_mem_cgroup(memcg, flags);
2327 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2328 move_unlock_mem_cgroup(memcg, flags);
2334 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
2336 struct page_cgroup *pc = lookup_page_cgroup(page);
2339 * It's guaranteed that pc->mem_cgroup never changes while
2340 * lock is held because a routine modifies pc->mem_cgroup
2341 * should take move_lock_mem_cgroup().
2343 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
2346 void mem_cgroup_update_page_stat(struct page *page,
2347 enum mem_cgroup_stat_index idx, int val)
2349 struct mem_cgroup *memcg;
2350 struct page_cgroup *pc = lookup_page_cgroup(page);
2351 unsigned long uninitialized_var(flags);
2353 if (mem_cgroup_disabled())
2356 VM_BUG_ON(!rcu_read_lock_held());
2357 memcg = pc->mem_cgroup;
2358 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2361 this_cpu_add(memcg->stat->count[idx], val);
2365 * size of first charge trial. "32" comes from vmscan.c's magic value.
2366 * TODO: maybe necessary to use big numbers in big irons.
2368 #define CHARGE_BATCH 32U
2369 struct memcg_stock_pcp {
2370 struct mem_cgroup *cached; /* this never be root cgroup */
2371 unsigned int nr_pages;
2372 struct work_struct work;
2373 unsigned long flags;
2374 #define FLUSHING_CACHED_CHARGE 0
2376 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2377 static DEFINE_MUTEX(percpu_charge_mutex);
2380 * consume_stock: Try to consume stocked charge on this cpu.
2381 * @memcg: memcg to consume from.
2382 * @nr_pages: how many pages to charge.
2384 * The charges will only happen if @memcg matches the current cpu's memcg
2385 * stock, and at least @nr_pages are available in that stock. Failure to
2386 * service an allocation will refill the stock.
2388 * returns true if successful, false otherwise.
2390 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2392 struct memcg_stock_pcp *stock;
2395 if (nr_pages > CHARGE_BATCH)
2398 stock = &get_cpu_var(memcg_stock);
2399 if (memcg == stock->cached && stock->nr_pages >= nr_pages)
2400 stock->nr_pages -= nr_pages;
2401 else /* need to call res_counter_charge */
2403 put_cpu_var(memcg_stock);
2408 * Returns stocks cached in percpu to res_counter and reset cached information.
2410 static void drain_stock(struct memcg_stock_pcp *stock)
2412 struct mem_cgroup *old = stock->cached;
2414 if (stock->nr_pages) {
2415 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2417 res_counter_uncharge(&old->res, bytes);
2418 if (do_swap_account)
2419 res_counter_uncharge(&old->memsw, bytes);
2420 stock->nr_pages = 0;
2422 stock->cached = NULL;
2426 * This must be called under preempt disabled or must be called by
2427 * a thread which is pinned to local cpu.
2429 static void drain_local_stock(struct work_struct *dummy)
2431 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2433 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2436 static void __init memcg_stock_init(void)
2440 for_each_possible_cpu(cpu) {
2441 struct memcg_stock_pcp *stock =
2442 &per_cpu(memcg_stock, cpu);
2443 INIT_WORK(&stock->work, drain_local_stock);
2448 * Cache charges(val) which is from res_counter, to local per_cpu area.
2449 * This will be consumed by consume_stock() function, later.
2451 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2453 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2455 if (stock->cached != memcg) { /* reset if necessary */
2457 stock->cached = memcg;
2459 stock->nr_pages += nr_pages;
2460 put_cpu_var(memcg_stock);
2464 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2465 * of the hierarchy under it. sync flag says whether we should block
2466 * until the work is done.
2468 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2472 /* Notify other cpus that system-wide "drain" is running */
2475 for_each_online_cpu(cpu) {
2476 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2477 struct mem_cgroup *memcg;
2479 memcg = stock->cached;
2480 if (!memcg || !stock->nr_pages)
2482 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2484 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2486 drain_local_stock(&stock->work);
2488 schedule_work_on(cpu, &stock->work);
2496 for_each_online_cpu(cpu) {
2497 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2498 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2499 flush_work(&stock->work);
2506 * Tries to drain stocked charges in other cpus. This function is asynchronous
2507 * and just put a work per cpu for draining localy on each cpu. Caller can
2508 * expects some charges will be back to res_counter later but cannot wait for
2511 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2514 * If someone calls draining, avoid adding more kworker runs.
2516 if (!mutex_trylock(&percpu_charge_mutex))
2518 drain_all_stock(root_memcg, false);
2519 mutex_unlock(&percpu_charge_mutex);
2522 /* This is a synchronous drain interface. */
2523 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2525 /* called when force_empty is called */
2526 mutex_lock(&percpu_charge_mutex);
2527 drain_all_stock(root_memcg, true);
2528 mutex_unlock(&percpu_charge_mutex);
2532 * This function drains percpu counter value from DEAD cpu and
2533 * move it to local cpu. Note that this function can be preempted.
2535 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2539 spin_lock(&memcg->pcp_counter_lock);
2540 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2541 long x = per_cpu(memcg->stat->count[i], cpu);
2543 per_cpu(memcg->stat->count[i], cpu) = 0;
2544 memcg->nocpu_base.count[i] += x;
2546 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2547 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2549 per_cpu(memcg->stat->events[i], cpu) = 0;
2550 memcg->nocpu_base.events[i] += x;
2552 spin_unlock(&memcg->pcp_counter_lock);
2555 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
2556 unsigned long action,
2559 int cpu = (unsigned long)hcpu;
2560 struct memcg_stock_pcp *stock;
2561 struct mem_cgroup *iter;
2563 if (action == CPU_ONLINE)
2566 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2569 for_each_mem_cgroup(iter)
2570 mem_cgroup_drain_pcp_counter(iter, cpu);
2572 stock = &per_cpu(memcg_stock, cpu);
2578 /* See mem_cgroup_try_charge() for details */
2580 CHARGE_OK, /* success */
2581 CHARGE_RETRY, /* need to retry but retry is not bad */
2582 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2583 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2586 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2587 unsigned int nr_pages, unsigned int min_pages,
2590 unsigned long csize = nr_pages * PAGE_SIZE;
2591 struct mem_cgroup *mem_over_limit;
2592 struct res_counter *fail_res;
2593 unsigned long flags = 0;
2596 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2599 if (!do_swap_account)
2601 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2605 res_counter_uncharge(&memcg->res, csize);
2606 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2607 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2609 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2611 * Never reclaim on behalf of optional batching, retry with a
2612 * single page instead.
2614 if (nr_pages > min_pages)
2615 return CHARGE_RETRY;
2617 if (!(gfp_mask & __GFP_WAIT))
2618 return CHARGE_WOULDBLOCK;
2620 if (gfp_mask & __GFP_NORETRY)
2621 return CHARGE_NOMEM;
2623 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2624 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2625 return CHARGE_RETRY;
2627 * Even though the limit is exceeded at this point, reclaim
2628 * may have been able to free some pages. Retry the charge
2629 * before killing the task.
2631 * Only for regular pages, though: huge pages are rather
2632 * unlikely to succeed so close to the limit, and we fall back
2633 * to regular pages anyway in case of failure.
2635 if (nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER) && ret)
2636 return CHARGE_RETRY;
2639 * At task move, charge accounts can be doubly counted. So, it's
2640 * better to wait until the end of task_move if something is going on.
2642 if (mem_cgroup_wait_acct_move(mem_over_limit))
2643 return CHARGE_RETRY;
2646 mem_cgroup_oom(mem_over_limit, gfp_mask, get_order(csize));
2648 return CHARGE_NOMEM;
2652 * mem_cgroup_try_charge - try charging a memcg
2653 * @memcg: memcg to charge
2654 * @nr_pages: number of pages to charge
2655 * @oom: trigger OOM if reclaim fails
2657 * Returns 0 if @memcg was charged successfully, -EINTR if the charge
2658 * was bypassed to root_mem_cgroup, and -ENOMEM if the charge failed.
2660 static int mem_cgroup_try_charge(struct mem_cgroup *memcg,
2662 unsigned int nr_pages,
2665 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2666 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2669 if (mem_cgroup_is_root(memcg))
2672 * Unlike in global OOM situations, memcg is not in a physical
2673 * memory shortage. Allow dying and OOM-killed tasks to
2674 * bypass the last charges so that they can exit quickly and
2675 * free their memory.
2677 if (unlikely(test_thread_flag(TIF_MEMDIE) ||
2678 fatal_signal_pending(current)))
2681 if (unlikely(task_in_memcg_oom(current)))
2684 if (gfp_mask & __GFP_NOFAIL)
2687 if (consume_stock(memcg, nr_pages))
2691 bool invoke_oom = oom && !nr_oom_retries;
2693 /* If killed, bypass charge */
2694 if (fatal_signal_pending(current))
2697 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch,
2698 nr_pages, invoke_oom);
2702 case CHARGE_RETRY: /* not in OOM situation but retry */
2705 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2707 case CHARGE_NOMEM: /* OOM routine works */
2708 if (!oom || invoke_oom)
2713 } while (ret != CHARGE_OK);
2715 if (batch > nr_pages)
2716 refill_stock(memcg, batch - nr_pages);
2720 if (!(gfp_mask & __GFP_NOFAIL))
2727 * mem_cgroup_try_charge_mm - try charging a mm
2728 * @mm: mm_struct to charge
2729 * @nr_pages: number of pages to charge
2730 * @oom: trigger OOM if reclaim fails
2732 * Returns the charged mem_cgroup associated with the given mm_struct or
2733 * NULL the charge failed.
2735 static struct mem_cgroup *mem_cgroup_try_charge_mm(struct mm_struct *mm,
2737 unsigned int nr_pages,
2741 struct mem_cgroup *memcg;
2744 memcg = get_mem_cgroup_from_mm(mm);
2745 ret = mem_cgroup_try_charge(memcg, gfp_mask, nr_pages, oom);
2746 css_put(&memcg->css);
2748 memcg = root_mem_cgroup;
2756 * Somemtimes we have to undo a charge we got by try_charge().
2757 * This function is for that and do uncharge, put css's refcnt.
2758 * gotten by try_charge().
2760 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2761 unsigned int nr_pages)
2763 if (!mem_cgroup_is_root(memcg)) {
2764 unsigned long bytes = nr_pages * PAGE_SIZE;
2766 res_counter_uncharge(&memcg->res, bytes);
2767 if (do_swap_account)
2768 res_counter_uncharge(&memcg->memsw, bytes);
2773 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2774 * This is useful when moving usage to parent cgroup.
2776 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2777 unsigned int nr_pages)
2779 unsigned long bytes = nr_pages * PAGE_SIZE;
2781 if (mem_cgroup_is_root(memcg))
2784 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2785 if (do_swap_account)
2786 res_counter_uncharge_until(&memcg->memsw,
2787 memcg->memsw.parent, bytes);
2791 * A helper function to get mem_cgroup from ID. must be called under
2792 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2793 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2794 * called against removed memcg.)
2796 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2798 /* ID 0 is unused ID */
2801 return mem_cgroup_from_id(id);
2804 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2806 struct mem_cgroup *memcg = NULL;
2807 struct page_cgroup *pc;
2811 VM_BUG_ON_PAGE(!PageLocked(page), page);
2813 pc = lookup_page_cgroup(page);
2814 lock_page_cgroup(pc);
2815 if (PageCgroupUsed(pc)) {
2816 memcg = pc->mem_cgroup;
2817 if (memcg && !css_tryget(&memcg->css))
2819 } else if (PageSwapCache(page)) {
2820 ent.val = page_private(page);
2821 id = lookup_swap_cgroup_id(ent);
2823 memcg = mem_cgroup_lookup(id);
2824 if (memcg && !css_tryget(&memcg->css))
2828 unlock_page_cgroup(pc);
2832 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2834 unsigned int nr_pages,
2835 enum charge_type ctype,
2838 struct page_cgroup *pc = lookup_page_cgroup(page);
2839 struct zone *uninitialized_var(zone);
2840 struct lruvec *lruvec;
2841 bool was_on_lru = false;
2844 lock_page_cgroup(pc);
2845 VM_BUG_ON_PAGE(PageCgroupUsed(pc), page);
2847 * we don't need page_cgroup_lock about tail pages, becase they are not
2848 * accessed by any other context at this point.
2852 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2853 * may already be on some other mem_cgroup's LRU. Take care of it.
2856 zone = page_zone(page);
2857 spin_lock_irq(&zone->lru_lock);
2858 if (PageLRU(page)) {
2859 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2861 del_page_from_lru_list(page, lruvec, page_lru(page));
2866 pc->mem_cgroup = memcg;
2868 * We access a page_cgroup asynchronously without lock_page_cgroup().
2869 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2870 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2871 * before USED bit, we need memory barrier here.
2872 * See mem_cgroup_add_lru_list(), etc.
2875 SetPageCgroupUsed(pc);
2879 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2880 VM_BUG_ON_PAGE(PageLRU(page), page);
2882 add_page_to_lru_list(page, lruvec, page_lru(page));
2884 spin_unlock_irq(&zone->lru_lock);
2887 if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
2892 mem_cgroup_charge_statistics(memcg, page, anon, nr_pages);
2893 unlock_page_cgroup(pc);
2896 * "charge_statistics" updated event counter. Then, check it.
2897 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2898 * if they exceeds softlimit.
2900 memcg_check_events(memcg, page);
2903 static DEFINE_MUTEX(set_limit_mutex);
2905 #ifdef CONFIG_MEMCG_KMEM
2906 static DEFINE_MUTEX(activate_kmem_mutex);
2908 static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg)
2910 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) &&
2911 memcg_kmem_is_active(memcg);
2915 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2916 * in the memcg_cache_params struct.
2918 static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
2920 struct kmem_cache *cachep;
2922 VM_BUG_ON(p->is_root_cache);
2923 cachep = p->root_cache;
2924 return cache_from_memcg_idx(cachep, memcg_cache_id(p->memcg));
2927 #ifdef CONFIG_SLABINFO
2928 static int mem_cgroup_slabinfo_read(struct seq_file *m, void *v)
2930 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
2931 struct memcg_cache_params *params;
2933 if (!memcg_can_account_kmem(memcg))
2936 print_slabinfo_header(m);
2938 mutex_lock(&memcg->slab_caches_mutex);
2939 list_for_each_entry(params, &memcg->memcg_slab_caches, list)
2940 cache_show(memcg_params_to_cache(params), m);
2941 mutex_unlock(&memcg->slab_caches_mutex);
2947 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size)
2949 struct res_counter *fail_res;
2952 ret = res_counter_charge(&memcg->kmem, size, &fail_res);
2956 ret = mem_cgroup_try_charge(memcg, gfp, size >> PAGE_SHIFT,
2957 oom_gfp_allowed(gfp));
2958 if (ret == -EINTR) {
2960 * mem_cgroup_try_charge() chosed to bypass to root due to
2961 * OOM kill or fatal signal. Since our only options are to
2962 * either fail the allocation or charge it to this cgroup, do
2963 * it as a temporary condition. But we can't fail. From a
2964 * kmem/slab perspective, the cache has already been selected,
2965 * by mem_cgroup_kmem_get_cache(), so it is too late to change
2968 * This condition will only trigger if the task entered
2969 * memcg_charge_kmem in a sane state, but was OOM-killed during
2970 * mem_cgroup_try_charge() above. Tasks that were already
2971 * dying when the allocation triggers should have been already
2972 * directed to the root cgroup in memcontrol.h
2974 res_counter_charge_nofail(&memcg->res, size, &fail_res);
2975 if (do_swap_account)
2976 res_counter_charge_nofail(&memcg->memsw, size,
2980 res_counter_uncharge(&memcg->kmem, size);
2985 static void memcg_uncharge_kmem(struct mem_cgroup *memcg, u64 size)
2987 res_counter_uncharge(&memcg->res, size);
2988 if (do_swap_account)
2989 res_counter_uncharge(&memcg->memsw, size);
2992 if (res_counter_uncharge(&memcg->kmem, size))
2996 * Releases a reference taken in kmem_cgroup_css_offline in case
2997 * this last uncharge is racing with the offlining code or it is
2998 * outliving the memcg existence.
3000 * The memory barrier imposed by test&clear is paired with the
3001 * explicit one in memcg_kmem_mark_dead().
3003 if (memcg_kmem_test_and_clear_dead(memcg))
3004 css_put(&memcg->css);
3008 * helper for acessing a memcg's index. It will be used as an index in the
3009 * child cache array in kmem_cache, and also to derive its name. This function
3010 * will return -1 when this is not a kmem-limited memcg.
3012 int memcg_cache_id(struct mem_cgroup *memcg)
3014 return memcg ? memcg->kmemcg_id : -1;
3017 static size_t memcg_caches_array_size(int num_groups)
3020 if (num_groups <= 0)
3023 size = 2 * num_groups;
3024 if (size < MEMCG_CACHES_MIN_SIZE)
3025 size = MEMCG_CACHES_MIN_SIZE;
3026 else if (size > MEMCG_CACHES_MAX_SIZE)
3027 size = MEMCG_CACHES_MAX_SIZE;
3033 * We should update the current array size iff all caches updates succeed. This
3034 * can only be done from the slab side. The slab mutex needs to be held when
3037 void memcg_update_array_size(int num)
3039 if (num > memcg_limited_groups_array_size)
3040 memcg_limited_groups_array_size = memcg_caches_array_size(num);
3043 static void kmem_cache_destroy_work_func(struct work_struct *w);
3045 int memcg_update_cache_size(struct kmem_cache *s, int num_groups)
3047 struct memcg_cache_params *cur_params = s->memcg_params;
3049 VM_BUG_ON(!is_root_cache(s));
3051 if (num_groups > memcg_limited_groups_array_size) {
3053 struct memcg_cache_params *new_params;
3054 ssize_t size = memcg_caches_array_size(num_groups);
3056 size *= sizeof(void *);
3057 size += offsetof(struct memcg_cache_params, memcg_caches);
3059 new_params = kzalloc(size, GFP_KERNEL);
3063 new_params->is_root_cache = true;
3066 * There is the chance it will be bigger than
3067 * memcg_limited_groups_array_size, if we failed an allocation
3068 * in a cache, in which case all caches updated before it, will
3069 * have a bigger array.
3071 * But if that is the case, the data after
3072 * memcg_limited_groups_array_size is certainly unused
3074 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3075 if (!cur_params->memcg_caches[i])
3077 new_params->memcg_caches[i] =
3078 cur_params->memcg_caches[i];
3082 * Ideally, we would wait until all caches succeed, and only
3083 * then free the old one. But this is not worth the extra
3084 * pointer per-cache we'd have to have for this.
3086 * It is not a big deal if some caches are left with a size
3087 * bigger than the others. And all updates will reset this
3090 rcu_assign_pointer(s->memcg_params, new_params);
3092 kfree_rcu(cur_params, rcu_head);
3097 int memcg_alloc_cache_params(struct mem_cgroup *memcg, struct kmem_cache *s,
3098 struct kmem_cache *root_cache)
3102 if (!memcg_kmem_enabled())
3106 size = offsetof(struct memcg_cache_params, memcg_caches);
3107 size += memcg_limited_groups_array_size * sizeof(void *);
3109 size = sizeof(struct memcg_cache_params);
3111 s->memcg_params = kzalloc(size, GFP_KERNEL);
3112 if (!s->memcg_params)
3116 s->memcg_params->memcg = memcg;
3117 s->memcg_params->root_cache = root_cache;
3118 INIT_WORK(&s->memcg_params->destroy,
3119 kmem_cache_destroy_work_func);
3121 s->memcg_params->is_root_cache = true;
3126 void memcg_free_cache_params(struct kmem_cache *s)
3128 kfree(s->memcg_params);
3131 void memcg_register_cache(struct kmem_cache *s)
3133 struct kmem_cache *root;
3134 struct mem_cgroup *memcg;
3137 if (is_root_cache(s))
3141 * Holding the slab_mutex assures nobody will touch the memcg_caches
3142 * array while we are modifying it.
3144 lockdep_assert_held(&slab_mutex);
3146 root = s->memcg_params->root_cache;
3147 memcg = s->memcg_params->memcg;
3148 id = memcg_cache_id(memcg);
3150 css_get(&memcg->css);
3154 * Since readers won't lock (see cache_from_memcg_idx()), we need a
3155 * barrier here to ensure nobody will see the kmem_cache partially
3161 * Initialize the pointer to this cache in its parent's memcg_params
3162 * before adding it to the memcg_slab_caches list, otherwise we can
3163 * fail to convert memcg_params_to_cache() while traversing the list.
3165 VM_BUG_ON(root->memcg_params->memcg_caches[id]);
3166 root->memcg_params->memcg_caches[id] = s;
3168 mutex_lock(&memcg->slab_caches_mutex);
3169 list_add(&s->memcg_params->list, &memcg->memcg_slab_caches);
3170 mutex_unlock(&memcg->slab_caches_mutex);
3173 void memcg_unregister_cache(struct kmem_cache *s)
3175 struct kmem_cache *root;
3176 struct mem_cgroup *memcg;
3179 if (is_root_cache(s))
3183 * Holding the slab_mutex assures nobody will touch the memcg_caches
3184 * array while we are modifying it.
3186 lockdep_assert_held(&slab_mutex);
3188 root = s->memcg_params->root_cache;
3189 memcg = s->memcg_params->memcg;
3190 id = memcg_cache_id(memcg);
3192 mutex_lock(&memcg->slab_caches_mutex);
3193 list_del(&s->memcg_params->list);
3194 mutex_unlock(&memcg->slab_caches_mutex);
3197 * Clear the pointer to this cache in its parent's memcg_params only
3198 * after removing it from the memcg_slab_caches list, otherwise we can
3199 * fail to convert memcg_params_to_cache() while traversing the list.
3201 VM_BUG_ON(!root->memcg_params->memcg_caches[id]);
3202 root->memcg_params->memcg_caches[id] = NULL;
3204 css_put(&memcg->css);
3208 * During the creation a new cache, we need to disable our accounting mechanism
3209 * altogether. This is true even if we are not creating, but rather just
3210 * enqueing new caches to be created.
3212 * This is because that process will trigger allocations; some visible, like
3213 * explicit kmallocs to auxiliary data structures, name strings and internal
3214 * cache structures; some well concealed, like INIT_WORK() that can allocate
3215 * objects during debug.
3217 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3218 * to it. This may not be a bounded recursion: since the first cache creation
3219 * failed to complete (waiting on the allocation), we'll just try to create the
3220 * cache again, failing at the same point.
3222 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3223 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3224 * inside the following two functions.
3226 static inline void memcg_stop_kmem_account(void)
3228 VM_BUG_ON(!current->mm);
3229 current->memcg_kmem_skip_account++;
3232 static inline void memcg_resume_kmem_account(void)
3234 VM_BUG_ON(!current->mm);
3235 current->memcg_kmem_skip_account--;
3238 static void kmem_cache_destroy_work_func(struct work_struct *w)
3240 struct kmem_cache *cachep;
3241 struct memcg_cache_params *p;
3243 p = container_of(w, struct memcg_cache_params, destroy);
3245 cachep = memcg_params_to_cache(p);
3248 * If we get down to 0 after shrink, we could delete right away.
3249 * However, memcg_release_pages() already puts us back in the workqueue
3250 * in that case. If we proceed deleting, we'll get a dangling
3251 * reference, and removing the object from the workqueue in that case
3252 * is unnecessary complication. We are not a fast path.
3254 * Note that this case is fundamentally different from racing with
3255 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3256 * kmem_cache_shrink, not only we would be reinserting a dead cache
3257 * into the queue, but doing so from inside the worker racing to
3260 * So if we aren't down to zero, we'll just schedule a worker and try
3263 if (atomic_read(&cachep->memcg_params->nr_pages) != 0)
3264 kmem_cache_shrink(cachep);
3266 kmem_cache_destroy(cachep);
3269 void mem_cgroup_destroy_cache(struct kmem_cache *cachep)
3271 if (!cachep->memcg_params->dead)
3275 * There are many ways in which we can get here.
3277 * We can get to a memory-pressure situation while the delayed work is
3278 * still pending to run. The vmscan shrinkers can then release all
3279 * cache memory and get us to destruction. If this is the case, we'll
3280 * be executed twice, which is a bug (the second time will execute over
3281 * bogus data). In this case, cancelling the work should be fine.
3283 * But we can also get here from the worker itself, if
3284 * kmem_cache_shrink is enough to shake all the remaining objects and
3285 * get the page count to 0. In this case, we'll deadlock if we try to
3286 * cancel the work (the worker runs with an internal lock held, which
3287 * is the same lock we would hold for cancel_work_sync().)
3289 * Since we can't possibly know who got us here, just refrain from
3290 * running if there is already work pending
3292 if (work_pending(&cachep->memcg_params->destroy))
3295 * We have to defer the actual destroying to a workqueue, because
3296 * we might currently be in a context that cannot sleep.
3298 schedule_work(&cachep->memcg_params->destroy);
3301 static struct kmem_cache *memcg_create_kmem_cache(struct mem_cgroup *memcg,
3302 struct kmem_cache *s)
3304 struct kmem_cache *new = NULL;
3305 static char *tmp_path = NULL, *tmp_name = NULL;
3306 static DEFINE_MUTEX(mutex); /* protects tmp_name */
3308 BUG_ON(!memcg_can_account_kmem(memcg));
3312 * kmem_cache_create_memcg duplicates the given name and
3313 * cgroup_name for this name requires RCU context.
3314 * This static temporary buffer is used to prevent from
3315 * pointless shortliving allocation.
3317 if (!tmp_path || !tmp_name) {
3319 tmp_path = kmalloc(PATH_MAX, GFP_KERNEL);
3321 tmp_name = kmalloc(NAME_MAX + 1, GFP_KERNEL);
3322 if (!tmp_path || !tmp_name)
3326 cgroup_name(memcg->css.cgroup, tmp_name, NAME_MAX + 1);
3327 snprintf(tmp_path, PATH_MAX, "%s(%d:%s)", s->name,
3328 memcg_cache_id(memcg), tmp_name);
3330 new = kmem_cache_create_memcg(memcg, tmp_path, s->object_size, s->align,
3331 (s->flags & ~SLAB_PANIC), s->ctor, s);
3333 new->allocflags |= __GFP_KMEMCG;
3337 mutex_unlock(&mutex);
3341 void kmem_cache_destroy_memcg_children(struct kmem_cache *s)
3343 struct kmem_cache *c;
3346 if (!s->memcg_params)
3348 if (!s->memcg_params->is_root_cache)
3352 * If the cache is being destroyed, we trust that there is no one else
3353 * requesting objects from it. Even if there are, the sanity checks in
3354 * kmem_cache_destroy should caught this ill-case.
3356 * Still, we don't want anyone else freeing memcg_caches under our
3357 * noses, which can happen if a new memcg comes to life. As usual,
3358 * we'll take the activate_kmem_mutex to protect ourselves against
3361 mutex_lock(&activate_kmem_mutex);
3362 for_each_memcg_cache_index(i) {
3363 c = cache_from_memcg_idx(s, i);
3368 * We will now manually delete the caches, so to avoid races
3369 * we need to cancel all pending destruction workers and
3370 * proceed with destruction ourselves.
3372 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3373 * and that could spawn the workers again: it is likely that
3374 * the cache still have active pages until this very moment.
3375 * This would lead us back to mem_cgroup_destroy_cache.
3377 * But that will not execute at all if the "dead" flag is not
3378 * set, so flip it down to guarantee we are in control.
3380 c->memcg_params->dead = false;
3381 cancel_work_sync(&c->memcg_params->destroy);
3382 kmem_cache_destroy(c);
3384 mutex_unlock(&activate_kmem_mutex);
3387 struct create_work {
3388 struct mem_cgroup *memcg;
3389 struct kmem_cache *cachep;
3390 struct work_struct work;
3393 static void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3395 struct kmem_cache *cachep;
3396 struct memcg_cache_params *params;
3398 if (!memcg_kmem_is_active(memcg))
3401 mutex_lock(&memcg->slab_caches_mutex);
3402 list_for_each_entry(params, &memcg->memcg_slab_caches, list) {
3403 cachep = memcg_params_to_cache(params);
3404 cachep->memcg_params->dead = true;
3405 schedule_work(&cachep->memcg_params->destroy);
3407 mutex_unlock(&memcg->slab_caches_mutex);
3410 static void memcg_create_cache_work_func(struct work_struct *w)
3412 struct create_work *cw;
3414 cw = container_of(w, struct create_work, work);
3415 memcg_create_kmem_cache(cw->memcg, cw->cachep);
3416 css_put(&cw->memcg->css);
3421 * Enqueue the creation of a per-memcg kmem_cache.
3423 static void __memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3424 struct kmem_cache *cachep)
3426 struct create_work *cw;
3428 cw = kmalloc(sizeof(struct create_work), GFP_NOWAIT);
3430 css_put(&memcg->css);
3435 cw->cachep = cachep;
3437 INIT_WORK(&cw->work, memcg_create_cache_work_func);
3438 schedule_work(&cw->work);
3441 static void memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3442 struct kmem_cache *cachep)
3445 * We need to stop accounting when we kmalloc, because if the
3446 * corresponding kmalloc cache is not yet created, the first allocation
3447 * in __memcg_create_cache_enqueue will recurse.
3449 * However, it is better to enclose the whole function. Depending on
3450 * the debugging options enabled, INIT_WORK(), for instance, can
3451 * trigger an allocation. This too, will make us recurse. Because at
3452 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3453 * the safest choice is to do it like this, wrapping the whole function.
3455 memcg_stop_kmem_account();
3456 __memcg_create_cache_enqueue(memcg, cachep);
3457 memcg_resume_kmem_account();
3460 * Return the kmem_cache we're supposed to use for a slab allocation.
3461 * We try to use the current memcg's version of the cache.
3463 * If the cache does not exist yet, if we are the first user of it,
3464 * we either create it immediately, if possible, or create it asynchronously
3466 * In the latter case, we will let the current allocation go through with
3467 * the original cache.
3469 * Can't be called in interrupt context or from kernel threads.
3470 * This function needs to be called with rcu_read_lock() held.
3472 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
3475 struct mem_cgroup *memcg;
3476 struct kmem_cache *memcg_cachep;
3478 VM_BUG_ON(!cachep->memcg_params);
3479 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
3481 if (!current->mm || current->memcg_kmem_skip_account)
3485 memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
3487 if (!memcg_can_account_kmem(memcg))
3490 memcg_cachep = cache_from_memcg_idx(cachep, memcg_cache_id(memcg));
3491 if (likely(memcg_cachep)) {
3492 cachep = memcg_cachep;
3496 /* The corresponding put will be done in the workqueue. */
3497 if (!css_tryget(&memcg->css))
3502 * If we are in a safe context (can wait, and not in interrupt
3503 * context), we could be be predictable and return right away.
3504 * This would guarantee that the allocation being performed
3505 * already belongs in the new cache.
3507 * However, there are some clashes that can arrive from locking.
3508 * For instance, because we acquire the slab_mutex while doing
3509 * kmem_cache_dup, this means no further allocation could happen
3510 * with the slab_mutex held.
3512 * Also, because cache creation issue get_online_cpus(), this
3513 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3514 * that ends up reversed during cpu hotplug. (cpuset allocates
3515 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3516 * better to defer everything.
3518 memcg_create_cache_enqueue(memcg, cachep);
3524 EXPORT_SYMBOL(__memcg_kmem_get_cache);
3527 * We need to verify if the allocation against current->mm->owner's memcg is
3528 * possible for the given order. But the page is not allocated yet, so we'll
3529 * need a further commit step to do the final arrangements.
3531 * It is possible for the task to switch cgroups in this mean time, so at
3532 * commit time, we can't rely on task conversion any longer. We'll then use
3533 * the handle argument to return to the caller which cgroup we should commit
3534 * against. We could also return the memcg directly and avoid the pointer
3535 * passing, but a boolean return value gives better semantics considering
3536 * the compiled-out case as well.
3538 * Returning true means the allocation is possible.
3541 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3543 struct mem_cgroup *memcg;
3549 * Disabling accounting is only relevant for some specific memcg
3550 * internal allocations. Therefore we would initially not have such
3551 * check here, since direct calls to the page allocator that are marked
3552 * with GFP_KMEMCG only happen outside memcg core. We are mostly
3553 * concerned with cache allocations, and by having this test at
3554 * memcg_kmem_get_cache, we are already able to relay the allocation to
3555 * the root cache and bypass the memcg cache altogether.
3557 * There is one exception, though: the SLUB allocator does not create
3558 * large order caches, but rather service large kmallocs directly from
3559 * the page allocator. Therefore, the following sequence when backed by
3560 * the SLUB allocator:
3562 * memcg_stop_kmem_account();
3563 * kmalloc(<large_number>)
3564 * memcg_resume_kmem_account();
3566 * would effectively ignore the fact that we should skip accounting,
3567 * since it will drive us directly to this function without passing
3568 * through the cache selector memcg_kmem_get_cache. Such large
3569 * allocations are extremely rare but can happen, for instance, for the
3570 * cache arrays. We bring this test here.
3572 if (!current->mm || current->memcg_kmem_skip_account)
3575 memcg = get_mem_cgroup_from_mm(current->mm);
3577 if (!memcg_can_account_kmem(memcg)) {
3578 css_put(&memcg->css);
3582 ret = memcg_charge_kmem(memcg, gfp, PAGE_SIZE << order);
3586 css_put(&memcg->css);
3590 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3593 struct page_cgroup *pc;
3595 VM_BUG_ON(mem_cgroup_is_root(memcg));
3597 /* The page allocation failed. Revert */
3599 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3603 pc = lookup_page_cgroup(page);
3604 lock_page_cgroup(pc);
3605 pc->mem_cgroup = memcg;
3606 SetPageCgroupUsed(pc);
3607 unlock_page_cgroup(pc);
3610 void __memcg_kmem_uncharge_pages(struct page *page, int order)
3612 struct mem_cgroup *memcg = NULL;
3613 struct page_cgroup *pc;
3616 pc = lookup_page_cgroup(page);
3618 * Fast unlocked return. Theoretically might have changed, have to
3619 * check again after locking.
3621 if (!PageCgroupUsed(pc))
3624 lock_page_cgroup(pc);
3625 if (PageCgroupUsed(pc)) {
3626 memcg = pc->mem_cgroup;
3627 ClearPageCgroupUsed(pc);
3629 unlock_page_cgroup(pc);
3632 * We trust that only if there is a memcg associated with the page, it
3633 * is a valid allocation
3638 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
3639 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3642 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3645 #endif /* CONFIG_MEMCG_KMEM */
3647 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3649 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3651 * Because tail pages are not marked as "used", set it. We're under
3652 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3653 * charge/uncharge will be never happen and move_account() is done under
3654 * compound_lock(), so we don't have to take care of races.
3656 void mem_cgroup_split_huge_fixup(struct page *head)
3658 struct page_cgroup *head_pc = lookup_page_cgroup(head);
3659 struct page_cgroup *pc;
3660 struct mem_cgroup *memcg;
3663 if (mem_cgroup_disabled())
3666 memcg = head_pc->mem_cgroup;
3667 for (i = 1; i < HPAGE_PMD_NR; i++) {
3669 pc->mem_cgroup = memcg;
3670 smp_wmb();/* see __commit_charge() */
3671 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
3673 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
3676 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3679 * mem_cgroup_move_account - move account of the page
3681 * @nr_pages: number of regular pages (>1 for huge pages)
3682 * @pc: page_cgroup of the page.
3683 * @from: mem_cgroup which the page is moved from.
3684 * @to: mem_cgroup which the page is moved to. @from != @to.
3686 * The caller must confirm following.
3687 * - page is not on LRU (isolate_page() is useful.)
3688 * - compound_lock is held when nr_pages > 1
3690 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3693 static int mem_cgroup_move_account(struct page *page,
3694 unsigned int nr_pages,
3695 struct page_cgroup *pc,
3696 struct mem_cgroup *from,
3697 struct mem_cgroup *to)
3699 unsigned long flags;
3701 bool anon = PageAnon(page);
3703 VM_BUG_ON(from == to);
3704 VM_BUG_ON_PAGE(PageLRU(page), page);
3706 * The page is isolated from LRU. So, collapse function
3707 * will not handle this page. But page splitting can happen.
3708 * Do this check under compound_page_lock(). The caller should
3712 if (nr_pages > 1 && !PageTransHuge(page))
3715 lock_page_cgroup(pc);
3718 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3721 move_lock_mem_cgroup(from, &flags);
3723 if (!anon && page_mapped(page)) {
3724 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
3726 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
3730 if (PageWriteback(page)) {
3731 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK],
3733 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK],
3737 mem_cgroup_charge_statistics(from, page, anon, -nr_pages);
3739 /* caller should have done css_get */
3740 pc->mem_cgroup = to;
3741 mem_cgroup_charge_statistics(to, page, anon, nr_pages);
3742 move_unlock_mem_cgroup(from, &flags);
3745 unlock_page_cgroup(pc);
3749 memcg_check_events(to, page);
3750 memcg_check_events(from, page);
3756 * mem_cgroup_move_parent - moves page to the parent group
3757 * @page: the page to move
3758 * @pc: page_cgroup of the page
3759 * @child: page's cgroup
3761 * move charges to its parent or the root cgroup if the group has no
3762 * parent (aka use_hierarchy==0).
3763 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3764 * mem_cgroup_move_account fails) the failure is always temporary and
3765 * it signals a race with a page removal/uncharge or migration. In the
3766 * first case the page is on the way out and it will vanish from the LRU
3767 * on the next attempt and the call should be retried later.
3768 * Isolation from the LRU fails only if page has been isolated from
3769 * the LRU since we looked at it and that usually means either global
3770 * reclaim or migration going on. The page will either get back to the
3772 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3773 * (!PageCgroupUsed) or moved to a different group. The page will
3774 * disappear in the next attempt.
3776 static int mem_cgroup_move_parent(struct page *page,
3777 struct page_cgroup *pc,
3778 struct mem_cgroup *child)
3780 struct mem_cgroup *parent;
3781 unsigned int nr_pages;
3782 unsigned long uninitialized_var(flags);
3785 VM_BUG_ON(mem_cgroup_is_root(child));
3788 if (!get_page_unless_zero(page))
3790 if (isolate_lru_page(page))
3793 nr_pages = hpage_nr_pages(page);
3795 parent = parent_mem_cgroup(child);
3797 * If no parent, move charges to root cgroup.
3800 parent = root_mem_cgroup;
3803 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
3804 flags = compound_lock_irqsave(page);
3807 ret = mem_cgroup_move_account(page, nr_pages,
3810 __mem_cgroup_cancel_local_charge(child, nr_pages);
3813 compound_unlock_irqrestore(page, flags);
3814 putback_lru_page(page);
3821 int mem_cgroup_newpage_charge(struct page *page,
3822 struct mm_struct *mm, gfp_t gfp_mask)
3824 unsigned int nr_pages = 1;
3825 struct mem_cgroup *memcg;
3828 if (mem_cgroup_disabled())
3831 VM_BUG_ON_PAGE(page_mapped(page), page);
3832 VM_BUG_ON_PAGE(page->mapping && !PageAnon(page), page);
3835 if (PageTransHuge(page)) {
3836 nr_pages <<= compound_order(page);
3837 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
3839 * Never OOM-kill a process for a huge page. The
3840 * fault handler will fall back to regular pages.
3845 memcg = mem_cgroup_try_charge_mm(mm, gfp_mask, nr_pages, oom);
3848 __mem_cgroup_commit_charge(memcg, page, nr_pages,
3849 MEM_CGROUP_CHARGE_TYPE_ANON, false);
3854 * While swap-in, try_charge -> commit or cancel, the page is locked.
3855 * And when try_charge() successfully returns, one refcnt to memcg without
3856 * struct page_cgroup is acquired. This refcnt will be consumed by
3857 * "commit()" or removed by "cancel()"
3859 static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm,
3862 struct mem_cgroup **memcgp)
3864 struct mem_cgroup *memcg = NULL;
3865 struct page_cgroup *pc;
3868 pc = lookup_page_cgroup(page);
3870 * Every swap fault against a single page tries to charge the
3871 * page, bail as early as possible. shmem_unuse() encounters
3872 * already charged pages, too. The USED bit is protected by
3873 * the page lock, which serializes swap cache removal, which
3874 * in turn serializes uncharging.
3876 if (PageCgroupUsed(pc))
3878 if (do_swap_account)
3879 memcg = try_get_mem_cgroup_from_page(page);
3881 memcg = get_mem_cgroup_from_mm(mm);
3882 ret = mem_cgroup_try_charge(memcg, mask, 1, true);
3883 css_put(&memcg->css);
3885 memcg = root_mem_cgroup;
3893 int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page,
3894 gfp_t gfp_mask, struct mem_cgroup **memcgp)
3896 if (mem_cgroup_disabled()) {
3901 * A racing thread's fault, or swapoff, may have already
3902 * updated the pte, and even removed page from swap cache: in
3903 * those cases unuse_pte()'s pte_same() test will fail; but
3904 * there's also a KSM case which does need to charge the page.
3906 if (!PageSwapCache(page)) {
3907 struct mem_cgroup *memcg;
3909 memcg = mem_cgroup_try_charge_mm(mm, gfp_mask, 1, true);
3915 return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp);
3918 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
3920 if (mem_cgroup_disabled())
3924 __mem_cgroup_cancel_charge(memcg, 1);
3928 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
3929 enum charge_type ctype)
3931 if (mem_cgroup_disabled())
3936 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
3938 * Now swap is on-memory. This means this page may be
3939 * counted both as mem and swap....double count.
3940 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
3941 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
3942 * may call delete_from_swap_cache() before reach here.
3944 if (do_swap_account && PageSwapCache(page)) {
3945 swp_entry_t ent = {.val = page_private(page)};
3946 mem_cgroup_uncharge_swap(ent);
3950 void mem_cgroup_commit_charge_swapin(struct page *page,
3951 struct mem_cgroup *memcg)
3953 __mem_cgroup_commit_charge_swapin(page, memcg,
3954 MEM_CGROUP_CHARGE_TYPE_ANON);
3957 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
3960 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
3961 struct mem_cgroup *memcg;
3964 if (mem_cgroup_disabled())
3966 if (PageCompound(page))
3969 if (PageSwapCache(page)) { /* shmem */
3970 ret = __mem_cgroup_try_charge_swapin(mm, page,
3974 __mem_cgroup_commit_charge_swapin(page, memcg, type);
3979 * Page cache insertions can happen without an actual mm
3980 * context, e.g. during disk probing on boot.
3983 memcg = root_mem_cgroup;
3985 memcg = mem_cgroup_try_charge_mm(mm, gfp_mask, 1, true);
3989 __mem_cgroup_commit_charge(memcg, page, 1, type, false);
3993 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
3994 unsigned int nr_pages,
3995 const enum charge_type ctype)
3997 struct memcg_batch_info *batch = NULL;
3998 bool uncharge_memsw = true;
4000 /* If swapout, usage of swap doesn't decrease */
4001 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
4002 uncharge_memsw = false;
4004 batch = ¤t->memcg_batch;
4006 * In usual, we do css_get() when we remember memcg pointer.
4007 * But in this case, we keep res->usage until end of a series of
4008 * uncharges. Then, it's ok to ignore memcg's refcnt.
4011 batch->memcg = memcg;
4013 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
4014 * In those cases, all pages freed continuously can be expected to be in
4015 * the same cgroup and we have chance to coalesce uncharges.
4016 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
4017 * because we want to do uncharge as soon as possible.
4020 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
4021 goto direct_uncharge;
4024 goto direct_uncharge;
4027 * In typical case, batch->memcg == mem. This means we can
4028 * merge a series of uncharges to an uncharge of res_counter.
4029 * If not, we uncharge res_counter ony by one.
4031 if (batch->memcg != memcg)
4032 goto direct_uncharge;
4033 /* remember freed charge and uncharge it later */
4036 batch->memsw_nr_pages++;
4039 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
4041 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
4042 if (unlikely(batch->memcg != memcg))
4043 memcg_oom_recover(memcg);
4047 * uncharge if !page_mapped(page)
4049 static struct mem_cgroup *
4050 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
4053 struct mem_cgroup *memcg = NULL;
4054 unsigned int nr_pages = 1;
4055 struct page_cgroup *pc;
4058 if (mem_cgroup_disabled())
4061 if (PageTransHuge(page)) {
4062 nr_pages <<= compound_order(page);
4063 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
4066 * Check if our page_cgroup is valid
4068 pc = lookup_page_cgroup(page);
4069 if (unlikely(!PageCgroupUsed(pc)))
4072 lock_page_cgroup(pc);
4074 memcg = pc->mem_cgroup;
4076 if (!PageCgroupUsed(pc))
4079 anon = PageAnon(page);
4082 case MEM_CGROUP_CHARGE_TYPE_ANON:
4084 * Generally PageAnon tells if it's the anon statistics to be
4085 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
4086 * used before page reached the stage of being marked PageAnon.
4090 case MEM_CGROUP_CHARGE_TYPE_DROP:
4091 /* See mem_cgroup_prepare_migration() */
4092 if (page_mapped(page))
4095 * Pages under migration may not be uncharged. But
4096 * end_migration() /must/ be the one uncharging the
4097 * unused post-migration page and so it has to call
4098 * here with the migration bit still set. See the
4099 * res_counter handling below.
4101 if (!end_migration && PageCgroupMigration(pc))
4104 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
4105 if (!PageAnon(page)) { /* Shared memory */
4106 if (page->mapping && !page_is_file_cache(page))
4108 } else if (page_mapped(page)) /* Anon */
4115 mem_cgroup_charge_statistics(memcg, page, anon, -nr_pages);
4117 ClearPageCgroupUsed(pc);
4119 * pc->mem_cgroup is not cleared here. It will be accessed when it's
4120 * freed from LRU. This is safe because uncharged page is expected not
4121 * to be reused (freed soon). Exception is SwapCache, it's handled by
4122 * special functions.
4125 unlock_page_cgroup(pc);
4127 * even after unlock, we have memcg->res.usage here and this memcg
4128 * will never be freed, so it's safe to call css_get().
4130 memcg_check_events(memcg, page);
4131 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
4132 mem_cgroup_swap_statistics(memcg, true);
4133 css_get(&memcg->css);
4136 * Migration does not charge the res_counter for the
4137 * replacement page, so leave it alone when phasing out the
4138 * page that is unused after the migration.
4140 if (!end_migration && !mem_cgroup_is_root(memcg))
4141 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
4146 unlock_page_cgroup(pc);
4150 void mem_cgroup_uncharge_page(struct page *page)
4153 if (page_mapped(page))
4155 VM_BUG_ON_PAGE(page->mapping && !PageAnon(page), page);
4157 * If the page is in swap cache, uncharge should be deferred
4158 * to the swap path, which also properly accounts swap usage
4159 * and handles memcg lifetime.
4161 * Note that this check is not stable and reclaim may add the
4162 * page to swap cache at any time after this. However, if the
4163 * page is not in swap cache by the time page->mapcount hits
4164 * 0, there won't be any page table references to the swap
4165 * slot, and reclaim will free it and not actually write the
4168 if (PageSwapCache(page))
4170 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
4173 void mem_cgroup_uncharge_cache_page(struct page *page)
4175 VM_BUG_ON_PAGE(page_mapped(page), page);
4176 VM_BUG_ON_PAGE(page->mapping, page);
4177 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
4181 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4182 * In that cases, pages are freed continuously and we can expect pages
4183 * are in the same memcg. All these calls itself limits the number of
4184 * pages freed at once, then uncharge_start/end() is called properly.
4185 * This may be called prural(2) times in a context,
4188 void mem_cgroup_uncharge_start(void)
4190 current->memcg_batch.do_batch++;
4191 /* We can do nest. */
4192 if (current->memcg_batch.do_batch == 1) {
4193 current->memcg_batch.memcg = NULL;
4194 current->memcg_batch.nr_pages = 0;
4195 current->memcg_batch.memsw_nr_pages = 0;
4199 void mem_cgroup_uncharge_end(void)
4201 struct memcg_batch_info *batch = ¤t->memcg_batch;
4203 if (!batch->do_batch)
4207 if (batch->do_batch) /* If stacked, do nothing. */
4213 * This "batch->memcg" is valid without any css_get/put etc...
4214 * bacause we hide charges behind us.
4216 if (batch->nr_pages)
4217 res_counter_uncharge(&batch->memcg->res,
4218 batch->nr_pages * PAGE_SIZE);
4219 if (batch->memsw_nr_pages)
4220 res_counter_uncharge(&batch->memcg->memsw,
4221 batch->memsw_nr_pages * PAGE_SIZE);
4222 memcg_oom_recover(batch->memcg);
4223 /* forget this pointer (for sanity check) */
4224 batch->memcg = NULL;
4229 * called after __delete_from_swap_cache() and drop "page" account.
4230 * memcg information is recorded to swap_cgroup of "ent"
4233 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
4235 struct mem_cgroup *memcg;
4236 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
4238 if (!swapout) /* this was a swap cache but the swap is unused ! */
4239 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
4241 memcg = __mem_cgroup_uncharge_common(page, ctype, false);
4244 * record memcg information, if swapout && memcg != NULL,
4245 * css_get() was called in uncharge().
4247 if (do_swap_account && swapout && memcg)
4248 swap_cgroup_record(ent, mem_cgroup_id(memcg));
4252 #ifdef CONFIG_MEMCG_SWAP
4254 * called from swap_entry_free(). remove record in swap_cgroup and
4255 * uncharge "memsw" account.
4257 void mem_cgroup_uncharge_swap(swp_entry_t ent)
4259 struct mem_cgroup *memcg;
4262 if (!do_swap_account)
4265 id = swap_cgroup_record(ent, 0);
4267 memcg = mem_cgroup_lookup(id);
4270 * We uncharge this because swap is freed.
4271 * This memcg can be obsolete one. We avoid calling css_tryget
4273 if (!mem_cgroup_is_root(memcg))
4274 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
4275 mem_cgroup_swap_statistics(memcg, false);
4276 css_put(&memcg->css);
4282 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4283 * @entry: swap entry to be moved
4284 * @from: mem_cgroup which the entry is moved from
4285 * @to: mem_cgroup which the entry is moved to
4287 * It succeeds only when the swap_cgroup's record for this entry is the same
4288 * as the mem_cgroup's id of @from.
4290 * Returns 0 on success, -EINVAL on failure.
4292 * The caller must have charged to @to, IOW, called res_counter_charge() about
4293 * both res and memsw, and called css_get().
4295 static int mem_cgroup_move_swap_account(swp_entry_t entry,
4296 struct mem_cgroup *from, struct mem_cgroup *to)
4298 unsigned short old_id, new_id;
4300 old_id = mem_cgroup_id(from);
4301 new_id = mem_cgroup_id(to);
4303 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
4304 mem_cgroup_swap_statistics(from, false);
4305 mem_cgroup_swap_statistics(to, true);
4307 * This function is only called from task migration context now.
4308 * It postpones res_counter and refcount handling till the end
4309 * of task migration(mem_cgroup_clear_mc()) for performance
4310 * improvement. But we cannot postpone css_get(to) because if
4311 * the process that has been moved to @to does swap-in, the
4312 * refcount of @to might be decreased to 0.
4314 * We are in attach() phase, so the cgroup is guaranteed to be
4315 * alive, so we can just call css_get().
4323 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
4324 struct mem_cgroup *from, struct mem_cgroup *to)
4331 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4334 void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
4335 struct mem_cgroup **memcgp)
4337 struct mem_cgroup *memcg = NULL;
4338 unsigned int nr_pages = 1;
4339 struct page_cgroup *pc;
4340 enum charge_type ctype;
4344 if (mem_cgroup_disabled())
4347 if (PageTransHuge(page))
4348 nr_pages <<= compound_order(page);
4350 pc = lookup_page_cgroup(page);
4351 lock_page_cgroup(pc);
4352 if (PageCgroupUsed(pc)) {
4353 memcg = pc->mem_cgroup;
4354 css_get(&memcg->css);
4356 * At migrating an anonymous page, its mapcount goes down
4357 * to 0 and uncharge() will be called. But, even if it's fully
4358 * unmapped, migration may fail and this page has to be
4359 * charged again. We set MIGRATION flag here and delay uncharge
4360 * until end_migration() is called
4362 * Corner Case Thinking
4364 * When the old page was mapped as Anon and it's unmap-and-freed
4365 * while migration was ongoing.
4366 * If unmap finds the old page, uncharge() of it will be delayed
4367 * until end_migration(). If unmap finds a new page, it's
4368 * uncharged when it make mapcount to be 1->0. If unmap code
4369 * finds swap_migration_entry, the new page will not be mapped
4370 * and end_migration() will find it(mapcount==0).
4373 * When the old page was mapped but migraion fails, the kernel
4374 * remaps it. A charge for it is kept by MIGRATION flag even
4375 * if mapcount goes down to 0. We can do remap successfully
4376 * without charging it again.
4379 * The "old" page is under lock_page() until the end of
4380 * migration, so, the old page itself will not be swapped-out.
4381 * If the new page is swapped out before end_migraton, our
4382 * hook to usual swap-out path will catch the event.
4385 SetPageCgroupMigration(pc);
4387 unlock_page_cgroup(pc);
4389 * If the page is not charged at this point,
4397 * We charge new page before it's used/mapped. So, even if unlock_page()
4398 * is called before end_migration, we can catch all events on this new
4399 * page. In the case new page is migrated but not remapped, new page's
4400 * mapcount will be finally 0 and we call uncharge in end_migration().
4403 ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
4405 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
4407 * The page is committed to the memcg, but it's not actually
4408 * charged to the res_counter since we plan on replacing the
4409 * old one and only one page is going to be left afterwards.
4411 __mem_cgroup_commit_charge(memcg, newpage, nr_pages, ctype, false);
4414 /* remove redundant charge if migration failed*/
4415 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
4416 struct page *oldpage, struct page *newpage, bool migration_ok)
4418 struct page *used, *unused;
4419 struct page_cgroup *pc;
4425 if (!migration_ok) {
4432 anon = PageAnon(used);
4433 __mem_cgroup_uncharge_common(unused,
4434 anon ? MEM_CGROUP_CHARGE_TYPE_ANON
4435 : MEM_CGROUP_CHARGE_TYPE_CACHE,
4437 css_put(&memcg->css);
4439 * We disallowed uncharge of pages under migration because mapcount
4440 * of the page goes down to zero, temporarly.
4441 * Clear the flag and check the page should be charged.
4443 pc = lookup_page_cgroup(oldpage);
4444 lock_page_cgroup(pc);
4445 ClearPageCgroupMigration(pc);
4446 unlock_page_cgroup(pc);
4449 * If a page is a file cache, radix-tree replacement is very atomic
4450 * and we can skip this check. When it was an Anon page, its mapcount
4451 * goes down to 0. But because we added MIGRATION flage, it's not
4452 * uncharged yet. There are several case but page->mapcount check
4453 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4454 * check. (see prepare_charge() also)
4457 mem_cgroup_uncharge_page(used);
4461 * At replace page cache, newpage is not under any memcg but it's on
4462 * LRU. So, this function doesn't touch res_counter but handles LRU
4463 * in correct way. Both pages are locked so we cannot race with uncharge.
4465 void mem_cgroup_replace_page_cache(struct page *oldpage,
4466 struct page *newpage)
4468 struct mem_cgroup *memcg = NULL;
4469 struct page_cgroup *pc;
4470 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4472 if (mem_cgroup_disabled())
4475 pc = lookup_page_cgroup(oldpage);
4476 /* fix accounting on old pages */
4477 lock_page_cgroup(pc);
4478 if (PageCgroupUsed(pc)) {
4479 memcg = pc->mem_cgroup;
4480 mem_cgroup_charge_statistics(memcg, oldpage, false, -1);
4481 ClearPageCgroupUsed(pc);
4483 unlock_page_cgroup(pc);
4486 * When called from shmem_replace_page(), in some cases the
4487 * oldpage has already been charged, and in some cases not.
4492 * Even if newpage->mapping was NULL before starting replacement,
4493 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4494 * LRU while we overwrite pc->mem_cgroup.
4496 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
4499 #ifdef CONFIG_DEBUG_VM
4500 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
4502 struct page_cgroup *pc;
4504 pc = lookup_page_cgroup(page);
4506 * Can be NULL while feeding pages into the page allocator for
4507 * the first time, i.e. during boot or memory hotplug;
4508 * or when mem_cgroup_disabled().
4510 if (likely(pc) && PageCgroupUsed(pc))
4515 bool mem_cgroup_bad_page_check(struct page *page)
4517 if (mem_cgroup_disabled())
4520 return lookup_page_cgroup_used(page) != NULL;
4523 void mem_cgroup_print_bad_page(struct page *page)
4525 struct page_cgroup *pc;
4527 pc = lookup_page_cgroup_used(page);
4529 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4530 pc, pc->flags, pc->mem_cgroup);
4535 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
4536 unsigned long long val)
4539 u64 memswlimit, memlimit;
4541 int children = mem_cgroup_count_children(memcg);
4542 u64 curusage, oldusage;
4546 * For keeping hierarchical_reclaim simple, how long we should retry
4547 * is depends on callers. We set our retry-count to be function
4548 * of # of children which we should visit in this loop.
4550 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
4552 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4555 while (retry_count) {
4556 if (signal_pending(current)) {
4561 * Rather than hide all in some function, I do this in
4562 * open coded manner. You see what this really does.
4563 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4565 mutex_lock(&set_limit_mutex);
4566 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4567 if (memswlimit < val) {
4569 mutex_unlock(&set_limit_mutex);
4573 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4577 ret = res_counter_set_limit(&memcg->res, val);
4579 if (memswlimit == val)
4580 memcg->memsw_is_minimum = true;
4582 memcg->memsw_is_minimum = false;
4584 mutex_unlock(&set_limit_mutex);
4589 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4590 MEM_CGROUP_RECLAIM_SHRINK);
4591 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4592 /* Usage is reduced ? */
4593 if (curusage >= oldusage)
4596 oldusage = curusage;
4598 if (!ret && enlarge)
4599 memcg_oom_recover(memcg);
4604 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
4605 unsigned long long val)
4608 u64 memlimit, memswlimit, oldusage, curusage;
4609 int children = mem_cgroup_count_children(memcg);
4613 /* see mem_cgroup_resize_res_limit */
4614 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
4615 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4616 while (retry_count) {
4617 if (signal_pending(current)) {
4622 * Rather than hide all in some function, I do this in
4623 * open coded manner. You see what this really does.
4624 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4626 mutex_lock(&set_limit_mutex);
4627 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4628 if (memlimit > val) {
4630 mutex_unlock(&set_limit_mutex);
4633 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4634 if (memswlimit < val)
4636 ret = res_counter_set_limit(&memcg->memsw, val);
4638 if (memlimit == val)
4639 memcg->memsw_is_minimum = true;
4641 memcg->memsw_is_minimum = false;
4643 mutex_unlock(&set_limit_mutex);
4648 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4649 MEM_CGROUP_RECLAIM_NOSWAP |
4650 MEM_CGROUP_RECLAIM_SHRINK);
4651 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4652 /* Usage is reduced ? */
4653 if (curusage >= oldusage)
4656 oldusage = curusage;
4658 if (!ret && enlarge)
4659 memcg_oom_recover(memcg);
4663 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
4665 unsigned long *total_scanned)
4667 unsigned long nr_reclaimed = 0;
4668 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
4669 unsigned long reclaimed;
4671 struct mem_cgroup_tree_per_zone *mctz;
4672 unsigned long long excess;
4673 unsigned long nr_scanned;
4678 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
4680 * This loop can run a while, specially if mem_cgroup's continuously
4681 * keep exceeding their soft limit and putting the system under
4688 mz = mem_cgroup_largest_soft_limit_node(mctz);
4693 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
4694 gfp_mask, &nr_scanned);
4695 nr_reclaimed += reclaimed;
4696 *total_scanned += nr_scanned;
4697 spin_lock(&mctz->lock);
4700 * If we failed to reclaim anything from this memory cgroup
4701 * it is time to move on to the next cgroup
4707 * Loop until we find yet another one.
4709 * By the time we get the soft_limit lock
4710 * again, someone might have aded the
4711 * group back on the RB tree. Iterate to
4712 * make sure we get a different mem.
4713 * mem_cgroup_largest_soft_limit_node returns
4714 * NULL if no other cgroup is present on
4718 __mem_cgroup_largest_soft_limit_node(mctz);
4720 css_put(&next_mz->memcg->css);
4721 else /* next_mz == NULL or other memcg */
4725 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
4726 excess = res_counter_soft_limit_excess(&mz->memcg->res);
4728 * One school of thought says that we should not add
4729 * back the node to the tree if reclaim returns 0.
4730 * But our reclaim could return 0, simply because due
4731 * to priority we are exposing a smaller subset of
4732 * memory to reclaim from. Consider this as a longer
4735 /* If excess == 0, no tree ops */
4736 __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess);
4737 spin_unlock(&mctz->lock);
4738 css_put(&mz->memcg->css);
4741 * Could not reclaim anything and there are no more
4742 * mem cgroups to try or we seem to be looping without
4743 * reclaiming anything.
4745 if (!nr_reclaimed &&
4747 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
4749 } while (!nr_reclaimed);
4751 css_put(&next_mz->memcg->css);
4752 return nr_reclaimed;
4756 * mem_cgroup_force_empty_list - clears LRU of a group
4757 * @memcg: group to clear
4760 * @lru: lru to to clear
4762 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4763 * reclaim the pages page themselves - pages are moved to the parent (or root)
4766 static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
4767 int node, int zid, enum lru_list lru)
4769 struct lruvec *lruvec;
4770 unsigned long flags;
4771 struct list_head *list;
4775 zone = &NODE_DATA(node)->node_zones[zid];
4776 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
4777 list = &lruvec->lists[lru];
4781 struct page_cgroup *pc;
4784 spin_lock_irqsave(&zone->lru_lock, flags);
4785 if (list_empty(list)) {
4786 spin_unlock_irqrestore(&zone->lru_lock, flags);
4789 page = list_entry(list->prev, struct page, lru);
4791 list_move(&page->lru, list);
4793 spin_unlock_irqrestore(&zone->lru_lock, flags);
4796 spin_unlock_irqrestore(&zone->lru_lock, flags);
4798 pc = lookup_page_cgroup(page);
4800 if (mem_cgroup_move_parent(page, pc, memcg)) {
4801 /* found lock contention or "pc" is obsolete. */
4806 } while (!list_empty(list));
4810 * make mem_cgroup's charge to be 0 if there is no task by moving
4811 * all the charges and pages to the parent.
4812 * This enables deleting this mem_cgroup.
4814 * Caller is responsible for holding css reference on the memcg.
4816 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
4822 /* This is for making all *used* pages to be on LRU. */
4823 lru_add_drain_all();
4824 drain_all_stock_sync(memcg);
4825 mem_cgroup_start_move(memcg);
4826 for_each_node_state(node, N_MEMORY) {
4827 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4830 mem_cgroup_force_empty_list(memcg,
4835 mem_cgroup_end_move(memcg);
4836 memcg_oom_recover(memcg);
4840 * Kernel memory may not necessarily be trackable to a specific
4841 * process. So they are not migrated, and therefore we can't
4842 * expect their value to drop to 0 here.
4843 * Having res filled up with kmem only is enough.
4845 * This is a safety check because mem_cgroup_force_empty_list
4846 * could have raced with mem_cgroup_replace_page_cache callers
4847 * so the lru seemed empty but the page could have been added
4848 * right after the check. RES_USAGE should be safe as we always
4849 * charge before adding to the LRU.
4851 usage = res_counter_read_u64(&memcg->res, RES_USAGE) -
4852 res_counter_read_u64(&memcg->kmem, RES_USAGE);
4853 } while (usage > 0);
4856 static inline bool memcg_has_children(struct mem_cgroup *memcg)
4858 lockdep_assert_held(&memcg_create_mutex);
4860 * The lock does not prevent addition or deletion to the list
4861 * of children, but it prevents a new child from being
4862 * initialized based on this parent in css_online(), so it's
4863 * enough to decide whether hierarchically inherited
4864 * attributes can still be changed or not.
4866 return memcg->use_hierarchy &&
4867 !list_empty(&memcg->css.cgroup->children);
4871 * Reclaims as many pages from the given memcg as possible and moves
4872 * the rest to the parent.
4874 * Caller is responsible for holding css reference for memcg.
4876 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
4878 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
4879 struct cgroup *cgrp = memcg->css.cgroup;
4881 /* returns EBUSY if there is a task or if we come here twice. */
4882 if (cgroup_has_tasks(cgrp) || !list_empty(&cgrp->children))
4885 /* we call try-to-free pages for make this cgroup empty */
4886 lru_add_drain_all();
4887 /* try to free all pages in this cgroup */
4888 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
4891 if (signal_pending(current))
4894 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
4898 /* maybe some writeback is necessary */
4899 congestion_wait(BLK_RW_ASYNC, HZ/10);
4904 mem_cgroup_reparent_charges(memcg);
4909 static int mem_cgroup_force_empty_write(struct cgroup_subsys_state *css,
4912 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4914 if (mem_cgroup_is_root(memcg))
4916 return mem_cgroup_force_empty(memcg);
4919 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
4922 return mem_cgroup_from_css(css)->use_hierarchy;
4925 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
4926 struct cftype *cft, u64 val)
4929 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4930 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(css_parent(&memcg->css));
4932 mutex_lock(&memcg_create_mutex);
4934 if (memcg->use_hierarchy == val)
4938 * If parent's use_hierarchy is set, we can't make any modifications
4939 * in the child subtrees. If it is unset, then the change can
4940 * occur, provided the current cgroup has no children.
4942 * For the root cgroup, parent_mem is NULL, we allow value to be
4943 * set if there are no children.
4945 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
4946 (val == 1 || val == 0)) {
4947 if (list_empty(&memcg->css.cgroup->children))
4948 memcg->use_hierarchy = val;
4955 mutex_unlock(&memcg_create_mutex);
4961 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
4962 enum mem_cgroup_stat_index idx)
4964 struct mem_cgroup *iter;
4967 /* Per-cpu values can be negative, use a signed accumulator */
4968 for_each_mem_cgroup_tree(iter, memcg)
4969 val += mem_cgroup_read_stat(iter, idx);
4971 if (val < 0) /* race ? */
4976 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
4980 if (!mem_cgroup_is_root(memcg)) {
4982 return res_counter_read_u64(&memcg->res, RES_USAGE);
4984 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
4988 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
4989 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
4991 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
4992 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
4995 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
4997 return val << PAGE_SHIFT;
5000 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
5003 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5008 type = MEMFILE_TYPE(cft->private);
5009 name = MEMFILE_ATTR(cft->private);
5013 if (name == RES_USAGE)
5014 val = mem_cgroup_usage(memcg, false);
5016 val = res_counter_read_u64(&memcg->res, name);
5019 if (name == RES_USAGE)
5020 val = mem_cgroup_usage(memcg, true);
5022 val = res_counter_read_u64(&memcg->memsw, name);
5025 val = res_counter_read_u64(&memcg->kmem, name);
5034 #ifdef CONFIG_MEMCG_KMEM
5035 /* should be called with activate_kmem_mutex held */
5036 static int __memcg_activate_kmem(struct mem_cgroup *memcg,
5037 unsigned long long limit)
5042 if (memcg_kmem_is_active(memcg))
5046 * We are going to allocate memory for data shared by all memory
5047 * cgroups so let's stop accounting here.
5049 memcg_stop_kmem_account();
5052 * For simplicity, we won't allow this to be disabled. It also can't
5053 * be changed if the cgroup has children already, or if tasks had
5056 * If tasks join before we set the limit, a person looking at
5057 * kmem.usage_in_bytes will have no way to determine when it took
5058 * place, which makes the value quite meaningless.
5060 * After it first became limited, changes in the value of the limit are
5061 * of course permitted.
5063 mutex_lock(&memcg_create_mutex);
5064 if (cgroup_has_tasks(memcg->css.cgroup) || memcg_has_children(memcg))
5066 mutex_unlock(&memcg_create_mutex);
5070 memcg_id = ida_simple_get(&kmem_limited_groups,
5071 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
5078 * Make sure we have enough space for this cgroup in each root cache's
5081 err = memcg_update_all_caches(memcg_id + 1);
5085 memcg->kmemcg_id = memcg_id;
5086 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
5087 mutex_init(&memcg->slab_caches_mutex);
5090 * We couldn't have accounted to this cgroup, because it hasn't got the
5091 * active bit set yet, so this should succeed.
5093 err = res_counter_set_limit(&memcg->kmem, limit);
5096 static_key_slow_inc(&memcg_kmem_enabled_key);
5098 * Setting the active bit after enabling static branching will
5099 * guarantee no one starts accounting before all call sites are
5102 memcg_kmem_set_active(memcg);
5104 memcg_resume_kmem_account();
5108 ida_simple_remove(&kmem_limited_groups, memcg_id);
5112 static int memcg_activate_kmem(struct mem_cgroup *memcg,
5113 unsigned long long limit)
5117 mutex_lock(&activate_kmem_mutex);
5118 ret = __memcg_activate_kmem(memcg, limit);
5119 mutex_unlock(&activate_kmem_mutex);
5123 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
5124 unsigned long long val)
5128 if (!memcg_kmem_is_active(memcg))
5129 ret = memcg_activate_kmem(memcg, val);
5131 ret = res_counter_set_limit(&memcg->kmem, val);
5135 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
5138 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5143 mutex_lock(&activate_kmem_mutex);
5145 * If the parent cgroup is not kmem-active now, it cannot be activated
5146 * after this point, because it has at least one child already.
5148 if (memcg_kmem_is_active(parent))
5149 ret = __memcg_activate_kmem(memcg, RES_COUNTER_MAX);
5150 mutex_unlock(&activate_kmem_mutex);
5154 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
5155 unsigned long long val)
5159 #endif /* CONFIG_MEMCG_KMEM */
5162 * The user of this function is...
5165 static int mem_cgroup_write(struct cgroup_subsys_state *css, struct cftype *cft,
5168 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5171 unsigned long long val;
5174 type = MEMFILE_TYPE(cft->private);
5175 name = MEMFILE_ATTR(cft->private);
5179 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
5183 /* This function does all necessary parse...reuse it */
5184 ret = res_counter_memparse_write_strategy(buffer, &val);
5188 ret = mem_cgroup_resize_limit(memcg, val);
5189 else if (type == _MEMSWAP)
5190 ret = mem_cgroup_resize_memsw_limit(memcg, val);
5191 else if (type == _KMEM)
5192 ret = memcg_update_kmem_limit(memcg, val);
5196 case RES_SOFT_LIMIT:
5197 ret = res_counter_memparse_write_strategy(buffer, &val);
5201 * For memsw, soft limits are hard to implement in terms
5202 * of semantics, for now, we support soft limits for
5203 * control without swap
5206 ret = res_counter_set_soft_limit(&memcg->res, val);
5211 ret = -EINVAL; /* should be BUG() ? */
5217 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
5218 unsigned long long *mem_limit, unsigned long long *memsw_limit)
5220 unsigned long long min_limit, min_memsw_limit, tmp;
5222 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
5223 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5224 if (!memcg->use_hierarchy)
5227 while (css_parent(&memcg->css)) {
5228 memcg = mem_cgroup_from_css(css_parent(&memcg->css));
5229 if (!memcg->use_hierarchy)
5231 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
5232 min_limit = min(min_limit, tmp);
5233 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5234 min_memsw_limit = min(min_memsw_limit, tmp);
5237 *mem_limit = min_limit;
5238 *memsw_limit = min_memsw_limit;
5241 static int mem_cgroup_reset(struct cgroup_subsys_state *css, unsigned int event)
5243 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5247 type = MEMFILE_TYPE(event);
5248 name = MEMFILE_ATTR(event);
5253 res_counter_reset_max(&memcg->res);
5254 else if (type == _MEMSWAP)
5255 res_counter_reset_max(&memcg->memsw);
5256 else if (type == _KMEM)
5257 res_counter_reset_max(&memcg->kmem);
5263 res_counter_reset_failcnt(&memcg->res);
5264 else if (type == _MEMSWAP)
5265 res_counter_reset_failcnt(&memcg->memsw);
5266 else if (type == _KMEM)
5267 res_counter_reset_failcnt(&memcg->kmem);
5276 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
5279 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
5283 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5284 struct cftype *cft, u64 val)
5286 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5288 if (val >= (1 << NR_MOVE_TYPE))
5292 * No kind of locking is needed in here, because ->can_attach() will
5293 * check this value once in the beginning of the process, and then carry
5294 * on with stale data. This means that changes to this value will only
5295 * affect task migrations starting after the change.
5297 memcg->move_charge_at_immigrate = val;
5301 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5302 struct cftype *cft, u64 val)
5309 static int memcg_numa_stat_show(struct seq_file *m, void *v)
5313 unsigned int lru_mask;
5316 static const struct numa_stat stats[] = {
5317 { "total", LRU_ALL },
5318 { "file", LRU_ALL_FILE },
5319 { "anon", LRU_ALL_ANON },
5320 { "unevictable", BIT(LRU_UNEVICTABLE) },
5322 const struct numa_stat *stat;
5325 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5327 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
5328 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
5329 seq_printf(m, "%s=%lu", stat->name, nr);
5330 for_each_node_state(nid, N_MEMORY) {
5331 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5333 seq_printf(m, " N%d=%lu", nid, nr);
5338 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
5339 struct mem_cgroup *iter;
5342 for_each_mem_cgroup_tree(iter, memcg)
5343 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
5344 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
5345 for_each_node_state(nid, N_MEMORY) {
5347 for_each_mem_cgroup_tree(iter, memcg)
5348 nr += mem_cgroup_node_nr_lru_pages(
5349 iter, nid, stat->lru_mask);
5350 seq_printf(m, " N%d=%lu", nid, nr);
5357 #endif /* CONFIG_NUMA */
5359 static inline void mem_cgroup_lru_names_not_uptodate(void)
5361 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
5364 static int memcg_stat_show(struct seq_file *m, void *v)
5366 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5367 struct mem_cgroup *mi;
5370 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5371 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5373 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
5374 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
5377 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
5378 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
5379 mem_cgroup_read_events(memcg, i));
5381 for (i = 0; i < NR_LRU_LISTS; i++)
5382 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
5383 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
5385 /* Hierarchical information */
5387 unsigned long long limit, memsw_limit;
5388 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
5389 seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
5390 if (do_swap_account)
5391 seq_printf(m, "hierarchical_memsw_limit %llu\n",
5395 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5398 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5400 for_each_mem_cgroup_tree(mi, memcg)
5401 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
5402 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
5405 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
5406 unsigned long long val = 0;
5408 for_each_mem_cgroup_tree(mi, memcg)
5409 val += mem_cgroup_read_events(mi, i);
5410 seq_printf(m, "total_%s %llu\n",
5411 mem_cgroup_events_names[i], val);
5414 for (i = 0; i < NR_LRU_LISTS; i++) {
5415 unsigned long long val = 0;
5417 for_each_mem_cgroup_tree(mi, memcg)
5418 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
5419 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
5422 #ifdef CONFIG_DEBUG_VM
5425 struct mem_cgroup_per_zone *mz;
5426 struct zone_reclaim_stat *rstat;
5427 unsigned long recent_rotated[2] = {0, 0};
5428 unsigned long recent_scanned[2] = {0, 0};
5430 for_each_online_node(nid)
5431 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
5432 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
5433 rstat = &mz->lruvec.reclaim_stat;
5435 recent_rotated[0] += rstat->recent_rotated[0];
5436 recent_rotated[1] += rstat->recent_rotated[1];
5437 recent_scanned[0] += rstat->recent_scanned[0];
5438 recent_scanned[1] += rstat->recent_scanned[1];
5440 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
5441 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
5442 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
5443 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
5450 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
5453 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5455 return mem_cgroup_swappiness(memcg);
5458 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
5459 struct cftype *cft, u64 val)
5461 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5462 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(&memcg->css));
5464 if (val > 100 || !parent)
5467 mutex_lock(&memcg_create_mutex);
5469 /* If under hierarchy, only empty-root can set this value */
5470 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5471 mutex_unlock(&memcg_create_mutex);
5475 memcg->swappiness = val;
5477 mutex_unlock(&memcg_create_mutex);
5482 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
5484 struct mem_cgroup_threshold_ary *t;
5490 t = rcu_dereference(memcg->thresholds.primary);
5492 t = rcu_dereference(memcg->memsw_thresholds.primary);
5497 usage = mem_cgroup_usage(memcg, swap);
5500 * current_threshold points to threshold just below or equal to usage.
5501 * If it's not true, a threshold was crossed after last
5502 * call of __mem_cgroup_threshold().
5504 i = t->current_threshold;
5507 * Iterate backward over array of thresholds starting from
5508 * current_threshold and check if a threshold is crossed.
5509 * If none of thresholds below usage is crossed, we read
5510 * only one element of the array here.
5512 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
5513 eventfd_signal(t->entries[i].eventfd, 1);
5515 /* i = current_threshold + 1 */
5519 * Iterate forward over array of thresholds starting from
5520 * current_threshold+1 and check if a threshold is crossed.
5521 * If none of thresholds above usage is crossed, we read
5522 * only one element of the array here.
5524 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
5525 eventfd_signal(t->entries[i].eventfd, 1);
5527 /* Update current_threshold */
5528 t->current_threshold = i - 1;
5533 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
5536 __mem_cgroup_threshold(memcg, false);
5537 if (do_swap_account)
5538 __mem_cgroup_threshold(memcg, true);
5540 memcg = parent_mem_cgroup(memcg);
5544 static int compare_thresholds(const void *a, const void *b)
5546 const struct mem_cgroup_threshold *_a = a;
5547 const struct mem_cgroup_threshold *_b = b;
5549 if (_a->threshold > _b->threshold)
5552 if (_a->threshold < _b->threshold)
5558 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
5560 struct mem_cgroup_eventfd_list *ev;
5562 list_for_each_entry(ev, &memcg->oom_notify, list)
5563 eventfd_signal(ev->eventfd, 1);
5567 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
5569 struct mem_cgroup *iter;
5571 for_each_mem_cgroup_tree(iter, memcg)
5572 mem_cgroup_oom_notify_cb(iter);
5575 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
5576 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
5578 struct mem_cgroup_thresholds *thresholds;
5579 struct mem_cgroup_threshold_ary *new;
5580 u64 threshold, usage;
5583 ret = res_counter_memparse_write_strategy(args, &threshold);
5587 mutex_lock(&memcg->thresholds_lock);
5590 thresholds = &memcg->thresholds;
5591 else if (type == _MEMSWAP)
5592 thresholds = &memcg->memsw_thresholds;
5596 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5598 /* Check if a threshold crossed before adding a new one */
5599 if (thresholds->primary)
5600 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5602 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
5604 /* Allocate memory for new array of thresholds */
5605 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
5613 /* Copy thresholds (if any) to new array */
5614 if (thresholds->primary) {
5615 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
5616 sizeof(struct mem_cgroup_threshold));
5619 /* Add new threshold */
5620 new->entries[size - 1].eventfd = eventfd;
5621 new->entries[size - 1].threshold = threshold;
5623 /* Sort thresholds. Registering of new threshold isn't time-critical */
5624 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
5625 compare_thresholds, NULL);
5627 /* Find current threshold */
5628 new->current_threshold = -1;
5629 for (i = 0; i < size; i++) {
5630 if (new->entries[i].threshold <= usage) {
5632 * new->current_threshold will not be used until
5633 * rcu_assign_pointer(), so it's safe to increment
5636 ++new->current_threshold;
5641 /* Free old spare buffer and save old primary buffer as spare */
5642 kfree(thresholds->spare);
5643 thresholds->spare = thresholds->primary;
5645 rcu_assign_pointer(thresholds->primary, new);
5647 /* To be sure that nobody uses thresholds */
5651 mutex_unlock(&memcg->thresholds_lock);
5656 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
5657 struct eventfd_ctx *eventfd, const char *args)
5659 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
5662 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
5663 struct eventfd_ctx *eventfd, const char *args)
5665 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
5668 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
5669 struct eventfd_ctx *eventfd, enum res_type type)
5671 struct mem_cgroup_thresholds *thresholds;
5672 struct mem_cgroup_threshold_ary *new;
5676 mutex_lock(&memcg->thresholds_lock);
5678 thresholds = &memcg->thresholds;
5679 else if (type == _MEMSWAP)
5680 thresholds = &memcg->memsw_thresholds;
5684 if (!thresholds->primary)
5687 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5689 /* Check if a threshold crossed before removing */
5690 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5692 /* Calculate new number of threshold */
5694 for (i = 0; i < thresholds->primary->size; i++) {
5695 if (thresholds->primary->entries[i].eventfd != eventfd)
5699 new = thresholds->spare;
5701 /* Set thresholds array to NULL if we don't have thresholds */
5710 /* Copy thresholds and find current threshold */
5711 new->current_threshold = -1;
5712 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
5713 if (thresholds->primary->entries[i].eventfd == eventfd)
5716 new->entries[j] = thresholds->primary->entries[i];
5717 if (new->entries[j].threshold <= usage) {
5719 * new->current_threshold will not be used
5720 * until rcu_assign_pointer(), so it's safe to increment
5723 ++new->current_threshold;
5729 /* Swap primary and spare array */
5730 thresholds->spare = thresholds->primary;
5731 /* If all events are unregistered, free the spare array */
5733 kfree(thresholds->spare);
5734 thresholds->spare = NULL;
5737 rcu_assign_pointer(thresholds->primary, new);
5739 /* To be sure that nobody uses thresholds */
5742 mutex_unlock(&memcg->thresholds_lock);
5745 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
5746 struct eventfd_ctx *eventfd)
5748 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
5751 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
5752 struct eventfd_ctx *eventfd)
5754 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
5757 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
5758 struct eventfd_ctx *eventfd, const char *args)
5760 struct mem_cgroup_eventfd_list *event;
5762 event = kmalloc(sizeof(*event), GFP_KERNEL);
5766 spin_lock(&memcg_oom_lock);
5768 event->eventfd = eventfd;
5769 list_add(&event->list, &memcg->oom_notify);
5771 /* already in OOM ? */
5772 if (atomic_read(&memcg->under_oom))
5773 eventfd_signal(eventfd, 1);
5774 spin_unlock(&memcg_oom_lock);
5779 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
5780 struct eventfd_ctx *eventfd)
5782 struct mem_cgroup_eventfd_list *ev, *tmp;
5784 spin_lock(&memcg_oom_lock);
5786 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
5787 if (ev->eventfd == eventfd) {
5788 list_del(&ev->list);
5793 spin_unlock(&memcg_oom_lock);
5796 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
5798 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
5800 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
5801 seq_printf(sf, "under_oom %d\n", (bool)atomic_read(&memcg->under_oom));
5805 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
5806 struct cftype *cft, u64 val)
5808 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5809 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(&memcg->css));
5811 /* cannot set to root cgroup and only 0 and 1 are allowed */
5812 if (!parent || !((val == 0) || (val == 1)))
5815 mutex_lock(&memcg_create_mutex);
5816 /* oom-kill-disable is a flag for subhierarchy. */
5817 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5818 mutex_unlock(&memcg_create_mutex);
5821 memcg->oom_kill_disable = val;
5823 memcg_oom_recover(memcg);
5824 mutex_unlock(&memcg_create_mutex);
5828 #ifdef CONFIG_MEMCG_KMEM
5829 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5833 memcg->kmemcg_id = -1;
5834 ret = memcg_propagate_kmem(memcg);
5838 return mem_cgroup_sockets_init(memcg, ss);
5841 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5843 mem_cgroup_sockets_destroy(memcg);
5846 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5848 if (!memcg_kmem_is_active(memcg))
5852 * kmem charges can outlive the cgroup. In the case of slab
5853 * pages, for instance, a page contain objects from various
5854 * processes. As we prevent from taking a reference for every
5855 * such allocation we have to be careful when doing uncharge
5856 * (see memcg_uncharge_kmem) and here during offlining.
5858 * The idea is that that only the _last_ uncharge which sees
5859 * the dead memcg will drop the last reference. An additional
5860 * reference is taken here before the group is marked dead
5861 * which is then paired with css_put during uncharge resp. here.
5863 * Although this might sound strange as this path is called from
5864 * css_offline() when the referencemight have dropped down to 0
5865 * and shouldn't be incremented anymore (css_tryget would fail)
5866 * we do not have other options because of the kmem allocations
5869 css_get(&memcg->css);
5871 memcg_kmem_mark_dead(memcg);
5873 if (res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0)
5876 if (memcg_kmem_test_and_clear_dead(memcg))
5877 css_put(&memcg->css);
5880 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5885 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5889 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5895 * DO NOT USE IN NEW FILES.
5897 * "cgroup.event_control" implementation.
5899 * This is way over-engineered. It tries to support fully configurable
5900 * events for each user. Such level of flexibility is completely
5901 * unnecessary especially in the light of the planned unified hierarchy.
5903 * Please deprecate this and replace with something simpler if at all
5908 * Unregister event and free resources.
5910 * Gets called from workqueue.
5912 static void memcg_event_remove(struct work_struct *work)
5914 struct mem_cgroup_event *event =
5915 container_of(work, struct mem_cgroup_event, remove);
5916 struct mem_cgroup *memcg = event->memcg;
5918 remove_wait_queue(event->wqh, &event->wait);
5920 event->unregister_event(memcg, event->eventfd);
5922 /* Notify userspace the event is going away. */
5923 eventfd_signal(event->eventfd, 1);
5925 eventfd_ctx_put(event->eventfd);
5927 css_put(&memcg->css);
5931 * Gets called on POLLHUP on eventfd when user closes it.
5933 * Called with wqh->lock held and interrupts disabled.
5935 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
5936 int sync, void *key)
5938 struct mem_cgroup_event *event =
5939 container_of(wait, struct mem_cgroup_event, wait);
5940 struct mem_cgroup *memcg = event->memcg;
5941 unsigned long flags = (unsigned long)key;
5943 if (flags & POLLHUP) {
5945 * If the event has been detached at cgroup removal, we
5946 * can simply return knowing the other side will cleanup
5949 * We can't race against event freeing since the other
5950 * side will require wqh->lock via remove_wait_queue(),
5953 spin_lock(&memcg->event_list_lock);
5954 if (!list_empty(&event->list)) {
5955 list_del_init(&event->list);
5957 * We are in atomic context, but cgroup_event_remove()
5958 * may sleep, so we have to call it in workqueue.
5960 schedule_work(&event->remove);
5962 spin_unlock(&memcg->event_list_lock);
5968 static void memcg_event_ptable_queue_proc(struct file *file,
5969 wait_queue_head_t *wqh, poll_table *pt)
5971 struct mem_cgroup_event *event =
5972 container_of(pt, struct mem_cgroup_event, pt);
5975 add_wait_queue(wqh, &event->wait);
5979 * DO NOT USE IN NEW FILES.
5981 * Parse input and register new cgroup event handler.
5983 * Input must be in format '<event_fd> <control_fd> <args>'.
5984 * Interpretation of args is defined by control file implementation.
5986 static int memcg_write_event_control(struct cgroup_subsys_state *css,
5987 struct cftype *cft, char *buffer)
5989 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5990 struct mem_cgroup_event *event;
5991 struct cgroup_subsys_state *cfile_css;
5992 unsigned int efd, cfd;
5999 efd = simple_strtoul(buffer, &endp, 10);
6004 cfd = simple_strtoul(buffer, &endp, 10);
6005 if ((*endp != ' ') && (*endp != '\0'))
6009 event = kzalloc(sizeof(*event), GFP_KERNEL);
6013 event->memcg = memcg;
6014 INIT_LIST_HEAD(&event->list);
6015 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
6016 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
6017 INIT_WORK(&event->remove, memcg_event_remove);
6025 event->eventfd = eventfd_ctx_fileget(efile.file);
6026 if (IS_ERR(event->eventfd)) {
6027 ret = PTR_ERR(event->eventfd);
6034 goto out_put_eventfd;
6037 /* the process need read permission on control file */
6038 /* AV: shouldn't we check that it's been opened for read instead? */
6039 ret = inode_permission(file_inode(cfile.file), MAY_READ);
6044 * Determine the event callbacks and set them in @event. This used
6045 * to be done via struct cftype but cgroup core no longer knows
6046 * about these events. The following is crude but the whole thing
6047 * is for compatibility anyway.
6049 * DO NOT ADD NEW FILES.
6051 name = cfile.file->f_dentry->d_name.name;
6053 if (!strcmp(name, "memory.usage_in_bytes")) {
6054 event->register_event = mem_cgroup_usage_register_event;
6055 event->unregister_event = mem_cgroup_usage_unregister_event;
6056 } else if (!strcmp(name, "memory.oom_control")) {
6057 event->register_event = mem_cgroup_oom_register_event;
6058 event->unregister_event = mem_cgroup_oom_unregister_event;
6059 } else if (!strcmp(name, "memory.pressure_level")) {
6060 event->register_event = vmpressure_register_event;
6061 event->unregister_event = vmpressure_unregister_event;
6062 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
6063 event->register_event = memsw_cgroup_usage_register_event;
6064 event->unregister_event = memsw_cgroup_usage_unregister_event;
6071 * Verify @cfile should belong to @css. Also, remaining events are
6072 * automatically removed on cgroup destruction but the removal is
6073 * asynchronous, so take an extra ref on @css.
6075 cfile_css = css_tryget_from_dir(cfile.file->f_dentry->d_parent,
6076 &memory_cgrp_subsys);
6078 if (IS_ERR(cfile_css))
6080 if (cfile_css != css) {
6085 ret = event->register_event(memcg, event->eventfd, buffer);
6089 efile.file->f_op->poll(efile.file, &event->pt);
6091 spin_lock(&memcg->event_list_lock);
6092 list_add(&event->list, &memcg->event_list);
6093 spin_unlock(&memcg->event_list_lock);
6105 eventfd_ctx_put(event->eventfd);
6114 static struct cftype mem_cgroup_files[] = {
6116 .name = "usage_in_bytes",
6117 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
6118 .read_u64 = mem_cgroup_read_u64,
6121 .name = "max_usage_in_bytes",
6122 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
6123 .trigger = mem_cgroup_reset,
6124 .read_u64 = mem_cgroup_read_u64,
6127 .name = "limit_in_bytes",
6128 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
6129 .write_string = mem_cgroup_write,
6130 .read_u64 = mem_cgroup_read_u64,
6133 .name = "soft_limit_in_bytes",
6134 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
6135 .write_string = mem_cgroup_write,
6136 .read_u64 = mem_cgroup_read_u64,
6140 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
6141 .trigger = mem_cgroup_reset,
6142 .read_u64 = mem_cgroup_read_u64,
6146 .seq_show = memcg_stat_show,
6149 .name = "force_empty",
6150 .trigger = mem_cgroup_force_empty_write,
6153 .name = "use_hierarchy",
6154 .flags = CFTYPE_INSANE,
6155 .write_u64 = mem_cgroup_hierarchy_write,
6156 .read_u64 = mem_cgroup_hierarchy_read,
6159 .name = "cgroup.event_control", /* XXX: for compat */
6160 .write_string = memcg_write_event_control,
6161 .flags = CFTYPE_NO_PREFIX,
6165 .name = "swappiness",
6166 .read_u64 = mem_cgroup_swappiness_read,
6167 .write_u64 = mem_cgroup_swappiness_write,
6170 .name = "move_charge_at_immigrate",
6171 .read_u64 = mem_cgroup_move_charge_read,
6172 .write_u64 = mem_cgroup_move_charge_write,
6175 .name = "oom_control",
6176 .seq_show = mem_cgroup_oom_control_read,
6177 .write_u64 = mem_cgroup_oom_control_write,
6178 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
6181 .name = "pressure_level",
6185 .name = "numa_stat",
6186 .seq_show = memcg_numa_stat_show,
6189 #ifdef CONFIG_MEMCG_KMEM
6191 .name = "kmem.limit_in_bytes",
6192 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
6193 .write_string = mem_cgroup_write,
6194 .read_u64 = mem_cgroup_read_u64,
6197 .name = "kmem.usage_in_bytes",
6198 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
6199 .read_u64 = mem_cgroup_read_u64,
6202 .name = "kmem.failcnt",
6203 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
6204 .trigger = mem_cgroup_reset,
6205 .read_u64 = mem_cgroup_read_u64,
6208 .name = "kmem.max_usage_in_bytes",
6209 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
6210 .trigger = mem_cgroup_reset,
6211 .read_u64 = mem_cgroup_read_u64,
6213 #ifdef CONFIG_SLABINFO
6215 .name = "kmem.slabinfo",
6216 .seq_show = mem_cgroup_slabinfo_read,
6220 { }, /* terminate */
6223 #ifdef CONFIG_MEMCG_SWAP
6224 static struct cftype memsw_cgroup_files[] = {
6226 .name = "memsw.usage_in_bytes",
6227 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6228 .read_u64 = mem_cgroup_read_u64,
6231 .name = "memsw.max_usage_in_bytes",
6232 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6233 .trigger = mem_cgroup_reset,
6234 .read_u64 = mem_cgroup_read_u64,
6237 .name = "memsw.limit_in_bytes",
6238 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6239 .write_string = mem_cgroup_write,
6240 .read_u64 = mem_cgroup_read_u64,
6243 .name = "memsw.failcnt",
6244 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6245 .trigger = mem_cgroup_reset,
6246 .read_u64 = mem_cgroup_read_u64,
6248 { }, /* terminate */
6251 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6253 struct mem_cgroup_per_node *pn;
6254 struct mem_cgroup_per_zone *mz;
6255 int zone, tmp = node;
6257 * This routine is called against possible nodes.
6258 * But it's BUG to call kmalloc() against offline node.
6260 * TODO: this routine can waste much memory for nodes which will
6261 * never be onlined. It's better to use memory hotplug callback
6264 if (!node_state(node, N_NORMAL_MEMORY))
6266 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
6270 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6271 mz = &pn->zoneinfo[zone];
6272 lruvec_init(&mz->lruvec);
6273 mz->usage_in_excess = 0;
6274 mz->on_tree = false;
6277 memcg->nodeinfo[node] = pn;
6281 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6283 kfree(memcg->nodeinfo[node]);
6286 static struct mem_cgroup *mem_cgroup_alloc(void)
6288 struct mem_cgroup *memcg;
6291 size = sizeof(struct mem_cgroup);
6292 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
6294 memcg = kzalloc(size, GFP_KERNEL);
6298 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
6301 spin_lock_init(&memcg->pcp_counter_lock);
6310 * At destroying mem_cgroup, references from swap_cgroup can remain.
6311 * (scanning all at force_empty is too costly...)
6313 * Instead of clearing all references at force_empty, we remember
6314 * the number of reference from swap_cgroup and free mem_cgroup when
6315 * it goes down to 0.
6317 * Removal of cgroup itself succeeds regardless of refs from swap.
6320 static void __mem_cgroup_free(struct mem_cgroup *memcg)
6324 mem_cgroup_remove_from_trees(memcg);
6327 free_mem_cgroup_per_zone_info(memcg, node);
6329 free_percpu(memcg->stat);
6332 * We need to make sure that (at least for now), the jump label
6333 * destruction code runs outside of the cgroup lock. This is because
6334 * get_online_cpus(), which is called from the static_branch update,
6335 * can't be called inside the cgroup_lock. cpusets are the ones
6336 * enforcing this dependency, so if they ever change, we might as well.
6338 * schedule_work() will guarantee this happens. Be careful if you need
6339 * to move this code around, and make sure it is outside
6342 disarm_static_keys(memcg);
6347 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6349 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
6351 if (!memcg->res.parent)
6353 return mem_cgroup_from_res_counter(memcg->res.parent, res);
6355 EXPORT_SYMBOL(parent_mem_cgroup);
6357 static void __init mem_cgroup_soft_limit_tree_init(void)
6359 struct mem_cgroup_tree_per_node *rtpn;
6360 struct mem_cgroup_tree_per_zone *rtpz;
6361 int tmp, node, zone;
6363 for_each_node(node) {
6365 if (!node_state(node, N_NORMAL_MEMORY))
6367 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
6370 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6372 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6373 rtpz = &rtpn->rb_tree_per_zone[zone];
6374 rtpz->rb_root = RB_ROOT;
6375 spin_lock_init(&rtpz->lock);
6380 static struct cgroup_subsys_state * __ref
6381 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
6383 struct mem_cgroup *memcg;
6384 long error = -ENOMEM;
6387 memcg = mem_cgroup_alloc();
6389 return ERR_PTR(error);
6392 if (alloc_mem_cgroup_per_zone_info(memcg, node))
6396 if (parent_css == NULL) {
6397 root_mem_cgroup = memcg;
6398 res_counter_init(&memcg->res, NULL);
6399 res_counter_init(&memcg->memsw, NULL);
6400 res_counter_init(&memcg->kmem, NULL);
6403 memcg->last_scanned_node = MAX_NUMNODES;
6404 INIT_LIST_HEAD(&memcg->oom_notify);
6405 memcg->move_charge_at_immigrate = 0;
6406 mutex_init(&memcg->thresholds_lock);
6407 spin_lock_init(&memcg->move_lock);
6408 vmpressure_init(&memcg->vmpressure);
6409 INIT_LIST_HEAD(&memcg->event_list);
6410 spin_lock_init(&memcg->event_list_lock);
6415 __mem_cgroup_free(memcg);
6416 return ERR_PTR(error);
6420 mem_cgroup_css_online(struct cgroup_subsys_state *css)
6422 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6423 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(css));
6425 if (css->cgroup->id > MEM_CGROUP_ID_MAX)
6431 mutex_lock(&memcg_create_mutex);
6433 memcg->use_hierarchy = parent->use_hierarchy;
6434 memcg->oom_kill_disable = parent->oom_kill_disable;
6435 memcg->swappiness = mem_cgroup_swappiness(parent);
6437 if (parent->use_hierarchy) {
6438 res_counter_init(&memcg->res, &parent->res);
6439 res_counter_init(&memcg->memsw, &parent->memsw);
6440 res_counter_init(&memcg->kmem, &parent->kmem);
6443 * No need to take a reference to the parent because cgroup
6444 * core guarantees its existence.
6447 res_counter_init(&memcg->res, NULL);
6448 res_counter_init(&memcg->memsw, NULL);
6449 res_counter_init(&memcg->kmem, NULL);
6451 * Deeper hierachy with use_hierarchy == false doesn't make
6452 * much sense so let cgroup subsystem know about this
6453 * unfortunate state in our controller.
6455 if (parent != root_mem_cgroup)
6456 memory_cgrp_subsys.broken_hierarchy = true;
6458 mutex_unlock(&memcg_create_mutex);
6460 return memcg_init_kmem(memcg, &memory_cgrp_subsys);
6464 * Announce all parents that a group from their hierarchy is gone.
6466 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup *memcg)
6468 struct mem_cgroup *parent = memcg;
6470 while ((parent = parent_mem_cgroup(parent)))
6471 mem_cgroup_iter_invalidate(parent);
6474 * if the root memcg is not hierarchical we have to check it
6477 if (!root_mem_cgroup->use_hierarchy)
6478 mem_cgroup_iter_invalidate(root_mem_cgroup);
6481 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
6483 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6484 struct mem_cgroup_event *event, *tmp;
6485 struct cgroup_subsys_state *iter;
6488 * Unregister events and notify userspace.
6489 * Notify userspace about cgroup removing only after rmdir of cgroup
6490 * directory to avoid race between userspace and kernelspace.
6492 spin_lock(&memcg->event_list_lock);
6493 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
6494 list_del_init(&event->list);
6495 schedule_work(&event->remove);
6497 spin_unlock(&memcg->event_list_lock);
6499 kmem_cgroup_css_offline(memcg);
6501 mem_cgroup_invalidate_reclaim_iterators(memcg);
6504 * This requires that offlining is serialized. Right now that is
6505 * guaranteed because css_killed_work_fn() holds the cgroup_mutex.
6507 css_for_each_descendant_post(iter, css)
6508 mem_cgroup_reparent_charges(mem_cgroup_from_css(iter));
6510 mem_cgroup_destroy_all_caches(memcg);
6511 vmpressure_cleanup(&memcg->vmpressure);
6514 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
6516 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6518 * XXX: css_offline() would be where we should reparent all
6519 * memory to prepare the cgroup for destruction. However,
6520 * memcg does not do css_tryget() and res_counter charging
6521 * under the same RCU lock region, which means that charging
6522 * could race with offlining. Offlining only happens to
6523 * cgroups with no tasks in them but charges can show up
6524 * without any tasks from the swapin path when the target
6525 * memcg is looked up from the swapout record and not from the
6526 * current task as it usually is. A race like this can leak
6527 * charges and put pages with stale cgroup pointers into
6531 * lookup_swap_cgroup_id()
6533 * mem_cgroup_lookup()
6536 * disable css_tryget()
6539 * reparent_charges()
6540 * res_counter_charge()
6543 * pc->mem_cgroup = dead memcg
6546 * The bulk of the charges are still moved in offline_css() to
6547 * avoid pinning a lot of pages in case a long-term reference
6548 * like a swapout record is deferring the css_free() to long
6549 * after offlining. But this makes sure we catch any charges
6550 * made after offlining:
6552 mem_cgroup_reparent_charges(memcg);
6554 memcg_destroy_kmem(memcg);
6555 __mem_cgroup_free(memcg);
6559 /* Handlers for move charge at task migration. */
6560 #define PRECHARGE_COUNT_AT_ONCE 256
6561 static int mem_cgroup_do_precharge(unsigned long count)
6564 int batch_count = PRECHARGE_COUNT_AT_ONCE;
6565 struct mem_cgroup *memcg = mc.to;
6567 if (mem_cgroup_is_root(memcg)) {
6568 mc.precharge += count;
6569 /* we don't need css_get for root */
6572 /* try to charge at once */
6574 struct res_counter *dummy;
6576 * "memcg" cannot be under rmdir() because we've already checked
6577 * by cgroup_lock_live_cgroup() that it is not removed and we
6578 * are still under the same cgroup_mutex. So we can postpone
6581 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
6583 if (do_swap_account && res_counter_charge(&memcg->memsw,
6584 PAGE_SIZE * count, &dummy)) {
6585 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
6588 mc.precharge += count;
6592 /* fall back to one by one charge */
6594 if (signal_pending(current)) {
6598 if (!batch_count--) {
6599 batch_count = PRECHARGE_COUNT_AT_ONCE;
6602 ret = mem_cgroup_try_charge(memcg, GFP_KERNEL, 1, false);
6604 /* mem_cgroup_clear_mc() will do uncharge later */
6612 * get_mctgt_type - get target type of moving charge
6613 * @vma: the vma the pte to be checked belongs
6614 * @addr: the address corresponding to the pte to be checked
6615 * @ptent: the pte to be checked
6616 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6619 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6620 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6621 * move charge. if @target is not NULL, the page is stored in target->page
6622 * with extra refcnt got(Callers should handle it).
6623 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6624 * target for charge migration. if @target is not NULL, the entry is stored
6627 * Called with pte lock held.
6634 enum mc_target_type {
6640 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
6641 unsigned long addr, pte_t ptent)
6643 struct page *page = vm_normal_page(vma, addr, ptent);
6645 if (!page || !page_mapped(page))
6647 if (PageAnon(page)) {
6648 /* we don't move shared anon */
6651 } else if (!move_file())
6652 /* we ignore mapcount for file pages */
6654 if (!get_page_unless_zero(page))
6661 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6662 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6664 struct page *page = NULL;
6665 swp_entry_t ent = pte_to_swp_entry(ptent);
6667 if (!move_anon() || non_swap_entry(ent))
6670 * Because lookup_swap_cache() updates some statistics counter,
6671 * we call find_get_page() with swapper_space directly.
6673 page = find_get_page(swap_address_space(ent), ent.val);
6674 if (do_swap_account)
6675 entry->val = ent.val;
6680 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6681 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6687 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
6688 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6690 struct page *page = NULL;
6691 struct address_space *mapping;
6694 if (!vma->vm_file) /* anonymous vma */
6699 mapping = vma->vm_file->f_mapping;
6700 if (pte_none(ptent))
6701 pgoff = linear_page_index(vma, addr);
6702 else /* pte_file(ptent) is true */
6703 pgoff = pte_to_pgoff(ptent);
6705 /* page is moved even if it's not RSS of this task(page-faulted). */
6706 page = find_get_page(mapping, pgoff);
6709 /* shmem/tmpfs may report page out on swap: account for that too. */
6710 if (radix_tree_exceptional_entry(page)) {
6711 swp_entry_t swap = radix_to_swp_entry(page);
6712 if (do_swap_account)
6714 page = find_get_page(swap_address_space(swap), swap.val);
6720 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
6721 unsigned long addr, pte_t ptent, union mc_target *target)
6723 struct page *page = NULL;
6724 struct page_cgroup *pc;
6725 enum mc_target_type ret = MC_TARGET_NONE;
6726 swp_entry_t ent = { .val = 0 };
6728 if (pte_present(ptent))
6729 page = mc_handle_present_pte(vma, addr, ptent);
6730 else if (is_swap_pte(ptent))
6731 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
6732 else if (pte_none(ptent) || pte_file(ptent))
6733 page = mc_handle_file_pte(vma, addr, ptent, &ent);
6735 if (!page && !ent.val)
6738 pc = lookup_page_cgroup(page);
6740 * Do only loose check w/o page_cgroup lock.
6741 * mem_cgroup_move_account() checks the pc is valid or not under
6744 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6745 ret = MC_TARGET_PAGE;
6747 target->page = page;
6749 if (!ret || !target)
6752 /* There is a swap entry and a page doesn't exist or isn't charged */
6753 if (ent.val && !ret &&
6754 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
6755 ret = MC_TARGET_SWAP;
6762 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6764 * We don't consider swapping or file mapped pages because THP does not
6765 * support them for now.
6766 * Caller should make sure that pmd_trans_huge(pmd) is true.
6768 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6769 unsigned long addr, pmd_t pmd, union mc_target *target)
6771 struct page *page = NULL;
6772 struct page_cgroup *pc;
6773 enum mc_target_type ret = MC_TARGET_NONE;
6775 page = pmd_page(pmd);
6776 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
6779 pc = lookup_page_cgroup(page);
6780 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6781 ret = MC_TARGET_PAGE;
6784 target->page = page;
6790 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6791 unsigned long addr, pmd_t pmd, union mc_target *target)
6793 return MC_TARGET_NONE;
6797 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
6798 unsigned long addr, unsigned long end,
6799 struct mm_walk *walk)
6801 struct vm_area_struct *vma = walk->private;
6805 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
6806 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
6807 mc.precharge += HPAGE_PMD_NR;
6812 if (pmd_trans_unstable(pmd))
6814 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6815 for (; addr != end; pte++, addr += PAGE_SIZE)
6816 if (get_mctgt_type(vma, addr, *pte, NULL))
6817 mc.precharge++; /* increment precharge temporarily */
6818 pte_unmap_unlock(pte - 1, ptl);
6824 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
6826 unsigned long precharge;
6827 struct vm_area_struct *vma;
6829 down_read(&mm->mmap_sem);
6830 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6831 struct mm_walk mem_cgroup_count_precharge_walk = {
6832 .pmd_entry = mem_cgroup_count_precharge_pte_range,
6836 if (is_vm_hugetlb_page(vma))
6838 walk_page_range(vma->vm_start, vma->vm_end,
6839 &mem_cgroup_count_precharge_walk);
6841 up_read(&mm->mmap_sem);
6843 precharge = mc.precharge;
6849 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
6851 unsigned long precharge = mem_cgroup_count_precharge(mm);
6853 VM_BUG_ON(mc.moving_task);
6854 mc.moving_task = current;
6855 return mem_cgroup_do_precharge(precharge);
6858 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6859 static void __mem_cgroup_clear_mc(void)
6861 struct mem_cgroup *from = mc.from;
6862 struct mem_cgroup *to = mc.to;
6865 /* we must uncharge all the leftover precharges from mc.to */
6867 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
6871 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6872 * we must uncharge here.
6874 if (mc.moved_charge) {
6875 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
6876 mc.moved_charge = 0;
6878 /* we must fixup refcnts and charges */
6879 if (mc.moved_swap) {
6880 /* uncharge swap account from the old cgroup */
6881 if (!mem_cgroup_is_root(mc.from))
6882 res_counter_uncharge(&mc.from->memsw,
6883 PAGE_SIZE * mc.moved_swap);
6885 for (i = 0; i < mc.moved_swap; i++)
6886 css_put(&mc.from->css);
6888 if (!mem_cgroup_is_root(mc.to)) {
6890 * we charged both to->res and to->memsw, so we should
6893 res_counter_uncharge(&mc.to->res,
6894 PAGE_SIZE * mc.moved_swap);
6896 /* we've already done css_get(mc.to) */
6899 memcg_oom_recover(from);
6900 memcg_oom_recover(to);
6901 wake_up_all(&mc.waitq);
6904 static void mem_cgroup_clear_mc(void)
6906 struct mem_cgroup *from = mc.from;
6909 * we must clear moving_task before waking up waiters at the end of
6912 mc.moving_task = NULL;
6913 __mem_cgroup_clear_mc();
6914 spin_lock(&mc.lock);
6917 spin_unlock(&mc.lock);
6918 mem_cgroup_end_move(from);
6921 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
6922 struct cgroup_taskset *tset)
6924 struct task_struct *p = cgroup_taskset_first(tset);
6926 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6927 unsigned long move_charge_at_immigrate;
6930 * We are now commited to this value whatever it is. Changes in this
6931 * tunable will only affect upcoming migrations, not the current one.
6932 * So we need to save it, and keep it going.
6934 move_charge_at_immigrate = memcg->move_charge_at_immigrate;
6935 if (move_charge_at_immigrate) {
6936 struct mm_struct *mm;
6937 struct mem_cgroup *from = mem_cgroup_from_task(p);
6939 VM_BUG_ON(from == memcg);
6941 mm = get_task_mm(p);
6944 /* We move charges only when we move a owner of the mm */
6945 if (mm->owner == p) {
6948 VM_BUG_ON(mc.precharge);
6949 VM_BUG_ON(mc.moved_charge);
6950 VM_BUG_ON(mc.moved_swap);
6951 mem_cgroup_start_move(from);
6952 spin_lock(&mc.lock);
6955 mc.immigrate_flags = move_charge_at_immigrate;
6956 spin_unlock(&mc.lock);
6957 /* We set mc.moving_task later */
6959 ret = mem_cgroup_precharge_mc(mm);
6961 mem_cgroup_clear_mc();
6968 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
6969 struct cgroup_taskset *tset)
6971 mem_cgroup_clear_mc();
6974 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6975 unsigned long addr, unsigned long end,
6976 struct mm_walk *walk)
6979 struct vm_area_struct *vma = walk->private;
6982 enum mc_target_type target_type;
6983 union mc_target target;
6985 struct page_cgroup *pc;
6988 * We don't take compound_lock() here but no race with splitting thp
6990 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6991 * under splitting, which means there's no concurrent thp split,
6992 * - if another thread runs into split_huge_page() just after we
6993 * entered this if-block, the thread must wait for page table lock
6994 * to be unlocked in __split_huge_page_splitting(), where the main
6995 * part of thp split is not executed yet.
6997 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
6998 if (mc.precharge < HPAGE_PMD_NR) {
7002 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
7003 if (target_type == MC_TARGET_PAGE) {
7005 if (!isolate_lru_page(page)) {
7006 pc = lookup_page_cgroup(page);
7007 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
7008 pc, mc.from, mc.to)) {
7009 mc.precharge -= HPAGE_PMD_NR;
7010 mc.moved_charge += HPAGE_PMD_NR;
7012 putback_lru_page(page);
7020 if (pmd_trans_unstable(pmd))
7023 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
7024 for (; addr != end; addr += PAGE_SIZE) {
7025 pte_t ptent = *(pte++);
7031 switch (get_mctgt_type(vma, addr, ptent, &target)) {
7032 case MC_TARGET_PAGE:
7034 if (isolate_lru_page(page))
7036 pc = lookup_page_cgroup(page);
7037 if (!mem_cgroup_move_account(page, 1, pc,
7040 /* we uncharge from mc.from later. */
7043 putback_lru_page(page);
7044 put: /* get_mctgt_type() gets the page */
7047 case MC_TARGET_SWAP:
7049 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
7051 /* we fixup refcnts and charges later. */
7059 pte_unmap_unlock(pte - 1, ptl);
7064 * We have consumed all precharges we got in can_attach().
7065 * We try charge one by one, but don't do any additional
7066 * charges to mc.to if we have failed in charge once in attach()
7069 ret = mem_cgroup_do_precharge(1);
7077 static void mem_cgroup_move_charge(struct mm_struct *mm)
7079 struct vm_area_struct *vma;
7081 lru_add_drain_all();
7083 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
7085 * Someone who are holding the mmap_sem might be waiting in
7086 * waitq. So we cancel all extra charges, wake up all waiters,
7087 * and retry. Because we cancel precharges, we might not be able
7088 * to move enough charges, but moving charge is a best-effort
7089 * feature anyway, so it wouldn't be a big problem.
7091 __mem_cgroup_clear_mc();
7095 for (vma = mm->mmap; vma; vma = vma->vm_next) {
7097 struct mm_walk mem_cgroup_move_charge_walk = {
7098 .pmd_entry = mem_cgroup_move_charge_pte_range,
7102 if (is_vm_hugetlb_page(vma))
7104 ret = walk_page_range(vma->vm_start, vma->vm_end,
7105 &mem_cgroup_move_charge_walk);
7108 * means we have consumed all precharges and failed in
7109 * doing additional charge. Just abandon here.
7113 up_read(&mm->mmap_sem);
7116 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
7117 struct cgroup_taskset *tset)
7119 struct task_struct *p = cgroup_taskset_first(tset);
7120 struct mm_struct *mm = get_task_mm(p);
7124 mem_cgroup_move_charge(mm);
7128 mem_cgroup_clear_mc();
7130 #else /* !CONFIG_MMU */
7131 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
7132 struct cgroup_taskset *tset)
7136 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
7137 struct cgroup_taskset *tset)
7140 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
7141 struct cgroup_taskset *tset)
7147 * Cgroup retains root cgroups across [un]mount cycles making it necessary
7148 * to verify sane_behavior flag on each mount attempt.
7150 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
7153 * use_hierarchy is forced with sane_behavior. cgroup core
7154 * guarantees that @root doesn't have any children, so turning it
7155 * on for the root memcg is enough.
7157 if (cgroup_sane_behavior(root_css->cgroup))
7158 mem_cgroup_from_css(root_css)->use_hierarchy = true;
7161 struct cgroup_subsys memory_cgrp_subsys = {
7162 .css_alloc = mem_cgroup_css_alloc,
7163 .css_online = mem_cgroup_css_online,
7164 .css_offline = mem_cgroup_css_offline,
7165 .css_free = mem_cgroup_css_free,
7166 .can_attach = mem_cgroup_can_attach,
7167 .cancel_attach = mem_cgroup_cancel_attach,
7168 .attach = mem_cgroup_move_task,
7169 .bind = mem_cgroup_bind,
7170 .base_cftypes = mem_cgroup_files,
7174 #ifdef CONFIG_MEMCG_SWAP
7175 static int __init enable_swap_account(char *s)
7177 if (!strcmp(s, "1"))
7178 really_do_swap_account = 1;
7179 else if (!strcmp(s, "0"))
7180 really_do_swap_account = 0;
7183 __setup("swapaccount=", enable_swap_account);
7185 static void __init memsw_file_init(void)
7187 WARN_ON(cgroup_add_cftypes(&memory_cgrp_subsys, memsw_cgroup_files));
7190 static void __init enable_swap_cgroup(void)
7192 if (!mem_cgroup_disabled() && really_do_swap_account) {
7193 do_swap_account = 1;
7199 static void __init enable_swap_cgroup(void)
7205 * subsys_initcall() for memory controller.
7207 * Some parts like hotcpu_notifier() have to be initialized from this context
7208 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
7209 * everything that doesn't depend on a specific mem_cgroup structure should
7210 * be initialized from here.
7212 static int __init mem_cgroup_init(void)
7214 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
7215 enable_swap_cgroup();
7216 mem_cgroup_soft_limit_tree_init();
7220 subsys_initcall(mem_cgroup_init);