1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * Kernel Memory Controller
14 * Copyright (C) 2012 Parallels Inc. and Google Inc.
15 * Authors: Glauber Costa and Suleiman Souhlal
17 * This program is free software; you can redistribute it and/or modify
18 * it under the terms of the GNU General Public License as published by
19 * the Free Software Foundation; either version 2 of the License, or
20 * (at your option) any later version.
22 * This program is distributed in the hope that it will be useful,
23 * but WITHOUT ANY WARRANTY; without even the implied warranty of
24 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
25 * GNU General Public License for more details.
28 #include <linux/res_counter.h>
29 #include <linux/memcontrol.h>
30 #include <linux/cgroup.h>
32 #include <linux/hugetlb.h>
33 #include <linux/pagemap.h>
34 #include <linux/smp.h>
35 #include <linux/page-flags.h>
36 #include <linux/backing-dev.h>
37 #include <linux/bit_spinlock.h>
38 #include <linux/rcupdate.h>
39 #include <linux/limits.h>
40 #include <linux/export.h>
41 #include <linux/mutex.h>
42 #include <linux/rbtree.h>
43 #include <linux/slab.h>
44 #include <linux/swap.h>
45 #include <linux/swapops.h>
46 #include <linux/spinlock.h>
47 #include <linux/eventfd.h>
48 #include <linux/poll.h>
49 #include <linux/sort.h>
51 #include <linux/seq_file.h>
52 #include <linux/vmpressure.h>
53 #include <linux/mm_inline.h>
54 #include <linux/page_cgroup.h>
55 #include <linux/cpu.h>
56 #include <linux/oom.h>
57 #include <linux/lockdep.h>
58 #include <linux/file.h>
62 #include <net/tcp_memcontrol.h>
65 #include <asm/uaccess.h>
67 #include <trace/events/vmscan.h>
69 struct cgroup_subsys mem_cgroup_subsys __read_mostly;
70 EXPORT_SYMBOL(mem_cgroup_subsys);
72 #define MEM_CGROUP_RECLAIM_RETRIES 5
73 static struct mem_cgroup *root_mem_cgroup __read_mostly;
75 #ifdef CONFIG_MEMCG_SWAP
76 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
77 int do_swap_account __read_mostly;
79 /* for remember boot option*/
80 #ifdef CONFIG_MEMCG_SWAP_ENABLED
81 static int really_do_swap_account __initdata = 1;
83 static int really_do_swap_account __initdata = 0;
87 #define do_swap_account 0
91 static const char * const mem_cgroup_stat_names[] = {
100 enum mem_cgroup_events_index {
101 MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */
102 MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */
103 MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */
104 MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */
105 MEM_CGROUP_EVENTS_NSTATS,
108 static const char * const mem_cgroup_events_names[] = {
115 static const char * const mem_cgroup_lru_names[] = {
124 * Per memcg event counter is incremented at every pagein/pageout. With THP,
125 * it will be incremated by the number of pages. This counter is used for
126 * for trigger some periodic events. This is straightforward and better
127 * than using jiffies etc. to handle periodic memcg event.
129 enum mem_cgroup_events_target {
130 MEM_CGROUP_TARGET_THRESH,
131 MEM_CGROUP_TARGET_SOFTLIMIT,
132 MEM_CGROUP_TARGET_NUMAINFO,
135 #define THRESHOLDS_EVENTS_TARGET 128
136 #define SOFTLIMIT_EVENTS_TARGET 1024
137 #define NUMAINFO_EVENTS_TARGET 1024
139 struct mem_cgroup_stat_cpu {
140 long count[MEM_CGROUP_STAT_NSTATS];
141 unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
142 unsigned long nr_page_events;
143 unsigned long targets[MEM_CGROUP_NTARGETS];
146 struct mem_cgroup_reclaim_iter {
148 * last scanned hierarchy member. Valid only if last_dead_count
149 * matches memcg->dead_count of the hierarchy root group.
151 struct mem_cgroup *last_visited;
154 /* scan generation, increased every round-trip */
155 unsigned int generation;
159 * per-zone information in memory controller.
161 struct mem_cgroup_per_zone {
162 struct lruvec lruvec;
163 unsigned long lru_size[NR_LRU_LISTS];
165 struct mem_cgroup_reclaim_iter reclaim_iter[DEF_PRIORITY + 1];
167 struct rb_node tree_node; /* RB tree node */
168 unsigned long long usage_in_excess;/* Set to the value by which */
169 /* the soft limit is exceeded*/
171 struct mem_cgroup *memcg; /* Back pointer, we cannot */
172 /* use container_of */
175 struct mem_cgroup_per_node {
176 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
180 * Cgroups above their limits are maintained in a RB-Tree, independent of
181 * their hierarchy representation
184 struct mem_cgroup_tree_per_zone {
185 struct rb_root rb_root;
189 struct mem_cgroup_tree_per_node {
190 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
193 struct mem_cgroup_tree {
194 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
197 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
199 struct mem_cgroup_threshold {
200 struct eventfd_ctx *eventfd;
205 struct mem_cgroup_threshold_ary {
206 /* An array index points to threshold just below or equal to usage. */
207 int current_threshold;
208 /* Size of entries[] */
210 /* Array of thresholds */
211 struct mem_cgroup_threshold entries[0];
214 struct mem_cgroup_thresholds {
215 /* Primary thresholds array */
216 struct mem_cgroup_threshold_ary *primary;
218 * Spare threshold array.
219 * This is needed to make mem_cgroup_unregister_event() "never fail".
220 * It must be able to store at least primary->size - 1 entries.
222 struct mem_cgroup_threshold_ary *spare;
226 struct mem_cgroup_eventfd_list {
227 struct list_head list;
228 struct eventfd_ctx *eventfd;
232 * cgroup_event represents events which userspace want to receive.
234 struct mem_cgroup_event {
236 * memcg which the event belongs to.
238 struct mem_cgroup *memcg;
240 * eventfd to signal userspace about the event.
242 struct eventfd_ctx *eventfd;
244 * Each of these stored in a list by the cgroup.
246 struct list_head list;
248 * register_event() callback will be used to add new userspace
249 * waiter for changes related to this event. Use eventfd_signal()
250 * on eventfd to send notification to userspace.
252 int (*register_event)(struct mem_cgroup *memcg,
253 struct eventfd_ctx *eventfd, const char *args);
255 * unregister_event() callback will be called when userspace closes
256 * the eventfd or on cgroup removing. This callback must be set,
257 * if you want provide notification functionality.
259 void (*unregister_event)(struct mem_cgroup *memcg,
260 struct eventfd_ctx *eventfd);
262 * All fields below needed to unregister event when
263 * userspace closes eventfd.
266 wait_queue_head_t *wqh;
268 struct work_struct remove;
271 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
272 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
275 * The memory controller data structure. The memory controller controls both
276 * page cache and RSS per cgroup. We would eventually like to provide
277 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
278 * to help the administrator determine what knobs to tune.
280 * TODO: Add a water mark for the memory controller. Reclaim will begin when
281 * we hit the water mark. May be even add a low water mark, such that
282 * no reclaim occurs from a cgroup at it's low water mark, this is
283 * a feature that will be implemented much later in the future.
286 struct cgroup_subsys_state css;
288 * the counter to account for memory usage
290 struct res_counter res;
292 /* vmpressure notifications */
293 struct vmpressure vmpressure;
296 * the counter to account for mem+swap usage.
298 struct res_counter memsw;
301 * the counter to account for kernel memory usage.
303 struct res_counter kmem;
305 * Should the accounting and control be hierarchical, per subtree?
308 unsigned long kmem_account_flags; /* See KMEM_ACCOUNTED_*, below */
312 atomic_t oom_wakeups;
315 /* OOM-Killer disable */
316 int oom_kill_disable;
318 /* set when res.limit == memsw.limit */
319 bool memsw_is_minimum;
321 /* protect arrays of thresholds */
322 struct mutex thresholds_lock;
324 /* thresholds for memory usage. RCU-protected */
325 struct mem_cgroup_thresholds thresholds;
327 /* thresholds for mem+swap usage. RCU-protected */
328 struct mem_cgroup_thresholds memsw_thresholds;
330 /* For oom notifier event fd */
331 struct list_head oom_notify;
334 * Should we move charges of a task when a task is moved into this
335 * mem_cgroup ? And what type of charges should we move ?
337 unsigned long move_charge_at_immigrate;
339 * set > 0 if pages under this cgroup are moving to other cgroup.
341 atomic_t moving_account;
342 /* taken only while moving_account > 0 */
343 spinlock_t move_lock;
347 struct mem_cgroup_stat_cpu __percpu *stat;
349 * used when a cpu is offlined or other synchronizations
350 * See mem_cgroup_read_stat().
352 struct mem_cgroup_stat_cpu nocpu_base;
353 spinlock_t pcp_counter_lock;
356 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
357 struct cg_proto tcp_mem;
359 #if defined(CONFIG_MEMCG_KMEM)
360 /* analogous to slab_common's slab_caches list. per-memcg */
361 struct list_head memcg_slab_caches;
362 /* Not a spinlock, we can take a lot of time walking the list */
363 struct mutex slab_caches_mutex;
364 /* Index in the kmem_cache->memcg_params->memcg_caches array */
368 int last_scanned_node;
370 nodemask_t scan_nodes;
371 atomic_t numainfo_events;
372 atomic_t numainfo_updating;
375 /* List of events which userspace want to receive */
376 struct list_head event_list;
377 spinlock_t event_list_lock;
379 struct mem_cgroup_per_node *nodeinfo[0];
380 /* WARNING: nodeinfo must be the last member here */
383 /* internal only representation about the status of kmem accounting. */
385 KMEM_ACCOUNTED_ACTIVE, /* accounted by this cgroup itself */
386 KMEM_ACCOUNTED_DEAD, /* dead memcg with pending kmem charges */
389 #ifdef CONFIG_MEMCG_KMEM
390 static inline void memcg_kmem_set_active(struct mem_cgroup *memcg)
392 set_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
395 static bool memcg_kmem_is_active(struct mem_cgroup *memcg)
397 return test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
400 static void memcg_kmem_mark_dead(struct mem_cgroup *memcg)
403 * Our caller must use css_get() first, because memcg_uncharge_kmem()
404 * will call css_put() if it sees the memcg is dead.
407 if (test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags))
408 set_bit(KMEM_ACCOUNTED_DEAD, &memcg->kmem_account_flags);
411 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup *memcg)
413 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD,
414 &memcg->kmem_account_flags);
418 /* Stuffs for move charges at task migration. */
420 * Types of charges to be moved. "move_charge_at_immitgrate" and
421 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
424 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
425 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
429 /* "mc" and its members are protected by cgroup_mutex */
430 static struct move_charge_struct {
431 spinlock_t lock; /* for from, to */
432 struct mem_cgroup *from;
433 struct mem_cgroup *to;
434 unsigned long immigrate_flags;
435 unsigned long precharge;
436 unsigned long moved_charge;
437 unsigned long moved_swap;
438 struct task_struct *moving_task; /* a task moving charges */
439 wait_queue_head_t waitq; /* a waitq for other context */
441 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
442 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
445 static bool move_anon(void)
447 return test_bit(MOVE_CHARGE_TYPE_ANON, &mc.immigrate_flags);
450 static bool move_file(void)
452 return test_bit(MOVE_CHARGE_TYPE_FILE, &mc.immigrate_flags);
456 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
457 * limit reclaim to prevent infinite loops, if they ever occur.
459 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
460 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
463 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
464 MEM_CGROUP_CHARGE_TYPE_ANON,
465 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
466 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
470 /* for encoding cft->private value on file */
478 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
479 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
480 #define MEMFILE_ATTR(val) ((val) & 0xffff)
481 /* Used for OOM nofiier */
482 #define OOM_CONTROL (0)
485 * Reclaim flags for mem_cgroup_hierarchical_reclaim
487 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
488 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
489 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
490 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
493 * The memcg_create_mutex will be held whenever a new cgroup is created.
494 * As a consequence, any change that needs to protect against new child cgroups
495 * appearing has to hold it as well.
497 static DEFINE_MUTEX(memcg_create_mutex);
499 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
501 return s ? container_of(s, struct mem_cgroup, css) : NULL;
504 /* Some nice accessors for the vmpressure. */
505 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
508 memcg = root_mem_cgroup;
509 return &memcg->vmpressure;
512 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
514 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
517 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
519 return (memcg == root_mem_cgroup);
523 * We restrict the id in the range of [1, 65535], so it can fit into
526 #define MEM_CGROUP_ID_MAX USHRT_MAX
528 static inline unsigned short mem_cgroup_id(struct mem_cgroup *memcg)
531 * The ID of the root cgroup is 0, but memcg treat 0 as an
532 * invalid ID, so we return (cgroup_id + 1).
534 return memcg->css.cgroup->id + 1;
537 static inline struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
539 struct cgroup_subsys_state *css;
541 css = css_from_id(id - 1, &mem_cgroup_subsys);
542 return mem_cgroup_from_css(css);
545 /* Writing them here to avoid exposing memcg's inner layout */
546 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
548 void sock_update_memcg(struct sock *sk)
550 if (mem_cgroup_sockets_enabled) {
551 struct mem_cgroup *memcg;
552 struct cg_proto *cg_proto;
554 BUG_ON(!sk->sk_prot->proto_cgroup);
556 /* Socket cloning can throw us here with sk_cgrp already
557 * filled. It won't however, necessarily happen from
558 * process context. So the test for root memcg given
559 * the current task's memcg won't help us in this case.
561 * Respecting the original socket's memcg is a better
562 * decision in this case.
565 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
566 css_get(&sk->sk_cgrp->memcg->css);
571 memcg = mem_cgroup_from_task(current);
572 cg_proto = sk->sk_prot->proto_cgroup(memcg);
573 if (!mem_cgroup_is_root(memcg) &&
574 memcg_proto_active(cg_proto) && css_tryget(&memcg->css)) {
575 sk->sk_cgrp = cg_proto;
580 EXPORT_SYMBOL(sock_update_memcg);
582 void sock_release_memcg(struct sock *sk)
584 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
585 struct mem_cgroup *memcg;
586 WARN_ON(!sk->sk_cgrp->memcg);
587 memcg = sk->sk_cgrp->memcg;
588 css_put(&sk->sk_cgrp->memcg->css);
592 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
594 if (!memcg || mem_cgroup_is_root(memcg))
597 return &memcg->tcp_mem;
599 EXPORT_SYMBOL(tcp_proto_cgroup);
601 static void disarm_sock_keys(struct mem_cgroup *memcg)
603 if (!memcg_proto_activated(&memcg->tcp_mem))
605 static_key_slow_dec(&memcg_socket_limit_enabled);
608 static void disarm_sock_keys(struct mem_cgroup *memcg)
613 #ifdef CONFIG_MEMCG_KMEM
615 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
616 * The main reason for not using cgroup id for this:
617 * this works better in sparse environments, where we have a lot of memcgs,
618 * but only a few kmem-limited. Or also, if we have, for instance, 200
619 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
620 * 200 entry array for that.
622 * The current size of the caches array is stored in
623 * memcg_limited_groups_array_size. It will double each time we have to
626 static DEFINE_IDA(kmem_limited_groups);
627 int memcg_limited_groups_array_size;
630 * MIN_SIZE is different than 1, because we would like to avoid going through
631 * the alloc/free process all the time. In a small machine, 4 kmem-limited
632 * cgroups is a reasonable guess. In the future, it could be a parameter or
633 * tunable, but that is strictly not necessary.
635 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
636 * this constant directly from cgroup, but it is understandable that this is
637 * better kept as an internal representation in cgroup.c. In any case, the
638 * cgrp_id space is not getting any smaller, and we don't have to necessarily
639 * increase ours as well if it increases.
641 #define MEMCG_CACHES_MIN_SIZE 4
642 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
645 * A lot of the calls to the cache allocation functions are expected to be
646 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
647 * conditional to this static branch, we'll have to allow modules that does
648 * kmem_cache_alloc and the such to see this symbol as well
650 struct static_key memcg_kmem_enabled_key;
651 EXPORT_SYMBOL(memcg_kmem_enabled_key);
653 static void disarm_kmem_keys(struct mem_cgroup *memcg)
655 if (memcg_kmem_is_active(memcg)) {
656 static_key_slow_dec(&memcg_kmem_enabled_key);
657 ida_simple_remove(&kmem_limited_groups, memcg->kmemcg_id);
660 * This check can't live in kmem destruction function,
661 * since the charges will outlive the cgroup
663 WARN_ON(res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0);
666 static void disarm_kmem_keys(struct mem_cgroup *memcg)
669 #endif /* CONFIG_MEMCG_KMEM */
671 static void disarm_static_keys(struct mem_cgroup *memcg)
673 disarm_sock_keys(memcg);
674 disarm_kmem_keys(memcg);
677 static void drain_all_stock_async(struct mem_cgroup *memcg);
679 static struct mem_cgroup_per_zone *
680 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
682 VM_BUG_ON((unsigned)nid >= nr_node_ids);
683 return &memcg->nodeinfo[nid]->zoneinfo[zid];
686 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
691 static struct mem_cgroup_per_zone *
692 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
694 int nid = page_to_nid(page);
695 int zid = page_zonenum(page);
697 return mem_cgroup_zoneinfo(memcg, nid, zid);
700 static struct mem_cgroup_tree_per_zone *
701 soft_limit_tree_node_zone(int nid, int zid)
703 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
706 static struct mem_cgroup_tree_per_zone *
707 soft_limit_tree_from_page(struct page *page)
709 int nid = page_to_nid(page);
710 int zid = page_zonenum(page);
712 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
716 __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
717 struct mem_cgroup_per_zone *mz,
718 struct mem_cgroup_tree_per_zone *mctz,
719 unsigned long long new_usage_in_excess)
721 struct rb_node **p = &mctz->rb_root.rb_node;
722 struct rb_node *parent = NULL;
723 struct mem_cgroup_per_zone *mz_node;
728 mz->usage_in_excess = new_usage_in_excess;
729 if (!mz->usage_in_excess)
733 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
735 if (mz->usage_in_excess < mz_node->usage_in_excess)
738 * We can't avoid mem cgroups that are over their soft
739 * limit by the same amount
741 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
744 rb_link_node(&mz->tree_node, parent, p);
745 rb_insert_color(&mz->tree_node, &mctz->rb_root);
750 __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
751 struct mem_cgroup_per_zone *mz,
752 struct mem_cgroup_tree_per_zone *mctz)
756 rb_erase(&mz->tree_node, &mctz->rb_root);
761 mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
762 struct mem_cgroup_per_zone *mz,
763 struct mem_cgroup_tree_per_zone *mctz)
765 spin_lock(&mctz->lock);
766 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
767 spin_unlock(&mctz->lock);
771 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
773 unsigned long long excess;
774 struct mem_cgroup_per_zone *mz;
775 struct mem_cgroup_tree_per_zone *mctz;
776 int nid = page_to_nid(page);
777 int zid = page_zonenum(page);
778 mctz = soft_limit_tree_from_page(page);
781 * Necessary to update all ancestors when hierarchy is used.
782 * because their event counter is not touched.
784 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
785 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
786 excess = res_counter_soft_limit_excess(&memcg->res);
788 * We have to update the tree if mz is on RB-tree or
789 * mem is over its softlimit.
791 if (excess || mz->on_tree) {
792 spin_lock(&mctz->lock);
793 /* if on-tree, remove it */
795 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
797 * Insert again. mz->usage_in_excess will be updated.
798 * If excess is 0, no tree ops.
800 __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
801 spin_unlock(&mctz->lock);
806 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
809 struct mem_cgroup_per_zone *mz;
810 struct mem_cgroup_tree_per_zone *mctz;
812 for_each_node(node) {
813 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
814 mz = mem_cgroup_zoneinfo(memcg, node, zone);
815 mctz = soft_limit_tree_node_zone(node, zone);
816 mem_cgroup_remove_exceeded(memcg, mz, mctz);
821 static struct mem_cgroup_per_zone *
822 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
824 struct rb_node *rightmost = NULL;
825 struct mem_cgroup_per_zone *mz;
829 rightmost = rb_last(&mctz->rb_root);
831 goto done; /* Nothing to reclaim from */
833 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
835 * Remove the node now but someone else can add it back,
836 * we will to add it back at the end of reclaim to its correct
837 * position in the tree.
839 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
840 if (!res_counter_soft_limit_excess(&mz->memcg->res) ||
841 !css_tryget(&mz->memcg->css))
847 static struct mem_cgroup_per_zone *
848 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
850 struct mem_cgroup_per_zone *mz;
852 spin_lock(&mctz->lock);
853 mz = __mem_cgroup_largest_soft_limit_node(mctz);
854 spin_unlock(&mctz->lock);
859 * Implementation Note: reading percpu statistics for memcg.
861 * Both of vmstat[] and percpu_counter has threshold and do periodic
862 * synchronization to implement "quick" read. There are trade-off between
863 * reading cost and precision of value. Then, we may have a chance to implement
864 * a periodic synchronizion of counter in memcg's counter.
866 * But this _read() function is used for user interface now. The user accounts
867 * memory usage by memory cgroup and he _always_ requires exact value because
868 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
869 * have to visit all online cpus and make sum. So, for now, unnecessary
870 * synchronization is not implemented. (just implemented for cpu hotplug)
872 * If there are kernel internal actions which can make use of some not-exact
873 * value, and reading all cpu value can be performance bottleneck in some
874 * common workload, threashold and synchonization as vmstat[] should be
877 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
878 enum mem_cgroup_stat_index idx)
884 for_each_online_cpu(cpu)
885 val += per_cpu(memcg->stat->count[idx], cpu);
886 #ifdef CONFIG_HOTPLUG_CPU
887 spin_lock(&memcg->pcp_counter_lock);
888 val += memcg->nocpu_base.count[idx];
889 spin_unlock(&memcg->pcp_counter_lock);
895 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
898 int val = (charge) ? 1 : -1;
899 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
902 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
903 enum mem_cgroup_events_index idx)
905 unsigned long val = 0;
909 for_each_online_cpu(cpu)
910 val += per_cpu(memcg->stat->events[idx], cpu);
911 #ifdef CONFIG_HOTPLUG_CPU
912 spin_lock(&memcg->pcp_counter_lock);
913 val += memcg->nocpu_base.events[idx];
914 spin_unlock(&memcg->pcp_counter_lock);
920 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
922 bool anon, int nr_pages)
927 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
928 * counted as CACHE even if it's on ANON LRU.
931 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
934 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
937 if (PageTransHuge(page))
938 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
941 /* pagein of a big page is an event. So, ignore page size */
943 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
945 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
946 nr_pages = -nr_pages; /* for event */
949 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
955 mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
957 struct mem_cgroup_per_zone *mz;
959 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
960 return mz->lru_size[lru];
964 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
965 unsigned int lru_mask)
967 struct mem_cgroup_per_zone *mz;
969 unsigned long ret = 0;
971 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
974 if (BIT(lru) & lru_mask)
975 ret += mz->lru_size[lru];
981 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
982 int nid, unsigned int lru_mask)
987 for (zid = 0; zid < MAX_NR_ZONES; zid++)
988 total += mem_cgroup_zone_nr_lru_pages(memcg,
994 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
995 unsigned int lru_mask)
1000 for_each_node_state(nid, N_MEMORY)
1001 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
1005 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
1006 enum mem_cgroup_events_target target)
1008 unsigned long val, next;
1010 val = __this_cpu_read(memcg->stat->nr_page_events);
1011 next = __this_cpu_read(memcg->stat->targets[target]);
1012 /* from time_after() in jiffies.h */
1013 if ((long)next - (long)val < 0) {
1015 case MEM_CGROUP_TARGET_THRESH:
1016 next = val + THRESHOLDS_EVENTS_TARGET;
1018 case MEM_CGROUP_TARGET_SOFTLIMIT:
1019 next = val + SOFTLIMIT_EVENTS_TARGET;
1021 case MEM_CGROUP_TARGET_NUMAINFO:
1022 next = val + NUMAINFO_EVENTS_TARGET;
1027 __this_cpu_write(memcg->stat->targets[target], next);
1034 * Check events in order.
1037 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
1040 /* threshold event is triggered in finer grain than soft limit */
1041 if (unlikely(mem_cgroup_event_ratelimit(memcg,
1042 MEM_CGROUP_TARGET_THRESH))) {
1044 bool do_numainfo __maybe_unused;
1046 do_softlimit = mem_cgroup_event_ratelimit(memcg,
1047 MEM_CGROUP_TARGET_SOFTLIMIT);
1048 #if MAX_NUMNODES > 1
1049 do_numainfo = mem_cgroup_event_ratelimit(memcg,
1050 MEM_CGROUP_TARGET_NUMAINFO);
1054 mem_cgroup_threshold(memcg);
1055 if (unlikely(do_softlimit))
1056 mem_cgroup_update_tree(memcg, page);
1057 #if MAX_NUMNODES > 1
1058 if (unlikely(do_numainfo))
1059 atomic_inc(&memcg->numainfo_events);
1065 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1068 * mm_update_next_owner() may clear mm->owner to NULL
1069 * if it races with swapoff, page migration, etc.
1070 * So this can be called with p == NULL.
1075 return mem_cgroup_from_css(task_css(p, mem_cgroup_subsys_id));
1078 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
1080 struct mem_cgroup *memcg = NULL;
1085 * Because we have no locks, mm->owner's may be being moved to other
1086 * cgroup. We use css_tryget() here even if this looks
1087 * pessimistic (rather than adding locks here).
1091 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1092 if (unlikely(!memcg))
1094 } while (!css_tryget(&memcg->css));
1100 * Returns a next (in a pre-order walk) alive memcg (with elevated css
1101 * ref. count) or NULL if the whole root's subtree has been visited.
1103 * helper function to be used by mem_cgroup_iter
1105 static struct mem_cgroup *__mem_cgroup_iter_next(struct mem_cgroup *root,
1106 struct mem_cgroup *last_visited)
1108 struct cgroup_subsys_state *prev_css, *next_css;
1110 prev_css = last_visited ? &last_visited->css : NULL;
1112 next_css = css_next_descendant_pre(prev_css, &root->css);
1115 * Even if we found a group we have to make sure it is
1116 * alive. css && !memcg means that the groups should be
1117 * skipped and we should continue the tree walk.
1118 * last_visited css is safe to use because it is
1119 * protected by css_get and the tree walk is rcu safe.
1122 if ((next_css->flags & CSS_ONLINE) && css_tryget(next_css))
1123 return mem_cgroup_from_css(next_css);
1125 prev_css = next_css;
1133 static void mem_cgroup_iter_invalidate(struct mem_cgroup *root)
1136 * When a group in the hierarchy below root is destroyed, the
1137 * hierarchy iterator can no longer be trusted since it might
1138 * have pointed to the destroyed group. Invalidate it.
1140 atomic_inc(&root->dead_count);
1143 static struct mem_cgroup *
1144 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter *iter,
1145 struct mem_cgroup *root,
1148 struct mem_cgroup *position = NULL;
1150 * A cgroup destruction happens in two stages: offlining and
1151 * release. They are separated by a RCU grace period.
1153 * If the iterator is valid, we may still race with an
1154 * offlining. The RCU lock ensures the object won't be
1155 * released, tryget will fail if we lost the race.
1157 *sequence = atomic_read(&root->dead_count);
1158 if (iter->last_dead_count == *sequence) {
1160 position = iter->last_visited;
1163 * We cannot take a reference to root because we might race
1164 * with root removal and returning NULL would end up in
1165 * an endless loop on the iterator user level when root
1166 * would be returned all the time.
1168 if (position && position != root &&
1169 !css_tryget(&position->css))
1175 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter *iter,
1176 struct mem_cgroup *last_visited,
1177 struct mem_cgroup *new_position,
1178 struct mem_cgroup *root,
1181 /* root reference counting symmetric to mem_cgroup_iter_load */
1182 if (last_visited && last_visited != root)
1183 css_put(&last_visited->css);
1185 * We store the sequence count from the time @last_visited was
1186 * loaded successfully instead of rereading it here so that we
1187 * don't lose destruction events in between. We could have
1188 * raced with the destruction of @new_position after all.
1190 iter->last_visited = new_position;
1192 iter->last_dead_count = sequence;
1196 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1197 * @root: hierarchy root
1198 * @prev: previously returned memcg, NULL on first invocation
1199 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1201 * Returns references to children of the hierarchy below @root, or
1202 * @root itself, or %NULL after a full round-trip.
1204 * Caller must pass the return value in @prev on subsequent
1205 * invocations for reference counting, or use mem_cgroup_iter_break()
1206 * to cancel a hierarchy walk before the round-trip is complete.
1208 * Reclaimers can specify a zone and a priority level in @reclaim to
1209 * divide up the memcgs in the hierarchy among all concurrent
1210 * reclaimers operating on the same zone and priority.
1212 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1213 struct mem_cgroup *prev,
1214 struct mem_cgroup_reclaim_cookie *reclaim)
1216 struct mem_cgroup *memcg = NULL;
1217 struct mem_cgroup *last_visited = NULL;
1219 if (mem_cgroup_disabled())
1223 root = root_mem_cgroup;
1225 if (prev && !reclaim)
1226 last_visited = prev;
1228 if (!root->use_hierarchy && root != root_mem_cgroup) {
1236 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1237 int uninitialized_var(seq);
1240 int nid = zone_to_nid(reclaim->zone);
1241 int zid = zone_idx(reclaim->zone);
1242 struct mem_cgroup_per_zone *mz;
1244 mz = mem_cgroup_zoneinfo(root, nid, zid);
1245 iter = &mz->reclaim_iter[reclaim->priority];
1246 if (prev && reclaim->generation != iter->generation) {
1247 iter->last_visited = NULL;
1251 last_visited = mem_cgroup_iter_load(iter, root, &seq);
1254 memcg = __mem_cgroup_iter_next(root, last_visited);
1257 mem_cgroup_iter_update(iter, last_visited, memcg, root,
1262 else if (!prev && memcg)
1263 reclaim->generation = iter->generation;
1272 if (prev && prev != root)
1273 css_put(&prev->css);
1279 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1280 * @root: hierarchy root
1281 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1283 void mem_cgroup_iter_break(struct mem_cgroup *root,
1284 struct mem_cgroup *prev)
1287 root = root_mem_cgroup;
1288 if (prev && prev != root)
1289 css_put(&prev->css);
1293 * Iteration constructs for visiting all cgroups (under a tree). If
1294 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1295 * be used for reference counting.
1297 #define for_each_mem_cgroup_tree(iter, root) \
1298 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1300 iter = mem_cgroup_iter(root, iter, NULL))
1302 #define for_each_mem_cgroup(iter) \
1303 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1305 iter = mem_cgroup_iter(NULL, iter, NULL))
1307 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1309 struct mem_cgroup *memcg;
1312 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1313 if (unlikely(!memcg))
1318 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1321 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1329 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1332 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1333 * @zone: zone of the wanted lruvec
1334 * @memcg: memcg of the wanted lruvec
1336 * Returns the lru list vector holding pages for the given @zone and
1337 * @mem. This can be the global zone lruvec, if the memory controller
1340 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1341 struct mem_cgroup *memcg)
1343 struct mem_cgroup_per_zone *mz;
1344 struct lruvec *lruvec;
1346 if (mem_cgroup_disabled()) {
1347 lruvec = &zone->lruvec;
1351 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1352 lruvec = &mz->lruvec;
1355 * Since a node can be onlined after the mem_cgroup was created,
1356 * we have to be prepared to initialize lruvec->zone here;
1357 * and if offlined then reonlined, we need to reinitialize it.
1359 if (unlikely(lruvec->zone != zone))
1360 lruvec->zone = zone;
1365 * Following LRU functions are allowed to be used without PCG_LOCK.
1366 * Operations are called by routine of global LRU independently from memcg.
1367 * What we have to take care of here is validness of pc->mem_cgroup.
1369 * Changes to pc->mem_cgroup happens when
1372 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1373 * It is added to LRU before charge.
1374 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1375 * When moving account, the page is not on LRU. It's isolated.
1379 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1381 * @zone: zone of the page
1383 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1385 struct mem_cgroup_per_zone *mz;
1386 struct mem_cgroup *memcg;
1387 struct page_cgroup *pc;
1388 struct lruvec *lruvec;
1390 if (mem_cgroup_disabled()) {
1391 lruvec = &zone->lruvec;
1395 pc = lookup_page_cgroup(page);
1396 memcg = pc->mem_cgroup;
1399 * Surreptitiously switch any uncharged offlist page to root:
1400 * an uncharged page off lru does nothing to secure
1401 * its former mem_cgroup from sudden removal.
1403 * Our caller holds lru_lock, and PageCgroupUsed is updated
1404 * under page_cgroup lock: between them, they make all uses
1405 * of pc->mem_cgroup safe.
1407 if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1408 pc->mem_cgroup = memcg = root_mem_cgroup;
1410 mz = page_cgroup_zoneinfo(memcg, page);
1411 lruvec = &mz->lruvec;
1414 * Since a node can be onlined after the mem_cgroup was created,
1415 * we have to be prepared to initialize lruvec->zone here;
1416 * and if offlined then reonlined, we need to reinitialize it.
1418 if (unlikely(lruvec->zone != zone))
1419 lruvec->zone = zone;
1424 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1425 * @lruvec: mem_cgroup per zone lru vector
1426 * @lru: index of lru list the page is sitting on
1427 * @nr_pages: positive when adding or negative when removing
1429 * This function must be called when a page is added to or removed from an
1432 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1435 struct mem_cgroup_per_zone *mz;
1436 unsigned long *lru_size;
1438 if (mem_cgroup_disabled())
1441 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1442 lru_size = mz->lru_size + lru;
1443 *lru_size += nr_pages;
1444 VM_BUG_ON((long)(*lru_size) < 0);
1448 * Checks whether given mem is same or in the root_mem_cgroup's
1451 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1452 struct mem_cgroup *memcg)
1454 if (root_memcg == memcg)
1456 if (!root_memcg->use_hierarchy || !memcg)
1458 return cgroup_is_descendant(memcg->css.cgroup, root_memcg->css.cgroup);
1461 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1462 struct mem_cgroup *memcg)
1467 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1472 bool task_in_mem_cgroup(struct task_struct *task,
1473 const struct mem_cgroup *memcg)
1475 struct mem_cgroup *curr = NULL;
1476 struct task_struct *p;
1479 p = find_lock_task_mm(task);
1481 curr = try_get_mem_cgroup_from_mm(p->mm);
1485 * All threads may have already detached their mm's, but the oom
1486 * killer still needs to detect if they have already been oom
1487 * killed to prevent needlessly killing additional tasks.
1490 curr = mem_cgroup_from_task(task);
1492 css_get(&curr->css);
1498 * We should check use_hierarchy of "memcg" not "curr". Because checking
1499 * use_hierarchy of "curr" here make this function true if hierarchy is
1500 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1501 * hierarchy(even if use_hierarchy is disabled in "memcg").
1503 ret = mem_cgroup_same_or_subtree(memcg, curr);
1504 css_put(&curr->css);
1508 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1510 unsigned long inactive_ratio;
1511 unsigned long inactive;
1512 unsigned long active;
1515 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1516 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1518 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1520 inactive_ratio = int_sqrt(10 * gb);
1524 return inactive * inactive_ratio < active;
1527 #define mem_cgroup_from_res_counter(counter, member) \
1528 container_of(counter, struct mem_cgroup, member)
1531 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1532 * @memcg: the memory cgroup
1534 * Returns the maximum amount of memory @mem can be charged with, in
1537 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1539 unsigned long long margin;
1541 margin = res_counter_margin(&memcg->res);
1542 if (do_swap_account)
1543 margin = min(margin, res_counter_margin(&memcg->memsw));
1544 return margin >> PAGE_SHIFT;
1547 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1550 if (!css_parent(&memcg->css))
1551 return vm_swappiness;
1553 return memcg->swappiness;
1557 * memcg->moving_account is used for checking possibility that some thread is
1558 * calling move_account(). When a thread on CPU-A starts moving pages under
1559 * a memcg, other threads should check memcg->moving_account under
1560 * rcu_read_lock(), like this:
1564 * memcg->moving_account+1 if (memcg->mocing_account)
1566 * synchronize_rcu() update something.
1571 /* for quick checking without looking up memcg */
1572 atomic_t memcg_moving __read_mostly;
1574 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1576 atomic_inc(&memcg_moving);
1577 atomic_inc(&memcg->moving_account);
1581 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1584 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1585 * We check NULL in callee rather than caller.
1588 atomic_dec(&memcg_moving);
1589 atomic_dec(&memcg->moving_account);
1594 * 2 routines for checking "mem" is under move_account() or not.
1596 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1597 * is used for avoiding races in accounting. If true,
1598 * pc->mem_cgroup may be overwritten.
1600 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1601 * under hierarchy of moving cgroups. This is for
1602 * waiting at hith-memory prressure caused by "move".
1605 static bool mem_cgroup_stolen(struct mem_cgroup *memcg)
1607 VM_BUG_ON(!rcu_read_lock_held());
1608 return atomic_read(&memcg->moving_account) > 0;
1611 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1613 struct mem_cgroup *from;
1614 struct mem_cgroup *to;
1617 * Unlike task_move routines, we access mc.to, mc.from not under
1618 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1620 spin_lock(&mc.lock);
1626 ret = mem_cgroup_same_or_subtree(memcg, from)
1627 || mem_cgroup_same_or_subtree(memcg, to);
1629 spin_unlock(&mc.lock);
1633 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1635 if (mc.moving_task && current != mc.moving_task) {
1636 if (mem_cgroup_under_move(memcg)) {
1638 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1639 /* moving charge context might have finished. */
1642 finish_wait(&mc.waitq, &wait);
1650 * Take this lock when
1651 * - a code tries to modify page's memcg while it's USED.
1652 * - a code tries to modify page state accounting in a memcg.
1653 * see mem_cgroup_stolen(), too.
1655 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1656 unsigned long *flags)
1658 spin_lock_irqsave(&memcg->move_lock, *flags);
1661 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1662 unsigned long *flags)
1664 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1667 #define K(x) ((x) << (PAGE_SHIFT-10))
1669 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1670 * @memcg: The memory cgroup that went over limit
1671 * @p: Task that is going to be killed
1673 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1676 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1679 * protects memcg_name and makes sure that parallel ooms do not
1682 static DEFINE_SPINLOCK(oom_info_lock);
1683 struct cgroup *task_cgrp;
1684 struct cgroup *mem_cgrp;
1685 static char memcg_name[PATH_MAX];
1687 struct mem_cgroup *iter;
1693 spin_lock(&oom_info_lock);
1696 mem_cgrp = memcg->css.cgroup;
1697 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1699 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1702 * Unfortunately, we are unable to convert to a useful name
1703 * But we'll still print out the usage information
1710 pr_info("Task in %s killed", memcg_name);
1713 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1721 * Continues from above, so we don't need an KERN_ level
1723 pr_cont(" as a result of limit of %s\n", memcg_name);
1726 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1727 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1728 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1729 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1730 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1731 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1732 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1733 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1734 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1735 res_counter_read_u64(&memcg->kmem, RES_USAGE) >> 10,
1736 res_counter_read_u64(&memcg->kmem, RES_LIMIT) >> 10,
1737 res_counter_read_u64(&memcg->kmem, RES_FAILCNT));
1739 for_each_mem_cgroup_tree(iter, memcg) {
1740 pr_info("Memory cgroup stats");
1743 ret = cgroup_path(iter->css.cgroup, memcg_name, PATH_MAX);
1745 pr_cont(" for %s", memcg_name);
1749 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1750 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1752 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1753 K(mem_cgroup_read_stat(iter, i)));
1756 for (i = 0; i < NR_LRU_LISTS; i++)
1757 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1758 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1762 spin_unlock(&oom_info_lock);
1766 * This function returns the number of memcg under hierarchy tree. Returns
1767 * 1(self count) if no children.
1769 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1772 struct mem_cgroup *iter;
1774 for_each_mem_cgroup_tree(iter, memcg)
1780 * Return the memory (and swap, if configured) limit for a memcg.
1782 static u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1786 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1789 * Do not consider swap space if we cannot swap due to swappiness
1791 if (mem_cgroup_swappiness(memcg)) {
1794 limit += total_swap_pages << PAGE_SHIFT;
1795 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1798 * If memsw is finite and limits the amount of swap space
1799 * available to this memcg, return that limit.
1801 limit = min(limit, memsw);
1807 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1810 struct mem_cgroup *iter;
1811 unsigned long chosen_points = 0;
1812 unsigned long totalpages;
1813 unsigned int points = 0;
1814 struct task_struct *chosen = NULL;
1817 * If current has a pending SIGKILL or is exiting, then automatically
1818 * select it. The goal is to allow it to allocate so that it may
1819 * quickly exit and free its memory.
1821 if (fatal_signal_pending(current) || current->flags & PF_EXITING) {
1822 set_thread_flag(TIF_MEMDIE);
1826 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1827 totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
1828 for_each_mem_cgroup_tree(iter, memcg) {
1829 struct css_task_iter it;
1830 struct task_struct *task;
1832 css_task_iter_start(&iter->css, &it);
1833 while ((task = css_task_iter_next(&it))) {
1834 switch (oom_scan_process_thread(task, totalpages, NULL,
1836 case OOM_SCAN_SELECT:
1838 put_task_struct(chosen);
1840 chosen_points = ULONG_MAX;
1841 get_task_struct(chosen);
1843 case OOM_SCAN_CONTINUE:
1845 case OOM_SCAN_ABORT:
1846 css_task_iter_end(&it);
1847 mem_cgroup_iter_break(memcg, iter);
1849 put_task_struct(chosen);
1854 points = oom_badness(task, memcg, NULL, totalpages);
1855 if (!points || points < chosen_points)
1857 /* Prefer thread group leaders for display purposes */
1858 if (points == chosen_points &&
1859 thread_group_leader(chosen))
1863 put_task_struct(chosen);
1865 chosen_points = points;
1866 get_task_struct(chosen);
1868 css_task_iter_end(&it);
1873 points = chosen_points * 1000 / totalpages;
1874 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1875 NULL, "Memory cgroup out of memory");
1878 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1880 unsigned long flags)
1882 unsigned long total = 0;
1883 bool noswap = false;
1886 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1888 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1891 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1893 drain_all_stock_async(memcg);
1894 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1896 * Allow limit shrinkers, which are triggered directly
1897 * by userspace, to catch signals and stop reclaim
1898 * after minimal progress, regardless of the margin.
1900 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1902 if (mem_cgroup_margin(memcg))
1905 * If nothing was reclaimed after two attempts, there
1906 * may be no reclaimable pages in this hierarchy.
1915 * test_mem_cgroup_node_reclaimable
1916 * @memcg: the target memcg
1917 * @nid: the node ID to be checked.
1918 * @noswap : specify true here if the user wants flle only information.
1920 * This function returns whether the specified memcg contains any
1921 * reclaimable pages on a node. Returns true if there are any reclaimable
1922 * pages in the node.
1924 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1925 int nid, bool noswap)
1927 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1929 if (noswap || !total_swap_pages)
1931 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1936 #if MAX_NUMNODES > 1
1939 * Always updating the nodemask is not very good - even if we have an empty
1940 * list or the wrong list here, we can start from some node and traverse all
1941 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1944 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1948 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1949 * pagein/pageout changes since the last update.
1951 if (!atomic_read(&memcg->numainfo_events))
1953 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1956 /* make a nodemask where this memcg uses memory from */
1957 memcg->scan_nodes = node_states[N_MEMORY];
1959 for_each_node_mask(nid, node_states[N_MEMORY]) {
1961 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1962 node_clear(nid, memcg->scan_nodes);
1965 atomic_set(&memcg->numainfo_events, 0);
1966 atomic_set(&memcg->numainfo_updating, 0);
1970 * Selecting a node where we start reclaim from. Because what we need is just
1971 * reducing usage counter, start from anywhere is O,K. Considering
1972 * memory reclaim from current node, there are pros. and cons.
1974 * Freeing memory from current node means freeing memory from a node which
1975 * we'll use or we've used. So, it may make LRU bad. And if several threads
1976 * hit limits, it will see a contention on a node. But freeing from remote
1977 * node means more costs for memory reclaim because of memory latency.
1979 * Now, we use round-robin. Better algorithm is welcomed.
1981 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1985 mem_cgroup_may_update_nodemask(memcg);
1986 node = memcg->last_scanned_node;
1988 node = next_node(node, memcg->scan_nodes);
1989 if (node == MAX_NUMNODES)
1990 node = first_node(memcg->scan_nodes);
1992 * We call this when we hit limit, not when pages are added to LRU.
1993 * No LRU may hold pages because all pages are UNEVICTABLE or
1994 * memcg is too small and all pages are not on LRU. In that case,
1995 * we use curret node.
1997 if (unlikely(node == MAX_NUMNODES))
1998 node = numa_node_id();
2000 memcg->last_scanned_node = node;
2005 * Check all nodes whether it contains reclaimable pages or not.
2006 * For quick scan, we make use of scan_nodes. This will allow us to skip
2007 * unused nodes. But scan_nodes is lazily updated and may not cotain
2008 * enough new information. We need to do double check.
2010 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
2015 * quick check...making use of scan_node.
2016 * We can skip unused nodes.
2018 if (!nodes_empty(memcg->scan_nodes)) {
2019 for (nid = first_node(memcg->scan_nodes);
2021 nid = next_node(nid, memcg->scan_nodes)) {
2023 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
2028 * Check rest of nodes.
2030 for_each_node_state(nid, N_MEMORY) {
2031 if (node_isset(nid, memcg->scan_nodes))
2033 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
2040 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
2045 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
2047 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
2051 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
2054 unsigned long *total_scanned)
2056 struct mem_cgroup *victim = NULL;
2059 unsigned long excess;
2060 unsigned long nr_scanned;
2061 struct mem_cgroup_reclaim_cookie reclaim = {
2066 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
2069 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
2074 * If we have not been able to reclaim
2075 * anything, it might because there are
2076 * no reclaimable pages under this hierarchy
2081 * We want to do more targeted reclaim.
2082 * excess >> 2 is not to excessive so as to
2083 * reclaim too much, nor too less that we keep
2084 * coming back to reclaim from this cgroup
2086 if (total >= (excess >> 2) ||
2087 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
2092 if (!mem_cgroup_reclaimable(victim, false))
2094 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
2096 *total_scanned += nr_scanned;
2097 if (!res_counter_soft_limit_excess(&root_memcg->res))
2100 mem_cgroup_iter_break(root_memcg, victim);
2104 #ifdef CONFIG_LOCKDEP
2105 static struct lockdep_map memcg_oom_lock_dep_map = {
2106 .name = "memcg_oom_lock",
2110 static DEFINE_SPINLOCK(memcg_oom_lock);
2113 * Check OOM-Killer is already running under our hierarchy.
2114 * If someone is running, return false.
2116 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
2118 struct mem_cgroup *iter, *failed = NULL;
2120 spin_lock(&memcg_oom_lock);
2122 for_each_mem_cgroup_tree(iter, memcg) {
2123 if (iter->oom_lock) {
2125 * this subtree of our hierarchy is already locked
2126 * so we cannot give a lock.
2129 mem_cgroup_iter_break(memcg, iter);
2132 iter->oom_lock = true;
2137 * OK, we failed to lock the whole subtree so we have
2138 * to clean up what we set up to the failing subtree
2140 for_each_mem_cgroup_tree(iter, memcg) {
2141 if (iter == failed) {
2142 mem_cgroup_iter_break(memcg, iter);
2145 iter->oom_lock = false;
2148 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
2150 spin_unlock(&memcg_oom_lock);
2155 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
2157 struct mem_cgroup *iter;
2159 spin_lock(&memcg_oom_lock);
2160 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
2161 for_each_mem_cgroup_tree(iter, memcg)
2162 iter->oom_lock = false;
2163 spin_unlock(&memcg_oom_lock);
2166 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
2168 struct mem_cgroup *iter;
2170 for_each_mem_cgroup_tree(iter, memcg)
2171 atomic_inc(&iter->under_oom);
2174 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
2176 struct mem_cgroup *iter;
2179 * When a new child is created while the hierarchy is under oom,
2180 * mem_cgroup_oom_lock() may not be called. We have to use
2181 * atomic_add_unless() here.
2183 for_each_mem_cgroup_tree(iter, memcg)
2184 atomic_add_unless(&iter->under_oom, -1, 0);
2187 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
2189 struct oom_wait_info {
2190 struct mem_cgroup *memcg;
2194 static int memcg_oom_wake_function(wait_queue_t *wait,
2195 unsigned mode, int sync, void *arg)
2197 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
2198 struct mem_cgroup *oom_wait_memcg;
2199 struct oom_wait_info *oom_wait_info;
2201 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
2202 oom_wait_memcg = oom_wait_info->memcg;
2205 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2206 * Then we can use css_is_ancestor without taking care of RCU.
2208 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
2209 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
2211 return autoremove_wake_function(wait, mode, sync, arg);
2214 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
2216 atomic_inc(&memcg->oom_wakeups);
2217 /* for filtering, pass "memcg" as argument. */
2218 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
2221 static void memcg_oom_recover(struct mem_cgroup *memcg)
2223 if (memcg && atomic_read(&memcg->under_oom))
2224 memcg_wakeup_oom(memcg);
2227 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
2229 if (!current->memcg_oom.may_oom)
2232 * We are in the middle of the charge context here, so we
2233 * don't want to block when potentially sitting on a callstack
2234 * that holds all kinds of filesystem and mm locks.
2236 * Also, the caller may handle a failed allocation gracefully
2237 * (like optional page cache readahead) and so an OOM killer
2238 * invocation might not even be necessary.
2240 * That's why we don't do anything here except remember the
2241 * OOM context and then deal with it at the end of the page
2242 * fault when the stack is unwound, the locks are released,
2243 * and when we know whether the fault was overall successful.
2245 css_get(&memcg->css);
2246 current->memcg_oom.memcg = memcg;
2247 current->memcg_oom.gfp_mask = mask;
2248 current->memcg_oom.order = order;
2252 * mem_cgroup_oom_synchronize - complete memcg OOM handling
2253 * @handle: actually kill/wait or just clean up the OOM state
2255 * This has to be called at the end of a page fault if the memcg OOM
2256 * handler was enabled.
2258 * Memcg supports userspace OOM handling where failed allocations must
2259 * sleep on a waitqueue until the userspace task resolves the
2260 * situation. Sleeping directly in the charge context with all kinds
2261 * of locks held is not a good idea, instead we remember an OOM state
2262 * in the task and mem_cgroup_oom_synchronize() has to be called at
2263 * the end of the page fault to complete the OOM handling.
2265 * Returns %true if an ongoing memcg OOM situation was detected and
2266 * completed, %false otherwise.
2268 bool mem_cgroup_oom_synchronize(bool handle)
2270 struct mem_cgroup *memcg = current->memcg_oom.memcg;
2271 struct oom_wait_info owait;
2274 /* OOM is global, do not handle */
2281 owait.memcg = memcg;
2282 owait.wait.flags = 0;
2283 owait.wait.func = memcg_oom_wake_function;
2284 owait.wait.private = current;
2285 INIT_LIST_HEAD(&owait.wait.task_list);
2287 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2288 mem_cgroup_mark_under_oom(memcg);
2290 locked = mem_cgroup_oom_trylock(memcg);
2293 mem_cgroup_oom_notify(memcg);
2295 if (locked && !memcg->oom_kill_disable) {
2296 mem_cgroup_unmark_under_oom(memcg);
2297 finish_wait(&memcg_oom_waitq, &owait.wait);
2298 mem_cgroup_out_of_memory(memcg, current->memcg_oom.gfp_mask,
2299 current->memcg_oom.order);
2302 mem_cgroup_unmark_under_oom(memcg);
2303 finish_wait(&memcg_oom_waitq, &owait.wait);
2307 mem_cgroup_oom_unlock(memcg);
2309 * There is no guarantee that an OOM-lock contender
2310 * sees the wakeups triggered by the OOM kill
2311 * uncharges. Wake any sleepers explicitely.
2313 memcg_oom_recover(memcg);
2316 current->memcg_oom.memcg = NULL;
2317 css_put(&memcg->css);
2322 * Currently used to update mapped file statistics, but the routine can be
2323 * generalized to update other statistics as well.
2325 * Notes: Race condition
2327 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2328 * it tends to be costly. But considering some conditions, we doesn't need
2329 * to do so _always_.
2331 * Considering "charge", lock_page_cgroup() is not required because all
2332 * file-stat operations happen after a page is attached to radix-tree. There
2333 * are no race with "charge".
2335 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2336 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2337 * if there are race with "uncharge". Statistics itself is properly handled
2340 * Considering "move", this is an only case we see a race. To make the race
2341 * small, we check mm->moving_account and detect there are possibility of race
2342 * If there is, we take a lock.
2345 void __mem_cgroup_begin_update_page_stat(struct page *page,
2346 bool *locked, unsigned long *flags)
2348 struct mem_cgroup *memcg;
2349 struct page_cgroup *pc;
2351 pc = lookup_page_cgroup(page);
2353 memcg = pc->mem_cgroup;
2354 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2357 * If this memory cgroup is not under account moving, we don't
2358 * need to take move_lock_mem_cgroup(). Because we already hold
2359 * rcu_read_lock(), any calls to move_account will be delayed until
2360 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2362 if (!mem_cgroup_stolen(memcg))
2365 move_lock_mem_cgroup(memcg, flags);
2366 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2367 move_unlock_mem_cgroup(memcg, flags);
2373 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
2375 struct page_cgroup *pc = lookup_page_cgroup(page);
2378 * It's guaranteed that pc->mem_cgroup never changes while
2379 * lock is held because a routine modifies pc->mem_cgroup
2380 * should take move_lock_mem_cgroup().
2382 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
2385 void mem_cgroup_update_page_stat(struct page *page,
2386 enum mem_cgroup_stat_index idx, int val)
2388 struct mem_cgroup *memcg;
2389 struct page_cgroup *pc = lookup_page_cgroup(page);
2390 unsigned long uninitialized_var(flags);
2392 if (mem_cgroup_disabled())
2395 VM_BUG_ON(!rcu_read_lock_held());
2396 memcg = pc->mem_cgroup;
2397 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2400 this_cpu_add(memcg->stat->count[idx], val);
2404 * size of first charge trial. "32" comes from vmscan.c's magic value.
2405 * TODO: maybe necessary to use big numbers in big irons.
2407 #define CHARGE_BATCH 32U
2408 struct memcg_stock_pcp {
2409 struct mem_cgroup *cached; /* this never be root cgroup */
2410 unsigned int nr_pages;
2411 struct work_struct work;
2412 unsigned long flags;
2413 #define FLUSHING_CACHED_CHARGE 0
2415 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2416 static DEFINE_MUTEX(percpu_charge_mutex);
2419 * consume_stock: Try to consume stocked charge on this cpu.
2420 * @memcg: memcg to consume from.
2421 * @nr_pages: how many pages to charge.
2423 * The charges will only happen if @memcg matches the current cpu's memcg
2424 * stock, and at least @nr_pages are available in that stock. Failure to
2425 * service an allocation will refill the stock.
2427 * returns true if successful, false otherwise.
2429 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2431 struct memcg_stock_pcp *stock;
2434 if (nr_pages > CHARGE_BATCH)
2437 stock = &get_cpu_var(memcg_stock);
2438 if (memcg == stock->cached && stock->nr_pages >= nr_pages)
2439 stock->nr_pages -= nr_pages;
2440 else /* need to call res_counter_charge */
2442 put_cpu_var(memcg_stock);
2447 * Returns stocks cached in percpu to res_counter and reset cached information.
2449 static void drain_stock(struct memcg_stock_pcp *stock)
2451 struct mem_cgroup *old = stock->cached;
2453 if (stock->nr_pages) {
2454 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2456 res_counter_uncharge(&old->res, bytes);
2457 if (do_swap_account)
2458 res_counter_uncharge(&old->memsw, bytes);
2459 stock->nr_pages = 0;
2461 stock->cached = NULL;
2465 * This must be called under preempt disabled or must be called by
2466 * a thread which is pinned to local cpu.
2468 static void drain_local_stock(struct work_struct *dummy)
2470 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2472 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2475 static void __init memcg_stock_init(void)
2479 for_each_possible_cpu(cpu) {
2480 struct memcg_stock_pcp *stock =
2481 &per_cpu(memcg_stock, cpu);
2482 INIT_WORK(&stock->work, drain_local_stock);
2487 * Cache charges(val) which is from res_counter, to local per_cpu area.
2488 * This will be consumed by consume_stock() function, later.
2490 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2492 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2494 if (stock->cached != memcg) { /* reset if necessary */
2496 stock->cached = memcg;
2498 stock->nr_pages += nr_pages;
2499 put_cpu_var(memcg_stock);
2503 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2504 * of the hierarchy under it. sync flag says whether we should block
2505 * until the work is done.
2507 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2511 /* Notify other cpus that system-wide "drain" is running */
2514 for_each_online_cpu(cpu) {
2515 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2516 struct mem_cgroup *memcg;
2518 memcg = stock->cached;
2519 if (!memcg || !stock->nr_pages)
2521 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2523 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2525 drain_local_stock(&stock->work);
2527 schedule_work_on(cpu, &stock->work);
2535 for_each_online_cpu(cpu) {
2536 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2537 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2538 flush_work(&stock->work);
2545 * Tries to drain stocked charges in other cpus. This function is asynchronous
2546 * and just put a work per cpu for draining localy on each cpu. Caller can
2547 * expects some charges will be back to res_counter later but cannot wait for
2550 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2553 * If someone calls draining, avoid adding more kworker runs.
2555 if (!mutex_trylock(&percpu_charge_mutex))
2557 drain_all_stock(root_memcg, false);
2558 mutex_unlock(&percpu_charge_mutex);
2561 /* This is a synchronous drain interface. */
2562 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2564 /* called when force_empty is called */
2565 mutex_lock(&percpu_charge_mutex);
2566 drain_all_stock(root_memcg, true);
2567 mutex_unlock(&percpu_charge_mutex);
2571 * This function drains percpu counter value from DEAD cpu and
2572 * move it to local cpu. Note that this function can be preempted.
2574 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2578 spin_lock(&memcg->pcp_counter_lock);
2579 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2580 long x = per_cpu(memcg->stat->count[i], cpu);
2582 per_cpu(memcg->stat->count[i], cpu) = 0;
2583 memcg->nocpu_base.count[i] += x;
2585 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2586 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2588 per_cpu(memcg->stat->events[i], cpu) = 0;
2589 memcg->nocpu_base.events[i] += x;
2591 spin_unlock(&memcg->pcp_counter_lock);
2594 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
2595 unsigned long action,
2598 int cpu = (unsigned long)hcpu;
2599 struct memcg_stock_pcp *stock;
2600 struct mem_cgroup *iter;
2602 if (action == CPU_ONLINE)
2605 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2608 for_each_mem_cgroup(iter)
2609 mem_cgroup_drain_pcp_counter(iter, cpu);
2611 stock = &per_cpu(memcg_stock, cpu);
2617 /* See __mem_cgroup_try_charge() for details */
2619 CHARGE_OK, /* success */
2620 CHARGE_RETRY, /* need to retry but retry is not bad */
2621 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2622 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2625 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2626 unsigned int nr_pages, unsigned int min_pages,
2629 unsigned long csize = nr_pages * PAGE_SIZE;
2630 struct mem_cgroup *mem_over_limit;
2631 struct res_counter *fail_res;
2632 unsigned long flags = 0;
2635 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2638 if (!do_swap_account)
2640 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2644 res_counter_uncharge(&memcg->res, csize);
2645 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2646 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2648 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2650 * Never reclaim on behalf of optional batching, retry with a
2651 * single page instead.
2653 if (nr_pages > min_pages)
2654 return CHARGE_RETRY;
2656 if (!(gfp_mask & __GFP_WAIT))
2657 return CHARGE_WOULDBLOCK;
2659 if (gfp_mask & __GFP_NORETRY)
2660 return CHARGE_NOMEM;
2662 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2663 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2664 return CHARGE_RETRY;
2666 * Even though the limit is exceeded at this point, reclaim
2667 * may have been able to free some pages. Retry the charge
2668 * before killing the task.
2670 * Only for regular pages, though: huge pages are rather
2671 * unlikely to succeed so close to the limit, and we fall back
2672 * to regular pages anyway in case of failure.
2674 if (nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER) && ret)
2675 return CHARGE_RETRY;
2678 * At task move, charge accounts can be doubly counted. So, it's
2679 * better to wait until the end of task_move if something is going on.
2681 if (mem_cgroup_wait_acct_move(mem_over_limit))
2682 return CHARGE_RETRY;
2685 mem_cgroup_oom(mem_over_limit, gfp_mask, get_order(csize));
2687 return CHARGE_NOMEM;
2691 * __mem_cgroup_try_charge() does
2692 * 1. detect memcg to be charged against from passed *mm and *ptr,
2693 * 2. update res_counter
2694 * 3. call memory reclaim if necessary.
2696 * In some special case, if the task is fatal, fatal_signal_pending() or
2697 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2698 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2699 * as possible without any hazards. 2: all pages should have a valid
2700 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2701 * pointer, that is treated as a charge to root_mem_cgroup.
2703 * So __mem_cgroup_try_charge() will return
2704 * 0 ... on success, filling *ptr with a valid memcg pointer.
2705 * -ENOMEM ... charge failure because of resource limits.
2706 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2708 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2709 * the oom-killer can be invoked.
2711 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2713 unsigned int nr_pages,
2714 struct mem_cgroup **ptr,
2717 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2718 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2719 struct mem_cgroup *memcg = NULL;
2723 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2724 * in system level. So, allow to go ahead dying process in addition to
2727 if (unlikely(test_thread_flag(TIF_MEMDIE)
2728 || fatal_signal_pending(current)))
2731 if (unlikely(task_in_memcg_oom(current)))
2734 if (gfp_mask & __GFP_NOFAIL)
2738 * We always charge the cgroup the mm_struct belongs to.
2739 * The mm_struct's mem_cgroup changes on task migration if the
2740 * thread group leader migrates. It's possible that mm is not
2741 * set, if so charge the root memcg (happens for pagecache usage).
2744 *ptr = root_mem_cgroup;
2746 if (*ptr) { /* css should be a valid one */
2748 if (mem_cgroup_is_root(memcg))
2750 if (consume_stock(memcg, nr_pages))
2752 css_get(&memcg->css);
2754 struct task_struct *p;
2757 p = rcu_dereference(mm->owner);
2759 * Because we don't have task_lock(), "p" can exit.
2760 * In that case, "memcg" can point to root or p can be NULL with
2761 * race with swapoff. Then, we have small risk of mis-accouning.
2762 * But such kind of mis-account by race always happens because
2763 * we don't have cgroup_mutex(). It's overkill and we allo that
2765 * (*) swapoff at el will charge against mm-struct not against
2766 * task-struct. So, mm->owner can be NULL.
2768 memcg = mem_cgroup_from_task(p);
2770 memcg = root_mem_cgroup;
2771 if (mem_cgroup_is_root(memcg)) {
2775 if (consume_stock(memcg, nr_pages)) {
2777 * It seems dagerous to access memcg without css_get().
2778 * But considering how consume_stok works, it's not
2779 * necessary. If consume_stock success, some charges
2780 * from this memcg are cached on this cpu. So, we
2781 * don't need to call css_get()/css_tryget() before
2782 * calling consume_stock().
2787 /* after here, we may be blocked. we need to get refcnt */
2788 if (!css_tryget(&memcg->css)) {
2796 bool invoke_oom = oom && !nr_oom_retries;
2798 /* If killed, bypass charge */
2799 if (fatal_signal_pending(current)) {
2800 css_put(&memcg->css);
2804 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch,
2805 nr_pages, invoke_oom);
2809 case CHARGE_RETRY: /* not in OOM situation but retry */
2811 css_put(&memcg->css);
2814 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2815 css_put(&memcg->css);
2817 case CHARGE_NOMEM: /* OOM routine works */
2818 if (!oom || invoke_oom) {
2819 css_put(&memcg->css);
2825 } while (ret != CHARGE_OK);
2827 if (batch > nr_pages)
2828 refill_stock(memcg, batch - nr_pages);
2829 css_put(&memcg->css);
2834 if (!(gfp_mask & __GFP_NOFAIL)) {
2839 *ptr = root_mem_cgroup;
2844 * Somemtimes we have to undo a charge we got by try_charge().
2845 * This function is for that and do uncharge, put css's refcnt.
2846 * gotten by try_charge().
2848 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2849 unsigned int nr_pages)
2851 if (!mem_cgroup_is_root(memcg)) {
2852 unsigned long bytes = nr_pages * PAGE_SIZE;
2854 res_counter_uncharge(&memcg->res, bytes);
2855 if (do_swap_account)
2856 res_counter_uncharge(&memcg->memsw, bytes);
2861 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2862 * This is useful when moving usage to parent cgroup.
2864 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2865 unsigned int nr_pages)
2867 unsigned long bytes = nr_pages * PAGE_SIZE;
2869 if (mem_cgroup_is_root(memcg))
2872 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2873 if (do_swap_account)
2874 res_counter_uncharge_until(&memcg->memsw,
2875 memcg->memsw.parent, bytes);
2879 * A helper function to get mem_cgroup from ID. must be called under
2880 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2881 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2882 * called against removed memcg.)
2884 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2886 /* ID 0 is unused ID */
2889 return mem_cgroup_from_id(id);
2892 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2894 struct mem_cgroup *memcg = NULL;
2895 struct page_cgroup *pc;
2899 VM_BUG_ON_PAGE(!PageLocked(page), page);
2901 pc = lookup_page_cgroup(page);
2902 lock_page_cgroup(pc);
2903 if (PageCgroupUsed(pc)) {
2904 memcg = pc->mem_cgroup;
2905 if (memcg && !css_tryget(&memcg->css))
2907 } else if (PageSwapCache(page)) {
2908 ent.val = page_private(page);
2909 id = lookup_swap_cgroup_id(ent);
2911 memcg = mem_cgroup_lookup(id);
2912 if (memcg && !css_tryget(&memcg->css))
2916 unlock_page_cgroup(pc);
2920 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2922 unsigned int nr_pages,
2923 enum charge_type ctype,
2926 struct page_cgroup *pc = lookup_page_cgroup(page);
2927 struct zone *uninitialized_var(zone);
2928 struct lruvec *lruvec;
2929 bool was_on_lru = false;
2932 lock_page_cgroup(pc);
2933 VM_BUG_ON_PAGE(PageCgroupUsed(pc), page);
2935 * we don't need page_cgroup_lock about tail pages, becase they are not
2936 * accessed by any other context at this point.
2940 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2941 * may already be on some other mem_cgroup's LRU. Take care of it.
2944 zone = page_zone(page);
2945 spin_lock_irq(&zone->lru_lock);
2946 if (PageLRU(page)) {
2947 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2949 del_page_from_lru_list(page, lruvec, page_lru(page));
2954 pc->mem_cgroup = memcg;
2956 * We access a page_cgroup asynchronously without lock_page_cgroup().
2957 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2958 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2959 * before USED bit, we need memory barrier here.
2960 * See mem_cgroup_add_lru_list(), etc.
2963 SetPageCgroupUsed(pc);
2967 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2968 VM_BUG_ON_PAGE(PageLRU(page), page);
2970 add_page_to_lru_list(page, lruvec, page_lru(page));
2972 spin_unlock_irq(&zone->lru_lock);
2975 if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
2980 mem_cgroup_charge_statistics(memcg, page, anon, nr_pages);
2981 unlock_page_cgroup(pc);
2984 * "charge_statistics" updated event counter. Then, check it.
2985 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2986 * if they exceeds softlimit.
2988 memcg_check_events(memcg, page);
2991 static DEFINE_MUTEX(set_limit_mutex);
2993 #ifdef CONFIG_MEMCG_KMEM
2994 static DEFINE_MUTEX(activate_kmem_mutex);
2996 static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg)
2998 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) &&
2999 memcg_kmem_is_active(memcg);
3003 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
3004 * in the memcg_cache_params struct.
3006 static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
3008 struct kmem_cache *cachep;
3010 VM_BUG_ON(p->is_root_cache);
3011 cachep = p->root_cache;
3012 return cache_from_memcg_idx(cachep, memcg_cache_id(p->memcg));
3015 #ifdef CONFIG_SLABINFO
3016 static int mem_cgroup_slabinfo_read(struct seq_file *m, void *v)
3018 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3019 struct memcg_cache_params *params;
3021 if (!memcg_can_account_kmem(memcg))
3024 print_slabinfo_header(m);
3026 mutex_lock(&memcg->slab_caches_mutex);
3027 list_for_each_entry(params, &memcg->memcg_slab_caches, list)
3028 cache_show(memcg_params_to_cache(params), m);
3029 mutex_unlock(&memcg->slab_caches_mutex);
3035 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size)
3037 struct res_counter *fail_res;
3038 struct mem_cgroup *_memcg;
3041 ret = res_counter_charge(&memcg->kmem, size, &fail_res);
3046 ret = __mem_cgroup_try_charge(NULL, gfp, size >> PAGE_SHIFT,
3047 &_memcg, oom_gfp_allowed(gfp));
3049 if (ret == -EINTR) {
3051 * __mem_cgroup_try_charge() chosed to bypass to root due to
3052 * OOM kill or fatal signal. Since our only options are to
3053 * either fail the allocation or charge it to this cgroup, do
3054 * it as a temporary condition. But we can't fail. From a
3055 * kmem/slab perspective, the cache has already been selected,
3056 * by mem_cgroup_kmem_get_cache(), so it is too late to change
3059 * This condition will only trigger if the task entered
3060 * memcg_charge_kmem in a sane state, but was OOM-killed during
3061 * __mem_cgroup_try_charge() above. Tasks that were already
3062 * dying when the allocation triggers should have been already
3063 * directed to the root cgroup in memcontrol.h
3065 res_counter_charge_nofail(&memcg->res, size, &fail_res);
3066 if (do_swap_account)
3067 res_counter_charge_nofail(&memcg->memsw, size,
3071 res_counter_uncharge(&memcg->kmem, size);
3076 static void memcg_uncharge_kmem(struct mem_cgroup *memcg, u64 size)
3078 res_counter_uncharge(&memcg->res, size);
3079 if (do_swap_account)
3080 res_counter_uncharge(&memcg->memsw, size);
3083 if (res_counter_uncharge(&memcg->kmem, size))
3087 * Releases a reference taken in kmem_cgroup_css_offline in case
3088 * this last uncharge is racing with the offlining code or it is
3089 * outliving the memcg existence.
3091 * The memory barrier imposed by test&clear is paired with the
3092 * explicit one in memcg_kmem_mark_dead().
3094 if (memcg_kmem_test_and_clear_dead(memcg))
3095 css_put(&memcg->css);
3099 * helper for acessing a memcg's index. It will be used as an index in the
3100 * child cache array in kmem_cache, and also to derive its name. This function
3101 * will return -1 when this is not a kmem-limited memcg.
3103 int memcg_cache_id(struct mem_cgroup *memcg)
3105 return memcg ? memcg->kmemcg_id : -1;
3108 static size_t memcg_caches_array_size(int num_groups)
3111 if (num_groups <= 0)
3114 size = 2 * num_groups;
3115 if (size < MEMCG_CACHES_MIN_SIZE)
3116 size = MEMCG_CACHES_MIN_SIZE;
3117 else if (size > MEMCG_CACHES_MAX_SIZE)
3118 size = MEMCG_CACHES_MAX_SIZE;
3124 * We should update the current array size iff all caches updates succeed. This
3125 * can only be done from the slab side. The slab mutex needs to be held when
3128 void memcg_update_array_size(int num)
3130 if (num > memcg_limited_groups_array_size)
3131 memcg_limited_groups_array_size = memcg_caches_array_size(num);
3134 static void kmem_cache_destroy_work_func(struct work_struct *w);
3136 int memcg_update_cache_size(struct kmem_cache *s, int num_groups)
3138 struct memcg_cache_params *cur_params = s->memcg_params;
3140 VM_BUG_ON(!is_root_cache(s));
3142 if (num_groups > memcg_limited_groups_array_size) {
3144 struct memcg_cache_params *new_params;
3145 ssize_t size = memcg_caches_array_size(num_groups);
3147 size *= sizeof(void *);
3148 size += offsetof(struct memcg_cache_params, memcg_caches);
3150 new_params = kzalloc(size, GFP_KERNEL);
3154 new_params->is_root_cache = true;
3157 * There is the chance it will be bigger than
3158 * memcg_limited_groups_array_size, if we failed an allocation
3159 * in a cache, in which case all caches updated before it, will
3160 * have a bigger array.
3162 * But if that is the case, the data after
3163 * memcg_limited_groups_array_size is certainly unused
3165 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3166 if (!cur_params->memcg_caches[i])
3168 new_params->memcg_caches[i] =
3169 cur_params->memcg_caches[i];
3173 * Ideally, we would wait until all caches succeed, and only
3174 * then free the old one. But this is not worth the extra
3175 * pointer per-cache we'd have to have for this.
3177 * It is not a big deal if some caches are left with a size
3178 * bigger than the others. And all updates will reset this
3181 rcu_assign_pointer(s->memcg_params, new_params);
3183 kfree_rcu(cur_params, rcu_head);
3188 int memcg_alloc_cache_params(struct mem_cgroup *memcg, struct kmem_cache *s,
3189 struct kmem_cache *root_cache)
3193 if (!memcg_kmem_enabled())
3197 size = offsetof(struct memcg_cache_params, memcg_caches);
3198 size += memcg_limited_groups_array_size * sizeof(void *);
3200 size = sizeof(struct memcg_cache_params);
3202 s->memcg_params = kzalloc(size, GFP_KERNEL);
3203 if (!s->memcg_params)
3207 s->memcg_params->memcg = memcg;
3208 s->memcg_params->root_cache = root_cache;
3209 INIT_WORK(&s->memcg_params->destroy,
3210 kmem_cache_destroy_work_func);
3212 s->memcg_params->is_root_cache = true;
3217 void memcg_free_cache_params(struct kmem_cache *s)
3219 kfree(s->memcg_params);
3222 void memcg_register_cache(struct kmem_cache *s)
3224 struct kmem_cache *root;
3225 struct mem_cgroup *memcg;
3228 if (is_root_cache(s))
3232 * Holding the slab_mutex assures nobody will touch the memcg_caches
3233 * array while we are modifying it.
3235 lockdep_assert_held(&slab_mutex);
3237 root = s->memcg_params->root_cache;
3238 memcg = s->memcg_params->memcg;
3239 id = memcg_cache_id(memcg);
3241 css_get(&memcg->css);
3245 * Since readers won't lock (see cache_from_memcg_idx()), we need a
3246 * barrier here to ensure nobody will see the kmem_cache partially
3252 * Initialize the pointer to this cache in its parent's memcg_params
3253 * before adding it to the memcg_slab_caches list, otherwise we can
3254 * fail to convert memcg_params_to_cache() while traversing the list.
3256 VM_BUG_ON(root->memcg_params->memcg_caches[id]);
3257 root->memcg_params->memcg_caches[id] = s;
3259 mutex_lock(&memcg->slab_caches_mutex);
3260 list_add(&s->memcg_params->list, &memcg->memcg_slab_caches);
3261 mutex_unlock(&memcg->slab_caches_mutex);
3264 void memcg_unregister_cache(struct kmem_cache *s)
3266 struct kmem_cache *root;
3267 struct mem_cgroup *memcg;
3270 if (is_root_cache(s))
3274 * Holding the slab_mutex assures nobody will touch the memcg_caches
3275 * array while we are modifying it.
3277 lockdep_assert_held(&slab_mutex);
3279 root = s->memcg_params->root_cache;
3280 memcg = s->memcg_params->memcg;
3281 id = memcg_cache_id(memcg);
3283 mutex_lock(&memcg->slab_caches_mutex);
3284 list_del(&s->memcg_params->list);
3285 mutex_unlock(&memcg->slab_caches_mutex);
3288 * Clear the pointer to this cache in its parent's memcg_params only
3289 * after removing it from the memcg_slab_caches list, otherwise we can
3290 * fail to convert memcg_params_to_cache() while traversing the list.
3292 VM_BUG_ON(!root->memcg_params->memcg_caches[id]);
3293 root->memcg_params->memcg_caches[id] = NULL;
3295 css_put(&memcg->css);
3299 * During the creation a new cache, we need to disable our accounting mechanism
3300 * altogether. This is true even if we are not creating, but rather just
3301 * enqueing new caches to be created.
3303 * This is because that process will trigger allocations; some visible, like
3304 * explicit kmallocs to auxiliary data structures, name strings and internal
3305 * cache structures; some well concealed, like INIT_WORK() that can allocate
3306 * objects during debug.
3308 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3309 * to it. This may not be a bounded recursion: since the first cache creation
3310 * failed to complete (waiting on the allocation), we'll just try to create the
3311 * cache again, failing at the same point.
3313 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3314 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3315 * inside the following two functions.
3317 static inline void memcg_stop_kmem_account(void)
3319 VM_BUG_ON(!current->mm);
3320 current->memcg_kmem_skip_account++;
3323 static inline void memcg_resume_kmem_account(void)
3325 VM_BUG_ON(!current->mm);
3326 current->memcg_kmem_skip_account--;
3329 static void kmem_cache_destroy_work_func(struct work_struct *w)
3331 struct kmem_cache *cachep;
3332 struct memcg_cache_params *p;
3334 p = container_of(w, struct memcg_cache_params, destroy);
3336 cachep = memcg_params_to_cache(p);
3339 * If we get down to 0 after shrink, we could delete right away.
3340 * However, memcg_release_pages() already puts us back in the workqueue
3341 * in that case. If we proceed deleting, we'll get a dangling
3342 * reference, and removing the object from the workqueue in that case
3343 * is unnecessary complication. We are not a fast path.
3345 * Note that this case is fundamentally different from racing with
3346 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3347 * kmem_cache_shrink, not only we would be reinserting a dead cache
3348 * into the queue, but doing so from inside the worker racing to
3351 * So if we aren't down to zero, we'll just schedule a worker and try
3354 if (atomic_read(&cachep->memcg_params->nr_pages) != 0) {
3355 kmem_cache_shrink(cachep);
3356 if (atomic_read(&cachep->memcg_params->nr_pages) == 0)
3359 kmem_cache_destroy(cachep);
3362 void mem_cgroup_destroy_cache(struct kmem_cache *cachep)
3364 if (!cachep->memcg_params->dead)
3368 * There are many ways in which we can get here.
3370 * We can get to a memory-pressure situation while the delayed work is
3371 * still pending to run. The vmscan shrinkers can then release all
3372 * cache memory and get us to destruction. If this is the case, we'll
3373 * be executed twice, which is a bug (the second time will execute over
3374 * bogus data). In this case, cancelling the work should be fine.
3376 * But we can also get here from the worker itself, if
3377 * kmem_cache_shrink is enough to shake all the remaining objects and
3378 * get the page count to 0. In this case, we'll deadlock if we try to
3379 * cancel the work (the worker runs with an internal lock held, which
3380 * is the same lock we would hold for cancel_work_sync().)
3382 * Since we can't possibly know who got us here, just refrain from
3383 * running if there is already work pending
3385 if (work_pending(&cachep->memcg_params->destroy))
3388 * We have to defer the actual destroying to a workqueue, because
3389 * we might currently be in a context that cannot sleep.
3391 schedule_work(&cachep->memcg_params->destroy);
3394 static struct kmem_cache *memcg_create_kmem_cache(struct mem_cgroup *memcg,
3395 struct kmem_cache *s)
3397 struct kmem_cache *new;
3398 static char *tmp_name = NULL;
3399 static DEFINE_MUTEX(mutex); /* protects tmp_name */
3401 BUG_ON(!memcg_can_account_kmem(memcg));
3405 * kmem_cache_create_memcg duplicates the given name and
3406 * cgroup_name for this name requires RCU context.
3407 * This static temporary buffer is used to prevent from
3408 * pointless shortliving allocation.
3411 tmp_name = kmalloc(PATH_MAX, GFP_KERNEL);
3417 snprintf(tmp_name, PATH_MAX, "%s(%d:%s)", s->name,
3418 memcg_cache_id(memcg), cgroup_name(memcg->css.cgroup));
3421 new = kmem_cache_create_memcg(memcg, tmp_name, s->object_size, s->align,
3422 (s->flags & ~SLAB_PANIC), s->ctor, s);
3425 new->allocflags |= __GFP_KMEMCG;
3429 mutex_unlock(&mutex);
3433 void kmem_cache_destroy_memcg_children(struct kmem_cache *s)
3435 struct kmem_cache *c;
3438 if (!s->memcg_params)
3440 if (!s->memcg_params->is_root_cache)
3444 * If the cache is being destroyed, we trust that there is no one else
3445 * requesting objects from it. Even if there are, the sanity checks in
3446 * kmem_cache_destroy should caught this ill-case.
3448 * Still, we don't want anyone else freeing memcg_caches under our
3449 * noses, which can happen if a new memcg comes to life. As usual,
3450 * we'll take the activate_kmem_mutex to protect ourselves against
3453 mutex_lock(&activate_kmem_mutex);
3454 for_each_memcg_cache_index(i) {
3455 c = cache_from_memcg_idx(s, i);
3460 * We will now manually delete the caches, so to avoid races
3461 * we need to cancel all pending destruction workers and
3462 * proceed with destruction ourselves.
3464 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3465 * and that could spawn the workers again: it is likely that
3466 * the cache still have active pages until this very moment.
3467 * This would lead us back to mem_cgroup_destroy_cache.
3469 * But that will not execute at all if the "dead" flag is not
3470 * set, so flip it down to guarantee we are in control.
3472 c->memcg_params->dead = false;
3473 cancel_work_sync(&c->memcg_params->destroy);
3474 kmem_cache_destroy(c);
3476 mutex_unlock(&activate_kmem_mutex);
3479 struct create_work {
3480 struct mem_cgroup *memcg;
3481 struct kmem_cache *cachep;
3482 struct work_struct work;
3485 static void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3487 struct kmem_cache *cachep;
3488 struct memcg_cache_params *params;
3490 if (!memcg_kmem_is_active(memcg))
3493 mutex_lock(&memcg->slab_caches_mutex);
3494 list_for_each_entry(params, &memcg->memcg_slab_caches, list) {
3495 cachep = memcg_params_to_cache(params);
3496 cachep->memcg_params->dead = true;
3497 schedule_work(&cachep->memcg_params->destroy);
3499 mutex_unlock(&memcg->slab_caches_mutex);
3502 static void memcg_create_cache_work_func(struct work_struct *w)
3504 struct create_work *cw;
3506 cw = container_of(w, struct create_work, work);
3507 memcg_create_kmem_cache(cw->memcg, cw->cachep);
3508 css_put(&cw->memcg->css);
3513 * Enqueue the creation of a per-memcg kmem_cache.
3515 static void __memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3516 struct kmem_cache *cachep)
3518 struct create_work *cw;
3520 cw = kmalloc(sizeof(struct create_work), GFP_NOWAIT);
3522 css_put(&memcg->css);
3527 cw->cachep = cachep;
3529 INIT_WORK(&cw->work, memcg_create_cache_work_func);
3530 schedule_work(&cw->work);
3533 static void memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3534 struct kmem_cache *cachep)
3537 * We need to stop accounting when we kmalloc, because if the
3538 * corresponding kmalloc cache is not yet created, the first allocation
3539 * in __memcg_create_cache_enqueue will recurse.
3541 * However, it is better to enclose the whole function. Depending on
3542 * the debugging options enabled, INIT_WORK(), for instance, can
3543 * trigger an allocation. This too, will make us recurse. Because at
3544 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3545 * the safest choice is to do it like this, wrapping the whole function.
3547 memcg_stop_kmem_account();
3548 __memcg_create_cache_enqueue(memcg, cachep);
3549 memcg_resume_kmem_account();
3552 * Return the kmem_cache we're supposed to use for a slab allocation.
3553 * We try to use the current memcg's version of the cache.
3555 * If the cache does not exist yet, if we are the first user of it,
3556 * we either create it immediately, if possible, or create it asynchronously
3558 * In the latter case, we will let the current allocation go through with
3559 * the original cache.
3561 * Can't be called in interrupt context or from kernel threads.
3562 * This function needs to be called with rcu_read_lock() held.
3564 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
3567 struct mem_cgroup *memcg;
3568 struct kmem_cache *memcg_cachep;
3570 VM_BUG_ON(!cachep->memcg_params);
3571 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
3573 if (!current->mm || current->memcg_kmem_skip_account)
3577 memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
3579 if (!memcg_can_account_kmem(memcg))
3582 memcg_cachep = cache_from_memcg_idx(cachep, memcg_cache_id(memcg));
3583 if (likely(memcg_cachep)) {
3584 cachep = memcg_cachep;
3588 /* The corresponding put will be done in the workqueue. */
3589 if (!css_tryget(&memcg->css))
3594 * If we are in a safe context (can wait, and not in interrupt
3595 * context), we could be be predictable and return right away.
3596 * This would guarantee that the allocation being performed
3597 * already belongs in the new cache.
3599 * However, there are some clashes that can arrive from locking.
3600 * For instance, because we acquire the slab_mutex while doing
3601 * kmem_cache_dup, this means no further allocation could happen
3602 * with the slab_mutex held.
3604 * Also, because cache creation issue get_online_cpus(), this
3605 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3606 * that ends up reversed during cpu hotplug. (cpuset allocates
3607 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3608 * better to defer everything.
3610 memcg_create_cache_enqueue(memcg, cachep);
3616 EXPORT_SYMBOL(__memcg_kmem_get_cache);
3619 * We need to verify if the allocation against current->mm->owner's memcg is
3620 * possible for the given order. But the page is not allocated yet, so we'll
3621 * need a further commit step to do the final arrangements.
3623 * It is possible for the task to switch cgroups in this mean time, so at
3624 * commit time, we can't rely on task conversion any longer. We'll then use
3625 * the handle argument to return to the caller which cgroup we should commit
3626 * against. We could also return the memcg directly and avoid the pointer
3627 * passing, but a boolean return value gives better semantics considering
3628 * the compiled-out case as well.
3630 * Returning true means the allocation is possible.
3633 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3635 struct mem_cgroup *memcg;
3641 * Disabling accounting is only relevant for some specific memcg
3642 * internal allocations. Therefore we would initially not have such
3643 * check here, since direct calls to the page allocator that are marked
3644 * with GFP_KMEMCG only happen outside memcg core. We are mostly
3645 * concerned with cache allocations, and by having this test at
3646 * memcg_kmem_get_cache, we are already able to relay the allocation to
3647 * the root cache and bypass the memcg cache altogether.
3649 * There is one exception, though: the SLUB allocator does not create
3650 * large order caches, but rather service large kmallocs directly from
3651 * the page allocator. Therefore, the following sequence when backed by
3652 * the SLUB allocator:
3654 * memcg_stop_kmem_account();
3655 * kmalloc(<large_number>)
3656 * memcg_resume_kmem_account();
3658 * would effectively ignore the fact that we should skip accounting,
3659 * since it will drive us directly to this function without passing
3660 * through the cache selector memcg_kmem_get_cache. Such large
3661 * allocations are extremely rare but can happen, for instance, for the
3662 * cache arrays. We bring this test here.
3664 if (!current->mm || current->memcg_kmem_skip_account)
3667 memcg = try_get_mem_cgroup_from_mm(current->mm);
3670 * very rare case described in mem_cgroup_from_task. Unfortunately there
3671 * isn't much we can do without complicating this too much, and it would
3672 * be gfp-dependent anyway. Just let it go
3674 if (unlikely(!memcg))
3677 if (!memcg_can_account_kmem(memcg)) {
3678 css_put(&memcg->css);
3682 ret = memcg_charge_kmem(memcg, gfp, PAGE_SIZE << order);
3686 css_put(&memcg->css);
3690 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3693 struct page_cgroup *pc;
3695 VM_BUG_ON(mem_cgroup_is_root(memcg));
3697 /* The page allocation failed. Revert */
3699 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3703 pc = lookup_page_cgroup(page);
3704 lock_page_cgroup(pc);
3705 pc->mem_cgroup = memcg;
3706 SetPageCgroupUsed(pc);
3707 unlock_page_cgroup(pc);
3710 void __memcg_kmem_uncharge_pages(struct page *page, int order)
3712 struct mem_cgroup *memcg = NULL;
3713 struct page_cgroup *pc;
3716 pc = lookup_page_cgroup(page);
3718 * Fast unlocked return. Theoretically might have changed, have to
3719 * check again after locking.
3721 if (!PageCgroupUsed(pc))
3724 lock_page_cgroup(pc);
3725 if (PageCgroupUsed(pc)) {
3726 memcg = pc->mem_cgroup;
3727 ClearPageCgroupUsed(pc);
3729 unlock_page_cgroup(pc);
3732 * We trust that only if there is a memcg associated with the page, it
3733 * is a valid allocation
3738 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
3739 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3742 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3745 #endif /* CONFIG_MEMCG_KMEM */
3747 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3749 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3751 * Because tail pages are not marked as "used", set it. We're under
3752 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3753 * charge/uncharge will be never happen and move_account() is done under
3754 * compound_lock(), so we don't have to take care of races.
3756 void mem_cgroup_split_huge_fixup(struct page *head)
3758 struct page_cgroup *head_pc = lookup_page_cgroup(head);
3759 struct page_cgroup *pc;
3760 struct mem_cgroup *memcg;
3763 if (mem_cgroup_disabled())
3766 memcg = head_pc->mem_cgroup;
3767 for (i = 1; i < HPAGE_PMD_NR; i++) {
3769 pc->mem_cgroup = memcg;
3770 smp_wmb();/* see __commit_charge() */
3771 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
3773 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
3776 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3779 void mem_cgroup_move_account_page_stat(struct mem_cgroup *from,
3780 struct mem_cgroup *to,
3781 unsigned int nr_pages,
3782 enum mem_cgroup_stat_index idx)
3784 /* Update stat data for mem_cgroup */
3786 __this_cpu_sub(from->stat->count[idx], nr_pages);
3787 __this_cpu_add(to->stat->count[idx], nr_pages);
3792 * mem_cgroup_move_account - move account of the page
3794 * @nr_pages: number of regular pages (>1 for huge pages)
3795 * @pc: page_cgroup of the page.
3796 * @from: mem_cgroup which the page is moved from.
3797 * @to: mem_cgroup which the page is moved to. @from != @to.
3799 * The caller must confirm following.
3800 * - page is not on LRU (isolate_page() is useful.)
3801 * - compound_lock is held when nr_pages > 1
3803 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3806 static int mem_cgroup_move_account(struct page *page,
3807 unsigned int nr_pages,
3808 struct page_cgroup *pc,
3809 struct mem_cgroup *from,
3810 struct mem_cgroup *to)
3812 unsigned long flags;
3814 bool anon = PageAnon(page);
3816 VM_BUG_ON(from == to);
3817 VM_BUG_ON_PAGE(PageLRU(page), page);
3819 * The page is isolated from LRU. So, collapse function
3820 * will not handle this page. But page splitting can happen.
3821 * Do this check under compound_page_lock(). The caller should
3825 if (nr_pages > 1 && !PageTransHuge(page))
3828 lock_page_cgroup(pc);
3831 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3834 move_lock_mem_cgroup(from, &flags);
3836 if (!anon && page_mapped(page))
3837 mem_cgroup_move_account_page_stat(from, to, nr_pages,
3838 MEM_CGROUP_STAT_FILE_MAPPED);
3840 if (PageWriteback(page))
3841 mem_cgroup_move_account_page_stat(from, to, nr_pages,
3842 MEM_CGROUP_STAT_WRITEBACK);
3844 mem_cgroup_charge_statistics(from, page, anon, -nr_pages);
3846 /* caller should have done css_get */
3847 pc->mem_cgroup = to;
3848 mem_cgroup_charge_statistics(to, page, anon, nr_pages);
3849 move_unlock_mem_cgroup(from, &flags);
3852 unlock_page_cgroup(pc);
3856 memcg_check_events(to, page);
3857 memcg_check_events(from, page);
3863 * mem_cgroup_move_parent - moves page to the parent group
3864 * @page: the page to move
3865 * @pc: page_cgroup of the page
3866 * @child: page's cgroup
3868 * move charges to its parent or the root cgroup if the group has no
3869 * parent (aka use_hierarchy==0).
3870 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3871 * mem_cgroup_move_account fails) the failure is always temporary and
3872 * it signals a race with a page removal/uncharge or migration. In the
3873 * first case the page is on the way out and it will vanish from the LRU
3874 * on the next attempt and the call should be retried later.
3875 * Isolation from the LRU fails only if page has been isolated from
3876 * the LRU since we looked at it and that usually means either global
3877 * reclaim or migration going on. The page will either get back to the
3879 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3880 * (!PageCgroupUsed) or moved to a different group. The page will
3881 * disappear in the next attempt.
3883 static int mem_cgroup_move_parent(struct page *page,
3884 struct page_cgroup *pc,
3885 struct mem_cgroup *child)
3887 struct mem_cgroup *parent;
3888 unsigned int nr_pages;
3889 unsigned long uninitialized_var(flags);
3892 VM_BUG_ON(mem_cgroup_is_root(child));
3895 if (!get_page_unless_zero(page))
3897 if (isolate_lru_page(page))
3900 nr_pages = hpage_nr_pages(page);
3902 parent = parent_mem_cgroup(child);
3904 * If no parent, move charges to root cgroup.
3907 parent = root_mem_cgroup;
3910 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
3911 flags = compound_lock_irqsave(page);
3914 ret = mem_cgroup_move_account(page, nr_pages,
3917 __mem_cgroup_cancel_local_charge(child, nr_pages);
3920 compound_unlock_irqrestore(page, flags);
3921 putback_lru_page(page);
3929 * Charge the memory controller for page usage.
3931 * 0 if the charge was successful
3932 * < 0 if the cgroup is over its limit
3934 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
3935 gfp_t gfp_mask, enum charge_type ctype)
3937 struct mem_cgroup *memcg = NULL;
3938 unsigned int nr_pages = 1;
3942 if (PageTransHuge(page)) {
3943 nr_pages <<= compound_order(page);
3944 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
3946 * Never OOM-kill a process for a huge page. The
3947 * fault handler will fall back to regular pages.
3952 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
3955 __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false);
3959 int mem_cgroup_newpage_charge(struct page *page,
3960 struct mm_struct *mm, gfp_t gfp_mask)
3962 if (mem_cgroup_disabled())
3964 VM_BUG_ON_PAGE(page_mapped(page), page);
3965 VM_BUG_ON_PAGE(page->mapping && !PageAnon(page), page);
3967 return mem_cgroup_charge_common(page, mm, gfp_mask,
3968 MEM_CGROUP_CHARGE_TYPE_ANON);
3972 * While swap-in, try_charge -> commit or cancel, the page is locked.
3973 * And when try_charge() successfully returns, one refcnt to memcg without
3974 * struct page_cgroup is acquired. This refcnt will be consumed by
3975 * "commit()" or removed by "cancel()"
3977 static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm,
3980 struct mem_cgroup **memcgp)
3982 struct mem_cgroup *memcg;
3983 struct page_cgroup *pc;
3986 pc = lookup_page_cgroup(page);
3988 * Every swap fault against a single page tries to charge the
3989 * page, bail as early as possible. shmem_unuse() encounters
3990 * already charged pages, too. The USED bit is protected by
3991 * the page lock, which serializes swap cache removal, which
3992 * in turn serializes uncharging.
3994 if (PageCgroupUsed(pc))
3996 if (!do_swap_account)
3998 memcg = try_get_mem_cgroup_from_page(page);
4002 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
4003 css_put(&memcg->css);
4008 ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
4014 int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page,
4015 gfp_t gfp_mask, struct mem_cgroup **memcgp)
4018 if (mem_cgroup_disabled())
4021 * A racing thread's fault, or swapoff, may have already
4022 * updated the pte, and even removed page from swap cache: in
4023 * those cases unuse_pte()'s pte_same() test will fail; but
4024 * there's also a KSM case which does need to charge the page.
4026 if (!PageSwapCache(page)) {
4029 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, memcgp, true);
4034 return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp);
4037 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
4039 if (mem_cgroup_disabled())
4043 __mem_cgroup_cancel_charge(memcg, 1);
4047 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
4048 enum charge_type ctype)
4050 if (mem_cgroup_disabled())
4055 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
4057 * Now swap is on-memory. This means this page may be
4058 * counted both as mem and swap....double count.
4059 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
4060 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
4061 * may call delete_from_swap_cache() before reach here.
4063 if (do_swap_account && PageSwapCache(page)) {
4064 swp_entry_t ent = {.val = page_private(page)};
4065 mem_cgroup_uncharge_swap(ent);
4069 void mem_cgroup_commit_charge_swapin(struct page *page,
4070 struct mem_cgroup *memcg)
4072 __mem_cgroup_commit_charge_swapin(page, memcg,
4073 MEM_CGROUP_CHARGE_TYPE_ANON);
4076 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
4079 struct mem_cgroup *memcg = NULL;
4080 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4083 if (mem_cgroup_disabled())
4085 if (PageCompound(page))
4088 if (!PageSwapCache(page))
4089 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
4090 else { /* page is swapcache/shmem */
4091 ret = __mem_cgroup_try_charge_swapin(mm, page,
4094 __mem_cgroup_commit_charge_swapin(page, memcg, type);
4099 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
4100 unsigned int nr_pages,
4101 const enum charge_type ctype)
4103 struct memcg_batch_info *batch = NULL;
4104 bool uncharge_memsw = true;
4106 /* If swapout, usage of swap doesn't decrease */
4107 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
4108 uncharge_memsw = false;
4110 batch = ¤t->memcg_batch;
4112 * In usual, we do css_get() when we remember memcg pointer.
4113 * But in this case, we keep res->usage until end of a series of
4114 * uncharges. Then, it's ok to ignore memcg's refcnt.
4117 batch->memcg = memcg;
4119 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
4120 * In those cases, all pages freed continuously can be expected to be in
4121 * the same cgroup and we have chance to coalesce uncharges.
4122 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
4123 * because we want to do uncharge as soon as possible.
4126 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
4127 goto direct_uncharge;
4130 goto direct_uncharge;
4133 * In typical case, batch->memcg == mem. This means we can
4134 * merge a series of uncharges to an uncharge of res_counter.
4135 * If not, we uncharge res_counter ony by one.
4137 if (batch->memcg != memcg)
4138 goto direct_uncharge;
4139 /* remember freed charge and uncharge it later */
4142 batch->memsw_nr_pages++;
4145 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
4147 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
4148 if (unlikely(batch->memcg != memcg))
4149 memcg_oom_recover(memcg);
4153 * uncharge if !page_mapped(page)
4155 static struct mem_cgroup *
4156 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
4159 struct mem_cgroup *memcg = NULL;
4160 unsigned int nr_pages = 1;
4161 struct page_cgroup *pc;
4164 if (mem_cgroup_disabled())
4167 if (PageTransHuge(page)) {
4168 nr_pages <<= compound_order(page);
4169 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
4172 * Check if our page_cgroup is valid
4174 pc = lookup_page_cgroup(page);
4175 if (unlikely(!PageCgroupUsed(pc)))
4178 lock_page_cgroup(pc);
4180 memcg = pc->mem_cgroup;
4182 if (!PageCgroupUsed(pc))
4185 anon = PageAnon(page);
4188 case MEM_CGROUP_CHARGE_TYPE_ANON:
4190 * Generally PageAnon tells if it's the anon statistics to be
4191 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
4192 * used before page reached the stage of being marked PageAnon.
4196 case MEM_CGROUP_CHARGE_TYPE_DROP:
4197 /* See mem_cgroup_prepare_migration() */
4198 if (page_mapped(page))
4201 * Pages under migration may not be uncharged. But
4202 * end_migration() /must/ be the one uncharging the
4203 * unused post-migration page and so it has to call
4204 * here with the migration bit still set. See the
4205 * res_counter handling below.
4207 if (!end_migration && PageCgroupMigration(pc))
4210 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
4211 if (!PageAnon(page)) { /* Shared memory */
4212 if (page->mapping && !page_is_file_cache(page))
4214 } else if (page_mapped(page)) /* Anon */
4221 mem_cgroup_charge_statistics(memcg, page, anon, -nr_pages);
4223 ClearPageCgroupUsed(pc);
4225 * pc->mem_cgroup is not cleared here. It will be accessed when it's
4226 * freed from LRU. This is safe because uncharged page is expected not
4227 * to be reused (freed soon). Exception is SwapCache, it's handled by
4228 * special functions.
4231 unlock_page_cgroup(pc);
4233 * even after unlock, we have memcg->res.usage here and this memcg
4234 * will never be freed, so it's safe to call css_get().
4236 memcg_check_events(memcg, page);
4237 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
4238 mem_cgroup_swap_statistics(memcg, true);
4239 css_get(&memcg->css);
4242 * Migration does not charge the res_counter for the
4243 * replacement page, so leave it alone when phasing out the
4244 * page that is unused after the migration.
4246 if (!end_migration && !mem_cgroup_is_root(memcg))
4247 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
4252 unlock_page_cgroup(pc);
4256 void mem_cgroup_uncharge_page(struct page *page)
4259 if (page_mapped(page))
4261 VM_BUG_ON_PAGE(page->mapping && !PageAnon(page), page);
4263 * If the page is in swap cache, uncharge should be deferred
4264 * to the swap path, which also properly accounts swap usage
4265 * and handles memcg lifetime.
4267 * Note that this check is not stable and reclaim may add the
4268 * page to swap cache at any time after this. However, if the
4269 * page is not in swap cache by the time page->mapcount hits
4270 * 0, there won't be any page table references to the swap
4271 * slot, and reclaim will free it and not actually write the
4274 if (PageSwapCache(page))
4276 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
4279 void mem_cgroup_uncharge_cache_page(struct page *page)
4281 VM_BUG_ON_PAGE(page_mapped(page), page);
4282 VM_BUG_ON_PAGE(page->mapping, page);
4283 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
4287 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4288 * In that cases, pages are freed continuously and we can expect pages
4289 * are in the same memcg. All these calls itself limits the number of
4290 * pages freed at once, then uncharge_start/end() is called properly.
4291 * This may be called prural(2) times in a context,
4294 void mem_cgroup_uncharge_start(void)
4296 current->memcg_batch.do_batch++;
4297 /* We can do nest. */
4298 if (current->memcg_batch.do_batch == 1) {
4299 current->memcg_batch.memcg = NULL;
4300 current->memcg_batch.nr_pages = 0;
4301 current->memcg_batch.memsw_nr_pages = 0;
4305 void mem_cgroup_uncharge_end(void)
4307 struct memcg_batch_info *batch = ¤t->memcg_batch;
4309 if (!batch->do_batch)
4313 if (batch->do_batch) /* If stacked, do nothing. */
4319 * This "batch->memcg" is valid without any css_get/put etc...
4320 * bacause we hide charges behind us.
4322 if (batch->nr_pages)
4323 res_counter_uncharge(&batch->memcg->res,
4324 batch->nr_pages * PAGE_SIZE);
4325 if (batch->memsw_nr_pages)
4326 res_counter_uncharge(&batch->memcg->memsw,
4327 batch->memsw_nr_pages * PAGE_SIZE);
4328 memcg_oom_recover(batch->memcg);
4329 /* forget this pointer (for sanity check) */
4330 batch->memcg = NULL;
4335 * called after __delete_from_swap_cache() and drop "page" account.
4336 * memcg information is recorded to swap_cgroup of "ent"
4339 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
4341 struct mem_cgroup *memcg;
4342 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
4344 if (!swapout) /* this was a swap cache but the swap is unused ! */
4345 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
4347 memcg = __mem_cgroup_uncharge_common(page, ctype, false);
4350 * record memcg information, if swapout && memcg != NULL,
4351 * css_get() was called in uncharge().
4353 if (do_swap_account && swapout && memcg)
4354 swap_cgroup_record(ent, mem_cgroup_id(memcg));
4358 #ifdef CONFIG_MEMCG_SWAP
4360 * called from swap_entry_free(). remove record in swap_cgroup and
4361 * uncharge "memsw" account.
4363 void mem_cgroup_uncharge_swap(swp_entry_t ent)
4365 struct mem_cgroup *memcg;
4368 if (!do_swap_account)
4371 id = swap_cgroup_record(ent, 0);
4373 memcg = mem_cgroup_lookup(id);
4376 * We uncharge this because swap is freed.
4377 * This memcg can be obsolete one. We avoid calling css_tryget
4379 if (!mem_cgroup_is_root(memcg))
4380 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
4381 mem_cgroup_swap_statistics(memcg, false);
4382 css_put(&memcg->css);
4388 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4389 * @entry: swap entry to be moved
4390 * @from: mem_cgroup which the entry is moved from
4391 * @to: mem_cgroup which the entry is moved to
4393 * It succeeds only when the swap_cgroup's record for this entry is the same
4394 * as the mem_cgroup's id of @from.
4396 * Returns 0 on success, -EINVAL on failure.
4398 * The caller must have charged to @to, IOW, called res_counter_charge() about
4399 * both res and memsw, and called css_get().
4401 static int mem_cgroup_move_swap_account(swp_entry_t entry,
4402 struct mem_cgroup *from, struct mem_cgroup *to)
4404 unsigned short old_id, new_id;
4406 old_id = mem_cgroup_id(from);
4407 new_id = mem_cgroup_id(to);
4409 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
4410 mem_cgroup_swap_statistics(from, false);
4411 mem_cgroup_swap_statistics(to, true);
4413 * This function is only called from task migration context now.
4414 * It postpones res_counter and refcount handling till the end
4415 * of task migration(mem_cgroup_clear_mc()) for performance
4416 * improvement. But we cannot postpone css_get(to) because if
4417 * the process that has been moved to @to does swap-in, the
4418 * refcount of @to might be decreased to 0.
4420 * We are in attach() phase, so the cgroup is guaranteed to be
4421 * alive, so we can just call css_get().
4429 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
4430 struct mem_cgroup *from, struct mem_cgroup *to)
4437 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4440 void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
4441 struct mem_cgroup **memcgp)
4443 struct mem_cgroup *memcg = NULL;
4444 unsigned int nr_pages = 1;
4445 struct page_cgroup *pc;
4446 enum charge_type ctype;
4450 if (mem_cgroup_disabled())
4453 if (PageTransHuge(page))
4454 nr_pages <<= compound_order(page);
4456 pc = lookup_page_cgroup(page);
4457 lock_page_cgroup(pc);
4458 if (PageCgroupUsed(pc)) {
4459 memcg = pc->mem_cgroup;
4460 css_get(&memcg->css);
4462 * At migrating an anonymous page, its mapcount goes down
4463 * to 0 and uncharge() will be called. But, even if it's fully
4464 * unmapped, migration may fail and this page has to be
4465 * charged again. We set MIGRATION flag here and delay uncharge
4466 * until end_migration() is called
4468 * Corner Case Thinking
4470 * When the old page was mapped as Anon and it's unmap-and-freed
4471 * while migration was ongoing.
4472 * If unmap finds the old page, uncharge() of it will be delayed
4473 * until end_migration(). If unmap finds a new page, it's
4474 * uncharged when it make mapcount to be 1->0. If unmap code
4475 * finds swap_migration_entry, the new page will not be mapped
4476 * and end_migration() will find it(mapcount==0).
4479 * When the old page was mapped but migraion fails, the kernel
4480 * remaps it. A charge for it is kept by MIGRATION flag even
4481 * if mapcount goes down to 0. We can do remap successfully
4482 * without charging it again.
4485 * The "old" page is under lock_page() until the end of
4486 * migration, so, the old page itself will not be swapped-out.
4487 * If the new page is swapped out before end_migraton, our
4488 * hook to usual swap-out path will catch the event.
4491 SetPageCgroupMigration(pc);
4493 unlock_page_cgroup(pc);
4495 * If the page is not charged at this point,
4503 * We charge new page before it's used/mapped. So, even if unlock_page()
4504 * is called before end_migration, we can catch all events on this new
4505 * page. In the case new page is migrated but not remapped, new page's
4506 * mapcount will be finally 0 and we call uncharge in end_migration().
4509 ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
4511 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
4513 * The page is committed to the memcg, but it's not actually
4514 * charged to the res_counter since we plan on replacing the
4515 * old one and only one page is going to be left afterwards.
4517 __mem_cgroup_commit_charge(memcg, newpage, nr_pages, ctype, false);
4520 /* remove redundant charge if migration failed*/
4521 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
4522 struct page *oldpage, struct page *newpage, bool migration_ok)
4524 struct page *used, *unused;
4525 struct page_cgroup *pc;
4531 if (!migration_ok) {
4538 anon = PageAnon(used);
4539 __mem_cgroup_uncharge_common(unused,
4540 anon ? MEM_CGROUP_CHARGE_TYPE_ANON
4541 : MEM_CGROUP_CHARGE_TYPE_CACHE,
4543 css_put(&memcg->css);
4545 * We disallowed uncharge of pages under migration because mapcount
4546 * of the page goes down to zero, temporarly.
4547 * Clear the flag and check the page should be charged.
4549 pc = lookup_page_cgroup(oldpage);
4550 lock_page_cgroup(pc);
4551 ClearPageCgroupMigration(pc);
4552 unlock_page_cgroup(pc);
4555 * If a page is a file cache, radix-tree replacement is very atomic
4556 * and we can skip this check. When it was an Anon page, its mapcount
4557 * goes down to 0. But because we added MIGRATION flage, it's not
4558 * uncharged yet. There are several case but page->mapcount check
4559 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4560 * check. (see prepare_charge() also)
4563 mem_cgroup_uncharge_page(used);
4567 * At replace page cache, newpage is not under any memcg but it's on
4568 * LRU. So, this function doesn't touch res_counter but handles LRU
4569 * in correct way. Both pages are locked so we cannot race with uncharge.
4571 void mem_cgroup_replace_page_cache(struct page *oldpage,
4572 struct page *newpage)
4574 struct mem_cgroup *memcg = NULL;
4575 struct page_cgroup *pc;
4576 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4578 if (mem_cgroup_disabled())
4581 pc = lookup_page_cgroup(oldpage);
4582 /* fix accounting on old pages */
4583 lock_page_cgroup(pc);
4584 if (PageCgroupUsed(pc)) {
4585 memcg = pc->mem_cgroup;
4586 mem_cgroup_charge_statistics(memcg, oldpage, false, -1);
4587 ClearPageCgroupUsed(pc);
4589 unlock_page_cgroup(pc);
4592 * When called from shmem_replace_page(), in some cases the
4593 * oldpage has already been charged, and in some cases not.
4598 * Even if newpage->mapping was NULL before starting replacement,
4599 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4600 * LRU while we overwrite pc->mem_cgroup.
4602 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
4605 #ifdef CONFIG_DEBUG_VM
4606 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
4608 struct page_cgroup *pc;
4610 pc = lookup_page_cgroup(page);
4612 * Can be NULL while feeding pages into the page allocator for
4613 * the first time, i.e. during boot or memory hotplug;
4614 * or when mem_cgroup_disabled().
4616 if (likely(pc) && PageCgroupUsed(pc))
4621 bool mem_cgroup_bad_page_check(struct page *page)
4623 if (mem_cgroup_disabled())
4626 return lookup_page_cgroup_used(page) != NULL;
4629 void mem_cgroup_print_bad_page(struct page *page)
4631 struct page_cgroup *pc;
4633 pc = lookup_page_cgroup_used(page);
4635 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4636 pc, pc->flags, pc->mem_cgroup);
4641 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
4642 unsigned long long val)
4645 u64 memswlimit, memlimit;
4647 int children = mem_cgroup_count_children(memcg);
4648 u64 curusage, oldusage;
4652 * For keeping hierarchical_reclaim simple, how long we should retry
4653 * is depends on callers. We set our retry-count to be function
4654 * of # of children which we should visit in this loop.
4656 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
4658 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4661 while (retry_count) {
4662 if (signal_pending(current)) {
4667 * Rather than hide all in some function, I do this in
4668 * open coded manner. You see what this really does.
4669 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4671 mutex_lock(&set_limit_mutex);
4672 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4673 if (memswlimit < val) {
4675 mutex_unlock(&set_limit_mutex);
4679 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4683 ret = res_counter_set_limit(&memcg->res, val);
4685 if (memswlimit == val)
4686 memcg->memsw_is_minimum = true;
4688 memcg->memsw_is_minimum = false;
4690 mutex_unlock(&set_limit_mutex);
4695 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4696 MEM_CGROUP_RECLAIM_SHRINK);
4697 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4698 /* Usage is reduced ? */
4699 if (curusage >= oldusage)
4702 oldusage = curusage;
4704 if (!ret && enlarge)
4705 memcg_oom_recover(memcg);
4710 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
4711 unsigned long long val)
4714 u64 memlimit, memswlimit, oldusage, curusage;
4715 int children = mem_cgroup_count_children(memcg);
4719 /* see mem_cgroup_resize_res_limit */
4720 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
4721 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4722 while (retry_count) {
4723 if (signal_pending(current)) {
4728 * Rather than hide all in some function, I do this in
4729 * open coded manner. You see what this really does.
4730 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4732 mutex_lock(&set_limit_mutex);
4733 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4734 if (memlimit > val) {
4736 mutex_unlock(&set_limit_mutex);
4739 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4740 if (memswlimit < val)
4742 ret = res_counter_set_limit(&memcg->memsw, val);
4744 if (memlimit == val)
4745 memcg->memsw_is_minimum = true;
4747 memcg->memsw_is_minimum = false;
4749 mutex_unlock(&set_limit_mutex);
4754 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4755 MEM_CGROUP_RECLAIM_NOSWAP |
4756 MEM_CGROUP_RECLAIM_SHRINK);
4757 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4758 /* Usage is reduced ? */
4759 if (curusage >= oldusage)
4762 oldusage = curusage;
4764 if (!ret && enlarge)
4765 memcg_oom_recover(memcg);
4769 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
4771 unsigned long *total_scanned)
4773 unsigned long nr_reclaimed = 0;
4774 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
4775 unsigned long reclaimed;
4777 struct mem_cgroup_tree_per_zone *mctz;
4778 unsigned long long excess;
4779 unsigned long nr_scanned;
4784 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
4786 * This loop can run a while, specially if mem_cgroup's continuously
4787 * keep exceeding their soft limit and putting the system under
4794 mz = mem_cgroup_largest_soft_limit_node(mctz);
4799 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
4800 gfp_mask, &nr_scanned);
4801 nr_reclaimed += reclaimed;
4802 *total_scanned += nr_scanned;
4803 spin_lock(&mctz->lock);
4806 * If we failed to reclaim anything from this memory cgroup
4807 * it is time to move on to the next cgroup
4813 * Loop until we find yet another one.
4815 * By the time we get the soft_limit lock
4816 * again, someone might have aded the
4817 * group back on the RB tree. Iterate to
4818 * make sure we get a different mem.
4819 * mem_cgroup_largest_soft_limit_node returns
4820 * NULL if no other cgroup is present on
4824 __mem_cgroup_largest_soft_limit_node(mctz);
4826 css_put(&next_mz->memcg->css);
4827 else /* next_mz == NULL or other memcg */
4831 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
4832 excess = res_counter_soft_limit_excess(&mz->memcg->res);
4834 * One school of thought says that we should not add
4835 * back the node to the tree if reclaim returns 0.
4836 * But our reclaim could return 0, simply because due
4837 * to priority we are exposing a smaller subset of
4838 * memory to reclaim from. Consider this as a longer
4841 /* If excess == 0, no tree ops */
4842 __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess);
4843 spin_unlock(&mctz->lock);
4844 css_put(&mz->memcg->css);
4847 * Could not reclaim anything and there are no more
4848 * mem cgroups to try or we seem to be looping without
4849 * reclaiming anything.
4851 if (!nr_reclaimed &&
4853 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
4855 } while (!nr_reclaimed);
4857 css_put(&next_mz->memcg->css);
4858 return nr_reclaimed;
4862 * mem_cgroup_force_empty_list - clears LRU of a group
4863 * @memcg: group to clear
4866 * @lru: lru to to clear
4868 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4869 * reclaim the pages page themselves - pages are moved to the parent (or root)
4872 static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
4873 int node, int zid, enum lru_list lru)
4875 struct lruvec *lruvec;
4876 unsigned long flags;
4877 struct list_head *list;
4881 zone = &NODE_DATA(node)->node_zones[zid];
4882 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
4883 list = &lruvec->lists[lru];
4887 struct page_cgroup *pc;
4890 spin_lock_irqsave(&zone->lru_lock, flags);
4891 if (list_empty(list)) {
4892 spin_unlock_irqrestore(&zone->lru_lock, flags);
4895 page = list_entry(list->prev, struct page, lru);
4897 list_move(&page->lru, list);
4899 spin_unlock_irqrestore(&zone->lru_lock, flags);
4902 spin_unlock_irqrestore(&zone->lru_lock, flags);
4904 pc = lookup_page_cgroup(page);
4906 if (mem_cgroup_move_parent(page, pc, memcg)) {
4907 /* found lock contention or "pc" is obsolete. */
4912 } while (!list_empty(list));
4916 * make mem_cgroup's charge to be 0 if there is no task by moving
4917 * all the charges and pages to the parent.
4918 * This enables deleting this mem_cgroup.
4920 * Caller is responsible for holding css reference on the memcg.
4922 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
4928 /* This is for making all *used* pages to be on LRU. */
4929 lru_add_drain_all();
4930 drain_all_stock_sync(memcg);
4931 mem_cgroup_start_move(memcg);
4932 for_each_node_state(node, N_MEMORY) {
4933 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4936 mem_cgroup_force_empty_list(memcg,
4941 mem_cgroup_end_move(memcg);
4942 memcg_oom_recover(memcg);
4946 * Kernel memory may not necessarily be trackable to a specific
4947 * process. So they are not migrated, and therefore we can't
4948 * expect their value to drop to 0 here.
4949 * Having res filled up with kmem only is enough.
4951 * This is a safety check because mem_cgroup_force_empty_list
4952 * could have raced with mem_cgroup_replace_page_cache callers
4953 * so the lru seemed empty but the page could have been added
4954 * right after the check. RES_USAGE should be safe as we always
4955 * charge before adding to the LRU.
4957 usage = res_counter_read_u64(&memcg->res, RES_USAGE) -
4958 res_counter_read_u64(&memcg->kmem, RES_USAGE);
4959 } while (usage > 0);
4962 static inline bool memcg_has_children(struct mem_cgroup *memcg)
4964 lockdep_assert_held(&memcg_create_mutex);
4966 * The lock does not prevent addition or deletion to the list
4967 * of children, but it prevents a new child from being
4968 * initialized based on this parent in css_online(), so it's
4969 * enough to decide whether hierarchically inherited
4970 * attributes can still be changed or not.
4972 return memcg->use_hierarchy &&
4973 !list_empty(&memcg->css.cgroup->children);
4977 * Reclaims as many pages from the given memcg as possible and moves
4978 * the rest to the parent.
4980 * Caller is responsible for holding css reference for memcg.
4982 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
4984 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
4985 struct cgroup *cgrp = memcg->css.cgroup;
4987 /* returns EBUSY if there is a task or if we come here twice. */
4988 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
4991 /* we call try-to-free pages for make this cgroup empty */
4992 lru_add_drain_all();
4993 /* try to free all pages in this cgroup */
4994 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
4997 if (signal_pending(current))
5000 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
5004 /* maybe some writeback is necessary */
5005 congestion_wait(BLK_RW_ASYNC, HZ/10);
5010 mem_cgroup_reparent_charges(memcg);
5015 static int mem_cgroup_force_empty_write(struct cgroup_subsys_state *css,
5018 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5020 if (mem_cgroup_is_root(memcg))
5022 return mem_cgroup_force_empty(memcg);
5025 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
5028 return mem_cgroup_from_css(css)->use_hierarchy;
5031 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
5032 struct cftype *cft, u64 val)
5035 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5036 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(css_parent(&memcg->css));
5038 mutex_lock(&memcg_create_mutex);
5040 if (memcg->use_hierarchy == val)
5044 * If parent's use_hierarchy is set, we can't make any modifications
5045 * in the child subtrees. If it is unset, then the change can
5046 * occur, provided the current cgroup has no children.
5048 * For the root cgroup, parent_mem is NULL, we allow value to be
5049 * set if there are no children.
5051 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
5052 (val == 1 || val == 0)) {
5053 if (list_empty(&memcg->css.cgroup->children))
5054 memcg->use_hierarchy = val;
5061 mutex_unlock(&memcg_create_mutex);
5067 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
5068 enum mem_cgroup_stat_index idx)
5070 struct mem_cgroup *iter;
5073 /* Per-cpu values can be negative, use a signed accumulator */
5074 for_each_mem_cgroup_tree(iter, memcg)
5075 val += mem_cgroup_read_stat(iter, idx);
5077 if (val < 0) /* race ? */
5082 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
5086 if (!mem_cgroup_is_root(memcg)) {
5088 return res_counter_read_u64(&memcg->res, RES_USAGE);
5090 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
5094 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
5095 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
5097 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
5098 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
5101 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
5103 return val << PAGE_SHIFT;
5106 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
5109 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5114 type = MEMFILE_TYPE(cft->private);
5115 name = MEMFILE_ATTR(cft->private);
5119 if (name == RES_USAGE)
5120 val = mem_cgroup_usage(memcg, false);
5122 val = res_counter_read_u64(&memcg->res, name);
5125 if (name == RES_USAGE)
5126 val = mem_cgroup_usage(memcg, true);
5128 val = res_counter_read_u64(&memcg->memsw, name);
5131 val = res_counter_read_u64(&memcg->kmem, name);
5140 #ifdef CONFIG_MEMCG_KMEM
5141 /* should be called with activate_kmem_mutex held */
5142 static int __memcg_activate_kmem(struct mem_cgroup *memcg,
5143 unsigned long long limit)
5148 if (memcg_kmem_is_active(memcg))
5152 * We are going to allocate memory for data shared by all memory
5153 * cgroups so let's stop accounting here.
5155 memcg_stop_kmem_account();
5158 * For simplicity, we won't allow this to be disabled. It also can't
5159 * be changed if the cgroup has children already, or if tasks had
5162 * If tasks join before we set the limit, a person looking at
5163 * kmem.usage_in_bytes will have no way to determine when it took
5164 * place, which makes the value quite meaningless.
5166 * After it first became limited, changes in the value of the limit are
5167 * of course permitted.
5169 mutex_lock(&memcg_create_mutex);
5170 if (cgroup_task_count(memcg->css.cgroup) || memcg_has_children(memcg))
5172 mutex_unlock(&memcg_create_mutex);
5176 memcg_id = ida_simple_get(&kmem_limited_groups,
5177 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
5184 * Make sure we have enough space for this cgroup in each root cache's
5187 err = memcg_update_all_caches(memcg_id + 1);
5191 memcg->kmemcg_id = memcg_id;
5192 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
5193 mutex_init(&memcg->slab_caches_mutex);
5196 * We couldn't have accounted to this cgroup, because it hasn't got the
5197 * active bit set yet, so this should succeed.
5199 err = res_counter_set_limit(&memcg->kmem, limit);
5202 static_key_slow_inc(&memcg_kmem_enabled_key);
5204 * Setting the active bit after enabling static branching will
5205 * guarantee no one starts accounting before all call sites are
5208 memcg_kmem_set_active(memcg);
5210 memcg_resume_kmem_account();
5214 ida_simple_remove(&kmem_limited_groups, memcg_id);
5218 static int memcg_activate_kmem(struct mem_cgroup *memcg,
5219 unsigned long long limit)
5223 mutex_lock(&activate_kmem_mutex);
5224 ret = __memcg_activate_kmem(memcg, limit);
5225 mutex_unlock(&activate_kmem_mutex);
5229 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
5230 unsigned long long val)
5234 if (!memcg_kmem_is_active(memcg))
5235 ret = memcg_activate_kmem(memcg, val);
5237 ret = res_counter_set_limit(&memcg->kmem, val);
5241 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
5244 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5249 mutex_lock(&activate_kmem_mutex);
5251 * If the parent cgroup is not kmem-active now, it cannot be activated
5252 * after this point, because it has at least one child already.
5254 if (memcg_kmem_is_active(parent))
5255 ret = __memcg_activate_kmem(memcg, RES_COUNTER_MAX);
5256 mutex_unlock(&activate_kmem_mutex);
5260 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
5261 unsigned long long val)
5265 #endif /* CONFIG_MEMCG_KMEM */
5268 * The user of this function is...
5271 static int mem_cgroup_write(struct cgroup_subsys_state *css, struct cftype *cft,
5274 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5277 unsigned long long val;
5280 type = MEMFILE_TYPE(cft->private);
5281 name = MEMFILE_ATTR(cft->private);
5285 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
5289 /* This function does all necessary parse...reuse it */
5290 ret = res_counter_memparse_write_strategy(buffer, &val);
5294 ret = mem_cgroup_resize_limit(memcg, val);
5295 else if (type == _MEMSWAP)
5296 ret = mem_cgroup_resize_memsw_limit(memcg, val);
5297 else if (type == _KMEM)
5298 ret = memcg_update_kmem_limit(memcg, val);
5302 case RES_SOFT_LIMIT:
5303 ret = res_counter_memparse_write_strategy(buffer, &val);
5307 * For memsw, soft limits are hard to implement in terms
5308 * of semantics, for now, we support soft limits for
5309 * control without swap
5312 ret = res_counter_set_soft_limit(&memcg->res, val);
5317 ret = -EINVAL; /* should be BUG() ? */
5323 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
5324 unsigned long long *mem_limit, unsigned long long *memsw_limit)
5326 unsigned long long min_limit, min_memsw_limit, tmp;
5328 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
5329 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5330 if (!memcg->use_hierarchy)
5333 while (css_parent(&memcg->css)) {
5334 memcg = mem_cgroup_from_css(css_parent(&memcg->css));
5335 if (!memcg->use_hierarchy)
5337 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
5338 min_limit = min(min_limit, tmp);
5339 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5340 min_memsw_limit = min(min_memsw_limit, tmp);
5343 *mem_limit = min_limit;
5344 *memsw_limit = min_memsw_limit;
5347 static int mem_cgroup_reset(struct cgroup_subsys_state *css, unsigned int event)
5349 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5353 type = MEMFILE_TYPE(event);
5354 name = MEMFILE_ATTR(event);
5359 res_counter_reset_max(&memcg->res);
5360 else if (type == _MEMSWAP)
5361 res_counter_reset_max(&memcg->memsw);
5362 else if (type == _KMEM)
5363 res_counter_reset_max(&memcg->kmem);
5369 res_counter_reset_failcnt(&memcg->res);
5370 else if (type == _MEMSWAP)
5371 res_counter_reset_failcnt(&memcg->memsw);
5372 else if (type == _KMEM)
5373 res_counter_reset_failcnt(&memcg->kmem);
5382 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
5385 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
5389 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5390 struct cftype *cft, u64 val)
5392 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5394 if (val >= (1 << NR_MOVE_TYPE))
5398 * No kind of locking is needed in here, because ->can_attach() will
5399 * check this value once in the beginning of the process, and then carry
5400 * on with stale data. This means that changes to this value will only
5401 * affect task migrations starting after the change.
5403 memcg->move_charge_at_immigrate = val;
5407 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5408 struct cftype *cft, u64 val)
5415 static int memcg_numa_stat_show(struct seq_file *m, void *v)
5419 unsigned int lru_mask;
5422 static const struct numa_stat stats[] = {
5423 { "total", LRU_ALL },
5424 { "file", LRU_ALL_FILE },
5425 { "anon", LRU_ALL_ANON },
5426 { "unevictable", BIT(LRU_UNEVICTABLE) },
5428 const struct numa_stat *stat;
5431 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5433 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
5434 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
5435 seq_printf(m, "%s=%lu", stat->name, nr);
5436 for_each_node_state(nid, N_MEMORY) {
5437 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5439 seq_printf(m, " N%d=%lu", nid, nr);
5444 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
5445 struct mem_cgroup *iter;
5448 for_each_mem_cgroup_tree(iter, memcg)
5449 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
5450 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
5451 for_each_node_state(nid, N_MEMORY) {
5453 for_each_mem_cgroup_tree(iter, memcg)
5454 nr += mem_cgroup_node_nr_lru_pages(
5455 iter, nid, stat->lru_mask);
5456 seq_printf(m, " N%d=%lu", nid, nr);
5463 #endif /* CONFIG_NUMA */
5465 static inline void mem_cgroup_lru_names_not_uptodate(void)
5467 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
5470 static int memcg_stat_show(struct seq_file *m, void *v)
5472 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5473 struct mem_cgroup *mi;
5476 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5477 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5479 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
5480 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
5483 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
5484 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
5485 mem_cgroup_read_events(memcg, i));
5487 for (i = 0; i < NR_LRU_LISTS; i++)
5488 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
5489 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
5491 /* Hierarchical information */
5493 unsigned long long limit, memsw_limit;
5494 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
5495 seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
5496 if (do_swap_account)
5497 seq_printf(m, "hierarchical_memsw_limit %llu\n",
5501 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5504 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5506 for_each_mem_cgroup_tree(mi, memcg)
5507 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
5508 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
5511 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
5512 unsigned long long val = 0;
5514 for_each_mem_cgroup_tree(mi, memcg)
5515 val += mem_cgroup_read_events(mi, i);
5516 seq_printf(m, "total_%s %llu\n",
5517 mem_cgroup_events_names[i], val);
5520 for (i = 0; i < NR_LRU_LISTS; i++) {
5521 unsigned long long val = 0;
5523 for_each_mem_cgroup_tree(mi, memcg)
5524 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
5525 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
5528 #ifdef CONFIG_DEBUG_VM
5531 struct mem_cgroup_per_zone *mz;
5532 struct zone_reclaim_stat *rstat;
5533 unsigned long recent_rotated[2] = {0, 0};
5534 unsigned long recent_scanned[2] = {0, 0};
5536 for_each_online_node(nid)
5537 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
5538 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
5539 rstat = &mz->lruvec.reclaim_stat;
5541 recent_rotated[0] += rstat->recent_rotated[0];
5542 recent_rotated[1] += rstat->recent_rotated[1];
5543 recent_scanned[0] += rstat->recent_scanned[0];
5544 recent_scanned[1] += rstat->recent_scanned[1];
5546 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
5547 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
5548 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
5549 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
5556 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
5559 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5561 return mem_cgroup_swappiness(memcg);
5564 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
5565 struct cftype *cft, u64 val)
5567 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5568 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(&memcg->css));
5570 if (val > 100 || !parent)
5573 mutex_lock(&memcg_create_mutex);
5575 /* If under hierarchy, only empty-root can set this value */
5576 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5577 mutex_unlock(&memcg_create_mutex);
5581 memcg->swappiness = val;
5583 mutex_unlock(&memcg_create_mutex);
5588 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
5590 struct mem_cgroup_threshold_ary *t;
5596 t = rcu_dereference(memcg->thresholds.primary);
5598 t = rcu_dereference(memcg->memsw_thresholds.primary);
5603 usage = mem_cgroup_usage(memcg, swap);
5606 * current_threshold points to threshold just below or equal to usage.
5607 * If it's not true, a threshold was crossed after last
5608 * call of __mem_cgroup_threshold().
5610 i = t->current_threshold;
5613 * Iterate backward over array of thresholds starting from
5614 * current_threshold and check if a threshold is crossed.
5615 * If none of thresholds below usage is crossed, we read
5616 * only one element of the array here.
5618 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
5619 eventfd_signal(t->entries[i].eventfd, 1);
5621 /* i = current_threshold + 1 */
5625 * Iterate forward over array of thresholds starting from
5626 * current_threshold+1 and check if a threshold is crossed.
5627 * If none of thresholds above usage is crossed, we read
5628 * only one element of the array here.
5630 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
5631 eventfd_signal(t->entries[i].eventfd, 1);
5633 /* Update current_threshold */
5634 t->current_threshold = i - 1;
5639 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
5642 __mem_cgroup_threshold(memcg, false);
5643 if (do_swap_account)
5644 __mem_cgroup_threshold(memcg, true);
5646 memcg = parent_mem_cgroup(memcg);
5650 static int compare_thresholds(const void *a, const void *b)
5652 const struct mem_cgroup_threshold *_a = a;
5653 const struct mem_cgroup_threshold *_b = b;
5655 if (_a->threshold > _b->threshold)
5658 if (_a->threshold < _b->threshold)
5664 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
5666 struct mem_cgroup_eventfd_list *ev;
5668 list_for_each_entry(ev, &memcg->oom_notify, list)
5669 eventfd_signal(ev->eventfd, 1);
5673 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
5675 struct mem_cgroup *iter;
5677 for_each_mem_cgroup_tree(iter, memcg)
5678 mem_cgroup_oom_notify_cb(iter);
5681 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
5682 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
5684 struct mem_cgroup_thresholds *thresholds;
5685 struct mem_cgroup_threshold_ary *new;
5686 u64 threshold, usage;
5689 ret = res_counter_memparse_write_strategy(args, &threshold);
5693 mutex_lock(&memcg->thresholds_lock);
5696 thresholds = &memcg->thresholds;
5697 else if (type == _MEMSWAP)
5698 thresholds = &memcg->memsw_thresholds;
5702 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5704 /* Check if a threshold crossed before adding a new one */
5705 if (thresholds->primary)
5706 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5708 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
5710 /* Allocate memory for new array of thresholds */
5711 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
5719 /* Copy thresholds (if any) to new array */
5720 if (thresholds->primary) {
5721 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
5722 sizeof(struct mem_cgroup_threshold));
5725 /* Add new threshold */
5726 new->entries[size - 1].eventfd = eventfd;
5727 new->entries[size - 1].threshold = threshold;
5729 /* Sort thresholds. Registering of new threshold isn't time-critical */
5730 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
5731 compare_thresholds, NULL);
5733 /* Find current threshold */
5734 new->current_threshold = -1;
5735 for (i = 0; i < size; i++) {
5736 if (new->entries[i].threshold <= usage) {
5738 * new->current_threshold will not be used until
5739 * rcu_assign_pointer(), so it's safe to increment
5742 ++new->current_threshold;
5747 /* Free old spare buffer and save old primary buffer as spare */
5748 kfree(thresholds->spare);
5749 thresholds->spare = thresholds->primary;
5751 rcu_assign_pointer(thresholds->primary, new);
5753 /* To be sure that nobody uses thresholds */
5757 mutex_unlock(&memcg->thresholds_lock);
5762 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
5763 struct eventfd_ctx *eventfd, const char *args)
5765 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
5768 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
5769 struct eventfd_ctx *eventfd, const char *args)
5771 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
5774 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
5775 struct eventfd_ctx *eventfd, enum res_type type)
5777 struct mem_cgroup_thresholds *thresholds;
5778 struct mem_cgroup_threshold_ary *new;
5782 mutex_lock(&memcg->thresholds_lock);
5784 thresholds = &memcg->thresholds;
5785 else if (type == _MEMSWAP)
5786 thresholds = &memcg->memsw_thresholds;
5790 if (!thresholds->primary)
5793 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5795 /* Check if a threshold crossed before removing */
5796 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5798 /* Calculate new number of threshold */
5800 for (i = 0; i < thresholds->primary->size; i++) {
5801 if (thresholds->primary->entries[i].eventfd != eventfd)
5805 new = thresholds->spare;
5807 /* Set thresholds array to NULL if we don't have thresholds */
5816 /* Copy thresholds and find current threshold */
5817 new->current_threshold = -1;
5818 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
5819 if (thresholds->primary->entries[i].eventfd == eventfd)
5822 new->entries[j] = thresholds->primary->entries[i];
5823 if (new->entries[j].threshold <= usage) {
5825 * new->current_threshold will not be used
5826 * until rcu_assign_pointer(), so it's safe to increment
5829 ++new->current_threshold;
5835 /* Swap primary and spare array */
5836 thresholds->spare = thresholds->primary;
5837 /* If all events are unregistered, free the spare array */
5839 kfree(thresholds->spare);
5840 thresholds->spare = NULL;
5843 rcu_assign_pointer(thresholds->primary, new);
5845 /* To be sure that nobody uses thresholds */
5848 mutex_unlock(&memcg->thresholds_lock);
5851 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
5852 struct eventfd_ctx *eventfd)
5854 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
5857 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
5858 struct eventfd_ctx *eventfd)
5860 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
5863 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
5864 struct eventfd_ctx *eventfd, const char *args)
5866 struct mem_cgroup_eventfd_list *event;
5868 event = kmalloc(sizeof(*event), GFP_KERNEL);
5872 spin_lock(&memcg_oom_lock);
5874 event->eventfd = eventfd;
5875 list_add(&event->list, &memcg->oom_notify);
5877 /* already in OOM ? */
5878 if (atomic_read(&memcg->under_oom))
5879 eventfd_signal(eventfd, 1);
5880 spin_unlock(&memcg_oom_lock);
5885 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
5886 struct eventfd_ctx *eventfd)
5888 struct mem_cgroup_eventfd_list *ev, *tmp;
5890 spin_lock(&memcg_oom_lock);
5892 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
5893 if (ev->eventfd == eventfd) {
5894 list_del(&ev->list);
5899 spin_unlock(&memcg_oom_lock);
5902 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
5904 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
5906 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
5907 seq_printf(sf, "under_oom %d\n", (bool)atomic_read(&memcg->under_oom));
5911 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
5912 struct cftype *cft, u64 val)
5914 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5915 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(&memcg->css));
5917 /* cannot set to root cgroup and only 0 and 1 are allowed */
5918 if (!parent || !((val == 0) || (val == 1)))
5921 mutex_lock(&memcg_create_mutex);
5922 /* oom-kill-disable is a flag for subhierarchy. */
5923 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5924 mutex_unlock(&memcg_create_mutex);
5927 memcg->oom_kill_disable = val;
5929 memcg_oom_recover(memcg);
5930 mutex_unlock(&memcg_create_mutex);
5934 #ifdef CONFIG_MEMCG_KMEM
5935 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5939 memcg->kmemcg_id = -1;
5940 ret = memcg_propagate_kmem(memcg);
5944 return mem_cgroup_sockets_init(memcg, ss);
5947 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5949 mem_cgroup_sockets_destroy(memcg);
5952 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5954 if (!memcg_kmem_is_active(memcg))
5958 * kmem charges can outlive the cgroup. In the case of slab
5959 * pages, for instance, a page contain objects from various
5960 * processes. As we prevent from taking a reference for every
5961 * such allocation we have to be careful when doing uncharge
5962 * (see memcg_uncharge_kmem) and here during offlining.
5964 * The idea is that that only the _last_ uncharge which sees
5965 * the dead memcg will drop the last reference. An additional
5966 * reference is taken here before the group is marked dead
5967 * which is then paired with css_put during uncharge resp. here.
5969 * Although this might sound strange as this path is called from
5970 * css_offline() when the referencemight have dropped down to 0
5971 * and shouldn't be incremented anymore (css_tryget would fail)
5972 * we do not have other options because of the kmem allocations
5975 css_get(&memcg->css);
5977 memcg_kmem_mark_dead(memcg);
5979 if (res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0)
5982 if (memcg_kmem_test_and_clear_dead(memcg))
5983 css_put(&memcg->css);
5986 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5991 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5995 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
6001 * DO NOT USE IN NEW FILES.
6003 * "cgroup.event_control" implementation.
6005 * This is way over-engineered. It tries to support fully configurable
6006 * events for each user. Such level of flexibility is completely
6007 * unnecessary especially in the light of the planned unified hierarchy.
6009 * Please deprecate this and replace with something simpler if at all
6014 * Unregister event and free resources.
6016 * Gets called from workqueue.
6018 static void memcg_event_remove(struct work_struct *work)
6020 struct mem_cgroup_event *event =
6021 container_of(work, struct mem_cgroup_event, remove);
6022 struct mem_cgroup *memcg = event->memcg;
6024 remove_wait_queue(event->wqh, &event->wait);
6026 event->unregister_event(memcg, event->eventfd);
6028 /* Notify userspace the event is going away. */
6029 eventfd_signal(event->eventfd, 1);
6031 eventfd_ctx_put(event->eventfd);
6033 css_put(&memcg->css);
6037 * Gets called on POLLHUP on eventfd when user closes it.
6039 * Called with wqh->lock held and interrupts disabled.
6041 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
6042 int sync, void *key)
6044 struct mem_cgroup_event *event =
6045 container_of(wait, struct mem_cgroup_event, wait);
6046 struct mem_cgroup *memcg = event->memcg;
6047 unsigned long flags = (unsigned long)key;
6049 if (flags & POLLHUP) {
6051 * If the event has been detached at cgroup removal, we
6052 * can simply return knowing the other side will cleanup
6055 * We can't race against event freeing since the other
6056 * side will require wqh->lock via remove_wait_queue(),
6059 spin_lock(&memcg->event_list_lock);
6060 if (!list_empty(&event->list)) {
6061 list_del_init(&event->list);
6063 * We are in atomic context, but cgroup_event_remove()
6064 * may sleep, so we have to call it in workqueue.
6066 schedule_work(&event->remove);
6068 spin_unlock(&memcg->event_list_lock);
6074 static void memcg_event_ptable_queue_proc(struct file *file,
6075 wait_queue_head_t *wqh, poll_table *pt)
6077 struct mem_cgroup_event *event =
6078 container_of(pt, struct mem_cgroup_event, pt);
6081 add_wait_queue(wqh, &event->wait);
6085 * DO NOT USE IN NEW FILES.
6087 * Parse input and register new cgroup event handler.
6089 * Input must be in format '<event_fd> <control_fd> <args>'.
6090 * Interpretation of args is defined by control file implementation.
6092 static int memcg_write_event_control(struct cgroup_subsys_state *css,
6093 struct cftype *cft, const char *buffer)
6095 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6096 struct mem_cgroup_event *event;
6097 struct cgroup_subsys_state *cfile_css;
6098 unsigned int efd, cfd;
6105 efd = simple_strtoul(buffer, &endp, 10);
6110 cfd = simple_strtoul(buffer, &endp, 10);
6111 if ((*endp != ' ') && (*endp != '\0'))
6115 event = kzalloc(sizeof(*event), GFP_KERNEL);
6119 event->memcg = memcg;
6120 INIT_LIST_HEAD(&event->list);
6121 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
6122 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
6123 INIT_WORK(&event->remove, memcg_event_remove);
6131 event->eventfd = eventfd_ctx_fileget(efile.file);
6132 if (IS_ERR(event->eventfd)) {
6133 ret = PTR_ERR(event->eventfd);
6140 goto out_put_eventfd;
6143 /* the process need read permission on control file */
6144 /* AV: shouldn't we check that it's been opened for read instead? */
6145 ret = inode_permission(file_inode(cfile.file), MAY_READ);
6150 * Determine the event callbacks and set them in @event. This used
6151 * to be done via struct cftype but cgroup core no longer knows
6152 * about these events. The following is crude but the whole thing
6153 * is for compatibility anyway.
6155 * DO NOT ADD NEW FILES.
6157 name = cfile.file->f_dentry->d_name.name;
6159 if (!strcmp(name, "memory.usage_in_bytes")) {
6160 event->register_event = mem_cgroup_usage_register_event;
6161 event->unregister_event = mem_cgroup_usage_unregister_event;
6162 } else if (!strcmp(name, "memory.oom_control")) {
6163 event->register_event = mem_cgroup_oom_register_event;
6164 event->unregister_event = mem_cgroup_oom_unregister_event;
6165 } else if (!strcmp(name, "memory.pressure_level")) {
6166 event->register_event = vmpressure_register_event;
6167 event->unregister_event = vmpressure_unregister_event;
6168 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
6169 event->register_event = memsw_cgroup_usage_register_event;
6170 event->unregister_event = memsw_cgroup_usage_unregister_event;
6177 * Verify @cfile should belong to @css. Also, remaining events are
6178 * automatically removed on cgroup destruction but the removal is
6179 * asynchronous, so take an extra ref on @css.
6184 cfile_css = css_from_dir(cfile.file->f_dentry->d_parent,
6185 &mem_cgroup_subsys);
6186 if (cfile_css == css && css_tryget(css))
6193 ret = event->register_event(memcg, event->eventfd, buffer);
6197 efile.file->f_op->poll(efile.file, &event->pt);
6199 spin_lock(&memcg->event_list_lock);
6200 list_add(&event->list, &memcg->event_list);
6201 spin_unlock(&memcg->event_list_lock);
6213 eventfd_ctx_put(event->eventfd);
6222 static struct cftype mem_cgroup_files[] = {
6224 .name = "usage_in_bytes",
6225 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
6226 .read_u64 = mem_cgroup_read_u64,
6229 .name = "max_usage_in_bytes",
6230 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
6231 .trigger = mem_cgroup_reset,
6232 .read_u64 = mem_cgroup_read_u64,
6235 .name = "limit_in_bytes",
6236 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
6237 .write_string = mem_cgroup_write,
6238 .read_u64 = mem_cgroup_read_u64,
6241 .name = "soft_limit_in_bytes",
6242 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
6243 .write_string = mem_cgroup_write,
6244 .read_u64 = mem_cgroup_read_u64,
6248 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
6249 .trigger = mem_cgroup_reset,
6250 .read_u64 = mem_cgroup_read_u64,
6254 .seq_show = memcg_stat_show,
6257 .name = "force_empty",
6258 .trigger = mem_cgroup_force_empty_write,
6261 .name = "use_hierarchy",
6262 .flags = CFTYPE_INSANE,
6263 .write_u64 = mem_cgroup_hierarchy_write,
6264 .read_u64 = mem_cgroup_hierarchy_read,
6267 .name = "cgroup.event_control", /* XXX: for compat */
6268 .write_string = memcg_write_event_control,
6269 .flags = CFTYPE_NO_PREFIX,
6273 .name = "swappiness",
6274 .read_u64 = mem_cgroup_swappiness_read,
6275 .write_u64 = mem_cgroup_swappiness_write,
6278 .name = "move_charge_at_immigrate",
6279 .read_u64 = mem_cgroup_move_charge_read,
6280 .write_u64 = mem_cgroup_move_charge_write,
6283 .name = "oom_control",
6284 .seq_show = mem_cgroup_oom_control_read,
6285 .write_u64 = mem_cgroup_oom_control_write,
6286 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
6289 .name = "pressure_level",
6293 .name = "numa_stat",
6294 .seq_show = memcg_numa_stat_show,
6297 #ifdef CONFIG_MEMCG_KMEM
6299 .name = "kmem.limit_in_bytes",
6300 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
6301 .write_string = mem_cgroup_write,
6302 .read_u64 = mem_cgroup_read_u64,
6305 .name = "kmem.usage_in_bytes",
6306 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
6307 .read_u64 = mem_cgroup_read_u64,
6310 .name = "kmem.failcnt",
6311 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
6312 .trigger = mem_cgroup_reset,
6313 .read_u64 = mem_cgroup_read_u64,
6316 .name = "kmem.max_usage_in_bytes",
6317 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
6318 .trigger = mem_cgroup_reset,
6319 .read_u64 = mem_cgroup_read_u64,
6321 #ifdef CONFIG_SLABINFO
6323 .name = "kmem.slabinfo",
6324 .seq_show = mem_cgroup_slabinfo_read,
6328 { }, /* terminate */
6331 #ifdef CONFIG_MEMCG_SWAP
6332 static struct cftype memsw_cgroup_files[] = {
6334 .name = "memsw.usage_in_bytes",
6335 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6336 .read_u64 = mem_cgroup_read_u64,
6339 .name = "memsw.max_usage_in_bytes",
6340 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6341 .trigger = mem_cgroup_reset,
6342 .read_u64 = mem_cgroup_read_u64,
6345 .name = "memsw.limit_in_bytes",
6346 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6347 .write_string = mem_cgroup_write,
6348 .read_u64 = mem_cgroup_read_u64,
6351 .name = "memsw.failcnt",
6352 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6353 .trigger = mem_cgroup_reset,
6354 .read_u64 = mem_cgroup_read_u64,
6356 { }, /* terminate */
6359 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6361 struct mem_cgroup_per_node *pn;
6362 struct mem_cgroup_per_zone *mz;
6363 int zone, tmp = node;
6365 * This routine is called against possible nodes.
6366 * But it's BUG to call kmalloc() against offline node.
6368 * TODO: this routine can waste much memory for nodes which will
6369 * never be onlined. It's better to use memory hotplug callback
6372 if (!node_state(node, N_NORMAL_MEMORY))
6374 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
6378 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6379 mz = &pn->zoneinfo[zone];
6380 lruvec_init(&mz->lruvec);
6381 mz->usage_in_excess = 0;
6382 mz->on_tree = false;
6385 memcg->nodeinfo[node] = pn;
6389 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6391 kfree(memcg->nodeinfo[node]);
6394 static struct mem_cgroup *mem_cgroup_alloc(void)
6396 struct mem_cgroup *memcg;
6399 size = sizeof(struct mem_cgroup);
6400 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
6402 memcg = kzalloc(size, GFP_KERNEL);
6406 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
6409 spin_lock_init(&memcg->pcp_counter_lock);
6418 * At destroying mem_cgroup, references from swap_cgroup can remain.
6419 * (scanning all at force_empty is too costly...)
6421 * Instead of clearing all references at force_empty, we remember
6422 * the number of reference from swap_cgroup and free mem_cgroup when
6423 * it goes down to 0.
6425 * Removal of cgroup itself succeeds regardless of refs from swap.
6428 static void __mem_cgroup_free(struct mem_cgroup *memcg)
6432 mem_cgroup_remove_from_trees(memcg);
6435 free_mem_cgroup_per_zone_info(memcg, node);
6437 free_percpu(memcg->stat);
6440 * We need to make sure that (at least for now), the jump label
6441 * destruction code runs outside of the cgroup lock. This is because
6442 * get_online_cpus(), which is called from the static_branch update,
6443 * can't be called inside the cgroup_lock. cpusets are the ones
6444 * enforcing this dependency, so if they ever change, we might as well.
6446 * schedule_work() will guarantee this happens. Be careful if you need
6447 * to move this code around, and make sure it is outside
6450 disarm_static_keys(memcg);
6455 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6457 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
6459 if (!memcg->res.parent)
6461 return mem_cgroup_from_res_counter(memcg->res.parent, res);
6463 EXPORT_SYMBOL(parent_mem_cgroup);
6465 static void __init mem_cgroup_soft_limit_tree_init(void)
6467 struct mem_cgroup_tree_per_node *rtpn;
6468 struct mem_cgroup_tree_per_zone *rtpz;
6469 int tmp, node, zone;
6471 for_each_node(node) {
6473 if (!node_state(node, N_NORMAL_MEMORY))
6475 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
6478 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6480 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6481 rtpz = &rtpn->rb_tree_per_zone[zone];
6482 rtpz->rb_root = RB_ROOT;
6483 spin_lock_init(&rtpz->lock);
6488 static struct cgroup_subsys_state * __ref
6489 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
6491 struct mem_cgroup *memcg;
6492 long error = -ENOMEM;
6495 memcg = mem_cgroup_alloc();
6497 return ERR_PTR(error);
6500 if (alloc_mem_cgroup_per_zone_info(memcg, node))
6504 if (parent_css == NULL) {
6505 root_mem_cgroup = memcg;
6506 res_counter_init(&memcg->res, NULL);
6507 res_counter_init(&memcg->memsw, NULL);
6508 res_counter_init(&memcg->kmem, NULL);
6511 memcg->last_scanned_node = MAX_NUMNODES;
6512 INIT_LIST_HEAD(&memcg->oom_notify);
6513 memcg->move_charge_at_immigrate = 0;
6514 mutex_init(&memcg->thresholds_lock);
6515 spin_lock_init(&memcg->move_lock);
6516 vmpressure_init(&memcg->vmpressure);
6517 INIT_LIST_HEAD(&memcg->event_list);
6518 spin_lock_init(&memcg->event_list_lock);
6523 __mem_cgroup_free(memcg);
6524 return ERR_PTR(error);
6528 mem_cgroup_css_online(struct cgroup_subsys_state *css)
6530 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6531 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(css));
6533 if (css->cgroup->id > MEM_CGROUP_ID_MAX)
6539 mutex_lock(&memcg_create_mutex);
6541 memcg->use_hierarchy = parent->use_hierarchy;
6542 memcg->oom_kill_disable = parent->oom_kill_disable;
6543 memcg->swappiness = mem_cgroup_swappiness(parent);
6545 if (parent->use_hierarchy) {
6546 res_counter_init(&memcg->res, &parent->res);
6547 res_counter_init(&memcg->memsw, &parent->memsw);
6548 res_counter_init(&memcg->kmem, &parent->kmem);
6551 * No need to take a reference to the parent because cgroup
6552 * core guarantees its existence.
6555 res_counter_init(&memcg->res, NULL);
6556 res_counter_init(&memcg->memsw, NULL);
6557 res_counter_init(&memcg->kmem, NULL);
6559 * Deeper hierachy with use_hierarchy == false doesn't make
6560 * much sense so let cgroup subsystem know about this
6561 * unfortunate state in our controller.
6563 if (parent != root_mem_cgroup)
6564 mem_cgroup_subsys.broken_hierarchy = true;
6566 mutex_unlock(&memcg_create_mutex);
6568 return memcg_init_kmem(memcg, &mem_cgroup_subsys);
6572 * Announce all parents that a group from their hierarchy is gone.
6574 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup *memcg)
6576 struct mem_cgroup *parent = memcg;
6578 while ((parent = parent_mem_cgroup(parent)))
6579 mem_cgroup_iter_invalidate(parent);
6582 * if the root memcg is not hierarchical we have to check it
6585 if (!root_mem_cgroup->use_hierarchy)
6586 mem_cgroup_iter_invalidate(root_mem_cgroup);
6589 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
6591 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6592 struct mem_cgroup_event *event, *tmp;
6595 * Unregister events and notify userspace.
6596 * Notify userspace about cgroup removing only after rmdir of cgroup
6597 * directory to avoid race between userspace and kernelspace.
6599 spin_lock(&memcg->event_list_lock);
6600 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
6601 list_del_init(&event->list);
6602 schedule_work(&event->remove);
6604 spin_unlock(&memcg->event_list_lock);
6606 kmem_cgroup_css_offline(memcg);
6608 mem_cgroup_invalidate_reclaim_iterators(memcg);
6609 mem_cgroup_reparent_charges(memcg);
6610 mem_cgroup_destroy_all_caches(memcg);
6611 vmpressure_cleanup(&memcg->vmpressure);
6614 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
6616 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6618 * XXX: css_offline() would be where we should reparent all
6619 * memory to prepare the cgroup for destruction. However,
6620 * memcg does not do css_tryget() and res_counter charging
6621 * under the same RCU lock region, which means that charging
6622 * could race with offlining. Offlining only happens to
6623 * cgroups with no tasks in them but charges can show up
6624 * without any tasks from the swapin path when the target
6625 * memcg is looked up from the swapout record and not from the
6626 * current task as it usually is. A race like this can leak
6627 * charges and put pages with stale cgroup pointers into
6631 * lookup_swap_cgroup_id()
6633 * mem_cgroup_lookup()
6636 * disable css_tryget()
6639 * reparent_charges()
6640 * res_counter_charge()
6643 * pc->mem_cgroup = dead memcg
6646 * The bulk of the charges are still moved in offline_css() to
6647 * avoid pinning a lot of pages in case a long-term reference
6648 * like a swapout record is deferring the css_free() to long
6649 * after offlining. But this makes sure we catch any charges
6650 * made after offlining:
6652 mem_cgroup_reparent_charges(memcg);
6654 memcg_destroy_kmem(memcg);
6655 __mem_cgroup_free(memcg);
6659 /* Handlers for move charge at task migration. */
6660 #define PRECHARGE_COUNT_AT_ONCE 256
6661 static int mem_cgroup_do_precharge(unsigned long count)
6664 int batch_count = PRECHARGE_COUNT_AT_ONCE;
6665 struct mem_cgroup *memcg = mc.to;
6667 if (mem_cgroup_is_root(memcg)) {
6668 mc.precharge += count;
6669 /* we don't need css_get for root */
6672 /* try to charge at once */
6674 struct res_counter *dummy;
6676 * "memcg" cannot be under rmdir() because we've already checked
6677 * by cgroup_lock_live_cgroup() that it is not removed and we
6678 * are still under the same cgroup_mutex. So we can postpone
6681 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
6683 if (do_swap_account && res_counter_charge(&memcg->memsw,
6684 PAGE_SIZE * count, &dummy)) {
6685 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
6688 mc.precharge += count;
6692 /* fall back to one by one charge */
6694 if (signal_pending(current)) {
6698 if (!batch_count--) {
6699 batch_count = PRECHARGE_COUNT_AT_ONCE;
6702 ret = __mem_cgroup_try_charge(NULL,
6703 GFP_KERNEL, 1, &memcg, false);
6705 /* mem_cgroup_clear_mc() will do uncharge later */
6713 * get_mctgt_type - get target type of moving charge
6714 * @vma: the vma the pte to be checked belongs
6715 * @addr: the address corresponding to the pte to be checked
6716 * @ptent: the pte to be checked
6717 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6720 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6721 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6722 * move charge. if @target is not NULL, the page is stored in target->page
6723 * with extra refcnt got(Callers should handle it).
6724 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6725 * target for charge migration. if @target is not NULL, the entry is stored
6728 * Called with pte lock held.
6735 enum mc_target_type {
6741 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
6742 unsigned long addr, pte_t ptent)
6744 struct page *page = vm_normal_page(vma, addr, ptent);
6746 if (!page || !page_mapped(page))
6748 if (PageAnon(page)) {
6749 /* we don't move shared anon */
6752 } else if (!move_file())
6753 /* we ignore mapcount for file pages */
6755 if (!get_page_unless_zero(page))
6762 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6763 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6765 struct page *page = NULL;
6766 swp_entry_t ent = pte_to_swp_entry(ptent);
6768 if (!move_anon() || non_swap_entry(ent))
6771 * Because lookup_swap_cache() updates some statistics counter,
6772 * we call find_get_page() with swapper_space directly.
6774 page = find_get_page(swap_address_space(ent), ent.val);
6775 if (do_swap_account)
6776 entry->val = ent.val;
6781 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6782 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6788 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
6789 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6791 struct page *page = NULL;
6792 struct address_space *mapping;
6795 if (!vma->vm_file) /* anonymous vma */
6800 mapping = vma->vm_file->f_mapping;
6801 if (pte_none(ptent))
6802 pgoff = linear_page_index(vma, addr);
6803 else /* pte_file(ptent) is true */
6804 pgoff = pte_to_pgoff(ptent);
6806 /* page is moved even if it's not RSS of this task(page-faulted). */
6807 page = find_get_page(mapping, pgoff);
6810 /* shmem/tmpfs may report page out on swap: account for that too. */
6811 if (radix_tree_exceptional_entry(page)) {
6812 swp_entry_t swap = radix_to_swp_entry(page);
6813 if (do_swap_account)
6815 page = find_get_page(swap_address_space(swap), swap.val);
6821 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
6822 unsigned long addr, pte_t ptent, union mc_target *target)
6824 struct page *page = NULL;
6825 struct page_cgroup *pc;
6826 enum mc_target_type ret = MC_TARGET_NONE;
6827 swp_entry_t ent = { .val = 0 };
6829 if (pte_present(ptent))
6830 page = mc_handle_present_pte(vma, addr, ptent);
6831 else if (is_swap_pte(ptent))
6832 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
6833 else if (pte_none(ptent) || pte_file(ptent))
6834 page = mc_handle_file_pte(vma, addr, ptent, &ent);
6836 if (!page && !ent.val)
6839 pc = lookup_page_cgroup(page);
6841 * Do only loose check w/o page_cgroup lock.
6842 * mem_cgroup_move_account() checks the pc is valid or not under
6845 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6846 ret = MC_TARGET_PAGE;
6848 target->page = page;
6850 if (!ret || !target)
6853 /* There is a swap entry and a page doesn't exist or isn't charged */
6854 if (ent.val && !ret &&
6855 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
6856 ret = MC_TARGET_SWAP;
6863 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6865 * We don't consider swapping or file mapped pages because THP does not
6866 * support them for now.
6867 * Caller should make sure that pmd_trans_huge(pmd) is true.
6869 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6870 unsigned long addr, pmd_t pmd, union mc_target *target)
6872 struct page *page = NULL;
6873 struct page_cgroup *pc;
6874 enum mc_target_type ret = MC_TARGET_NONE;
6876 page = pmd_page(pmd);
6877 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
6880 pc = lookup_page_cgroup(page);
6881 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6882 ret = MC_TARGET_PAGE;
6885 target->page = page;
6891 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6892 unsigned long addr, pmd_t pmd, union mc_target *target)
6894 return MC_TARGET_NONE;
6898 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
6899 unsigned long addr, unsigned long end,
6900 struct mm_walk *walk)
6902 struct vm_area_struct *vma = walk->private;
6906 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
6907 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
6908 mc.precharge += HPAGE_PMD_NR;
6913 if (pmd_trans_unstable(pmd))
6915 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6916 for (; addr != end; pte++, addr += PAGE_SIZE)
6917 if (get_mctgt_type(vma, addr, *pte, NULL))
6918 mc.precharge++; /* increment precharge temporarily */
6919 pte_unmap_unlock(pte - 1, ptl);
6925 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
6927 unsigned long precharge;
6928 struct vm_area_struct *vma;
6930 down_read(&mm->mmap_sem);
6931 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6932 struct mm_walk mem_cgroup_count_precharge_walk = {
6933 .pmd_entry = mem_cgroup_count_precharge_pte_range,
6937 if (is_vm_hugetlb_page(vma))
6939 walk_page_range(vma->vm_start, vma->vm_end,
6940 &mem_cgroup_count_precharge_walk);
6942 up_read(&mm->mmap_sem);
6944 precharge = mc.precharge;
6950 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
6952 unsigned long precharge = mem_cgroup_count_precharge(mm);
6954 VM_BUG_ON(mc.moving_task);
6955 mc.moving_task = current;
6956 return mem_cgroup_do_precharge(precharge);
6959 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6960 static void __mem_cgroup_clear_mc(void)
6962 struct mem_cgroup *from = mc.from;
6963 struct mem_cgroup *to = mc.to;
6966 /* we must uncharge all the leftover precharges from mc.to */
6968 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
6972 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6973 * we must uncharge here.
6975 if (mc.moved_charge) {
6976 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
6977 mc.moved_charge = 0;
6979 /* we must fixup refcnts and charges */
6980 if (mc.moved_swap) {
6981 /* uncharge swap account from the old cgroup */
6982 if (!mem_cgroup_is_root(mc.from))
6983 res_counter_uncharge(&mc.from->memsw,
6984 PAGE_SIZE * mc.moved_swap);
6986 for (i = 0; i < mc.moved_swap; i++)
6987 css_put(&mc.from->css);
6989 if (!mem_cgroup_is_root(mc.to)) {
6991 * we charged both to->res and to->memsw, so we should
6994 res_counter_uncharge(&mc.to->res,
6995 PAGE_SIZE * mc.moved_swap);
6997 /* we've already done css_get(mc.to) */
7000 memcg_oom_recover(from);
7001 memcg_oom_recover(to);
7002 wake_up_all(&mc.waitq);
7005 static void mem_cgroup_clear_mc(void)
7007 struct mem_cgroup *from = mc.from;
7010 * we must clear moving_task before waking up waiters at the end of
7013 mc.moving_task = NULL;
7014 __mem_cgroup_clear_mc();
7015 spin_lock(&mc.lock);
7018 spin_unlock(&mc.lock);
7019 mem_cgroup_end_move(from);
7022 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
7023 struct cgroup_taskset *tset)
7025 struct task_struct *p = cgroup_taskset_first(tset);
7027 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7028 unsigned long move_charge_at_immigrate;
7031 * We are now commited to this value whatever it is. Changes in this
7032 * tunable will only affect upcoming migrations, not the current one.
7033 * So we need to save it, and keep it going.
7035 move_charge_at_immigrate = memcg->move_charge_at_immigrate;
7036 if (move_charge_at_immigrate) {
7037 struct mm_struct *mm;
7038 struct mem_cgroup *from = mem_cgroup_from_task(p);
7040 VM_BUG_ON(from == memcg);
7042 mm = get_task_mm(p);
7045 /* We move charges only when we move a owner of the mm */
7046 if (mm->owner == p) {
7049 VM_BUG_ON(mc.precharge);
7050 VM_BUG_ON(mc.moved_charge);
7051 VM_BUG_ON(mc.moved_swap);
7052 mem_cgroup_start_move(from);
7053 spin_lock(&mc.lock);
7056 mc.immigrate_flags = move_charge_at_immigrate;
7057 spin_unlock(&mc.lock);
7058 /* We set mc.moving_task later */
7060 ret = mem_cgroup_precharge_mc(mm);
7062 mem_cgroup_clear_mc();
7069 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
7070 struct cgroup_taskset *tset)
7072 mem_cgroup_clear_mc();
7075 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
7076 unsigned long addr, unsigned long end,
7077 struct mm_walk *walk)
7080 struct vm_area_struct *vma = walk->private;
7083 enum mc_target_type target_type;
7084 union mc_target target;
7086 struct page_cgroup *pc;
7089 * We don't take compound_lock() here but no race with splitting thp
7091 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
7092 * under splitting, which means there's no concurrent thp split,
7093 * - if another thread runs into split_huge_page() just after we
7094 * entered this if-block, the thread must wait for page table lock
7095 * to be unlocked in __split_huge_page_splitting(), where the main
7096 * part of thp split is not executed yet.
7098 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
7099 if (mc.precharge < HPAGE_PMD_NR) {
7103 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
7104 if (target_type == MC_TARGET_PAGE) {
7106 if (!isolate_lru_page(page)) {
7107 pc = lookup_page_cgroup(page);
7108 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
7109 pc, mc.from, mc.to)) {
7110 mc.precharge -= HPAGE_PMD_NR;
7111 mc.moved_charge += HPAGE_PMD_NR;
7113 putback_lru_page(page);
7121 if (pmd_trans_unstable(pmd))
7124 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
7125 for (; addr != end; addr += PAGE_SIZE) {
7126 pte_t ptent = *(pte++);
7132 switch (get_mctgt_type(vma, addr, ptent, &target)) {
7133 case MC_TARGET_PAGE:
7135 if (isolate_lru_page(page))
7137 pc = lookup_page_cgroup(page);
7138 if (!mem_cgroup_move_account(page, 1, pc,
7141 /* we uncharge from mc.from later. */
7144 putback_lru_page(page);
7145 put: /* get_mctgt_type() gets the page */
7148 case MC_TARGET_SWAP:
7150 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
7152 /* we fixup refcnts and charges later. */
7160 pte_unmap_unlock(pte - 1, ptl);
7165 * We have consumed all precharges we got in can_attach().
7166 * We try charge one by one, but don't do any additional
7167 * charges to mc.to if we have failed in charge once in attach()
7170 ret = mem_cgroup_do_precharge(1);
7178 static void mem_cgroup_move_charge(struct mm_struct *mm)
7180 struct vm_area_struct *vma;
7182 lru_add_drain_all();
7184 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
7186 * Someone who are holding the mmap_sem might be waiting in
7187 * waitq. So we cancel all extra charges, wake up all waiters,
7188 * and retry. Because we cancel precharges, we might not be able
7189 * to move enough charges, but moving charge is a best-effort
7190 * feature anyway, so it wouldn't be a big problem.
7192 __mem_cgroup_clear_mc();
7196 for (vma = mm->mmap; vma; vma = vma->vm_next) {
7198 struct mm_walk mem_cgroup_move_charge_walk = {
7199 .pmd_entry = mem_cgroup_move_charge_pte_range,
7203 if (is_vm_hugetlb_page(vma))
7205 ret = walk_page_range(vma->vm_start, vma->vm_end,
7206 &mem_cgroup_move_charge_walk);
7209 * means we have consumed all precharges and failed in
7210 * doing additional charge. Just abandon here.
7214 up_read(&mm->mmap_sem);
7217 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
7218 struct cgroup_taskset *tset)
7220 struct task_struct *p = cgroup_taskset_first(tset);
7221 struct mm_struct *mm = get_task_mm(p);
7225 mem_cgroup_move_charge(mm);
7229 mem_cgroup_clear_mc();
7231 #else /* !CONFIG_MMU */
7232 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
7233 struct cgroup_taskset *tset)
7237 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
7238 struct cgroup_taskset *tset)
7241 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
7242 struct cgroup_taskset *tset)
7248 * Cgroup retains root cgroups across [un]mount cycles making it necessary
7249 * to verify sane_behavior flag on each mount attempt.
7251 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
7254 * use_hierarchy is forced with sane_behavior. cgroup core
7255 * guarantees that @root doesn't have any children, so turning it
7256 * on for the root memcg is enough.
7258 if (cgroup_sane_behavior(root_css->cgroup))
7259 mem_cgroup_from_css(root_css)->use_hierarchy = true;
7262 struct cgroup_subsys mem_cgroup_subsys = {
7264 .subsys_id = mem_cgroup_subsys_id,
7265 .css_alloc = mem_cgroup_css_alloc,
7266 .css_online = mem_cgroup_css_online,
7267 .css_offline = mem_cgroup_css_offline,
7268 .css_free = mem_cgroup_css_free,
7269 .can_attach = mem_cgroup_can_attach,
7270 .cancel_attach = mem_cgroup_cancel_attach,
7271 .attach = mem_cgroup_move_task,
7272 .bind = mem_cgroup_bind,
7273 .base_cftypes = mem_cgroup_files,
7277 #ifdef CONFIG_MEMCG_SWAP
7278 static int __init enable_swap_account(char *s)
7280 if (!strcmp(s, "1"))
7281 really_do_swap_account = 1;
7282 else if (!strcmp(s, "0"))
7283 really_do_swap_account = 0;
7286 __setup("swapaccount=", enable_swap_account);
7288 static void __init memsw_file_init(void)
7290 WARN_ON(cgroup_add_cftypes(&mem_cgroup_subsys, memsw_cgroup_files));
7293 static void __init enable_swap_cgroup(void)
7295 if (!mem_cgroup_disabled() && really_do_swap_account) {
7296 do_swap_account = 1;
7302 static void __init enable_swap_cgroup(void)
7308 * subsys_initcall() for memory controller.
7310 * Some parts like hotcpu_notifier() have to be initialized from this context
7311 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
7312 * everything that doesn't depend on a specific mem_cgroup structure should
7313 * be initialized from here.
7315 static int __init mem_cgroup_init(void)
7317 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
7318 enable_swap_cgroup();
7319 mem_cgroup_soft_limit_tree_init();
7323 subsys_initcall(mem_cgroup_init);